GenABEL tutorial

Copyright 2014, the GenABEL project developers

Copyright 2007-2013, Yurii Aulchenko

August 28, 2014

2

THIS WORK IS LICENSED UNDER THE CREATIVE COMMONS ATTRIBUTION-

SHAREALIKE 3.0 UNPORTED LICENSE. TO VIEW A COPY OF THIS LI-
CENSE, VISIT HTTP://[Link]/LICENSES/BY-SA/3.0/
OR SEND A LETTER TO CREATIVE COMMONS, 444 CASTRO STREET,
SUITE 900, MOUNTAIN VIEW, CALIFORNIA, 94041, USA.
Contents

1 Overview 7
1.1 Download necessary files . . . . . . . . . . . . . . . . . . . . . . . 9

2 Introduction to R 11
2.1 Basic R data types and operations . . . . . . . . . . . . . . . . . 11
2.2 Data frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Exploratory analysis of qualitative and quantitative traits . . . . 28
2.4 Regression analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.5 Answers to exercises . . . . . . . . . . . . . . . . . . . . . . . . . 40

3 Introduction to genetic association analysis in R 47

3.1 Characterisation of genetic data . . . . . . . . . . . . . . . . . . . 47
3.2 Exploring genetic data with library genetics . . . . . . . . . . . 47
3.3 Genetic association analysis . . . . . . . . . . . . . . . . . . . . . 54
3.4 Example association analysis . . . . . . . . . . . . . . . . . . . . 54
3.5 Exercise: Exploring genetic data using library genetics . . . . . 58
3.6 Answers to exercises . . . . . . . . . . . . . . . . . . . . . . . . . 59

4 Introduction to the GenABEL-package 79

4.1 General description of [Link]-class . . . . . . . . . . . . . . 79
4.2 Accessing and modifying phenotypic data . . . . . . . . . . . . . 83
4.3 Sub-setting and coercing [Link] . . . . . . . . . . . . . . . . . 85
4.4 Exploring genetic data . . . . . . . . . . . . . . . . . . . . . . . . 89
4.5 Answers to exercises . . . . . . . . . . . . . . . . . . . . . . . . . 97

5 Genome-wide association analysis 103

5.1 Data descriptives and first round of GWA analysis . . . . . . . . 104
5.2 Genetic data QC . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.3 Finding genetic sub-structure . . . . . . . . . . . . . . . . . . . . 116
5.4 GWA association analysis . . . . . . . . . . . . . . . . . . . . . . 121
5.5 Genome-wide association analysis exercise . . . . . . . . . . . . . 126
5.6 Answers to exercises . . . . . . . . . . . . . . . . . . . . . . . . . 127

6 GWA analysis in presence of stratification: theory 137

6.1 Genetic structure of populations . . . . . . . . . . . . . . . . . . 138
6.1.1 Hardy-Weinberg equilibrium . . . . . . . . . . . . . . . . 139
6.1.2 Inbreeding . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.1.3 Mixture of genetic populations: Wahlund’s effect . . . . . 145

3
4 CONTENTS

6.2 Effects of population structure on standard tests for association . 148

6.2.1 Standard tests for genetic association . . . . . . . . . . . 148
6.2.2 Effects of genetic structure on standard tests . . . . . . . 151
6.2.3 Genomic control . . . . . . . . . . . . . . . . . . . . . . . 154
6.3 Analysis of structured populations . . . . . . . . . . . . . . . . . 157
6.3.1 Structured association . . . . . . . . . . . . . . . . . . . . 158
6.3.2 Mixed models based approach . . . . . . . . . . . . . . . . 160
6.3.3 Estimation of kinship matrix from genomic data . . . . . 162
6.3.4 EIGENSTRAT and related methods . . . . . . . . . . . . 163
6.3.5 Summary: what method to use? . . . . . . . . . . . . . . 165
6.4 Links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

7 GWA in presence of genetic stratification: practice 167

7.1 Analysis with ethnic admixture . . . . . . . . . . . . . . . . . . . 167
7.2 Analysis of family data . . . . . . . . . . . . . . . . . . . . . . . . 172
7.3 Example GWA analysis using family-based data . . . . . . . . . 176
7.4 Exercise: analysis of family data . . . . . . . . . . . . . . . . . . 186
7.5 Answers to exercises . . . . . . . . . . . . . . . . . . . . . . . . . 186

8 Imperfect knowledge about genotypes 193

8.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.2 Input files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
8.2.1 SNP information file . . . . . . . . . . . . . . . . . . . . . 194
8.2.2 Genomic predictor file . . . . . . . . . . . . . . . . . . . . 194
8.2.3 Phenotypic file . . . . . . . . . . . . . . . . . . . . . . . . 195
8.2.4 Optional map file . . . . . . . . . . . . . . . . . . . . . . . 196
8.3 Running an analysis . . . . . . . . . . . . . . . . . . . . . . . . . 196
8.3.1 Basic analysis options . . . . . . . . . . . . . . . . . . . . 197
8.3.2 Advanced analysis options . . . . . . . . . . . . . . . . . . 197
8.3.3 Running multiple analyses at once: [Link] . . . . . 198
8.4 Output file format . . . . . . . . . . . . . . . . . . . . . . . . . . 198
8.5 Preparing input files . . . . . . . . . . . . . . . . . . . . . . . . . 199
8.6 Memory use and performance . . . . . . . . . . . . . . . . . . . . 199
8.7 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
8.7.1 Analysis of population-based data . . . . . . . . . . . . . 199
8.7.2 Analysis of pedigree data . . . . . . . . . . . . . . . . . . 202
8.8 How to cite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204

9 Analysis of imputed data: an example 205

9.1 Analysis of 500 directly typed SNPs . . . . . . . . . . . . . . . . 205
9.2 Analysis of imputed data with ProbABEL-package . . . . . . . . 208
9.3 Analysis of imputed data with MixABEL-package . . . . . . . . . 212
9.4 Answers to exercises . . . . . . . . . . . . . . . . . . . . . . . . . 213

10 Meta-analysis of GWA scans 219

10.1 Standard meta-analysis methods . . . . . . . . . . . . . . . . . . 219
10.2 Exercise: meta-analysis of literature data . . . . . . . . . . . . . 224
10.3 Reporting GWA results for future meta-analysis . . . . . . . . . 225
10.4 Meta-analysis with MetABEL-package . . . . . . . . . . . . . . . 232
10.5 Answers to the exercise . . . . . . . . . . . . . . . . . . . . . . . 235
CONTENTS 5

10.5.1 Exercise 9: . . . . . . . . . . . . . . . . . . . . . . . . . . 238

11 Analysis of selected region 241

11.1 Exploring linkage disequilibrium . . . . . . . . . . . . . . . . . . 241
11.2 Haplotype analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 241
11.3 Analysis of interactions . . . . . . . . . . . . . . . . . . . . . . . 241

A Importing data to GenABEL-package 243

A.1 Converting from preferred format . . . . . . . . . . . . . . . . . . 244
A.2 Converting PLINK tped files . . . . . . . . . . . . . . . . . . . . 247
A.3 Converting linkage-like files . . . . . . . . . . . . . . . . . . . . . 249
A.4 Converting from MACH format . . . . . . . . . . . . . . . . . . . 253
A.5 Converting from text format . . . . . . . . . . . . . . . . . . . . . 253

B GenABEL internals 255

B.1 Internal structure of [Link]-class . . . . . . . . . . . . . . . 255
6 CONTENTS
Chapter 1

Overview

This introduction is outdated: now the GenABEL-package is the project,

the suite, and the package, see [Link]
GenABEL-package is an R library developed to facilitate Genome-Wide Asso-
ciation (GWA) analysis of binary and quantitative traits. GenABEL-package is
implemented as an R library. R is a free, open source language and environment
for general-purpose statistical analysis (available at [Link]
org/). It implements powerful data management and analysis tools. Though it
is not strictly necessary to learn everything about R to run GenABEL-package,
it is highly recommended as this knowledge will improve flexibility and quality
of your analysis.
Originally GenABEL-package was developed to facilitate GWA analysis of
quantitative traits using data coming from extended families and/or collected
form genetically isolated populations. At the same time GenABEL-package im-
plements a large number of procedures used in analysis of population-based
data; it supports analysis of binary and quantitative tarits, and of survival
(time-till-event) data. Most up-to-date information about GenABEL-package
can be found at the web site [Link]
This tutorial was originally written to serve as a set of exercises for the
”Advances in population-based studies of complex genetic disorders” (GE03)
course of the Netherlands Institute of Health Sciences (Nihes).
If you read this tutorial not as a part of the GE03 course, and you are eager
to start with you GWA analysis without reading all the not-so-strictly-necessary
staff, start directly from the section 5 (”Genome-wide association analysis”).
Otherwise, you can start with R basics and simple association analyses us-
ing few SNPs in section 2, ”Introduction to R”. In the next section, 4 (”In-
troduction to the GenABEL-package”) you will learn how to work with the
[Link]-class, which is used to store GWA data in GenABEL-package and
will perform some simple large-scale analyses.
In the next section, 5 (”Genome-wide association analysis”), you will do
quality control of genetic data and do association analysis under realistic con-
ditions. This section is the core of this tutorial.
The section 7 (”GWA in presence of genetic stratification: practice”) is
dedicated to analysis in the presence of population stratification and analysis of
family-based data.
Genetic data imputations are covered in the section ??, ”??”.

7
8 CHAPTER 1. OVERVIEW

The last section, 11 (”Analysis of selected region”), is dedicated to analysis

of haplotype association and analysis of SNP interactions.
Information on importing the data from different file formats to GenABEL-package
is given in appendix A (”Importing data to GenABEL-package”). Answers to
exercises are provided at the end of the respective chapters.
Experienced R users start directly with the section (4, ”Introduction to the
GenABEL-package”).
1.1. DOWNLOAD NECESSARY FILES 9

1.1 Download necessary files

This code needs to be run prior to other parts of tutorial. We reccommend that
prior to any actions you create a new directory, say, ’exercisesGenABEL’, to
keep all of your working tutorial files there.
Start R and make sure that your working directory is set to a proper location.
Your current working directory can be queried by command ’getwd()’. Use
’setwd’ command to set the working directory.
The next lines of code kill the ’RData’ directory if it is present in your
working directory (danger! danger!) to make new clean data installation. Paste
this code into R:
unlink("RData",recursive=TRUE,force=TRUE)
[Link]("RData")
Now, fetch the necessary data from the server. First, define the download
procedure
myDownloads <- function(baseUrl,baseLocal,files) {
for (cFile in files) {
cFileUrl <- paste(baseUrl,cFile,sep="")
cFileLocal <- paste(baseLocal,cFile,sep="")
tryDownload <- try(
[Link](url=cFileUrl,destfile=cFileLocal)
)
if ( is(tryDownload,"try-error") )
stop(paste("can not download",cFileUrl,"into",cFileLocal,":",tryDownload))
}
}
Second, download data files:
baseUrl <- "[Link]
baseLocal <- "RData/"
dataFiles <- c(
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]",
"[Link]")
myDownloads(baseUrl,baseLocal,dataFiles)
That’s it! - now you are fully set to start with the GenABEL tutorial!
10 CHAPTER 1. OVERVIEW
Chapter 2

Introduction to R

In this section we will consider the basic R data types and operations, as well
as tools for the analysis of qualitative and quantitative traits. Only basic R
functionality – the things which are crucial to know before we can proceed to
genetic association analysis – will be covered within this section. If you want
to make most of your data, though, we strongly recommend that you improve
your knowledge of R using books other than this one. A number of excellent
manuals (’An introduction to R’, ’Simple R’, ’Practical Regression and Anova
using R’, and others) is available free of charge from the R project web-site
([Link]
In the first part of this chapter you will learn about the most important
R data types and will learn how to work with R data. Next, we will cover ex-
ploratory data analysis. The chapter will end with an introduction to regression
analysis.

2.1 Basic R data types and operations

In contrast with many other statistical analysis packages, analysis in R is not
based on a graphical user interface, but is command line-based. When you
first start R, a command prompt appears. To get help and overview of R,
type [Link]() on the command line and press enter. This will start your
default internet browser and open the main page of the R documentation.
Let us first use R as a powerful calculator. You can directly operate with
numbers in R. Try multiplying two by three:
> 2 * 3

[1] 6

Other standard arithmetic operations can be performed in a similar manner:

> 2 / 3

[1] 0.6666667

(division)
> 2^3

11
12 CHAPTER 2. INTRODUCTION TO R

[1] 8

(power)
> 2 - 3

[1] -1

(subtraction)
> 2 + 3

[1] 5

(summation)1 .
Mathematical functions, such as square roots, base-10 logarithm, and expo-
nentiation, are available in R as well:
> sqrt(5)

[1] 2.236068

> log10(2.24)

[1] 0.350248

> exp(0.35)

[1] 1.419068

Here, we have computed e to the power of base-10 logarithm of the square

root of the sum of two and three. After each operation, we have rounded the
result to the two digits after the floating point – just to do less typing.
The arithmetic operations and functions can be nested and therefore we can
obtain the above result in one line, and without the 2nd -digit approximation:
> exp(log10(sqrt(2+3)))

[1] 1.418337

R functions include not only the standard mathematical ones, but also a
wide range of statistical function, for example, probability density functions of
many probability distributions. We will make extensive use of these at a later
stage, when computing significance and estimating statistical power.
For any function with a name say ’fun’, help may be obtained by typing
’help(fun)’ (or ?fun) on the command line2 .
R help pages have a standard layout, documenting usage of the function,
explaining function arguments, providing details of implementation and/or us-
age, explaining the value returned by the function, and giving references and
examples of the function use.
Most of the documented functions have examples of their usage at the
end of the ’help’ page, and these examples can be evaluated in R. E.g. try
’example(log10)’.
1 For
a complete list of arithmetic operations try help("+").
2 If
you run R directly from the Linux, OS X or Windows command line (i.e. not in RStudio
or the Windows R environment), you can exit the help by typing q.
2.1. BASIC R DATA TYPES AND OPERATIONS 13

Exercise 1. Explore help for Wilcoxon test

Explore the help page for the Wilcoxon test (function: [Link]) and
answer the following questions:
1. When is the exact Wilcoxon test computed by default?
2. If the default conditions for the exact test are not satisfied, what approx-
imation is used?
If you do not know the exact name for the function you look for, try
’[Link]("query")’, where query is a keyword related to what you are
looking for.

Exercise 2. Finding functions and help pages

Try to find out what are the functions to do the
1. Fisher exact test
2. T-test
One of the important R operations is assignment, which is done with the
’<-’ operator. A (new) variable name should be provided on the left-hand side
of this operator and on the right-hand side, there must be either the name of
an already existing variable or an expression. For example, we if want to assign
the value ’2’ to the variable ’a’, and the value ’3’ to the variable ’b’ we would
use the assignment operator in the following way:

> a <- 2
> b <- 3

Typing the variable name on the R command line will return its value, e.g.

> b

[1] 3

Evaluation of the expression

> exp(log10(sqrt(a+b)))

[1] 1.418337

gives the expected result we have obtained earlier using numerical arguments.
While the variables ’a’ and ’b’ contain single numeric values, variables in
general can be multi-dimensional; a one-dimensional example of such is the
vector (or array). Let us create an example vector and experiment with it:

> v <- c(1, 3, 5, 7, 11)

Here, ’c()’ is a function, which combines its arguments to make a vector. This
vector is then assigned to a variable named ’v’.
Now, let us try different operations with this vector:

> v + 1

[1] 2 4 6 8 12
14 CHAPTER 2. INTRODUCTION TO R

It is easy to see that the result is a vector, which is obtained by adding one to
each element of the original vector v. Other arithmetic operations and mathe-
matical functions behave in the same way, e.g. the operation is performed for
each element of the vector, and the results are returned:

> 1/v

[1] 1.00000000 0.33333333 0.20000000 0.14285714 0.09090909

> log(v)

[1] 0.000000 1.098612 1.609438 1.945910 2.397895

What happens if two vectors are supplied as function arguments? Let us

define a new vector
> ov <- c(1, 2, 3, 4, 5)
and add it to the vector v:
> v + ov

[1] 2 5 8 11 16

You can see that the summation was done element-wise, i.e. the first element of
the result vector is obtained as the sum of the first elements of v and ov, the
second is the sum of the second elements, and so forth.
Other arithmetic operations with two vectors are performed in the same
element-wise manner:
> v * ov

[1] 1 6 15 28 55

(multiplication)
> v^ov

[1] 1 9 125 2401 161051

(power).
The vector operations considered above returned a same-length vector as
output. There are others – statistical and summary – functions which evaluate
a vector as a whole and return a single value as output. For example, to obtain
a sum of elements of a vector, use
> sum(v)

[1] 27

Other examples of such functions are length, returning the number of ele-
ments of a vector, mean, returning the mean, var, returning the variance, etc.:
> length(v)

[1] 5
2.1. BASIC R DATA TYPES AND OPERATIONS 15

> mean(v)

[1] 5.4

> var(v)

[1] 14.8

One of the basic, and probably most used, data operations in R is sub-setting.
This refers to an operation which helps you deriving a subset of the data. Let
us create a short vector and play a bit with sub-setting. This vector will contain
5 simple character strings:

> a <- c("I am element 1", "I am element 2", "I am element 3",
+ "I am element 4", "I am element 5")
> a

[1] "I am element 1" "I am element 2" "I am element 3" "I am element 4"
[5] "I am element 5"

To find out what is the value of the i-th element of this vector, you can
sub-set it by a[i]. For example the 3rd elements is:
> a[3]

[1] "I am element 3"

You can also select a bigger sub-set, e.g. all elements from 2 to 4:
> a[c(2:4)]

[1] "I am element 2" "I am element 3" "I am element 4"

Here, the operation c(2:4) stands for ’combine numbers from 2 to 4 into a
vector’. An equivalent result is obtained by
> a[c(2, 3, 4)]

[1] "I am element 2" "I am element 3" "I am element 4"

We can also easily get disjoint elements; e.g. if you want to retrieve elements
1, 3, and 5, you can do that with
> dje <- c(1, 3, 5)
> dje

[1] 1 3 5

> a[dje]

[1] "I am element 1" "I am element 3" "I am element 5"

One of the very attractive features of R data objects is the possibility to

derive a sub-set based on some condition. Let us consider two vectors, tmphgt,
containing the height of some subjects, and tmpids, containing their identifica-
tion codes (IDs):
16 CHAPTER 2. INTRODUCTION TO R

> tmphgt <- c(150, 175, 182, 173, 192, 168)

> tmphgt

[1] 150 175 182 173 192 168

> tmpids <- c("fem1", "fem2", "man1", "fem3", "man2", "man3")

> tmpids

[1] "fem1" "fem2" "man1" "fem3" "man2" "man3"

Imagine you need to derive the IDs of the people with height over 170 cm.
To do that, we need to combine several operations. First, we should run the
logical function > 170 on the height data:

> vec <- (tmphgt > 170)

> vec

[1] FALSE TRUE TRUE TRUE TRUE FALSE

This returns a logical vector whose elements are ’TRUE’, when a particular
element of the tmphgt satisfies the condition > 170. The returned logical vector,
in turn, can be applied to sub-set any other vector of the same length3 , including
itself. Thus if you want to see the heights in people that are taller than 170 cm,
you can use

> tmphgt[vec]

[1] 175 182 173 192

As you can see, only the elements of tmphgt for which the corresponding value
of vec was ’TRUE’, are returned. In the same manner, the logical vector vec can
be applied to select elements of the vector of IDs:

> tmpids[vec]

[1] "fem2" "man1" "fem3" "man2"

You can combine more than one logical condition to derive sub-sets. For
example, to see what are the IDs of people taller than 170 but shorter than 190
cm, you can use

> vec <- (tmphgt>170 & tmphgt<190)

> vec

[1] FALSE TRUE TRUE TRUE FALSE FALSE

> tmpids[vec]

[1] "fem2" "man1" "fem3"

3 Actually, you can apply it to a longer vector too, and then the logical vector will be

”expanded” to the total length by repeating the original vector head-to-tail. However, we will
not use this in our exercises.
2.1. BASIC R DATA TYPES AND OPERATIONS 17

A better4 way to do logical sub-setting is to use the which() function on top

of the logical vector. This function reports which elements are TRUE. To obtain
the aforementioned result you can run:

> vec <- which(tmphgt>170 & tmphgt<190)

> vec

[1] 2 3 4

> tmpids[vec]

[1] "fem2" "man1" "fem3"

You can see that now vec contains a vector whose elements are the indices of
the elements of tmphgt for which the logical condition holds.
Sub-setting for 2D objects (matrices) is done in a similar manner. Let us
construct a simple matrix and do several sub-setting operations on it:

> a <- matrix(c(11, 12, 13,

+ 21, 22, 23,
+ 31, 32, 33
+ ),
+ nrow=3, ncol=3)
> a

[,1] [,2] [,3]

[1,] 11 21 31
[2,] 12 22 32
[3,] 13 23 33

To obtain the element in the 2nd row and 2nd column, you can use
> a[2, 2]

[1] 22

To access the element from the second row and third column, use
> a[2, 3]

[1] 32

Note that here, the row index (2) comes first, and the column index (3) comes
second.
To obtain the 2 × 2 set of elements contained in upper left corner, you can
do
> a[1:2, 1:2]

[,1] [,2]
[1,] 11 21
[2,] 12 22
4 Because it treats NAs for you
18 CHAPTER 2. INTRODUCTION TO R

1 2 3
1 1 4 7
2 2 5 8
3 3 6 9

Table 2.1: Vector representation of a matrix. Elements in the table are the
vector indices of the matrix elements.

Or you can even get the variables that reside in corners:

> a[c(1, 3), c(1, 3)]

[,1] [,2]
[1,] 11 31
[2,] 13 33

If one of the dimensions is not specified, a complete vector is returned for

this dimension. For example, here we retrieve the first row

> a[1,]

[1] 11 21 31

. . . and the third column

> a[, 3]

[1] 31 32 33

. . . or columns 1 and 3:

> a[, c(1, 3)]

[,1] [,2]
[1,] 11 31
[2,] 12 32
[3,] 13 33

Another way to address elements of a matrix is to use a one-dimensional

index. For example, if you want to access the element in the 2nd row and 2nd
column, instead of

> a[2, 2]

[1] 22

you can use

> a[5]

[1] 22
2.1. BASIC R DATA TYPES AND OPERATIONS 19

This way of accessing the elements of a matrix is based on the fact that
each matrix can be represented as a vector whose elements are numbered con-
secutively: the element in the upper-left corner has index 1, the element in
the second row of the first column has index 2, and the last element in the
bottom-right corner has the maximal value, as shown in Table 2.1.
You can sub-set matrices using logical conditions or indexes like you can
with vectors. For example, if we want to see which elements of a are greater
than 21, we can run

> a > 21

[,1] [,2] [,3]

[1,] FALSE FALSE TRUE
[2,] FALSE TRUE TRUE
[3,] FALSE TRUE TRUE

or, better

> which(a > 21)

[1] 5 6 7 8 9

Note that in the latter case, a vector whose elements give the 1-D indicess of the
matrix, is returned. This vector indicates the elements of matrix a, for which
the condition (a > 21) is satisfied.
You can obtain the values of the matrix’s elements for which the condition
is fulfilled either by

> a[a > 21]

[1] 22 23 31 32 33

or using

> a[which(a > 21)]

[1] 22 23 31 32 33

Once again, the latter method should be preferred. Consider the example
where some elements of the matrix are missing (NA) – a situation which is
common in real data analysis. Let us replace element number 5 with NA and
perform sub-setting operations on the resulting matrix:

> a

[,1] [,2] [,3]

[1,] 11 21 31
[2,] 12 22 32
[3,] 13 23 33

> a[5] <- NA

> a
20 CHAPTER 2. INTRODUCTION TO R

[,1] [,2] [,3]

[1,] 11 21 31
[2,] 12 NA 32
[3,] 13 23 33

> a[a > 21]

[1] NA 23 31 32 33

> a[which(a > 21)]

[1] 23 31 32 33

You can see that when a[a > 21] was used, not only the elements which are
greater than 21 were returned, but also NA was. As a rule, this is not what you
want, and which should be used unless you do want to make some use of the
NA elements.
In this section, we have generated a number of R data objects. Some of these
were numeric (e.g. vector of heights, tmphgt) and some were character, or string
(e.g. vector of study IDs, tmpids). Sometimes you need to figure out what the
class of a certain object is. This can be done using the class() function. For
example,

> tmphgt

[1] 150 175 182 173 192 168

> class(tmphgt)

[1] "numeric"

> tmpids

[1] "fem1" "fem2" "man1" "fem3" "man2" "man3"

> class(tmpids)

[1] "character"

What happens if we try to find out the class of

> a

[,1] [,2] [,3]

[1,] 11 21 31
[2,] 12 NA 32
[3,] 13 23 33

– an object, which contains a matrix?

> class(a)

[1] "matrix"
2.1. BASIC R DATA TYPES AND OPERATIONS 21

Results are expected – we find out that a is a matrix, which is correct. At

the same time, a matrix is an upper-level class, which contains a number of
elements, belonging to some lower-level (e.g. character/numeric) class. To see
what is the class of the matrix elements, try

> a[1, ]

[1] 11 21 31

> class(a[1, ])

[1] "numeric"

which says that elements (at least of the first row) are numeric. Because all
elements of a matrix should have the same class, we can conclude that a is a
matrix containing numeric values.
At this point, it is worthwile inspecting what data objects were created
during our work. This can be done with the ls() command:

> ls()

[1] "a" "b" "dje" "ov" "tmphgt" "tmpids" "v" "vec"

Obviously, this ”list” command is very useful – you will soon find that it is
just too easy to forget the name of a variable which took a long time to create.
Sometimes you may wish to remove some of the data objects because you do
not need then anymore. You can remove an object using the rm() command,
where the names of objects to be deleted are listed as arguments. For example,
to remove the tmphgt and tmpids variables you can use

> rm(tmphgt, tmpids)

If you now look up what data objects are still left in you workspace with the
ls() command

> ls()

[1] "a" "b" "dje" "ov" "v" "vec"

you find that you have successfully deleted tmphgt and tmpids.
At this point, you can exit R by typing q() on the command line and pressing
Enter.

Summary:
• You can get access to the top-level R documentation via the [Link]()
command. To search help for some keyword keywrd, you can use the
[Link](keywrd) command. To get a description of some function
fun, use help(fun).
• You can use R as a powerful calculator
22 CHAPTER 2. INTRODUCTION TO R

• It is possible to get sub-sets of vectors and matrices by specifying an index

value or a logical condition (of the same length as the vector / matrix)
between square brackets ([, ])
• When you obtain an element of a matrix with [i, j], i is the row and j
is the column of the matrix.
• The function which(A) returns the index of the elements of A which are
TRUE
• You can see the objects available in your workspace by using the ls()
command
• Unnecessary objects (say, tmphgt) can be deleted from the workspace
using the rm command, e.g. rm(tmphgt)
• You can leave R using the q() command

Exercise 3. Exploring srdta

In this exercise, you will explore a few vectors representing different data on
study subjects described in hte srdta example data set supplied together with
GenABEL-package. First, you need to load GenABEL-package by typing
> library(GenABEL)
and load the data by
> data(srdta)
The vector containing the study subjects’ sex can be accessed through male(srdta);
this vector’s value is 1 when the corresponding person is male and 0 otherwise.
The vector containing SNP names can be accessed via snpnames(srdta), chro-
mosome ID – through chromosome(srdta) and map – through map(srdta).
Explore these vectors and answer the questions.
1. What is the ID and sex of the first person in the data set?
2. Of the 22nd person?
3. How many males are observed among the first hundred subjects?
4. How many FEMALES are among the 4th hundred?
5. What is the male proportion in the first 1000 people?
6. What is the FEMALE proportion in second 1000 (1001:2000) people?
7. What is name, chromosome and map position of 33rd marker?
8. What is distance between markers 25 and 26?

2.2 Data frames

A data frame is a class of R data which, basically, is a data table. In such
tables, it is usually assumed that rows correspond to subjects (observations) and
2.2. DATA FRAMES 23

columns correspond to variables (characteristics) measured on these subjects.

A nice feature of data frames is that columns (variables) have names, and the
data can be addressed by referencing to these names5 .
We will explore R data frames using the example data set assoc. Start R
with a double-click on the file named [Link]. You can see the names
of the loaded objects by using the ”list” command:
> ls()
[1] "assoc"
Thus, only one object is loaded. The class of this object is:
> class(assoc)
[1] "[Link]"
– a data frame.
The dimensionality of a data frame (or a matrix) can be determined by using
the dim() command:
> dim(assoc)
[1] 250 7
Here, the first number corresponds to the number of rows (subjects) and the
second to the number of columns (variables). Thus, the data frame assoc
contains the data on 250 subjects, who are characterised by 7 variables each.
Let us now figure out what the names are of the 7 variables present in the
data frame. To see what the variable names are, use the command names():
> names(assoc)
[1] "subj" "sex" "aff" "qt" "snp4" "snp5" "snp6"
These variables correspond to the personal identifier (ID, variable subj),
sex, affection status, quantitative trait qt and several SNPs. Each variable can
have its own type (numeric, character, logical), but all variables must have the
same length – thus forming a matrix-like data structure.
A variable from a data frame (say, fram), which has some name (say, nam)
can be accessed through fram$nam. This will return a conventional vector,
containing the values of the variable. For example, to see the affection status
(aff) in the data frame assoc, use
> assoc$aff
[1] 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 1 0 0 1 1 0 0 0 0 0 0
[38] 0 0 1 0 0 1 0 0 0 0 0 0 0 1 1 0 1 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 1 0 0 0 0
[75] 0 0 0 1 1 1 0 0 0 0 0 0 0 0 1 1 0 0 0 0 1 1 0 0 0 0 1 0 0 1 1 0 1 0 0 0 0
[112] 0 0 1 1 0 0 0 0 0 0 0 0 1 0 1 0 0 0 1 0 0 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 1
[149] 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 1
[186] 1 1 1 0 0 1 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0
[223] 0 0 0 0 0 0 1 1 0 0 0 0 0 1 0 0 0 0 1 0 0 0 1 0 0 0 0 0
5 This may also be true for matrices; however, a more fundamental difference is that a

matrix always contains variables of the same data type, e.g. character or numeric, whereas a
data frame may contain variables of different types
24 CHAPTER 2. INTRODUCTION TO R

The aff (affected) variable here codes for a case/control status. Conventi-
nally, cases are coded as 1 and controls as 0. You can also see several ”NA”s,
which denotes a missing observation.

Exercise 4. Exploring assoc

1. Investigate the types of the variables present in data frame assoc. For
each variable, write down the class.
A data frame may be thought of as a matrix which is a collection of (potentily
different-type) vectors. All sub-setting operations discussed before for matrices
are applicable to a data frame, while all operations dicussed for vectors are
applicable to a data frame’s variables.
Thus, as any particular variable present in a data frame is a conventional
vector, its elements can be accessed using the vector’s indices. For example, if
you would like to know what are the ID, sex and affection status for the person
with index 75, you can request
> assoc$subj[75]
[1] 75
> assoc$sex[75]
[1] 1
> assoc$aff[75]
[1] 0
Alternatively, using the matrix-style of sub-setting, you can see all the data
for person 75:
> assoc[75,]
subj sex aff qt snp4 snp5 snp6
75 75 1 0 1.014664 A/B B/A B/B
In the same manner as with matrices, you can get data for e.g. subjects 5 to
15 by
> assoc[5:15,]
subj sex aff qt snp4 snp5 snp6
5 5 0 0 0.1009220 A/B B/A B/A
6 6 1 0 -0.1724321 A/B A/A A/A
7 7 0 0 -0.3378473 B/B A/A A/A
8 8 0 0 -1.7112925 A/A B/B <NA>
9 9 1 0 -0.4815822 A/B B/A B/A
10 10 1 0 1.2281232 A/A B/B B/B
11 11 0 0 0.5993945 A/B B/A B/A
12 12 0 0 1.9792190 A/A B/B B/B
13 13 1 0 1.5435921 A/A B/B B/B
14 14 0 0 -1.6242738 A/B B/A B/A
15 15 0 0 -0.5160331 A/A B/B B/B
2.2. DATA FRAMES 25

The result is actually a new data frame containing data only on people with
index ranging from 5 to 15:
> x <- assoc[5:15,]
> class(x)

[1] "[Link]"

> dim(x)

[1] 11 7

As well as with matrices and vectors, it is possible to sub-set elements of a

data frame based on (a combination of) logical conditions. For example, if you
are interested in people who have qt values over 1.4, you can find out what the
indices of these people are:
> vec <- which(assoc$qt>1.4)
> vec

[1] 12 13 33 41 54 68 72 76 89 106 118 142 156 161 175 181 193 219 241

and then show the complete data with

> assoc$subj[vec]

[1] 12 13 33 41 54 68 72 76 89 106 118 142 156 161 175 181 193 219 241

At the same time, if you only want to check what the IDs of these people are,
try
> assoc$subj[vec]

[1] 12 13 33 41 54 68 72 76 89 106 118 142 156 161 175 181 193 219 241

Or, if we are interested to find what the IDs and the SNP genotypes are of
these people, we can try
> assoc[vec, c(1, 5, 6, 7)]

subj snp4 snp5 snp6

12 12 A/A B/B B/B
13 13 A/A B/B B/B
33 33 A/A B/B B/B
41 41 A/A B/A B/A
54 54 A/B B/A B/A
68 68 A/A B/B B/B
72 72 A/A B/A B/A
76 76 A/B B/A B/A
89 89 A/A B/B B/B
106 106 A/B B/A B/A
118 118 A/B B/A B/A
142 142 A/B B/A B/A
156 156 A/A B/B B/B
26 CHAPTER 2. INTRODUCTION TO R

161 161 A/B B/A B/A

175 175 A/B B/A B/A
181 181 A/B B/A B/A
193 193 A/A B/B B/B
219 219 A/B B/A B/A
241 241 B/B A/A A/A

here, we select people identified by vec in the first dimension (subjects), and by
c(1, 5, 6, 7) we select the first, fifth, sixth and seventh column (variable).
The same result can be obtained using variable names instead of the vari-
ables’ indices. To remind you the variable names can be found with:
> names(assoc)

[1] "subj" "sex" "aff" "qt" "snp4" "snp5" "snp6"

And now make a vector of the variable names of interest and filter the data
based on it:
> namstoshow <- c("subj", "snp4", "snp5", "snp6")
> assoc[vec, namstoshow]

subj snp4 snp5 snp6

12 12 A/A B/B B/B
13 13 A/A B/B B/B
33 33 A/A B/B B/B
41 41 A/A B/A B/A
54 54 A/B B/A B/A
68 68 A/A B/B B/B
72 72 A/A B/A B/A
76 76 A/B B/A B/A
89 89 A/A B/B B/B
106 106 A/B B/A B/A
118 118 A/B B/A B/A
142 142 A/B B/A B/A
156 156 A/A B/B B/B
161 161 A/B B/A B/A
175 175 A/B B/A B/A
181 181 A/B B/A B/A
193 193 A/A B/B B/B
219 219 A/B B/A B/A
241 241 B/B A/A A/A

A more convenient way to access data presented in a data frame is through

”attaching” it to the R search path by
> attach(assoc)
After that, the variables can be accessed directly, e.g.
> subj[75]

[1] 75
2.2. DATA FRAMES 27

instead of assoc$subj[75].
While it is possible to explore the data presented in a data frame using
the sub-setting operations and screen output, and modify certain data elements
using the assignment (”<-”) operation, you can also explore and modify the data
contained in a data frame6 by using the fix() command (e.g. try fix(assoc)).
However, normally this is not necessary.
With attached data frames, a possible complication is that later on you may
have several data frames which contain variables with the same names. The
variable which will be used when you directly use the name would be the one
from the data frame attached last. You can use the detach() function to remove
a certain data frame from the search path, e.g. after
> detach(assoc)
we cannot use a direct reference to the name (try subj[75]) anymore, but have
to use the full path instead:
> assoc$subj[75]

[1] 75

Summary:
• The list of available objects can be viewed with ls(); the class of some
object obj can be examined with class(obj).

• Simple summary statistics for numeric variables can be generated by using

the summary function
• A histogram for some variable var can be generated by hist(var).
• A variable with name name from a data frame frame, can be accessed
through frame$name.

• You can attach a data frame to the search path by attach(frame). Then
the variables contained in this data frame may be accessed directly. To
detach the data frame (because, e.g. , you are now interested in another
data frame), use detach(frame).

Exercise 5. Explore the phenotypic part of srdta

Load the srdta data object supplied with GenABEL by loading the package
with library(GenABEL) and then loading the data with data(srdta). The
srdta object contains a data frame with phenotypes. This data frame may
be accessed through phdata(srdta). Explore this data frame and answer the
questions
1. What is the value of the 4th variable for subject number 75?
6 and also a matrix
28 CHAPTER 2. INTRODUCTION TO R

2. What is the value of variable 1 for person 75? Check the value of this
variable for the first ten people. Can you guess what the first variable
is?
3. What is the sum of variable 2? Can you guess what data variable 2
contains?

2.3 Exploratory analysis of qualitative and quan-

titative traits
Let us now attach the data frame asscoc
> attach(assoc)
and explore it.
Let us first check how many of the subjects are males. In the sex variable,
males are coded with ”1” and females with ”0”. Therefore to see the total
number of males, you can use
> sum(sex==1)

[1] 129

and to determine what is the proportion of males you can use

> sum(sex==1)/length(sex)

[1] 0.516

This way to compute the proportion would only work correctly if there are no
missing observations (lenght() returns the total length of a variable, including
NAs).
Because of the way the males are coded, the same answer is reached by
> mean(sex)

[1] 0.516

However, that would not have worked if the sex was coded differently, e.g. with
”1” for males and ”2” for females.
Let us now try to find out the mean of the quantitative trait qt. By def-
inition, the mean of a variable, say x (with the i-th element denoted as xi )
is
ΣN x i
x̄ = i=1
N
where N is the number of measurements.
If we try to find out the mean of qt by direct use of this formula, we first need
to find out the sum of the elements of qt. The sum() function of R precisely
does the operation we need. However, if we try it
> sum(qt)

[1] -29.79333
2.3. EXPLORATORY ANALYSIS OF QUALITATIVE AND QUANTITATIVE TRAITS29

this returns ”NA”. The problem is that the qt variable contains ”NA”s (try qt
to see these) and then, by default, ”NA” is returned. We can, however, instruct
the sum() function to remove ”NA”s from consideration:
> sum(qt, [Link]=TRUE)

[1] -29.79333

where [Link]=TRUE tells R that missing variables should be be removed (Non-

[Link]=True)7 .
We can now try to compute the mean with
> sum(qt, [Link]=TRUE)/length(qt)

[1] -0.1191733

This result, however, is not correct. The length() function returns the total
length of a vector, which includes ”NA”s as well. Thus we need to compute the
number of elements in qt that are not missing.
For this, we can use R function [Link](). This function returns TRUE if
the supplied argument is missing (NA) and FALSE otherwise. Let us apply this
function to the vector assoc$qt:
> [Link](qt)

[1] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[13] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[25] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[37] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[49] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[61] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[73] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[85] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[97] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[109] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[121] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[133] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[145] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[157] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[169] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[181] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[193] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[205] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[217] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[229] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
[241] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE

Indeed, the 0 missing elements are correctly identified. However, we are

interested in elements which are not missing. To get these, we can use the
logical function NOT (!), which changes all FALSE to TRUE and visa versa:
7 The same argument works for a number of R statistical functions such as mean, median,

var, etc.
30 CHAPTER 2. INTRODUCTION TO R

> 

[1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[16] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[31] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[46] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[61] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[76] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[91] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[106] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[121] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[136] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[151] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[166] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[181] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[196] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[211] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[226] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[241] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE

Thus the number of elements which are not missing8 is

> sum()

[1] 250

Finally, we can compute the mean of the qt with

> sum(qt, [Link]=TRUE)/sum()

[1] -0.1191733

While this way of computing the mean is enlightening in the sense of how
missing values are treated, the same correct result should be normally achieved
by supplying the [Link]=TRUE argument to the mean() function:

> mean(qt, [Link]=TRUE)

[1] -0.1191733

The function table(x) produces a frequency table for the variable x. Thus,
we can use

> table(sex)

sex
0 1
121 129
8A hidden trick here is that arithmetic operations treat TRUE as one and FALSE as zero.
2.3. EXPLORATORY ANALYSIS OF QUALITATIVE AND QUANTITATIVE TRAITS31

which, again, tells us that there are 129 males and 121 females in this data set.
This function excludes missing observations.
Tables of other qualitative variables, such as affection status and SNPs, can
be generated in the same manner.
As with arithmetic operations and mathematical functions, most of the R
operations can be combined within a single line. Let us try to combine logical
conditions and the table() command to check the distribution of number of
affected in men and women separately:
> table(aff[which(sex==1)])

0 1
98 31

> table(aff[which(sex==0)])

0 1
96 25

On the R command line pressing the “up-arrow” button makes the last
typed command re-appear (pressing it one more time will bring you to the
one before the last, so on). This is very handy when you have to repeat the
same analysis of different variables

Exercise 6. Explore assoc

Explore the phenotypic variables present in assoc
1. How many affected and unaffected are present in the data set?
2. What is the proportion of affected?
3. What is the distribution of snp4 (how many different genotype classes
are present and what are the counts)?
Contingency tables for pairs of variables (cross-tables) can be generated in R
using the table command we have used in previous section to explore frequency
distributions. For example, if you want cross-tabulate sex and affection status
in the data frame assoc, you can use
> table(sex, aff)

aff
sex 0 1
0 96 25
1 98 31

Here, the first variable (sex) is presented in rows and the second (affection
status) in columns.
As is usually the case with R, the output may be saved as a new object (of
class ’table’, which is a variety of a matrix):
32 CHAPTER 2. INTRODUCTION TO R

> a <- table(sex, aff)

> class(a)

[1] "table"

> a

aff
sex 0 1
0 96 25
1 98 31

and this object may be analysed further.

For example, we can easily get the number of affected male with
> a[2, 2]

[1] 31

Alternatively, we can analyse the resulting contingency table a with more

complex functions. If we want to see proportions in this table, we can use
> [Link](a)

aff
sex 0 1
0 0.384 0.100
1 0.392 0.124

Needless to say, this is equivalent to

> [Link](table(assoc$sex, assoc$aff))

0 1
0 0.384 0.100
1 0.392 0.124

In the above table, we see what proportion of people belong to four different
classes (affected male, affected female, unaffected male and unaffected female).
We may also be interested in the proportion of males in affected and unaffected.
This may be achieved by
> [Link](a, 2)

aff
sex 0 1
0 0.4948454 0.4464286
1 0.5051546 0.5535714

telling us that 55.4% of affected individuals are male.

Alternatively, we may be interested in the proportion of affected among
males/females. To answer this question, run
> [Link](a, 1)
2.3. EXPLORATORY ANALYSIS OF QUALITATIVE AND QUANTITATIVE TRAITS33

aff
sex 0 1
0 0.7933884 0.2066116
1 0.7596899 0.2403101

telling us that 55.4% of male are affected.

Another useful contingency table analysis function is [Link], which
implements the Fisher Exact Test of independence:
> [Link](a)

Fisher's Exact Test for Count Data

data: a
p-value = 0.547
alternative hypothesis: true odds ratio is not equal to 1
95 percent confidence interval:
0.6409648 2.3156591
sample estimates:
odds ratio
1.213747

Exploration of genetic data within base R, though possible, may be a bit of a

pain. For example, we can easily generate contingency table of SNP5 vs. affected
status:
> a <- table(aff, snp5)
> a

snp5
aff A/A B/A B/B
0 31 88 71
1 9 26 17

We can also look up the proportion of affected among different genotypic groups
> [Link](a, 2)

snp5
aff A/A B/A B/B
0 0.7750000 0.7719298 0.8068182
1 0.2250000 0.2280702 0.1931818

showing that proportion of cases is similar in the ’A/A’ and ’A/B’ genotypic
groups and somewhat decreased in ’B/B’. It is easy to test if this affection is
statistically independent of genotype by doing a χ2 test
> [Link](a)

Pearson's Chi-squared test

data: a
X-squared = 0.3874, df = 2, p-value = 0.8239
34 CHAPTER 2. INTRODUCTION TO R

which gives a (insignificant) genotypic association test with two degrees of free-
dom.
However, testing Hardy-Weinberg equilibrium, testing allelic effects, and
even computation of allelic frequency is not so straightforward. Such specific
genetic tests are implemented in special R libraries, such as genetics and
GenABEL-package and will be covered in later sections of this document.
At this moment we will switch to exploratory analysis of quantitative traits.
We will make use of the srdta data supplied with the GenABEL-package. As you
can remember from an earlier exercise, that library is loaded with library(GenABEL)
and the data are loaded with data(srdta). Then the phenotypic data frame
may be accessed through phdata(srdta).

Exercise 7. Explore phenotypes in srdta

Explore the phenotypic data content of srdta object (phdata(srdta)).
1. How many observations and variables are present in the data frame?
2. What are the classes of these variables?
As was mentioned before, the function summary() generates a summary
statistics for an object. For example, to see the summary for trait qt1, we can
use
> summary( phdata(srdta)$qt1 )
Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
-4.6000 -0.9500 -0.3100 -0.2981 0.3800 3.2000 3

summary is quite a useful function which may operate in different ways for
objects of different classes. Try summary(phdata(srdta)).

With R, it is also easy to explore the data graphically. For example, a

histogram for qt1 may be generated by
> hist( phdata(srdta)$qt1 )
The resulting histogram is shown in Figure 2.1.
In a similar manner scatter-plots may be generated. To see the relation
between qt1 and qt3, you can run
> plot( phdata(srdta)$qt1, phdata(srdta)$qt3 )
The resulting plot is shown in Figure 2.2.
The mean, median, minimum and maximum of the distribution of a trait
may be found using the functions mean, median, min and max, respectively. The
variance and standard deviation can be computed with var and sd.
To compute the correlation between two variables (or all variables in a ma-
trix/data frame), use cor.
In GenABEL-package, there is a special function designed to facilitate phe-
notypic quality control. This function takes the names of variables and a data
frame as an input, and returns summary statistics, list of outliers (using False
Discovery Rate) and graphs.
For example, to do QC of sex, age and qt3, try
2.3. EXPLORATORY ANALYSIS OF QUALITATIVE AND QUANTITATIVE TRAITS35

Histogram of phdata(srdta)$qt1
400
300
Frequency

200
100
0

−4 −2 0 2

phdata(srdta)$qt1

Figure 2.1: Histogram of qt1.

36 CHAPTER 2. INTRODUCTION TO R

● ●
●
6

● ●
● ● ●
● ● ● ● ● ●
●● ●●●●●● ● ●●
●
●● ● ●●● ● ● ●●●
● ●
● ●● ● ● ● ● ●●●
●● ●●● ● ● ● ●●●● ● ●
● ● ●
●● ● ● ●● ●●● ●●
●●
●
●● ●●● ●●●●●●● ● ●● ● ● ●
● ● ●● ● ●● ●● ●
●●●●
●●●● ●●●
●
●●● ●● ●●●● ●● ● ●● ●●
● ●● ●●
● ●● ●●● ●● ●●● ● ● ●● ●
● ●
● ●
● ●●●●
●● ●
●●●●● ●●●●● ● ●● ●●● ● ●●●●
● ●●●● ●●●
●
●●
●● ● ●
●● ●
●● ●●● ●●● ●● ●●●●●● ●●●●
●
●●
●
●● ●●●● ●● ● ●
●
4

●● ● ● ●●● ● ●●
● ●● ●
●●●●
● ●●
●● ●●●●● ●●●● ●●
●●●●●●●●● ●●
●● ●● ●●●● ● ● ●●● ● ● ●●●●●●
● ● ●● ●●
●●
●●●● ● ●● ●
●● ● ● ●●● ● ●
●●●●●●●●●
●●
●
●●
●
●●
●●
●●
●●
●
●
●●●●
●●●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●●
●●
●●
●●●●
●
●●●
●●
●
●●
●●●●●
● ●
●●
●
●●●●●● ● ●● ●
● ●● ●
●● ● ● ● ●●●
●●
● ● ●
●● ● ●
● ●● ●●●●● ●● ● ●
● ●
● ●
●● ● ● ●
phdata(srdta)$qt3

●●● ●
●● ●●
● ●
●●●● ●●●●
● ●●
● ● ●●●●● ●
●● ● ● ● ● ● ●● ●●●● ●
● ● ●
●●● ●●●●
● ●
●
●●●●
●●●●●●●●
●●
●●
●●●●●
●
●
●
●●
●
●
●●
●●●
●
●●
●●
●
●●●● ●
●
●
●
●●
●●●●
●
●●●●●●
●●
●
●●
●●●
●●
●
●
●
●●●
● ●
●●●●● ●●
●●●
● ● ● ●
●
●●●● ●● ● ●●●
●
●●
●●●
●
●●●
●●●
●
●●●●●
●●●●●
●●●
●●
● ●
● ●●●●● ● ●●
●
●● ●● ●
● ● ● ●
●● ●●● ●●● ● ●● ●
●●●● ● ●●●●● ●● ●●●●●●
●
●● ●
●● ●● ●●
●● ●
●● ●● ●● ● ●● ●● ●
●● ● ●
●
● ●●● ●●●
●
●●
●●●●●●●
● ●●●
● ●
● ●
●
●●
● ● ●● ● ● ● ● ● ●
●● ●●
● ●● ●
●● ●
● ●●●
● ● ●
●
●●
●●
●●
●
●●●
●●●
●
●
●
● ●
●●●
●
●
●
● ●●
●
●●●●●
●●
●●●
●
● ●
●
●
●
●
●
●●●●
●
●
●●
●●
●●
●
●●
●
●
●
●
●
● ●●
●●●
●●
● ●●
●
●
●●●●
●●●●
●●●●●● ● ●● ●
●●● ●● ●●●
●●●●
●●●●
●
●●
● ●●●●
●
●●
●●● ●●● ●
●●
●●●
●
●●
●●●
●
●●
●●
●● ●●
●●
●●●●
●
●
●●
●● ●●
●●
●●
● ●●
●●●●● ●
● ●●
●● ●
● ● ● ● ●● ● ● ●●
●
●
●
●●
●● ●● ●●
●
●●●●●●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●●●
●
●●
●●
●
●
●●
●
●
●●●
●●●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●●
●●
●●
●●
●
●
●
●
●●●
●
●
●●
●
●●
●●●●●●
●
●●●●●
● ●●●●
●●●
●● ● ● ● ● ● ●
● ●●●● ●
●●●
●●
●
●
●●●
●●
●●
● ●
● ●
●
●
●● ●
●●
●●
●● ●●●●●
● ●●
●●
●● ●●●●
●
● ●
●●
●● ●
●● ● ●●
● ●●
● ● ●● ● ●●●● ● ●
●
●●●●●● ●
●● ●●● ● ●
●●● ●
●●● ●●
● ●●
●●●● ●
●
●●●●●●
●●●
●● ●●
●●
● ●●●●●●● ●●
●
● ● ●●● ● ●
●
●●●
● ● ●
● ●●
● ●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●●●●
●●
●
●●●
●
●●●
●
●
●
●
●
●
●●●●
●
●●●
●●
●
●●
●
●
●●
●
●●●
●●●
●
●
●
●●
●●●●
●
●
●●●
● ●
●●
●●●●● ●● ● ● ● ●
●
● ●●● ●● ●●●●●
● ● ● ●● ●●
●●● ● ●●
● ● ●● ●●● ● ●● ●
2

●
●●● ● ● ●● ●
●●
●
● ●●
●●● ●●●
●●
●●
●●
●●●
●●●●
●
●●
●●●●
●●●●● ●●
●● ●●●●
● ● ●●
● ●●● ●
●●●
● ●●
● ●●
●● ● ● ● ● ● ●● ● ●●
● ● ● ● ● ●●●● ●
● ● ● ● ●● ● ●●●●●
●
●●●●●
● ● ●●
● ●
●● ●●● ●●●●
●●●
● ●●●●
●●●●●● ●
● ●
●● ●
●
● ●● ●● ●● ● ● ●●
●●
●●
●
●●
●
●
●●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
● ●
●●
●●
●
●●
●●
● ●●●
●
●●
●
●
●
●●
● ●
●
●
●
●●
●●●●
●
●
●●●
● ●●●
●●●
●●●●
●● ●
● ●● ●●
●● ● ● ● ● ●●
● ● ●●●● ● ●
●● ●●● ●
●●
● ●● ● ● ●●
●●
● ●● ●●●●● ●●●●●
● ●● ●●● ●● ●
●●
●●
●●
●●
●●●
●●
●
●●●●
●
●
●● ●●
● ●
●●●●
● ●
●
● ● ●
●● ●●●
● ●●●●
●●●●● ●● ●● ●●
● ● ●● ●●
● ● ●
● ●
●● ●●● ●
●●●● ●● ● ●
●●● ●
●●●● ●
●●●
●●●●
●● ●●
●● ● ● ● ●
● ●
● ● ● ●
● ●● ●●● ●
● ●●
●●●● ●● ● ●
●● ●
● ● ● ●
●●
● ● ●
● ●●● ● ●●●●● ● ●●● ●● ●●●● ● ● ●●● ●
● ● ●● ●●
●● ● ●●●●●
●●●● ●
● ●
● ●
● ● ●
●● ●●
●● ●●
●●● ●
●
●● ● ● ●●● ●●
●●● ● ●● ●● ●●●●● ●●
●
●● ● ●
● ● ●●
●
●
●●●●●● ●●● ●● ● ●●● ● ●
● ●● ● ● ●●● ● ● ●● ●● ● ● ●
● ●
● ● ● ● ●●
●
● ● ● ● ●● ● ● ●
●●
● ●●
0

● ● ●● ● ● ●
● ● ●
● ● ●
● ●
●
−2

−4 −2 0 2

phdata(srdta)$qt1

Figure 2.2: Scatter-plot of qt1 against qt3.

2.3. EXPLORATORY ANALYSIS OF QUALITATIVE AND QUANTITATIVE TRAITS37

●● ●●
30 40 50 60 70

30 40 50 60 70
● ●●●● ●●
●
●
●
●
●
●● ● ●● ●●●
● ● ●●●●● ●
● ● ●●●●● ● ●
phdata(srdta)$age

phdata(srdta)$age
●
●● ●●●● ●● ●● ●
● ●● ●
●● ●
●
●
●
●
●
●
●
●
●● ●
●●●
●●
●
●●
●
●●
●
●
●
●
● ●
●
●●
●
●● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●● ●
●
●
●●
●
● ●● ● ●
●
●●
●●●
● ●●●
●
●
●●●
●●
●●
●●●●
●
●
●●●●●
●●
● ●●
●
●●
●
●●
●
●●
●
●
●●
●●
●
● ●
●
●●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●●●
●●● ●
●
●●
●
●●
●
●● ●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●●●●●
●●●
●
●●
●
●● ●
●● ●
●●
●
●●
●
●●
●
●●
●●
●●
●
●●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●●●
●
●●
●
●
●
●●
●●
●
●●
●
●●●●●
●
●●
●
●● ●● ●
●●●
●
●
●●
●●
●●
●
●●
●
●●●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
●●●
●
●
●
●●
●
●●●
●
●
●●●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●●
●
●
●●●
●
● ●● ●
●
●
●●
● ●● ●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●●
●●
●
●
●
●
●●●●
●●
●●● ●●●
● ● ●
●
●●
●
● ● ●
●
●●
●●
●●
●●
●
●●
●●
●
●●
●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●
●●
●●
●
●
●●
●
●
●
● ●
●●
●
●
●●●
●
●●
●
●
●●● ● ●
●●
●
●● ●
● ●
●
● ●
●●
●
●●
●
●
●●●
●●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●
●●
●
●
●●●
●
●●
●
●
●
●●
●
●
●●
●●●
●●
●
● ●●
●●
●
●●
●● ● ●
●●
●
●●
● ●●● ●●
●●
●
●
●●
●
●
● ●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
● ●
●
●
●
●●
●●
●
●●
●●
●●
●
●●●
●
● ● ● ●
●
●
●●
● ● ●●
●
●●●
●
●●●
●
●●●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●●
●
●●
●
●●
●●
●●
●
●
●●●
●
●
●●
●
●●
●
●
●●
●
●●●
●●
●●
●●
● ●
●
●
●●
●
● ● ● ●
●
●●●
●●
●
●●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●●●
●●
●
●
●●
●
●
●●●
●●
●●
●●●●●●●
●
●
●●
●
● ●●●●●
●●
●●
●●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●●
●●
●
●
●
●
●●
●●
●
●
●
●●
●●
●
●●
●●
●
●
●
●
●
●●
●●
●
●●
●
●
●●●● ●
●
●
●●
●
●
●
●
● ● ●●
●●●●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●● ●●
●●
● ●
●●●
●●
●
●
●
●● ●
●
●●
●
●●
●
● ●
●
●●●●
●
●●
●●
●
● ●●
●
●
●
●●
●●● ●
●●
●
●●●●● ●●
●
●●
●●
●
●● ●●●●● ●● ●
●●
●
●
●●●●●
●●●
●●
●●●
●●
●● ● ●
●
●●● ●
●●●●●
●● ●
●● ● ● ●
● ●
● ●

30 40 50 60 70 −2 0 2 4 6

phdata(srdta)$age phdata(srdta)$qt3

● ●
●● ●●
●
6

6
●●●● ● ●●
● ●●●●● ● ●● ●●
●
●●
phdata(srdta)$qt3

phdata(srdta)$qt3

● ● ●●● ●
● ● ●● ●● ●
●●
●●●●●●
●● ●●
● ●
● ●
●●
●●
●● ●
●●
●● ●●
●●
●● ●
●
●
●
●
●
●
●● ●● ●● ●●
●
●
●●
●
●
●●●
●●
●
●●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
● ●●
●
●● ● ●
●●
●● ●
●
●●
●
●
●
●
●
●
●
● ●●● ● ●●
●●
● ●●●●●
●●
● ●
●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●●
●●
● ●●● ● ●●
●
●
●●
●
● ● ● ●
●●●
● ● ● ● ●● ●
4

●●●●●●● ●●
● ●
● ●
●●●
●
●●
●●
●
●●
●●
●
●●
●
●●
●
●●
●●●●
●●●
●
●●
●●
●●
●●●
●●
● ●●● ●● ● ●
●
●
●●
●●● ●
●●●
● ●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
● ●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●●●
●●
●●
● ● ●●
●
●
●
●
●
●
●
●●●● ●●●●● ●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●●
● ●●
●● ● ●●
●
●
●
●
●
●
●●
●
●
●●●●● ●
●
● ●
●●●
●
●●
●
●
●●
●●
●●
●
●●
●
●●
●●
●
●●
●
●●
●
●●
●●
●
●●
●●
●
●●
●
●●
●●
●●
●
●
●
●●
●●
●
●
●
●●●
●
●●●●●●● ●
●●
●
●
● ●●●●●●
● ●
●●
●
●
●●
●
●
●
●
●●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
● ● ●●
●● ● ●●
●
●
●
●
●
●
●●
●
● ●●●●
●●● ●●●
●●● ●
●●
●
●●
●
●●
●
●●
●●●
●
●●
●
●
●●
●
●
●●
●
●●
●●
●●
●
●
●
●
●●
●
●
●●
●●
●
●●
●
●●
●●●●
●
●●
●●
●
●●●
●●●● ●● ●
●
●
●●
●
●
2

● ●● ● ●
● ●
●
●●
● ●
●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●●●
●●
●●
●
●
●●
●
●●
●●
●
●
●●
●
●
●●
●●
●
●●●
●
●●
●●
●
●
●●●
●●●
●
●
●●●● ● ● ●
●●
●
●
●
●●●● ●●● ●
●●
●
● ●
●
●
●●
●
●●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
● ●
●
●●
●
●
●
●●●● ● ●
●●
●
●
●
●
●
●
●●
●
●
● ●● ●● ●
● ●
● ●
●
●● ●
●
● ●●
●● ●
●●
●
●
●●
●●●
●● ●
●●
● ●●●
●● ●
●
●
●
●●●
●●
●●● ●
●●● ●●●●●
● ●
●
●●
●●
●●
●●●
●
●●
●●
●
●
●●●
●
● ●
●
●●
●●●●
●●
● ●●● ● ●
●
●●
●
●
●
● ● ●●● ● ●●
●● ●●●
● ●●
●●
●
●●●
●
●●
●●
●
●●
●●
●●
● ●●
●●●●
●
●●●
●● ● ●
●●
●
●
●●
● ●●
●●●●●● ●
●●●●
●●
●
●
●
●
●
●
●●
●
● ●
●
●●●●
● ●
●●●●
●●
●
●
●
●
●
●
●●
●
●● ●●● ● ●
●
0

●
●
●● ● ●
●
●
●
●●● ● ●● ●●
●
● ● ●●
−2

−2

● ●

30 40 50 60 70 −2 0 2 4 6

phdata(srdta)$age phdata(srdta)$qt3

Figure 2.3: Quality control graphs for sex, age, qt3

> [Link]( c("sex", "age", "qt3"), phdata(srdta) )

--------------------------------
Trait sex has 2500 measurements
Missing: 0 ( 0 %)
Mean = 0.51 ; s.d. = 0.5
NO outliers discovered for trait sex
--------------------------------
Trait age has 2500 measurements
Missing: 0 ( 0 %)
Mean = 50.0378 ; s.d. = 7.060125
NO outliers discovered for trait age
--------------------------------
Trait qt3 has 2489 measurements
Missing: 11 ( 0.44 %)
Mean = 2.60859 ; s.d. = 1.101154
NO outliers discovered for trait qt3

The corresponding graph is shown in figure 2.3.

Before you start with the exercise: if a function returns unexpected results,
and you are confident that syntax was right, checking the help page is always
38 CHAPTER 2. INTRODUCTION TO R

a good idea!

Exercise 8. Exploring the phenotypic part of srdta

Explore the phdata part of the srdta object
1. How many people have age over 65 years?
2. What is the sex distribution (proportion of males) in the people over 65
years old?
3. What are the mean, median, minimum and maximum ages in the sample?

4. Compare the distribution of qt3 in people younger and older than 65

years. Use the function sd(A) to get standard deviation of A.
5. Produce distributions of different traits. Do you see something special?
6. What is the correlation between qt3 and age?

2.4 Regression analysis

While contingency tables, bi-plots and correlation are powerful tools to analyse
relations between pairs of variables, a more general framework allowing investi-
gation of the relation of an outcome to multiple predictors is regression. In R,
the function lm implements linear regression modelling, and the function glm
implements generalised linear regression. In this section, we will use these two
functions to analyse quantitative and binary outcomes.
You can do linear regression to check if trait qt2 has a relation with sex and
age by

> a <- lm(phdata(srdta)$qt2 ~ phdata(srdta)$age + phdata(srdta)$sex)

The results of this analysis are stored in object ’a’, which has class ’lm’ and
contains many sub-objects:

> class(a)

[1] "lm"

> names(a)

[1] "coefficients" "residuals" "effects" "rank"

[5] "[Link]" "assign" "qr" "[Link]"
[9] "xlevels" "call" "terms" "model"

At this moment you do not need to understand all these sub-objects; the mean-
ingful summary of analysis is produced with

> summary(a)
2.4. REGRESSION ANALYSIS 39

Call:
lm(formula = phdata(srdta)$qt2 ~ phdata(srdta)$age + phdata(srdta)$sex)

Residuals:
Min 1Q Median 3Q Max
-5.65 -1.80 -1.03 -0.31 883.08

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -1.55892 4.41667 -0.353 0.724
phdata(srdta)$age 0.14022 0.08668 1.618 0.106
phdata(srdta)$sex 1.30377 1.22393 1.065 0.287

Residual standard error: 30.59 on 2497 degrees of freedom

Multiple R-squared: 0.001518, Adjusted R-squared: 0.0007181
F-statistic: 1.898 on 2 and 2497 DF, p-value: 0.1501
You can see that qt2 is not associated with age or sex.
As before, to make easy access to your data (basically, to avoid typing
phdata(srdta) before every trait name, you may attach the data to the search
path:
> attach(phdata(srdta))
Then, the above expression to run linear regression analysis simplifies to:
> summary( lm(qt2 ~ age + sex) )
Call:
lm(formula = qt2 ~ age + sex)

Residuals:
Min 1Q Median 3Q Max
-5.65 -1.80 -1.03 -0.31 883.08

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -1.55892 4.41667 -0.353 0.724
age 0.14022 0.08668 1.618 0.106
sex 1.30377 1.22393 1.065 0.287

Residual standard error: 30.59 on 2497 degrees of freedom

Multiple R-squared: 0.001518, Adjusted R-squared: 0.0007181
F-statistic: 1.898 on 2 and 2497 DF, p-value: 0.1501
with the same results.
Analysis of binary outcomes may be performed using glm function, using
binomial family for the error distribution and the link function. For example,
to figure out if your binary trait (bt) is associated with sex and age, you need
to tell that this is binary trait:
> a <- glm(bt ~ age + sex, family="binomial")
> summary(a)
40 CHAPTER 2. INTRODUCTION TO R

Call:
glm(formula = bt ~ age + sex, family = "binomial")

Deviance Residuals:
Min 1Q Median 3Q Max
-1.992 -1.091 -0.444 1.094 1.917

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -4.639958 0.330519 -14.038 < 2e-16 ***
age 0.088860 0.006463 13.749 < 2e-16 ***
sex 0.379593 0.084138 4.512 6.44e-06 ***
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 3450.5 on 2488 degrees of freedom

Residual deviance: 3216.5 on 2486 degrees of freedom
(11 observations deleted due to missingness)
AIC: 3222.5

Number of Fisher Scoring iterations: 4

There is strong association between bt and sex and age. If you want to charac-
terise the strength of association to a binary trait with Odds Ratios, take the
exponents of the regression coefficient. For example, the odds ratio associated
with male is

> exp(0.3796)

[1] 1.4617

2.5 Answers to exercises

Answer of Exercise 1. Explore help for Wilcoxon test

By default (if ’exact’ is not specified), an exact p-value is computed if the
samples contain less than 50 finite values and there are no ties. Otherwise, a
normal approximation is used.

Answer of Exercise 2. Finding functions and help pages

Try [Link]("Fisher") and [Link]("Student t-test"). You will
find that the corresponding functions are [Link] [Link].

Answer of Exercise 3. Exploring srdta

For the first person the id is ”p1” and the sex code is 1 (1=male, 0=female)

> idnames(srdta)[1]
2.5. ANSWERS TO EXERCISES 41

[1] "p1"

> male(srdta)[1]

p1
1

The id for the 22nd person is ”p22” and sex code is 1:

> idnames(srdta)[22]

[1] "p22"

> male(srdta)[22]

p22
1

Among the first 100 subjects, there are 53 males:

> sum(male(srdta)[1:100])

[1] 53

Among the 4th hundred subjects there are 45 females:

> 100 - sum(male(srdta)[301:400])

[1] 45

The male proportion among the first 1000 people is

> mean(male(srdta)[1:1000])

[1] 0.508

The female proportion among the second 1000 people is

> 1 - mean(male(srdta)[1001:2000])

[1] 0.476

Name, chromosome and map position of the 33rd marker are:

> snpnames(srdta)[33]

[1] "rs422"

> chromosome(srdta)[33]

rs422
"1"

> map(srdta)[33]

rs422
105500
42 CHAPTER 2. INTRODUCTION TO R

The map positions for and distance between markers 25 and 26 are:
> pos25 <- map(srdta)[25]
> pos25

rs365
91250

> pos26 <- map(srdta)[26]

> pos26

rs372
92750

> pos26 - pos25

rs372
1500

Answer of Exercise 4. Exploring assoc

Here is an script which automatically explores the classes of variables in assoc:
> for (i in names(assoc)) {
+ cat("Variable '", i ,"' has class '", class(assoc[, i]), "'\n", sep="")
+ }

Variable 'subj' has class 'integer'

Variable 'sex' has class 'numeric'
Variable 'aff' has class 'numeric'
Variable 'qt' has class 'numeric'
Variable 'snp4' has class 'genotypefactor'
Variable 'snp5' has class 'genotypefactor'
Variable 'snp6' has class 'genotypefactor'

In this so-called for-loop the variable i cycles through all names in assoc and
for each of them it uses the cat function to print the name of the variable and
its class. The \n is the code for a new line.

Answer of Exercise 5. Explore the phenotypic part of srdta

Load the data and look at the few first rows of the phenotypic data frame:
> data(srdta)
> phdata(srdta)[1:5, ]

id sex age qt1 qt2 qt3 bt

p1 p1 1 43.4 -0.58 4.46 1.43 0
p2 p2 1 48.2 0.80 6.32 3.90 1
p3 p3 0 37.9 -0.52 3.26 5.05 1
p4 p4 1 53.8 -1.55 888.00 3.76 1
p5 p5 1 47.5 0.25 5.70 2.89 1

The value of the 4th variable of person 75:

2.5. ANSWERS TO EXERCISES 43

> phdata(srdta)[75, 4]

[1] -0.04

The value for variable 1 is

> phdata(srdta)[75, 1]

[1] "p75"

Also, if we check the first 10 elements we see

> phdata(srdta)[1:10, 1]

[1] "p1" "p2" "p3" "p4" "p5" "p6" "p7" "p8" "p9" "p10"

This is the individual ID.

The sum for variable 2 is
> sum(phdata(srdta)[, 2])

[1] 1275

This is the sex variable – so there are 1275 males in the data set.
Answer of Exercise 6. Explore assoc
The number of affected (coded with ’1’) and unaffected (’0’) is
> table(aff)

aff
0 1
194 56

The proportion of unaffected and affected is

> [Link](table(aff))

aff
0 1
0.776 0.224

Distribution of the ’snp4’ is

> t <- table(snp4)
> t

snp4
A/A A/B B/B
109 105 29

> [Link](t)

snp4
A/A A/B B/B
0.4485597 0.4320988 0.1193416
44 CHAPTER 2. INTRODUCTION TO R

Answer of Exercise 7. Explore phenotypes in srdta

Number of people:
> nids(srdta)

[1] 2500

Number of variables:
> length( names( phdata(srdta) ) )

[1] 7

The same – dimensions of phenotypic data frame:

> dim(phdata(srdta))

[1] 2500 7

The class of variables in phenotypic data frame:

> for (i in names(phdata(srdta))) {
+ cat("The class of variable '", i ,"' is '",
+ class(phdata(srdta)[, i]), "'\n", sep="")
+ }

The class of variable 'id' is 'character'

The class of variable 'sex' is 'integer'
The class of variable 'age' is 'numeric'
The class of variable 'qt1' is 'numeric'
The class of variable 'qt2' is 'numeric'
The class of variable 'qt3' is 'numeric'
The class of variable 'bt' is 'integer'

Answer of Exercise 8. Exploring the phenotypic part of srdta

To obtain the number of people with age > 65 y.o., you can use any of the
following
> sum( phdata(srdta)$age > 65 )

[1] 48

> vec <- which( phdata(srdta)$age > 65 )

> length(vec)

[1] 48

To get sex of these people use any of:

> sx65 <- phdata(srdta)$sex[ phdata(srdta)$age > 65 ]
> sx65

[1] 1 1 0 0 0 0 1 1 1 1 0 0 1 1 0 1 1 1 0 0 0 0 1 1 1 0 1 1 1 1 0 1 1 1 0 0 0 0
[39] 1 0 1 0 0 0 0 1 1 1
2.5. ANSWERS TO EXERCISES 45

> sx65 <- phdata(srdta)$sex[vec]

> sx65

[1] 1 1 0 0 0 0 1 1 1 1 0 0 1 1 0 1 1 1 0 0 0 0 1 1 1 0 1 1 1 1 0 1 1 1 0 0 0 0
[39] 1 0 1 0 0 0 0 1 1 1

Thus, number of males is:

> sum(sx65)

[1] 26

To conclude, the proportion of males is 0.541666666666667.

The distributions of qt3 in people younger and older than 65 are:
> summary(phdata(srdta)$qt3[vec])

Min. 1st Qu. Median Mean 3rd Qu. Max.

0.730 2.690 3.480 3.499 4.265 5.840

> sd( phdata(srdta)$qt3[vec], [Link]=TRUE )

[1] 1.128701

> young <- which(phdata(srdta)$age<65)

> summary( phdata(srdta)$qt3[young] )

Min. 1st Qu. Median Mean 3rd Qu. Max. NA's

-1.97 1.83 2.58 2.59 3.35 6.34 11

> sd( phdata(srdta)$qt3[young], [Link]=TRUE )

[1] 1.093374

The Mean, median, min and max of age:

> summary( phdata(srdta)$age )

Min. 1st Qu. Median Mean 3rd Qu. Max.

24.10 45.10 50.00 50.04 54.80 71.60

The histogram for qt2 looks strange (you can generate it using hist(phdata(srdta)$qt2)):
it seems there are a few very strong outliers. You can also see that with summary:
> summary( phdata(srdta)$qt2 )

Min. 1st Qu. Median Mean 3rd Qu. Max.

0.000 4.220 5.045 6.122 5.910 888.000
46 CHAPTER 2. INTRODUCTION TO R
Chapter 3

Introduction to genetic
association analysis in R

When analyzing several (dozens of) SNPs, facilities of base R are sufficient
and efficient for data storage and analysis. Few specific test, such as these
of Hardy-Weinberg Equilibrium (HWE) and Linkage Disequilibrium (LD), are
implemented in different libraries, e.g. genetics and GenABEL-package.
In this section, we will describe library genetics and will make use of it to
guide you through simple genetic analysis exercise using a small example data
set. In the last part, you will investigate a bigger data set as based on the
knowledge obtained in the first part, and will answer the questions.

3.1 Characterisation of genetic data

3.2 Exploring genetic data with library genetics
Library genetics was written by Gregory R. Warnes to facilitate analysis of
genetic data in R. This library

• Implements genetic analysis tests, such as test for Hardy-Weinberg equi-

librium and Linkage disequilibrium.

• Implements new data classes, such as genotype, haplotype and [Link].

• Modifies default R functions, such as summary and plot to correctly anal-

yse and present these new classes.

• Facilitates export of the data from R to the formats supported by such

genetic analysis packages as GenePop and QTDT.

Start R by double-click on the file [Link]. THIS FILE IS THE

SAME AS THE [Link], so I load and patch it below Load
library genetics, which we will need for testing HWE and computations of LD
by

> library(genetics)

47
48CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

The file you have loaded contains single data frame assocg. Let us load
> #load("RData/[Link]")
> # load assocbase data and convert snps into proper 'genetics' format
> load("RData/[Link]")
> assocg <- assoc
> assocg$snp4 <- [Link](assocg$snp4)
> assocg$snp5 <- [Link](assocg$snp5)
> assocg$snp6 <- [Link](assocg$snp6)
and briefly explore it:
> class(assocg)
[1] "[Link]"
> names(assocg)
[1] "subj" "sex" "aff" "qt" "snp4" "snp5" "snp6"
> dim(assocg)
[1] 250 7
You can see that assocg looks remarkably similar to the previously explored
data frame assoc (section 2.2, page 22). Indeed, they are almost equivalent.
Let us present the data for the subjects 5 to 15 and compare this output to that
presented on page 25:
> assocg[5:15,]
subj sex aff qt snp4 snp5 snp6
5 5 0 0 0.1009220 A/B B/A B/A
6 6 1 0 -0.1724321 A/B A/A A/A
7 7 0 0 -0.3378473 B/B A/A A/A
8 8 0 0 -1.7112925 A/A B/B <NA>
9 9 1 0 -0.4815822 A/B B/A B/A
10 10 1 0 1.2281232 A/A B/B B/B
11 11 0 0 0.5993945 A/B B/A B/A
12 12 0 0 1.9792190 A/A B/B B/B
13 13 1 0 1.5435921 A/A B/B B/B
14 14 0 0 -1.6242738 A/B B/A B/A
15 15 0 0 -0.5160331 A/A B/B B/B
The data are identical. However, the SNP data presented in the new data frame
have special class genotype, as implemented in genetics library:
> class(assocg$snp4)
[1] "genotype" "factor"
Previously, the SNP genotypes were coded as characters. This new way of
presentation allows library genetics to recognise the SNP data as genetic and
analyse them accordingly.
Let us attach the assocg data frame and explore what data analysis advan-
tages are achieved by application of library genetics.
3.2. EXPLORING GENETIC DATA WITH LIBRARY GENETICS 49

> attach(assocg)
As we noted in section 2.2, testing Hardy-Weinberg equilibrium, testing
allelic effects, and even computation of allelic frequency is not so straightforward
in base R. These tests, are, however, easy with library genetics. To see the
allelic frequencies and other summary statistics for a SNP, you can use
> summary(snp4)

Number of samples typed: 243 (97.2%)

Allele Frequency: (2 alleles)

Count Proportion
A 323 0.66
B 163 0.34
NA 14 NA

Genotype Frequency:
Count Proportion
B/B 29 0.12
A/B 105 0.43
A/A 109 0.45
NA 7 NA

Heterozygosity (Hu) = 0.4467269

Poly. Inf. Content = 0.3464355

To check these characteristics in controls and cases separately, you can use
> summary(snp4[aff==0])

Number of samples typed: 190 (97.9%)

Allele Frequency: (2 alleles)

Count Proportion
A 255 0.67
B 125 0.33
NA 8 NA

Genotype Frequency:
Count Proportion
B/B 22 0.12
A/B 81 0.43
A/A 87 0.46
NA 4 NA

Heterozygosity (Hu) = 0.4426469

Poly. Inf. Content = 0.3440288

> summary(snp4[aff==1])
50CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

Number of samples typed: 53 (94.6%)

Allele Frequency: (2 alleles)

Count Proportion
A 68 0.64
B 38 0.36
NA 6 NA

Genotype Frequency:
Count Proportion
B/B 7 0.13
A/B 24 0.45
A/A 22 0.42
NA 3 NA

Heterozygosity (Hu) = 0.4643306

Poly. Inf. Content = 0.3541731

Let us check if HWE holds for the SNPs described in this data frame. We
can do exact test for HWE by
> [Link](snp4)

Exact Test for Hardy-Weinberg Equilibrium

data: snp4
N11 = 109, N12 = 105, N22 = 29, N1 = 323, N2 = 163, p-value = 0.666

If you want to check HWE using controls only, you can do it by

> [Link](snp4[aff==0])

Exact Test for Hardy-Weinberg Equilibrium

data: snp4[aff == 0]
N11 = 87, N12 = 81, N22 = 22, N1 = 255, N2 = 125, p-value = 0.6244

Let us check if the there is LD between snp4 and snp5:

> LD(snp4,snp5)

Pairwise LD
-----------
D D' Corr
Estimates: 0.2009042 0.9997352 0.8683117

X^2 P-value N
LD Test: 354.3636 0 235

The output shows results of the test for significance of LD, and estimates of the
magnitude of LD (D′ and correlation, r). To obtain r2 , you can either square
the correlation manually
3.2. EXPLORING GENETIC DATA WITH LIBRARY GENETICS 51

> 0.8683117*0.8683117

[1] 0.7539652

or simply ask LD() to report it by

> LD(snp4,snp5)$"R^2"

[1] 0.7539652

The latter command is possible because the LD() function actually computes
more things than it reports. This is quite common for R functions. You can
apply names() function to the analysis objects to see (at least part of ) what
was actually computed. Try
> ld45 <- LD(snp4,snp5)
and check what are the sub-objects contained in this analysis object
> names(ld45)
[1] "call" "D" "D'" "r" "R^2" "n" "X^2"
[8] "P-value"
Any of these variables can be accessed through object$var syntax, e.g. to
check D ′ we can use
> ld45$"D'"
[1] 0.9997352

To check LD for more that two SNPs, we can compute an LD analysis object
by

> ldall <- LD([Link](snp4,snp5,snp6))

and later check

> ldall$"P-value"

snp4 snp5 snp6

snp4 NA 0 0
snp5 NA NA 0
snp6 NA NA NA

to see significance,

> ldall$"D'"

snp4 snp5 snp6

snp4 NA 0.9997352 0.8039577
snp5 NA NA 0.9997231
snp6 NA NA NA

for D′ and

> ldall$"R^2"
52CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

1.2
1.0
0.8
0.6
0.4
0.2
0.0
−0.2

−0.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2

Figure 3.1: r2 plot for snp4, snp5 and snp6

snp4 snp5 snp6

snp4 NA 0.7539652 0.5886602
snp5 NA NA 0.8278328
snp6 NA NA NA

for r2 .
You can also present e.g. r2 matrix as a plot by
> image(ldall$"R^2")
A more neat way to present it requires specification of the set of threshold
(break points) and colors to be used (you do not need to try this example if you
do not want):
> image(ldall$"R^2",breaks=c(0.5,0.6,0.7,0.8,0.9,1),col=[Link](5))
Resulting plot is shown at figure 3.1.

For any R command, you can get help by typing help(command). Try
help(image) if you are interested to understand what are ”breaks” and ”col”;
or try help([Link]) to figure this color schema out.

Similar to our HWE checks, we may want to compute (and compare) LD in

cases and controls separately:
3.2. EXPLORING GENETIC DATA WITH LIBRARY GENETICS 53

1.2
1.0
0.8
0.6
0.4
0.2
0.0
−0.2

−0.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2

Figure 3.2: r2 plot for snp4, snp5 and snp6. Above diagonal: LD in cases;
below: controls

> ldcases <- LD([Link](snp4,snp5,snp6)[aff==1,])

> ldcases$"R^2"

snp4 snp5 snp6

snp4 NA 0.7615923 0.6891558
snp5 NA NA 0.8943495
snp6 NA NA NA

> ldcontr <- LD([Link](snp4,snp5,snp6)[aff==0,])

> ldcontr$"R^2"

snp4 snp5 snp6

snp4 NA 0.7512458 0.5616395
snp5 NA NA 0.8075894
snp6 NA NA NA

and even present it results for cases and controls on the same graph (you do not
need to produce this graph, which is presented at the figure 3.2):

> image(ldcases$"R^2",breaks=c(0.5,0.6,0.7,0.8,0.9,1),col=[Link](5))
> image(t(ldcontr$"R^2"),breaks=c(0.5,0.6,0.7,0.8,0.9,1),col=[Link](5),add=T)
54CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

3.3 Genetic association analysis

3.4 Example association analysis
Now, after we have described genetic and phenotypic data separately, we are
ready to test association between these two. In previous sections, we showed
that association between a binary trait and genotype may be analysed using
contingency tables (functions table, [Link], [Link], etc.). The
association between a quantitative trait and genotype may be done using cor-
relations, T-test, etc.
However, a more flexible analysis is possible when using regression modelling.
First, we will investigate relation between the quantitative trait qt and the SNPs
by using linear regression
> mg <- lm(qt~snp4)
The lm command fits linear regression model to the data and returns an analysis
object. The summary of analysis may be generated with
> summary(mg)

Call:
lm(formula = qt ~ snp4)

Residuals:
Min 1Q Median 3Q Max
-2.63700 -0.62291 -0.01225 0.58922 3.05561

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.081114 0.092517 -0.877 0.382
snp4A/B -0.108366 0.132079 -0.820 0.413
snp4B/B -0.006041 0.201820 -0.030 0.976

Residual standard error: 0.9659 on 240 degrees of freedom

(7 observations deleted due to missingness)
Multiple R-squared: 0.003049, Adjusted R-squared: -0.005259
F-statistic: 0.367 on 2 and 240 DF, p-value: 0.6932

From the summary output, it is clear that the model assumes arbitrary (esti-
mated) effects of the genotypes AA, AB and BB. Neither effect of AB nor BB
is significant in this case. The global test on two degrees of freedom (bottom of
the output) is also not significant.
If you want to include some covariate into your model, e.g. sex, you can
easily do that by adding the term to the formula:
> summary(lm(qt~sex+snp4))

Call:
lm(formula = qt ~ sex + snp4)

Residuals:
3.4. EXAMPLE ASSOCIATION ANALYSIS 55

Min 1Q Median 3Q Max

-2.66442 -0.62417 -0.00875 0.59705 3.08086

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.110298 0.115260 -0.957 0.340
sex 0.053018 0.124493 0.426 0.671
snp4A/B -0.104429 0.132628 -0.787 0.432
snp4B/B -0.002452 0.202340 -0.012 0.990

Residual standard error: 0.9676 on 239 degrees of freedom

(7 observations deleted due to missingness)
Multiple R-squared: 0.003805, Adjusted R-squared: -0.0087
F-statistic: 0.3043 on 3 and 239 DF, p-value: 0.8223
You can also allow for interaction by using the ”*” operator
> summary(lm(qt~sex*snp4))
Call:
lm(formula = qt ~ sex * snp4)

Residuals:
Min 1Q Median 3Q Max
-2.57049 -0.64596 -0.00264 0.61094 3.01970

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.20579 0.13834 -1.487 0.138
sex 0.22649 0.18647 1.215 0.226
snp4A/B 0.05222 0.19024 0.274 0.784
snp4B/B 0.18071 0.28576 0.632 0.528
sex:snp4A/B -0.30191 0.26566 -1.136 0.257
sex:snp4B/B -0.35508 0.40531 -0.876 0.382

Residual standard error: 0.9684 on 237 degrees of freedom

(7 observations deleted due to missingness)
Multiple R-squared: 0.01041, Adjusted R-squared: -0.01047
F-statistic: 0.4984 on 5 and 237 DF, p-value: 0.7773
Note that both main effects of sex and snp4, and also effects of interaction are
estimated in this model.
Of interest in genetic studies may be three other models: additive, dominant
and recessive.
The additive model assumes that the difference between mean trait’s values
between ’AA’ and ’BB’ is twice the difference between ’AA’ and ’BB’, that is
the mean value of the trait for heterozygous genotypes is right in between the
two homozygotes. To test additive model, we first need to recode the predictor
(genotype) as a numeric factor to be used as covariate. This can be easy done
with function [Link]:
> add4 <- [Link](snp4)-1
56CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

We can check if recoding was done correctly by producing the table

> table(snp4,add4)

add4
snp4 0 1 2
A/A 109 0 0
A/B 0 105 0
B/B 0 0 29

Now to test the additive model run

> summary(lm(qt~add4))

Call:
lm(formula = qt ~ add4)

Residuals:
Min 1Q Median 3Q Max
-2.54813 -0.62104 -0.02754 0.60584 3.00652

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.10476 0.08710 -1.203 0.230
add4 -0.03563 0.09133 -0.390 0.697

Residual standard error: 0.9651 on 241 degrees of freedom

(7 observations deleted due to missingness)
Multiple R-squared: 0.0006313, Adjusted R-squared: -0.003516
F-statistic: 0.1522 on 1 and 241 DF, p-value: 0.6968

The model assuming dominant action of the ’A’ allele means that the means
of genotypes ’AA’ and ’AB’ are the same. This is equivalent to the model of
recessive action of ’B’ allele. To code SNP4 according to this model, we can use
function replace:

> dom4 <- add4

> dom4[dom4==2] <- 1
> table(snp4,dom4)

dom4
snp4 0 1
A/A 109 0
A/B 0 105
B/B 0 29

To test association with a binary outcome, we will use function glm with
binomial family:

> summary(glm(aff~snp4,family="binomial"))
3.4. EXAMPLE ASSOCIATION ANALYSIS 57

Call:
glm(formula = aff ~ snp4, family = "binomial")

Deviance Residuals:
Min 1Q Median 3Q Max
-0.7433 -0.7204 -0.6715 -0.6715 1.7890

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.3749 0.2386 -5.761 8.35e-09 ***
snp4A/B 0.1585 0.3331 0.476 0.634
snp4B/B 0.2297 0.4952 0.464 0.643
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 254.91 on 242 degrees of freedom

Residual deviance: 254.58 on 240 degrees of freedom
(7 observations deleted due to missingness)
AIC: 260.58

Number of Fisher Scoring iterations: 4

To make a test of global significance of the SNP effect, you can use

> anova(glm(aff~snp4,family="binomial"),test="Chisq")

Analysis of Deviance Table

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 242 254.91
snp4 2 0.32894 240 254.58 0.8483

In the manner similar to that described for quantitative traits, additive and
dominance/recessive models can be tested by proper coding of the genotypic
variable, e.g. to test the additive model, use

> summary(glm(aff~[Link](snp4),family="binomial"))

Call:
glm(formula = aff ~ [Link](snp4), family = "binomial")

Deviance Residuals:
58CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

Min 1Q Median 3Q Max

-0.7549 -0.7139 -0.6747 -0.6747 1.7842

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.4913 0.4164 -3.581 0.000342 ***
[Link](snp4) 0.1272 0.2268 0.561 0.574994
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 254.91 on 242 degrees of freedom

Residual deviance: 254.60 on 241 degrees of freedom
(7 observations deleted due to missingness)
AIC: 258.6

Number of Fisher Scoring iterations: 4

Now you have learned all commands necessary to answer the questions of
the next section.
Exit R by typing q() command (do not save image) and and proceed to the
self exercise.

3.5 Exercise: Exploring genetic data using li-

brary genetics
Start R by double-click over the file [Link] (Windows) or by changing
to the directory containing the data, starting R and loading the data set with
load("[Link]") (Linux). Explore the data frame present and answer
the questions.

Ex. 1 — How many SNPs are described in this data frame?

Ex. 2 — What is the frequency (proportion) of snp1 allele ’A’ ?

Ex. 3 — What is its frequency of ’A’ in affected (aff==1)?

Ex. 4 — How many cases and controls are present in the data set?

Ex. 5 — If all subjects are used to test HWE, are there any SNPs out of HWE
at nominal P ≤ 0.05? Which ones?

Ex. 6 — If only controls are used to test the SNPs which are out of HWE in
total sample, are these still out of HWE?

Ex. 7 — Which SNP pairs are in strong LD (r2 ≥ 0.8)?

Ex. 8 — For SNPs in strong LD, what is r2 for separate samples of cases and
controls?
3.6. ANSWERS TO EXERCISES 59

Ex. 9 — Is there significant association between affection status and sex?

What is P -value for association?
Ex. 10 — Is association between the disease and qt significant?

Ex. 11 — Which SNPs are significantly associated with affection status at

nominal p-value ≤ 0.05? Use general genotypic (2 d.f.) model.

Ex. 12 — Test association between aff and snp5 and snp10, allowing for the
SNPs interaction effect. Use arbitrary (not an additive) model. Do you observe
significant interaction? How can you describe the model of concert action of
snp5 and snp10?

Ex. 13 — Test for association between the quantitative trait qt and SNPs
1-10 using additive model. Which SNPs are associated at nominal P ≤ 0.05?

Ex. 14 — OPTIONAL, difficulty is medium, but may be time-consuming.

If you adjust the analysis under additive model for sex, how do the findings
change? Before doing the exercise, please check the answer to previous exer-
cise – it shows a quick way to do testing for all 10 SNPs.

Ex. 15 — Which SNPs are associated with the quantitative trait qt at nom-
inal P ≤ 0.05 when general genotypic (2 d.f. test) model is used?

Ex. 16 — ADVANCED: How can you describe the model of action of the
significant SNPs? Test if the data are compatible with additive/dominant/recessive
model.

3.6 Answers to exercises

Answer (Ex. 1) — The answer is 10 snps:

> attach(popdat)
> names(popdat)
[1] "subj" "sex" "aff" "qt" "snp1" "snp2" "snp3" "snp4" "snp5"
[10] "snp6" "snp7" "snp8" "snp9" "snp10"

Answer (Ex. 2) — The frequency of ’A’ in all subjects is 0.73:

> summary(snp1)
Number of samples typed: 2374 (95%)

Allele Frequency: (2 alleles)

Count Proportion
A 3462 0.73
B 1286 0.27
NA 252 NA
60CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

Genotype Frequency:
Count Proportion
B/B 199 0.08
A/B 888 0.37
A/A 1287 0.54
NA 126 NA

Heterozygosity (Hu) = 0.3950646

Poly. Inf. Content = 0.3169762

Answer (Ex. 3) — The frequency of A in affected subjects is 0.7:

> summary(snp1[aff==1])
Number of samples typed: 519 (94.5%)

Allele Frequency: (2 alleles)

Count Proportion
A 729 0.7
B 309 0.3
NA 60 NA

Genotype Frequency:
Count Proportion
B/B 48 0.09
A/B 213 0.41
A/A 258 0.50
NA 30 NA

Heterozygosity (Hu) = 0.4185428

Poly. Inf. Content = 0.3307192

Answer (Ex. 4) — There are 549 cases and 1951 controls:

> table(aff)
aff
0 1
1951 549

Answer (Ex. 5) — Only SNP 1 is out of HWE in the total sample. Here is
a sciript testing all SNPs (no need to reproduce that, just check the results):
> for (i in 1:10) {
+ snpname <- paste("snp",i,sep="")
+ cat("HWE P-value for SNP",snpname,"is",[Link](get(snpname))$[Link],"\n")
+ }
HWE P-value for SNP snp1 is 0.01083499
HWE P-value for SNP snp2 is 1
3.6. ANSWERS TO EXERCISES 61

HWE P-value for SNP snp3 is 0.4197772

HWE P-value for SNP snp4 is 0.8960298
HWE P-value for SNP snp5 is 0.2960967
HWE P-value for SNP snp6 is 0.5207056
HWE P-value for SNP snp7 is 0.6284575
HWE P-value for SNP snp8 is 0.1309458
HWE P-value for SNP snp9 is 0.4457363
HWE P-value for SNP snp10 is 0.7897327

Answer (Ex. 6) — Yes, SNP one is out of HWE also in controls:

> [Link](snp1[aff==0])
Exact Test for Hardy-Weinberg Equilibrium

data: snp1[aff == 0]
N11 = 1029, N12 = 675, N22 = 151, N1 = 2733, N2 = 977, p-value =
0.008393

Answer (Ex. 7) — SNP pairs 4-5 and 5-6 have r2 ≥ 0.8:

> LD(popdat[,5:14])$"R^2"
snp1 snp2 snp3 snp4 snp5 snp6 snp7 snp8 snp9 snp10
snp1 NA 0.016 0.235 0.206 0.258 0.227 0.152 0.117 0.090 0.000
snp2 NA NA 0.004 0.004 0.005 0.004 0.000 0.000 0.000 0.000
snp3 NA NA NA 0.602 0.457 0.346 0.641 0.031 0.042 0.001
snp4 NA NA NA NA 0.803 0.650 0.729 0.027 0.037 0.002
snp5 NA NA NA NA NA 0.874 0.586 0.034 0.046 0.002
snp6 NA NA NA NA NA NA 0.670 0.030 0.040 0.002
snp7 NA NA NA NA NA NA NA 0.020 0.027 0.003
snp8 NA NA NA NA NA NA NA NA 0.002 0.000
snp9 NA NA NA NA NA NA NA NA NA 0.001
snp10 NA NA NA NA NA NA NA NA NA NA

Answer (Ex. 8) — For controls,

> #LD(popdat[aff==0,8:10])$"R^2"
> LD([Link](snp4,snp5,snp6)[aff==0,])$"R^2"
snp4 snp5 snp6
snp4 NA 0.806591 0.6419715
snp5 NA NA 0.8661005
snp6 NA NA NA
For cases,
> #LD(popdat[aff==1,8:10])$"R^2"
> LD([Link](snp4,snp5,snp6)[aff==1,])$"R^2"
snp4 snp5 snp6
snp4 NA 0.7951475 0.6773275
62CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

snp5 NA NA 0.9083237
snp6 NA NA NA
Note that the fact that LD is higher in cases may mean nothing because the
estimates of LD are biased upwards with smaller sample sizes. For example in a
small sample (5 people) of controls we expect even higher LD because of strong
upward bias:
> LD(popdat[which(aff==0)[1:5],8:10])$"R^2"
snp4 snp5 snp6
snp4 NA 0.9995876 0.9995876
snp5 NA NA 0.9995876
snp6 NA NA NA
More elaborate methods, such as that by Zaykin et al. (2006), are required to
contrast LD between sample of unequal size.

Answer (Ex. 9) — There is no significant association:

> t <- table(aff,sex)
> t
sex
aff 0 1
0 973 978
1 260 289
> [Link](t)
Fisher's Exact Test for Count Data

data: t
p-value = 0.3104
alternative hypothesis: true odds ratio is not equal to 1
95 percent confidence interval:
0.9107753 1.3430565
sample estimates:
odds ratio
1.105811
> summary(glm(aff~sex,family=binomial()))
Call:
glm(formula = aff ~ sex, family = binomial())

Deviance Residuals:
Min 1Q Median 3Q Max
-0.7196 -0.7196 -0.6882 -0.6882 1.7644

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.31970 0.06981 -18.90 <2e-16 ***
sex 0.10062 0.09673 1.04 0.298
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
3.6. ANSWERS TO EXERCISES 63

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 2632.0 on 2499 degrees of freedom

Residual deviance: 2630.9 on 2498 degrees of freedom
AIC: 2634.9

Number of Fisher Scoring iterations: 4

Answer (Ex. 10) — There is no significant association:

> summary(glm(aff~qt,family=binomial()))
Call:
glm(formula = aff ~ qt, family = binomial())

Deviance Residuals:
Min 1Q Median 3Q Max
-0.7326 -0.7079 -0.7012 -0.6905 1.7675

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.26769 0.04832 -26.238 <2e-16 ***
qt -0.02514 0.04862 -0.517 0.605
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 2632.0 on 2499 degrees of freedom

Residual deviance: 2631.7 on 2498 degrees of freedom
AIC: 2635.7

Number of Fisher Scoring iterations: 4

Answer (Ex. 11) — SNPs 5 and 10 are significantly associated:

> for (i in 1:10) {
+ snpname <- paste("snp",i,sep="")
+ cat("\nTesting association between aff and SNP",snpname,":\n")
+ print(anova(glm(aff~get(snpname),family=binomial),test="Chisq"))
+ #print(summary(lm(qt~get(snpname)))$coef)
+ }
Testing association between aff and SNP snp1 :
Analysis of Deviance Table

Model: binomial, link: logit

Response: aff
64CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2373 2493.4
get(snpname) 2 5.4094 2371 2488.0 0.06689 .
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Testing association between aff and SNP snp2 :

Analysis of Deviance Table

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2373 2485.8
get(snpname) 1 0.29367 2372 2485.5 0.5879

Testing association between aff and SNP snp3 :

Analysis of Deviance Table

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2377 2503.0
get(snpname) 2 2.6087 2375 2500.4 0.2714

Testing association between aff and SNP snp4 :

Analysis of Deviance Table

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2389 2519.1
3.6. ANSWERS TO EXERCISES 65

get(snpname) 2 5.2755 2387 2513.8 0.07152 .

---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Testing association between aff and SNP snp5 :

Analysis of Deviance Table

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2382 2440.4
get(snpname) 2 9.2395 2380 2431.2 0.009855 **
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Testing association between aff and SNP snp6 :

Analysis of Deviance Table

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2379 2498.9
get(snpname) 2 1.7969 2377 2497.1 0.4072

Testing association between aff and SNP snp7 :

Analysis of Deviance Table

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2367 2487.9
get(snpname) 2 1.3604 2365 2486.6 0.5065

Testing association between aff and SNP snp8 :

Analysis of Deviance Table
66CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2370 2489.4
get(snpname) 2 5.5375 2368 2483.9 0.06274 .
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Testing association between aff and SNP snp9 :

Analysis of Deviance Table

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2360 2476.8
get(snpname) 2 1.1891 2358 2475.6 0.5518

Testing association between aff and SNP snp10 :

Analysis of Deviance Table

Model: binomial, link: logit

Response: aff

Terms added sequentially (first to last)

Df Deviance Resid. Df Resid. Dev Pr(>Chi)

NULL 2383 2475.1
get(snpname) 2 6.7328 2381 2468.4 0.03451 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Answer (Ex. 12) — It appears that SNP10 genotype is only relevant in these
who are homozygous for the low-risk A allele at the SNP5; in such cases SNP 10
allele B is risk increasing. In these homozygous for SNP 5 A, we observe highly
significant increase in risk for heterozygotes for SNP10 and increased (though
3.6. ANSWERS TO EXERCISES 67

not significantly) risk for SNP 10 BB:

> summary(glm(aff~snp5*snp10,family=binomial()))
Call:
glm(formula = aff ~ snp5 * snp10, family = binomial())

Deviance Residuals:
Min 1Q Median 3Q Max
-0.9906 -0.7340 -0.6323 -0.5215 2.0310

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.50840 0.08905 -16.938 < 2e-16 ***
snp5A/A -0.41802 0.19722 -2.120 0.0340 *
snp5B/B 0.33441 0.13360 2.503 0.0123 *
snp10A/B -0.01403 0.18251 -0.077 0.9387
snp10B/B -0.14983 0.55277 -0.271 0.7863
snp5A/A:snp10A/B 1.48369 0.32750 4.530 5.89e-06 ***
snp5B/B:snp10A/B 0.12989 0.27441 0.473 0.6360
snp5A/A:snp10B/B 0.82348 0.98963 0.832 0.4053
snp5B/B:snp10B/B -0.28562 1.23104 -0.232 0.8165
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 2282.4 on 2268 degrees of freedom

Residual deviance: 2242.9 on 2260 degrees of freedom
(231 observations deleted due to missingness)
AIC: 2260.9

Number of Fisher Scoring iterations: 4

Answer (Ex. 13) — SNPs 1, 4, 5, 6 and 9 are significantly associated at

nominal P ≤ 0.05. Here is a testing script (no need to reproduce that, just
check the results):
> for (i in 1:10) {
+ snpname <- paste("snp",i,sep="")
+ cat("\nTesting association between qt and SNP",snpname,":\n")
+ testmodel <- lm(qt~[Link](get(snpname)))
+ print(summary(testmodel)$coef)
+ }
Testing association between qt and SNP snp1 :
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.11874800 0.05260279 -2.257447 0.024070746
[Link](get(snpname)) 0.08859657 0.03147693 2.814651 0.004923315

Testing association between qt and SNP snp2 :

68CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

Estimate Std. Error t value Pr(>|t|)

(Intercept) 0.09149145 0.1841115 0.4969352 0.6192808
[Link](get(snpname)) -0.07749967 0.1806078 -0.4291047 0.6678860

Testing association between qt and SNP snp3 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.06382773 0.05629376 1.1338331 0.2569789
[Link](get(snpname)) -0.02517149 0.02894125 -0.8697443 0.3845280

Testing association between qt and SNP snp4 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.14005988 0.05612775 2.495377 0.01264938
[Link](get(snpname)) -0.07284539 0.02982557 -2.442380 0.01466282

Testing association between qt and SNP snp5 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.14350846 0.06620992 -2.167477 0.03029734
[Link](get(snpname)) 0.07404874 0.02941437 2.517434 0.01188645

Testing association between qt and SNP snp6 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.12737489 0.06768096 -1.881990 0.05995937
[Link](get(snpname)) 0.06724115 0.02924840 2.298969 0.02159304

Testing association between qt and SNP snp7 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.08884244 0.05475991 1.622399 0.1048511
[Link](get(snpname)) -0.03774136 0.03152701 -1.197112 0.2313829

Testing association between qt and SNP snp8 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.05214665 0.08116533 -0.6424744 0.5206274
[Link](get(snpname)) 0.06942327 0.07222881 0.9611576 0.3365710

Testing association between qt and SNP snp9 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.2201742 0.07115711 -3.094199 0.0019965937
[Link](get(snpname)) 0.2110978 0.06112112 3.453761 0.0005625794

Testing association between qt and SNP snp10 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.01474695 0.05749473 -0.2564921 0.7975930
[Link](get(snpname)) 0.03140888 0.04251458 0.7387789 0.4601141

Answer (Ex. 14) — Generally, results do not change much: still, SNPs 1, 4,
5, 6 and 9 are significantly associated, and p-values are close to these observed
without adjustment For SNPs
> for (i in 1:10) {
3.6. ANSWERS TO EXERCISES 69

+ snpname <- paste("snp",i,sep="")

+ cat("\nTesting sex-adjusted association between qt and SNP",snpname,":\n")
+ testmodel <- lm(qt~sex+[Link](get(snpname)))
+ print(summary(testmodel)$coef)
+ }

Testing sex-adjusted association between qt and SNP snp1 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.12783200 0.05660623 -2.2582672 0.024019551
sex 0.01766209 0.04061074 0.4349117 0.663666095
[Link](get(snpname)) 0.08868826 0.03148302 2.8170191 0.004887301

Testing sex-adjusted association between qt and SNP snp2 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.08169692 0.18525892 0.4409878 0.6592621
sex 0.01977118 0.04101612 0.4820343 0.6298261
[Link](get(snpname)) -0.07770464 0.18063757 -0.4301688 0.6671120

Testing sex-adjusted association between qt and SNP snp3 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.05518829 0.06028776 0.9154146 0.3600669
sex 0.01626276 0.04057000 0.4008568 0.6885616
[Link](get(snpname)) -0.02491788 0.02895328 -0.8606239 0.3895321

Testing sex-adjusted association between qt and SNP snp4 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.12611774 0.06040009 2.0880388 0.03690009
sex 0.02553529 0.04083273 0.6253633 0.53179244
[Link](get(snpname)) -0.07227905 0.02984312 -2.4219671 0.01551079

Testing sex-adjusted association between qt and SNP snp5 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.15272175 0.06880690 -2.2195704 0.02654187
sex 0.02009534 0.04076087 0.4930057 0.62205408
[Link](get(snpname)) 0.07357452 0.02943476 2.4995791 0.01250094

Testing sex-adjusted association between qt and SNP snp6 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.14110680 0.07002874 -2.0149841 0.04401871
sex 0.03117861 0.04077551 0.7646406 0.44456146
[Link](get(snpname)) 0.06634602 0.02927437 2.2663517 0.02351939

Testing sex-adjusted association between qt and SNP snp7 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.06697438 0.05879604 1.139097 0.2547782
sex 0.04174468 0.04087001 1.021401 0.3071689
[Link](get(snpname)) -0.03723286 0.03153066 -1.180846 0.2377825

Testing sex-adjusted association between qt and SNP snp8 :

Estimate Std. Error t value Pr(>|t|)
70CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

(Intercept) -0.06807682 0.08455606 -0.8051087 0.4208378

sex 0.02760968 0.04102873 0.6729354 0.5010541
[Link](get(snpname)) 0.07124407 0.07228780 0.9855614 0.3244491

Testing sex-adjusted association between qt and SNP snp9 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.22558672 0.07313030 -3.0847229 0.0020610506
sex 0.01311497 0.04074702 0.3218632 0.7475848790
[Link](get(snpname)) 0.21002414 0.06122367 3.4304402 0.0006129721

Testing sex-adjusted association between qt and SNP snp10 :

Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.02805887 0.06107457 -0.4594198 0.6459747
sex 0.02628396 0.04064009 0.6467496 0.5178563
[Link](get(snpname)) 0.03150834 0.04252005 0.7410231 0.4587525

Answer (Ex. 15) — SNPs 1, 4, 5 an 9 are significantly associated at nominal

P ≤ 0.05. SNP 6 is only marginally significantly associated unde the general
genotypic model. Here is a testing script (no need to reproduce that, just check
the results):
> for (i in 1:10) {
+ snpname <- paste("snp",i,sep="")
+ cat("\nTesting association between qt and SNP",snpname,":\n")
+ print(anova(lm(qt~get(snpname)),test="Chisq"))
+ #print(summary(lm(qt~get(snpname)))$coef)
+ }
Testing association between qt and SNP snp1 :
Analysis of Variance Table

Response: qt
Df Sum Sq Mean Sq F value Pr(>F)
get(snpname) 2 7.8 3.8995 3.9845 0.01873 *
Residuals 2371 2320.4 0.9787
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Testing association between qt and SNP snp2 :

Analysis of Variance Table

Response: qt
Df Sum Sq Mean Sq F value Pr(>F)
get(snpname) 1 0.18 0.18376 0.1841 0.6679
Residuals 2372 2367.23 0.99799

Testing association between qt and SNP snp3 :

Analysis of Variance Table

Response: qt
3.6. ANSWERS TO EXERCISES 71

Df Sum Sq Mean Sq F value Pr(>F)

get(snpname) 2 3.24 1.61986 1.658 0.1907
Residuals 2375 2320.41 0.97701

Testing association between qt and SNP snp4 :

Analysis of Variance Table

Response: qt
Df Sum Sq Mean Sq F value Pr(>F)
get(snpname) 2 7.68 3.8417 3.8628 0.02114 *
Residuals 2387 2373.94 0.9945
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Testing association between qt and SNP snp5 :

Analysis of Variance Table

Response: qt
Df Sum Sq Mean Sq F value Pr(>F)
get(snpname) 2 6.48 3.2418 3.2798 0.03781 *
Residuals 2380 2352.48 0.9884
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Testing association between qt and SNP snp6 :

Analysis of Variance Table

Response: qt
Df Sum Sq Mean Sq F value Pr(>F)
get(snpname) 2 5.49 2.74680 2.7808 0.06219 .
Residuals 2377 2347.91 0.98776
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Testing association between qt and SNP snp7 :

Analysis of Variance Table

Response: qt
Df Sum Sq Mean Sq F value Pr(>F)
get(snpname) 2 4.02 2.01212 2.0368 0.1307
Residuals 2365 2336.31 0.98787

Testing association between qt and SNP snp8 :

Analysis of Variance Table

Response: qt
Df Sum Sq Mean Sq F value Pr(>F)
get(snpname) 2 1.38 0.68987 0.6924 0.5005
Residuals 2368 2359.23 0.99630
72CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

Testing association between qt and SNP snp9 :

Analysis of Variance Table

Response: qt
Df Sum Sq Mean Sq F value Pr(>F)
get(snpname) 2 15.6 7.8014 7.9982 0.0003453 ***
Residuals 2358 2300.0 0.9754
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Testing association between qt and SNP snp10 :

Analysis of Variance Table

Response: qt
Df Sum Sq Mean Sq F value Pr(>F)
get(snpname) 2 1.19 0.59456 0.6041 0.5467
Residuals 2381 2343.47 0.98424

Answer (Ex. 16) — For ’snp1’, though the data are compatible with either
additive, dominant or recessive model, the additive model provides best fit to
the data (largest p-value), while the recessive ’B’ model provide the wors fit
(almost significantly worse than the general model):
> table(snp1,[Link](snp1))
snp1 1 2 3
A/A 1287 0 0
A/B 0 888 0
B/B 0 0 199
> table(snp1,([Link](snp1)>=2))
snp1 FALSE TRUE
A/A 1287 0
A/B 0 888
B/B 0 199
> table(snp1,([Link](snp1)>=3))
snp1 FALSE TRUE
A/A 1287 0
A/B 888 0
B/B 0 199
> model_gen <- lm(qt~snp1)
> summary(model_gen)
Call:
lm(formula = qt ~ snp1)

Residuals:
Min 1Q Median 3Q Max
-3.5261 -0.6643 -0.0111 0.6765 3.5462
3.6. ANSWERS TO EXERCISES 73

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.02846 0.02758 -1.032 0.3022
snp1A/B 0.08200 0.04316 1.900 0.0575 .
snp1B/B 0.18644 0.07536 2.474 0.0134 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.9893 on 2371 degrees of freedom

(126 observations deleted due to missingness)
Multiple R-squared: 0.00335, Adjusted R-squared: 0.002509
F-statistic: 3.985 on 2 and 2371 DF, p-value: 0.01873
> model_add <- lm(qt~[Link](snp1))
> model_dom <- lm(qt~I([Link](snp1)>=2))
> model_rec <- lm(qt~I([Link](snp1)>=3))
> anova(model_add,model_gen,test="Chisq")
Analysis of Variance Table

Model 1: qt ~ [Link](snp1)
Model 2: qt ~ snp1
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2372 2320.5
2 2371 2320.4 1 0.04886 0.8232
> anova(model_dom,model_gen,test="Chisq")
Analysis of Variance Table

Model 1: qt ~ I([Link](snp1) >= 2)

Model 2: qt ~ snp1
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2372 2322.2
2 2371 2320.4 1 1.7733 0.1783
> anova(model_rec,model_gen,test="Chisq")
Analysis of Variance Table

Model 1: qt ~ I([Link](snp1) >= 3)

Model 2: qt ~ snp1
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2372 2324.0
2 2371 2320.4 1 3.5332 0.05743 .
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
For SNPs 4, 5, 6, and 9 results are:
> for (i in c(4,5,6,9)) {
+ snpname <- paste("snp",i,sep="")
+ cat("\nTesting SNP",snpname,":\n")
+ cursnp <- get(snpname)
+ model_gen <- lm(qt~cursnp)
74CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

+ print(summary(model_gen))
+ model_add <- lm(qt~[Link](cursnp))
+ model_dom <- lm(qt~I([Link](cursnp)>=2))
+ model_rec <- lm(qt~I([Link](cursnp)>=3))
+ print(anova(model_add,model_gen,test="Chisq"))
+ print(anova(model_dom,model_gen,test="Chisq"))
+ print(anova(model_rec,model_gen,test="Chisq"))
+ }

Testing SNP snp4 :

Call:
lm(formula = qt ~ cursnp)

Residuals:
Min 1Q Median 3Q Max
-3.4311 -0.6636 -0.0013 0.6737 3.5489

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.02132 0.02972 0.717 0.4733
cursnpA/A 0.02953 0.04423 0.668 0.5044
cursnpB/B -0.14481 0.06192 -2.339 0.0194 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.9973 on 2387 degrees of freedom

(110 observations deleted due to missingness)
Multiple R-squared: 0.003226, Adjusted R-squared: 0.002391
F-statistic: 3.863 on 2 and 2387 DF, p-value: 0.02114

Analysis of Variance Table

Model 1: qt ~ [Link](cursnp)
Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2388 2375.7
2 2387 2373.9 1 1.7489 0.1848
Analysis of Variance Table

Model 1: qt ~ I([Link](cursnp) >= 2)

Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2388 2379.4
2 2387 2373.9 1 5.4391 0.01936 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Analysis of Variance Table

Model 1: qt ~ I([Link](cursnp) >= 3)

3.6. ANSWERS TO EXERCISES 75

Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2388 2374.4
2 2387 2373.9 1 0.44342 0.5043

Testing SNP snp5 :

Call:
lm(formula = qt ~ cursnp)

Residuals:
Min 1Q Median 3Q Max
-3.4719 -0.6589 -0.0084 0.6622 3.5285

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.01401 0.02878 0.487 0.6264
cursnpA/A -0.09667 0.05611 -1.723 0.0851 .
cursnpB/B 0.05727 0.04607 1.243 0.2140
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.9942 on 2380 degrees of freedom

(117 observations deleted due to missingness)
Multiple R-squared: 0.002749, Adjusted R-squared: 0.00191
F-statistic: 3.28 on 2 and 2380 DF, p-value: 0.03781

Analysis of Variance Table

Model 1: qt ~ [Link](cursnp)
Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2381 2352.7
2 2380 2352.5 1 0.22152 0.6359
Analysis of Variance Table

Model 1: qt ~ I([Link](cursnp) >= 2)

Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2381 2354.0
2 2380 2352.5 1 1.5273 0.2138
Analysis of Variance Table

Model 1: qt ~ I([Link](cursnp) >= 3)

Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2381 2355.4
2 2380 2352.5 1 2.9335 0.08494 .
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
76CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R

Testing SNP snp6 :

Call:
lm(formula = qt ~ cursnp)

Residuals:
Min 1Q Median 3Q Max
-3.4784 -0.6753 -0.0064 0.6703 3.5324

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.07617 0.05085 -1.498 0.1343
cursnpB/A 0.09417 0.05886 1.600 0.1097
cursnpB/B 0.14351 0.06096 2.354 0.0186 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.9939 on 2377 degrees of freedom

(120 observations deleted due to missingness)
Multiple R-squared: 0.002334, Adjusted R-squared: 0.001495
F-statistic: 2.781 on 2 and 2377 DF, p-value: 0.06219

Analysis of Variance Table

Model 1: qt ~ [Link](cursnp)
Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2378 2348.2
2 2377 2347.9 1 0.27462 0.598
Analysis of Variance Table

Model 1: qt ~ I([Link](cursnp) >= 2)

Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2378 2349.1
2 2377 2347.9 1 1.1967 0.271
Analysis of Variance Table

Model 1: qt ~ I([Link](cursnp) >= 3)

Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2378 2350.4
2 2377 2347.9 1 2.5284 0.1096

Testing SNP snp9 :

Call:
lm(formula = qt ~ cursnp)
3.6. ANSWERS TO EXERCISES 77

Residuals:
Min 1Q Median 3Q Max
-3.5482 -0.6673 0.0074 0.6546 3.6061

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.006298 0.021562 -0.292 0.77026
cursnpA/B 0.162230 0.065729 2.468 0.01365 *
cursnpB/B 1.002439 0.313057 3.202 0.00138 **
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.9876 on 2358 degrees of freedom

(139 observations deleted due to missingness)
Multiple R-squared: 0.006738, Adjusted R-squared: 0.005896
F-statistic: 7.998 on 2 and 2358 DF, p-value: 0.0003453

Analysis of Variance Table

Model 1: qt ~ [Link](cursnp)
Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2359 2303.9
2 2358 2300.0 1 3.9528 0.04411 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Analysis of Variance Table

Model 1: qt ~ I([Link](cursnp) >= 2)

Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2359 2306.8
2 2358 2300.0 1 6.7911 0.008324 **
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Analysis of Variance Table

Model 1: qt ~ I([Link](cursnp) >= 3)

Model 2: qt ~ cursnp
[Link] RSS Df Sum of Sq Pr(>Chi)
1 2359 2305.9
2 2358 2300.0 1 5.942 0.01358 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
78CHAPTER 3. INTRODUCTION TO GENETIC ASSOCIATION ANALYSIS IN R
Chapter 4

Introduction to the
GenABEL-package

In this section, you will become familiar with the GenABEL-package library,
designed for GWA analysis. Compared to the genetics package, it provides
specific facilities for storage and manipulation of large amounts of data, very
fast tests for GWA analysis, and special functions to analyse and graphically
present the results of GWA analysis (thus ”analysis of analysis”).
Start R and load the GenABEL-package library using the command
> library(GenABEL)
After that, load the example data set using the command
> data(srdta)

4.1 General description of [Link]-class

The object you have loaded, srdta, belongs to the [Link] class. This is a
special R class developed to facilitate GWA analysis.
In GWA analysis, different types of data are used. These include the pheno-
typic and genotypic data on the study participants and chromosome and location
of every SNP. For every SNP, it is desirable to know the details of coding (how
are the alleles coded? – A, T, G, C? – as well as the strand – ’+’ or ’-’, ’top’ or
’bot’ ? – this coding is for).
One could attempt to store all phenotypes and genotypes together in a single
table, using, e.g. one row per study subject; than the columns will correspond
to study phenotypes and SNPs. For a typical GWA data set, this would lead
to a table of few thousands rows and few hundreds of thousands to millions of
columns. Such a format is generated when one downloads HapMap data for a
region. To store GWA data in such tables internally, within R, proves to be
inefficient. In GenABEL-package, a special data class, [Link]-class is used
to store GWA data.
You may consider an object of [Link]-class as a ’black box’ from which
you can get specific data using specific functions. If you are interested in the in-
ternal structure of the [Link]-class, you can find the description in section
B.1 (Internal structure of [Link]-class).

79
80 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

The data frame, which contains all phenotypic data in the study may be
accessed using the phdata function. Let us have a look at the first few rows of
the phenotypic data frame of srdta:
> phdata(srdta)[1:5,]
id sex age qt1 qt2 qt3 bt
p1 p1 1 43.4 -0.58 4.46 1.43 0
p2 p2 1 48.2 0.80 6.32 3.90 1
p3 p3 0 37.9 -0.52 3.26 5.05 1
p4 p4 1 53.8 -1.55 888.00 3.76 1
p5 p5 1 47.5 0.25 5.70 2.89 1
The rows of this data frame correspond to th estudy subjects, and the columns
correspond to the variables. There are two default variables, which are always
present in phdata. The first of these is “id”, which contains study subject
identification code. This identification code can be arbitrary character, numer,
or alphanumeric combination, but every person must be coded with an unique
ID. The second default variable is “sex”, where males are coded with ones (“1”)
and females are coded with zero (“0”).
It is important to understand that this data frame is not supposed to be
directly modified by the user, as its structure is coupled to the structure of
genotypic data. If at some point you need to manipulate (add/delete) the phe-
notypes included in phdata, you need to use such GenABEL-package functions
as [Link] and [Link] (see section 4.2).
The other part of an object of [Link]-class is gtdata, which contains
all GWA genetic information in an object of class [Link] class. It is not
supposed to be modified directly by user. The genotypic data can be accessed
through the gtdata function, e.g.
> gtdata(srdta[1:10, 1:10])
@nids = 10
@nsnps = 10
@nbytes = 3
@idnames = p1 p2 p3 p4 p5 p6 p7 p8 p9 p10
@snpnames = rs10 rs18 rs29 rs65 rs73 rs114 rs128 rs130 rs143 rs150
@chromosome = 1 1 1 1 1 1 1 1 1 1
@coding = 08 0b 0c 03 04 03 0c 04 08 0f
@strand = 01 01 02 01 01 01 02 01 01 01
@map = 2500 3500 5750 13500 14250 24500 27000 27250 31000 33250
@male = 1 1 0 1 1 0 0 1 0 0
@gtps =
40 40 40 80 40 40 40 40 c0 c0
40 40 00 00 40 40 40 c0 40 40
40 40 00 80 40 40 40 40 c0 c0
As you can see, these data are of little direct use as these are stored in an
internal format – you need to coerce that to another data type if you want
to manipulate/analyse these data using non-GenABEL-package functions (see
section ??).
The number of individuals described in an object of [Link]-class can
be accessed through nids function, e.g.
4.1. GENERAL DESCRIPTION OF [Link]-CLASS 81

> nids(srdta)

[1] 2500

and the number of SNPs using the nsnps function:

> nsnps(srdta)

[1] 833

The IDs of the individuals included in the study can be accessed via the
idnames function, for example the IDs of the first 7 individuals in the study are
> idnames(srdta)[1:7]

[1] "p1" "p2" "p3" "p4" "p5" "p6" "p7"

The sex of the individuals can be accessed using the male function:
> male(srdta)[1:7]

p1 p2 p3 p4 p5 p6 p7
1 1 0 1 1 0 0

where males (heterogametic sex) are assigned with “1” and a homogametic sex
(females) are assigned the value “0”.
Names of SNPs can be accessed using the snpnames function; for example
the names of the first 10 SNPs in the srdta are
> snpnames(srdta)[1:10]

[1] "rs10" "rs18" "rs29" "rs65" "rs73" "rs114" "rs128" "rs130" "rs143"
[10] "rs150"

SNP annotation includes (shown for the first 10 SNPs only):

• Chromosome:
> chromosome(srdta)[1:10]

rs10 rs18 rs29 rs65 rs73 rs114 rs128 rs130 rs143 rs150
"1" "1" "1" "1" "1" "1" "1" "1" "1" "1"

• Map position
> map(srdta)[1:10]

rs10 rs18 rs29 rs65 rs73 rs114 rs128 rs130 rs143 rs150
2500 3500 5750 13500 14250 24500 27000 27250 31000 33250

• Coding (where the second allele is the “effect” or “coded” one):

> coding(srdta)[1:10]

rs10 rs18 rs29 rs65 rs73 rs114 rs128 rs130 rs143 rs150
"TG" "GA" "GT" "AT" "AG" "AT" "GT" "AG" "TG" "CA"
82 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

For every SNP, the coding is represented with a pair of characters, for
example “AG”. For an “AG” polymorphism, you may expect “AA”, “AG”
and “GG” genotypes to be found in your population. The order (that is
“AG” vs. “GA”) is important – the first allele reported is the one which
will be used as a reference in association analysis, and thus the effects are
reported for the second allele. You can also access the reference allele with
the method refallele
> refallele(srdta)[1:10]
rs10 rs18 rs29 rs65 rs73 rs114 rs128 rs130 rs143 rs150
"T" "G" "G" "A" "A" "A" "G" "A" "T" "C"
and the effective (or ’coded’) allelel with
> effallele(srdta)[1:10]
rs10 rs18 rs29 rs65 rs73 rs114 rs128 rs130 rs143 rs150
"G" "A" "T" "T" "G" "T" "T" "G" "G" "A"
• The strand on which the coding is reported (’+’, ’-’ or missing, ’u’):
> strand(srdta)[1:10]
rs10 rs18 rs29 rs65 rs73 rs114 rs128 rs130 rs143 rs150
"+" "+" "-" "+" "+" "+" "-" "+" "+" "+"

Summary:
• GenABEL-package uses a special data class, [Link]-class, to store
GWA data.
• To access the content of an object of [Link]-class, a number of
functions is used

Exercise 1. Exploring IDs in srdta

Explore srdta.
1. How many people are included in the study?
2. How many of these are males?
3. How many are females?
4. What is male proportion?

Exercise 2. Exploring SNPs in srdta

Explore the SNPs contained in srdta using the functions to access SNP names
(snpnames) and map (map) location
1. What are names of markers located after 2,490,000 b.p.?
2. Between 1,100,000 and 1,105,000 b.p.?
4.2. ACCESSING AND MODIFYING PHENOTYPIC DATA 83

4.2 Accessing and modifying phenotypic data

As was already mentioned, the object returned by phdata contains phenotypic
data and is a conventional data frame, wich obligatory includes ’id’ and ’sex’
variables, and ordered an a way that it couples to the genotypic data.
Being a data frame, phdata may be accessed using the corresponding meth-
ods:

> phdata(srdta)[1:5, ]

id sex age qt1 qt2 qt3 bt

p1 p1 1 43.4 -0.58 4.46 1.43 0
p2 p2 1 48.2 0.80 6.32 3.90 1
p3 p3 0 37.9 -0.52 3.26 5.05 1
p4 p4 1 53.8 -1.55 888.00 3.76 1
p5 p5 1 47.5 0.25 5.70 2.89 1

> class(phdata(srdta))

[1] "[Link]"

> phdata(srdta)[1:5, 2]

[1] 1 1 0 1 1

> phdata(srdta)[1:5, "sex"]

[1] 1 1 0 1 1

> phdata(srdta)$sex[1:5]

[1] 1 1 0 1 1

The modification of the phenotypic data is performed using special meth-

ods, because of specific restrictions on phenotypic data frames. There are two
main functions which allow you to add ([Link]) and delete ([Link])
phenotypes from phdata part of an object of [Link]-class.
For example, if you want to add a variable (say, the square of age) computed
from the “age” variable of srdta

> phdata(srdta)[1:5, ]

id sex age qt1 qt2 qt3 bt

p1 p1 1 43.4 -0.58 4.46 1.43 0
p2 p2 1 48.2 0.80 6.32 3.90 1
p3 p3 0 37.9 -0.52 3.26 5.05 1
p4 p4 1 53.8 -1.55 888.00 3.76 1
p5 p5 1 47.5 0.25 5.70 2.89 1

> age2 <- phdata(srdta)$age^2

you need to use the [Link] function:

84 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

> srdta <- [Link](srdta, newph=age2, name="age_squared")

> phdata(srdta)[1:5, ]

id sex age qt1 qt2 qt3 bt age_squared

p1 p1 1 43.4 -0.58 4.46 1.43 0 1883.56
p2 p2 1 48.2 0.80 6.32 3.90 1 2323.24
p3 p3 0 37.9 -0.52 3.26 5.05 1 1436.41
p4 p4 1 53.8 -1.55 888.00 3.76 1 2894.44
p5 p5 1 47.5 0.25 5.70 2.89 1 2256.25

You can add more than one variable at once using the same function, how-
ever, in this case the second (“newph”) argument of the function should be a
data frame, which contains an ’id’ variable specifing the IDs of the individuals.
Imagine we have the data for individuals ’p1’, ’p2’ and ’p7’ (we will generate
random data for them; pay attention only to the result):
> newvalues <- matrix( rnorm(3*5), 3, 5 )
> newdata <- [Link](id=c("p1", "p2", "p7"),
+ ph1=1, ph2=1, ph3=1, ph4=1, ph5=1)
> newdata[, c(2:6)] <- newvalues
> newdata

id ph1 ph2 ph3 ph4 ph5

1 p1 -0.3659861 1.0874469 -0.09242666 -1.0272750 0.04431896
2 p2 1.0150075 1.7965355 -0.70273473 0.4395144 -0.72266042
3 p7 0.1499759 0.2848736 -0.04109438 1.3974057 1.43608206

These data can be added to the phenotypic data with

> srdta <- [Link](srdta,newdata)
> phdata(srdta)[1:10, ]

id sex age qt1 qt2 qt3 bt age_squared ph1 ph2

p1 p1 1 43.4 -0.58 4.46 1.43 0 1883.56 -0.3659861 1.0874469
p2 p2 1 48.2 0.80 6.32 3.90 1 2323.24 1.0150075 1.7965355
p3 p3 0 37.9 -0.52 3.26 5.05 1 1436.41 NA NA
p4 p4 1 53.8 -1.55 888.00 3.76 1 2894.44 NA NA
p5 p5 1 47.5 0.25 5.70 2.89 1 2256.25 NA NA
p6 p6 0 45.0 0.15 4.65 1.87 0 2025.00 NA NA
p7 p7 0 52.0 -0.56 4.64 2.49 0 2704.00 0.1499759 0.2848736
p8 p8 1 42.5 NA 5.77 2.68 1 1806.25 NA NA
p9 p9 0 29.7 -2.26 0.71 1.45 0 882.09 NA NA
p10 p10 0 45.8 -1.32 3.26 0.85 0 2097.64 NA NA
ph3 ph4 ph5
p1 -0.09242666 -1.0272750 0.04431896
p2 -0.70273473 0.4395144 -0.72266042
p3 NA NA NA
p4 NA NA NA
p5 NA NA NA
p6 NA NA NA
p7 -0.04109438 1.3974057 1.43608206
p8 NA NA NA
4.3. SUB-SETTING AND COERCING [Link] 85

p9 NA NA NA
p10 NA NA NA

Finally, if you need, you can delete some phenotypes from the phdata using
[Link] function. Let us delete the phenotypes we have just added:
> srdta <- [Link](srdta,
+ c("age_squared", "ph1", "ph2", "ph3", "ph4", "ph5"))
> phdata(srdta)[1:10, ]

id sex age qt1 qt2 qt3 bt

p1 p1 1 43.4 -0.58 4.46 1.43 0
p2 p2 1 48.2 0.80 6.32 3.90 1
p3 p3 0 37.9 -0.52 3.26 5.05 1
p4 p4 1 53.8 -1.55 888.00 3.76 1
p5 p5 1 47.5 0.25 5.70 2.89 1
p6 p6 0 45.0 0.15 4.65 1.87 0
p7 p7 0 52.0 -0.56 4.64 2.49 0
p8 p8 1 42.5 NA 5.77 2.68 1
p9 p9 0 29.7 -2.26 0.71 1.45 0
p10 p10 0 45.8 -1.32 3.26 0.85 0

Summary:

• Phenotypic data contained in an object of [Link]-class can be ac-

cessed using the phdata functions
• You can add phenotypes using the [Link] function
• You can delete phenotypes using the [Link] function

4.3 Sub-setting and coercing [Link]

It is possible to sub-set the object, which stores the GWA data in the man-
ner similar to that used for conventional R matrices and data frames. Very
primitively, you may think of an object of class [Link] as a matrix whose
rows correspond to study subjects and columns correspond to SNPs studied
(though the actual object is more complicated). For example, if we would like
to investigate the content of srdta for the first 5 people and 3 SNPs, we can
run
> ssubs <- srdta[1:5, 1:3]
> class(ssubs)

[1] "[Link]"
attr(,"package")
[1] "GenABEL"
86 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

As you can see, by sub-setting we obtained a smaller object of [Link]-class.

The two major parts it contains are phenotypic data, which can be accessed
through phdata (discussed in section 4.2):
> phdata(ssubs)

id sex age qt1 qt2 qt3 bt

p1 p1 1 43.4 -0.58 4.46 1.43 0
p2 p2 1 48.2 0.80 6.32 3.90 1
p3 p3 0 37.9 -0.52 3.26 5.05 1
p4 p4 1 53.8 -1.55 888.00 3.76 1
p5 p5 1 47.5 0.25 5.70 2.89 1

and genotypc data, which can be accessed via the gtdata function:
> gtdata(ssubs)

@nids = 5
@nsnps = 3
@nbytes = 2
@idnames = p1 p2 p3 p4 p5
@snpnames = rs10 rs18 rs29
@chromosome = 1 1 1
@coding = 08 0b 0c
@strand = 01 01 02
@map = 2500 3500 5750
@male = 1 1 0 1 1
@gtps =
40 40 40
40 40 00

whose content is not quite straightforward to read.

To get human-readable information, a genotypic object should be coerced to
a regular R data type, e.g. character, using the [Link]() function:
> [Link](gtdata(ssubs))

rs10 rs18 rs29

p1 "T/T" "G/G" "G/G"
p2 "T/T" "G/G" NA
p3 "T/T" "G/G" NA
p4 "T/T" "G/G" NA
p5 "T/T" "G/A" "G/G"

Another useful coercion is to ”numeric”:

> [Link](gtdata(ssubs))

rs10 rs18 rs29

p1 0 0 0
p2 0 0 NA
p3 0 0 NA
p4 0 0 NA
p5 0 1 0
4.3. SUB-SETTING AND COERCING [Link] 87

Note that conversion to numeric happened according to the underlying genotype

and the rules specified by SNP coding:
> coding(ssubs)
rs10 rs18 rs29
"TG" "GA" "GT"
– the genotype, which is made of the ’first’ allele of the ’code’ is converted to
“0”, the heterozygote to “1” and a homozygote for the second allele is converted
to “2”.
For example, when the coding is “GA”, as it is for the rs18 (the second SNP),
homozygotes for the first allele, as specified by coding (“G”) are converted to
zeros (“0”), heterozygotes are converted to ones (“1”), and homozygotes for
the second allele (“A”) are converted to twos (“2”). Clearly, when numerically
converted data are used for association analysis, the effects will be estimated
for the second allele, while first will be used as a reference.
Genotypic data converted to standard R format can be used in any further
analysis.
Several useful genetic analysis libraries were developed for R. These include
genetics (analysis of linkage disequilibrium and many other useful functions)
and [Link] (analysis of association between traits and haplotypes). These
use their own genetic data formats.
One can translate GenABEL-package genetic data to the format used by
”genetics” library using [Link]():
> [Link](gtdata(ssubs))
rs10 rs18 rs29
p1 T/T G/G G/G
p2 T/T G/G <NA>
p3 T/T G/G <NA>
p4 T/T G/G <NA>
p5 T/T G/A G/G
To translate GenABEL-package data to the format used by ”[Link]” you
can use function [Link]()
> [Link](gtdata(ssubs))
rs10.a1 rs10.a2 rs18.a1 rs18.a2 rs29.a1 rs29.a2
p1 1 1 1 1 1 1
p2 1 1 1 1 NA NA
p3 1 1 1 1 NA NA
p4 1 1 1 1 NA NA
p5 1 1 1 2 1 1
Actually, most users will not need the latter function, as GenABEL-package
provides a functional interface to ”[Link]” (such GenABEL-package func-
tions as [Link]() and [Link].2D()).
It is possible to select sub-sets of [Link]-class based not only on index
(e.g. first 10 people and SNP number 33), but also based on names.
For example, if we would like to retrieve phenotypic data on people with IDs
”p141”, ”p147” and ”p2000”, we can use
88 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

> phdata( srdta[c("p141", "p147", "p2000"), ] )

id sex age qt1 qt2 qt3 bt

p141 p141 0 47.2 0.51 5.23 2.17 0
p147 p147 0 43.2 0.14 4.47 1.73 0
p2000 p2000 0 43.1 -1.53 2.78 2.70 1

here, the first part of expression sub-sets srdta on selected IDs, and the sec-
ond tells which part of the retrieved sub-set we want to see. You can try
srdta[c("p141", "p147", "p2000"),], but be prepared to see long output,
as all information will be reported.
In a similar manner, we can also select on SNP name. For example, if we
are interested to see information on SNPs ”rs10” and ”rs29” for above people,
we can run
> phdata(srdta[c("p141", "p147", "p2000"), c("rs10", "rs29")])

id sex age qt1 qt2 qt3 bt

p141 p141 0 47.2 0.51 5.23 2.17 0
p147 p147 0 43.2 0.14 4.47 1.73 0
p2000 p2000 0 43.1 -1.53 2.78 2.70 1

> gtdata(srdta[c("p141", "p147", "p2000"), c("rs10", "rs29")])

@nids = 3
@nsnps = 2
@nbytes = 1
@idnames = p141 p147 p2000
@snpnames = rs10 rs29
@chromosome = 1 1
@coding = 08 0c
@strand = 01 02
@map = 2500 5750
@male = 0 0 0
@gtps =
40 40

To see the actual genotypes for the above three people and two SNPs, use
> [Link](srdta[c("p141", "p147", "p2000"), c("rs10", "rs29")])

rs10 rs29
p141 "T/T" "G/G"
p147 "T/T" "G/G"
p2000 "T/G" "G/T"

or
> [Link](srdta[c("p141", "p147", "p2000"), c("rs10", "rs29")])

rs10 rs29
p141 0 0
p147 0 0
p2000 1 1
4.4. EXPLORING GENETIC DATA 89

Exercise 3. Exploring rs114

Explore genotypes for SNP ”rs114”
1. What is the coding and which allele is the reference one?
2. What is the frequency of non-reference (“effect”) allele in total sample?

3. What is the frequency of the effect allele in males?

4. What is the frequency of the effect allele in females?
5. What is the frequency of the reference allele in the total sample, males
and females?

Summary:
• It is possible to obtain subsets of objects of [Link]-class and
[Link]-class using the standard 2D sub-setting model [i, j], where
i corresponds to study subjects and j corresponds to SNPs.

• It is possible to provide ID and SNP names instead of indexes when sub-

setting an object of class [Link]-class.
• The function [Link]() converts genotypic data from
[Link]-class to regular integer numbers, which can be used in
analysis with R.

• The function [Link]() converts genotypic data from

[Link]-class to character format.
• The function [Link]() converts genotypic data from
[Link]-class to the format used by the genetics library.

• The function [Link]() converts genotypic data from [Link]-class

to the format used by the [Link] library.

4.4 Exploring genetic data

Implementation of the function summary() to summarize the genotypic part of
a [Link]-class object is very useful in genetic data exploration and quality
control (QC). Let us try application of this function to the ssubs:
> a <- summary(ssubs)
> a

Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11 P.12 P.22

rs10 1 2500 + T G 5 1.0 0.0 5 0 0
rs18 1 3500 + G A 5 1.0 0.1 4 1 0
rs29 1 5750 - G T 2 0.4 0.0 2 0 0
Pexact Fmax Plrt
rs10 1 0.0000000 1.0000000
90 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

rs18 1 -0.1111111 0.7386227

rs29 1 0.0000000 1.0000000

In the first section, a summary is generated for the phenotypic data. In

the second section, a summary is generated for the genotypic data. In this
section, NoMeasured refers to the number of genotypes scores, CallRate to the
proportion of these, Q.2 is the frequency of the ’B’ allele. The counts in three
genotypic classes are provided next. Pexact refers to exact P -value for the test
of Hardy-Weinberg equilibrium.
As you have seen above, an object of the class [Link]-class is sub-
settable in the standard manner: [i, j], where i is the index of a study subject
and j is the index of a SNP. Importantly, i could be a list of indexes:
> vec <- which(phdata(srdta)$age >= 65)
> vec

[1] 64 122 186 206 207 286 385 386 492 514 525 536 545 565 613
[16] 632 649 673 701 779 799 981 1008 1131 1186 1223 1281 1383 1471 1489
[31] 1501 1565 1584 1673 1679 1782 1821 1832 1866 1891 1953 2081 2085 2140 2224
[46] 2268 2291 2384 2420 2453

> summary( gtdata(srdta[vec, 1:3]) )

Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11 P.12

rs10 1 2500 + T G 48 0.96 0.1354167 36 11
rs18 1 3500 + G A 47 0.94 0.2765957 25 18
rs29 1 5750 - G T 45 0.90 0.1555556 32 12
P.22 Pexact Fmax Plrt
rs10 1 1.0000000 0.02131603 0.8843626
rs18 4 0.7245853 0.04298643 0.7697067
rs29 1 1.0000000 -0.01503759 0.9188943

This shows summary of first three genotypes for people with age greater then or
equal to 65 y.o. The same result may be achieved by sub-setting using a vector
of logical values:
> vec <- (phdata(srdta)$age >= 65)
> table(vec)

vec
FALSE TRUE
2450 50

> summary( gtdata(srdta[vec, 1:3]) )

Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11 P.12

rs10 1 2500 + T G 48 0.96 0.1354167 36 11
rs18 1 3500 + G A 47 0.94 0.2765957 25 18
rs29 1 5750 - G T 45 0.90 0.1555556 32 12
P.22 Pexact Fmax Plrt
rs10 1 1.0000000 0.02131603 0.8843626
rs18 4 0.7245853 0.04298643 0.7697067
rs29 1 1.0000000 -0.01503759 0.9188943
4.4. EXPLORING GENETIC DATA 91

or a list with IDs of study subjects:

> vec1 <- idnames(srdta)[vec]
> vec1

[1] "p64" "p122" "p186" "p206" "p207" "p286" "p385" "p386" "p492"
[10] "p514" "p525" "p536" "p545" "p565" "p613" "p632" "p649" "p673"
[19] "p701" "p779" "p799" "p981" "p1008" "p1131" "p1186" "p1223" "p1281"
[28] "p1383" "p1471" "p1489" "p1501" "p1565" "p1584" "p1673" "p1679" "p1782"
[37] "p1821" "p1832" "p1866" "p1891" "p1953" "p2081" "p2085" "p2140" "p2224"
[46] "p2268" "p2291" "p2384" "p2420" "p2453"

> summary( gtdata(srdta[vec1, 1:3]) )

Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11 P.12

rs10 1 2500 + T G 48 0.96 0.1354167 36 11
rs18 1 3500 + G A 47 0.94 0.2765957 25 18
rs29 1 5750 - G T 45 0.90 0.1555556 32 12
P.22 Pexact Fmax Plrt
rs10 1 1.0000000 0.02131603 0.8843626
rs18 4 0.7245853 0.04298643 0.7697067
rs29 1 1.0000000 -0.01503759 0.9188943

Let us explore the object returned by the summary function when applied to
[Link] class in more detail:
> a <- summary( gtdata(srdta[vec1, 1:3]) )
> class(a)

[1] "[Link]"

Thus, the object returned is a [Link]. Therefore it should have dimensions

and names:
> dim(a)

[1] 3 14

> names(a)

[1] "Chromosome" "Position" "Strand" "A1" "A2"

[6] "NoMeasured" "CallRate" "Q.2" "P.11" "P.12"
[11] "P.22" "Pexact" "Fmax" "Plrt"

Indeed, we derived 8 characteristics (”NoMeasured”, ”CallRate”, ”Q.2”, ”P.11”,

”P.12”, ”P.22”, ”Pexact”, ”Chromosome”) for the first 3 SNPs.

Exercise 4. Testing HWE for 10 SNPs

Test if Hardy-Weinberg equilibrium holds for the first 10 SNPs
1. Total sample
2. In cases (bt is 1)
92 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

Histogram of crate

150
100
Frequency

50
0

0.935 0.940 0.945 0.950 0.955 0.960 0.965

crate

Figure 4.1: Histogram of the call rate.

3. In controls (bt is 0)
Let us analyse the distribution of call rate in the whole study. For this, we
first need to obtain the vector of call rates:
> sumgt <- summary(gtdata(srdta))
> crate <- sumgt[, "CallRate"]
This vector may be depicted by a histogram
> hist(crate)
which shows that most SNPs have call rate between 93 and 97% (Figure 4.1).
As a next step, you would like to produce a summary table, showing how
many markers had call rate lower than, say, 93%, between 93 and 95%, between
95 and 99% and more than 99%. You can use the catable() command for that:
> catable(crate, c(.93, .95, .99))

X<=0.93 0.93<X<=0.95 0.95<X<=0.99 X>0.99

No 0 415.000 418.000 0
Prop 0 0.498 0.502 0

A similar procedure may be applied to see deviation from HWE:

> hwp <- sumgt[, "Pexact"]
> catable(hwp, c(0.05/nsnps(srdta), 0.01, 0.05, 0.1))
4.4. EXPLORING GENETIC DATA 93

X<=6.00240096038415e-05 6.00240096038415e-05<X<=0.01 0.01<X<=0.05

No 2.000 7.000 23.000
Prop 0.002 0.008 0.028
0.05<X<=0.1 X>0.1
No 31.000 770.000
Prop 0.037 0.924

The first cut-off category will detect SNPs which are deviating from HWE at
the Bonferroni-corrected P -level.
However, for these data it will make more sense to table the cumulative
distribution:
> catable(hwp, c(0.05/nsnps(srdta), 0.01, 0.05, 0.1), cum=TRUE)

X<=6.00240096038415e-05 X<=0.01 X<=0.05 X<=0.1 all X

No 2.000 9.000 32.000 63.000 833
Prop 0.002 0.011 0.038 0.076 1

If you would like to investigate the minor allele frequency (MAF) distribu-
tion, the same logic would apply. First, derive the MAF with
> afr <- sumgt[, "Q.2"]
> maf <- pmin(afr, (1. - afr))
Next, generate histograms for frequency and MAF:
> par(mfcol=c(2,1))
> hist(afr)
> hist(maf)
(shown in Figure 4.2) and then generate a table describing the frequency distri-
bution:
> catable(afr, c(0.01, 0.05, 0.1, 0.2, 0.5, 0.8, 0.9, 0.95, 0.99))

X<=0.01 0.01<X<=0.05 0.05<X<=0.1 0.1<X<=0.2 0.2<X<=0.5 0.5<X<=0.8

No 22.000 53.000 99.000 132.000 313.000 187.000
Prop 0.026 0.064 0.119 0.158 0.376 0.224
0.8<X<=0.9 0.9<X<=0.95 0.95<X<=0.99 X>0.99
No 18.000 8.00 1.000 0
Prop 0.022 0.01 0.001 0

> catable(maf, c(0, 0.01, 0.05, 0.1, 0.2), cum=TRUE)

X<=0 X<=0.01 X<=0.05 X<=0.1 X<=0.2 all X

No 0 22.000 76.000 183.00 333.0 833
Prop 0 0.026 0.091 0.22 0.4 1

Note that we used “0” as the first category – this will give you the number of
monomorphic SNPs which we recommend to exclude from analysis.
Another function, [Link], produces summary SNP statistics per
person. Let us try producing this summary for the first 10 people:
> [Link](srdta[1:10, ])
94 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

Histogram of afr Histogram of maf

100
150

80
100
Frequency

Frequency

60
40
50

20
0

0.0 0.4 0.8 0.0 0.2 0.4

afr maf

Figure 4.2: Histogram of the call rate.

4.4. EXPLORING GENETIC DATA 95

NoMeasured NoPoly Hom E(Hom) Var F CallPP

p1 790 707 0.7987342 0.6600319 0.4048662 0.407986159 0.9483794
p2 792 714 0.7474747 0.6585152 0.5090002 0.260508049 0.9507803
p3 783 700 0.6206897 0.6618209 0.4332890 -0.121625581 0.9399760
p4 789 705 0.6070976 0.6601276 0.5251900 -0.156029161 0.9471789
p5 790 707 0.6658228 0.6619821 0.5288936 0.011362319 0.9483794
p6 787 703 0.7662008 0.6622227 0.3770418 0.307830275 0.9447779
p7 794 709 0.6309824 0.6587669 0.4527349 -0.081423884 0.9531813
p8 793 711 0.7023960 0.6587232 0.5163296 0.127968868 0.9519808
p9 788 711 0.6675127 0.6573272 0.5599395 0.029723748 0.9459784
p10 797 713 0.6587202 0.6614644 0.4889042 -0.008105999 0.9567827
Het
p1 0.2012658
p2 0.2525253
p3 0.3793103
p4 0.3929024
p5 0.3341772
p6 0.2337992
p7 0.3690176
p8 0.2976040
p9 0.3324873
p10 0.3412798

This table lists the number of genotypes scored for the person, call rate, and
heterozygosity. The outliers who have increased average heterozygosity may be
suggestive of contaminated DNA samples.
Let us analyse the distribution of heterozygosity:

> het <- [Link](srdta)$Het

> mean(het)

[1] 0.3309457

> catable(het, c(0.1, 0.25, 0.3, 0.35, 0.5))

X<=0.1 0.1<X<=0.25 0.25<X<=0.3 0.3<X<=0.35 0.35<X<=0.5 X>0.5

No 7.000 73.000 339.000 1281.000 800.00 0
Prop 0.003 0.029 0.136 0.512 0.32 0

> hist(het)

The resulting histogram is shown in Figure 4.3. It is easy to see that a few
people have very low heterozygosity, but there are no outliers with extremely
high values.
In this section, we covered the low-level functions summary and [Link].
On these functions a higher level genetic data quality control function, [Link],
is based. That function will be covered in the next section.

Summary:
96 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

Histogram of het
1200
1000
800
Frequency

600
400
200
0

0.0 0.1 0.2 0.3 0.4

het

Figure 4.3: Histogram of heterozygosity.

4.5. ANSWERS TO EXERCISES 97

• When the summary() function is applied to a gtdata subset of

[Link]-class, it returns summary statistics for the SNPs, including
an exact test for Hardy-Weinberg equilibrium.
• When the [Link]() function is applied to an object of
[Link]-class (or the gtdata part of it), it returns per-person sum-
mary statistics, including the call rate within this person and its heterozy-
gosity.

Exercise 5. Characterizing call rate

Characterise the distribution of call rates within study subjects and produce
a histogram. How many people have a call rate below 93%?

4.5 Answers to exercises

Answer (Ex. 1) — Load the data with

> data(srdta)
Number of people:
> nids(srdta)
[1] 2500
Number of males:
> sum(male(srdta))
[1] 1275
Number of females:
> nids(srdta) - sum(male(srdta))
[1] 1225
. . . you could get the same answer like this1 :
> sum(male(srdta) == 0)
[1] 1225
The proportion of males can be computed using above results
> sum(male(srdta)) / nids(srdta)
[1] 0.51
or by using the mean() function:
> mean(male(srdta))
[1] 0.51

1 This is something covered later in the section 4.3 (”Sub-setting and coercing [Link]”)
98 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

Answer (Ex. 2) — The names of markers located after 2,490,000 b.p. are
> vec <- (map(srdta) > 2490000)
> snpnames(srdta)[vec]
[1] "rs9273" "rs9277" "rs9279" "rs9283"
The names of markers located between 1,100,000 and 1,105,000 b.p. are:
> vec <- (map(srdta) > 1100000 & map(srdta) < 1105000)
> snpnames(srdta)[vec]
[1] "rs4180" "rs4186" "rs4187"

Answer (Ex. 3) — To learn what allele of “rs114” is the reference you need
to run
> coding(srdta)["rs114"]
rs114
"AT"
Here, the first (“A”) allele is the reference and thus the second (“T”) is the
effect allele. Remember that when using the [Link] function to convert
the genotypes to human-readable and R-operatable format, the homozygotes
for reference will be coded as “0”, heterozygotes as “1” and the non-reference
(“effect”) homozygotes will be coded as “2”:
> table([Link]( gtdata(srdta[, "rs114"] )),
+ [Link]( gtdata(srdta[,"rs114"] )))
0 1 2
A/A 1868 0 0
A/T 0 491 0
T/T 0 0 34
To compute the frequency of the effect allele of SNP “rs114” in the total sample,
you can go two ways. First, we can try to take a sum of all “rs114” genotypes
and divide it by twice the number of people:
> a <- [Link](gtdata(srdta[, "rs114"]))
> sum(a)
[1] NA
This, however, returns NA, because some of the genotypes are missing. We can
deal with this problem by running sum() with the option [Link]=TRUE:
> sum(a, [Link]=TRUE)
[1] 559
so the number of “effect” alleles is 559.
However, now we do not know what was the number of people for whom the
genotype was measured! – nids would return the total number of people, but
not the number of those measured for “rs114”.
This problem can be dealt with through using the [Link](A) function which
returns true when some element of A is not measured. Thus, the number of
people with measured genotype for “rs114” is
> nids(srdta)
[1] 2500
4.5. ANSWERS TO EXERCISES 99

> nmeasured <- sum()

> nmeasured
[1] 2393
(note the “!” before [Link], which means NOT, so we get these elements which
are not NA). Consequently, the frequency of the “effect” allele is
> sum(a, [Link]=TRUE) / (2 * nmeasured)
[1] 0.116799
An easier way would be to compute mean value of “rs114” with the mean( ...,
[Link]=TRUE) function and divide it by 2:
> mean(a, [Link]=TRUE)/2
[1] 0.116799
To compute the frequency of the effect allele of “rs114” in males, you can use
> amale <- [Link]( gtdata(srdta[male(srdta)==1, "rs114"]) )
> mean(amale, [Link]=TRUE)/2
[1] 0.1164216
To compute the frequency of the effect allele in females, you can use
> afemale <- [Link]( gtdata(srdta[male(srdta)==0, "rs114"]) )
> mean(afemale, [Link]=TRUE)/2
[1] 0.1171942
The frequencies of the reference allele are computed very simply as one minus
the frequency of the effective allele:
> 1 - mean(a, [Link]=TRUE)/2
[1] 0.883201
> 1 - mean(amale, [Link]=TRUE)/2
[1] 0.8835784
> 1 - mean(afemale, [Link]=TRUE)/2
[1] 0.8828058

Answer (Ex. 4) — To test for HWE in first 10 SNPs in total sample

> summary( gtdata(srdta[, 1:10]) )
Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11 P.12
rs10 1 2500 + T G 2384 0.9536 0.13255034 1792 552
rs18 1 3500 + G A 2385 0.9540 0.28029350 1232 969
rs29 1 5750 - G T 2374 0.9496 0.13774221 1763 568
rs65 1 13500 + A T 2378 0.9512 0.71972246 182 969
rs73 1 14250 + A G 2385 0.9540 0.01341719 2331 44
rs114 1 24500 + A T 2393 0.9572 0.11679900 1868 491
rs128 1 27000 - G T 2391 0.9564 0.02488499 2281 101
rs130 1 27250 + A G 2379 0.9516 0.69377890 222 1013
rs143 1 31000 + T G 2377 0.9508 0.47728229 655 1175
rs150 1 33250 + C A 2369 0.9476 0.65998312 267 1077
P.22 Pexact Fmax Plrt
100 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE

rs10 40 7.897327e-01 -0.006880004 7.355343e-01

rs18 184 7.608230e-01 -0.007017332 7.315304e-01
rs29 43 7.955141e-01 -0.007241148 7.227853e-01
rs65 1227 6.475412e-01 -0.010016746 6.246577e-01
rs73 10 1.792470e-12 0.303150234 1.281239e-12
rs114 34 7.663683e-01 0.005487764 7.894076e-01
rs128 9 9.408599e-06 0.129600629 1.000431e-05
rs130 1144 9.615127e-01 -0.002140946 9.168114e-01
rs143 547 6.512540e-01 0.009313705 6.497695e-01
rs150 1025 5.518478e-01 -0.012948436 5.281254e-01
To test it in cases
> summary( gtdata(srdta[phdata(srdta)$bt==1, 1:10]) )
Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11
rs10 1 2500 + T G 1197 0.9622186 0.13700919 888
rs18 1 3500 + G A 1189 0.9557878 0.28511354 605
rs29 1 5750 - G T 1176 0.9453376 0.14285714 859
rs65 1 13500 + A T 1185 0.9525723 0.72700422 83
rs73 1 14250 + A G 1187 0.9541801 0.01053075 1167
rs114 1 24500 + A T 1190 0.9565916 0.12184874 918
rs128 1 27000 - G T 1183 0.9509646 0.02409129 1129
rs130 1 27250 + A G 1188 0.9549839 0.68392256 117
rs143 1 31000 + T G 1192 0.9581994 0.48489933 320
rs150 1 33250 + C A 1182 0.9501608 0.66624365 127
P.12 P.22 Pexact Fmax Plrt
rs10 290 19 4.635677e-01 -0.024514202 3.871421e-01
rs18 490 94 7.759191e-01 -0.010949158 7.052930e-01
rs29 298 19 2.832575e-01 -0.034722222 2.214580e-01
rs65 481 621 4.647357e-01 -0.022595469 4.348023e-01
rs73 15 5 3.988770e-08 0.393614304 2.423624e-08
rs114 254 18 8.924018e-01 0.002606831 9.285104e-01
rs128 51 3 2.747904e-02 0.083175674 3.157174e-02
rs130 517 554 8.407527e-01 -0.006569292 8.207476e-01
rs143 588 284 6.848365e-01 0.012522119 6.654994e-01
rs150 535 520 5.568363e-01 -0.017756050 5.409408e-01
in controls
> summary( gtdata(srdta[phdata(srdta)$bt==0, 1:10]) )
Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11
rs10 1 2500 + T G 1177 0.9453815 0.12744265 897
rs18 1 3500 + G A 1185 0.9518072 0.27426160 623
rs29 1 5750 - G T 1188 0.9542169 0.13215488 897
rs65 1 13500 + A T 1183 0.9502008 0.71344041 98
rs73 1 14250 + A G 1188 0.9542169 0.01641414 1154
rs114 1 24500 + A T 1192 0.9574297 0.11157718 941
rs128 1 27000 - G T 1197 0.9614458 0.02589808 1141
rs130 1 27250 + A G 1181 0.9485944 0.70491109 104
rs143 1 31000 + T G 1174 0.9429719 0.46805792 334
rs150 1 33250 + C A 1176 0.9445783 0.65306122 139
P.12 P.22 Pexact Fmax Plrt
4.5. ANSWERS TO EXERCISES 101

rs10 260 20 7.933317e-01 0.006751055 8.178295e-01

rs18 474 88 9.418133e-01 -0.004812165 8.683219e-01
rs29 268 23 5.288436e-01 0.016525913 5.737373e-01
rs65 482 603 8.871139e-01 0.003540522 9.031273e-01
rs73 29 5 6.941219e-06 0.244001185 5.537568e-06
rs114 236 15 8.846527e-01 0.001356081 9.627084e-01
rs128 50 6 7.745807e-05 0.172107564 7.552399e-05
rs130 489 588 8.887439e-01 0.004728114 8.710047e-01
rs143 581 259 8.604122e-01 0.006165442 8.326938e-01
rs150 538 499 7.968462e-01 -0.009574142 7.424986e-01
SNPs ’rs73’ and ’rs128’ are out of HWE (at p ≤ 0.05) in the total sample, and
also in cases and controls.
Answer (Ex. 5) — To characterize ID call rate, you can run the following
commands:
> idsummary <- [Link](srdta)
> idsummary[1:5, ]
NoMeasured NoPoly Hom E(Hom) Var F CallPP
p1 790 790 0.7987342 0.6696986 0.5448255 0.3906601 0.9483794
p2 792 792 0.7474747 0.6685502 0.5390602 0.2381191 0.9507803
p3 783 783 0.6206897 0.6712102 0.4888671 -0.1536561 0.9399760
p4 789 789 0.6070976 0.6700900 0.4077382 -0.1909382 0.9471789
p5 790 790 0.6658228 0.6710232 0.4340010 -0.0158077 0.9483794
Het
p1 0.2012658
p2 0.2525253
p3 0.3793103
p4 0.3929024
p5 0.3341772
> idcall <- idsummary$Call
> idcall[1:5]
[1] 0.9483794 0.9507803 0.9399760 0.9471789 0.9483794
> catable(idcall, c(0.9, 0.93, 0.95, 0.98, 0.99))
X<=0.9 0.9<X<=0.93 0.93<X<=0.95 0.95<X<=0.98 0.98<X<=0.99 X>0.99
No 0 13.000 1186.000 1301.00 0 0
Prop 0 0.005 0.474 0.52 0 0
> table(idcall < 0.93)
FALSE TRUE
2487 13
To produce a histogram of call rates, use hist(idcall)
102 CHAPTER 4. INTRODUCTION TO THE GENABEL-PACKAGE
Chapter 5

Genome-wide association
analysis

In the first parts of this section you will be guided through a GWA analysis of a
small data set. In the last part you will investigate a larger data set by yourself,
do a verification study and will answer the questions. All data sets used assume
a study in a relatively homogeneous population. Try to finish the first part in
the morning and the second part in the afternoon.
Though only few thousands of markers located at four small chromosomes
are used in the scan, we still going to call it Genome-Wide (GW), as the amount
of data we will use is approaches the amount to be expected in a real experiment.
However, because the regions are small, and the LD between SNPs is high, some
specific features (e.g. relatively high residual inflation, which occurs because
large proportion of SNPs are in LD with the really associated ones) are specific
features of this data set, which are not observed in true GWA studies.
Start R and load the GenABEL-package library by typing

> library(GenABEL)

and load the data which we will use in this section by

> data(ge03d2ex)

Investigate the objects loaded by the command

> ls()

[1] "ge03d2ex"

The ge03d2ex is an object of the class [Link]:

> class(ge03d2ex)

[1] "[Link]"
attr(,"package")
[1] "GenABEL"

To check what are the names of variables in the phenotypic data frame, use

103
104 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

> names(phdata(ge03d2ex))

[1] "id" "sex" "age" "dm2" "height" "weight" "diet" "bmi"

We can attach this data frame to the R search path by

> attach(phdata(ge03d2ex))

5.1 Data descriptives and first round of GWA

analysis
Let us investigate which traits are present in the loaded data frame and what are
the characteristics of the distribution by using the GenABEL-package function
[Link]:
> [Link](ge03d2ex)

No Mean SD
id 136 NA NA
sex 136 0.529 0.501
age 136 49.069 12.926
dm2 136 0.632 0.484
height 135 169.440 9.814
weight 135 87.397 25.510
diet 136 0.059 0.236
bmi 135 30.301 8.082

You can see that the phenotypic frame contains data on 136 people; the
data on sex, age, height, weight, diet and body mass index (BMI) are available.
Our trait of interest is dm2 (type 2 diabetes). Note that every single piece
of information in this data set is simulated; however, we tried to keep our
simulations in a way we think the control of T2D may work.
You can produce a summary for cases and controls separately and compare
distributions of the traits by
> [Link](ge03d2ex, by=dm2)

No([Link]=0) Mean SD No([Link]=1) Mean SD Ptt Pkw

id 50 NA NA 86 NA NA NA NA
sex 50 0.420 0.499 86 0.593 0.494 0.053 0.052
age 50 47.038 13.971 86 50.250 12.206 0.179 0.205
dm2 50 NA NA 86 NA NA NA NA
height 49 167.671 8.586 86 170.448 10.362 0.097 0.141
weight 49 76.534 17.441 86 93.587 27.337 0.000 0.000
diet 50 0.060 0.240 86 0.058 0.235 0.965 0.965
bmi 49 27.304 6.463 86 32.008 8.441 0.000 0.001
Pexact
id NA
sex 0.074
age NA
dm2 NA
5.1. DATA DESCRIPTIVES AND FIRST ROUND OF GWA ANALYSIS105

height NA
weight NA
diet 1.000
bmi NA

Here, the by argument specifies the grouping variable. You can see that cases
and controls are different in weight, which is expected, as T2D is associated
with obesity.
Similarly, you can produce grand GW descriptives of the marker data by
using
> [Link](ge03d2ex)

$`Minor allele frequency distribution`

X<=0.01 0.01<X<=0.05 0.05<X<=0.1 0.1<X<=0.2 X>0.2
No 146.000 684.000 711.000 904.000 1555.000
Prop 0.036 0.171 0.178 0.226 0.389

$`Cumulative distr. of number of SNPs out of HWE, at different alpha`

X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 46.000 71.000 125.000 275.000 4000
Prop 0.012 0.018 0.031 0.069 1

$`Distribution of proportion of successful genotypes (per person)`

X<=0.9 0.9<X<=0.95 0.95<X<=0.98 0.98<X<=0.99 X>0.99
No 1.000 0 0 135.000 0
Prop 0.007 0 0 0.993 0

$`Distribution of proportion of successful genotypes (per SNP)`

X<=0.9 0.9<X<=0.95 0.95<X<=0.98 0.98<X<=0.99 X>0.99
No 37.000 6.000 996.000 1177.000 1784.000
Prop 0.009 0.002 0.249 0.294 0.446

$`Mean heterozygosity for a SNP`

[1] 0.2582298

$`Standard deviation of the mean heterozygosity for a SNP`

[1] 0.1592255

$`Mean heterozygosity for a person`

[1] 0.2476507

$`Standard deviation of mean heterozygosity for a person`

[1] 0.04291038

It is of note that we can see inflation of the proportion of the tests for HWE at
a particular threshold, as compared to the expected. This may indicate poor
genotyping quality and/or genetic stratification.
We can test the GW marker characteristics in controls by
> [Link](ge03d2ex, ids=(dm2==0))
106 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

$`Minor allele frequency distribution`

X<=0.01 0.01<X<=0.05 0.05<X<=0.1 0.1<X<=0.2 X>0.2
No 233.000 676.000 671.000 898.000 1522.00
Prop 0.058 0.169 0.168 0.224 0.38

$`Cumulative distr. of number of SNPs out of HWE, at different alpha`

X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 0 3.000 14.000 98.000 4000
Prop 0 0.001 0.003 0.024 1

$`Distribution of proportion of successful genotypes (per person)`

X<=0.9 0.9<X<=0.95 0.95<X<=0.98 0.98<X<=0.99 X>0.99
No 0 0 0 50 0
Prop 0 0 0 1 0

$`Distribution of proportion of successful genotypes (per SNP)`

X<=0.9 0.9<X<=0.95 0.95<X<=0.98 0.98<X<=0.99 X>0.99
No 37.000 49.000 1523.000 0 2391.000
Prop 0.009 0.012 0.381 0 0.598

$`Mean heterozygosity for a SNP`

[1] 0.2555009

$`Standard deviation of the mean heterozygosity for a SNP`

[1] 0.1618707

$`Mean heterozygosity for a person`

[1] 0.252572

$`Standard deviation of mean heterozygosity for a person`

[1] 0.04714886
Apparently, HWE distribution holds better in controls than in the total sample.
Let us check whether there are indications that deviation from HWE is due
to cases. At this stage we are only interested in the HWE distribution table,
and therefore will ask to report the distrbution for cases (dm2==1) and report
only table two:
> [Link](ge03d2ex, ids=(dm2==1))[2]
$`Cumulative distr. of number of SNPs out of HWE, at different alpha`
X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 45.000 79.00 136.000 268.000 4000
Prop 0.011 0.02 0.034 0.067 1
and for the controls
> [Link](ge03d2ex, ids=(dm2==0))[2]
$`Cumulative distr. of number of SNPs out of HWE, at different alpha`
X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 0 3.000 14.000 98.000 4000
Prop 0 0.001 0.003 0.024 1
5.1. DATA DESCRIPTIVES AND FIRST ROUND OF GWA ANALYSIS107

It seems that indeed excessive number of markers are out of HWE in cases. If no
laboratory procedure (e.g. DNA extraction, genotyping, calling) were done for
cases and controls separately, this may indicate possible genetic heterogeneity
specific for cases.
In essence, the ’[Link]’ function uses the ’summary’ function
to generate the HW P -values distribution. It may be interesting to generate
this distribution using the ’summary’ function You do not need to do so, but this
example shows how you can generate summaries from underlying SNP-tables.
First, we need to compute summary SNP statistics by

> s <- summary( gtdata( ge03d2ex[(dm2==1), ] ) )

Note the you have produced the summary for the gtdata slot of ge03d2ex;
this is the slot which actually contain all genetic data in special compressed
format.
You can see the first 5 rows of this very long summary table by

> s[1:5, ]

Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11

rs1646456 1 653 + C G 85 0.9883721 0.32941176 34
rs4435802 1 5291 + C A 86 1.0000000 0.09302326 70
rs946364 1 8533 - T C 84 0.9767442 0.24404762 46
rs299251 1 10737 + A G 85 0.9883721 0.03529412 79
rs2456488 1 11779 + G C 85 0.9883721 0.33529412 38
P.12 P.22 Pexact Fmax Plrt
rs1646456 46 5 0.05089075 -0.22493734 0.03233812
rs4435802 16 0 1.00000000 -0.10256410 0.19978917
rs946364 35 3 0.37413083 -0.12924909 0.21777250
rs299251 6 0 1.00000000 -0.03658537 0.63937447
rs2456488 37 10 0.81059718 0.02344356 0.82918345

Note that the column ’Pexact’ provides exact HWE test P -values we need. We
can extract these to a separate vector by

> pexcas <- s[, "Pexact"]

and do characterization of the cummulative distribution by

> catable(pexcas, c(0.001,0.01,0.05,0.1,0.5), cumulative=TRUE)

X<=0.001 X<=0.01 X<=0.05 X<=0.1 X<=0.5 all X

No 79.00 136.000 268.000 390.000 1359.00 4000
Prop 0.02 0.034 0.067 0.098 0.34 1

You can generate the distribution for controls in similar manner.

Let us first try do a GWA scan using the raw (before quality control)
data. We will use the score test, as implemented in the qtscore()1 function of
GenABEL-package for testing:

> an0 <- qtscore(dm2, ge03d2ex, trait="binomial")

1 consider ’mlreg’ function if you want to run a true linear regression
108 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

The first argument used describes the model; here it is rather simple — the
affection status, dm2, is supposed to depend on SNP genotype only.
You can see what information is computed by this function by using

> an0

***** '[Link]' object *****

*** Produced with:
qtscore(formula = dm2, data = ge03d2ex, [Link] = "binomial")
*** Test used: binomial
*** no. IDs used: 136 ( id199 id287 id300 , ... )
*** Lambda: 1.033102
*** Results table contains 4000 rows and 9 columns
*** Output for 10 first rows is:
N effB se_effB chi2.1df P1df effAB effBB
rs1646456 135 0.9487666 5.0501959 0.0352941176 0.8509807 1.5558824 0.4831933
rs4435802 134 2.6822601 1.6765631 2.5595403237 0.1096305 2.5142857 NA
rs946364 134 0.6376645 0.4012104 2.5260391714 0.1119810 0.7277883 0.2869565
rs299251 135 0.5592122 0.5740215 0.9490674319 0.3299568 0.5569620 NA
rs2456488 135 0.8669860 1.5393112 0.3172278551 0.5732784 0.9736842 0.6907895
rs3712159 133 0.8282737 2.4803134 0.1115153279 0.7384255 0.5641026 Inf
rs4602970 136 1.5227297 2.2683336 0.4506421219 0.5020302 1.5131579 NA
rs175910 134 0.9949826 57.3970087 0.0003005055 0.9861693 0.5600000 4.0727273
rs1919938 136 0.9303079 3.5515057 0.0686164762 0.7933619 0.1160000 0.1958333
rs8892781 133 1.0953022 14.9055712 0.0053997133 0.9414220 1.0952381 NA
chi2.2df P2df
rs1646456 3.885667042 0.14329734
rs4435802 2.559540324 0.10963046
rs946364 3.020970009 0.22080286
rs299251 0.949067432 0.32995679
rs2456488 0.493411146 0.78137072
rs3712159 1.358996877 0.50687116
rs4602970 0.450642122 0.50203018
rs175910 5.012993241 0.08155345
rs1919938 6.001763944 0.04974318
rs8892781 0.005399713 0.94142198
...
___ Use 'results(object)' to get complete results table ___

Here, let us look at the ’Results table’. P1df, P2df and Pc1df are most inter-
esting; the first two are vectors of 1 and 2 d.f. P -values obtained in the GWA
analysis, the last one is 1 d.f. P -value corrected for inflation factor λ (which
is in the lambda object). effB corresponds to the (approximate) Odds Ratio
estimate for the SNP.
Let us see if there is evidence for the inflation of the test statistics; for that
let us obtain λ with

> lambda(an0)

$estimate
[1] 1.033102
5.1. DATA DESCRIPTIVES AND FIRST ROUND OF GWA ANALYSIS109

● ●
12

● ●
●
●●
10

●
●●
●
●●●
●●
●●●
Observed χ2

●
8

●
●●
●●
●
●
●●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●●
6

●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●●
●
●●
●
●●
●
●
●●
●
4

●●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
2

●●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
0

0 5 10 15

Expected χ2

Figure 5.1: χ2 − χ2 plot for a GWA scan. Black line of slope 1: expected under
no inflation; Red line: fitted slope.

$se
[1] 0.0005639231

The estimate of λ is 1.03, suggesting inflation of the test and some degree of
stratification. Though the value obtained seems to be small, it should be noted
that λ grows linearly with sample size, so for this small number of cases and
controls the value is worrisome.
The λ is computed by regression in a Q-Q plot. Both estimation of λ and
production of the χ2 − χ2 plot can be done using the estlambda function; this
was already done automatically when running qtscore function, but let us
repeat this manually:

> estlambda(an0[, "P1df"], plot=TRUE)

$estimate
[1] 1.033102

$se
[1] 0.0005639231

The corresponding χ2 − χ2 plot is shown in Figure 5.1.

110 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

qtscore(dm2, ge03d2ex, binomial)

− log10(P − value)
● ●
●

3
● ● ●
●
●● ● ● ● ● ● ●
●
●
● ●● ● ●
● ●
● ● ● ● ●● ● ●● ●
● ●●● ● ● ●● ●● ● ● ● ●●●●
●

2
●●● ● ● ●●
● ●● ● ●●● ● ● ●●
●
●
● ●
● ●● ●
● ●●● ● ● ● ●●● ● ●
● ●●● ●● ●●●
● ● ●●●● ● ● ● ● ●● ● ●
●●●
● ●● ●
●● ● ●
●●
●●●●● ● ● ● ●●● ●●● ●
● ● ● ●●● ●●●●● ●●
●
●
●●● ●●●●●● ●● ●
● ●● ●● ● ●
●●●●●
●● ●
●●●●●● ● ●●●●
●●●●
●●
● ●●●
● ● ● ● ●●●●
● ●● ●
● ●
●●●●●●
● ●
●●●
●● ●●●●
● ●
●●● ●
●●●●●
● ●●
●● ●
●
●●●
● ● ●●●●
●●● ●
●
●●
● ●
●● ●
● ●●
●●●
● ●
●●
●
●
●
●●
●●●● ●●
●●●●
●●● ●●●
●●
●●●●
●
●
●
●●
● ●●
●
●
● ●
●● ● ● ●●●● ●
● ●● ● ● ●●
● ● ● ● ●●●● ● ●
●
●

1
●
●
●●● ●
●●●●●●●●
●●
●
● ● ● ●
●●
●●●
●●●
●●
●
● ●●●
●●●
●
●
● ●●
●
●●
●●
●
●●
● ●
●
●●
●●●
●●
●●
●
● ●
●●
●
●●● ●
●
●●●●●
●●●●●●
●
●●●●
●●
●
●
●●●
● ●●●●
●●
●●●
●● ●●●
●●●
●●
●
●●
●
● ●●
●
●●●●
●
●●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●●●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●●
●
● ●
●
●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●
●
●●●●
●
●
●
●
●●●●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●●
●
●●
●
●●
●●
●●
●
●
●●
●
●
● ●
●●
●
●●
●
●●
●●
●
●●
●●
●
●
●●●
●
●●
●
●●●
●
●
●
●●
●
●
●
●●
●●
● ●
●●
●●
●●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●
●● ●●
●
●
●●
●●
●
●
●●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
● ●
●
●●
●●
●
●●
●
●●
●
●●
●●
●●
●●
●
●●
●
●●
●●
● ●
●
●●
●
●●
●
●●
●
●
●●
●●●
●●●●
●
●●
●
●●
●
●●
●
●
●
●●
●●
●
●
●
●●●
● ●
●
●
●●
●
●●
●
●
●●
●●
●
●
●
●
●
●●●
●
●
●
●●●
●
●
●●
●
●
●
●
●●
●●●
●
●●●
●
●
●
●
●●
●
●●
●
●●●
●
●
●●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●●
●
●
●●
●●●
●●
●
●●
●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●●
●
●
●
●●●
●
●●
●
●●
●●
●
●
●●
● ●
●
●●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
● ●
●
●
●
●●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●●
●
● ●
●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●●
●●
●
●
●●
●
●●
●
●
●
●●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●●
●●
●
●
●●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●
●●
●●
●
●●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●
●●
● ●●
●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
● ●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●●
●
●●
●
●
● ●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
● ●
●●
●●●
●●
● ●
●
● ●
●●●●●●
●
●●●
●●●
●●●
●● ●●●
●●●
●●●●
●●
●●
●●
●
●●
●●
●
●●●
●●
●●
●
●●
●● ●●●
●●●
●●
●●●
● ● ●
●
●●●●●●●
●●● ●
●
●●
●● ●●
●●●
●●●●● ●●
●●
●●
●●
●●
● ●
●
●●
●
●●●
●
●●
●●
●
0

1 2 3 X

Chromosome

qtscore(dm2, ge03d2ex, binomial)

− log10(P − value)

● ●
3

●● ● ●●
●● ● ● ● ● ● ●
●
●
● ●● ● ●
● ●
● ● ● ● ●● ● ●● ●
●● ● ● ● ●● ● ● ● ●●●●
●
2

● ●● ● ● ●●● ● ● ●● ●● ● ●●
●
●
●●●
● ●● ●● ● ●●
● ● ●●●● ● ●●
● ● ●● ● ●● ● ● ● ●● ●
● ●● ●
● ●
●● ●● ● ● ● ● ●
●●● ● ●●● ● ●● ● ●● ●● ●
●
●●●
● ● ●●● ●●
●
● ●●
●● ●●● ●●
●●●
●●
●● ●●
●●● ● ● ●●●
●
●●●● ●
●● ● ● ●●
● ● ●●● ● ● ●●
●
● ●
●●●●●
●●●● ●● ●●●●● ●
●●●● ●
●●●●●
● ●●
●● ●
●
●●
●●● ●●● ●
●●
● ●
●
● ●
●●●●●
●● ● ●●●●
●●●
●
● ●●
●
●●●●● ●
●●●
● ●●
●● ●●●●
●●
●●●●●
●
● ●
●●
●
●●●●
●
● ●
● ● ●●●●● ●●●●
● ● ●● ●● ● ● ●●
● ● ● ●
●● ● ●● ● ●
●● ●● ●
1

●
●
●●●●
● ●●
●●
● ●●●●
●●●
●● ●
●●
●●●
●●
●●●
●●
● ●●●
●●●
● ●●
●
●●
●●●
●
●● ●
●
●●
● ●●
●●
●●
●
● ●
●
●
●●●●
●● ●
●
● ●● ●
●●
●●●●●
●
●●●
●●●
●
●
●●
●●● ●●●●●
● ●
●● ●●●●
●●
●●
●●
●
●●
●
● ●
●
●
●●
●● ●●
●●●
●
●●●
●●
●●
●●●●
●●
●● ●●
●●
● ●
●●
●●
●
●●●
●●
●
●●● ●●●
●●●
●
●●
●●●●
●
●●
●
●
●
●●
●
●●●
●
●
●●
●● ●●●
●● ●
●●
●
●●
●
●●
●
●●●●●
●
●
●●
●●●
● ●●
●●
●
●● ●
●
●●
●
●●
●
●
●●
●●●●
●
●
●
● ●●
●
●●●●
●
●●●●
●●●
●●●
●
●
● ●
●●
●●
●
●●
● ●●
●●
●●
●
●●●
●
●●
●●
●
●●●
●
●
●●
●
●●
●
●
●●
●●
●●●●●
●●●
●
●
●●
●
●
●●●
●●
●
●
●●
●●●
●●
●●
●
●
●●
●
●
●
● ●●●●
●● ●
●
●
●●
●
●
●
●●
●
●●
●
●●
●●●
●●●
●
●●●
●
●●●●
●
●●
●● ●
●●
●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●●
●●●●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●●
●
●●●
● ●
●●
●●
●
●●
●
●●
●
●●
●●
●●●●
●
●●●
●
●●
●
●
●●
●
●●
●
●●
●● ●
●●
●
●●
●
●
●●
●
●●●
●
●
●
●
●●
●●
●●
●
●●
●
●●
●
●●
●●
●●
●
● ●
●
●
●●
●
●●
●●
●
●●
●
●●
●
●
●●
●●
●●
●
●
●
●●
●
●●●
●●
●
●
●●
●
●
● ●
●
●●●●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●● ●
●●
●●
●
●
●
●● ●
●
●
●●
●
●●
●
●●
●
●
●
●●
●●
●●
●
●
●
●●
●●
●
●●
●
●
●●
●●
●
●●
●●
●
●●
●
●●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●●
● ●
●
●
●
●●
●
●
●●
●●
●
●
●
●●●●
●●
●
●
●●
●
●●
●
●●
●
●●●
●
●
●
●●
●●
●
●●
● ●
●
●●
●
●●
●
●
●●●
●
●●●
●
●
●
●●●
●
●●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●●
●
●●●
●●
●
●
●
●●
●
●●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●●●
●●
●●
●●
●
●●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●●●
●
●
●
●●
●
●●
● ●
●●●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●●
●
●
●
● ●
●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
● ●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●●●
●
●●
●
●●
● ●
●
●●
●
●
●
●
●●
●
●
●
●●
●●●
●●
●
●●
●
●
●●●●
● ●●
●
●●
●
●●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●●● ●
●●●
●●●
●
●●●●●
●●●
●
●●●●
●
●
●●
●●
●●
●
●●
● ●
●●
●
●
●
●●
●
●●
●
●
●●●
●
●
●●
●
●
●●
● ●
●●
●
●
●●●
●
●
●●●●●
●
●●
●●●
●● ●
●
●
●●
●●
●●●
●
●
●●●●
●●
●
●●
●
●
●
●●
●
●●
●●
●
●●●
● ●●
●
●●
●
●●
●
●●
●
●●
●
●●
●●● ●
●
●●
●●
●
●●●●
●●
●
●
●
●●●
●
●
●
●●
●
●●
●
●
●● ●
●
●
●
●●
●
●●
●
●
●●●
●
●
●●
●
●●●●
●
●●
●
●
●
●●●
●
●●
●
●●
●
●
●●●
●
●
●●●
●●
●
●
●
●●●
● ●
●
●
●
●●
● ●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●●●
●●
●●●
●
●
●
●
●
0

1 2 3 X

Chromosome

Figure 5.2: − log10 (P -value) from the genome scan before QC procedure. Raw
analysis: darker circles; corrected analysis: lighter circles

The ’se’ produced by estlambda can not be used to test if inflation is signif-
icant and make conclusions about the presence of significant or insignificant
stratification.

We can also present the obtained results using the ”Manhatten plot”, where
the SNP genomic position is on the horizontal axis and − log10 of the P -value
is shown on the vertical axis:
> plot(an0)
The resulting plot is show in Figure 5.2. By default, − log10 (P -value) of the
uncorrected 1 d.f. test are shown; see thehelp to figure out how this behaviour
can be changed.
We can also add the corrected P -values to the plot with
> [Link](an0, df="Pc1df", col=c("lightblue", "lightgreen"))
You can see that the P -values corrected by genomic control are uniformly lower
than the P -values from ’raw’ analysis. This is to be expected as genomic control
simply divides the ’raw’ χ2 statistics by a constant λ for all SNPs.
You can also generate a descriptive table for the ”top” (as ranked by P -value)
results by
5.1. DATA DESCRIPTIVES AND FIRST ROUND OF GWA ANALYSIS111

> [Link](an0)

Summary for top 10 results, sorted by P1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs1719133 1 4495479 + T A 136 0.33729339 0.09282784 13.202591
rs2975760 3 10518480 + A T 134 3.80380024 1.05172986 13.080580
rs7418878 1 2808520 + A T 136 3.08123060 0.93431795 10.875745
rs5308595 3 10543128 - C G 133 3.98254950 1.21582875 10.729452
rs4804634 1 2807417 + C G 132 0.43411456 0.13400290 10.494949
rs3224311 2 6009769 + G C 135 3.15831710 0.98401491 10.301681
rs26325 3 10617781 + A C 135 0.09742793 0.03035964 10.298495
rs8835506 2 6010852 + A T 132 3.17720829 1.00274087 10.039543
rs3925525 2 6008501 + C G 135 2.98416931 0.96286458 9.605423
rs2521089 3 10487652 - T C 135 2.50239493 0.81179595 9.502064
P1df effAB effBB chi2.2df P2df Pc1df
rs1719133 0.0002795623 0.4004237 0.000000 14.729116 0.0006333052 0.0003504258
rs2975760 0.0002983731 3.4545455 10.000000 13.547345 0.0011434877 0.0003732694
rs7418878 0.0009743183 3.6051282 4.871795 12.181064 0.0022642036 0.0011762545
rs5308595 0.0010544366 3.3171429 Inf 10.766439 0.0045930101 0.0012699705
rs4804634 0.0011970132 0.5240642 0.173913 11.200767 0.0036964462 0.0014362332
rs3224311 0.0013290907 3.4151786 4.250000 11.658283 0.0029405999 0.0015897278
rs26325 0.0013313876 0.1097724 NA 10.298495 0.0013313876 0.0015923930
rs8835506 0.0015321522 3.4903846 4.125000 11.513206 0.0031618340 0.0018248521
rs3925525 0.0019400358 3.2380952 4.121212 10.782867 0.0045554384 0.0022944719
rs2521089 0.0020524092 2.5717703 4.772727 9.933387 0.0069661425 0.0024233145

or, equivalently, by ’summary(an0)’

Here you see top 10 results, sorted by P -value with 1 d.f. If you want to sort
by the corrected P -value, you can use [Link](an0, sort="Pc1df");
to see more than 10 (e.g. 25) top results, use [Link](an0, top=25).
You can combine all these options. Large part of results reports NA as effect
estimates and 9.99 as P -value for 2 d.f. test – for these markers only two out
of the three possible genotypes were observed, and consequently the 2 d.f. test
could not be performed.
Now let us apply the qtscore() function with times argument, which tells
it to compute empirical GW (or experiment-wise) significance
> an0.e <- qtscore(dm2, ge03d2ex, times=200, quiet=TRUE)

|
| | 0%
|
|======================================================================| 100%

(you may skip the ’quiet=TRUE’ argument, then you will see progress)
Now let us generate the summary of the results
> [Link](an0.e, sort="Pc1df")

Summary for top 10 results, sorted by Pc1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs1719133 1 4495479 + T A 136 -0.2652064 0.07298850 13.202591
112 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

rs2975760 3 10518480 + A T 134 0.2340655 0.06471782 13.080580

rs7418878 1 2808520 + A T 136 0.2089098 0.06334746 10.875745
rs5308595 3 10543128 - C G 133 0.2445516 0.07465893 10.729452
rs4804634 1 2807417 + C G 132 -0.2050449 0.06329344 10.494949
rs3224311 2 6009769 + G C 135 0.2133633 0.06647611 10.301681
rs26325 3 10617781 + A C 135 -0.4875367 0.15192190 10.298495
rs8835506 2 6010852 + A T 132 0.2112000 0.06665565 10.039543
rs505606 1 2807572 - G T 134 0.2247738 0.07384377 9.265392
rs3925525 2 6008501 + C G 135 0.2057095 0.06637371 9.605423
P1df Pc1df effAB effBB chi2.2df P2df
rs1719133 0.440 0.525 -0.2080882 -0.7375000 14.72912 NA
rs2975760 0.460 0.545 0.2755102 0.4090909 13.54734 NA
rs7418878 0.865 0.920 0.2807405 0.3268398 12.18106 NA
rs5308595 0.885 0.930 0.2564832 0.4623656 10.76644 NA
rs4804634 0.925 0.945 -0.1193830 -0.3845238 11.20077 NA
rs3224311 0.930 0.955 0.2778634 0.3151515 11.65828 NA
rs26325 0.930 0.960 -0.4875367 NA 10.29849 NA
rs8835506 0.955 0.975 0.2796221 0.3076923 11.51321 NA
rs505606 0.995 0.995 0.2731481 0.2814815 10.27936 NA
rs3925525 0.980 0.995 0.2660834 0.3074627 10.78287 NA

Experimental-wise significance is computed by an empirical procedure, thus

we consider P -values ≤ 0.05 to be GW-significant. However, none of the SNPs
hits GW significance. If, actually, any did pass the threshold, we could not
trust the results, because the distribution of the HWE test and presence of
inflation factor for the association test statistics suggest that the data may
contain multiple errors (indeed they do). Therefore before association analysis
we need to do rigorous Quality Control (QC).
Note that at a certain SNP, the corrected P -values become equal to 1 – at
this point the order in the list is arbitrary because sorting could not be done.

Summary:
• The descriptives family of functions was developed to facilitate the
production of tables which can be directly used in a manuscript — it is
possible to save the output as a file, which can be open by Excel or Word.
See e.g. help([Link]) for details.

• The inflation of test statistics compared to null (1 d.f.) may be estimated

with estlambda function.

5.2 Genetic data QC

The major genetic data QC function of GenABEL-package is [Link]().
We will now use that to perform our data QC; the output is rather self-explaining.
Because of possible genetic heterogeneity of the study data it is a good idea to
5.2. GENETIC DATA QC 113

skip Hardy-Weinberg checks in the first round of QC. This can be achieved by
setting HWE P -value selection threshold to zero ([Link]=0):
> qc1 <- [Link](ge03d2ex, [Link]=0)
Excluding people/markers with extremely low call rate...
4000 markers and 136 people in total
0 people excluded because of call rate < 0.1
6 markers excluded because of call rate < 0.1
Passed: 3994 markers and 136 people

Running sex chromosome checks...

197 heterozygous X-linked male genotypes found
1 X-linked markers are likely to be autosomal (odds > 1000 )
2 male are likely to be female (odds > 1000 )
0 female are likely to be male (odds > 1000 )
0 people have intermediate X-chromosome inbreeding (0.5 > F > 0.5)
If these people/markers are removed, 0 heterozygous male genotypes are left
Passed: 3993 markers and 134 people

no X/Y/mtDNA-errors to fix

RUN 1
3993 markers and 134 people in total
304 (7.613323%) markers excluded as having low (<1.865672%) minor allele frequency
36 (0.9015778%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (P <0)
1 (0.7462687%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2747262 (s.e. 0.03721277)
3 (2.238806%) people excluded because too high autosomal heterozygosity (FDR <1%)
Excluded people had HET >= 0.4856887
Mean IBS is 0.7709982 (s.e. 0.02094168), as based on 2000 autosomal markers
2 (1.492537%) people excluded because of too high IBS (>=0.95)
In total, 3653 (91.4851%) markers passed all criteria
In total, 128 (95.52239%) people passed all criteria

RUN 2
3653 markers and 128 people in total
80 (2.189981%) markers excluded as having low (<1.953125%) minor allele frequency
0 (0%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (P <0)
0 (0%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2748341 (s.e. 0.01695461)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7745814 (s.e. 0.01793637), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 3573 (97.81002%) markers passed all criteria
In total, 128 (100%) people passed all criteria

RUN 3
114 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

3573 markers and 128 people in total

0 (0%) markers excluded as having low (<1.953125%) minor allele frequency
0 (0%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (P <0)
0 (0%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2748341 (s.e. 0.01695461)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7714028 (s.e. 0.01812413), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 3573 (100%) markers passed all criteria
In total, 128 (100%) people passed all criteria

Note that normally you will NEVER run this simple form of the QC function –
you should always provide a number of thresholds specific to the platform you
used for genotyping. See help to [Link]() for detailed list of arguments.
The default values used by the function are rather relaxed compared to the
thresholds routinely used nowadays with most of the platforms.

The computation of all pairwise proportion of alleles identical-by-state (IBS)

by ibs() function, which is also called by [Link]() may take quite some
time, which is proportional to the square of the number of subjects. This is
not a problem with the small number of people we use for this example or
when modern computers are used. However, the computers in the computer
room are very old. Therefore be prepared to wait for long time when you
will do a self-exercise with 1,000 people.

From the output you can see that QC starts with checking the data for SNPs
and people with extremely low call rate. Six markers are excluded from further
analysis due to very low call rate. Next, X-chromosomal errors are identified.
The function finds out that all errors (heterozygous male X-genotypes) are due
to two people with wrong sex assigned and one marker, which looks like an
autosomal one. This actually could be a marker from the pseudoautosomal
region, which should have been arranged as a separate ”autosome”. Nine people
are found to have intermediate inbreeding at the X-chromosome and are also
excluded from analysis.
Then, the procedure finds the markers with low call rate (≤ 0.95 by default)
across people, markers with low MAF (by default, low MAF is defined as less
than a few copies of the rare allele, see help for details); people with low call
rate (default value: ≤ 0.95) across SNPs, people with extreme heterozygosity
(at FDR 0.01) and those who have GW IBS ≥ 0.95. These default parameters
may be changed if you wish (consult the help).
Because some of the people fail to pass the tests, the data set is not guar-
anteed to be really ”clean” after single iteration, e.g. some marker may not
pass the call threshold after we exclude few informative (but apparently having
low quality) samples. Therefore the QC is repeated iteratively until no further
errors are found.
You can generate a short summary of the QC by marker and by person
through
5.2. GENETIC DATA QC 115

> summary(qc1)
$`Per-SNP fails statistics`
NoCall NoMAF NoHWE Redundant Xsnpfail
NoCall 42 0 0 0 0
NoMAF NA 384 0 0 0
NoHWE NA NA 0 0 0
Redundant NA NA NA 0 0
Xsnpfail NA NA NA NA 1

$`Per-person fails statistics`

IDnoCall HetFail IBSFail isfemale ismale isXXY otherSexErr
IDnoCall 1 0 0 0 0 0 0
HetFail NA 3 0 0 0 0 0
IBSFail NA NA 2 0 0 0 0
isfemale NA NA NA 2 0 0 0
ismale NA NA NA NA 0 0 0
isXXY NA NA NA NA NA 0 0
otherSexErr NA NA NA NA NA NA 0
Note that the original data, ge03d2ex, are not modified during the procedure;
rather, [Link]() generate a list of markers and people which pass or
do not pass certain QC criteria. The objects returned by [Link]() are:
> names(qc1)
[1] "nofreq" "nocall" "nohwe" "Xmrkfail" "hetfail"
[6] "idnocall" "ibsfail" "isfemale" "ismale" "otherSexErr"
[11] "snpok" "idok" "call"
The element idok provides the list of people who passed all QC criteria,
and snpok provides the list of SNPs which passed all criteria. You can easily
generate a new data set, which will consist only of these people and markers by
> data1 <- ge03d2ex[qc1$idok, qc1$snpok]
If there are any residual sporadic X-errors (male heterozygosity), these can
(and should!) be fixed (set to NA) by
> data1 <- Xfix(data1)

no X/Y/mtDNA-errors to fix
Applying this function does not make any difference for the example data set,
but you will need to use it for the bigger data set.
At this point, we are ready to work with the new, cleaned, data set data1.
However, if we try
> length(dm2)
[1] 136
we can see that the original phenotypic data are attached to the search path
(there are only 128 people left in the ’clean’ data set). Therefore we need to
detach the data by
116 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

> detach(phdata(ge03d2ex))
and attach new data by
> attach(phdata(data1))
At this stage, let us check if the first round of QC improves the fit of genetic
data to HWE, which may have been violated due to by genotyping errors which
we hopefully (at least partly!) eliminated:
> [Link](data1)[2]

$`Cumulative distr. of number of SNPs out of HWE, at different alpha`

X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 44.000 66.000 117.000 239.000 3573
Prop 0.012 0.018 0.033 0.067 1

> [Link](data1[dm2==1, ])[2]

$`Cumulative distr. of number of SNPs out of HWE, at different alpha`

X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 46.000 70.00 127.000 228.000 3573
Prop 0.013 0.02 0.036 0.064 1

> [Link](data1[dm2==0, ])[2]

$`Cumulative distr. of number of SNPs out of HWE, at different alpha`

X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 0 0 7.000 91.000 3573
Prop 0 0 0.002 0.025 1

You can see that the fit to HWE improved, but cases are still have excess number
of markers out of HWE. This may be due to genetic sub-structure.

5.3 Finding genetic sub-structure

Now, we are ready for the second round of QC – detection of genetic outliers
which may contaminate our results. We will detect genetic outliers using a
technique, which resembles the one suggested by Price at al.
As the first step, we will compute a matrix of genomic kinship between all
pairs of individuals, using only autosomal2 markers by
> [Link] <- ibs(data1[, autosomal(data1)], weight="freq")

This step may take few minutes on large data sets or when using old com-
puters!

You can see the 5 × 5 upper left sub-matrix by

2 the list of autosomal markers contained in data is returned by the autosomal(data)

function
5.3. FINDING GENETIC SUB-STRUCTURE 117

> [Link][1:5, 1:5]

id199 id300 id403 id415 id666

id199 0.494427766 3255.00000000 3253.00000000 3241.00000000 3257.0000000
id300 -0.011754266 0.49360296 3261.00000000 3250.00000000 3264.0000000
id403 -0.012253378 -0.01262949 0.50541775 3247.00000000 3262.0000000
id415 -0.001812109 0.01388179 -0.02515438 0.53008236 3251.0000000
id666 -0.018745051 -0.02127344 0.02083723 -0.02014175 0.5306584

The numbers below the diagonal show the genomic estimate of kinship (aka
’genomic kinship’ or ’genome-wide IBD’), the numbers on the diagonal corre-
spond to 0.5 plus the genomic homozigosity, and the numbers above the diagonal
tell how many SNPs were typed successfully for both subjects (thus the IBD
estimate is derived using this number of SNPs).
Second, we transform this matrix to a distance matrix using standard R
command
> [Link] <- [Link]([Link])
Finally, we perform Classical Multidimensional Scaling by
> [Link] <- cmdscale([Link])
By default, the first two principal components are computed and returned.

This may take few minutes on large data sets or when using old computers!

We can present the results graphically by

> plot([Link])
The resulting plot is shown in Figure 5.3. Each point on the plot corresponds
to a person, and the 2D distances between points were fitted to be as close as
possible to those presented in the original IBS matrix. You can see that study
subjects clearly cluster in two groups.
You can identify the points belonging to clusters by
> km <- kmeans([Link], centers=2, nstart=1000)
> cl1 <- names(which(km$cluster==1))
> cl2 <- names(which(km$cluster==2))
> if (length(cl1) > length(cl2)) {x<-cl2; cl2<-cl1; cl1<-x}
> cl1

[1] "id2097" "id6954" "id2136" "id858"

> cl2

[1] "id199" "id300" "id403" "id415" "id666" "id689" "id765" "id830"

[9] "id908" "id980" "id994" "id1193" "id1423" "id1505" "id1737" "id1827"
[17] "id1841" "id2068" "id2094" "id2151" "id2317" "id2618" "id2842" "id2894"
[25] "id2985" "id3354" "id3368" "id3641" "id3831" "id3983" "id4097" "id4328"
[33] "id4380" "id4395" "id4512" "id4552" "id4710" "id4717" "id4883" "id4904"
118 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

0.15
●

●
●●
●

0.10
●
● ●
● ●● ●
● ● ●
● ● ●● ●
● ●● ● ●
0.05
● ● ●
● ●● ●
● ●
● ● ●● ●● ● ●●
● ● ●
[Link][,2]

● ● ●●
● ●
●
●● ● ●
● ● ●● ●
0.00

● ● ●●●●
● ● ●●
● ●● ●
● ●
● ●● ●
●
● ●●● ● ●
−0.05

●● ● ● ●
● ● ●
● ●
● ●
● ● ●
● ● ● ●
● ● ●
●
● ● ●
●● ● ●
−0.10

● ● ●
●
●
●
−0.15

−0.3 −0.2 −0.1 0.0 0.1

[Link][,1]

Figure 5.3: Mapping samples to the space of the first two Principle Components
resulting from analysis of genomic kinship. Red dots identify genetic outliers.

[41] "id4934" "id4961" "id5014" "id5078" "id5274" "id5275" "id5454" "id5853"

[49] "id5926" "id5969" "id6237" "id6278" "id6352" "id6501" "id6554" "id6663"
[57] "id6723" "id7499" "id7514" "id7541" "id7598" "id7623" "id7949" "id8059"
[65] "id8128" "id8281" "id8370" "id8400" "id8433" "id8772" "id8880" "id8890"
[73] "id8957" "id8996" "id9082" "id9901" "id9930" "id1857" "id2528" "id4862"
[81] "id9184" "id5677" "id6407" "id5472" "id2135" "id8545" "id4333" "id1670"
[89] "id1536" "id6917" "id6424" "id3917" "id9628" "id9635" "id4729" "id5190"
[97] "id6399" "id6062" "id620" "id1116" "id6486" "id41" "id677" "id4947"
[105] "id9749" "id6428" "id7488" "id5949" "id2924" "id5783" "id4096" "id903"
[113] "id9049" "id185" "id1002" "id362" "id9014" "id5044" "id2749" "id2286"
[121] "id4743" "id4185" "id8330" "id6934"

Four outliers are shown in the smaller cluster.

Now you will need to use the BIGGER cluster for to select study subjects.
Whether this will be cl1 or cl2 in you case, is totally random.

We can form a data set which is free from outliers by using only people from
the bigger cluster:
> data2 <- data1[cl2, ]
5.3. FINDING GENETIC SUB-STRUCTURE 119

After we dropped the outliers, we need to repeat QC using [Link]().

At this stage, we want to allow for HWE checks (we will use only controls and
exclude markers with FDR ≤ 0.2):
> qc2 <- [Link](data2, hweids=(phdata(data2)$dm2==0), fdr=0.2)

Excluding people/markers with extremely low call rate...

3573 markers and 124 people in total
0 people excluded because of call rate < 0.1
0 markers excluded because of call rate < 0.1
Passed: 3573 markers and 124 people

Running sex chromosome checks...

0 heterozygous X-linked male genotypes found
0 X-linked markers are likely to be autosomal (odds > 1000 )
0 male are likely to be female (odds > 1000 )
0 female are likely to be male (odds > 1000 )
0 people have intermediate X-chromosome inbreeding (0.5 > F > 0.5)
If these people/markers are removed, 0 heterozygous male genotypes are left
Passed: 3573 markers and 124 people

no X/Y/mtDNA-errors to fix

RUN 1
3573 markers and 124 people in total
40 (1.119507%) markers excluded as having low (<2.016129%) minor allele frequency
0 (0%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (FDR <0.2)
0 (0%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2780246 (s.e. 0.01642372)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7704698 (s.e. 0.01247442), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 3533 (98.88049%) markers passed all criteria
In total, 124 (100%) people passed all criteria

RUN 2
3533 markers and 124 people in total
0 (0%) markers excluded as having low (<2.016129%) minor allele frequency
0 (0%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (FDR <0.2)
0 (0%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2780246 (s.e. 0.01642372)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.771093 (s.e. 0.01246105), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 3533 (100%) markers passed all criteria
In total, 124 (100%) people passed all criteria

> summary(qc2)
120 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

$`Per-SNP fails statistics`

NoCall NoMAF NoHWE Redundant Xsnpfail
NoCall 0 0 0 0 0
NoMAF NA 40 0 0 0
NoHWE NA NA 0 0 0
Redundant NA NA NA 0 0
Xsnpfail NA NA NA NA 0

$`Per-person fails statistics`

IDnoCall HetFail IBSFail isfemale ismale isXXY otherSexErr
IDnoCall 0 0 0 0 0 0 0
HetFail NA 0 0 0 0 0 0
IBSFail NA NA 0 0 0 0 0
isfemale NA NA NA 0 0 0 0
ismale NA NA NA NA 0 0 0
isXXY NA NA NA NA NA 0 0
otherSexErr NA NA NA NA NA NA 0

If the procedure did not run, check the previous Note.

Indeed, in the updated data set several markers do not pass our QC criteria
and we need to drop a few markers. This is done by
> data2 <- data2[qc2$idok, qc2$snpok]
This is going to be our final analysis data set, therefore let us attach the phe-
notypic data to the search path, so we do not need to type phdata(data2)$...
to access dm2 status or other variables:
> detach(phdata(data1))
> attach(phdata(data2))
> ####
> #[Link] <- data2
> #save([Link], file="[Link]")
> ####
Before proceeding to GWA, let us check if complete QC improved the fit of
genetic data to HWE:
> [Link](data2)[2]
$`Cumulative distr. of number of SNPs out of HWE, at different alpha`
X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 1 2.000 2e+01 101.000 3533
Prop 0 0.001 6e-03 0.029 1
> [Link](data2[phdata(data2)$dm2==1, ])[2]
$`Cumulative distr. of number of SNPs out of HWE, at different alpha`
X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 0 1 17.000 79.000 3533
Prop 0 0 0.005 0.022 1
5.4. GWA ASSOCIATION ANALYSIS 121

> [Link](data2[phdata(data2)$dm2==0, ])[2]

$`Cumulative distr. of number of SNPs out of HWE, at different alpha`

X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 0 0 7.000 91.000 3533
Prop 0 0 0.002 0.026 1

You can see that now there is no excessive number of SNPs out of HWE in the
sample (total, or cases, or controls)

5.4 GWA association analysis

Let us start again with descriptives of the phenotypic and marker data
> [Link](data2, by=dm2)

No([Link]=0) Mean SD No([Link]=1) Mean SD Ptt Pkw

id 47 NA NA 77 NA NA NA NA
sex 47 0.426 0.500 77 0.597 0.494 0.065 0.064
age 47 45.752 13.313 77 50.593 12.465 0.047 0.062
dm2 47 NA NA 77 NA NA NA NA
height 46 167.911 8.689 77 170.423 10.646 0.157 0.213
weight 46 77.015 17.528 77 94.160 26.963 0.000 0.000
diet 47 0.064 0.247 77 0.065 0.248 0.981 0.981
bmi 46 27.424 6.598 77 32.235 8.335 0.001 0.001
Pexact
id NA
sex 0.067
age NA
dm2 NA
height NA
weight NA
diet 1.000
bmi NA

You can see that the relation to weight is maintained in this smaller, but
hopefully cleaner, data set; moreover, the relation to age becomes borderline
significant.
If you check descriptives of markers (only HWE part shown)
> [Link](data2)[2]

$`Cumulative distr. of number of SNPs out of HWE, at different alpha`

X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 1 2.000 2e+01 101.000 3533
Prop 0 0.001 6e-03 0.029 1

you can see that the problems with HWE are apparently fixed; we may guess
that these were caused by Wahlund’s effect.
Run the score test on the cleaned data by
> [Link] <- qtscore(dm2, data2, trait="binomial")
122 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

qtscore(dm2, data2, binomial)

3
●
●
●
● ●
● ●
● ●
● ●
● ● ●
● ●
− log10(P − value)

● ● ● ●
● ● ●
●● ●●
●
● ● ●● ●
2

● ● ●
● ● ● ● ●
● ● ●● ●
● ●
● ● ● ● ●
●
● ●
● ● ● ●● ● ●●
● ●●
● ● ●● ●
●
●● ● ●●● ● ● ●● ● ● ● ●●
● ● ● ● ● ● ● ● ●
● ●● ●●●● ● ● ● ● ●● ● ●●●
● ●● ●● ● ● ●●
●● ●●●
● ●● ● ● ●●● ● ● ●
● ● ● ●
●
●
● ● ●● ● ● ●● ● ● ●● ●● ● ●● ●
● ●●
● ●
● ● ● ● ●●● ● ● ●● ●
●●● ● ●●●
●●● ● ●● ● ●
●●●●● ●●
● ●●●
● ● ●●● ● ● ● ● ● ●●●● ● ● ●●
● ●● ● ● ● ● ● ●●●● ●● ● ●●
● ●●● ● ● ●● ●
●
●● ●
● ●●●
●●●●● ●●● ●●●● ●● ●
●● ● ●●●
● ●● ● ●
●●● ● ●●
● ●● ●●● ●
●●●
● ● ● ● ●●● ●●●
● ●●● ●● ● ●●●● ●●
● ● ●● ●
●
●● ● ●●● ● ● ●● ●
●●
●
●●●●● ● ● ●●● ● ● ●●
●●
●● ● ● ●●●
1

● ●● ●● ●●● ● ● ●● ●
●
●● ● ● ● ● ●●●● ● ●●●●● ● ●● ●●● ●
●●
● ●
●●
●●●●● ● ● ● ● ● ●●●
● ●
● ●
●●
●
● ●● ●● ● ● ● ●●● ●● ●●●●●● ●● ●●
●
●● ●●
●● ●●●● ●● ●
● ●●
● ●●●●
●●●●●●
●●● ● ● ●
● ●●●● ● ●● ●●● ●●
●●●●●●●
●●●●
● ● ●●●●● ●● ●●●●
●●
●●●● ●●● ●● ● ●●
●●●●●●●
● ●
●● ● ●
●●●● ● ● ●●●●● ●● ●● ●●● ● ●● ● ● ●●● ●
●
●● ●●●● ●● ●
●●●
●●● ● ●● ●● ●● ●● ● ● ● ●● ●●● ●● ● ●● ● ● ● ●
●
●●
●● ●●
● ● ●●●●●
● ●● ● ●●●● ●●
● ●
●●● ●
● ●
●●
●● ●●●●● ●
●
● ●●●● ●● ●● ●
●●●●● ●● ●●● ●
● ●● ●
● ●● ●
●
●●●● ●
●●
●
● ●●●● ●
●●
●●
●●
●
●
●
●
●
●●●● ●●
● ●●●
●●
●●
●
● ●
●●●●●
●
● ●
● ●
●
●
●
●
●
●
●●
●●●
●●●●
●●
●●●●
●●●
●
●●●
● ● ●●●● ●●●● ●
●
●●
●
●● ●●●
●
●
●
●●●
●●●●
●
●
● ●
●●●●
●
●●
●
●
●●
●●●
●●●●●
●
●●
●
●●●
●
●
●
●
●
●●
● ●
● ●●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●● ●●
● ●
●
●
● ●
●
●
●
●
● ●●
●●●●●
●
● ●
●●
●
●●
●
● ●●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●●●●●
●● ●
●●●
●
●
●●●●
●
●●
●
● ●●
●
●
●●
●●
●
●
●●●
●
●
●●
● ●●●
●
●
●●
●
●
●
●
●●●
●
●●●
●
●●
●
● ●● ●
●
●
●●
●●
●
●●●●●●
●
●● ●●●● ●● ●
● ●●
●
●● ●●●
●●●
● ●●● ●
●●●
●
●●●●●●● ●● ●
●
●● ● ●●●●●● ● ●
●●●
●●●● ●
●●
● ●●●
● ●● ●● ●
● ●
●●●●
●●
● ●● ● ● ●●
●
● ● ● ●●●●●● ● ●●●
●●● ●
● ● ●●
●
● ●● ● ● ●●● ● ●● ●●
●●● ●● ● ●●● ●●●● ●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●●
●
● ●●
●
●
● ●
●●●
●
●
●●
●●●
●
●
●●
●
●
●●
●
●
●●●●
●
●●
●
●
●
●●
●●●
●●●●
●●
●
●●
●
●
●
●●●
●
●
●
●
●●●● ●
●
●
●
●
●
●
●
●●
●
●●●●
●●●
●
●
●
●● ●
●
●●●
●●●
● ●
●●●
●●●●
●
●●●●
●
●
●
●
●●●
●
●●
●
●
●
●●●●
● ● ●●●
● ●●● ●●● ●
●
●●
●
●●
●●
●
●
●
●
●
●●
●● ●
●●●
●●
●
●
●●
●
●
●
●●
●
●
●
●●●
●
●●●
●●
● ●●
●● ●●
● ●
●
●● ●
●
●
●●
●●● ●
●
●
●●●
●●●
●
● ●●
●●●●
●●●
●
●●●
●●●
●●
●●●●
● ● ●
● ●●●
● ●
●
●●
●
●●
●●
● ● ●
●● ●●
● ●●●●
●●●
●●
●●●
● ●●
●
● ● ●
● ●●
●●●
● ●
●●
●●
●
●● ● ●●
●●●●
●
● ●
●●
●
● ●●●●
●
● ●●●● ●
●● ●
●●●
● ●
●●●
● ●●
●●●● ● ●
●
●●●●
● ●●
●●●●
●
● ● ●
●●
●●● ●
●●●●
●
●
●●●●●●●
●●●
●
●●
●●
●●●●
●●●●
●
●
●● ●●
●●
●●
●●
● ●
●
● ●●●
●●●●●●
●●●●●
● ● ●
●●●
●●●
●
●● ● ● ●
●●●● ● ● ● ●●● ● ●
● ●● ● ● ●● ● ●
●● ● ●● ● ● ●● ●
●●● ●●● ● ● ●●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
● ●
●●
●
●
●
●●
●
●
●●
●
●
●●●
● ●●
●
●●
● ●
●●
●●
●
●
●●
●●
● ●
●●
●
●
●●●
●
●
●●
●●
●
●
●
●●
●
●
●●
●●
●
●●●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●●
●●●
●
●
●
●●
●
●
●●
●
●●
●●●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●●●●
●●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●●
●●
●
●
●●●
●● ●
●●
●● ●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●●●
●●
●
●
●
●
●
●●●
●
●●
●●
●
●●●
●
●●
●●
●
●●
●
●●●
●
●●●●
●●
●
●●
●●
●
●●
●●
● ●
●
●●
●●
●●
● ●●
●
●●
●
●●
● ●●● ●
●●
●●
●●
●
●●●
●
●●●
●●●
●
●●
●● ●●●●●
●
●
●●●
●
● ●
●●●
●
●●
●
●●●●●
●
●
●●
●●
●●
●●
●
●●
●
●
●●●
●
●
●
●●
●●
●
●●
●
● ●
●
●●
●●
●
●
●
●●●
●
●●
●●●●
●●
●●
●
●●
●●●
●
●●●
●
●●
● ●
●●●
●● ●
●
●●
●●
●●
●
●
●●●
●●
●
●
●
●
●●
●
●
●●
●
●●
●●
●
●
●●
●
●
● ●
●
●●
●●●
●●●●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●●
●●
●
●●
●
●
●●●●●
●
●
●●
●
●●
●●●
●●
●
●
●●●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●●
●● ●
●
●
●●
●
●
● ●
●
●●
● ●●
●
●
●●
●●
●
●
●
●
●
●
●●●
●●
●●
●●
●●●
●●●●
●
●
●
●●
●
●
●
●●
● ●
●●●
●
●●●
●
●●
●
●
●
●
●● ●●
●
●●
●●●
●●●●●
●
●
●
●●
●
●
●●
●
●
●
●●
●●
●
●●●
●●
●
●●●
●●
●
●
●
●●
● ●
●
●●
●●
●
●
●●
●●
●●
●●
●
●
●
●●
● ●
●●●
●
●●
●●
●●
●●
●●
●
●●●
●●●● ●
●●●
●
●● ●●
●●
●●●
●●●●●●
●
●
●● ●●
●●●●
●
●●
●
●
●●●● ●● ●●● ●●
●●●●●●●●
●●●
● ●●
●● ●
●● ●
●
●●
●●
●●●
●●
●●●
●●
●
● ●
●●
●● ●
●●●●●●
●●
●
●●●●
●●●● ●● ●
●●●●
●●
●●
● ●
0

1 2 3 X

Chromosome

Figure 5.4: − log10 (Corrected P -value) from the genome scan after the QC
procedure.

and check lambda

> lambda([Link])

$estimate
[1] 1.036958

$se
[1] 0.0007178

there is still some inflation, which is explained by the fact that we investigate
only a few short chromosomes with high LD and few causative variants.
Produce the association analysis plot by

> plot([Link], df="Pc1df")

(Figure 5.4).
Produce the scan summary by

> [Link]([Link], sort="Pc1df")

Summary for top 10 results, sorted by Pc1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs1719133 1 4495479 + T A 124 0.3167801 0.08614528 13.522368
5.4. GWA ASSOCIATION ANALYSIS 123

rs4804634 1 2807417 + C G 121 0.4119844 0.12480696 10.896423

rs8835506 2 6010852 + A T 121 3.5378209 1.08954331 10.543448
rs4534929 1 4474374 + C G 123 0.4547151 0.14160410 10.311626
rs1013473 1 4487262 + A T 124 2.7839368 0.86860745 10.272393
rs3925525 2 6008501 + C G 124 3.2807631 1.03380675 10.070964
rs3224311 2 6009769 + G C 124 3.2807631 1.03380675 10.070964
rs2975760 3 10518480 + A T 123 3.1802120 1.00916993 9.930784
rs2521089 3 10487652 - T C 123 2.7298775 0.87761175 9.675679
rs1048031 1 4485591 + G T 122 0.4510793 0.14548378 9.613391
P1df effAB effBB chi2.2df P2df Pc1df
rs1719133 0.0002357368 0.3740771 0.0000000 14.677906 0.0006497303 0.0003048399
rs4804634 0.0009635013 0.6315789 0.1739130 12.375590 0.0020543516 0.0011885463
rs8835506 0.0011660066 4.0185185 4.0185185 12.605556 0.0018312105 0.0014292471
rs4534929 0.0013219476 0.4830918 0.1739130 10.510272 0.0052206352 0.0016136479
rs1013473 0.0013503553 3.0495868 5.8441558 10.926296 0.0042401869 0.0016471605
rs3925525 0.0015062424 3.6923077 4.0000000 11.765985 0.0027864347 0.0018306610
rs3224311 0.0015062424 3.6923077 4.0000000 11.765985 0.0027864347 0.0018306610
rs2975760 0.0016253728 3.0000000 8.0000000 10.172522 0.0061810866 0.0019704699
rs2521089 0.0018672326 3.0147059 5.0000000 10.543296 0.0051351403 0.0022533033
rs1048031 0.0019316360 0.4844720 0.1714286 9.965696 0.0068545128 0.0023284084

Comparison with the top 10 from the scan before QC shows that results
changed substantially with only few markers overlapping.
You can see similar results when assessing empirical GW significance:
> [Link] <- qtscore(dm2,
+ data2, times=200, quiet=TRUE, trait="binomial")

|
| | 0%
|
|======================================================================| 100%

> [Link]([Link], sort="Pc1df")

Summary for top 10 results, sorted by Pc1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs1719133 1 4495479 + T A 124 0.3167801 0.08614528 13.522368
rs4804634 1 2807417 + C G 121 0.4119844 0.12480696 10.896423
rs8835506 2 6010852 + A T 121 3.5378209 1.08954331 10.543448
rs4534929 1 4474374 + C G 123 0.4547151 0.14160410 10.311626
rs1013473 1 4487262 + A T 124 2.7839368 0.86860745 10.272393
rs3925525 2 6008501 + C G 124 3.2807631 1.03380675 10.070964
rs3224311 2 6009769 + G C 124 3.2807631 1.03380675 10.070964
rs2975760 3 10518480 + A T 123 3.1802120 1.00916993 9.930784
rs1048031 1 4485591 + G T 122 0.4510793 0.14548378 9.613391
rs8258863 1 4735725 + T A 123 0.1895736 0.06222538 9.281554
P1df Pc1df effAB effBB chi2.2df P2df
rs1719133 0.330 0.435 0.3740771 0.0000000 14.677906 0.555
rs4804634 0.840 0.880 0.6315789 0.1739130 12.375590 0.950
rs8835506 0.870 0.925 4.0185185 4.0185185 12.605556 0.940
124 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

rs4534929 0.920 0.945 0.4830918 0.1739130 10.510272 1.000

rs1013473 0.920 0.945 3.0495868 5.8441558 10.926296 1.000
rs3925525 0.935 0.965 3.6923077 4.0000000 11.765985 0.975
rs3224311 0.935 0.965 3.6923077 4.0000000 11.765985 0.975
rs2975760 0.945 0.965 3.0000000 8.0000000 10.172522 1.000
rs1048031 0.965 0.975 0.4844720 0.1714286 9.965696 1.000
rs8258863 0.975 0.975 0.1967593 NA 9.281554 1.000

Again, none of the SNPs hits GW 5% significance. Still, you can see that
after QC the top markers achieve somewhat “better” significance.
In the last part, we will do several adjusted and stratified analyses. Only
empirical P -values will be estimated to make the story shorter. To adjust for
sex and age, we can
> [Link] <- qtscore(dm2~sex+age,
+ data2, times=200, quiet=TRUE, trait="binomial")

|
| | 0%
|
|======================================================================| 100%

> [Link]([Link])

Summary for top 10 results, sorted by P1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs1719133 1 4495479 + T A 124 0.5057834 0.1386464 13.307976
rs2398949 1 4828375 - A C 122 0.4163599 0.1253217 11.037861
rs4804634 1 2807417 + C G 121 0.5811444 0.1769232 10.789433
rs7522488 3 11689797 - G A 123 1.8718644 0.5707656 10.755565
rs3925525 2 6008501 + C G 124 2.0202686 0.6406356 9.944799
rs3224311 2 6009769 + G C 124 2.0202686 0.6406356 9.944799
rs8835506 2 6010852 + A T 121 2.0575431 0.6501899 10.014228
rs1037237 3 11690145 + C G 124 1.8144282 0.5735660 10.007206
rs4534929 1 4474374 + C G 123 0.6262025 0.2045254 9.374220
rs1048031 1 4485591 + G T 122 0.6128584 0.1993057 9.455418
P1df Pc1df effAB effBB chi2.2df P2df
rs1719133 0.400 0.545 0.4853657 0.2242383 13.367455 0.860
rs2398949 0.830 0.925 0.3110042 2.8197228 16.111010 0.390
rs4804634 0.865 0.945 0.7420448 0.3389525 12.112794 0.960
rs7522488 0.865 0.945 1.3995557 3.1160361 11.256933 0.995
rs3925525 0.960 0.980 2.3028971 2.2451502 11.988964 0.960
rs3224311 0.960 0.980 2.3028971 2.2451502 11.988964 0.960
rs8835506 0.960 0.980 2.3758119 2.2365358 12.302956 0.935
rs1037237 0.960 0.980 1.3400938 2.9836475 10.674189 1.000
rs4534929 0.980 0.995 0.6022249 0.3747712 9.381086 1.000
rs1048031 0.980 0.995 0.6453091 0.3417444 9.840158 1.000

You can see that there is little difference between adjusted and unadjusted
analysis, but this is not always the case; adjustment may make your study much
more powerful when covariates explain a large proportion of environmental trait
variation.
5.4. GWA ASSOCIATION ANALYSIS 125

Finally, let us do stratified (by BMI) analysis. We will contracts obese

(BMI ≥ 30) cases to all controls.
> [Link] <- qtscore(dm2~sex+age,
+ data2, ids=((bmi>30 & dm2==1) | dm2==0),
+ times=200, quiet=TRUE, trait="binomial")

|
| | 0%
|
|======================================================================| 100%

> [Link]([Link], sort="Pc1df")

Summary for top 10 results, sorted by Pc1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs7522488 3 11689797 - G A 88 1.7561171 0.5560485 9.9742822
rs1037237 3 11690145 + C G 88 1.7561171 0.5560485 9.9742822
rs9630764 1 3897972 + T A 88 1.7595598 0.5615384 9.8185891
rs1891586 1 2297398 - C T 88 0.5340528 0.1733886 9.4869655
rs3215698 X 13559835 - T A 88 0.4806715 0.1571512 9.3553776
rs794264 1 2534738 - C A 87 1.7252112 0.5832358 8.7497608
rs2884479 X 13618173 + A C 86 0.5896309 0.1990222 8.7772281
rs1646456 1 653 + C G 87 0.8853536 1.3981894 0.4009605
rs4435802 1 5291 + C A 86 1.7961795 1.1926370 2.2682082
rs946364 1 8533 - T C 86 0.7682453 0.5994013 1.6427231
P1df Pc1df effAB effBB chi2.2df P2df
rs7522488 0.930 0.945 1.3215457 3.1245402 11.299048 1.00
rs1037237 0.930 0.945 1.3215457 3.1245402 11.299048 1.00
rs9630764 0.945 0.950 1.9346110 3.3161548 9.831616 1.00
rs1891586 0.970 0.970 0.5836292 0.3065275 9.518213 1.00
rs3215698 0.970 0.980 0.3406693 0.2701336 9.941713 1.00
rs794264 0.985 0.995 2.2543966 3.5369544 8.965654 1.00
rs2884479 0.985 0.995 0.2872964 0.4639136 14.266996 0.55
rs1646456 1.000 1.000 1.1455730 0.6013127 2.201546 1.00
rs4435802 1.000 1.000 1.7919867 NA 2.268208 1.00
rs946364 1.000 1.000 0.8791195 0.4696496 2.188118 1.00

Again, nothing interesting at the GW significance level. If we would have had

found something, naturally, we would not have known if we mapped a T2D or
obesity gene (or a gene for obesity in presence of T2D, or the one for T2D in
presence of obesity).
Let us save the ’data1’, ’[Link]’, ’[Link]’ and ’cl1’ objects now
> save(data1, cl1, [Link], [Link], file="[Link]")
These data will be used later in section 7, ”GWA in presence of genetic strat-
ification: practice”, in which we will perform GWA using different methods to
account for stratification, and will compare the results with these obtaining by
removing outliers (’[Link]’).
Let us also save the cleaned ’data2’ object, which will be later used in section
10 (”Meta-analysis of GWA scans”):
126 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

> save(data2, file="[Link]")

At this point, you acquired the knowledge necessary for the self-exercise.
Please close R by q() command and proceed to the next section.

5.5 Genome-wide association analysis exercise

During the exercise, you will work with a larger data set (approximately 1,000
people and 7,000+ SNPs). You are to do the complete three-round QC; perform
GWA analysis with dm2 as the outcome of interest and identify 10 SNPs which
you would like to take to the stage 2 (replication) scan. You will do replication
analysis using a confirmatory data set. If you did everything right, the SNPs
which you identified as significant or replicated will be located in known T2D
genes.
Please keep in mind that the data are simulated, and do not take your
findings too seriously!
Start R by going to ”Start -¿ Programs -¿ R -¿ R-?.?.?”. Load the GenABEL-package
library by
> library(GenABEL)
The two data sets we will use in this exercise are part of the GenABEL-package
distribution. The first one (the ”discovery” set) can be loaded by
> data(ge03d2)
Please move along the lines detailed in the guided exercise and try to answer
following questions:
Ex. 1 — How many cases and controls are present in the original data set?

Ex. 2 — How many markers are present in the original data set?

Ex. 3 — Is there evidence for inflation of the HWE test statistics?

Ex. 4 — Analyse empirical GW significance. How many SNPs pass the genome-
wide significance threshold, after correction for the inflation factor? Write down
the names of these SNPs for further comparison.

Ex. 5 — Perform first steps of the genetic data QC.

Ex. 6 — How many males are ’genetically’ females?

Ex. 7 — How many females are ’genetically’ males?

Ex. 8 — How many people are quessed to have ’XXY’ genotype?

Ex. 9 — How many sporadic X errors do you still observe even when the
female male and non-X X-markers are removed? (do not forget to Xfix(s)
these
Ex. 10 — How many ”twin” DNAs did you discover?

Ex. 11 — Perform second step of QC.

Ex. 12 — How many genetic outliers did you discover?

5.6. ANSWERS TO EXERCISES 127

Ex. 13 — How many cases and controls are presented in the data after QC?

Ex. 14 — How many markers are presented in the data after QC?

Ex. 15 — Is there evidence for inflation of the HWE test staistics?

Ex. 16 — Perform GWA analysis of the cleaned data, using asimptotic test
and plot the results. What is the estimate of λ for the 1 d.f. test?

Ex. 17 — Analyse empirical GW significance. How many SNPs pass genome-

wide significance threshold, after correction for the inflation factor?

Ex. 18 — Do these SNPs overlap much with the ones ranked at the top before
the QC? If not, what could be the reason?

Ex. 19 — Select 10 SNPs which you would like to follow-up. Say, you’ve se-
lected rs1646456, rs7950586, rs4785242, rs4435802, rs2847446, rs946364, rs299251,
rs2456488, rs1292700, and rs8183220. Make a vector of these SNPs with
> #vec12<-c("rs1646456", "rs7950586", "rs4785242", "rs4435802", "rs2847446",
> # "rs946364", "rs299251", "rs2456488", "rs1292700", "rs8183220")
Load the stage 2 (replicaton) data set by
> #data(ge03d2c)
and select the subset of SNPs you need by
> #confdat <- ge03d2c[, vec12]
Analyse the confdat for association with dm2.

Ex. 20 — Given the two-stage design, and applying the puristic criteria spec-
ified in the lecture, for how many SNPs you can claim a significant finding?

Ex. 21 — Using the same criteria, for how many SNPs you can claim a repli-
cated finding?

Ex. 22 — If time permits, characterise the mode of inheritance of the signifi-

cant SNPs. You can convert data from GenABEL-package format to the format
used by [Link] and genetics libraries by using [Link]() func-
tion. Consult the help for details. Please do not attempt to convert more than
a few dozen SNPs: the format of genetics is not compressed, which means
conversion may take long and your low-memory computer may even crash if
you attempt to convert the whole data set.

Ex. 23 — If time permits, do analysis with adjustment for covariates and

stratified analysis.

Ex. 24 — If time permits, try to do first round of QC allowing for HWE

checks (assume FDR of 0.1 for total sample). In this case, can you still detect
stratification in the ”cleaned” data?

5.6 Answers to exercises

128 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

Answer (Ex. 1) — :
> table(phdata(ge03d2)$dm2)
0 1
487 463

Answer (Ex. 2) — :
> nsnps(ge03d2)
[1] 7589

Answer (Ex. 3) — Yes:

> [Link](ge03d2)[2]
$`Cumulative distr. of number of SNPs out of HWE, at different alpha`
X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 331.000 367.000 479.000 807.000 7589
Prop 0.044 0.048 0.063 0.106 1

Answer (Ex. 4) — :
> res0 <- qtscore(dm2, data=ge03d2, times=200,
+ quiet=TRUE, trait="binomial")
|
| | 0%
|
|======================================================================| 100%
> ### something funny is going on here
> ### disabeling next line
> #lambda(res0)
> ds <- [Link](res0)
Summary for top 10 results, sorted by P1df
> ds
Chromosome Position Strand A1 A2 N effB se_effB chi2.1df P1df Pc1df
rs1646456 1 653 + C G 938 NaN NaN NaN NA NA
rs7950586 1 849 - T A 938 NaN NaN NaN NA NA
rs4785242 1 1766 - T C 936 NaN NaN NaN NA NA
rs4435802 1 5291 + C A 943 NaN NaN NaN NA NA
rs2847446 1 5555 + T A 937 NaN NaN NaN NA NA
rs9308393 1 6739 + T C 937 NaN NaN NaN NA NA
rs946364 1 8533 - T C 938 NaN NaN NaN NA NA
rs299251 1 10737 + A G 942 NaN NaN NaN NA NA
rs2456488 1 11779 + G C 934 NaN NaN NaN NA NA
rs1292700 1 12710 - A C 941 NaN NaN NaN NA NA
effAB effBB chi2.2df P2df
rs1646456 NaN NaN 0 NA
rs7950586 NaN NaN 0 NA
5.6. ANSWERS TO EXERCISES 129

rs4785242 NaN NaN 0 NA

rs4435802 NaN NaN 0 NA
rs2847446 NaN NaN 0 NA
rs9308393 NaN NA 0 NA
rs946364 NaN NaN 0 NA
rs299251 NaN NaN 0 NA
rs2456488 NaN NaN 0 NA
rs1292700 NaN NaN 0 NA
Thus, there are no genome-wide empirically significant results. The ’top’ 10
SNPs are
> snps0 <- rownames(ds)
> snps0
[1] "rs1646456" "rs7950586" "rs4785242" "rs4435802" "rs2847446" "rs9308393"
[7] "rs946364" "rs299251" "rs2456488" "rs1292700"
(note that if the empirical P = 1, the rank is assigned quite arbitrarily)

Answer (Ex. 5) — First step of QC

> qc1 <- [Link](ge03d2, call=0.95, [Link]=0.95,
+ [Link]=0, [Link]="both")
Excluding people/markers with extremely low call rate...
7589 markers and 950 people in total
0 people excluded because of call rate < 0.1
7 markers excluded because of call rate < 0.1
Passed: 7582 markers and 950 people

Running sex chromosome checks...

1934 heterozygous X-linked male genotypes found
2 X-linked markers are likely to be autosomal (odds > 1000 )
10 male are likely to be female (odds > 1000 )
6 female are likely to be male (odds > 1000 )
0 people have intermediate X-chromosome inbreeding (0.5 > F > 0.5)
If these people/markers are removed, 8 heterozygous male genotypes are left
... these will be considered missing in analysis.
... Use Xfix() to fix these problems.
Passed: 7580 markers and 934 people

... 8 X/Y/mtDNA ( 8 0 0 ) impossible heterozygotes and female Ys set as missing

RUN 1
7580 markers and 934 people in total
73 (0.9630607%) markers excluded as having low (<0.267666%) minor allele frequency
75 (0.9894459%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (P <0)
4 (0.4282655%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2558271 (s.e. 0.02102863)
4 (0.4282655%) people excluded because too high autosomal heterozygosity (FDR <1%)
130 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

Excluded people had HET >= 0.4702949

Mean IBS is 0.7885754 (s.e. 0.01734777), as based on 2000 autosomal markers
8 (0.856531%) people excluded because of too high IBS (>=0.95)
In total, 7432 (98.04749%) markers passed all criteria
In total, 918 (98.28694%) people passed all criteria

RUN 2
7432 markers and 918 people in total
42 (0.5651238%) markers excluded as having low (<0.2723312%) minor allele frequency
0 (0%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (P <0)
0 (0%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2562716 (s.e. 0.0151377)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7872975 (s.e. 0.01587625), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 7390 (99.43488%) markers passed all criteria
In total, 918 (100%) people passed all criteria

RUN 3
7390 markers and 918 people in total
0 (0%) markers excluded as having low (<0.2723312%) minor allele frequency
0 (0%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (P <0)
0 (0%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2562716 (s.e. 0.0151377)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7844231 (s.e. 0.01656417), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 7390 (100%) markers passed all criteria
In total, 918 (100%) people passed all criteria
> summary(qc1)
$`Per-SNP fails statistics`
NoCall NoMAF NoHWE Redundant Xsnpfail
NoCall 82 0 0 0 0
NoMAF NA 115 0 0 0
NoHWE NA NA 0 0 0
Redundant NA NA NA 0 0
Xsnpfail NA NA NA NA 2

$`Per-person fails statistics`

IDnoCall HetFail IBSFail isfemale ismale isXXY otherSexErr
IDnoCall 4 0 0 0 0 0 0
HetFail NA 4 0 0 0 0 0
IBSFail NA NA 8 0 0 0 0
isfemale NA NA NA 10 0 0 0
ismale NA NA NA NA 6 0 0
isXXY NA NA NA NA NA 0 0
otherSexErr NA NA NA NA NA NA 0
5.6. ANSWERS TO EXERCISES 131

> data1 <- ge03d2[qc1$idok, qc1$snpok]

> data1 <- Xfix(data1)
... 7 X/Y/mtDNA ( 7 0 0 ) impossible heterozygotes and female Ys set as missing
> qc2 <- [Link](data1, call=0.95, [Link]=0.95, [Link]=0)
Excluding people/markers with extremely low call rate...
7390 markers and 918 people in total
0 people excluded because of call rate < 0.1
0 markers excluded because of call rate < 0.1
Passed: 7390 markers and 918 people

Running sex chromosome checks...

0 heterozygous X-linked male genotypes found
0 X-linked markers are likely to be autosomal (odds > 1000 )
0 male are likely to be female (odds > 1000 )
0 female are likely to be male (odds > 1000 )
0 people have intermediate X-chromosome inbreeding (0.5 > F > 0.5)
If these people/markers are removed, 0 heterozygous male genotypes are left
Passed: 7390 markers and 918 people

no X/Y/mtDNA-errors to fix

RUN 1
7390 markers and 918 people in total
0 (0%) markers excluded as having low (<0.2723312%) minor allele frequency
0 (0%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (P <0)
0 (0%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2562716 (s.e. 0.0151377)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7873545 (s.e. 0.01636484), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 7390 (100%) markers passed all criteria
In total, 918 (100%) people passed all criteria
> summary(qc2)
$`Per-SNP fails statistics`
NoCall NoMAF NoHWE Redundant Xsnpfail
NoCall 0 0 0 0 0
NoMAF NA 0 0 0 0
NoHWE NA NA 0 0 0
Redundant NA NA NA 0 0
Xsnpfail NA NA NA NA 0

$`Per-person fails statistics`

IDnoCall HetFail IBSFail isfemale ismale isXXY otherSexErr
IDnoCall 0 0 0 0 0 0 0
HetFail NA 0 0 0 0 0 0
IBSFail NA NA 0 0 0 0 0
132 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

isfemale NA NA NA 0 0 0 0
ismale NA NA NA NA 0 0 0
isXXY NA NA NA NA NA 0 0
otherSexErr NA NA NA NA NA NA 0

Answer (Ex. 6) — The list of genetic females who are coded as males is
> qc1$isfemale
[1] "id3374" "id6263" "id6835" "id8410" "id8509" "id8519" "id8542" "id2701"
[9] "id6494" "id3100"

Answer (Ex. 7) — The list of genetic males who are coded as females is
> qc1$ismale
[1] "id193" "id8475" "id2461" "id5669" "id7245" "id8301"

Answer (Ex. 8) — The number of ’XXY’ people is 0

Answer (Ex. 9) — Eight ’sporadic’ X-errors are left after removing people
with likely sex code errors (seven in the data set after first step of QC)

Answer (Ex. 10) — The list of IDs failing IBS checks (’twin’ DNAs) is
> qc1$ibsfail
[1] "id3368" "id9668" "id5437" "id956" "id386" "id660" "id2115" "id8370"

Answer (Ex. 11) — The second step of QC:

> [Link] <- ibs(data1[, autosomal(data1)], weight="freq")
> [Link] <- [Link](0.5 - [Link])
> [Link] <- cmdscale([Link])
> km <- kmeans([Link], centers=2, nstart=1000)
> cl1 <- names( which(km$cluster==1) )
> cl2 <- names( which(km$cluster==2) )
> if (length(cl1) > length(cl2)) {x<-cl2; cl2<-cl1; cl1<-x}
> cl1
[1] "id2097" "id2126" "id2878" "id3021" "id3176" "id4554" "id4756" "id7436"
[9] "id7533" "id9396" "id9546" "id4021" "id2171" "id6954" "id2136" "id5056"
[17] "id1751" "id6626" "id2970" "id1300" "id8639" "id1729" "id9398" "id9904"
[25] "id858"
> data2 <- data1[cl2,]
> qc2 <- [Link](data2, hweids=(phdata(data2)$dm2==0), fdr=0.2)
Excluding people/markers with extremely low call rate...
7390 markers and 893 people in total
0 people excluded because of call rate < 0.1
5.6. ANSWERS TO EXERCISES 133

0 markers excluded because of call rate < 0.1

Passed: 7390 markers and 893 people

Running sex chromosome checks...

0 heterozygous X-linked male genotypes found
0 X-linked markers are likely to be autosomal (odds > 1000 )
0 male are likely to be female (odds > 1000 )
0 female are likely to be male (odds > 1000 )
0 people have intermediate X-chromosome inbreeding (0.5 > F > 0.5)
If these people/markers are removed, 0 heterozygous male genotypes are left
Passed: 7390 markers and 893 people

no X/Y/mtDNA-errors to fix

RUN 1
7390 markers and 893 people in total
5 (0.067659%) markers excluded as having low (<0.2799552%) minor allele frequency
0 (0%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (FDR <0.2)
0 (0%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2565083 (s.e. 0.01505982)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7882472 (s.e. 0.01153768), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 7385 (99.93234%) markers passed all criteria
In total, 893 (100%) people passed all criteria

RUN 2
7385 markers and 893 people in total
0 (0%) markers excluded as having low (<0.2799552%) minor allele frequency
0 (0%) markers excluded because of low (<95%) call rate
0 (0%) markers excluded because they are out of HWE (FDR <0.2)
0 (0%) people excluded because of low (<95%) call rate
Mean autosomal HET is 0.2565083 (s.e. 0.01505982)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7904758 (s.e. 0.01152922), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 7385 (100%) markers passed all criteria
In total, 893 (100%) people passed all criteria
> summary(qc2)
$`Per-SNP fails statistics`
NoCall NoMAF NoHWE Redundant Xsnpfail
NoCall 0 0 0 0 0
NoMAF NA 5 0 0 0
NoHWE NA NA 0 0 0
Redundant NA NA NA 0 0
Xsnpfail NA NA NA NA 0
134 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

$`Per-person fails statistics`

IDnoCall HetFail IBSFail isfemale ismale isXXY otherSexErr
IDnoCall 0 0 0 0 0 0 0
HetFail NA 0 0 0 0 0 0
IBSFail NA NA 0 0 0 0 0
isfemale NA NA NA 0 0 0 0
ismale NA NA NA NA 0 0 0
isXXY NA NA NA NA NA 0 0
otherSexErr NA NA NA NA NA NA 0
> data2 <- data2[qc2$idok, qc2$snpok]
> data2 <- Xfix(data2)
no X/Y/mtDNA-errors to fix
> ####
> #[Link] <- data2
> #save([Link], file="[Link]")
> ####

Answer (Ex. 12) — The list of genetic outliers is

> cl1
[1] "id2097" "id2126" "id2878" "id3021" "id3176" "id4554" "id4756" "id7436"
[9] "id7533" "id9396" "id9546" "id4021" "id2171" "id6954" "id2136" "id5056"
[17] "id1751" "id6626" "id2970" "id1300" "id8639" "id1729" "id9398" "id9904"
[25] "id858"

Answer (Ex. 13) — :

> table(phdata(data2)$dm2)
0 1
472 421

Answer (Ex. 14) — :

> nsnps(data2)
[1] 7385

Answer (Ex. 15) — No:

> [Link](data2)[2]
$`Cumulative distr. of number of SNPs out of HWE, at different alpha`
X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 1 4.000 44.000 258.000 7385
Prop 0 0.001 0.006 0.035 1
FROM THIS POINT ON THE RUNNING OF THE TUTORIAL FAILS (2013.02.20,
USING GENABEL DEV VERSION 1.7-4). I COMMENTED OUT THE R
CODE BELOW.
5.6. ANSWERS TO EXERCISES 135

Answer (Ex. 16) — :

> #qts <- qtscore(dm2, data2, trait="binomial")
> #lambda(qts)

Answer (Ex. 17) — :

> #res1 <- qtscore(dm2, data=data2, times=200, quiet=TRUE, trait="binomial")
> #ds1 <- [Link](res1)
> #ds1
There are SNPs which are empirically genome-wide significant in the data. To
get the list of ’top’ 10 SNPs:
> #snps1 <- rownames(ds1)
> #snps1

Answer (Ex. 18) — There is little overlap between SNPs before and after
QC:
> #snps0
> #snps1

Answer (Ex. 19) — :

> #data(ge03d2c)
> #snps1
> #confdat <- ge03d2c[, snps1]
> #rep <- qtscore(dm2, confdat, times=10000, quiet=TRUE)
> #[Link](rep)

Answer (Ex. 20) — Two-stage P -value is

> # snps1
> # finres <- matrix(NA, 10, 3)
> # colnames(finres) <- c("Stage 1", "Replication", "Combined")
> # rownames(finres) <- snps1
> # for (i in 1:10) {
> # finres[i, 1] <- res1[which(snpnames(data2)==snps1[i]), "Pc1df"]
> # finres[i, 2] <- rep[which(snpnames(confdat)==snps1[i]), "P1df"]
> # finres[i, 3] <- finres[i, 1]*finres[i, 2]
> # }
> # finres
> # for (i in 1:10) {
> # if (finres[i, 3] <= 0.05) {
> # print(c("---------", rownames(finres)[i], "-------"))
> #
> # print(c(rownames(finres)[i], "stage 1:"))
> # ph <- phdata(data2)$dm2
> # gt <- [Link](data2[, rownames(finres)[i]])
136 CHAPTER 5. GENOME-WIDE ASSOCIATION ANALYSIS

> # print(summary( glm(ph~gt, family=binomial) )$coef)

> #
> # print(c(rownames(finres)[i], "stage 2:"))
> # ph <- phdata(confdat)$dm2
> # gt <- [Link](confdat[, rownames(finres)[i]])
> # print(summary(glm(ph~gt, family=binomial))$coef)
> #
> # print(c(rownames(finres)[i], "Joint:"))
> # ph <- c(phdata(data2)$dm2, phdata(confdat)$dm2)
> # gt <- c([Link](data2[, rownames(finres)[i]]),
> # [Link](confdat[, rownames(finres)[i]]))
> # print(summary( glm(ph~gt, family=binomial) )$coef)
> # }
> # }
> #replicatedsnps <- rownames(finres)[finres[, "Stage 1"] <= 0.05 &
> # finres[, "Replication"] <= 0.05 &
> # finres[, "Combined"] <= 0.05]
> #replicatedsnps
> #sigsnps <- rownames(finres)[finres[, "Combined"] <= 0.05]
> #sigsnps
At the first glance, NWSexprlength(replicatedsnps) SNPs may be claimed as
replicated because both first stage and replication P -values are ≤ 0.05 and ef-
fects are consistent, and additionally NWSexprlength(sigsnps) - length(replicatedsnps)
may be claimed as ’significant’ because joint P -values are ≤ 0.05 and the effects
are consistent. Generally, a more thorough simulation experiment should be
performed.

Answer (Ex. 21) — SNP “rs7903146” had empirical P -value ≤ 0.05 at both
stages, and very strong joint significance. It can be claimed as replicated.
You can check if any of the SNPs you have identified as significant or repli-
cated are the ones which were simulated to be associated with dm2 by using
the command [Link](c("snpname1", "snpname2", "snpname3")) where
snpnameX stands for the name of your identified SNP. The ”true” SNPs can be
found on NCBI and some are located in known T2D genes (just because we
used these names to name the ”significant” ones).
Chapter 6

GWA analysis in presence

of stratification: theory

In genetic association studies, we look for association between a genetic poly-

morphism and the value of a trait of interest. The best scenario – the one we
always hope for – is that the observed association results from causation, that
is the polymorphism studied is functionally involved in the control of the trait.
However, association has no direction, and making causal inference in epidemi-
ology in general and in genetic epidemiology in particular is usually not possible
based on statistical analysis only.
In fact, most associations observed in genetic studies are due to a confounder
– an (unobserved) factor which is associated with both the genetic polymorphism
and the trait analysed. Presence of such factor leads to induced, “secondary”
correlation between the trait and the polymorphism; if we would have controlled
for that factor in the association model, the relation between the polymorphism
and the trait would have gone.
There are two major types of confounders leading to induced correlation in
genetic association studies. One type is “good” confounding of association by
the real, unobserved functional variant, which is, as a rule, not present on the
SNP array, but is in linkage disequilibrium (LD) with typed SNP. Under this
scenario, the functional variant is associated with the trait because of causative
relation; at the same time it is associated with a typed polymorphism located
nearby because of LD. This confounding induces secondary correlation between
the typed polymorphism and the trait, making localistion of the true functional
polymorphism (LD mapping) possible.
Other major type of confounding observed in genetic association studies is
confounding by population (sub)structure. Let us consider a study in which sub-
jects come from two distantly related populations, say Chinese and European.
Due to genetic drift, these two populations will have very different frequencies at
many loci throughout the genome. At the same time, these two populations are
different phenotypically (prevalence of different disease, mean value of quantita-
tive traita) due to accumulated genetic and cultural differences. Therefore any
of these traits will show association with multiple genomic loci. While some
of these associations may be genuine genetic associations in a sense that ei-
ther the polymorphisms themselves, or the polymorphisms close by are causally

137
138CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

involved, most of these associations will be genetically false positives – noise as-
sociations generated by strong genetic and phenotypic divergence between the
two populations.
The scenario described above is extreme and indeed it is hard to imagine a
genetic asociation study in which two very distinct populations are so blintly
mixed and analysed not taking this mixture into account. However, a more
subtle scenario where several slighly genetically different populations are mixed
in the same study is frequently the case and a matter of concern in GWA studies.
In this chapter, we will define what is genetic structure, and how it can be
quantified (section 6.1); what are the effects of genetic structure on the standard
association tests (section 6.2) and specific association tests which take possible
genetic structure into account (section 6.3).
The text of this chapter is in large part based on a chapter of a book
published by Elsvier. We thank Elsvier for the permission to repro-
duce this material. COPYRIGHT NOTICE: Reprinted from ”Analy-
sis of Complex Disease Association Studies: A Practical Guide”, Yurii
Aulchenko, chapter ”Effects of Population Structure in Genome-wide
Association Studies”, 123-156, Copyright 2011, with permission from
Elsevier

6.1 Genetic structure of populations

A major unit of genetic structure is a genetic population. Different definitions
of genetic population are available, for example Wikipedia defines population
(biol.) as ”the collection of inter-breeding organisms of a particular species”.
The genetics of populations is ”the study of the allele frequency distribution and
change under the influence of . . . evolutionary processes”. In the framework of
population genetics, the main characteristics of interest of a group of individuals
are their genotypes, frequencies of alleles in this group, and the dynamics of
these distributions in time. While the units of interest of population genetics
are alleles, the units of evolutionary processes are acting upon are organisms.
Therefore a definition of a genetic population should be based on the chance
that different alleles, present in the individuals in question can mix together; if
such chance is zero, we may consider such groups as different populations, each
described by its own genotypic and allelic frequencies and their dynamic. Based
on these considerations, a genetic population may be defined a in the following
way:
Two individuals, I1 and I2 , belong to the same population if (a) the prob-
ability that they would have an offspring in common is greater then zero and
(b) this probability is much higher than the probability of I1 and I2 having an
offspring in common with some individual I3 , which is said to belong to other
genetic population.
Here, to have an offspring in common does not imply having a direct off-
spring, but rather a common descendant in a number of generations.
However, in gene discovery in general and GWA studies in particular we are
usually not interested in future dinamics of alleles and genotypes distributions.
What is the matter of concern in genetic association studies is potential common
ancestry – that is that individuals may share common ancestors and thus share
in common the alleles, which are exact copies of the same ancestral allele. Such
6.1. GENETIC STRUCTURE OF POPULATIONS 139

alleles are called ”identical-by-descent”, or IBD for short. If the chance of IBD
is high, this reflects high degree of genetic relationship. As a rule, relatives share
many features, both environmental and genetic, which may lead to confounding.
Genetic relationship between a pair of individuals is quantified using the
”coefficient of kinship”, which measures that chance that gametes, sampled at
random from these individuals, are IBD.
Thus for the purposes of gene-discovery we can define genetic population use
retrospective terms and based on the concept of IBD:
Two individuals, I1 and I2 , belong to the same genetic population if (a)
their genetic relationship, measured with the coefficient of kinship, is greater
then zero and (b) their kinship is much higher than kinship between them and
some individual I3 , which is said to belong to other genetic population.
One can see that this definition is quantitative and rather flexible (if not
to say arbitrary): what we call a ”population” depends on the choice of the
threshold for the ”much-higher” probability. Actually, what you define as ”the
same” genetic population depends in large part on the scope aims of your study.
In human genetics literature you may find references to a particular genetically
isolated population, population of some country (e.g. ”German population”,
”population of United Kingdom”), European, Caucasoid or even general hu-
man population. Defining a population is about deciding on some probability
threshold.
In genetic association studies, it is frequently assumed that study partic-
ipants are ”unrelated” and ”come from the same genetic population”. Here,
”unrelated” means, that while study participants come from the same popula-
tion (so, there is non-zero kinship between them!), this kinship is so low that it
has very little effect on the statistical testing procedures used to study associa-
tion between genes and phenotypes.
In the following sections we will consider the effects of population structure
on the istribution of genotypes in a study population. We will start with as-
sumption of zero kinship between study participants, which would allow us to
formulate Hary-Weinberg principle (section 6.1.1). In effect, there is no such
thing as zero kinship between any two organisms, however, when kinship is very
low, the effects of kinship on genotypic distribution are minimal, as we will see
in section 6.1.2. The effects of substructure – that is when study sample consist
of several genetic populations – onto genotypic distribution will be considered
in section 6.1.3. Finally, we will generalize the obtained results for the case
of arbitrary structures and will see what are the effects of kinship onto joint
distribution of genotypes and phenotypes in section ??.

6.1.1 Hardy-Weinberg equilibrium

To describe genetic structure of populations we will use rather simplistic model
approximating genetic processes in natural populations. Firstly, we will assume
that the population under consideration has infinitely large size, which implies
that we can work in terms of probabilities, and no random process take place.
Secondly, we accept non-overlapping

generation ⇒ gametic pool ⇒ generation

140CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

model. This model assumes that a set of individuals contributes gametes to

genetic pool, and dies out. The gametes are sampled randomly from this pool
in pairs to form individuals of the second generation. The selection acts on indi-
viduals, while mutation occurs when the gametic pool is formed. The key point
of this model is the abstract of gametic pool: if you use that, you do not need to
consider all pair-wise mating between male and female individuals; you rather
consider some abstract infinitely large pool, where gametes are contributed to
with the frequency proportional to that in previous generation. Interestingly,
this rather artificial construct has a great potential to describe the phenomena
we indeed observe in nature.
In this section, we will derive Hardy-Weinberg low (this analog of the Mendel’s
low for populations). The question to be answered is, if some alleles at some
locus segregate according to Mendel’s lows and aggregate totally at random,
what would be genotypic distribution in a population?
Let us consider two alleles, wild type normal allele (N ) and a mutant (D),
segregating at some locus in the population and apply the ”generation ⇒ ga-
metic pool ⇒ generation” model. Let us denote the ferquency of the D allele
in the gametic pool as q, and the frequency of the other allele, N , as p = 1 − q.
Gametes containing alleles N and D are sampled at random to form diploid
individuals of the next generation. The probability to sample a ”N ” gamete is
p, and the probability that the second sampled gamete is also ”N ” is also p.
According to the rule, which states that joint probability of two independent
events is a product of their probabilities, the probability to sample ”N ” and
”N ” is p · p = p2 . In the same manner, the probability to sample ”D” and then
”D” is q · q = q 2 . The probability to sample first the mutant and then normal
allele is q · p, the same is the probability to sample ”D” first and ”N ” second.
In most situations, we do not (and can not) distinguish heterozygous genotypes
DN and N D and refer to both of them as ”N D”. In this notation, frequency
of N D will be q · p + p · q = 2 · p · q. Thus, we have computed the genotypic
distribution for a population formed from a gametic pool in which the frequency
of D allele was q.
To obtain the next generation, the next gametic pool is generated. The
frequency of D in the nect gametic pool is q 2 + 12 · 2 · p · q. Here, q 2 is the
probability that a gamete-contributing individual has genotype DD; 2 · p · q
is the probability that a gamete-contributing individual is N D, and 21 is the
probability that N D individual contributes D allele (only half of the gametes
contributed by individuals with N D genotype are D); see Figure 6.1. Thus the
freqeuncy of D in the gametic pool is q 2 + 21 · 2 · p · q = q · (q + p) = q – exactly
the same as it was in previous gametic pool.
Thus, if assumptions of random segregation and aggregation hold, the ex-
pected frequency of N N , N D and DD genotypes are stable over generations
and can be related to the allelic frequencies using the follwoing relation

P (N N ) = (1 − q) · (1 − q) = p2 ,
P (N D) = q · (1 − q) + (1 − q) · q = 2 · p · q, (6.1)
P (DD) =q·q = q2

which is known as Hardy-Weinberg equlibrium (HWE) point.

There are many reasons, in which random segregation and aggregation, and,
consequently, Hardy-Weinberg equilibrium, are violated. It is very important
6.1. GENETIC STRUCTURE OF POPULATIONS 141

Individuals
P(DD) P(DN) P(NN)

D D D N N N

P(D) P(N)=1-P(D)
Alleles
Figure 6.1: Genotypic and allelic frequency distribution in a population; q =
P (D) = P (DD) + 12 · P (DN ).

to realize that, especially if the study participants are believed to come from
the same genetic population, most of the times when deviation from HWE is
detected, this deviation is due to technical reasons, i.e. genotyping error. There-
fore testing for HWE is a part of the genotypic quality control procedure in most
studies. Only when the possibility of technical errors is eliminated, other possi-
ble explanations may be considered. In a case when deviation from HWE can
not be explained by technical reasons, the most frequent explanation would be
that the sample tested is composed of representatives of different genetic pop-
ulations, or more subtle genetic structure. However, unless study participants
represent a mixture of very distinct genetic populations – the chances of which
coming unnoticed are low – the efffects of genetic structure on HWE are difficult
to detect, at least for any single marker, as you will see in the next sections.

6.1.2 Inbreeding
Inbreeding is preferential breeding between (close) relatives. An extreme ex-
ample of inbreeding is a selfing, a breeding system, observed in some plants.
The inbreeding is not uncommon in animal and human populations. Here, the
main reason for inbreeding are usually geographical (e.g. mice live in very small
interbred colonies – dems – which are usually established by few mice and are
quite separated from other dems) or cultural (e.g. noble families of Europe).
Clearly, such preferential breeding between relatives violates the assumption
of random aggregation, underling Hardy-Weinberg principle. Relatives are likely
to share the same alleles, inherited from common ancestors. Therefore their
progeny has an increased chance of being autozygous – that is to inherit a copy
of exactly the same ancestral allele from both parents. An autozygous genotype
is always homozygous, therefore inbreeding should increase the frequency of
homozygous, and decrease the frequency of heterozygous, genotypes.
Inbreeding is quantified by the coefficient of inbreeding, which is defined as
the probability of autozygosity. This coefficient may characterize an individual,
142CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

A A B B A B
P=1/2 P=1/2

C D E F C D E F
P=1/2 P=1/2 P=1
P=1

G H G H

P=1/2 P=1/2

J J
6
P=1/2 =1/64

Figure 6.2: Inbred family structure (A) and probability of individual ”G” being
autozygous for the ”Red” ancestral allele

or a population in general, in which case this is expectation that a random

individual from the population is autozygous at a random locus. The coefficient
of inbreeding is closely related to the coefficient of kinship, defined earlier for a
pair of individuals as the probability that two alleles sampled at random from
these individuals, are IBD. It is easy to see that the coefficient of inbreeding for
a person is the same as the kinship between its parents.
Let us compute the inbreeding coefficient for the person J depicted at figure
6.2. J is a child of G and H, who are cousins. J could be autozygous at
for example ”red” allele of founder grand-grand-parent A, which could have
been transmitted through the meioses A ⇒ D, D ⇒ G, and G ⇒ J, and
also through the path A ⇒ E, E ⇒ H, and H ⇒ J (Figure 6.2 B). What
is the chance for J to be autozygous for the ”red” allele? The probability
that this particular founder allele is transmitted to D is 1/2, the same is the
probability that the allele is transmitted from D to G, and the probability that
the allele is transmitted from G to J. Thus the probability that the ”red” allele
is transmitted from A to J is 1/2·1/2·1/2 = 1/23 = 1/8. The same is the chance
that that allele is transmitted from A to E to H to J, therefore the probability
that J would be autozygous for the red allele is 1/23 · 1/23 = 1/26 = 1/64.
However, we are interested in autozygosity for any founder allele; and there are
four such alleles (”red”, ”green”, ”yellow” and ”blue”, figure 6.2 B). For any of
these the probability of autozygosity is the same, thus the total probability of
autozygosity for J is 4 · 1/64 = 1/24 = 1/16.
Now we shall estimate the expected genotypic probability distribution for a
person characterized with some arbitrary coefficient of inbreeding, F – or for a
population in which average inbreeding is F . Consider a locus with two alleles,
A and B, with frequency of B denoted as q, and frequency of A as p = 1 − q.
If the person is autozygous for some founder allele, the founder allele may be
either A, leading to autozygous genotype AA, or the founder allele may be B,
6.1. GENETIC STRUCTURE OF POPULATIONS 143

leading to genotype BB. The chance that the founder allele is A is p, and the
chance that the founder allele is B is q. If the person is not autozygous, then the
expected genotypic frequencies follow HWE. Thus, the probability of genotype
AA is (1 − F ) · p2 + F · p, where the first term corresponds to probability that
the person is AA given it is not inbred (p2 ), multiplied by the probability that
it is not inbred (1 − F ), and the second term corresponds to probability that a
person is AA given it is inbred (p), multiplied by the probability that the person
is inbred (F ). This computations can be easily done for all genotypic classes
leading to the expression for HWE under inbreeding.

P (AA) = (1 − F ) · p2 + F · p = p2 + p · q · F
P (AB) = (1 − F ) · 2 · p · q + F · 0 = 2 · p · q · (1 − F ) (6.2)
P (BB) = (1 − f ) · q 2 + F · q = q2 + p · q · F

How much is inbreeding expected to modify genotypic distribution in human

populations? The levels of inbreeding observed in human genetically isolated
populations typically vary between 0.001 (low inbreeding) to 0.05 (relatively
high), see Pardo et al. (2005); Rudan et al. (2003). The genotypic distribution
for q = 0.5 and different values of the inbreeding coefficient is shown at the figure
6.3.
What is the power to detect deviation from HWE due to inbreeding? For
that, we need to estimate the expectation of the χ2 statistics (the non-centrality
parameter, NCP) used to test for HWE. The test for HWE is performed using
standard formula
X (Oi − Ei )2
T2 = (6.3)
i
Ei

where summation is performed over all classes (genotypes); Oi is the count

observed in i-th class, and Ei is the count expected under the null hypothesis
(HWE). Under the null hypothesis, this test statistic is distributed as χ2 with
number of degrees of freedom equal to the number of genotypes minus the
number of alleles.
Thus the expectation of this test statistic for some q, F , and N (sample size)
is

(N (q 2 +pqF )−N q 2 )2 )−N 2pq)2 (N (p2 +pqF )−N p2 )2

E[T 2 ] = N q2 + (N 2pq(1−F
N 2pq + N p2
(N pqF )2 pqF )2 )2
= N q2 + (−2N
N 2pq + (NNpqF
p2
2 2
= N p F + 2N pqF + N q F 2 2 2 (6.4)
= N F 2 (p2 + 2pq + q 2 )
= N · F2

Interestingly, the non-centrality parameter does not depend on the allelic

frequency. Given the non-centrality parameter, it is easy to compute the power
to detect deviation from HWE for any given F . For example, to achieve the
power of > 0.8 at α = 0.05, for a test with one degree of freedom the non-
centrality parameter should be > 7.85. Thus, if F = 0.05, to have 80% power,
N · F 2 > 7.85, that is the required sample size should be N > 7.85 7.85
F 2 = 0.0025 =
3140 people.
144CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

0.5
0.4
0.3
P(g)

0.2
0.1
0.0

AA AB BB

Genotype, g

Figure 6.3: Genotypic probability distribution for a locus with 50% frequency
of the B allele; black bar, no inbreeding; red, F = 0.001; green, F = 0.01; blue,
F = 0.05
6.1. GENETIC STRUCTURE OF POPULATIONS 145

Table 6.1: Expected proportion of markers deviating from HWE in a sample of

N people coming from a population with average inbreeding F . Proportion of
markers is shown for particular test statistic threshold, corresponding to nominal
significance α.
α
N F 0.05 10−4 5 · 10−8
−4
0.001 0.0501 1.008 · 10 5.077 · 10−8
−4
1,000 0.005 0.0529 1.205 · 10 7.025 · 10−8
−4
0.010 0.0615 1.885 · 10 14.503 · 10−8
−4
0.001 0.0511 1.081 · 10 5.784 · 10−8
−4
10,000 0.005 0.0790 3.544 · 10 36.991 · 10−8
−4
0.010 0.1701 19.231 · 10 426.745 · 10−8

Thus, even in populations with strong inbreeding, rather large sample sizes
are required to detect the effects of inbreeding on HWE at a particular locus,
even at relatively weak significance level of 5%.
While the chance that deviation from HWE due to inbreeding will be sta-
tistically significant is relatively small, inbreeding may have clear effects on the
results of HWE testing in GWA study. Basically, if testing is performed at
a threshold corresponding to nominal significance α, a proportion of markers
which show significant deviation will be larger than α. Clearly, how large this
proportion will be depends on the inbreeding and on size of the study – expec-
tation of T 2 is a function of both N an F . A proportion of markers showing
significant deviation form HWE at different values of inbreeding, sample size,
and nominal significance threshold, is shown in table 6.1. While deviation of
this proportion from nominal one is minimal at large α’s and small sample sizes
and coefficients of inbreeding, it may be 10-fold and even 100-fold higher than
the nominal level at reasonable values of N and F for smaller thresholds.

6.1.3 Mixture of genetic populations: Wahlund’s effect

Consider the following artificial example. Imagine that recruitment of study
participants occurs at a hospital, which serves two equally size villagec(V1 and
(V2 ); however, the villages are very distinct because of cultural reasons, and
most marriages occur within a village. Thus these two villages represent two
genetically distinct populations. Let us consider a locus with two alleles, A and
B. The frequency of A is 0.9 in V1 and it is 0.2 in V2 . In each population,
marriages occur at random, and HWE holds for the locus. What genotypic
distribution is expected in a sample ascertained in the hospital, which represents
a 1 : 1 mixture of the two populations?
The expected gentypic proportions are presented in table 6.2. First, assum-
ing that HWE holds for each of the populations, we can compute genotypic
proportions within these (rows 1 and 2 of table 6.2). If our sample represents
a 1 : 1 mixture of these populations, then the frequency of some genotype is
also 1 : 1 mixture of the respective frequencies. For example, frequency of AA
genotype would be 0.81 0.04
2 + 2 = 0.425, and so on. The frequency of the A allele
in pooled sample will be 0.425 + 0.25
2 = 0.55. Based on this frequency we would
expect genotypic frequency distribution of 0.3, 0.5 and 0.2, for AA, AB, and
146CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

Table 6.2: Genotypic proportions in a mixed population

Village %Sample p(A) P (AA) P (AB) P (BB)
V1 50 0.9 0.81 0.18 0.01
V2 50 0.2 0.04 0.32 0.64
Observed
Pooled 100 0.55 0.425 0.25 0.325
Expected
0.30 0.50 0.20
Difference
0.125 −0.250 0.125

Table 6.3: Genotypic proportions of P P ARGenABEL-packagemma P ro12Ala

genotype in a mixed population
Ethnics %Sample p(P ro) P (P ro/P ro) P (P ro/Ala) P (Ala/Ala)
Caucasian 50 0.85 0.7225 0.2550 0.0225
Afro-American 50 0.99 0.9801 0.0198 0.0.001
Observed
Pooled 100 0.92 0.8513 0.1374 0.0113
Expected
0.8464 0.1472 0.0064
Difference
0.0049 −0.0098 0.0049

BB, respectively. As you can see the observed distribution has much higher
frequencies of homozygous genotypes – excess of homozygotes.
It is notable, that the differences between the observed homozygotes fre-
quencies and these expected under HWE are both 0.125, and, consequently, the
observed heterozygosity is less than that expected by 0.125 · 2 = 0.25.
The phenomenon of deviation from HWE due to the fact that considered
population consist of two sub-populations, is known as ”Wahlund’s effect”, after
the scientist who has first considered and quantified genotypic distribution under
such modelWahlund (1928).
Such marked differences between observed and expected under HWE are
very easily detected; for the above example, a sample of ≈ 35 people is enough
to reject the hypothesis of HWE (power > 80% at α = 0.05).
However, the differences we can see in real life are not so marked. For
example, the common Pro allele at position 12 of the peroxisome proliferator-
activated receptor gamma is associated with increased risk for type 2 diabetes.
The frequency of the Pro allele is about 85% in European populations and
Caucasian-Americans, about 97% in Japan and 99% in African-American (see
table 1 from Ruiz-Narvez (2005)). Table 6.3 shows hypothetical observed and
expected genotypic proportions in a sample composed of 50% Caucasians and
50% African-American.
You can see that observed distribution and the one expected under HWE are
very similar; only a sample as large as 1,800 people would allow detection of the
deviation from HWE (power > 80% at α = 0.05). The situation is similar for
most genes observed in real life – while the frequencies may be (or may be not)
6.1. GENETIC STRUCTURE OF POPULATIONS 147

Table 6.4: Expected genotypic proportions in a mixed population; Fst is defined

by equation 6.6
Population Prop. p(B) P (AA) P (AB) P (BB)
P1 m q1 p21 2p1 q1 q12
2
P2 (1 − m) q2 p2 2p2 q2 q22
Observed
Pooled 1.0 q = mq1 mp21 2mp1 q1 mq12
+(1 − m)q2 ; +(1 − m)p2 ; +2(1 − m)p2 q2 ; +(1 − m)q22
2

Expected
p2 2pq q2
Difference
pqFst −2pqFst pqFst

very different for populations, which diverged long time ago, for relatively close
populations expected frequency differences are small and large sample sizes are
required to detect deviation from HWE due to Wahlund’s effect at a particular
fixed locus.
Let us summarize, what genotypic proportions are expected in a sample,
which is a mixture of two populations. Let each population is in HWE, and
the frequency of the B allele is q1 in population one and q2 in population two.
Let the proportion of individuals coming from population one is m in the mixed
population, and consequently the proportion of individuals from population two
is (1 − m). The allelic frequencies, and genotypic distributions in the original
and mixed populations are presented in tale 6.4.
The frequency of the B allele in the mixed population is just the weighted
average of the allelic frequencies in the two populations, q = m · q1 + (1 − m) · q2 .
Let us denote the frequency of the A allele as p = 1 − q. It can be demonstrated
that the genotypic frequency distribution in the mixed sample is the function of
the frequency of allele B in the sample, q, and ”disequilibrium” parameter D:

P (AA) = p2 + p · q · Fst
P (AB) = 2 · p · q · (1 − Fst ) (6.5)
P (BB) = q 2 + p · q · Fst
where
m · (1 − m) · (q1 − q2 )2
Fst = (6.6)
p·q
You can see that equation 6.5, expressing the genotypic frequencies distri-
bution under Wahlund’s effect, is remarkably similar (actually, is specifically
re-written in a form similar) to the equation 6.2, expressing the genotypic pro-
portions under the effects of inbreeding. Again, the reason is that Fst (as well
as F of equation 6.2) is easily estimated from the data as the ratio between
the observed and expected variances of the genotypic distributions. Then the
2
expected non-centrality parameter for the test of HWE is simply N · Fst , where
N is the sample size. Therefore our results concerning the proportion of tests
expected to pass a particular significance threshold when genome-wide data are
analyzed (table 6.1) hold, with replacement of F with Fst .
We can compute that the values of Fst , corresponding to the population
mixtures presented in tables 6.2 and 6.3 are 0.49 and 0.067, respectively, which
148CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

Table 6.5: Counts of cases and controls with different genotypes

Genotype
Status AA AB BB Total
Case r0 r1 r2 R
Control s0 s1 s2 S
Total n0 n1 n2 N

Table 6.6: Counts of alleles in cases and controls

Status A B Total
Case 2 · r0 + r1 r1 + 2 · r2 2·R
Control 2 · s0 + s1 s1 + 2 · s2 2·S
Total 2 · n0 + n1 n1 + 2 · n2 2·N

gives us a shortcut to estimate the sample size required to detect deviation from
HWE due to Wahlund’s effect (at α = 0.05 and power 80%): N > 7.85/0.492 ≈
32 and N > 7.85/0.0672 ≈ 1771.
A typical value of Fst for European populations is about 0.002 (up to
0.023Nelis et al. (2009)); very large sample sizes are required to detect de-
viation from HWE at any given locus at such small Fst ’s. However, the effects
onto the proportion of markers failing to pass HWE test in GWA may be visibly
inflated (table 6.1).

6.2 Effects of population structure on standard

tests for association
6.2.1 Standard tests for genetic association
Standard tests for association between genes and a binary trait are the test
for allele frequency difference between cases and controls, and the Armitage’s
trend test for proportions (that the proportion of cases changes across genotypic
groups). For quantitative traits, one of the standard tests is the score test for
association, which is closely related – even equivalent – to the Armitage’s trend
test.
We will start with presenting the study data as 2x3 table, where the rows cor-
respond to the case/control status and columns correspond to genotypic groups,
and the cells contain counts of events (table 6.5). For example, r0 is the number
of cases with genotype AA, s0 is the number of controls with genotype AA and
so on.
This table can be re-arranged in a 2x2 allelic table, presented in table 6.6.
Each cell of this table contains the counts of alleles present in cases and controls,
e.g. total number of A alleles in cases is 2·r0 (twice the number of cases who are
homozygous for the A allele) plus the number of A alleles present in heterozygous
cases (r1 ).
Based on these tables, we can test if the allelic frequency is different between
the cases and controls, using standard χ2 test, formulated as
6.2. EFFECTS OF POPULATION STRUCTURE ON STANDARD TESTS FOR ASSOCIATION149

X (Oi − Ei )2
T2 = (6.7)
i
Ei

where summation is performed over all cells (defined by combination of geno-

type/allele and phenotype); Oi is the count observed in i-th class, and Ei is
the count expected under the null hypothesis (equal frequencies in cases and
controls). Under the null hypothesis, this test statistic is distributed as χ2 with
number of degrees of freedom equal to the number of independent classes.
The null hypothesis assumes that the frequency of the A allele is the same in
both cases and controls, and is equal to the frequency for A in the total sample:
p = 2·n2·N
0 +n1
. Thus the expected count of A in cases is 2 · R · p, and the expected
count of B alleles is 2 · R · (1 − p). Similarly, for cases, the expected count of A
is 2 · S · p, and the expected count of B is 2 · S · (1 − p).
Now, for the table 6.6 we can re-write the allelic test as
((2r0 +r1 )−2Rp)2 2
TA2 = 2Rp + ((r1 +2r2R(1−p)
2 )−2R(1−p))

2 2
+ ((2s0 +s2Sp
1 )−2Sp)
+ ((s1 +2s2S(1−p)
2 )−2S(1−p))

With some algebra, it can be demonstarted that this expression simplifies to

2N {2N (r1 + 2r2 ) − 2R(n1 + 2n2 )}2

Ta2 = (6.8)
(2R)2(N − R){2N (n1 + 2n2 ) − (n1 + 2n2 )2 }

Under the null hypothesis that the frequency of alleles is the same in cases
and controls (and, as you will see later, that HWE halds in total sample) the
test statistic is distributed as χ2 with one degree of freedom.
An alternative, Armitage’s trend test for proportions, can be used to test
the null hypothesis. This test is performed using the 2x3 genotypic table (6.5).
The null hypothesis assumes that the frequency of cases is the same in all
genotypic groups; alternative is that the frequency is not the same, but are,
however, not totally arbitrary. As it follows from the name of the test, a trend
in proportions is assumed, that is the frequency of cases among people with
heterozygous genotypes AB should be exactly between the frequencies of cases
in two homozygous classes. This hypothesis may be formalized using parameters
beta0 , the expected frequency of cases in the AA group, and β1 , the increase
in frequency of cases in the AB group. The expected frequency of cases in the
AB group is then (β0 + β1 ); the frequency of cases in the BB group is assumed
to be (β0 + 2 · β1 ). Parameters p0 and p1 can be estimated using the maximum
(r1 +2r2 )−pR
likelihood; β̂1 = (n 1 +4n2 )−pN
, where p = 2n2N
2 +n1
is the frequency of the B allele
R
in the total sample. Then, β0 is estimated as β̂0 = N − β̂1 p. Is can be shown
that the chi-square test 6.7 based on these expectations takes the form

N {N (r1 + 2r2 ) − R(n1 + 2n2 )}2

Tt2 = (6.9)
R(N − R){N (n1 + 4n2 ) − (n1 + 2n2 )2 }
√
It can be shown that Tt2 = Ta2 if n1 = 2 n0 · n2 , which is equivalent to the
condition that HWE haldspexactly, in which pcase the counts of heterozygotes
should be 2N pq, with p = n0 /N and q = n2 /N . When HWE holds for the
total sample, the two tests are (at least assymptotically) equivalent. Thus the
150CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

test Ta2 is the test for association in presence of HWE; when HWE does not
hold, even in absence of association, the values of Ta2 are greater then the values
of Tt2 test, possibly leading to false positive conclusions about association in
presence of deviations from HWE. Therefore the trend test Tt2 is to be prefered
when testing for genetic association Sasieni (1997).
For study of association between genotype and a quantitative traits, linear
regression analysis is performed. Let us denote the vector of phenotypes as
y, with particular values yi (i = 1, . . . , N ). Let code the genotypes with a
quantitative variable, which reflects the number of B allels. Thus, we will code
AA as 0, AB as 1, and BB as 2. Let us denote the vector of genotypes as g,
with particular values gi (i = 1, . . . , N ) taking the value of 0, 1 or 2. The linear
regression model assumes that the expectation of the trait is

E[yi ] = µ + βg · gi

where µ is intercept and βg is the coefficient of regression of the genotype on

the phenotype.
The estimate of βg is provided by a well-known expresion

xi gi /N − xi gi /N 2
P P P P P P
Cov(y, g) N xi g i − xi g i
β̂ = = P 2 = (6.10)
gi /N − ( gi )2 /N 2 N gi2 − ( gi )2
P P P
V ar(g)
V ar(y)
Under the null hypothesis, the variance of β̂ is V ar0 (β̂g ) = N ·V ar(g) , and
the score test statistic

β̂g2 Cov(y, g)2 N · V ar(g) Cov(y, g)2

Tq2 = = · = N (6.11)
V ar0 (β̂g ) V ar(g)2 V ar(y) V ar(y) · V ar(g)

is distributed as χ2 with one degree of freedom. Here, V ar0 β̂g denotes the
variance under the null hypothesis. Denote the correlation between y and g,
√ Cov(y,g) , as r. Then Tq2 = N · r2 .
V ar(y)·V ar(g)
It is worth to note that Armitage’s trend test can be expressed as a linear
regression test. In the table 6.5 Let us code the genotypic groups based on the
number of B allels. Thus, we will code AA as 0, AB as 1, and BB as 2. Let
us also code the cases with ”1” and controls with ”0”. Let y denote the vector
of phenotypes and g the vector of genotypes coded in this way. Then, if we
perform linear regression of the case-control status onto the number of B alleles
in the genotype, the estimate of the regression coefficient is (following 6.10)
P P P
xi gi − xi gi /N (r1 + 2r2 ) − pR
β̂ = =
gi2 − ( gi )2 /N
P P
(n1 + 4n2 ) − pN
Thus, the expected proportions in the Armitage’s trend test are provided
by solution of linear regression equation, in which regression of case-control
status, coded as ”0” and ”1” is performed onto the number of B alleles. It can
be further demonstrated that the trend test statistics 6.9 can be expressed as
Tt2 = N · r2 , where r2 is squared coefficient of correlation between y and g.
In the next two sections we will consider the effects of genetic structure on
the standard tests for association described above. The extensive treatment of
the problem is mainly due to the seminal works of Devlin, Roeder, Bacanu, and
6.2. EFFECTS OF POPULATION STRUCTURE ON STANDARD TESTS FOR ASSOCIATION151

colleaguesBacanu et al. (2000, 2002); Devlin and Roeder (1999); Devlin

et al. (2001); here, some of their results are repeated.
The genetic structure of the sudy population will be characterized by the
kinship matrix – a square matrix, which, for all pairs of individuals in question,
provides their pair-wise kinship coefficients (defined at page 139). The kinship
coefficient between persons i and j will be defined as fij .

6.2.2 Effects of genetic structure on standard tests

As it was demonstrated above, the Armitage’s trend test for association can
β̂ 2
be expressed as Ta2 = V ar( β̂)
. Genetic structure affects both numerator and
denominator of this expression. Indeed, if measurments are dependent, the
varaiance of the estimate is likely to be under-estimated if such dependence is
not accounted for, leading to the inflation of the tests statistic. Secondly, even
in absence of association, the estimate of the effect may be biased, again, leading
to increased value of the test statistic.
Let us consider following artificial example. Let study cases are closely
related individuals – sibs1 – coming from one family, and study contols are sibs
from other family. In essence, we compare the squared frequency difference
between the two groups to the variance of this difference. If groups are formed
by independent individuals, we can detect arbitrary small frequency differences
by increasing the sample size, and consequently, decreasing the variance of the
frequency estimate. This definitely is not the case when our study sample
consist of a large sibship – whatever is the sibship size, only four alleles may be
present in the parents, and the precision of the estimate of the allele frequency
in the general population is limited – as if we had two people at maximum. If
we do not take this consideration into account, we are likely to over-estimate
the precision of the frequency estimate; the denominator of the test statistic
becomes small, and the statistic becomes large. Secondly, the parents of the
”case” and ”control” sibships have a high chance to be genotypically different,
just by chance; and any genotypic configuration leading to the different number
of B alleles will be reflected in diffence in frequencies between sibships, given
the sibships are large. Assuming HWE, the chance that two parental couples
will have different number of B alleles is one minus the chance that they will
have the same number of B alleles:

P (#B1 6= #B2 ) = 1 − (P (#B1 = #B2 = 0) + P (#B1 = #B2 = 1) + . . . + P (#B1 = #B2 = 4))

= 1 − (P (#B1 = 0)P (#B2 = 0) + . . . + P (#B1 = 4)P (#B2 = 4))
= 1 − (q 8 + 16p2 q 6 + 36p4 q 4 + 16p6 q 2 + p8 )

which, for a common allele with frequency of 0.2 translates to the probability
0.64.
Let us quantify these two sources of bias. Following Devlin et al. (2001)
let us assume that the number of cases and the number of controls is the same,
N . Let us denote the vector of genotypes, coded as 0, 1, and 2 of cases as X,
with Xi (i = 1, . . . , N ) being the genotype of the i-th case, and the vector of
genotypes of controls as Y (with Yj , j = 1, . . . , N being the genotype of j-th
control). The trend and the allelic tests statistic are proportional to the square
1 brothers and sisters
152CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

P P
of T = Xi − Yj . The variance of this statistic, in general form, is
PN PN
V ar(T ) = P V ar(Xi ) + i=1 V ar(Y
i=1 P i)
+2 i<j Cov(Xi , Xj ) + 2 i<j Cov(Yi , Yj ) (6.12)
P P
−2 i j Cov(Xi , Yj )

Let us consider a sitatuation in which cases come from one popualtion, and
controls from the other population; each of the populations is in HWE and
the difference between the populations is characterized with Fst (see equation
6.6). Under this model, V ar(Xi ) = V ar(Yi ) = 2pq(1 + Fst ), and the covariance
between any pair of genotypes from the same population is 4pqFst . Then
V ar(T ) = 2pq(1 + Fst ) · N + 2pq(1 + Fst ) · N
+2 · 2pqFst · N (N − 1)/2 + 2 · 2pqFst · N (N − 1)/2 + 0
= 4pqN · (1 + Fst + 2Fst (N − 1)) ≈ 4pqN (1 + 2N Fst )
(6.13)
here, 4pqN corresponds to the binomial variance of T in absence of genetic
structure, while Fst + 2 · Fst · (N − 1) reflects the inflation of the variance.
As the second term is the function of the sample size, large inflation may be
achieved even with small vaules of Fst . Note that here the sample size is 2N ,
as we assumed N cases and N controls.
Above we have considered an example in which we know the Fst between
the population of cases and the population of controls. In a practical study, a
number of cases and controls is usually sampled from each genetically different
population. Let the proportion of individuals sampled from population c among
cases is ac and the proportion of individuals coming from that population among
controls is uc . Then, it can be shown that
X
V ar(T ) ≈ 4pqN (1 + 2 · N · Fst · (ac − uc )2 ) (6.14)
c

As it follows from this equation, the variance of the estimated frequency dif-
ference depends on the composition of the sample. Maximal inflation is achieved
when the cases and controls are sampled from different populations, while if
2
P
c (ac − uc ) = 0 – which is achieved by sampling equal number of cases and
controls from each sub-population – the variance inflation is minimal.
These results canbe generalized to arbitrary relations between cases and con-
X
trols. Let us denote kinship between cases i and j as fij , kinship between con-
Y XY
trols as fij , and the kinship between a case and a control as fij . ThenDevlin
et al. (2001)
V ar(T ) = 4pqN · (1 +Fst + N2 i<j fij X
P
2
P Y 2
P P XY (6.15)
+ N i<j fij − N i j fij )

This equation demonstrates that the variance is inflated not only in a case
when cases and controls come from genetically distinct populations, but that
X
also ”criptic relations” among either cases (some fij > 0) or controls (some
Y
fij > 0) may lead to increased variance of the test. Because summation (e.g.
X
P
i<j fij ) is performed overl all pairs of cases/controls, cryptic relations may
have a strong impact onto inflation of the variance.
Generally, the variance of the T can be expressed as V ar(T ) = 4N pq(1+Fst +
D(f X , f Y , f XY )). Here, 4N pq corresponds to the binomial variance assuming
6.2. EFFECTS OF POPULATION STRUCTURE ON STANDARD TESTS FOR ASSOCIATION153

HWE and independence between cases and controls, the second term – Fst –
accounts for increase in variance due to deviation from HWE, and the last,
which is a function of kinship – D(f X , f Y , f XY ) – accounts for dependencies
T2
bewteen cases and controls. The test 4N pq(1+Fst +D(f X ,f Y ,f XY )) is distributed

as χ2 with one degree fo freedom.

T2
The allelic and Armitage’s trend statistics can be expressed as Ta2 = 4N pq
2
T
and Tt2 = 4N pq·(1+F st )
. What is their distribution when there is relatedness /
stratification bewteen cases and controls? Clearly, both of them are distributed
as τ 2 · χ2 , where τ 2 > 1 is the variance inflatiopn factor. It easy to show
that it is equal to τ 2 = 1 + Fst + D(f X , f Y , f XY ) for the allelic and τ 2 =
1+Fst +D(f X ,f Y ,f XY )
1+Fst for the trend test.
In above, we have accounted for the over-dispersion of the test statistic
due to possible relatedness between the cases and controls, and demonstrated
that the allelic and the trend test are inflated by some constant τ 2 , which
reflects the inflation of the variance, V ar(T ) = σ 2 · τ 2 , where σ 2 is the variance
estimated under the assumption of independence. However, in the beginning of
this section we have discussed another source of the infaltion of the test statistic
– the bias in frequencies between cases and controls due to confounding. Let
cases and controls are sampled from different populations, whose difference is
characterized by Fst , and let for some marker the allele frequency difference
– which occures entirely due to genetic structure of the sampels – between
casese and controls is d. The test statistic T is then distributed as Normal with
mean N d and variance computed in previous section, σ 2 · τ 2 , N (N d, σ 2 τ 2 ). To
address the issue of bias, we need to figure out the distribution of the mean
– 2N d. In general, the expected frequency difference is zero – indeed, if we
consider a large number of subpopulations, or a large number of markers, if
allelic frequencies are determined by random effects, the deviation is likely to
occure in any direction, resultin in zero on average. The variance of the d is
2pqFst , thus the variance of 2N d is (2N )2 2pqFst .2 Thus finally, T is expected to
be distributed as N (0, σ 2 · τ 2 + 8pqFst N 2 ). If we consider the variance of allelic
test, σ 2 = 4pqN , then T ∼ N (0, σ 2 (τ 2 + 2Fst N )), and, denoting 2Fst N as ν 2 ,
T ∼ N (0, σ 2 (τ 2 + ν 2 )). Similar expression may be obtained for the Armitage’s
trend test.
Thus, in presence of genetic structure both allelic and the trend test are
distributed as (τ 2 + ν 2 ) · χ2 . Note that while σ 2 is a function of allelic frequency,
the inflation factors τ 2 and ν 2 depend on the differentiation between the popu-
lations in question, as measured by Fst , sample size, N , and the composition of
the case-control sample (ac , uc ), but does not depend on the allele frequency.
As it was mentioned earlier, the Armitage’s trend test can be re-formulated
as a regression-based score test, in which genotypes are coded as 0, 1, and 2,
and phenotypes as 0 and 1. Therefore all above arguments apply to the analysis
β̂g2
of quantitative traits as well. The test statistic Tq2 = V ar(β̂g )
in which β̂ and
V ar(β̂) are estimated using independence assumption is inflated because of over-
dispersion (V ar(β̂) is under-estimated if the relatedness is not accounted for),
and also β̂g may be biased, because different genetic populations may well be
2 For the case when some proportion, a , among cases comes from population c, and some
c P
proportion of controls, uc , comes from population c V ar(d) = 2pqFst c (ac − uc )2
154CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

characterized by different mean value of the trait (e.g. because of environmental

influences). It can be shown that also for quantitative traits Tq2 is distributed as
(τ 2 + ν 2 ) · χ2 , where τ 2 is inflation due to over-dispersion, and ν 2 is the inflation
because of bias.

6.2.3 Genomic control

From previous section it follows that in presence of genetic structure the stan-
dard association statistics may be inflated and the distribution is described as
(τ 2 + ν 2 ) · χ2 . Let us denote (τ 2 + ν 2 ) as λ for short. λ depends on the ge-
netic structure of the sample, as characterized by pairwise kinship, and sample
size, N . For binary trait analysis, it also depends on the composition of the
case-control sample, as expressed by the proportio of cases/conrols coming from
particular population, c (ac , uc ). For quantitative trait, it depends of neritabil-
ity of the trait, and environmentally detemined differences co-occuring with
the difference in kinship. However, λ does not depend on the allele frequency.
Therefore, for any particular study sample, if Fst is constant over the genome,
λ is also a constant.
Therefore λ can be estimated from the genomic data, using ”null” loci – a
set of random markers, which are believed not to be associated with the trait.
This estimate can then be used to correct the values of the test statistic at
the tested loci – a procedure, known as ”genomic control”. The test statistics
computed from these loci thus estimates the distrbution of the test statistic
under the null hypothesis of no association. Let us consider M ”null” markers,
and denote the test statistic obtained from i-th marker as Ti2 . Given genetic
structure determines inflation λ, T 2 ∼ λ · χ21 . The mean of a random variable
coming from χ21 is equal to 1; if a random variable comes from λ · χ21 , the mean
would be λ. Thus we can estimate λ as the mean of the obtained ”null” tests.
In practice, it is recommended to use the ratio between the observed median
and the one expected for χ21 :

M edian(Ti2 )
λ̂ = (6.16)
0.4549

For the tested markers, the corrected value of the test statistic is obtained by
2 2
simple division of the original test statistic value on λ̂, Tcorrected = Toriginal /λ.
Note that this procedure is correct only if the same number of study participants
was typed for any marker, as you will see later.
As you can see, the genomic control procedure is computationally extremely
simple – one needs to compute the test statistic using a simple test (e.g. score
test), compute the median to estimate λ, and divide the original test statistics
values onto λ̂.
How to choose ”null loci” is a GWA study? In a genome, we expect that
a small proportion of markers is truly associated with the trait. Therefore in
practice, all loci are used to estimate λ. Of cause, if very strong (or multiple
weak) true associations are present, true association will increase the average
value of the test, and genomic control correction will be conservative. To relax
this, it has been suggestedSladek et al. (2007) to use, say, 95% of the least
significant associations for estimation of the inflation factor; however, the ques-
6.2. EFFECTS OF POPULATION STRUCTURE ON STANDARD TESTS FOR ASSOCIATION155

tion of selection of cut-off is not obvious, and all markers are used in practice
in most studies.
We have observed that the inflation factor λ is a function of sample size, N
– the bigger is the sample size, the larger is inflation. If this is the case, how
can we compare inflations between different GWA studies? A good idea is to
use a standardized inflation, say, inflation per 1,000 subjects. For a quantitative
trait analysis, inflation factor can be expressed as λ ≈ (1 + N ∗ D(sample)),
where N ∗ D(sample) is a term, which grows linear with sample size, at some
rate determined by sample characteristics.
Therefore a standarsized inflation factor can be etimated as

λ̂ − 1
λ̂1000 = 1 + · 1000,
N

where N is the sample size and λ̂ is the estimate from the total sample.
What about a case-control sample? For such samples, a standardized λ is
computed for a fixed number of cases and controls, say 1000; we will denote such
standardized inflation as λ1000,1000 to distinguish it from the one computed for
a quantitative traits. If we denote the number of cases as Na and the number
of controls as Nu , then

500 · (λ̂ − 1) · (Na + Nu )

λ̂1000,1000 = 1 +
Na · Nu

Clearly, we have assumed that in our GWA study the sample size was equal
for all studied markers; this fact allows for very simple estimation of λ̂ (equation
6.16) and correction of the test statistic. It may happen, however, that the
number of participants typed for different marker loci is different; for example,
this may happen when all study participans were typed for SNP panel one, and
then a part of the study was additionally typed at a different panel. In such
situation, expectation of a Tq2 test for a quantitative trait can be expressed as

N
E[Tq2 ] = 1 + (λ1000 − 1) ·
1000
It is straightforward to obtain an estimate of λ1000 by performing linear re-
gression of the observed test statistic values onto the sample size used; note
that the intercept should be fixed to 1 in this procedure. Other, more effective
procedures may be thought of.
For binary traits, if sample is composed from Na cases and Nu controls, the
expression for the expected value of the test statistic is similar to that obtained
for quantitative traits, however, geometric mean of the number of cases and
controls are usedFreedman et al. (2004):
1 1
+
E[Tt2 ] = 1 + (λ1000,1000 − 1) · 1000
1
1000
1
Na + Nu

This relation can be used to estimate λ1000,1000 .

We have considered the genomic control procedure for the tests assuming
additive effects of the locus onto phenotype. Under this model, inflation of the
tests (λ) does not depend on allelic frequency, which allows a straightforward
156CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

60
50
40
Mean T^2

30
20
10
0

0.0 0.2 0.4 0.6 0.8 1.0

Figure 6.4: Mean Tt2 test statistic as frequency of the B allele, q. Models
considered: additive, green; recessive, red; dominant, blue; over-dominant, cyan.

estimation of λ and further correction of the tests. Does the same apply to other
genetic models?
In figure 6.4 the results for additive, dominant, recessive, and over-dominant
model tests are presented. Simulated study consisted of 1000 cases and 1000
controls; the differentiation between the case and control populations was as-
sumed Fst = 0.002. Armitage’s trend test was used in analysis. One can see
that when a common polymorphism is studies (0.05 < q < 0.95), the additive
model test (green line) does not depend on the frequency of the B. The range
of allele frequencies in which the additive model does not depend on these de-
pends of the values of Fst and the sample size: the larger differentiation, and
the smaller the sample size, the narrower is the range.
However, other one degree of freedom tests (recessive, red; dominant, blue;
over-dominant, cyan) do depend on the allele frequency very much across all
the range of allelic frequencies (figure 6.4, see Zheng et al. (2005) for details).
Thus, it is important to remember that the genomic control was developed
for, and works with additive genetic models; even in this framework, behaviour
6.3. ANALYSIS OF STRUCTURED POPULATIONS 157

of λ depends on allelic frequencies when theses are too low or too high – thus
application of genome-wide derived inflation factor to such marker loci may be
incorrect for loci with low minor allele frequency. What is somewaht sustaining
statistically is the fact that such correction will be conservative, and not liberal
– but the same may be worrysome biologically (e.g. missed true positives). For
other types of model, lambda can, at least in theory, be estimated taking into
account the allelic frequency of the locus in question. However, these methods
are not implemented yet in packages for genome-wide association analyses.
Other important thing to remember about genomic control is that it assumes
uniorm Fst across the genome. This may not be the case for some (e.g. selected)
genomic regions. Such regions may still generate very high test statistic, even
after appropritate genomic control correctionCampbell et al. (2005); basically,
the frequency distribution at such loci may vary from population to popula-
tion much more then the average across the genome. A number of methods
taking genetic structure of study population into account directly (structured
association, EIGENSTRAT, reviewed later), can deal with such situations.
Finally, what levels of genomic control inflation parameter are acceptable in
GWA studies? As the question is of how much the total test statistic is affected,
non-standardized inflation may be used to address the question. Despite of lack
of clear guidance, the general practice in the field is to consider the values of
λ̂ < 1.01 as small, λ̂ < 1.05 as moderate and still acceptable. If λ̂ > 1.1,
this suggest strong influence of genetic structure or other design factors on the
test statistic. While GC statistically correct method, such analysis lacks power;
good practice is to consider use of other methods, which take the structure of
the sample into account in diect manner.

6.3 Analysis of structured populations

A study populaion may be structured in at least two ways. Firstly, while study
participants may be ”independent” in that kinship between them is expected to
be very low (e.g. random sample from a large outbred population), they may
belong to rather different genetic populations. An extreme example of such
study is a study aiming to combine the data from a sample obtained by random
ascertainment from the population of Bejing and a sample from Amsterdam.
While in any of the samples people are only remotedly related (”independent”),
the two samples in question are characterized by high genetic differences; also
the environmental influences may be very different for the two populations. The
distribution of the trait under the study may be essentially different between
the two populations: not only the mean values, but also variances.
On the other hand, consider a family-based sample from genetically isolated
population. Here, environmental influences are more or less uniform for all study
participants; the distribution of the trait may be assumed to be governed by
the same set of rules for all study participants. However, they are characterized
by high and variable kinship between each other.
These two flavours of stratification may be expressed in terms of kinship.
When we talk about different ”populations”, we assume very low kinship and
high genetic differentiation between the members of these; when we talk about
”relatedness”, we assume high kinship and small degree of differentiation. This
is illustrated at figure 6.5. Roughly speaking, when we expect that study par-
158CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

Kinship

1 2 3
Population A Population B

Figure 6.5: Three samples from two populations. 1: Family-based sample; 2,3:
random sample of ”independent” people

ticipants share common ancestors few (1-4) generation ago, the kinship between
study participants is high, and the study can be classified as a family-based one
(Study 1 of figure 6.5). When common ancestors between study participants is
expected say ¿5 generations ago, the kinship between study participants is low,
and such study may be classified as a population-based sample of ”indepen-
dent” people (Studies 2 and 3 of figure 6.5). Now consider two groups of study
participants (samples 2 and 3). Any person from group 2 is expected to share
a common ancestor with a person from the same study group with much higher
probability than with a person form group 3; thus these represent two genetic
population (see page 139 for the retrospective definition of genetic population).
It shouldbe kept in mind that if kinship is very low (expected common ances-
tors dozns of generations ago), this would translate in high degree of genetic
differentiation. Moreover, this reflects a long history of isolation, which usually
means geographic sepration, and accumulation of cultural and other environ-
mentsl differences between the populations, which may be crucial in association
studies.
Of cause, a particular study is usually characterized by some mixture of
populations and some relatednes between study participants. We will, however,
first consider the two extreme scenarios – analysis of samples of ”independent”
subjects from different populations, and analysis of a family-based study.

6.3.1 Structured association

If study populayion consist of a mixture of several distinct genetic populayions,
and, within each sub-population, the study participants are remotedly related,
structured asoiation analysis may be the method of choice. In such analysis,
6.3. ANALYSIS OF STRUCTURED POPULATIONS 159

effect and its variance are estimated within each strata separately, and then
these estimates are pooled to generate global statistics. The strata can be known
from design (e.g. place of birth or ethnicity of parents) or estimated from GWA
data. By doing this, we allow for arbitrary trait distribution, characterized
by stratum-specific mean and variance, in each stratum. Clearly, this may be
crucial when different populations characterized by different environments are
included in analysis.
Combining the evidence across strata may be done using a number of meth-
ods, e.g. Cochran-Mantel-Haenszel test for binary outcomes. One of the most
simple ways to combine the evidences coming from multiple strata – or studies
– is to use fixed effects inverse variance meta-analysis.
In essence, this method is equivalent to combining likelihoods coming from
separate studies, using quadratic approximation. Denote coefficients of regres-
sion estimated in N studies/strata as βi , and associated squared standard errors
of the estimates as s2i where i ∈ 1, 2, ..., N . Note that the regression coefficient
should be reported on the same scale, e.g. centimeters, meters, or using obser-
vations reported on the standard normal scale. Define weights for individual
studies as
1
wi = 2
si
Then the pooled estimate of the regression coefficient is
PN
w i βi
β = Pi=1
N
i=1 wi

As you can see, the weights have straightforward interpretation: the bigger
the weight of the study (meaning the small is the standard error in the study),
the larger is the contribution from this study onto the pooled estimate.
The standard error of the pooled estimate is computed as
1
s2 = PN
i=1 wi

and the χ2 -test for association is computed in standard manner as

P 2
N
β2 i=1 w i βi
T2 = 2 = PN
s i=1 wi

or, alternatively, the Z-test is

PN
β i=1 wi βi
Z = = qP
s N
i=1 wi

When binary traits are studied, and results are expressed as Odds Ratios
with P − values, it is also possible to apply inverse variance method. For this,
you need to transform your Odds Ratios using natural logarithm, and, on this
scale, estimate the standard error. Generic inverse variance pooling may be
applied to the data transformed this way; the final results are back-transformed
onto Odds Ratio scale using exponentiation.
160CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

In meta-analysis procedure, it is assumed that study populations are com-

posed of genetically homogenous ”unrelated” individuals. Each study is first
analysed seprately, and genomic control should either show no inflation of the
test statistic, or a small one, which should be corrected with GC prior to com-
binaning of strata/studies. If some of the stata demonstrate large residual in-
flation, this may suggest further sub-structure present in that strata – it should
be dealt with using further sub-division and structure association, or methods
described further in this chapter, prior to attempting the pooling of results.
The meta-analytic methodology to perform structured association analysis
outlined above has advantage of simplicity. However, it assumes goodness of
quadratic approximation of the likelihood function, which is achieved if num-
bers are large. If numbers are not large (e.g. rare polymorphisms), a more com-
plicated strategy can be used. A regression model, allowing for study-specific
mean, effects of nuisance parameters, and residual variances, but unique (across
the studies) effect of interest may be formulated. Analysis of such model is
equivalent to the above-described meta-analysis procedure, under large num-
bers assumption.

6.3.2 Mixed models based approach

In this section we will consider the methods used to analyse samples in which
close family relations may be persent; theses may be family-based samples from
outbred populations, or random samples from genetically siaolted populations,
where, due to limited population size, relatives are likely to be present by chance.
The mixed model based ideology stems from the classical animal breeding
and human heritability analysis methodology, dating back to works of FisherFisher
(1918). He proposed a model in which a very large number of Mendeian genes
contribute small effects to the phenotype (infenitisimal model). We may express
this model as
yi = µ + Gi + e i (6.17)
– the value of the trait of i-th person is sum of grand population mean, contri-
bution from multiple additively acting genes of small effect (Gi ), and residual
error (ei ).
Under this assumption, the distribution of the trait in a pedigree is described
by a multivariate normal distribution with number of dimensions equal to the
number of phenotyped people, with the expectation of the trait value for some
individual i equal to
E[yi ] = µ
where µ is grand population mean (intercept). The variance-covariance matrix is
defined through its elements Vij – covariance between the phenotypes of person
i and person j:  2 2
σe + σG if i = j
Vij = 2 (6.18)
2 · fij · σG if i 6= j
2
where σG is variance due genes, fij is kinship between persons i and j, and σe2
is the residual variance. The proportion of variance explainable by the additive
genetic effects,
σ2
h2 = 2 G 2
σG + σe
6.3. ANALYSIS OF STRUCTURED POPULATIONS 161

is termed (narrow-sense) ”heritability”.

Fixed effects of some factor, e.g. SNP, may be included into the model by
modifying the expression for the expectation, e.g.

E[yi ] = µ + βg · g

leading to so-called ”measured genotypes” modelBoerwinkle et al. (1986).

In essence, this is a linear mixed effects model – the one containing both
fixed (e.g. SNP) and random (polygenic) effects. A large body of literature is
dedicated to this class of models, and finding the solution for such regression
represents no great methodological challenge. Applicability of this model to
GWA analysis was first proposed by Yu and colleaguesYu et al. (2006). Two
problems occure with application of this model to GWA analysis: firstly, if
standard Maximum Likelihood (or Restricted Maximum Lileihood) approach is
used to estimate this model, the computational time is excessiveAulchenko
et al. (2007a); secondly, the pedigree – and thus the kinship matrix – may be
not known for study participant.
Let us first consider exisiting solutions to the first problem. High compu-
tational complexity occures if both random and fixed effects part of the model
are estimated simulatneously, and this procedure is repeated for every SNP
analysed. However, we may assume that the effects of SNP are liklely to be
small. Thus, inclusion of the SNP into the model is not likely to change the es-
timates of the components of variance. Therefore a two-step procedure was sug-
gestedAulchenko and Struchalin (submitted); Chen and Abecasis (2007):
firstly, the mixed model containing all terms but these involving SNPs, is esti-
mated (this includes estimation of the components of variance, and possibly the
nuisance fixed effects); secondly, these estimates are then used to compute a test
for association for every SNP in analysis. For details on this test, see Chen and
Abecasis (2007) for original formulation, and Aulchenko and Struchalin
(submitted) for extension to multiple variables and interaction effects tested
in step two. The model is implemented in MACH and ProbABEL software
packages.
Another method, bearing close resemblance to the method described above,
is GRAMMAR – Genome-wide Rapid Analysis using Mixed Models and Regres-
sionAulchenko et al. (2007a). Here, in step one, not only the mixed model is
estimated, but also the environmental residuals,

êi = yi − (µ̂ + Ĝi )

are estimated. These are free of correlations, and can be used as the new ”trait”
in further analysis. This procedue allows for very fast GWA analysis in step 2
(for large sample sizes, much faster then above-described two-step approach),
also it has a great flexibility as the environmental residuals can be analysed
using a large variaty of methods. At the same time, the GRAMMAR test for
association is conservative, and the estimates of the effects obtained in step 2
are downward biasedAulchenko et al. (2007a). While conservativity of the
test may be dealt with using ”reverse genomic control”Amin et al. (2007), the
issue of bias in effect estimates has no solution yet. The GRAMMAR model is
implemented in GenABEL software packageAulchenko et al. (2007b).
Strictly speaking, above described two-step tests are correct if the distribu-
tions of covariates in the first and the second parts of the model are independent
162CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

conditional on the estimated phenotypic variance-covariance matrix. This as-

sumption is most likely to be true when the covariates included in the base
model are environmental ones, and thus are not expected to exhibit conditional
correlation with SNPs. However, when endogenous risk factors, such as e.g.
body mass index, are included as the covariates in the base model, some SNPs
are expected to exhibit covariance with this covariate. In such situation the
two-step test is not strictly correct, but given relatively weak SNP-phenotype
correlations normally observed in GWAS should normally keep good statistical
properties in most situations. The properties of the test under such conditions
are still awaiting their description.
The methodology described here applies to study of association of quantita-
tive traits. For binary traits, no formal practical GWA solution is available yet.
The best current strategy to analyze binary traits in samples of relatives may
be to treat the binary outcome as if it was quantitative. As you have seen with
Armitage’s trend test, which is equaivalent to the score test for a quantitative
variable when oucome is coded with ”0” and ”1” (page 150), this approach leads
to correct inferences about significance.

6.3.3 Estimation of kinship matrix from genomic data

Another problem with application of mixed models to analysis of GWA data
is that pedigree is required to estimate the kinship matrix. The expectation
of kinship can be estimated from pedigree data using standard methods, for
example the kinship for two outbred sibs is 1/4, for grandchild-grandparent is
1/8, etc. For an outbred person, the kinship coefficient is 1/2 – that is two
gametes sampled from this person at random are IBD only if the same gamete
is sampled. However, in many situations, pedigree information may be absent,
incomplete, or not reliable. Moreover, the estimates obtained using pedigree
data reflect the expectation of kinship, while the true realization of kinship may
vary around this expectation. In presence of genomic data it may therefore be
desirable to estimate the kinship coefficient from these, and not from pedigree.
It can be demonstrated that unbiased and positive semi-definite estimator of the
kinship matrixAstle and Balding ((in press) can be obtained by computing
the kinship coefficients between individuals i and j with

L
1 X (gl,i − pl )(gl,j − pl )
fˆij = (6.19)
L pl (1 − pl )
l=1

where L is the number of loci, pl is the allelic frequency at l-th locus and gl,j
is the genotype of j-th person at the l-th locus, coded as 0, 1/2, and 1, cor-
responding to the homozygous, heterozygous, and other type of homozygous
[Link] et al. (2007); Astle and Balding ((in press); Price et al.
(2006) The frequency is computed for the allele which, when homozygous, cor-
responds to the genotype coded as ”1’.
Inetrestingly, use of the kinship matrix instead of pedigree kinship (when
available) may lead to higher power, especially when ”dense” pedigres and traits
with high heritability are considered (YSA, unpublished data). This is likely to
happen because genomic-based kinship is likely to reflect true genetic relations
better than (possibly not completely correct) pedigree expectations.
6.3. ANALYSIS OF STRUCTURED POPULATIONS 163

6.3.4 EIGENSTRAT and related methods

EIGENSTRATPrice et al. (2006) method combines the ideas underlying struc-
tured association and the genomic kinship matrix. Basically, it allows for struc-
tured association, with strata and more subtle structure identified through the
use of genomic data.
The genomic kinship matrix is computed as described earlier (equation 6.19).
This matrix reflects genetic similarities between study participants. A reverse
metric, distance matrix, is computed as 0.5 − f . Classical multi-dimentional
scaling (CMDS) is applied to identify a number (say, k) of first principal com-
ponents (PCs) of variation of the distance matrix. The idea behind CMDS is
mapping of the all pair-wise distances defined by the original matrix to some
k-dimentional space. Each study participant is presented as a dot in this space,
and these dots are constructed in such way, that the distance between these is
maximally close to the distances observed in the original distance matrix. Thus,
if two genetically distinct populations are present in a sample, the first principal
axe of variation (k = 1) will identify these. Generally, of n distinct populations
are present, these can be identified by k = n − 1 first axes.
To demonstrate this principle, results of CMDS of the genomic distance
matrix for the HapMap participants, and a number of people from Europe, are
presented at figure 6.6. On panel A, the results of CMDS of the distance matrix
to to the space of the first two PCs are present. Inter-continental differences
are obvious – there are three distinct clusters, corresponding to Yoruba (yellow-
brown), Asian (green and dark brown), and European (blue and red) study
participant. You can see that mapping to the first two principal components
perfectly distinguishes these three populations. The same methodology can be
used to distinguish between more subtly divided populations – at panel B of
figure 6.6 the mapping of European populations is present. The panel C presents
even further detalisation of genetic distances – to the level of single country.
Note that configuration of European cluster in panel A is not exactly the
same as in panel B. This happens because the distance matrix, computed based
on formula 6.19, depends on allelic frequencies, as estimated in the total sample.
Clearly, these are different, when all continents or Europeans only are used.
Consequtive analysis of sub-samples allows distinguishing more and more subtle
differences between populations.
Interestingly, the genetic distances co-incide very well with geographic dis-
tances, i.e. the populations close geographically are also close genetically. Thus,
most isolation between human populations is explained by geographic distance.
How these principal components may be incorporated in association analy-
sis? Different populations may have diferent mean values (or prevalences) of the
trait under the study, which, coupled with genetic differences, leads to confound-
ing and false positives in association study. A simple method would be to try to
account for these differences in the means by incorporating the principal com-
ponents as covariates into association analysisSladek et al. (2007). Another,
related method would be to perform linear adjustment of both the phenotype
and the genotype onto PCs (i.e. compute the residuals from linear regression of
genotypes/phenotypes onto PCs). Then simple score test for association (Tq2 ,
equation 6.11) can be used to analyze association. This comprises the origi-
nal EIGENSTRAT methodology suggested by Price and colleaguesPrice et al.
(2006).
164CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

Figure 6.6: The European genetic structure (based on 273,464 SNPs). Three
levels of structure as revealed by PC analysis are shown: A) inter-continental; B)
intra-continental; and C) inside a single country (Estonia), where median values
of the PC1&2 are shown. D) European map illustrating the origin of sample
and population size. CEU – Utah residents with ancestry from Northern and
Western Europe, CHB – Han Chinese from Beijing, JPT - Japanese from Tokyo,
and YRI – Yoruba from Ibadan, Nigeria. (reproduced from Nelis et al. (2009))
6.3. ANALYSIS OF STRUCTURED POPULATIONS 165

Different
Mixed
Structured Models
Association +
Structured
Population

Association
EIGENSTRAT,
PC-adjustment

Genomic Mixed
Control Models

Same

“Independent” First degree

(remote) Relatedness relatives

Figure 6.7: Applicability of different methods for association analysis.

A practically interesting question is how many PCs to include into analy-

sis? By default, EIGENSTRAT uses 10 PCs, but this is rather arbitrary num-
ber. In general, ”significant” PCs should be included into analysis. Patterson
and colleaguesPatterson et al. (2006) suggested including the leading PCs,
which significantly explain progressively more genetic variation. Alternatively,
it was argued that PCs significantly associated with the phenotype should be
includedNovembre and Stephens (2008), as inclusion of other PCs is not only
unnecessary, but may also reduce power.
Note that methods described here allow for differences in means between
the populations, but not for the difference in variances. It would be interesting
to address the question what effects violation of this assumption may have
on association studies, and develop extended methods allowing for difference
in variances. This would first probably require a study of inter-populatinal
variances.

6.3.5 Summary: what method to use?

What methods should be used for analysis of particular study? An simplified
overview is provided at figure 6.7.
Basically, if study participants come from the same genetic population, are
characterozed by low degree of kinship, and confounding, as measured with
genomic control (GC) inflation parameter λ is small, GC is enough to correct
for residual bias.
If study participants come from substantially different populations, and, for
each population, above conditions hold, structured association may be used.
166CHAPTER 6. GWA ANALYSIS IN PRESENCE OF STRATIFICATION: THEORY

If study participants are characterized by high degree of kinship, mixed-

model based methods should be used.
EIGENSTRAT and PC-based adjustment methods should be used in some-
what intermediate situation when differentiation between the populations and
kinship between study participants is not too high: while substantial population
differentiation may mean differences in variances, which EIGENSTRAT/PCBA
does not account for, it is also known that these methods do not perform very
well in pedigrees (YSA, Najaf Amin, unpublished data).
Clearly, a particular study may have its own specifics and should be consid-
ered separately. As difference between outlined scenarios is quantitative rather
than qualitative, there is no single recipe for analysis when genetic structure is
present in a study sample.

6.4 Links
Here are some useful links to software which can be used for analysis in struc-
tured populations:
EIGENSTRAT: [Link]

GenABEL: [Link]
MACH: [Link]
PLINK: [Link]
Chapter 7

GWA in presence of genetic

stratification: practice

Both ethnic admixture and presence of close relationships represents examples

of confounding in association analysis. However, the methods to correct for
stratification as resulting from mixture of subjects coming from different ge-
netic populations, and methods to correct for family relations may be slightly
different, and will be described separately in the next sections.

7.1 Analysis with ethnic admixture

In previous section we detected genetic stratification by analysis of genomic
kinship matrix and excluded genetic outliers from our further analysis. When
there are only a few such outliers, exclusion them from analysis is a good option.
However, in large studies cases and controls are usually selected across a number
of locations and genetic populations, and stratification is expected by design.
In such case, analysis of association should account for this stratification.
Let us do structured association analysis using the data1 data derived in
previous section.
If you are not running R yet, start R and load GenABEL-package library by
typing
> library(GenABEL)
and load the ’data1’ workspace generated in section 5 (”Genome-wide associa-
tion analysis”).
> load("[Link]")
> ls()

[1] "cl1" "data1" "[Link]" "[Link]"

First, let us check how much test statistic inflation is there if we ignore
stratification.
> [Link] <- qtscore(dm2,data1)
> lambda([Link])

167
168CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

$estimate
[1] 1.051744

$se
[1] 0.000756613

We now will consider several ways to account for stratification, namely, stru-
cutured association analysis, method of Price et al. (EIGENSTRAT), a similar
method based on adjusting for the principal components of variation of genomic
kinship matrix, and use of a mixed model.
One of the ways to do that is to perform structured association analysis. In
such analysis, effect and its variance are estimated within each strata separately,
and then these estimates are pooled to generate global statistics. The strata can
be known from design (e.g. place of birth or ethnicity of parents) or estimated
from GWA data.
To do structured association analysis we need to define a variable which will
tell what population the study subjects belong to. In previous section, we stored
the names of ’outlier’ subjects in variable cl1:

> cl1

[1] "id2097" "id6954" "id2136" "id858"

We can use function %in% to find out what names of subjects are in cl1:

> pop <- [Link](idnames(data1) %in% cl1)

Let us check how the ’population’ is distributed among the cases and con-
trols:

> table(pop,phdata(data1)$dm2)

pop 0 1
0 47 77
1 0 4

As we have seen before, one of the clusters contains only the cases.
Now, structured association may be done with qtscore function by specify-
ing strata argument:

> [Link] <- qtscore(dm2,data=data1,strata=pop)

> lambda([Link])

$estimate
[1] 1.03431

$se
[1] 0.0007059588

We can compare the original results, results of analysis excluding outliers, and
structured association analysis by
7.1. ANALYSIS WITH ETHNIC ADMIXTURE 169

− log10(P − value) qtscore(dm2, data1)

5

●
●
●
3

● ● ●
● ●
●
● ● ● ● ● ●●
●● ● ●
●● ● ●●●●● ●● ●
● ●
●
● ●
●
●● ● ● ●●● ●
●●● ● ● ● ● ●● ●● ● ● ● ●●●
●●● ●● ● ● ●
●●●
●● ●● ●● ●●
● ● ●●● ●●
● ● ●● ●● ●● ●
● ● ●● ●●●
● ●
● ●●●●●● ● ● ●●●
●
●●
● ●
●● ● ● ●
●●●●●
● ●
●●●● ● ● ● ● ●
● ●● ●●
●●●●●
● ●●● ●●
●● ● ●●
●●
● ●●
● ●● ●● ● ●●● ●● ●●● ●● ●● ● ● ● ●● ●
● ●●
●●●
●
● ●●●●●●●● ●
●
●●●● ●
●●●●●●●
●●●●
●●●
●●●
●●
●●
● ●●
●● ●
●●●
● ●●● ●
● ●
●
●●●●
●●
●●●●●
● ●
●●●●
● ●
●
●● ●● ●
●●●●●● ●●
●●
●●●●
●●
●●●●●
●●●●
●●
● ● ●
●●● ● ●●●
●●
●●●
●
● ●●●●● ●
●●●
●●●
●
●●
●● ●●●
●
●
●●●●
● ●●
●
1

●
●●
● ●●
● ●●
● ●●● ●
●●
● ●● ●
●●●
● ● ● ●● ●●●●
● ●
●● ●
●●● ●●● ●
●● ●●● ●● ● ● ●●● ●● ●
●●● ● ● ●● ● ● ● ● ●● ● ●●
●●●
● ●●
●●
●
● ●
●
●●
●
●●●
●●
●●
●
● ● ●●●●
●
●
●●●
●
●●
●
●●
●●●
●
●●
●●●
●●
●●●
●●●
●
●
●
● ●●●●
●
●●
●
●●●
●●
●●
●
●●
●●
●
●
●●
●●●
●●
●●●
●●
●●
●
●●
●
●
●
● ●● ●● ● ●
●●
●
●●●
●●
●
●●
●
●●●●
●
●●● ●
●●
●
●●
●●
●●
●●
●
●
●●
● ●
●●
●●
● ●●
●●
●
●●●
●
●●
● ● ●●●
●●
●
●●
●●
●
●●
●
● ●●
●●●
● ●
●●●●
●● ●
●●
●●
●
●●
●
●
●●
●
●
●
●●
●●●●
●●
●
●●
● ●
●●
●●
●●●●
●
●●●
●
● ●
●
●●
●
●●
●
●●●
●●
●
●
●●
●●●
●●●●● ●
●● ●●
●●●
●●
●
●
●●
●●
●●
● ●● ●
● ●
●●
●
●●
●
●
●●
●●
●
●
●●
●●
● ●
●
●
●
●
●●
●
●●●●●●
●
●
●●

1 2 3 X

Chromosome

qtscore(dm2, data2, binomial)

− log10(P − value)

●
●
3

●
● ●● ●
● ●
●
● ● ● ●● ●
●
● ●
●● ● ●●● ● ●
●● ●
● ● ● ●
● ● ● ●● ●
●
●● ●
● ●
●
●
●
●●
●
●●● ● ●
●
●●●●● ●● ● ● ● ●●● ●●●
● ● ● ● ●
● ● ● ● ●●●
● ● ●
● ●●●
●● ● ●
●● ●
●●● ●●●
●
● ●●●●●
● ● ●●
● ●● ● ● ●●
●
●● ●●●
●● ●● ●●● ●
● ●●●
●● ●●●●● ● ●●●●● ● ● ●●●●
●● ● ● ● ●
● ●● ●
●●●● ●●
● ●●●
●
●●
●●●
●
●
●
● ●● ●
●●●
●
●●
●●●
●●
●
●●
● ●
●
●●●
● ●●● ●●●
● ●
●
●●●●●
●
● ●
●●●
●●
● ●●
●●
●
●
●●
●●●● ●
●
●●●
●● ●●●
●
●
●
●
●●● ●
●
●●●
●
●
●● ●●● ●●●●●●● ●●●●● ●
●●●
●●●
●●
●
●
● ●●
●
● ●
● ●
● ●● ●
●● ●
●● ●● ●●
●
●●●●
●● ●●●●
●
● ●
●●
●●
● ●
●●●●●
1

● ● ● ● ● ● ● ● ● ●●
●
●●●●●
●
● ●●
●
●●
●●
●
●
●
●
●
●
●
●●
●●
●● ●●
●
●●
● ●●
●
●
●●●●
●●●
●
●●
●
●
●
●●●
●
●
●●●
●
●●
●
● ●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●●
●
●●●
●●
●●
●
●
●
●
●
●●
●●●●
●
●
●
●
●●
●
●●
●
●
●●
●●●●●
●●
●
●
●
●●●
●
●
●
● ●
●
●●
●
●
● ●
●●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
● ●
●
●●
●
●●
●
●●
●●●
●●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●● ●●
●●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●●●
●
●
●
●
●●
●
●
●●
●
●●
● ●
●●
●
●
●●●● ●●●
●
●
●
●●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●
●
●●
●
●
●
●●
●
●●●
●●●
●
●
●●
●
●●●
●
●
●
●
●
●●
●
●●
●●
●
●
● ●
●
●
●●●
●
● ●
●●
●● ●
●●
●
●

1 2 3 X

Chromosome

qtscore(dm2, data1, pop)

− log10(P − value)

●
●
3

●
● ●● ●
● ●
●
● ● ● ●● ●
●
● ●
●● ● ●●● ● ●
●● ●
● ● ● ●
● ● ● ●● ●
●
●● ●
● ●
●
●
●
●
●●
●●● ● ●
●
●●●●● ●● ● ● ● ●● ● ●●●
● ● ● ● ●
● ● ● ● ● ●●● ● ●
● ●●●
●● ● ●
●●●●
●●● ●●●
●
● ●●●●●
● ● ●●
● ●● ● ● ●●
●
●● ●●●
●● ●● ●●●● ●
●●
●●●●●●●● ● ●●●●● ● ●
●●●●●●●●●●●● ●● ● ●
●●●● ●●
● ●●●
●
●●
●●●
●
●
●
● ●● ●
●●●
●●
●●●
●●
●
●
● ●
●
●●●
●
● ●●● ●●●
● ●
●
●●●●●
●
● ●
● ●●
●● ●● ●
●
●●●
●●●● ●
●
●●●
●● ●●●
●
●
●
●
●●● ●
●
●●●
●
●
●● ●●● ●●●●●●● ●● ●● ●●
● ●●●
●
●
● ●
●●
●●● ●
● ●
● ●● ●●●
●● ● ●●
● ●●
●● ●●●●
●
● ●
●●
●●
● ●
●●●●●
1

● ● ● ● ● ● ● ● ● ●●
●
●●●●●
●
● ●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●●
●● ●●
●
●●
● ●●
●
●
●●●●
●●●
●
●●
●
●
●
●●●
●
●
●●●
●
●●
●
● ●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●●
●
●●
●●●
●●
●
●
●
●
●
●
●●
●●●●
●
●
●
●
●●
●
●●
●
●
●●
●●●●●
●●
●
●
●
●●●
●
●
●
● ●
●
●●
●
●
● ●
●●
●
●
●
●
●
●
●
●●●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●●●
●
●
●
●
●
●●●
●
●●
●
●●
●
●
●
●● ●
●
●
●●
●
●
●●
●●
●
●●●
●●
●
●
●●●●●
●
● ●●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●●●
●
●
●
●●
●
●
●
●●
●
●●
●●
●●●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●●
●
●●
●●
●
●
● ●
●
●
●●●
●
● ●
●●
●●●●
●●
●
●

1 2 3 X

Chromosome

Figure 7.1: Comparison of the original results, results of analysis excluding

outliers, and structured association analysis.

> par(mfcol=c(3,1))
> plot([Link],ylim=c(1,6))
> plot([Link],ylim=c(1,6))
> plot([Link],ylim=c(1,6))
> par(mfcol=c(1,1))
The resulting plot is presented at figure 7.1. In this case, there is little difference,
because all people belonging to the smaller sub-population are cases.
Other way to adjust for genetic (sub)structure is to apply the method of
Price et al. (EIGENSTRAT ), which make use of principal components of the
genomic kinship matrix to adjust both phenotypes and genotypes for possible
stratification. In GenABEL-package, such analysis is done using egscore func-
tion.
> [Link] <- egscore(dm2,data=data1,kin=[Link])
> lambda([Link])

$estimate
[1] 1.102747

$se
[1] 0.001043038

The analysis plot may be added to the previous one by

170CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

> par(mfcol=c(3,1))
> plot([Link],ylim=c(1,6))

Now let us apply adjustment for the stratification by use of the principal
components of genetic variation. For that we first need to extract the principal
components of genetic variation by constructing the distance matrix

> dst <- [Link]([Link])

and performing the classical multidimensional scaling

> pcs <- cmdscale(dst,k=10)

> pcs[1:5,]

[,1] [,2] [,3] [,4] [,5]

id199 0.030289769 0.057952824 0.10342206 -0.020496319 -0.02678162
id300 0.031798583 0.004789923 0.02947431 0.033227508 0.05977593
id403 0.060654557 0.079249013 -0.06111279 -0.081417523 0.03294180
id415 -0.004913697 -0.030714900 0.07807686 0.017468144 0.03078079
id666 -0.012639670 0.016176179 -0.08784632 -0.002322116 0.02957059
[,6] [,7] [,8] [,9] [,10]
id199 0.08021104 0.01144582 -0.020228061 -0.001507913 -0.060483058
id300 0.03702791 0.02887831 0.005142389 -0.005340720 0.094991803
id403 -0.01481834 -0.01259170 0.040037267 -0.048689378 0.007304146
id415 -0.02472941 0.07900292 0.075300986 0.028609310 0.021747288
id666 -0.04742543 -0.11594090 0.106539735 0.026380262 -0.022472669

Now we can use these PCs for adjustment:

> [Link] <- qtscore(dm2~pcs[,1]+pcs[,2]+pcs[,3],data1)

> lambda([Link])

$estimate
[1] 1.004946

$se
[1] 0.0007356663

> plot([Link],ylim=c(1,6))

Finally, let us use the full genomic kinship matrix for the adjustemnt for
pupulational structure. First, let us estimate the polygenic model with

> h2a <- polygenic(dm2,data1,kin=[Link])

The resulting ’heritability’ estimate is

> h2a$esth2

[1] 0.2547813

Now we can perform mixed model approximation analysis using mmscore func-
tion
7.1. ANALYSIS WITH ETHNIC ADMIXTURE 171

− log10(P − value) egscore(dm2, data1, [Link])

●
● ●
3
● ● ●
● ●
●
● ● ● ●● ●
●
● ● ● ●● ● ● ● ●●●
●● ●●
● ●● ● ● ●● ●●
● ●
2

●
●
●●● ●● ● ●● ● ●● ●● ● ●● ●●●●● ● ● ●● ● ● ● ●●
●
●● ● ●●
●
●
● ● ●●●● ●● ● ●
● ●● ● ● ● ●
● ●●●
● ● ●● ● ● ●● ● ●● ●● ● ●●
●
●
●● ●
●●●●
●
●●
●
●●
●
●
● ●●
●
●●
● ●● ●
●●● ●●
● ●●●●●
●
●●●
●● ● ● ●
●●
●●● ●
●
●
●
●●
●
● ● ●●
● ● ●
●●●
●
● ●●
● ●● ● ●●●
●●
● ● ● ●
●●
●● ● ● ● ●●● ●● ●
●
●
●
●●
● ●●●
●● ● ●
● ●
●
●
●
●● ● ●● ●● ●
●
●
●●●
●
●●
●● ●●● ● ●●
●●
● ● ●●●
● ● ●●
● ●●● ●●
● ● ● ●● ●
●●●●
●
● ●● ●
●
● ● ●● ●●●●●●
●
● ● ●●●●● ● ●● ●
● ● ●● ●●●● ●● ●●
● ●● ●● ● ●●● ●
●● ●●●●●
● ● ● ●●●●●● ●●● ●●●● ● ● ● ●
● ●
●●●●
●●● ●●
●●●●
● ● ● ●● ●● ● ● ●● ● ●● ●● ●●
● ●
●●● ●●●●●
● ● ●
● ●●
1

●●● ●● ●
●●
●●●●●●●●● ●● ●●
●● ●●●
●● ●
●●●● ●
●●●
● ●
●●●●●●●●
● ●● ●
●● ●● ●
●● ●●
●●●● ●●
●●● ●
●●● ●●
● ●
●
●●
●●● ●●
●● ●
●●●● ●● ●●
●● ●
●●● ● ●●
●● ● ●
● ●●
● ●●●
● ● ●
●●
●●
●
● ●● ● ● ●●●●●● ●● ●● ●
●● ●
●● ●●
● ● ●
● ● ●
● ●
●●●
●● ● ●● ●
●● ●●
● ●● ●●
● ● ● ● ●● ● ● ●● ●● ● ●●● ●●●
● ● ● ●
●●
● ●
●●
●●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●●●
●
●●
●
●
●●
●
●
●●
●
●●●●
●●●●●
●●
● ●
●●
●●●●●
●
●●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●●
●
●
●
●
●●●
●
●●
●●
●●
●
● ●●
●
●●
●
●
●
●
●
●●
●
● ●●
●
●
●
●
●●
● ●
●●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●● ●
●
●
●
●●●
●
●●
●●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●
●●●
●●
●
●●
●
●●
●
●●
● ●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●●
●● ●
●●
●●●
●
●●●
●
●●●
●●
●
●
●●
●● ●
●
●● ●
●●
●●
●
●
●●●●
●
●
●●●
●
●
●
●
●
●
●●●
● ●●●
●
●●
●
●
●●
●
●●
●●
●
●
●●●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●●
●
● ●
●●
●
●
●
●
●●
● ●
●
●●
●●
●
●●
●
●●
●
●
●
●●●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●
●●
●●
●●
●
●●
●●
●●●
●
●●
●
●
●
●
●
●
●●
●●
●●
●●
●●
●
●●●
●●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●●●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●●●
●
●
●●●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
● ●
●
●
●●
●
●
●
●●
● ●
●
●●
●
●
●●
●
●●
●●●
●●
●
●●
●
●●
●●
●● ●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●●
●
●●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●●
●●
●●
●
●
● ●●
●●
●●●
●
●●
●●
●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●●
●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●●●
●
●●●
●●●
●
●
●●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●● ●
●●●●
●
●●
● ●●●
●●
●●
●●
●●
●●● ●●
●
● ●
●●●
●●
●●
●●
●●●
● ●●●
●●
●●●●●●●●
●●
● ●
●
● ●●●●
●●●●●●●
●●
●
●
●●●
●● ●
●
●●●
●●
●●
●
●●●
●
●●●●●
●●
● ●●
●
● ●
●●●
●●
●●●●●●●●●●
● ●●
●
●●●
●●●●
●●
●●
●●●
●●
●
● ●●
●
● ●●
●●
●●
●●
●●● ●
●●
●●●●●●●●
●● ●
●
●
●●
●●●●●
● ●
● ●●
●●●●●
●●
●● ●●
●●●
●
● ●●●●
●
●
●●●
●
●
●
●●●
●●
●●●
●
●
●●
● ●
●
●●
●●
●
●
●●
●●
●●●
●
●
●●
●
●
●●●●
● ●●●
●
●●
●
●●
●
●
●●●
●
●●
●
●●●●
● ●●●
●
●●●
●
●
●
●●
●
●
●●
●●
●●●
● ●
●
●●
●●
●
●●
●
●
●●
●
●●
●●
●●
●●●●
●
●●
●●
●●●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●
●●●●
●●
●●
●●
●
●● ●●●●
● ●
●
●●
●●●●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
● ●●
●●●
●●
●
●●
●●
●
●
●●
●●
●●
●●
●
●●●●
●●
●●
●
●●
●●
●
●●●
●●
●
●●
●●
●
●●
●●●
●●
●●
●●
●
●
●●
●
●
●● ●
●
●
●●
●●
●●
●●●
●
●
●
●●●
● ●●
●●
●
●●
●●
●
●
●
●●●●●
● ●
●●●●
●●
●
● ●
●●
●●
●
●
●●
●
● ●
●
●●
●
●●
●
●●
●●●
●
●●
●●
●●
● ●
●
●
0

1 2 3 X

Chromosome

qtscore(dm2 pcs + pcs + pcs, data1)

− log10(P − value)

●
3

● ● ● ● ● ●
● ● ● ●● ●
● ● ● ●
● ●● ● ● ● ●● ●●● ● ●
● ●● ●
2

● ● ● ●● ● ● ●
●●
● ● ●
● ●●● ●●●●●● ●● ● ●●●● ● ●● ● ● ● ● ● ●●● ● ● ● ● ●●●
● ●
● ● ●● ●● ● ●●●●
● ● ● ●● ●●
● ●●
● ●●●●
● ● ● ●● ● ● ●● ●●
●
●●
●
●●
● ● ●
●●●●
● ●●●
● ● ● ●● ●
● ●●
●
●●
● ●●●
●● ● ●●
●● ●●
●
●
●●
●
● ● ● ●
● ●● ●
●●
●●
●
● ● ●●●● ● ●
●● ● ●●
● ●●
●● ●
●
●●●
● ●●● ● ●● ●●●●
●
●●
●●● ●●●
●● ●● ●●● ●●●
●
●
●
●
●●● ● ●
●
●●●●
● ●● ●●●●●●
● ● ●●
● ● ●
●● ●●●● ●●●● ● ●● ●●
●● ●●●●●●●●● ●
● ●● ●
● ●
●●● ● ● ● ● ● ●●
● ●●●●●
●● ● ●● ●● ●●●●●● ●
● ●● ●● ●
●● ●● ● ● ● ●● ● ● ●
1

●● ●●●
● ● ●●●●● ● ●●
●●● ●● ●●● ● ●● ● ● ●●●
●●
● ●
●●● ●
●●●● ●●● ● ●● ●
● ● ● ●●●● ●● ● ●
● ● ●● ●●●●● ● ●●
●●
●●●
● ●●
● ●●●
●
●●
●●●●●
● ●● ●● ● ●
●
●
●
●
●●●●
●●●
●
●●●
●●
●
●●
●
●
●●
● ●
●
● ● ●
●●●
●
●
●
●
●●●●●
●●
●●
● ●
●●●
●
●
●
● ●●
●
●
●
●●●●
●
●●●●
●●
●●
●
●●
●●●
●
●
●
●●
●
●●● ●●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●●●
● ●
●●●
●
●
●
●●
●●
●
●
● ●●
●● ●
●
●●
●
●●
● ●●
●● ●
●●●●● ● ●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●●●● ●●
●●
●●●●
●●●●
●
●
●●
●●
●
●●
●● ●●
●
●
●●
●
●
●●
●●●
●●
●
●
●
●
●
●●●●
●
●●
●●●
●●
●●
●
●●●
●●
●
●●
●●
●●●
●●
●
●●●●
●●
●●
●
●●●●
●●
●●
●
●●
● ●
●●
●●
●●
●●●
●
●
●●●
●
●● ●
●●
●●●●
●●●
● ●
●●
●●●
●●
●●
●
●●●
●●
●
●●
● ●●
●●●●
●
●● ●
●●
●
●●●
●●
●
●
●● ●●●
●
●●
●●●
●
●●●
●●
●●●
● ● ●
●
●
●●
●
●●
●●●
●●
●●●●●
●●
●●
●
●●
● ●
●●●
●●●●
● ●●● ●
● ●
●●
●●
●
●●●
●●
●●
●
●
●●
●
● ● ●
●
●
● ●
●
●●
●●
●
●●●
●
●●
●
●
●
●●
●●●●
●●
●●●
● ●●●
●
●
●●
●●●
● ●
●●
●●
●
●
●●●●●
●●
●
●●
●
●●
●
●●
●
●●●●
●
●●
●●
●
●
●●●
●
●●
●●
●●●
●
●●
●●
●
●●●
●
●
●●●
●
●●●
●●●●
●
●
●
●
●●
●●
●
● ●
●
●●
●●
●
●
●●
●
●
●●●
●
●
●●
●●●●●
●
●●●
●
●
●●
●●
●●
●●
●
●●
●●●
●
●
●●
●●
●●
●
●
●●
●
● ●
●
●●
●●
●●●●●
●●
●●●
●● ●●●
●●
●●
●
●●
●●
●●
●
●
●●
●●
●
●●
●●
●●
●●
●
●●
●●
●●
●
●●
●●
●●●●
● ●
●●●●
●
●
●
●●
●●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●
●
● ●
●
●
●
●●●●
●●●
●●
●
● ●
●
●●●
●
●
●●
●
●●
●
●
●●
●
●●
●●
●●
●●
●●
● ●
●
●●
●●
●
●
●●
●
●●
●
●●●
●●
●
●●
●
●
●●●
● ●●
●●
●
●● ●
●
●●
●
●
●●●
●
●●
●●
●
●
●●●
●
●●
●
●●
●
●
●●
●
●●●●
●
●●
●
●
●●
●●
●
●●
●●
●●
●
●●
●
●●
●●●
●
●
●
●●
●
●●
●●
●●
● ●
●●●
●
●
●●●
●
●●●
●
●
●
●●
●
●●●
●
●●●
●
●●
●●
●
●
●●
●
●
●●
●●
●●
●●
●
●●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●
● ●●●
●
●●
●
●●●
●
●●●
●
●●
●
●
●●
●
●
●●
●●
● ●
●
●
●●
●
●●
●●
●
●●
●
● ●
●
●●
●●
●
●
●●
●
●●
●● ●●
●●
●
●
● ●
●
●
●●
●
●●
●●
●
●
●●●
●●
●●●
●
●●
●●
●●
●
●●
●
●●
●
●●
●
●●
●●
●
●●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●●
●●
●
●
●●
●●●
●
●
●●
●
●
●●
●●●
●
●
●●
●
●
●● ●
●
●
●●
●
●
●
●●
●●●
●
●
●●
●
●●
●●
●
●
●●●
●●
●
●
●●
●●
●
●
●●
●●●
●
●
●●
●●
●●
●●
●●
●
●
●
●
●●
●● ●
●
●●
●
●
●●
●
●
●●
●
● ●
●
●●
●●
●
●●
●●
●
●
●●●
●●
●●
●
●
●
●●
●
●
●
●●●
●●
●
●●●
● ●●
●●
● ●
●●●
●●●
●
●
●●●
●
●
●●
●
●●
●
●●●●●
●●
●
●
●
●●
●
●●●
●
●
●
●●
●●
● ●●
●● ●●
●
●
●●
●●
● ●●
●● ●
●
●●
●●
●
●●
●
●●
●●
● ●
●●●
●●●
●●
●●●
●
●●●
●
●●
●
●●●
●●
●
●●
●
●
●●
●
●●●
●●●●
●●●●
●●
●●
●●
● ● ●●
●●
●●●
●
●●
●●
●●
●
●●
●
●●
●●
●●●●
●●
●
●●
●
●●
● ●
●●●
●
●●●●
●
● ●●●
● ●
●●
●
●
●●
●●
●●
●●
●●
●●
● ●●
●
●●
●●
●
●●●
●● ●
●
●
●●●●
●
●●
●●
●●●●
●
●
● ●
●●
●●●
●●●
●
●●
●●●●●●
● ●
●●●●
●●●
●
●●
●
● ●
●●
●
●
●
●●
●●
●●●●
●●
●●
●
●
●●
●
●
0

1 2 3 X

Chromosome

mmscore(h2a, data1)
− log10(P − value)

●
3

● ● ●
●
●●●
● ● ●
● ● ●
●
● ●● ●● ● ● ●
2

● ● ● ●
●● ●● ● ●
● ●●● ● ●● ●●●● ● ●●● ●
●●● ● ● ● ● ● ●●● ●● ● ●
● ●
●● ● ● ● ●
● ●
●
● ●●●● ●●●●● ● ● ● ● ●●● ● ●●
●● ●● ● ● ●● ● ● ● ●●●●● ● ●● ● ●
●● ●●
●
●
●●●
●● ●
●
●
● ●●● ●●●
● ●●
●
● ●● ●●●●
● ●●●●
●●●
●● ●
● ●●●● ●●●●●● ● ●
●●●●
●
●●●● ●● ●
●●
●●●●●
●●● ●
●
●
●● ●● ● ●
●●●●● ● ● ●●●●
● ●● ●● ●● ●
●●
●
● ● ●
● ● ●
●●
●
● ●● ●
● ●●
● ●●●●
● ●● ● ●● ●● ● ● ● ●
● ●● ● ● ●●● ●● ● ●●●● ● ●●●●
●●●●● ● ●●● ●●●●●
● ● ●●
●● ● ●●
● ● ● ● ●● ●●● ● ● ●●
1

● ●
● ●
●● ●
●●●●
●● ●●● ● ● ●●
● ●
●● ●● ● ● ● ●●●●●
●● ●●● ● ●● ● ●● ● ● ● ●●●
● ●
● ●●
●
● ●
● ● ●
●● ●
●
● ●● ● ●●● ● ●
●●●●
●
●●
●●●
●
●
●
●
●●
● ●●
●●
●●
●●●
●
●
●
●
●
●
●
●●●
●
●●
●●●●●●
●
●
●●
●
●
●●●●
●
●
●●
●●
●
●
●
●
●
●●●
●
●
●
●●●●
●
●
●
●
●
●
●●
●
●●●●
●●●
●●●
●●●
●●
●●
●
●●
●●●
●
●●●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●●●
●●●
●
●
●
●● ●●
●
●
●
●
●
●●●
●
● ●
●
●
●●●●
●
● ●
●●●
●●
●
●
●
●
●
●
●
●
●
●●●●
●● ●●●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●●
●●
●
●●
●●
●
●
●
●●
●●●
● ●●●●● ●●●
●
●●
●
●
●
●
●
●
●●
●●
●●●
●
●●●
●
●
●●● ●●●●
●
●●●●●
●
●● ●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●●
●●
●
●●
●
● ●●
●
●● ●●
●
●
●●
●●
●●●
●
●●
●
● ●●
●●
●
●
●●
●●
●●
●●●
●●● ●
●
●
●●
●●●
●
●●
●●
●●
●●
●
●
●●●●
●
●●
● ●
●
●
●●
●●●●
●
●●●●
● ●
●●
● ●
●
●●
●
●●●
●
●●●
●●
●●●
●●
●●
●●●●
●
●●
●●
●
● ●
●
●
● ●●
●●●●
●●
●●
●
●●●●
●●
●
●●
●
●●
●
●●●
●
●●●
●● ●
●●
●● ●
● ●
●
●●
●
●●●●●
●●●●
●
●●
●●●
● ● ●
●● ●●
● ●
● ●●●
●
●
●●
●●
●●
●●
●●● ●●●●
●●●
●
● ●
●
●●
●●
●
●●
●
●
●●
●●●●
●●●
●
●● ●●
●●
●
●●
●●●●
●●
●●●
●
●
●●
●●●
●
●●
●●
●
●
●
●●
●
●
●●●●
● ●
●
●●
●
●
●●●
●●
●●●
●●●
●
●●
●●
●●
●●
●●
●
●
●
●●
●
●●●
●●
●●
●●●
●
●●
●
●●
●
●●
●
●●
●
●●●●
●
●
●●●
●●●
● ●
●
●●
●●●
●
●●
●
●●
●
●●
●●
●●
●
●
●●
●
●
●●
●●
●●
●
●●
●● ●
●
●●
●
● ●
●
●●●
●●
●●
●
●● ●●
●●
●
●
●
●●●
● ●
●●
●
●
●●
●●
●
●●
●●
●
●●
●●●
●
●●
●●
●
●
●●●●
●●●
● ●●●
● ●
● ●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●●
● ●
●●
●●●
●● ●
●
●
●●
● ●
●
●
●
●●●
●●
●
●
●●●
●●
●
●●●
●●●●
●●
●
● ●●
●●
●●
●●
●●
●
●●
●●
● ●
●
●
●●
●●
●
●● ●●
●
●
●●
●
●
● ●
●
●
●●●
●
●●●
●
●
●●
●
●
●●
●
●
●●●●
●
●
●●
●●
●●
●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●●●●
●
●●●
●●
●
●
●●
●●
●
●
●●●
●
●●
●
●
●●
●
●●●
●
●●
●●●●
●
●
●●●
●
●●
●●●
●
● ●●
●
●
●●
●●
●●
●
●●●
●●
●
●●
●
●●
● ●
●
●●
●●
●
●
●
●
●●●
●
●●
●
●● ●
●
●
●●
●
●●
●
●
●●
●●
●
●●●
●●
●
●
●●
●
●●
●
●● ●●
●
●● ●
●●
●
●●
●●
●
●
●●
●●
●●
●●
●●●
●
●
●●
●●●●●
●
● ●
●
●●
●●
●
●●
●●
●●
●●
●
●
● ●
●
●●
●
●●
●●
●
●●
●●
●●
●
●●
●●
●
●●
●●●
●●
●
●
●●
●● ●
●
●●
●●
●
●●
●
●●
●
●
●●
●●●
●●
● ●●
●
●●
●
●
●
●●
●●
●
●
●●
●
●
●●●
●●
●●
●
●●●
●
●●●
●●
●
●●
●●
● ●
●●●●●
●
●
● ●
● ●
●●
●
●●
●
●●
●●
●●●●
●●
● ●
●●
●●●
●
●●●●
●
●
●●
● ●●
●
●●●
●
●●●●●●●
●
●●●●●●●
●
●
●●●
●●●
●
●●
●
●●
●
●
● ●
●
●●●●
●●●
●●●
● ●
●●●
● ●
●●
● ●
●
●●●
●●
●
●
●
● ●
●●
●●
● ●●
●●
●●●●●●
● ●
●
●
●●
●●
●
●
●●
●
●●
●●
●
●
●●
●●●
●
●●● ●
●●
●● ●
●
●
●●●
●
●●
●
●
● ●
●●●●●
●
●
●●
●●
●●●●
●● ●
●
●●●●●●
● ●
●●●
●
●●●
●●●
●
●●
●●
●
●
● ●●
●●
●●●●●●
●
●●
● ●●●●●●
●●
●●● ●●
●●
●●
●●
●●●●
●
●● ●
●
●● ●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●●
●●
●
●
●●●
●
●●
●
●
●●
●●
●
●●●
●●●
●●
●
●●
●●
●●
●●●
● ●
●
●●
●●
●
●●
●
● ●
●
●●
●●
●
●
●●
●
●●●●
●●
● ●
●
●
●●●
●
●
●
●●
●
●
0

1 2 3 X

Chromosome

Figure 7.2: Comparison of method of Price et al., principal components adjust-

ment and mixed modelling.

> [Link] <- mmscore(h2a,data1)

> lambda([Link])

$estimate
[1] 1

$se
[1] NA

> plot([Link],ylim=c(1,6))
> par(mfcol=c(1,1))

The resulting plot is presented at figure 7.2.

Again, the difference between three analysis methods is marginal because
there are no highly differentiated SNPs in the data set, and one sub-population
is presented by cases only. Still, the signals at chromosome one and three slightly
improved, while these at two and X went down.
Load and analyse the data set presented in file [Link]. GWA
data presented in this file concern a study containing data from several pop-
ulations. All these populations originate from the same base population some
generations ago. Some of these populations mantained large size and some were
small. There was little (2.5%) migration between populations.
172CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

Two traits (’quat’ and ’bint’) ara available for analysis. Investigate relations
between phenotypes and covariates. Perform association analysis. Answer the
questions
DATAPROBLEM: ok, here is a problem; replacing the ’strdat’
with data set available within GenABEL. Hence comments to the
analysis below are not making sense anymore
Ex. 1 — What covariates are significantly associated with the traits?

Ex. 2 — How many SNPs and IDs are presented in the data set?

Ex. 3 — How many SNPs and IDs pass the quality control (use SNP and ID
call rate of 0.98)?

Ex. 4 — Is there evidence for stratification coming from the distribution of

GW test for HWE (what is λ?)

Ex. 5 — Is there evidence that the test statistics for trait quat is inflated
(what is λ?)

Ex. 6 — Is there evidence that the test statistics for trait bint is inflated
(what is λ?)

Ex. 7 — How many genetically distinct populations are present in the data
set? How many people belong to each population?

Ex. 8 — Is the case/control and quantitative trait disbalance between popu-

lations?
Ex. 9 — For the quantitative trait, what method corrects best for stratifica-
tion (in terms of minimal residual inflation)?

Ex. 10 — What is the strongest SNP associated with trait quat? What model
(method and covariates used) gives best results? Is the finding GW-significant?

Ex. 11 — What is the strongest SNP associated with trait bint? What model
(method and covariates used) gives best results? Is the finding GW-significant?

7.2 Analysis of family data

In this section we will consider analysis of quantitative traits in a family-based
cohort, where participants were not selected for the value of the trait under
analysis. Such data may be generated in any study selecting participants based
on kinship (e.g. collections of sibships, nuclear or extended families); also any
study in a genetically isolated population is likely to end up with a large propor-
tion of relatively closely related individuals, even if ascertainment was random
with the respect to kinship.
In pedigree-based association analysis the pedigree works as a confounder –
exactly in the same manner as ethnic origin may work in a population-based
study. Any genetic polymorphism is inherited through genealogy, and therefore
genotypes are more similar between close relatives. In the same manner, any
other heritable trait will be also more similar between relatives, and therefore
certain degree of association is expected between any genetic marker and any
7.2. ANALYSIS OF FAMILY DATA 173

heritable trait in a family-based sample. If additive 1 d.f. test for association

is considered, the effect of confounding by pedigree can be shown to inflate the
resulting null distribution of presumably χ21 test statistics by a certain constant
λ.
As you remember, this is exactly what happens when simple test for as-
sociation is applied to a population-based data with ethnic admixture. In a
population-based study with strong admixture (both in terms of the proportion
and ethnic ”distance”), some genomic regions may have been differentially se-
lected in different populations. In such situation, use of genomic control does
not prevent false-positive association between a trait and these regions, and
other methods, such as EIGENSTRAT or Structured Association, are to be
used.
For pedigree-based data coming from (relatively) genetically homogeneous
population it can be shown that λ is a function of trait’s heritability and pedigree
structure, expressed as kinship matrix. Thus, genomic control is a simple and
valid method to study association in genetically homogeneous families. However,
this method reduces (or summarises if you prefer) all the abundant information
about heritability and relationship into a single parameter λ, therefore it is not
the most powerful method.
In quantitative genetics, a mixed polygenic model of inheritance may be
considered as ”industrial standard” – this model has sound theoretical bases
and is proven by time to describe well inheritance of complex quantitative traits.
This model describes the vector of observed quantitative traits as

Y =µ+G+e (7.1)

where µ is the intercept, G is contribution from polygene, and e is random

residual.
It is assumed that for each individual its ”personal” random residual ei
is distributed as Normal with mean zero and variance σe2 . As these residuals
are independent between pedigree members, the joint distribution of residuals
in the pedigree can be modelled using multivariate normal distribution with
variance-covariance matrix proportional to the identity matrix I (this is a matrix
with diagonal elements equal to 1, and off-diagonal elements equal to zero):
e ∼ M V N (0, Iσe2 ).
The polygenic component G describes the contribution from multiple inde-
pendently segregating genes all having a small additive effect onto the trait (in-
finitesimal model). For a person for whom parents are not known, it is assumed
2
that Gi is distributed as Normal with mean zero and variance σG . Assuming
model of infinitely large number of genes, it can be shown that given polygenic
values for parents, the distribution of polygene in offspring follows Normal dis-
2
tribution with mean (Gm + Gf )/2 and variance σG /2, where Gm is maternal
and Gf is paternal polygenic values. From this, it can be shown that jointly the
distribution of polygenic component in a pedigree can be described as multi-
variate normal with variance-covariance matrix proportional to the relationship
2
matrix Φ: G ∼ M V N (0, ΦσG ).
Thus the log-likelihood for this model can be written as a function of three
174CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

parameters:

2 1
L(µ, σG , σe2 ) = − 2
· loge (Φ · σG + I · σe2 )
2
2
+ (Y − µ)T · (Φ · σG + I · σe2 )−1 · (Y − µ) (7.2)
2
where µ is intercept, σG is the proportion of variance explained by the polygenic
2
component, and σe is the residual variance.
Covariates such as sex, age, or a genetic marker studied for association can
be easily included into the model:
X
Y =µ+ βj · Cj + G + e
j

Here, Cj is the vector containing j-th covariate and betaj is the coefficient
of regression of Y onto that covariate.
This mixed morel leads to likleihood

2 1
L(µ, σG , σe2 , β1 , β2 , ...) = − 2
· loge (Φ · σG + I · σe2 )
2
 T
X
2

−1
+ Y − (µ + β j · C j )  · Φ · σG + I · σe2
j
 
X
· Y − (µ + βj · Cj )  (7.3)
j

This general formulation can be easily adopted to test genetic association;

for example, an effect of a SNP can be incorporated into regression model

Y = µ + βg · g + G + e

where g is the vector containing genotypic values. In this mode, you can specify
a variety of 1 d.f. models by different coding of the vector g. For example, if you
consider an ”AG” polymorphism and want to estimate and test additive effect
of the allele ”G”, you should code ”AA” as 0 (zero), ”AG” as 1 and ”GG” as
2. Under this coding, the βg will estimate additive contribution from the ”G”
allele. If you are willing to consider dominant model for G, you should code
”AA” and ”AG” as 0 and ”GG” as 1. Recessive and over-dominant models can
be specified in a similar manner. If, however, you want to estimate general 2
d.f. model, the specification should be different:

Y = µ + βa · g + βd · Ig=2 + G + e

where g is coded as 0, 1 or 2, exactly the same as in the additive model, and Ig=2
is the binary indicator which takes value of one when g is equal to 2 and zero
otherwise. In this model, βa will estimate the additive and βd – the dominance
effect. There may be other, alternative coding(s) allowing for essentially the
same model, for example

Y = µ + β1 · Ig=1 + β2 · Ig=2 + G + e
7.2. ANALYSIS OF FAMILY DATA 175

would estimate trait’s deviation in these with g = 1 (β1 ) and these with g = 2
(β2 ) from the reference (g = 0).
The classical way to estimate mixed polygenic model and test for significance
is Maximum Likelihood (ML) or Restricted ML (REML) using equation (7.3).
However, when large pedigrees are analysed, ML/REML solution may take pro-
hibitively long time, i.e. from minutes to hours for single SNP analysis, making
study of hundreds of thousand of SNPs impossible. Therefore fast approximate
tests were developed for the purposes of GWA association analysis in samples
of relatives.
Here we will cover two of fast approximations available, FAmily-based Score
Test for Association (FASTA, Chen & Abecasis, 2007) and Genome-wide Rapid
Analysis using Mixed Models And Score test (GRAMMAS, Amin et al, 2007).
Both tests are based on the classical polygenic mixed model and are performed
in two steps.
First, polygenic model as specified by equation (7.1) and likelihood (7.2) is
estimated using available data.
Secondly, the maximum likelihood estimates (MLEs) of the intercept, µ̂,
2
proportion of variance explained by the polygenic component, σ̂G , and residual
2
variance, σ̂e , are used to compute the FASTA test statistics

2

2
(g − E[g])T · (Φ · σ̂G + I · σ̂e2 )−1 · (Y − µ̂)
TF2 = 2 + I · σ̂ 2 )−1 · (g − E[g])
(g − E[g])T · (Φ · σ̂G e

It can be shown that TF2 follows χ21 when pedigree structure is 100% complete
and 100% correct. As this is never actually the case, application of GC to correct
for residual inflation is recommended.
FASTA test results in unbiased estimates of the SNP effect and correct
P − values. Please keep in mind that this is correct – as for any score test
– only when alternative is reasonably close to the null, i.e. when the SNP
explains small proportion of trait’s variance. Disadvantages of this test are
that is can be relatively slow when thousands of study subjects are analysed,
and that permutation procedures can not be applied to estimate genome-wide
significance, because the data structure is not exchangeable.
Other test, GRAMMAS, also exploits MLEs from the polygenic model (7.1).
However, these are used to first compute the vector of environmental residuals
ê, using standard equation

ê = σ̂e2 · (Φ · σ̂G
2
+ I · σ̂e2 )−1 · (Y − µ̂)

These residuals, in turn, are used to run simple score test:

2
(g − E[g])T · ê)
TG2 =
(g − E[g])T · (g − E[g])

This test is conservative, but GC can be used to correct for the deflation of
the test statistics.
The fact that environmental residuals ê are not dependent on pedigree struc-
ture leads to a nice property of the GRAMMAS test: the data structure be-
comes exchangeable and permutations may be used to estimate genome-wide
significance. When used in combination with GC, P − values derived from
176CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

GRAMMAS test are correct; however, there is a downward bias in estimates of

SNP effects.
When using FASTA or GRAMMAS test, it is recommended to estimate ge-
nomic kinship matrix from available genome-wide data, and use it in analysis
instead of pedigree kinship. This solution firstly does not rely on the complete-
ness and quality of pedigree; secondly, genomic kinship is more likely to give a
better estimate of a true covariance between individual genomes, while pedigree
kinship provides one with expectation. Therefore use of genomic kinship is ex-
pected to lead to better estimates of polygenic model, and thus better power to
detect association in GWA analysis. This being said, we generally advocate use
of genomic, and not pedigree kinship. Of cause, you can only implement this
solution when you have GWA data; in a candidate gene study you will have to
rely on the pedigree structure to estimate kinship matrix.

7.3 Example GWA analysis using family-based

data
DATAPROBLEM In this section, we will explore small data set (notWork-
ingSexprnids(erfs) people, notWorkingSexprnsnps(erfs) SNPs). Let us load and
explore it: DATAPROBLEM trying to keep things working technically
- the interpretation is not sane anymore

> #load("RData/[Link]")
> data([Link])
> erfs <- [Link][1:100,]
> phdata(erfs)$qtbas <- phdata(erfs)$weight
> pkins <- matrix(rnorm(nids(erfs)^2,sd=0.01),ncol=nids(erfs),nrow=nids(erfs))
> ls()

[1] "erfs" "[Link]" "pkins"

> class(erfs)

[1] "[Link]"
attr(,"package")
[1] "GenABEL"

> class(pkins)

[1] "matrix"

You can see that there a two objects, erfs and pkins, presented in the data.
The class of the first object is standard GenABEL-package’s [Link]-class;
this is the object containing GWA data. The other object contains kinship
matrix, as estimated from pedigree data.
You can check the number of people and SNPs in the data set with

> nids(erfs)

[1] 100
7.3. EXAMPLE GWA ANALYSIS USING FAMILY-BASED DATA 177

> nsnps(erfs)
[1] 7374
As usual, it is advisable to check the distribution of SNPs by chromosome:
> table(chromosome(erfs))
1 2 3 X
3482 1927 1417 548
(here, 23 stays for pseudo-autosomal region of the X chromosome); you can see
that markers are evenly spread over the chromosomes.
Summary marker statistics can be generated by
> [Link](gtdata(erfs))
$`Minor allele frequency distribution`
X<=0.01 0.01<X<=0.05 0.05<X<=0.1 0.1<X<=0.2 X>0.2
No 336.000 1289.000 1275.000 1625.00 2849.000
Prop 0.046 0.175 0.173 0.22 0.386

$`Cumulative distr. of number of SNPs out of HWE, at different alpha`

X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 0 1 21.000 182.000 7374
Prop 0 0 0.003 0.025 1

$`Distribution of proportion of successful genotypes (per person)`

X<=0.9 0.9<X<=0.95 0.95<X<=0.98 0.98<X<=0.99 X>0.99
No 0 0 0 56.00 44.00
Prop 0 0 0 0.56 0.44

$`Distribution of proportion of successful genotypes (per SNP)`

X<=0.9 0.9<X<=0.95 0.95<X<=0.98 0.98<X<=0.99 X>0.99
No 0 28.000 1967.000 2710.000 2669.000
Prop 0 0.004 0.267 0.368 0.362

$`Mean heterozygosity for a SNP`

[1] 0.2565366

$`Standard deviation of the mean heterozygosity for a SNP`

[1] 0.1627121

$`Mean heterozygosity for a person`

[1] 0.2460929

$`Standard deviation of mean heterozygosity for a person`

[1] 0.01891617
You can see that the quality of genotypic data is quite reasonable: call rate is
generally high, both per-person and per SNP, and there is little deviation from
Hardy-Weinberg equilibrium.
Let us explore pedigree kinship matrix. First, let us just look how this
matrix looks like by displaying few elements from the upper-left corner:
178CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

> pkins[1:5,1:5]

[,1] [,2] [,3] [,4] [,5]

[1,] -0.0128760758 0.002467059 -0.000566331 0.002766224 0.0026094789
[2,] 0.0140021209 -0.001057102 0.018525944 0.007620800 0.0005592727
[3,] 0.0053502144 0.006628210 -0.008691123 -0.011589443 0.0049269473
[4,] -0.0004255615 0.003463868 0.005752267 0.005118552 0.0058317597
[5,] 0.0064096985 -0.007207738 0.010857333 -0.004234360 0.0139396821

By definition, pedigree kinship should take values between 0 and 0.5 (plus
some small amount from inbreeding); kinship between (non-inbred) sibs or an
offspring and the parent is 1/4. You can see that in the upper-left corner there
is one inbred sib-pair (or parent-offspring pair; ”id2” and ”id5”). You can also
see that this matrix is symmetric around the diagonal.
Let us summarise the distribution of kinship coefficients; in doing this we
want to generate the summary for every off-diagonal element only once. Func-
tion [Link] can be used to get the ”lower triangle” sub-matrix elements:

> summary(pkins[[Link](pkins)])

Min. 1st Qu. Median Mean 3rd Qu. Max.

-3.311e-02 -6.732e-03 -4.235e-05 5.381e-05 6.761e-03 3.410e-02

As you can see, average relationship corresponds to that expected between

second cousins (1/64 = 0.015625) and third cousins (1/256 = 0.00390625).
We can also draw a histogram of the distribution of the kinship coefficients
(shown at figure 7.4A):

> hist(pkins[[Link](pkins)])

and see that most elations are indeed remote.

Let us estimate genomic kinship matrix using autosomal data with the com-
mand ibs, and look up the elements in the upper-left corner:

> gkins <- ibs(erfs[,autosomal(erfs)],weight="freq")

> gkins[1:5,1:5]

id4 id10 id25 id33 id35

id4 0.443821169 6.587000e+03 6.592000e+03 6601.0000000 6587.000000
id10 0.007417817 4.624733e-01 6.582000e+03 6592.0000000 6577.000000
id25 -0.005339590 3.822318e-03 4.783133e-01 6595.0000000 6583.000000
id33 0.001082992 1.240865e-02 -2.744355e-02 0.4946951 6590.000000
id35 -0.029822882 1.299173e-02 5.400605e-03 0.0161940 0.494195

Here, the estimated kinship is shown below the diagonal, and the number of
informative SNP pairs used for estimation is shown above the diagonal.
You can see that ”genomic kinship” coefficients may take values lower than
zero, which is consequence of the fact that in effect ”genomic kinship” is simply
covariance between the vectors of individual genotypes. This quantity, though
it provides an unbiased estimate of kinship, can be lower than zero.

> summary(gkins[[Link](gkins)])
7.3. EXAMPLE GWA ANALYSIS USING FAMILY-BASED DATA 179

0.15
●

● ● ●●
● ●
●
0.10

●
●● ● ●
●
● ● ● ●● ● ●●●● ●
● ●● ●●● ●● ● ●
● ● ●● ●●● ●
●●● ● ●● ● ●● ●● ● ● ●
●● ●●●● ● ●● ●
0.05

●
● ●● ● ●●● ●● ● ● ● ● ●● ●
●●●● ●● ●● ●
● ● ●
●● ● ●
●
● ● ● ●●●
●●● ● ●●● ●●● ● ●●
●● ● ●●
● ● ●● ●●● ●
● ● ● ● ● ●● ●● ●
pcs[, 1:2][,2]

●● ● ●● ●● ● ●●● ● ● ● ● ● ●
● ● ● ●● ● ●● ●● ● ●● ● ● ●
●●
●●● ●●●●
●● ●●
●●●●● ●● ●● ●●●●●● ●● ● ●● ●● ●
●● ●● ●●
● ● ● ●●●● ●●
● ●●
●● ● ●
● ●●●●● ●
● ●●
● ●● ●●● ● ●● ● ●
● ● ● ●●●● ●
●
● ●
● ●● ●
● ●●●●
●
●
● ●
● ●● ●● ●●● ● ●●●
●
●● ● ● ●
● ● ●
●●●● ● ● ●● ●● ●●● ● ●
● ●● ●●●● ●●●●●● ● ●●● ●●●● ●●
●●● ●● ● ●
0.00

● ●●●● ●
●●
● ●
●
●● ●● ●
●
●
●● ●●●●
●●●
● ● ●● ●●●●
● ●● ● ●● ●●
● ●● ●● ● ●● ●
● ● ●● ● ●●● ●●●● ● ● ●● ●●● ●●●●● ●● ● ●●
●●
●● ● ● ● ●● ●
● ● ●● ● ● ● ●●
●● ●●●● ● ●
● ●● ●● ●● ● ●
● ●●● ●● ● ● ●
●
●● ●
●
●
●
●
●● ●●●●●●●●
●
●● ● ●
● ● ● ●● ●● ●● ● ●●●●●
● ●●●●●●●●● ●● ● ●●● ●● ●●●● ● ●●●●● ●
● ●
● ●
● ● ● ●● ●● ●
●● ●● ●● ●
●
● ●●●●●
●
●●● ●●
● ● ●●●●
● ●
● ●
●●●● ●●
●●● ●● ● ●● ●
● ●●● ● ●● ● ● ●●
● ●● ●● ●
−0.05

● ● ● ● ●● ● ●
● ●● ●●●● ● ●●
● ●●●●● ● ●
● ● ●● ● ●
● ●
●
● ● ● ●● ●● ● ● ●
●
●●● ● ●
●
● ● ● ● ● ●
●● ● ● ● ●●● ● ●● ● ● ●●
● ●
●
● ● ● ●● ● ● ●
● ● ● ●●● ● ●●● ●●
● ● ●● ●● ● ● ● ● ● ●
● ●●
● ● ●● ●
● ●● ●●● ● ●
−0.10

● ● ●● ●● ● ●
● ● ●
●
● ● ● ●
● ●● ●
●●

−0.15 −0.10 −0.05 0.00 0.05 0.10 0.15

pcs[, 1:2][,1]

Figure 7.3: Population structure of ’strdat1’ data

Min. 1st Qu. Median Mean 3rd Qu. Max.

-0.074800 -0.020410 -0.008065 -0.005080 0.006761 0.154700

here, the average is quite close to that obtained with pedigree kinship.
We can also draw a histogram of the distribution of ”genomic kinship” co-
efficients (shown at figure 7.4B):

> hist(gkins[[Link](gkins)])

and can easily graphically present relations between genomic and pedigree kin-
ship with

> plot(pkins[[Link](pkins)],gkins[[Link](gkins)])

(shown at figure 7.5), and estimate correlation between the two with

> cor(pkins[[Link](pkins)],gkins[[Link](gkins)])

[1] 0.01196389

From the graph, you can clearly see that, though there is a very strong
correlation between genomic and pedigree kinships, these are not identical.
In real data, you may find that there are some points where pedigree data
clearly suggest relation different from that suggested by genomic data. Which
one to believe? Generally, pedigrees are more prone to errors than genotypic
180CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

Histogram of pkins[[Link](pkins)]

800
Frequency

400
0

−0.03 −0.02 −0.01 0.00 0.01 0.02 0.03

pkins[[Link](pkins)]

Histogram of gkins[[Link](gkins)]
2000
Frequency

1000
0

−0.05 0.00 0.05 0.10 0.15

gkins[[Link](gkins)]

Figure 7.4: Distribution of the pedigree (upper histogram) and genomic (lower
histogram) kinship coefficients for erfs data set.

data. In the data containing close relatives it is better to rely on ”genomic

kinship”.
Let us first analyse the data using plain GC method:

> qts <- qtscore(qtbas,data=erfs)

You can check the estimate of the inflation factor λ with

> lambda(qts)$est

[1] 1.033775

This is relatively high value, suggesting presence of close relatives in data and
high heritability of the trait.
The top 10 hits from GWA analysis can be displayed with

> [Link](qts,sort="Pc1df")

Summary for top 10 results, sorted by Pc1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs5941875 1 3640023 + T A 97 19.68681 4.942356 15.86656
rs8994834 1 1845763 + T G 98 14.48396 3.898478 13.80335
rs4550412 1 1844249 + A G 99 15.14198 4.147116 13.33131
rs9253916 1 3655610 + C A 100 15.37248 4.237537 13.16015
7.3. EXAMPLE GWA ANALYSIS USING FAMILY-BASED DATA 181

rs9154666 1 3689697 + G C 99 14.77000 4.093217 13.02062

rs762401 1 3689720 + T C 99 -14.69329 4.090225 12.90457
rs7881009 2 7390224 - C A 100 -33.38444 9.454799 12.46762
rs3361134 3 11192711 + G A 99 -13.37212 3.801274 12.37491
rs6927636 1 385190 - G A 99 15.40149 4.421136 12.13551
rs1289801 1 4041666 + G T 100 26.26488 7.627108 11.85854
P1df effAB effBB chi2.2df P2df Pc1df
rs5941875 6.796909e-05 22.154991 24.12716 16.84336 0.0002200451 8.940686e-05
rs8994834 2.029744e-04 13.336702 29.93850 13.88632 0.0009652165 2.580954e-04
rs4550412 2.610112e-04 12.319392 35.27284 13.97540 0.0009231695 3.293272e-04
rs9253916 2.859663e-04 21.651670 18.72325 16.14362 0.0003122177 3.597997e-04
rs9154666 3.080797e-04 13.685849 31.22516 13.11489 0.0014195048 3.867340e-04
rs762401 3.277815e-04 -17.539309 -31.11841 13.00428 0.0015002246 4.106811e-04
rs7881009 4.140675e-04 -33.384438 NA 12.46762 0.0004140675 5.150733e-04
rs3361134 4.351428e-04 -6.364411 -26.42515 14.11220 0.0008621359 5.404643e-04
rs6927636 4.947085e-04 16.864623 27.51095 12.35393 0.0020767259 6.120307e-04
rs1289801 5.739765e-04 26.264881 NA 11.85854 0.0005739765 7.068666e-04

here, nominal P − values after genomic control are given in column named
”Pc1df”.
We can estimate genome-wide empirical significance by using the same func-
tion with times argument, which tells the number of permutations:

> qts.e <- qtscore(qtbas,data=erfs,times=200,quiet=TRUE)

|
| | 0%
|
|======================================================================| 100%

> [Link](qts.e,sort="Pc1df")

Summary for top 10 results, sorted by Pc1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs5941875 1 3640023 + T A 97 19.68681 4.942356 15.86656
rs8994834 1 1845763 + T G 98 14.48396 3.898478 13.80335
rs4550412 1 1844249 + A G 99 15.14198 4.147116 13.33131
rs9253916 1 3655610 + C A 100 15.37248 4.237537 13.16015
rs9154666 1 3689697 + G C 99 14.77000 4.093217 13.02062
rs762401 1 3689720 + T C 99 -14.69329 4.090225 12.90457
rs7881009 2 7390224 - C A 100 -33.38444 9.454799 12.46762
rs3361134 3 11192711 + G A 99 -13.37212 3.801274 12.37491
rs6927636 1 385190 - G A 99 15.40149 4.421136 12.13551
rs1289801 1 4041666 + G T 100 26.26488 7.627108 11.85854
P1df Pc1df effAB effBB chi2.2df P2df
rs5941875 0.150 0.195 22.154991 24.12716 16.84336 NA
rs8994834 0.450 0.520 13.336702 29.93850 13.88632 NA
rs4550412 0.530 0.635 12.319392 35.27284 13.97540 NA
rs9253916 0.590 0.660 21.651670 18.72325 16.14362 NA
rs9154666 0.610 0.690 13.685849 31.22516 13.11489 NA
rs762401 0.635 0.705 -17.539309 -31.11841 13.00428 NA
182CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

rs7881009 0.710 0.755 -33.384438 NA 12.46762 NA

rs3361134 0.720 0.780 -6.364411 -26.42515 14.11220 NA
rs6927636 0.745 0.805 16.864623 27.51095 12.35393 NA
rs1289801 0.790 0.850 26.264881 NA 11.85854 NA

(argument ”quiet” supress warning messages; this is used for the purposes of
this tutorial, normally you do not need to specify this option)
As you can see, in this analysis nothing comes even close to genome-wide sig-
nificance, as indicated by genome-wide corrected P − values (column ”Pc1df”)
all >> 0.05 .
Let us estimate polygenic model with
> h2 <- polygenic(qtbas,kin=gkins,data=erfs)
The results of estimation are contained in ”h2an” element of the resulting
analysis object:
> h2$h2an

$minimum
[1] 747.5489

$estimate
[1] 92.3114671 0.4273438 663.0750850

$gradient
[1] -1.421085e-07 -5.684342e-08 1.278977e-07

$code
[1] 1

$iterations
[1] 7

In the ”estimate” list, the MLEs shown correspond to intercept µ̂, heritability
ĥ2 = σ̂G
2 2
/(σ̂G + σ̂e2 ), and total variance σ̂T2 = σ̂G
2
+ σ̂e2 . You can see that
heritability of the trait is indeed high – almost 80%.
Under these conditions (hight heritability, presence of close relatives) we
may expect that FASTA and GRAMMAS analysis exploiting heritability model
and relationship matrix in exact manner may have better power compared to
simple GC.
Let us run FASTA test using estimated polygenic model, as specified by h2
object:
> mms <- mmscore(h2,data=erfs)
There is little residual inflation left when we use ”genomic kinship” matrix:
> lambda(mms)$est

[1] 1

And the significance of ”top” hit becomes an order of magnitude better

compared to plain GC:
7.3. EXAMPLE GWA ANALYSIS USING FAMILY-BASED DATA 183

> [Link](mms,sort="Pc1df")

Summary for top 10 results, sorted by Pc1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs5941875 1 3640023 + T A 97 20.31359 5.201615 15.25096
rs1289801 1 4041666 + G T 100 29.17214 7.817672 13.92457
rs4550412 1 1844249 + A G 99 15.94462 4.311427 13.67686
rs8994834 1 1845763 + T G 98 14.75871 4.061652 13.20357
rs9253916 1 3655610 + C A 100 16.17211 4.596761 12.37740
rs6927636 1 385190 - G A 99 15.46655 4.437391 12.14875
rs6191668 1 3630425 - G A 100 17.23958 5.060666 11.60481
rs3361134 3 11192711 + G A 99 -13.05935 3.912352 11.14210
rs9154414 1 1836956 - A C 100 14.27818 4.307012 10.98990
rs7881009 2 7390224 - C A 100 -31.57798 9.543831 10.94771
P1df Pc1df effAB effBB chi2.2df P2df
rs5941875 0.0000941288 0.0000941288 NA NA 0 NA
rs1289801 0.0001902950 0.0001902950 NA NA 0 NA
rs4550412 0.0002171135 0.0002171135 NA NA 0 NA
rs8994834 0.0002794163 0.0002794163 NA NA 0 NA
rs9253916 0.0004345614 0.0004345614 NA NA 0 NA
rs6927636 0.0004912069 0.0004912069 NA NA 0 NA
rs6191668 0.0006578149 0.0006578149 NA NA 0 NA
rs3361134 0.0008439010 0.0008439010 NA NA 0 NA
rs9154414 0.0009160997 0.0009160997 NA NA 0 NA
rs7881009 0.0009371951 0.0009371951 NA NA 0 NA

If you compare these results to that obtained with simple GC, you can also
see that the ranks of top hits have changed quite a bit; unbiased estimated of
genetic effects were obtained.
However, we can not estimate genome-wide significance with FASTA, be-
cause the data structure is not exchangeble.
Using GRAMMAS method, you can estimate nominal P − values by
> grs <- qtscore(h2$pgres,data=erfs,clam=FALSE)
> lambda(grs)$est

[1] 0.8998862

In the above analysis, note that the estimated ”inflation” factor λ is less
than one, i.e. now it is the GRAMMAS deflation factor. In order to obtain
non-concervative test statistics, we had to say to qtscore that deflation is OK
(parameter clam=FALSE).
We can see ”top” nominal corrected P − values with
> [Link](grs,sort="Pc1df")

Summary for top 10 results, sorted by Pc1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs5941875 1 3640023 + T A 97 11.049817 2.892067 14.598002
rs1289801 1 4041666 + G T 100 16.025348 4.475877 12.819130
rs4550412 1 1844249 + A G 99 8.433205 2.431798 12.026249
rs6927636 1 385190 - G A 99 8.772838 2.592604 11.450066
184CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

rs8994834 1 1845763 + T G 98 7.733678 2.293640 11.368989

rs7881009 2 7390224 - C A 100 -17.886196 5.548436 10.391897
rs9253916 1 3655610 + C A 100 7.931660 2.486748 10.173366
rs3361134 3 11192711 + G A 99 -7.058117 2.231245 10.006525
rs6191668 1 3630425 - G A 100 8.580199 2.757873 9.679351
rs9154414 1 1836956 - A C 100 7.501683 2.411571 9.676480
P1df effAB effBB chi2.2df P2df Pc1df
rs5941875 0.0001330556 12.468348 13.337066 15.54028 0.0004221544 0.0000563345
rs1289801 0.0003430932 16.025348 NA 12.81913 0.0003430932 0.0001604628
rs4550412 0.0005245653 7.599102 18.340676 12.18983 0.0022543047 0.0002564749
rs6927636 0.0007149146 9.564196 15.765123 11.63588 0.0029737318 0.0003610113
rs8994834 0.0007468052 7.076555 16.023283 11.44763 0.0032672281 0.0003788329
rs7881009 0.0012656955 -17.886196 NA 10.39190 0.0012656955 0.0006782175
rs9253916 0.0014248395 11.762204 8.529627 13.39739 0.0012325173 0.0007729162
rs3361134 0.0015598655 -3.716155 -13.964066 11.15333 0.0037851766 0.0008541251
rs6191668 0.0018635041 11.907381 3.293695 13.63234 0.0010959101 0.0010393086
rs9154414 0.0018664189 6.073566 17.517735 10.15451 0.0062370161 0.0010411021

By comparing this output to that from FASTA test, you can see that P −
values are quite close, but the effects are underestimated with GRAMMAS, as
expected.
However, the streangths of GRAMMAS test is not only its speed, but also
possiblity to estimate genome-wide significance. This can be done by

> grs.e <- qtscore(h2$pgres,data=erfs,times=200,clam=FALSE,quiet=TRUE)

|
| | 0%
|
|======================================================================| 100%

> [Link](grs.e,sort="Pc1df")

Summary for top 10 results, sorted by Pc1df

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs5941875 1 3640023 + T A 97 11.049817 2.892067 14.598002
rs1289801 1 4041666 + G T 100 16.025348 4.475877 12.819130
rs4550412 1 1844249 + A G 99 8.433205 2.431798 12.026249
rs6927636 1 385190 - G A 99 8.772838 2.592604 11.450066
rs8994834 1 1845763 + T G 98 7.733678 2.293640 11.368989
rs7881009 2 7390224 - C A 100 -17.886196 5.548436 10.391897
rs9253916 1 3655610 + C A 100 7.931660 2.486748 10.173366
rs3361134 3 11192711 + G A 99 -7.058117 2.231245 10.006525
rs9154414 1 1836956 - A C 100 7.501683 2.411571 9.676480
rs6191668 1 3630425 - G A 100 8.580199 2.757873 9.679351
P1df Pc1df effAB effBB chi2.2df P2df
rs5941875 0.290 0.120 12.468348 13.337066 15.54028 NA
rs1289801 0.595 0.330 16.025348 NA 12.81913 NA
rs4550412 0.770 0.475 7.599102 18.340676 12.18983 NA
rs6927636 0.860 0.615 9.564196 15.765123 11.63588 NA
rs8994834 0.880 0.635 7.076555 16.023283 11.44763 NA
7.3. EXAMPLE GWA ANALYSIS USING FAMILY-BASED DATA 185

0.15
●
● ●

●
● ●
●
● ●
●
0.10

● ● ●● ●
●● ● ● ● ●
●● ●
● ● ●
●
●● ●●
gkins[[Link](gkins)]

●●
● ● ●● ●
●
● ●●● ●
● ●●
● ●
●● ●
● ●● ● ●● ● ●
●
● ●● ● ● ●●
● ●● ●● ● ● ● ● ●● ● ●● ● ●● ●● ●●
0.05

●●● ●● ● ●
● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ●
●● ● ● ●●● ● ● ●
● ●● ●●●● ● ● ● ●●
●● ●
●●● ● ●●●●●●●● ●
● ●●●
●
●● ●●● ●● ●
●● ● ●
●● ●●●●●● ●● ●●●● ●●
● ●●●
●● ●
●●
●●● ●● ●
● ● ●●● ●
●
●● ●●
● ● ●●●●
● ●●
●
● ● ●●
●
●
●●●●●●
●
●
●
●
●●
●
● ●●●
●
● ● ●●●●●●●●
●● ●●● ● ● ● ●
● ●● ●● ●
●●● ● ● ●●● ●●●
● ●
●● ● ●● ● ●●●●●
●●
●●●●
● ●●●
●●●●
●●● ●● ●●
● ●
●●●●●● ●
●●●●●●●●●●
● ●●●
●● ●●●
● ●●
●●● ●●
●●●
●
●● ● ● ●
● ● ●●●● ●●
●●
●●●
●●
● ● ●
●●
●● ●●
●●●●
● ●
● ●
● ●
●
● ●●
●●●●●●● ●● ●● ●●●● ●
●
● ● ● ●●● ●
●●●●
●
●●
● ●●●
●
●
●
●●●●●●●●
●
●
●●●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
● ● ●
●
●●
● ● ●
●●●
●
● ●
●
●
●
●
●
●
●●
●● ●●
●●
● ●
●● ●
●●●
●
●●
●●
●
●●●●
● ●● ●
● ● ●● ●●
●● ● ●● ●●●
●● ●● ●
●
●●●●
● ●
●●●●●
●
●●
●● ●
● ●●
●●●●
● ● ●●●● ●
● ● ● ● ●
●● ●
●
●●●●●●
● ● ●●●●●●●●●●●
● ●●
●● ●●●● ●
●●●
●
●●
●●
●●●
● ●
●●
●
●●
●●
●●●●●
●●●
●● ●
●●
● ●●●●
●●
●●● ●● ● ●
●●●●● ● ● ●
● ●●● ● ●●●●●●●
●●●
●●
●●●●●●●●
●●
●●●
●
●●
●
●
●●
●●●● ●
●●
●●
●●●
●
●●
●
●●
● ●
●
●●
●●●
●
●
●●●
●
● ●
●
●●
●
●●
●●
●●
●●●●
●
● ●●
●
●●●●
● ●●●●●
●●●
●● ●●●●
●●●● ● ●● ●
● ●●
● ●●
●● ●●● ●
●●●
●●
●●
●
●●
●
●
●
●
●
●●● ●
●●
●
●
●●●●
●●●
●
● ●
●
●●●
●
●
●●
●
●●
●
●
●
●●
●●●
●
●●
●●
●
●●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●●●
●
●
●
●
●
●
●
●
● ●
●
●
●●●
●
●●●
●
●●●
●●●●● ●
●● ●●●●●
●● ●● ● ●●
●●● ● ● ●
●●●● ● ●● ● ● ●● ● ●● ● ● ●● ●
● ● ● ●● ●
0.00

● ●
● ●● ● ● ●● ● ●●
● ●●●
●
●●●●●●●
●
●●●●
●● ●
●●
● ●●
●●
●
●●
●
●●●●
●●
●
●
●
●●
●●●
●
●●
●
●●
●
●
●●
●●●●
●
●● ●
●●
●●
●●
●
●●
●●
●
●●●
●
●
●●
●
● ● ●
● ●
●●
● ●
●●
●●● ●
●●● ● ●
●
● ● ●●● ● ● ● ●● ●● ●●●● ●●●●●
●●●●● ●● ●●●●
● ● ●
● ●
●● ●●●●●
● ●
●●●●● ●
●●● ● ●
●● ● ●
●
●● ●●● ●
●●●
●● ● ●●
● ●
●●
●
●●
●●
●●
●●●
●●
●●
●
●●
●
●
●●
●
●●●
●
●●
●● ●
●●●
●
●●●
●
●●
●
●●
●
●●●
●
●●
●●
●
●●●
●
●●
●
●
●●
●
●●
●
●●
●●
●●
●●
●●
●
●●●
●●●
● ●
●
●●●●
●
●●
●●●
● ●●
●● ●●●● ●●●●● ● ● ●
●
●● ●
● ●●●
●
●●●●●●● ●●●●
●● ●●
●
●●●● ●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●
●●
●●
●
●●
●●
●
●●
●
●●
●●
●
●
●●
●●● ●
●●●
●
●
●●
●
●
●●
●●
●
●
●●●
●
●●●● ●
●●●●●● ●●
●● ●
● ●
●
● ● ●●●●●
●
●
●●
●
●●
●
●●●
●
●
●
●
●
●●
●
●●●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●●●
●
●
●●●
●●
●
●
●●●
●●
●
●
●●
●●
●
●
●
●
●● ●
●●● ●
●●
●●●
●●●
●● ●●
● ●
● ● ●● ●●●
●●●●
●●●●
●●
●
●●
●
●●
●●
●●
●●
●
●●
●●●
●
●
●●
●
●●
●
●
● ●
●
●●●
●
●●●
●
●
●●
●
●●
●●
●●●
●
●
●●
●●
●●●
●
●●
●●
●●●
●
●●
●
●
●●●
●●
●●
●
●●
●●
●
●●●
●
●●●
●
●
●
●●●●●●
●●●●●
●●
●●
● ●●●●
●
●●
● ● ●
● ●●
● ● ● ●
● ●● ● ● ●●
●
● ●●
●● ●●
●● ●●●●●
● ●
●
●●
●●●
●●●
●
●●●
●
●●
●
●
●●
●●●●
●
● ●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●●●
●●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●●●
●
●
●●
●●
●●
●
●●
●
●●●
●
●
●● ●
●
●
●●
●
●
● ●
●
●
●
●●
●
●●
●●
●
●
●
●●
●
●●
●●●
● ●●
●●●
●
●●●● ● ●●
●
●● ●● ●●●●
●●● ● ● ● ●● ●●
● ●
●● ●●●●
●●●
●●●●●
●●●
●
●●
●●●●
●●
●●
●●
●
● ●
●●
●●
●●●●
●
●
●
●●
●●●●
●
●●
●●
●●
●●
●●
●
●
●●●
●● ● ●●●
●
●●●● ●●
● ●●●
●●●
●●
●● ●●●●
● ●
●
●
●●●●● ●
●●●●
●
●●●
●●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●
●
●
●●● ●
●●●●●
●●●●
● ●
●
●● ●● ●●
● ●
●● ●
● ●●●●
●●
●
● ●
●●
●●●
●
● ●●
●●
●●●●
●
●● ●
●●
●●
●
●
● ●●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●
●●●
●●
●
●●
●●
●
●
●
●
●
●●●
●
●
●●●●
●
●●
●
●
●●
●●
●●●
●
●●
●●
●● ● ●
●●
●●
●
●● ●
●●●●
●● ●●● ●
●●● ●
●
● ●●●●●●
● ●●● ●● ●
●●
●●●● ●●
●●●
●● ●
●
●●
●
●●●
●●
● ●
●●●
●
● ●
● ●
●
●●●
●●●
●
●●
●●●●●●●●
●●●
●
●●●● ● ●●
●●●●●●● ●●●
● ● ● ●●● ● ● ●●
● ●●● ●●● ●●● ●● ● ●●●●●●
●● ●●
●●
●
●
●●
●
●●●●
●
●●
●●
●●
●
●
●●
●
●●●●
●●
●
●
● ●
●
●●●
●
●
●
●
●
●●
●
●
●
●●
● ●
●●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●● ●
●
●●●●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●●
● ●●
●●
●●●
●● ●● ●● ●
●
●●●●● ●
●
●
●●●
●● ●
● ● ● ●●●●●●
●
●●
●
●●●
●●● ●
●●●
●●●
●●
●●
●
●●●●
● ● ●
● ●●●
●
●●
● ●● ●● ●●●●●●● ●
●
●●● ●●●● ●●
● ●● ●● ●●●● ●● ●● ●● ●
●●●● ●
●●● ●●● ●●● ●● ●●●●
−0.05

● ●● ● ● ●● ●● ● ● ● ● ●●●
● ● ● ● ●● ●●
● ●● ● ●●●● ● ●
●●●● ● ●●
● ●●●● ● ●●
●● ● ●●
● ● ● ● ●
● ● ● ● ● ● ●● ●●●
● ● ●
● ● ● ● ● ●
● ●
● ● ● ●●
● ●●●
● ● ● ●●● ● ●
● ●● ●
● ●●
● ● ●● ●● ●
● ●
●
●●
●

−0.03 −0.02 −0.01 0.00 0.01 0.02 0.03

pkins[[Link](pkins)]

Figure 7.5: Scatter-plot relating pedigree and genomic kinships for erfs data
set.

rs7881009 0.985 0.855 -17.886196 NA 10.39190 NA

rs9253916 0.990 0.890 11.762204 8.529627 13.39739 NA
rs3361134 0.995 0.925 -3.716155 -13.964066 11.15333 NA
rs9154414 1.000 0.965 6.073566 17.517735 10.15451 NA
rs6191668 1.000 0.965 11.907381 3.293695 13.63234 NA

As you can see, now the ”top” hit starts approaching genome-wide signif-
icance (genome-wide P − value ∼ 10%), showing the power of kinship-based
methods under high heritability model.
Finally, let us plot −log10 nominal P − values from different methods across
the genome. Let black dots correspond to GC, green to GRAMMAS and red to
FASTA (figure 7.6):

> plot(mms,df="Pc1df")
> [Link](grs,df="Pc1df",col=c("lightgreen","lightblue"),cex=1.2)
> [Link](qts)

You can see that there is a great degree of correlation between the FASTA
and GRAMMAS P − values, while plain GC really stands apart.
186CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

7.4 Exercise: analysis of family data

Exercise 1 Repeat heritability estimation, FASTA and GRAMMAS analysis of
previous section using pediree kinship (pkins object). Discuss the results.

In the next section, you will explore a small (695 people) subset of people
from ERF, a family-based study with participants coming from a genetically
isolated population and sampled based on kinship (all living descendants of 22
couples living in the area in mid-XIXth century). The study participants were
genotyped using Illumina 6K ”linkage” array. QC was already performed. Your
trait of interest is ”qtbas”.
Explore the data set and answer the questions:

Exercise 2 Describe the trait ”qt”. Can you detect significant outliers at visual
inspection? Is trait distributed normally? What are significant covariates?

Exercise 3 Explore relations between genomic and pedigree kinship (these are
provided in data as gkin and pkin data objects, respectively). What are your
conclusions? Which matrix would you use later on?

Exercise 4 What is the heritability of the trait (take care: polygenic analysis
may rung for a long while)? Based on heritability analysis, how would you rank
different methods of GWA analysis for this trait (and why)?

Exercise 5 Do GWA analysis using simple score test with genomic control.
Estimate genome-wide significance. What are your conclusions?
Run GWA analysis using the ”best” method and model as you have decided
in previous exercises. Estimate genome-wide significance. What are your con-
clusions? Did they change compared to simple analysis?

Exercise 6 Repeat the last ”best” analysis using pedigree kinship. How your
results change?

Exercise 7 If you have any time left – repeat analysis using ”qt” trait. This
one is much more fun, but also more laborous to analyse.

7.5 Answers to exercises

Answer (Ex. 1) — Basically, all covariates are significantly associated with

the traits:
> summary(lm(quat~sex+age+age2,data=phdata(strdat)))
Call:
lm(formula = quat ~ sex + age + age2, data = phdata(strdat))

Residuals:
Min 1Q Median 3Q Max
-21.3585 -4.9637 0.0959 4.7896 22.4212

Coefficients:
7.5. ANSWERS TO EXERCISES 187

Estimate Std. Error t value Pr(>|t|)

(Intercept) 167.935224 3.139965 53.483 <2e-16 ***
sex 12.102771 0.505588 23.938 <2e-16 ***
age -0.025454 0.127427 -0.200 0.842
age2 -0.001697 0.001247 -1.361 0.174
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 7.222 on 816 degrees of freedom

(2 observations deleted due to missingness)
Multiple R-squared: 0.4517, Adjusted R-squared: 0.4497
F-statistic: 224 on 3 and 816 DF, p-value: < 2.2e-16
> summary(glm(bint~sex+age+age2,data=phdata(strdat),family=binomial))
Call:
glm(formula = bint ~ sex + age + age2, family = binomial, data = phdata(strdat))

Deviance Residuals:
Min 1Q Median 3Q Max
-1.3318 -1.1169 -0.8151 1.1907 1.5775

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -3.1234742 0.9295943 -3.360 0.000779 ***
sex 0.4913847 0.1424384 3.450 0.000561 ***
age 0.0976423 0.0373181 2.616 0.008884 **
age2 -0.0007977 0.0003623 -2.202 0.027697 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 1137.4 on 821 degrees of freedom

Residual deviance: 1110.6 on 818 degrees of freedom
AIC: 1118.6

Number of Fisher Scoring iterations: 4

Answer (Ex. 2) — How many SNPs and IDs are presented in the data set?
> nsnps(strdat)
[1] 7374
> nids(strdat)
[1] 822

Answer (Ex. 3) — Let us perform the QC:

> qc <- [Link](strdat,call=0.98,[Link]=0.98,[Link]=0)
188CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

Excluding people/markers with extremely low call rate...

7374 markers and 822 people in total
0 people excluded because of call rate < 0.1
0 markers excluded because of call rate < 0.1
Passed: 7374 markers and 822 people

Running sex chromosome checks...

0 heterozygous X-linked male genotypes found
0 X-linked markers are likely to be autosomal (odds > 1000 )
0 male are likely to be female (odds > 1000 )
0 female are likely to be male (odds > 1000 )
0 people have intermediate X-chromosome inbreeding (0.5 > F > 0.5)
If these people/markers are removed, 0 heterozygous male genotypes are left
Passed: 7374 markers and 822 people

no X/Y/mtDNA-errors to fix

RUN 1
7374 markers and 822 people in total
0 (0%) markers excluded as having low (<0.3041363%) minor allele frequency
32 (0.4339571%) markers excluded because of low (<98%) call rate
0 (0%) markers excluded because they are out of HWE (P <0)
0 (0%) people excluded because of low (<98%) call rate
Mean autosomal HET is 0.2564845 (s.e. 0.01498989)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7899338 (s.e. 0.01170803), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 7342 (99.56604%) markers passed all criteria
In total, 822 (100%) people passed all criteria

RUN 2
7342 markers and 822 people in total
0 (0%) markers excluded as having low (<0.3041363%) minor allele frequency
0 (0%) markers excluded because of low (<98%) call rate
0 (0%) markers excluded because they are out of HWE (P <0)
0 (0%) people excluded because of low (<98%) call rate
Mean autosomal HET is 0.2564845 (s.e. 0.01498989)
0 people excluded because too high autosomal heterozygosity (FDR <1%)
Mean IBS is 0.7888525 (s.e. 0.01175626), as based on 2000 autosomal markers
0 (0%) people excluded because of too high IBS (>=0.95)
In total, 7342 (100%) markers passed all criteria
In total, 822 (100%) people passed all criteria
> strdat1 <- strdat[qc$idok,qc$snpok]
The number of IDs and SNPs passing are
> nsnps(strdat1)
[1] 7342
> nids(strdat1)
7.5. ANSWERS TO EXERCISES 189

[1] 822

Answer (Ex. 4) — Not really:

> [Link](strdat1)[2]
$`Cumulative distr. of number of SNPs out of HWE, at different alpha`
X<=1e-04 X<=0.001 X<=0.01 X<=0.05 all X
No 2 5.000 64.000 316.000 7342
Prop 0 0.001 0.009 0.043 1

Answer (Ex. 5) — Yes:

> qts.q <- qtscore(quat~sex+age+age2,strdat1)
> lambda(qts.q)
$estimate
[1] 1.29321

$se
[1] 0.005757743

Answer (Ex. 6) — Yes:

> qts.b <- qtscore(bint~sex+age+age2,strdat1)
> lambda(qts.b)
$estimate
[1] 1.153616

$se
[1] 0.0011015

Answer (Ex. 7) — Two genetically distinct populations are present in the

data set. These can be visualised with
> gkin <- ibs(strdat1,w="freq")
> dst <- [Link](0.5-gkin)
> pcs <- cmdscale(dst,k=10)
> plot(pcs[,1:2])
(graph presented in figure 7.3)
The populations can be dustinguished with
> pop <- 1*(pcs[,1]>0)
and the number of people belonging to each can be computed with
> table(pop)
pop
0 1
408 414
190CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

mmscore(h2, erfs)

− log10(P − value)
●

4
●
● ●
● ●
●
● ●
●●

3
● ● ● ●
● ● ●● ● ●
● ● ● ●● ● ● ●●● ● ● ●
●
● ●● ● ●● ●● ●
● ● ● ● ●● ●● ● ●● ●● ● ● ●

2
● ●● ● ●● ●● ●● ● ●●●● ●● ●
●● ● ●●●●
● ●● ●● ●●● ● ● ●●●●●● ●●● ●● ●● ●● ● ●● ● ●●● ●● ● ●●
●●●● ●●●● ● ●● ●●●
●
● ●●●
●
●● ●● ●●● ●●●●● ●
●●●●●●● ●● ● ●
● ●●● ●●●● ● ●●● ●●●
●
● ● ● ●● ●●●● ●● ●● ●●● ● ●
●
●
●●
● ●●
●●
●
●
●
●
● ●●●
●
●
●
●
● ●●
●● ● ●●●
● ●●
●
●
●
●●
●
●●●●●
●
●
●●●
● ●●● ●
●●
●●●●
●●
●
●
●
●●
●
● ●
●
●
●
●
●
●
● ● ●
●●●
●
●●
●●●●
●●●●●
●● ●
●
●
●
●●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●● ●● ●● ●
● ●
●●●●● ●●● ●●
●
●●●
●
●
●
●● ●
●
●●●●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●●●●●●●
●
●●
●
●●
●●●
●●●●
●●●●●●●
●●
●● ●●●●
●●
●
●
●
● ●●
● ● ● ●●● ● ●●
●●●
●●
● ●●●
● ●●●●●●● ●●●
● ●
●●●●●●
● ●
●●●● ● ●●●●
●● ●
●●●● ●●●●
●●●
●●●
● ●● ●● ●
●● ●●●
●● ●●
● ●●●
● ● ●
●● ●● ●
●●●
● ● ● ●● ● ●● ●● ●●

1
●
●
●●
●
●
●
●
●●●
●●●
●
●
●●
●●●
● ●
●●
●●
●
●●●●●
●
●●
●
●●
●
●●
●●●
●● ●
●●●●
●●
●
●●
●
●●●●
● ●
●
●●●
● ●
●●
● ●
●●
●●●●
●●
●●●
●
●
●
●●
●●
●●
● ●
● ●
●
●●
●●
●●
●
● ●
●●
●
●●
●●
● ●
●
● ●
●
● ●
●●
● ●
●●
●●●
●
●●
●
●●
●●
●●●
●●●●●●
●
●●●
●●
●
●●●
●●
●
●●
●●●●
●●
●
●●●
● ●
●●●
●●
●●
● ●●
●
●●
●
●●
●●
●●
●●
●● ●●●●
● ●
●
●
●
●●●●
●●
●
●●
●●
●●
●
●● ●
●●●
● ●
● ●
●●
●
●●
● ●
● ●●
●●
●●
●●
●
●●●●
●●
●
●●
●●
●●
●●
●
●●●●
●
●
●●
●
●
●●
●●
●●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●●●●●
●●
●
●
●●
●●●
●
●●
●●
●●
● ●
●●
●●●
●
● ●
●
●
●●●●●
●
●
●
●●●●
●
●
●●
●
●●
●●
●
●●
●
●
●●
●●
●●
●
●●●
●●
●●
●●
●
●
● ●● ●
●●
●
●●●
●
● ●
●●
●
●
●
● ●
●
●●
●●
●
●
●●●●
●
●●
●
●●●
●
●●
● ●
●
●●●●
●●
●●
●
●●●
●
●
●●
●●
●●●
●
●
●●
●
●●
●
●●
●
●●
●●
●●
●
●●
●●
●●
●●
●
●●●
●
●●
●
●●
●
●
● ●
●
●●
●●
●
●●
●●●●●
●●
●●●●
●
●
●
●●
●●
●●
●●●●
●●
●
●
●●●
●●
●
●●
●
●●
●●
●●
●●●●
● ●●●
●
●
●●
●●
●
●●
●●
●
●●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●●
●
●●
●●
●
●●
●●●
●
●●
●●
●●
●●
●
●
●
●●●
●●
●●
●
●●
●●
●
●
●
●●
●●●
●
●
●
●
● ●
●
●●
●
●
●●
●●
●
●
●●
●
●●
●●●
●
●
●●
●
●●
●
●
●
●●
●●
●●
● ●
●
●●
● ●●
●●
●
●
●●
●●
●
●
●●
●
●●●
●
●●
●
●
●●●
●
●
●●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●
●●
●
●
●
● ●
● ●
●●
●
●●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●●
●●●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●●
●●
●
●●
●
●●
●●
●●
●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●●●
●
●
● ●
●
●●
●●
●
●●
●
●
●●●
●
●●
●
●●
●●
●
● ●
●●
●
●●
●
●●
●
●●
●●
●●●●●
●
●●
●●
●●
●
●
●
●●
●●
●
●●
●●
●
●●●
●
●●
●
● ●
●
●
●
●●
●●
●
●●
●●
●
●
●●●● ●
●●
●
●
●●
●
●●
●
●
●
●●
●●
●
●●●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●●
●
●●
●●
●●
●●
●●
●
●
●●
●●
●●
●●
●
●●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●●
●
●
●●
●
●
●●
●●●●
●
●
●●
●
●●
●●
●●
●
●
●●
●
●●
●
●
●●
●●●
●
●
●●
●
●●
●
●●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●●
● ●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●●
●●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●●●
●
●
●●
●
●
●●
●●
●
●
●●
●●
●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●●
●
●●
●
●●
●
●
●
●●
● ●
●
●
●
●●
●
●●●
●
●
●
●●
●
●
●●
●
●
●
●●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●
●
●●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●● ●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●
●●
●●
●
●●
●
●
●●
●
●
●
●●
●●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
● ●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●●
●●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
● ●
●
●
●
●
●●●
●
●
●
●●
●
●
●●
●●
●
●
●
●●
●●●
●
●
●
●
●●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●●
●
●
●
●●
●
●●
●
●
●● ●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
● ●
●●
●
●
●
●●
●
●
●●
●
●●●
●●●●●
●●
●
●●● ●●●
●
●●●
●
●●●
●●●
●●
●●
●
●●
● ●
● ●●●●
●●
●●●●
●●
●●●●
●●● ●
●● ●
●●● ●● ●
●
●●
● ●
●●●
●●
●● ●
●●●
● ●
● ●
●
●●●●
●● ● ●
●●●
●
●●●●
●●
●●●
●●●
●●●
●
● ●●
●●●
●●
● ●
●●
●●●
●
● ●
●
●●●●
●
● ●●
● ●
●
●●
●●
●
●●
●
●●●
● ●●●
●●
●●
●●
●●
●●●
● ●
●● ●●
●●
●
●●
●●
●●●●● ●
●●
●●
●●
●●
●●
●●●●
●
●●
●

0
1 2 3 X

Chromosome

qtscore($(h2, pgres), erfs, )

− log10(P − value)

●
0 1 2 3 4

● ●
● ●
● ●
● ● ●
● ● ● ●
● ●● ● ● ●
● ● ●● ● ●
● ● ● ●● ●● ● ● ● ●●
● ●●● ● ●● ●● ●● ● ●●● ● ● ●●●●
● ● ●●●● ●●● ● ● ●● ●
●
● ●● ●● ● ● ●●●●●●● ● ●● ●●● ●● ●●● ● ● ●● ●● ● ● ●● ●●● ●
●
●●
●●
●●●●
●● ●
●
●● ● ●●
●
● ●●
●
●●
●
●
●●●
●
●
●
●●●
● ●●
● ●●●
● ●
●● ● ●
●●●
●●●
●
●●●
● ●●●●
●
● ●●●
●● ●
●
●●
●
●
●
●
●●
●
●
●
●● ●●
● ●●
●●●● ●●●●●●●●●●
● ●●●
● ●●●●●
●
● ●●
● ●●
●●
●●
●
●
●●
● ●●
●
●
●
●
●
●● ●●
●
●
●●
● ●
●●●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●●
●●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●●●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●●●
●
●●
●●●
●
●●
●
●
●
●●●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●●
●●
●
●
●●
●
●●
●
●
●
●●
●
●●
●●
●
●
●●
●●
●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
●●
●
●●●
●
●
●
●●
●
●
●●
●
●●
●●
●●●
●
●●●
●
●●
●
●●
●
●●
●
●●
●●●
●
●
●
●●
●
●●●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●●●●●
●
●
●●
●●
● ●
●
●●
●
●●
● ●
●
●●
●
● ●
●●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●●
●
●●●
●
●●
●●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●●
●●
●●
●
●●
●
●
●●
●
●●
●
●●●
●
●●
●
●
●●
●
●●
● ●
●
●●
●●
●
●●
●●
●●
●
●
●
●●
●●
●
●●
●
●●
●●
●●
●●
●
●●
●●
●
●
●●
●
●●
●
●
●●
●●●
●
●●● ●
●
●
●●
●
●
●●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●
●●●
●
●
●
●●
●●
●
●
●
●●
●●
●
●●
●●
●●
●●
●●
●●
●●
●
●
●●●
●●
●
●
●
●
●●
●●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
● ●
●
●
●●●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●●●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●●
●
●●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
● ●
●
●●
●●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●●
● ●
●
●●
●
●
●●
●
●
●●
●
●●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
● ●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
● ●
●
●
●
●●
●
●●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
● ●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●● ●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●●
●
●●
●●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●●
●
●
●●
●●
●
●●
●●
●
●●●
●
●●
●
●●
●
●●
●●
●
●
●●
●●
● ●
●
●●
●●
●●
●●
●●
●●
●
●●
●●
●
●
●●
●
●●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●●
● ●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●
●●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●● ●
●
●
●
●●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●

1 2 3 X

Chromosome

qtscore(qtbas, erfs)
− log10(P − value)

●
4

●
● ●●
● ● ● ●
● ●
3

●
● ● ● ● ●● ●
●● ●● ● ●● ●
● ●
● ●● ● ●● ● ●●● ● ● ●● ●
●● ●● ● ● ● ●● ●
● ●●
●● ● ●●
2

●● ●● ●● ● ● ●●● ●●● ● ● ●● ●●●● ●● ●●

● ●●● ● ● ●●● ● ●●●●●●
● ● ● ● ●●●●● ●● ●●
●
●●
● ● ●●●●●●
●
●● ● ●
● ●● ● ●● ● ●●●● ●● ● ●
● ●●●●● ●● ●
●
●● ●
● ●●●●●
●● ●
●
●●● ● ● ● ●●●
● ●● ●●●●●●●
●●●●● ●
● ●● ● ●●●●●●●●●●
●●●
●
●● ●●● ●
● ●● ●
● ●
●●●●●●●● ● ●●●●● ● ●●●
●●●●● ● ●●
●● ●●●●●● ● ●●
●● ●●
●
● ●● ●
● ●● ●● ●
● ●●●●● ●●
●
●●
● ●
●●
●●
●
● ●●
●
●
●● ●●
●●●● ●
●●●●
●●
●●
●●
●●
●●●●
●
●●
●●●●● ●●● ●
●● ●●
●
●
●●●
●●
●●●
●
●
●●●
● ●●●●
●●●●
●
●
● ● ●●●
●
●
● ● ●●
●
●●
●
●
●●●●
●
●● ●
●
●
●●●
●
●● ● ●●●● ●
● ●●●
●●● ● ●●
● ●
● ●
●●●
●●
●●
●●●
●●
●
●
●●
● ●● ●●
●●
● ●●
●
● ●●
●● ●● ●
●●
●● ●●●●●●●
●
●●●
●
● ●●● ● ● ● ● ●● ● ● ●●● ●● ●●● ● ● ●● ●
●●
● ●● ● ● ●● ●
● ●●● ●
● ●●
●
●●● ● ● ●● ●● ●●●●
● ● ●● ●●● ● ● ●
●●●● ● ● ● ●● ●● ● ●● ● ●● ● ● ● ●
1

●
●● ● ●●● ●●●
●●●
● ●●
●
● ●
●●
● ●●
●● ● ● ●
●●● ●
●●
● ●● ●●
● ●●●● ● ●
● ●● ●●●
● ●●●● ●
●
● ● ● ●●
●●●
●● ●● ● ●
●● ●● ● ●●
● ●● ● ●
●● ●
● ●●●●●
●●●●●● ●
● ●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●● ●
●
●●
●
●
●●
●
●●
●
●●●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
● ●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●●
●●
●
●
●●
●●
●
●●
●●
●
●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●●●
●
●
●●
●
●
●
●
●●●
●●
●●●
●●
●
●●
●
●
●●
●
●●
●●
●●●
●
●●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●●
●
●
●●
●●
●
●
●●
●
●●
●●
●●
●●
●●
●
●●
●
●●
●
●
●●
●●
●
●●
●
●
●●
●●
●●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
● ●
●● ●●
●
●
●
●●
●
●●
●
●
●
●●
●●
● ●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●● ●●
●●
●
●●
●
●
●●
●
●●
●●
●
●●
●
●●
●●
●
●
●●
●●
●
●
●
●●
●●
●
●
●●
●
●●
●●
●●
●●
●●
●
●●
●●●
●
●
●
●● ●
●
●●●
●●
●
●
●●
●●●
●
●
●●
●
●●
●●
●●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●
●●
●●●
●●
●
●●
●●
●●
●
●●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
● ●
●
●●
●●
●
●●●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●●●
●
●
●
●●
●●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●●●
●
●
●
●●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●●
●●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●●
●●
●
●●
●
●
●
●●
●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●
●●
●●
●
●●
●
●●●
●●
●
●●
●
●●
●
●
●
●●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
● ●
●●
●
●
●
●●
●
●●
●●
●
●●
●
●●
●●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●
●●
●●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●●
●●
●
●●
●●
●●
●●
●
●●
●
●
●
●●
●●
●●
●
●
●●
●●
●
●●
●
●
●●●
●
●●
●
●●
●
●●
●●
●
●
●
●
●●
●●
●
●
●
●●
●
●●
●
●●
●
●● ●
●
●
●●
●●
●
●
●
●●
●●
●
●
●●
●●
●
●●
●
●
●
●●
●●
●●
●
●
●●
●●●
●
●●
●
●
●●
●
●
●●●
●
●
●
●
●●
●●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
● ●
●
●●
●
●
●●
●
●
●●
●
●●
●●
●
●
●●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●●●
●●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●● ●
●●
●
●
●●
●
●
●●
●●
●
●
●●●
●
●
●●
●
●●
●
●●●
●
●
●●●
●●
●●
●
●
●
●●
●●
●
●
●●●
●
●
●●
●
●
●
●●
●
●
●
●●●
●
●
●
●●●
●
●
●●
●●
●
●●
●
●●
● ●
●
●
●●●
●
●
●
●●
●
●
●
●●
●●
●
●●●●
●
●●●●
●
●
●●
●●
●
●
●
●●
● ●
●
●
●
●●
●
●
●
●●●
●●
●●
●●
●
●
●
●●●
●
●●●
●
●
●●
● ●
●
●
●
●
●●
●●
●
●
●
●●
●●
●
●
●●
● ●●
●
●
●●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●●
●
●●
● ●●
●●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
● ●
●
●●
●
●
●●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
0

1 2 3 X

Chromosome

Figure 7.6: Comparison of FASTA (red), GRAMMAS (green), and plain GC

(black).

Answer (Ex. 8) — Yes, there is significant case/control disbalance:

> [Link](strdat1,by=pop)
No([Link]=0) Mean SD No([Link]=1) Mean SD Ptt
id 408 NA NA 414 NA NA NA
sex 408 0.547 0.498 414 0.519 0.500 0.434
age 408 49.609 12.890 414 50.355 13.082 0.410
dm2 408 0.444 0.497 414 0.505 0.501 0.079
height 406 168.755 9.518 414 168.399 9.952 0.601
weight 406 84.043 23.348 414 85.685 24.093 0.322
diet 408 0.047 0.211 414 0.043 0.204 0.831
bmi 406 29.421 7.574 414 30.198 8.285 0.161
quat 406 168.755 9.518 414 168.399 9.952 0.601
bint 408 0.444 0.497 414 0.505 0.501 0.079
age2 408 2626.745 1322.704 414 2706.340 1331.713 0.390
Pkw Pexact
id NA NA
sex 0.434 0.443
age 0.313 NA
dm2 0.079 0.081
height 0.559 NA
weight 0.325 NA
7.5. ANSWERS TO EXERCISES 191

diet 0.831 0.868

bmi 0.231 NA
quat 0.559 NA
bint 0.079 0.081
age2 0.313 NA

Answer (Ex. 9) — MMSCORE corrects best for stratification in terms of

minimal residual inflation:
> sa <- qtscore(quat~sex+age+age2,strdat1,strat=pop)
> lambda(sa)$est
[1] 1.30266
> es <- egscore(quat~sex+age+age2,strdat1,kin=gkin)
> lambda(es)$est
[1] 1.332255
> pcs <- cmdscale([Link](0.5-gkin),k=10)
> pc <- qtscore(quat~sex+age+age2+pcs[,1]+pcs[,2]+pcs[,3],strdat1,strat=pop)
> lambda(pc)$est
[1] 1.303465
> h2an <- polygenic(quat~sex+age+age2,data=strdat1,kin=gkin)
> h2an$h2an
$minimum
[1] 4053.123

$estimate
[1] 168.305955085 12.169348264 -0.042230533 -0.001535005 0.100481616
[6] 51.749954932

$gradient
[1] 0.018562837 0.441838116 0.225665709 5.919720007 0.009112349 0.090638959

$code
[1] 4

$iterations
[1] 100
> mm <- mmscore(h2an,strdat1)
> lambda(mm)$est
[1] 1.120987

Answer (Ex. 10) — The best results are achieved with mmscore:
> summary(mm)
Summary for top 10 results, sorted by P1df
Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
192CHAPTER 7. GWA IN PRESENCE OF GENETIC STRATIFICATION: PRACTICE

rs3436694 2 8921418 - C G 812 -4.624369 0.6359106 52.88261

rs70099 2 8857747 + C A 810 -5.192696 0.7278379 50.89986
rs3074653 2 8915495 - G C 813 -3.453183 0.5282050 42.73997
rs1801282 2 8931192 + C T 809 -2.989985 0.5435349 30.26099
rs3175719 2 8950441 + A C 812 -2.656511 0.5556511 22.85698
rs6392986 2 8935924 + A T 810 1.743557 0.4232678 16.96844
rs7217010 1 2495320 + C A 812 4.400790 1.0719972 16.85288
rs4277955 2 7088955 - C T 806 -1.930872 0.4910079 15.46430
rs645280 2 7087627 + C A 815 1.652257 0.4471302 13.65486
rs1351516 2 8602074 + G A 815 -2.158293 0.5853190 13.59676
P1df Pc1df effAB effBB chi2.2df P2df
rs3436694 3.540899e-13 6.492229e-12 NA NA 0 NA
rs70099 9.720075e-13 1.601211e-11 NA NA 0 NA
rs3074653 6.252160e-11 6.628397e-10 NA NA 0 NA
rs1801282 3.776445e-08 2.039882e-07 NA NA 0 NA
rs3175719 1.745141e-06 6.315750e-06 NA NA 0 NA
rs6392986 3.800639e-05 9.998193e-05 NA NA 0 NA
rs7217010 4.039208e-05 1.055941e-04 NA NA 0 NA
rs4277955 8.407854e-05 2.038515e-04 NA NA 0 NA
rs645280 2.196720e-04 4.827615e-04 NA NA 0 NA
rs1351516 2.265757e-04 4.963642e-04 NA NA 0 NA
Chapter 8

Imperfect knowledge about

genotypes

This chapter is mostly copy-pasted from the ProbABEL-package manual – the

user is encouraged to check it directly for the latest version1 .

8.1 Motivation
Many statistical and experimental techniques, such as imputations and high-
throughput sequencing, generate data which are informative for genome-wide
association analysis and are probabilistic in the nature.
When we work with directly genotyped markers using such techniques as
SNP or microsatellite typing, we would normally know the genotype of a par-
ticular person at a particular locus with very high degree of confidence, and, in
case of biallelic marker, can state whether genotype is AA, AB or BB.
On the contrary, when dealing with imputed or high-throughput sequencing
data, for many of the genomic loci we are quite uncertain about the genotypic
status of the person. Instead of dealing with known genotypes we work with
a probability distribution that is based on observed information, and we have
estimates that true underlying genotype is either AA, AB or BB. The degree
of confidence about the real status is measured with the probability distribution
{P (AA), P (AB), P (BB)}.
Several techniques may be applied to analyse such data. The most sim-
plistic approach would be to pick up the genotype with highest probability,
i.e. maxg [P (g = AA), P (g = AB), P (g = BB)] and then analyse the data as
if directly typed markers were used. The disadvantage of this approach is that
it does not take into account the probability distribution – i.e. the uncertainty
about the true genotypic status. Such analysis is statistically wrong: the esti-
mates of association parameters (regression coefficients, odds or hazard ratios,
etc.) are biased, and the bias becomes more pronounced with greater probability
distribution uncertainty (entropy).
One of the solutions that generate unbiased estimates of association param-
eters and takes the probability distribution into account is achieved by perform-
1 The ProbABEL-package manual can be found at [Link]

193
194 CHAPTER 8. IMPERFECT KNOWLEDGE ABOUT GENOTYPES

ing association analysis by means of regression of the outcome of interest onto

estimated genotypic probabilities.
The ProbABEL-package package was designed to perform such regression in
a fast, memory-efficient and consequently genome-wide feasible manner. Cur-
rently, ProbABEL-package implements linear, logistic regression, and Cox pro-
portional hazards models. The corresponding analysis programs are called
palinear, palogist, and pacoxph.

8.2 Input files

ProbABEL-package takes three files as input: a file containing SNP information
(e.g. the MLINFO file of MACH), a file with genome- or chromosome-wide
predictor information (e.g. the MLDOSE or MLPROB file of MACH), and a
file containing the phenotype of interest and covariates.
Optionally, the map information can be supplied (e.g. the ”legend” files of
HapMap).
The dose/probability file may be supplied in filevector format in which
case ProbABEL-package will operate much faster, and in low-RAM mode (ap-
prox. ≈ 128 MB). See the R libraries GenABEL-package and DatABEL-package
on how to convert MACH and IMPUTE files to filevector format (functions:
mach2databel() and impute2databel(), respectively).

8.2.1 SNP information file

In the simplest scenario, the SNP information file is an MLINFO file generated
by MACH. This must be a space or tab-delimited file containing SNP name,
coding for allele 1 and 2 (e.g. A, T, G or C), frequency of allele 1, minor allele
frequency and two quality metrics (“Quality”, the average maximum posterior
probability and “Rsq”, the proportion of variance decrease after imputations).
Actually, for ProbABEL-package, it does not matter what is written in this
file – this information is just brought forward to the output. However, it is
critical that the number of columns is seven and the number of lines in the file
is equal to the number of SNPs in the corresponding DOSE file (plus one for
the header line).
The example of SNP information file content follows here (also to be found
in ProbABEL/examples/[Link])
Note that header line is present in the file. The file describes five SNPs.

8.2.2 Genomic predictor file

Again, in the simplest scenario this is an MLDOSE or MLPROB file gener-
ated by MACH. Such file starts with two special columns plus, for each of the
SNPs under consideration, a column containing the estimated allele 1 dose (ML-
DOSE). In an MLPROB file, two columns for each SNP correspond to posterior
probability that person has two (PA1 A1 ) or one (PA1 A2 ) copies of allele 1. The
first “special” column is made of the sequential id, followed by an arrow followed
by study ID (the one specified in the MACH input files). The second column
contains the method keyword (e.g. “MLDOSE”).
8.2. INPUT FILES 195

An example of the few first lines of an MLDOSE file for five SNPs described
in SNP information file follows here (also to be found in the file ProbABEL/examples/[Link])
The order of SNPs in the SNP information file and DOSE-file must
be the same. This should be the case if you just used MACH outputs.
Therefore, by all means, the number of columns in the genomic predictor file
must be the same as the number of lines in the SNP information file plus one.
The dose/probability file may be supplied in filevector format (.fvi and
.fvd files) in which case ProbABEL will operate much faster, and in low-RAM
mode (approx. 128 MB). On the command line simply specify the .fvi file as
argument for the --dose option (cf. section 8.3 for more information on the
options accepted by ProbABEL). See the R libraries GenABEL and DatABEL
on how to convert MACH and IMPUTE files to filevector format (functions:
mach2databel() and impute2databel(), respectively).

8.2.3 Phenotypic file

The phenotypic data file contains phenotypic data, but also specifies the analysis
model. There is a header line, specifying the variable names. The first column
should contain personal study IDs. It is assumed that both the total number
and the order of these IDs are exactly the same as in the genomic
predictor (MLDOSE) file described in previous section. This is not
difficult to arrange using e.g. R; an example is given in the ProbABEL/examples
directory.
Missing data should be coded with ’NA’, ’N’ or ’NaN’ codes. Any
other coding will be converted to some number which will be used in analysis!
E.g. coding missing as ’-999.9’ will result in an analysis which will consider
-999.9 as indeed a true measurements of the trait/covariates.
In the case of linear or logistic regression (programs palinear and palogist,
respectively), the second column specifies the trait under analysis, while the
third, fourth, etc. provide information on covariates to be included into analysis.
An example few lines of phenotypic information file designed for linear regression
analysis follow here (also to be found in ProbABEL/examples/[Link])
Note again that the order of IDs is the same between the MLDOSE file and
the phenotypic data file. The model specified by this file is height ∼ µ+sex+age,
where µ is the intercept.
Clearly, you can for example include sex × age interaction terms by spec-
ifying another column having a product of sex and age here.
For logistic regression, it is assumed that in the second column cases are
coded as “1” and controls as “0”. An couple of example lines of a phenotypic
information file designed for logistic regression analysis follow here (also to be
found in ProbABEL/examples/logist [Link])
You can see that in the first 10 people, there are three cases, as indicated
by ”chd” equal to one. The model specified by this file is chd ∼ µ + sex + age +
othercov.
In case of the Cox proportional hazards model, the composition of the
phenotypic input file is a bit different. In the second column and third col-
umn, you need to specify the outcome in terms of follow-up time (column
two) and event (column three, “1” if an event occurred and zero if censored).
Columns starting from four (inclusive) specify covariates to be included into
the analysis. An example few lines of a phenotypic information file designed for
196 CHAPTER 8. IMPERFECT KNOWLEDGE ABOUT GENOTYPES

the Cox proportional hazards model analysis follow here (also to be found in
ProbABEL/examples/coxph [Link])
You can see that for the first ten people, the event occurs for three of them,
while for the other seven there is no event during the follow-up time, as indicated
by the “chd” column. Follow-up time is specified in the preceding column. The
covariates included into the model are age (presumably at baseline), sex and
“othercov”; thus the model, in terms of R/survival is
Surv(fuptime chd, chd) ∼ sex + age + othercov.

8.2.4 Optional map file

If you would like map information (e.g. base pair position) to be included in
your outputs, you can supply a map file. These follow HapMap ”legend” file
format. For example, for the five SNPs we considered the map-file may look
like (example can be found in ProbABEL/examples/[Link])
The order of the SNPs in the map file should follow that in the SNP infor-
mation file. Only information from the second column – the SNP location – is
actually used to generate the output.

8.3 Running an analysis

To run linear regression, you should use the program called palinear; for logistic
analysis use palogist, and for the Cox proportional hazards model use pacoxph
(all are found in the ProbABEL/bin/ directory after you have compiled the
program).
There are in total 11 command line options you can specify to the ProbABEL-package
analysis functions palinear or palogist. If you run either program without
any argument, you will get a short explanation to command line options:
user@server:~$ palogist

Usage: ../bin/palogist options

Options:
--pheno : phenotype file name
--info : information (e.g. MLINFO) file name
--dose : predictor (e.g. MLDOSE/MLPROB) file name
--map : [optional] map file name
--nids : [optional] number of people to analyse
--chrom : [optional] chromosome (to be passed to output)
--out : [optional] output file name (default is [Link])
--skipd : [optional] how many columns to skip in predictor
(dose/prob) file (default 2)
--ntraits : [optional] how many traits are analysed (default 1)
--ngpreds : [optional] how many predictor columns per marker
(default 1 = MLDOSE; else use 2 for MLPROB)
--separat : [optional] character to separate fields (default is space)
--score : use score test
--no-head : do not report header line
--allcov : report estimates for all covariates (large outputs!)
8.3. RUNNING AN ANALYSIS 197

--interaction : which covariate to use for interaction with SNP

(default is no ineraction, 0)
--mmscore : score test for association between a trait and genetic
polymorphism, in samples of related individuals
--robust : report robust (aka sandwich, aka Hubert-White) standard
errors
--help : print help

8.3.1 Basic analysis options

However, for a simple run only three options are mandatory, which specify the
necessary files needed to run the regression analysis.
These options are --dose (or -d), specifying the genomic predictor/MLDOSE
file described in sub-section 8.2.2; --pheno (or -p), specifying the phenotypic
data file described in sub-section 8.2.3; and --info (or -i), specifying the SNP
information file described in sub-section 8.2.1.
If you change to the ProbABEL/examples directory you can run an analysis
of height by running

user@server:~/ProbABEL/examples/$ ../bin/palinear -p [Link]

-d [Link] -i [Link]

Output from the analysis will be directed to the [Link] file.

The analysis of a binary trait (e.g. chd) can be run with

user@server:~/ProbABEL/examples/$ ../bin/palogist -p logist_data.txt

-d [Link] -i [Link]

To run a Cox proportional hazards model, try

user@server:~/ProbABEL/examples/$ ../bin/pacoxph -p coxph_data.txt

-d [Link] -i [Link]

Please have a look at the shell script files example [Link], example [Link]
and example [Link] to have a better overview of the analysis options.
To run an analysis with MLPROB files, you need specify the MLPROB file
with the -d option and also specify that there are two genetic predictors per
SNP, e.g. you can run linear model with

user@server:~/ProbABEL/examples/$ ../bin/palinear -p [Link]

-d [Link] -i [Link]
--ngpreds=2

8.3.2 Advanced analysis options

The option --interaction allows you to include interaction between SNPs and
any covariate. If for example your model is

trait ∼ sex + age + SNP,

running the program with the option --interaction=2 will model

trait ∼ sex + age + SNP + age × SNP.

198 CHAPTER 8. IMPERFECT KNOWLEDGE ABOUT GENOTYPES

The option --robust allows you to compute so-called “robust” (a.k.a. “sand-
wich”, a.k.a. Hubert-White) standard errors (cf. section 8.7 “Methodology” for
details).
With the option --mmscore a score test for association between a trait and
genetic polymorphisms in samples of related individuals is performed. A file with
the inverse of the variance-covarince matrix goes as input parameter with that
option, e.g. --mmscore <filename>. The file has to contain the first column
with id names exactly like in phenotype file, BUT OMITTING people with
no measured phenotype. The rest is a matrix. The phenotype file in case of
using the --mmscore argument may contain any amount of covariates (this is
different from previous versions). The first column contains id names, the second
the trait. The others are covariates.
An example of how a polygenic object estimated by GenABEL-package can
be used with ProbABEL is provided in ProbABEL/examples/mmscore.R
Though technically --mmscore allows for inclusion of multiple covariates,
these should be kept to minimum as this is a score test. We suggest that any
covariates explaining an essential proportion of variance should be fit as part of
GenABEL-package’s polygenic procedure.

8.3.3 Running multiple analyses at once: [Link]

The Perl script bin/[Link] example represents a handy wraper for ProbABEL-package
functions. To start using it the configuration file bin/probabel [Link] example
needs to be edited. The configuration file consists of five columns. Each column
except the first is a pattern for files produced by MACH (imputation software).
The column named “cohort” is an identifying name of a population (“ERGO” in
this example), the column “mlinfo path” is the full path to mlinfo files, includ-
ing a pattern where the chromosome number has been replaced by . chr . .
The columns “mldose path”, “mlprobe path” and “legend path” are paths and
patterns for “mldose”, “mlprob” and “legend” files, respectively. These also
need to include the pattern for the chromosome as used in the column for the
“mlinfo” files. Probably you also have to change the variable $config in the
script to point to the full path of the configuration file and the variable @anprog
to point full path to the ProbABEL-package scripts.

8.4 Output file format

Let us consider what comes out of the linear regression analysis described in
the previous section. After the analysis has run, in the output file you will find
something like
Here, only the first three lines of output have been shown. Note that lines
starting with +> are actually the ones continuing the previous line – they have
just been wrapped so we can see these long lines.
The header provides a short description of what can be found in a specific
column. The first column provides the SNP name and next six are descrip-
tions which were taken directly from the SNP information file. Therefore, these
describe allele frequencies and the quality in your total imputations, not neces-
sarily in the data under analysis.
8.5. PREPARING INPUT FILES 199

In contrast, starting with the next column, named n, the output concerns the
data analysed. Column 8 (n) tells the number of subjects for whom complete
phenotypic information was available. At this point, unless you have complete
measurements on all subjects, you should feel alarmed if the number here is
exactly the number of people in the file – this may indicate you did not code
missing values according to ProbABEL-package format (’NA’, ’NaN’, or ’N’).
The next column, nine (“Mean predictor allele”), gives the estimated fre-
quency of the predictor allele (A1) in subjects with complete phenotypic data.
If the --chrom option was used, in the next column you will find the value
specified by this option. If --map option was used, in the subsequent column
you will find map location taken from the map-file. The subsequent columns
provide coefficients of regression of the phenotype onto genotype, corresponding
standard errors, and Wald χ2 test value.

8.5 Preparing input files

In the ProbABEL/bin directory you can find the prepare data.R file – an R
script that arranges phenotypic data in right format. Please read this script for
details.

8.6 Memory use and performance

Maximum likelihood regression is implemented in ProbABEL-package. With
6,000 people and 2.5 millions SNPs, a genome-wide scan is completed in less
that an hour for a linear model with 1-2 covariates and overnight for logistic
regression or the Cox proportional hazards model (figures for a PC bought back
in 2007).
Memory may be an issue with ProbABEL-package if you use MACH text
dose/probability files, e.g. for large chromosomes, such as chromosome one
consumed up to 5 GB of RAM with 6,000 people.
We suggest that dose/probability file is to be supplied in filevector format
in which case ProbABEL-package will operate about 2-3 times faster, and in
low-RAM mode (approx. 128 MB). See the R libraries GenABEL-package and
DatABEL-package on how to convert MACH and IMPUTE files to filevector
format (functions: mach2databel() and impute2databel(), respectively).
When ’–mmscore’ option is used, the analysis may take quite some time.

8.7 Methodology
8.7.1 Analysis of population-based data
Linear regression assuming normal distribution
Standard linear regression theory is used to estimate coefficients of regression
and their standard errors. We assume a linear model with expectation

E[Y] = X β (8.1)
200 CHAPTER 8. IMPERFECT KNOWLEDGE ABOUT GENOTYPES

and variance-covariance matrix

V = σ 2 I,

where Y is the vector of phenotypes of interest, X is the design matrix, β is the

vector of regression parameters, σ 2 is the variance and I is the identity matrix.
The maximum likelihood estimates (MLEs) for the regression parameters
are given by
β̂ = (XT X)−1 XT Y (8.2)
and the MLE of the residual variance is
(Y − Xβ̂)T (Y − Xβ̂)
σ̂ 2 = , (8.3)
N − rX
where N is the number of observations and rX is the rank of X (i.e. the number
of columns of the design matrix).
The variance-covariance matrix for the parameter estimates under alterna-
tive hypothesis can be computed as

varβ̂ = σ̂ 2 (XT X)−1 . (8.4)

For the j-the element β̂(j) of the vector of estimates the standard error
under the alternative hypothesis is given by the square root of the corresponding
diagonal element of the above matrix, varβ̂ (jj), and the Wald test can be
computed with
β̂(j)2
T 2 (j) = ,
varβ̂ (jj)
which asymptotically follows the χ2 distribution with one degree of freedom
under the null hypothesis.
When testing significance for more than one parameter simultaneously, sev-
eral alternatives are available. Let us first partition the vector of parameters
into two components, β = (βg , βx ), and our interest is testing the parameters
contained in βg (SNP effects), while βx (e.g. effects of sex, age, etc.) are consid-
ered nuisance parameters. Let us define the vector of the parameters of interest
which are fixed to certain values under the null hypothesis as βg,0 .
Firstly, the likelihood ratio test can be obtained with
 
LRT = 2 logLik(β̂g , β̂x ) − logLik(βg,0 , β̂x ) ,

which under the null hypothesis is asymptotically distributed as χ2 with the

number of degrees of freedom equal to the number of parameters specified by
βg . Assuming the normal distribution, the log-likelihood of a model specified
by the vector of parameters β and residual variance σ 2 can be computed as
1
logLik(β, σ 2 ) = − N · loge σ 2 + (Y − βX)T (I/σ 2 )(Y − βX) .


2
Secondly, the Wald test can be used; for that the inverse variance-covariance
matrix of β̂g should be computed as
 −1
var−1
β̂
= var −1
β̂
(g, g) − var −1
β̂
(g, x) var −1
β̂
(x, x) var−1
β̂
(x, g),
g
8.7. METHODOLOGY 201

where var−1
β̂
(a, b) correspond to sub-matrices of the inverse of the variance-
covariance matrix of β̂, involving either only parameters of interest (g, g), nui-
sance parameters (x, x) or combination of these (x, g), (g, x).
The Wald test statistics is then computed as

W 2 = (β̂g − βg,0 )T var−1

β̂
(β̂g − βg,0 ),
g

which asymptotically follows the χ2 distribution with the number of degrees

of freedom equal to the number of parameters specified by βg . The Wald test
generally is computationally easier than the LRT, because it avoids estimation
of the model specified by the parameter’s vector (βg,0 , β̂x ).
Lastly, similar to the Wald test, the score test can be performed by use of
var(βg,0 ,β̂x ) instead of varβ̂ .

Logistic regression
For logistic regression, the procedure to obtain parameters estimates, their
variance-covariance matrix, and tests are similar to these outlined above with
several modifications.
The expectation of the binary trait is defined as the expected probability of
the event as defined by the logistic function
1
E[Y] = π = .
1 + e−(Xβ)
The estimates of the parameters are obtained not in one step, as is the case of
the linear model, but using an iterative procedure (iteratively re-weighted least
squares). This procedure is not described here for the sake of brevity.
The log-likelihood of the data is computed using the binomial probability
formula:
logLik(β) = YT loge π + (1 − Y)T loge (1 − π),
where loge π is a vector obtained by taking the natural logarithm of every value
contained in the vector π.

Robust variance-covariance matrix of parameter estimates

For a linear model, these are computed using formula

varr = (XT X)−1 (XT RX)(XT X)−1 ,

where R is a diagonal matrix containing squares of residuals of Y. The same

formula may be used for “standard” analysis, in which case the elements of the
R matrix are constant, namely mean residual sum of squares (the estimate of
σ 2 ).
Similar to that, the robust matrix is computed for logistic regression with

varr = (XT WX)−1 (XT RX)(XT WX)−1 ,

where 1 is the vector of ones and W is the diagonal matrix of ”weights” used
in logistic regression.
202 CHAPTER 8. IMPERFECT KNOWLEDGE ABOUT GENOTYPES

Cox proportional hazards model

The implementation of the Cox proportional hazard model used in ProbABEL-package
is entirely based on the code of R library survival developed by Thomas Lumley
(function coxfit2), and is therefore not described here.
Many thanks to Thomas for making his code available under GNU GPL!

8.7.2 Analysis of pedigree data

The framework for analysis of pedigree data follows the two-step logic developed
in the works of Aulchenko et al. (2007) and Chen and Abecasis (2007). General
analysis model is a linear mixed model which defines the expectation of the trait
as
E[Y] = Xβ,
identical to that defined for linear model (cf. section 8.1). To account for cor-
relations between the phenotypes of relatives which may be induced by family
relations the variance-covariance matrix is defined to be proportional to the
linear combination of the identity matrix I and the relationship matrix Φ:
Vσ2 ,h2 = σ 2 2h2 Φ + (1 − h2 )I ,

where h2 is the heritability of the trait. The relationship matrix Φ is twice

the matrix containing the coefficients of kinship between all pairs of individuals
under consideration; its estimation is discussed in a separate section ”8.7.2”
(8.7.2).
Estimation of a model defined in such a way is possible by numerical max-
imization of the likelihood function, however, the estimation of this model for
large pedigrees is laborious, and is not computationally feasible for hundreds of
thousands to millions of SNPs to be tested in the context of GWAS, as we have
demonstrated previously (Aulchenko et al., 2007).

Two-step score test for association

A two-step score test approach is therefore used to decrease the computational
burden. Let us first re-define the expectation of the trait by splitting the design
matrix in two parts, the ”base” part Xx , which includes all terms not changing
across all SNP models fit in GWAS (e.g. effects of sex, age, etc.), and the part
including SNP information, Xg :
E[Y] = Xx βx + Xg βg .
Note that the latter design matrix may include not only the main SNP effect,
but e.g. SNP by environment interaction terms.
At the first step, a linear mixed model not including SNP effects
E[Y] = Xx βx
is fitted. The maximum likelihood estimates (MLEs) of the model parameters
(regression coefficients for the fixed effects β̂x , the residual variance σ̂x2 and the
heritability ĥ2x ) can be obtained by numerical maximization of the likelihood
function
1 
logLik(βx , h2 , σ 2 ) = − loge |Vσ2 ,h2 | + (Y − βx Xx )T Vσ−1
2 ,h2 (Y − β X
x x ) ,
2
8.7. METHODOLOGY 203

where Vσ−12 ,h2 is the inverse and |Vσ 2 ,h2 | is the determinant of the variance-

covariance matrix.
At the second step, the unbiased estimates of the fixed effects of the terms
involving SNP are obtained with

β̂g = (XTg Vσ̂−1

2 ,ĥ2
Xg )−1 XTg Vσ̂−1
2 ,ĥ2
Rβ̂x ,

where Vσ̂−1
2 ,ĥ2
is the variance-covariance matrix at the point of the MLE esti-
mates of ĥ2x and σ̂x2 and Rβ̂x = Y− β̂x Xx is the vector of residuals obtained from
the base regression model. Under the null model, the inverse variance-covariance
matrix of the parameter’s estimates is defined as

varβ̂g = σ̂x2 (XTg Vσ̂−1

2 ,ĥ2
Xg )−1 .

Thus the score test for joint significance of the terms involving SNP can be
obtained with
T 2 = (β̂g − βg,0 )T var−1
β̂
(β̂g − βg,0 ),
g

where βg,0 are the values of parameters fixed under the null model. This test
statistics under the null hypothesis asymptotically follows the χ2 distribution
with the number of degrees of freedom equal to the number of parameters tested.
The significance of an individual j-the elements of the vector β̂g can be tested
with
Tj2 = β̂g2 (j) varβ̂−1 (jj),
g

where β̂g2 (j) is the square of the j-th element of the vector of estimates β̂g ,
and var−1
β̂
(jj) corresponds to the j-th diagonal element of varβ̂−1 . The latter
g g

statistics asymptotically follows χ21 .

Estimation of the kinship matrix

The relationship matrix Φ used in estimation of the linear mixed model for
pedigree data is twice the matrix containing the coefficients of kinship between
all pairs of individuals under consideration. This coefficient is defined as the
probability that two gametes randomly sampled from each member of the pair
are identical-by-descent (IBD), that is they are copies of exactly the same an-
cestral allele. The expectation of kinship can be estimated from pedigree data
using standard methods, for example the kinship for two outbred sibs is 1/4, for
grandchild-grandparent is 1/8, etc. For an outbred person, the kinship coeffi-
cient is 1/2 – that is two gametes sampled from this person at random are IBD
only if the same gamete is sampled. However, if the person is inbred, there is
a chance that a maternal and paternal chromosomes are also IBD. The proba-
bility of this is characterized by kinship between individual’s parents, which is
defined as the individual’s inbreeding coefficient, F . In this case, the kinship
coefficient for the individual is F + 1/2. Similar logic applies to computation of
the kinship coefficient for other types of pairs in inbred pedigrees.
The kinship matrix can be computed using the pedigree data using standard
methods. However, in many cases, pedigree information may be absent, incom-
plete, or not reliable. Moreover, the estimates obtained using pedigree data
reflect the expectation of the kinship, while the true realization of kinship may
204 CHAPTER 8. IMPERFECT KNOWLEDGE ABOUT GENOTYPES

vary around this expectation. In presence of genomic data it may therefore be

desirable to estimate the kinship coefficient from these, and not from pedigree.
It can be demonstrated that unbiased and positive semi-definite estimator of the
kinship matrix can be obtained (Astle and Balding, 2010; Amin et al., 2007) by
computing the kinship coefficients between individuals i and j with
L
1 X (gl,i − pl )(gl,j − pl )
K̂ij =
L pl (1 − pl )
l=1

where L is the number of loci, pl is the allelic frequency at l-th locus and
gl,j is the genotype of j-th person at the l-th locus, coded as 0, 1/2, and 1,
corresponding to the homozygous, heterozygous, and other type of homozygous
genotype. The frequency is computed for the allele which, when homozygous,
corresponds to the genotype coded as “1”.

8.8 How to cite

If you used ProbABEL-package for your analysis please give a link to the GenABEL
project home page

[Link]

and cite the ProbABEL-package paper to give us some credit:

Aulchenko YS, Struchalin MV, van Duijn CM.
ProbABEL package for genome-wide association analysis of imputed
data.
BMC Bioinformatics. 2010, 11:134.
A proper reference may look like
For the analysis of imputed data, we used the ProbABEL-package
from the GenABEL suite of programs (Aulchenko et al., 2010).

If you have used the Cox proportional hazard model, please mention the
R package survival by Thomas Lumley. Additionally to the above citation,
please tell that
The Cox proportional hazards model implemented in ProbABEL-package
makes use of the source code of the R package ”survival” as imple-
mented by T. Lumley.
Chapter 9

Analysis of imputed data:

an example

In this chapter, you will perform an analysis of imputed data set. In this set of
120 individuals, 4500 SNPs are imputed based on information on 500 directly
typed SNPs. You will first analyse 500 directly typed SNPs and then proceed
to the analysis of imputed data. Finally, you will have a possibility to compare
your results to the results of analysis in case all 5000 SNPs were directly typed.

9.1 Analysis of 500 directly typed SNPs

Load the GenABEL-package library and load the data set we will use:
> library(GenABEL)
> load("RData/[Link]")
> ls()

[1] "df500" "df5k" "map5k" "rcT"

Here, df500 contains the data including 500 directly typed SNPs, and rcT
is the vector containing the value of the trait of interest:
> nids(df500)

[1] 120

> nsnps(df500)

[1] 500

> length(rcT)

[1] 120

> rcT[1:10]

[1] -0.5560057 1.1798113 0.2721125 0.2174227 -0.8193852 -1.8739403

[7] 0.9530623 0.9022811 -1.0906900 0.5575454

205
206 CHAPTER 9. ANALYSIS OF IMPUTED DATA: AN EXAMPLE

In all analysis that follow, do disregard the Genomic Control and GC-
corrected results: as we will analyse a small region with strong association,
the GC can not be applied.
Let us start with analysis of directly typed SNPs. For that, we will use
mlreg function of GenABEL-package. This function implements ML-regression
and Wald test of significance1 . This will later on allow us direct comparison with
the results of ProbABEL-package, which implements the same testing procedure.
We can run regression of rcT on SNPs region-wise using

> qts500 <- mlreg(rcT~1,df500)

> qts500[1:5,]

Chromosome Position Strand A1 A2 N effB se_effB

rs6139074 20 11244 + A C 120 0.488853649 0.2079789
rs13043000 20 13288 + G T 120 -0.002890489 0.4394487
rs6037629 20 43093 + T G 120 0.504885204 0.4040614
rs6052070 20 44931 + A G 120 0.171428327 0.3037772
rs6116135 20 46930 + G A 120 -0.239668333 0.4235491
chi2.1df P1df Pc1df effAB effBB chi2.2df P2df
rs6139074 5.524833e+00 0.0187484 0.2123725 NA NA NA NA
rs13043000 4.326395e-05 0.9947519 0.9972156 NA NA NA NA
rs6037629 1.561315e+00 0.2114728 0.5073673 NA NA NA NA
rs6052070 3.184599e-01 0.5725346 0.7646316 NA NA NA NA
rs6116135 3.201945e-01 0.5714908 0.7640104 NA NA NA NA

The summary for the SNPs, which show most significant association can be
produced with

> bestHits500 <- [Link](qts500,top=10)

Summary for top 10 results, sorted by P1df

> bestHits500

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df

rs6039167 20 846271 + G A 120 -5.9451984 0.7731855 59.12420
rs7261762 20 853448 + A G 120 -6.7335160 0.9748760 47.70732
rs511582 20 868752 + A G 120 -0.9872044 0.1883796 27.46290
rs967789 20 2876445 + A G 120 -0.9289317 0.1949364 22.70814
rs642758 20 2831146 + A G 120 -0.9194347 0.1993352 21.27521
rs676749 20 2974069 + A T 120 -0.8711208 0.1900255 21.01517
rs577116 20 2818285 + G A 120 -0.9291733 0.2032178 20.90595
rs570673 20 2819015 + C G 120 -0.8728347 0.1973862 19.55377
rs6112914 20 2048348 + A G 120 -0.6753674 0.1543434 19.14715
rs6106161 20 1995179 + T G 120 -0.6253343 0.1609140 15.10208
P1df Pc1df effAB effBB chi2.2df P2df
rs6039167 1.480274e-14 4.512105e-05 NA NA NA NA
rs7261762 4.948399e-12 2.477531e-04 NA NA NA NA
rs511582 1.601368e-07 5.429559e-03 NA NA NA NA
rs967789 1.885672e-06 1.146299e-02 NA NA NA NA
1 In general the score test may be preferred because it is faster and more robust.
9.1. ANALYSIS OF 500 DIRECTLY TYPED SNPS 207

mlreg(rcT 1, df500)
14
●
12

●
10
− log10(P − value)

●
6

●
● ●
●
● ●

●●●
4

● ● ●
● ●
● ●● ●
● ●
● ● ● ● ● ●
● ● ● ●
● ●
● ● ● ● ●
● ●
●
● ● ●● ●● ●● ● ●● ● ●
●
●●
2

● ●● ● ● ● ● ●● ●● ●●● ● ● ●
●●
●
● ● ●
●
●● ● ●● ● ●● ● ● ● ● ● ●
●
● ●● ● ●● ●● ● ● ●
●●● ●
●
● ● ●● ● ● ●●●●● ● ● ● ●● ● ●● ● ●● ● ● ● ● ●●
● ● ● ● ● ● ● ● ●● ● ● ● ●● ●
● ● ●● ●● ● ● ●●
● ● ●
● ●
●
●●●●
● ●●●●●
●●
● ●●●
● ●●
● ●
● ●
● ●● ● ● ●●
●●
●
●●●
● ●
●
●● ●
●
● ●●● ● ●●
● ●● ● ●●● ●●●
● ● ●
● ●●● ●●● ●
●●● ● ● ●
●●● ●● ●● ●●● ● ● ● ● ● ●●
●●
●●
●
●● ●●●
●● ●
●●● ●
●
●
●●
●●
●●●● ●● ● ●
●●●
● ●●●●
●●
●
●● ● ●●
●●●●●●
● ●●●
●
● ●
●●
● ● ●●●●●
●●
●●
●●
●● ● ●● ● ●
●●
● ●●
●
●
0

0e+00 1e+06 2e+06 3e+06 4e+06

Map position

Figure 9.1: Manhattan plot for 500 directly typed SNPs

rs642758 3.978434e-06 1.439751e-02 NA NA NA NA

rs676749 4.556610e-06 1.500792e-02 NA NA NA NA
rs577116 4.823944e-06 1.527219e-02 NA NA NA NA
rs570673 9.780761e-06 1.897119e-02 NA NA NA NA
rs6112914 1.210181e-05 2.025583e-02 NA NA NA NA
rs6106161 1.018509e-04 3.922563e-02 NA NA NA NA

and the plot of the results with

> plot(qts500)
> abline(h=-log10(5e-8))

(see figure 9.1).

Finally, we can produce descriptive statistics for the SNPs, which demon-
strated genome-wide significance:

> gwsSnps500 <- rownames(bestHits500)[bestHits500$P1df<=6e-8]

> gwsSnps500

[1] "rs6039167" "rs7261762"

> summary(gtdata(df500[,gwsSnps500 ]))

Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11

rs6039167 20 846271 + G A 120 1 0.012500000 117
208 CHAPTER 9. ANALYSIS OF IMPUTED DATA: AN EXAMPLE

rs7261762 20 853448 + A G 120 1 0.008333333 118

P.12 P.22 Pexact Fmax Plrt
rs6039167 3 0 1 -0.012658228 0.8454919
rs7261762 2 0 1 -0.008403361 0.8968501

Exercise 1.
It is known that rare variation in the presence of outliers can generate spurious
associations. Do you believe this is a true association in this particular case?
What you can do to check whether this is a true association or not?

9.2 Analysis of imputed data with ProbABEL-package

Here, you will analyse imputed data. In the RData directory, you will find the
necessary files: [Link] and [Link] (these files represent
mldose data produced by mach1, converted into DatABEL-package format2 ) and
[Link], which contains information for the imputed SNPs generated
in the process of imputations (such as ’Rsq’, etc.).
We will start with producing a phenotypic data file for the use with ProbABEL-package:

> [Link]([Link](id=idnames(df500), rcT=rcT),

+ file="[Link]", quote=FALSE, [Link]=FALSE)

next, try the command ’system("head [Link]")’ to check the few first lines
of the file.
At this moment, leave R (or, rather, start new console!), copy the mach-files
to the working directory with

yourname@server> cp RData/mach1* .

and run ProbABEL-package analysis with

yourname@server> palinear --pheno [Link] --info [Link] /

--dose [Link]

Do not forget to check that you start the analysis in right directory, i.e. all files
([Link], [Link], [Link], [Link]) are present
in the working directory (use the ’ls’ command from the console).
Now, you can return to R and load the analysis results:

> qtsPal <- [Link]("regression_add.[Link]", head=T, strings=F)

> qtsPal[1:5,]

name A1 A2 Freq1 MAF Quality Rsq n Mean_predictor_allele

1 rs4814683 T G 0.4921 0.4921 0.7334 0.5650 120 0.492088
2 rs6076506 T G 0.7723 0.2277 0.6511 0.2415 120 0.772271
3 rs6139074 C A 0.3384 0.3384 0.9904 0.9803 120 0.338388
4 rs1418258 T C 0.5009 0.4991 0.7208 0.5474 120 0.500917
5 rs7274499 C A 0.9745 0.0255 0.9505 0.1011 120 0.974450
2 using mach2databel
9.2. ANALYSIS OF IMPUTED DATA WITH PROBABEL-PACKAGE 209

beta_SNP_add sebeta_SNP_add chi2_SNP

1 0.618307 0.269166 5.249680
2 0.722713 0.499781 2.107910
3 0.501472 0.210870 5.617680
4 0.636760 0.273391 5.393680
5 1.348270 2.085770 0.424181

As you see, there is not P -value produced in the ProbABEL-package output,

and we need to compute it:

> qtsPal$Chisq <- (qtsPal$beta/qtsPal$sebeta)^2

> qtsPal$"P-value" <- pchisq(qtsPal$Chisq,1,low=F)
> qtsPal[1:5,]

name A1 A2 Freq1 MAF Quality Rsq n Mean_predictor_allele

1 rs4814683 T G 0.4921 0.4921 0.7334 0.5650 120 0.492088
2 rs6076506 T G 0.7723 0.2277 0.6511 0.2415 120 0.772271
3 rs6139074 C A 0.3384 0.3384 0.9904 0.9803 120 0.338388
4 rs1418258 T C 0.5009 0.4991 0.7208 0.5474 120 0.500917
5 rs7274499 C A 0.9745 0.0255 0.9505 0.1011 120 0.974450
beta_SNP_add sebeta_SNP_add chi2_SNP Chisq P-value
1 0.618307 0.269166 5.249680 5.2767671 0.02161184
2 0.722713 0.499781 2.107910 2.0910877 0.14816055
3 0.501472 0.210870 5.617680 5.6554059 0.01740165
4 0.636760 0.273391 5.393680 5.4247924 0.01985280
5 1.348270 2.085770 0.424181 0.4178505 0.51801156

Let us have a look at the top 20 associated SNPs:

> qtsPal[order(qtsPal$Chisq,decreasing=T)[1:20],]

name A1 A2 Freq1 MAF Quality Rsq n Mean_predictor_allele

1009 rs6039167 G A 0.9875 0.0125 0.9995 0.9800 120 0.987537
1022 rs7273309 C T 0.9985 0.0015 0.9970 0.1896 120 0.998550
1026 rs7265788 A G 0.9886 0.0114 0.9929 0.6884 120 0.988546
1025 rs8123328 G T 0.9902 0.0098 0.9964 0.8126 120 0.990246
1024 rs7267882 G A 0.9911 0.0089 0.9986 0.9230 120 0.991167
4009 rs566570 C T 0.5307 0.4693 0.8814 0.8074 120 0.530712
1023 rs7261762 A G 0.9913 0.0087 0.9991 0.9472 120 0.991346
2973 rs6035871 A G 0.9387 0.0613 0.8962 0.2582 120 0.938696
2969 rs6047425 A G 0.9388 0.0612 0.8964 0.2573 120 0.938767
3877 rs2325971 T C 0.5811 0.4189 0.8526 0.7807 120 0.581042
1017 rs553378 G A 0.9987 0.0013 0.9974 0.0005 120 0.998704
3880 rs873711 A G 0.5888 0.4112 0.8597 0.7934 120 0.588750
3886 rs6037425 G C 0.5871 0.4129 0.8651 0.8034 120 0.587125
3958 rs6037443 A G 0.4975 0.4975 0.8918 0.8340 120 0.497446
4027 rs6076466 T C 0.7040 0.2960 0.6570 0.3365 120 0.703937
3884 rs6051434 G T 0.5186 0.4814 0.8427 0.7732 120 0.518621
1018 rs554362 A G 0.6496 0.3504 0.5748 0.2481 120 0.649604
4032 rs2326056 C A 0.7338 0.2662 0.6422 0.2403 120 0.733767
4029 rs11697448 G A 0.7414 0.2586 0.6464 0.2370 120 0.741396
210 CHAPTER 9. ANALYSIS OF IMPUTED DATA: AN EXAMPLE

1560 rs6033304 A C 0.8234 0.1766 0.7408 0.2813 120 0.823429

beta_SNP_add sebeta_SNP_add chi2_SNP Chisq P-value
1009 5.98851 0.781065 48.5097 58.78455 1.759171e-14
1022 38.75330 5.101300 47.7785 57.71061 3.036607e-14
1026 7.39562 0.977268 47.4767 57.26929 3.800428e-14
1025 7.12447 0.984098 44.1039 52.41172 4.500298e-13
1024 6.85041 0.977608 41.7507 49.10250 2.429299e-12
4009 1.29428 0.185797 41.3364 48.52651 3.258536e-12
1023 6.80755 0.977629 41.3086 48.48792 3.323302e-12
2973 4.88622 0.721348 39.4165 45.88348 1.254997e-11
2969 4.89489 0.722872 39.3938 45.85251 1.275000e-11
3877 1.29728 0.192726 38.9951 45.30925 1.682530e-11
1017 728.01800 108.683000 38.6722 44.87051 2.105045e-11
3880 1.26975 0.192071 37.8095 43.70316 3.821541e-11
3886 1.25994 0.190625 37.7960 43.68577 3.855647e-11
3958 1.21385 0.184428 37.5238 43.31879 4.650954e-11
4027 1.98273 0.339011 30.5459 34.20579 4.958132e-09
3884 1.14070 0.199652 29.3077 32.64341 1.107140e-08
1018 2.12435 0.376830 28.6184 31.78053 1.726140e-08
4032 2.33945 0.427909 27.0940 29.88992 4.572850e-08
4029 2.36276 0.435773 26.6943 29.39805 5.893679e-08
1560 2.39452 0.451563 25.6485 28.11902 1.140790e-07

and produce the plot of the results with

> plot(map5k,-log10(qtsPal[,"P-value"]))
> abline(h=-log10(5e-8))

(here, map5k contains map information for all 5000 SNPs).

To compare with the results obtained using 500 directly typed SNPs only,
we can add the points with

> points(map(qts500),-log10(qts500[,"P1df"]),col="red",pch=19,cex=0.5)

(see figure 9.2).

Now, we can also check different characteristics of the SNPs, which are
claimed to be significant in our analysis:

> gwsSnpsImp <- qtsPal$name[which(qtsPal[,"P-value"]<=5e-8)]

> gwsSnpsImp

[1] "rs6039167" "rs553378" "rs554362" "rs7273309" "rs7261762" "rs7267882"

[7] "rs8123328" "rs7265788" "rs6047425" "rs6035871" "rs2325971" "rs873711"
[13] "rs6051434" "rs6037425" "rs6037443" "rs566570" "rs6076466" "rs2326056"

> mlinfo <- [Link]("[Link]",head=T,strings=F)

> mlinfo[mlinfo$SNP %in% gwsSnpsImp,]

SNP Al1 Al2 Freq1 MAF Quality Rsq

1009 rs6039167 G A 0.9875 0.0125 0.9995 0.9800
1017 rs553378 G A 0.9987 0.0013 0.9974 0.0005
1018 rs554362 A G 0.6496 0.3504 0.5748 0.2481
9.2. ANALYSIS OF IMPUTED DATA WITH PROBABEL-PACKAGE 211

14
●
●

●
●

●
12

●
● ●
●

● ●
●
●
● ●
−log10(qtsPal[, "P−value"])

10

●
8

●
●
●
●
●● ● ●
●
●
● ●
● ●
●
● ●
6

●●●
● ●● ●
●
● ●●
● ●
●
●
●●
● ●
●● ●
● ● ●
●
● ●
●
●
●
●
●● ●
● ●● ●● ●●
●
● ●● ● ●
● ● ●●● ●● ●
● ● ●●
● ●
●
●● ●●●
●● ● ● ● ●
● ● ●●● ● ● ●
● ● ● ●●
● ●
4

● ● ●●
●
● ● ● ● ●● ●
● ● ● ●
● ●
●●●● ●
● ●● ● ● ●● ●●● ● ●● ●
● ●
●● ●
● ●
● ●●●● ●●● ●●● ●
●
●
●●● ● ●
● ● ● ●● ●● ● ●●
● ●● ●
● ●
●●
●● ●
● ● ●
●● ● ● ● ●●●
●
●● ● ●
● ● ●
●
●●● ● ●●● ●
● ●●●●
●
●●
● ● ●●
●●
●●● ● ●●● ● ● ● ●●●
●● ●●●●
●
●● ● ●
●
●● ● ●
● ●●● ●●● ●
●●
● ● ● ●●
● ●●●
● ● ●● ●
●● ●● ●
●
●
●●● ● ●●●● ● ●● ● ●●
● ●
●●●●
●●
● ● ● ● ●
●
● ● ●● ● ● ● ● ●
● ●●
●
●
●
● ●● ● ●● ● ●
●
●●●
●●●
● ●
●● ●●●
●
●●●
● ●
●
●● ● ● ●●
●●●● ● ● ●● ● ● ●● ●●
●
● ●
● ●
●●● ● ●●● ● ●●●
● ●● ●●
●●
● ● ● ●
● ●● ●● ● ● ● ●
● ● ● ● ●●●
●
● ●● ● ●
●
●
● ●● ● ● ●● ● ● ● ● ●
●● ● ● ●
2

●
● ● ●● ● ● ●
●●
● ●● ●●
● ●
●
●●●
●
●●●
● ●●●●●
● ●●
● ●●
● ●● ● ● ● ● ● ● ● ● ●● ●● ●
● ● ●● ● ●
● ●●●●● ●● ● ● ●● ●● ●●●● ●●●
●●●● ● ●● ●
●
●●
●● ●●
●●● ●●
●●●
●● ●
●● ● ● ●● ●●● ●
●
●●
●
●
●
●
● ●●
●●●● ●● ●●● ●● ●●●
●●
●●● ●● ●● ●●● ●●●
●●●
● ●
●●
●
●●● ●● ●● ●● ●
● ● ● ●● ● ● ● ● ●●●
●●●● ●● ● ●
●
●
●
●●● ●● ● ●
●●
● ● ● ●● ● ● ●● ● ●
●
● ● ●● ● ● ●● ● ● ●
●● ●
●
● ●● ● ●●● ●
●
●
● ● ●●●●
● ●● ● ● ● ● ●●● ●● ● ●● ● ●●
● ●● ●●
● ● ● ● ●●● ● ●● ● ●●
●
● ● ●● ●
●
● ●● ●●●
● ●●●●●●
●
●
● ● ●●
●●
●●●
●● ● ●
●●●●
●
●●
● ● ●
●●●
●
●
●
●●●●● ●
● ●
●
●
●
●●●● ●●●
●
●
●
●●● ●●● ●●●● ●
● ●
●●●
●
●●● ●● ● ●
●●●
● ● ● ● ● ●●●

●●●
●
●
●●●
●●
● ●●● ●● ●●● ●
●●●●●●
●
●●●● ● ●●
●
●●
●
●●
●● ●
●●
●
●
●●
●
●
●●
●● ●
●
●●● ●
●
●●●●
● ●
●●
● ●
●
●
●●
● ●● ●●
●●●
● ●●● ●
● ●●
●
●●●
●
●●● ●
●
● ●●
●
●●
●
●●●● ●●●
●●
●●● ●
●●
●● ●● ● ●●
● ● ● ●
● ●
●
●
●●●● ●
● ●●
●
●
●●
●
●
● ● ●
●
●●
● ●
●
●
● ●●
●●
●
●●
● ●
● ●●●●●●●●● ●● ●●
● ● ● ●●● ●
● ●
● ●●
● ●
●●●
●● ●●●
●●
●●●● ● ● ● ● ● ● ● ● ●
●●
●
●●
● ●●●
● ● ●● ●● ● ●● ●●
●●●
●
●
●●● ● ● ● ●
●
●●●● ●●
●●●
●
● ●
●●●● ● ● ●●
● ●● ● ● ●● ●
●
●
●●●●
●●●●●
●
●●●
● ● ●
●
● ●● ●●
●● ●●●
●
●●
● ●
●
●●●
●●
●●●
●
●
●
●●●
●
● ●
●●
● ●●
● ●
●●●
●
●
●● ●●● ●
● ●
● ●
●●
●
●●
●
●●● ●
●●●● ●
●●
●●
●●● ●●
●
●●
●●●● ●
●● ● ●
●●
●●●●
●●
●●
●
●●● ●
●
● ●● ● ●
●
●
●●● ●● ●
● ●●● ●●●
●
●
●
●
●●●
●
● ●● ●●●● ●
●●●● ●●
●●
●
●●
●●
●
●●●
●
● ●●
●
●
● ●●●
●
●●●●●
●
● ●●●●
● ●●
●
●●● ● ●
●
●
●
●
●●●●
●
●●
●●
●
●●
●●
●
●●
●●●
● ●
●●●●
●
●
●●●
●●● ●●● ●
●●● ●
●●●
●● ●● ●
● ●
●
●
● ●● ● ●
●●
●● ●●●●
●●
●
●●●●
●●
●●●●●●
● ●●
●
●●
●
●●
●●●
●
●●
●●
●
●●
●
●●●●
●
●
●●
●●
●
●●● ●
●●●
●
●
●●● ●●● ●● ●●
●●●●
●
● ●
●●●●●●
● ●
● ●
●● ●●● ●● ●●●
●● ● ●●●
● ●
●●
●●
●●●●●●
● ● ●
●●
●
● ●●
●
● ● ● ● ● ● ● ●● ●● ●
● ●
●●●
●●●
●●●●●
●●●●●
●
●
●●●●
●●●●
●●●
●●
●
●
●
●●●
●●● ●
●●
●● ●
● ●●●●
●
●●●●
●
●●
●
●
●
●
●
● ●
●●
●
●●
●
●●
●
●
●
●
●●●
●
●●●
●●
●●
● ●● ●●●
●●●
●
●● ●
●●●
●
●●
●●
●
●●●
●● ●●
●
●●
●
●
●●
●
●
●
●●
●
●●●●
●●●
●●●
●●
●
● ●
●
●
●
●●●
●
●
●●●●
●●●
●●
● ●●●●
●●
●
●
●●●
●
●
●●●
●
●●
●
●
●●●
●
●
●●●
●
●●●
● ●
●●
●● ●● ●●
●●●
●
● ● ● ●
●●● ●
●●
●
●
●●●●
●●
●
●●
●
●
●
●
●● ● ● ●
●● ●●●
●●●
●
●●
●● ●
●
●●
●
●
● ●
● ●
●●●
●●●
●●●
●
●●●
●
●●
●●●
●
●●
● ●
●●●
●
●●●●
●
●● ●
●●
●
●●
●
●●
● ●
●●
●●
●
● ●
●
●
●●●
●●●
●● ●●●
●●
●
●●
●
●●●●
●
●
●
●●
●●
●
●●
●●
●
● ●
●
●●
●●● ●
●● ●
●● ●
●
●●● ●●●● ●
● ●●
● ● ●
●●
●
●
●●●
●
●
●●
●●●
●
● ●
●●●●●●●●● ●●
●
●●●●●●●●
●●
●●
● ●
●
●● ● ●
●
●
●
●●●●●●
●
●●●●
●
●
●
●
●
●●●●●● ● ● ●
●●●
● ●●●
●●
●
●
●●
● ●
●
●●
●
●●
● ●
●
●●●
●●●
● ●●
●●
●●
●●●
●●●● ●
●
●●
●
●●
●
●
●●●
●●
● ●●
●
●●●●
●
●
●
●●
●●
●
●●●●
● ●
●
●●●
●
●
●
●
●●●●
●●●● ● ●
●
● ●●
●
●●●●●●●
●
●●●●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●●●●●
● ● ●
●●●
●
●●
●●
●●
●●
●●
●●
●
● ●●
●●● ●●●
●●
● ●●
●●
●●●●
●●●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●● ●
●
●●
●●
● ●●
● ●●● ●
●●
●●●
●●
●
●
●
●●
●
●●
●●
●
●
●
●●●
●
● ●
●●●
●
●
●●
●●●●
● ●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●● ●
●●●
●● ●●
●●
●
●●
●
●
●●
●
●●●
●
●●
●●●
●●
●
●●
●
●
●●
●
●
●●
●
●●●
●
●
●
●
●
●●●
●
●●
●●
●●
●
●●
●●●
●●
●●
●●
●
●
●
●●
●
●
●●● ●●●
●
● ●
●
●
●●
●
●●
●
●●●
●●
0

● ●
●● ● ●● ●●● ● ● ● ● ● ●●
●● ●● ● ● ● ● ● ● ●
●● ● ● ●

0e+00 1e+06 2e+06 3e+06 4e+06

map5k

Figure 9.2: Manhattan plot for imputed (black empty circles) and 500 directly
typed (red solid circles) SNPs

1022 rs7273309 C T 0.9985 0.0015 0.9970 0.1896

1023 rs7261762 A G 0.9913 0.0087 0.9991 0.9472
1024 rs7267882 G A 0.9911 0.0089 0.9986 0.9230
1025 rs8123328 G T 0.9902 0.0098 0.9964 0.8126
1026 rs7265788 A G 0.9886 0.0114 0.9929 0.6884
2969 rs6047425 A G 0.9388 0.0612 0.8964 0.2573
2973 rs6035871 A G 0.9387 0.0613 0.8962 0.2582
3877 rs2325971 T C 0.5811 0.4189 0.8526 0.7807
3880 rs873711 A G 0.5888 0.4112 0.8597 0.7934
3884 rs6051434 G T 0.5186 0.4814 0.8427 0.7732
3886 rs6037425 G C 0.5871 0.4129 0.8651 0.8034
3958 rs6037443 A G 0.4975 0.4975 0.8918 0.8340
4009 rs566570 C T 0.5307 0.4693 0.8814 0.8074
4027 rs6076466 T C 0.7040 0.2960 0.6570 0.3365
4032 rs2326056 C A 0.7338 0.2662 0.6422 0.2403

Exercise 2.
Please classify the associations obtained into ’true’ and ’false’ (meaning ’I
believe it’ and ’I do not believe’). Of cause ultimate answer is replications.
Still, the object df5k provides genotypes for all 5000 SNPs, directly typed! So,
212 CHAPTER 9. ANALYSIS OF IMPUTED DATA: AN EXAMPLE

14
●
●
●

12
● ● ●
● ●
●
● ●
−log10(qtsPal[, "P−value"])

10

●
8

●
●
●
●
● ● ●
●
●
● ●
● ●
6

●
● ●
●
●
●
●●
●●●●
●● ●
● ● ●● ●●
● ● ● ●●
●●● ● ● ●● ● ●
● ● ● ●●
● ●●
● ● ●●
●●●● ● ● ●● ●
●● ●
●● ● ●● ● ●
●●
●
● ●●
●● ●● ●●●●
4

● ●● ● ● ●●
●●●
●
●
●●
●●●
● ●●
●
●●●● ●
●● ● ●●●● ●● ●
●● ●
●
●
●
●●●
●●●
●
●
●●
●●●●● ● ●● ● ●
●● ●●
●●
● ●
●●
●●
● ●
● ● ●● ● ● ●● ●
●●
●
●
●
●●●
●●● ● ●
● ● ●● ●●
●●●
● ●● ● ●
●
●
●
●
●●
●●
●●●
●
●●● ●
●
●●
●●
●●
●
●
●
●
●●●●
●●
●
●
●
●
●●
●
●
●
●
●●●
●● ●● ● ●
● ●●
● ●●
●●●●●●●
●●● ●
●●
●
● ●
●●●●●● ● ●●
●●
● ●●●●●
●●● ●●
●
● ●●
●●●
●
●
●● ●
●●●
●
●●
●●●
●●●●
● ●
● ●
●
●
●●
● ●●
●
●
●●●
●●
● ●● ● ●●
2

● ●● ●●
● ●●
●●
●●
●●
●●
●
●
●●
●●●
●●●
● ●
●●● ● ● ● ●
● ●●
●●●
●●●●●
●
●●
●
●●
●●
●●●●●●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●●
●●●●●●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
● ●
●
●
●
●
●●●
●
● ●●●●●●
● ● ●● ● ●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●●
●
●
●●
●●
●
●
●
●●
●
●
●
●●
●
●●
●●
●●
●
●
●●●●●●
● ●
●● ● ● ●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●●
●●
●
●●●●●
● ● ●●●●●
●
●
●●●
●●
●
●
●●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●●
●●●
●
●
● ● ● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●●
●
●
●
●
●
●●
●●●
●
●
●●●●●●
● ● ● ●
●
●
●●
●●
●●
●
●●●
●●
●
●●
●●
●●
●
●●●
●●●●●● ●●●●● ● ●●●●●
●
●
●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●●
●
●
●●
●●
●●
●
● ●
●
0

0 5 10 15

−log10(qts5k[, "P1df"])

Figure 9.3: Cross-plot of the results from analysis of imputed and directly typed
data (see answers to exercises).

you can run analysis, and cross-check which SNPs are confirmed as significant.
Make a cross table: SNPs you thought were truly associated vs. SNPs indeed
associated in directly typed data set.

9.3 Analysis of imputed data with MixABEL-package

In case you are interested in quantitative traits, MixABEL-package provides
much faster and more flexible analysis tools, compared to the ProbABEL-package3 .
Here the above analysis is repeated using MixABEL-package.
Load necessary libraries:
> library(DatABEL)
> library(MixABEL)

MixABEL v 0.1-2 (November 23, 2012) loaded

Load the genotypic data:

> imp5k <- databel("[Link]")
Run analysis and look up 10 top results:
3 However, at the moment MixABEL-package misses functionality to analyse binary and

time-till-event traits
9.4. ANSWERS TO EXERCISES 213

> qtsImp <- GWFGLS(rcT~SNP,geno=imp5k,data=NULL)

> qtsImp[order(qtsImp[,"Chisq"],decreasing=T)[1:10],]

n Chisq df P-value beta_SNP se_SNP mean_SNP

rs6039167 120 58.78462 1 1.759110e-14 5.988511 0.7810647 1.975075
rs7273309 120 57.71070 1 3.036458e-14 38.753300 5.1012957 1.997100
rs7265788 120 57.26926 1 3.800491e-14 7.395619 0.9772682 1.977092
rs8123328 120 52.41165 1 4.500458e-13 7.124468 0.9840984 1.980492
rs7267882 120 49.10249 1 2.429317e-12 6.850411 0.9776083 1.982333
rs566570 120 48.52652 1 3.258526e-12 1.294281 0.1857972 1.061425
rs7261762 120 48.48790 1 3.323324e-12 6.807552 0.9776294 1.982692
rs6035871 120 45.88351 1 1.254980e-11 4.886221 0.7213479 1.877392
rs6047425 120 45.85245 1 1.275033e-11 4.894889 0.7228723 1.877533
rs2325971 120 45.30891 1 1.682820e-11 1.297276 0.1927261 1.162083

Look up the information for SNPs with P -value less than 6e − 8:

> #plot(map5k,-log10(qtsImp[,"P-value"]))
> #abline(h=-log10(5e-8))
> gwsSnpsImp <- rownames(qtsImp)[which(qtsImp[,"P-value"]<=6e-8)]
> mlinfo <- [Link]("[Link]",head=T,strings=F)
> mlinfo[mlinfo$SNP %in% gwsSnpsImp,]

SNP Al1 Al2 Freq1 MAF Quality Rsq

1009 rs6039167 G A 0.9875 0.0125 0.9995 0.9800
1017 rs553378 G A 0.9987 0.0013 0.9974 0.0005
1018 rs554362 A G 0.6496 0.3504 0.5748 0.2481
1022 rs7273309 C T 0.9985 0.0015 0.9970 0.1896
1023 rs7261762 A G 0.9913 0.0087 0.9991 0.9472
1024 rs7267882 G A 0.9911 0.0089 0.9986 0.9230
1025 rs8123328 G T 0.9902 0.0098 0.9964 0.8126
1026 rs7265788 A G 0.9886 0.0114 0.9929 0.6884
2969 rs6047425 A G 0.9388 0.0612 0.8964 0.2573
2973 rs6035871 A G 0.9387 0.0613 0.8962 0.2582
3877 rs2325971 T C 0.5811 0.4189 0.8526 0.7807
3880 rs873711 A G 0.5888 0.4112 0.8597 0.7934
3884 rs6051434 G T 0.5186 0.4814 0.8427 0.7732
3886 rs6037425 G C 0.5871 0.4129 0.8651 0.8034
3958 rs6037443 A G 0.4975 0.4975 0.8918 0.8340
4009 rs566570 C T 0.5307 0.4693 0.8814 0.8074
4027 rs6076466 T C 0.7040 0.2960 0.6570 0.3365
4029 rs11697448 G A 0.7414 0.2586 0.6464 0.2370
4032 rs2326056 C A 0.7338 0.2662 0.6422 0.2403

9.4 Answers to exercises

Answer (Ex. 1) — Firstly, you can check (by producing a cross-plot of geno-
type vs. phenotype) if association is indeed due to extreme phenotypic out-
liers. A related question is whether the distribution is skewed. Additionally,
214 CHAPTER 9. ANALYSIS OF IMPUTED DATA: AN EXAMPLE

a permutation-based test can help establishing correct p-value, taking into ac-
count the nature of the data in question.
However, to give an ultimate answer, a replication study is needed, in which
these rare SNPs are to be typed in a large independent sample.

Answer (Ex. 2) — Here is the sequence of commands leading you to the an-
swer:
> qts5k <- mlreg(rcT~1,df5k)
> bestHits5k <- [Link](qts5k,top=20)
Summary for top 20 results, sorted by P1df
> bestHits5k
Chromosome Position Strand A1 A2 N effB se_effB chi2.1df
rs566570 20 2965113 + T C 120 1.341584 0.1585342 71.61267
rs6039167 20 846271 + G A 120 -5.945198 0.7731855 59.12420
rs6037443 20 2909408 + G A 120 1.194178 0.1628897 53.74655
rs7261762 20 853448 + A G 120 -6.733516 0.9748760 47.70732
rs554362 20 850002 + A G 120 -5.431215 0.8045925 45.56613
rs2104741 20 787070 + G A 120 -9.203756 1.3888630 43.91485
rs7273309 20 853154 + C T 120 -9.203756 1.3888630 43.91485
rs7267882 20 853785 + G A 120 -9.203756 1.3888630 43.91485
rs8123328 20 855045 + G T 120 -9.203756 1.3888630 43.91485
rs7265788 20 855426 + A G 120 -9.203756 1.3888630 43.91485
rs7263171 20 855655 + C T 120 -9.203756 1.3888630 43.91485
rs6110342 20 1458950 + T G 120 -9.203756 1.3888630 43.91485
rs6105340 20 1493635 + G A 120 -9.203756 1.3888630 43.91485
rs11905071 20 1557640 + T C 120 -9.203756 1.3888630 43.91485
rs6080013 20 1590861 + G A 120 -9.203756 1.3888630 43.91485
rs6074978 20 1595216 + T C 120 -9.203756 1.3888630 43.91485
rs2325971 20 2774477 + T C 120 -1.055991 0.1697536 38.69752
rs873711 20 2775468 + A G 120 -1.055991 0.1697536 38.69752
rs6037425 20 2785379 + G C 120 -1.055991 0.1697536 38.69752
rs6051434 20 2781237 + T G 120 0.938854 0.1769196 28.16076
P1df Pc1df effAB effBB chi2.2df P2df
rs566570 2.618695e-17 5.551496e-06 NA NA NA NA
rs6039167 1.480274e-14 3.663781e-05 NA NA NA NA
rs6037443 2.280947e-13 8.301445e-05 NA NA NA NA
rs7261762 4.948399e-12 2.090409e-04 NA NA NA NA
rs554362 1.475718e-11 2.904523e-04 NA NA NA NA
rs2104741 3.429767e-11 3.745369e-04 NA NA NA NA
rs7273309 3.429767e-11 3.745369e-04 NA NA NA NA
rs7267882 3.429767e-11 3.745369e-04 NA NA NA NA
rs8123328 3.429767e-11 3.745369e-04 NA NA NA NA
rs7265788 3.429767e-11 3.745369e-04 NA NA NA NA
rs7263171 3.429767e-11 3.745369e-04 NA NA NA NA
rs6110342 3.429767e-11 3.745369e-04 NA NA NA NA
rs6105340 3.429767e-11 3.745369e-04 NA NA NA NA
rs11905071 3.429767e-11 3.745369e-04 NA NA NA NA
rs6080013 3.429767e-11 3.745369e-04 NA NA NA NA
rs6074978 3.429767e-11 3.745369e-04 NA NA NA NA
9.4. ANSWERS TO EXERCISES 215

rs2325971 4.948386e-10 8.395801e-04 NA NA NA NA

rs873711 4.948386e-10 8.395801e-04 NA NA NA NA
rs6037425 4.948386e-10 8.395801e-04 NA NA NA NA
rs6051434 1.116452e-07 4.389549e-03 NA NA NA NA
> plot(qts5k)
> abline(h=-log10(5e-8))
> gwsSnps5k <- rownames(bestHits5k)[bestHits5k$P1df<=6e-8]
> summary(gtdata(df5k[,gwsSnps5k ]))
Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2
rs566570 20 2965113 + T C 120 1 0.495833333
rs6039167 20 846271 + G A 120 1 0.012500000
rs6037443 20 2909408 + G A 120 1 0.491666667
rs7261762 20 853448 + A G 120 1 0.008333333
rs554362 20 850002 + A G 120 1 0.012500000
rs2104741 20 787070 + G A 120 1 0.004166667
rs7273309 20 853154 + C T 120 1 0.004166667
rs7267882 20 853785 + G A 120 1 0.004166667
rs8123328 20 855045 + G T 120 1 0.004166667
rs7265788 20 855426 + A G 120 1 0.004166667
rs7263171 20 855655 + C T 120 1 0.004166667
rs6110342 20 1458950 + T G 120 1 0.004166667
rs6105340 20 1493635 + G A 120 1 0.004166667
rs11905071 20 1557640 + T C 120 1 0.004166667
rs6080013 20 1590861 + G A 120 1 0.004166667
rs6074978 20 1595216 + T C 120 1 0.004166667
rs2325971 20 2774477 + T C 120 1 0.470833333
rs873711 20 2775468 + A G 120 1 0.470833333
rs6037425 20 2785379 + G C 120 1 0.470833333
P.11 P.12 P.22 Pexact Fmax Plrt
rs566570 33 55 32 0.3648758 0.083269672 0.3614010
rs6039167 117 3 0 1.0000000 -0.012658228 0.8454919
rs6037443 35 52 33 0.1467972 0.133092526 0.1442654
rs7261762 118 2 0 1.0000000 -0.008403361 0.8968501
rs554362 117 3 0 1.0000000 -0.012658228 0.8454919
rs2104741 119 1 0 1.0000000 -0.004184100 0.9484250
rs7273309 119 1 0 1.0000000 -0.004184100 0.9484250
rs7267882 119 1 0 1.0000000 -0.004184100 0.9484250
rs8123328 119 1 0 1.0000000 -0.004184100 0.9484250
rs7265788 119 1 0 1.0000000 -0.004184100 0.9484250
rs7263171 119 1 0 1.0000000 -0.004184100 0.9484250
rs6110342 119 1 0 1.0000000 -0.004184100 0.9484250
rs6105340 119 1 0 1.0000000 -0.004184100 0.9484250
rs11905071 119 1 0 1.0000000 -0.004184100 0.9484250
rs6080013 119 1 0 1.0000000 -0.004184100 0.9484250
rs6074978 119 1 0 1.0000000 -0.004184100 0.9484250
rs2325971 38 51 31 0.1032427 0.147097763 0.1065638
rs873711 38 51 31 0.1032427 0.147097763 0.1065638
rs6037425 38 51 31 0.1032427 0.147097763 0.1065638
> plot(-log10(qts5k[,"P1df"]),-log10(qtsPal[,"P-value"]))
216 CHAPTER 9. ANALYSIS OF IMPUTED DATA: AN EXAMPLE

> abline(h=-log10(5e-8))
> abline(v=-log10(5e-8))
> directNotImp <- gwsSnps5k[!(gwsSnps5k %in% gwsSnpsImp)]
> directNotImp
[1] "rs2104741" "rs7263171" "rs6110342" "rs6105340" "rs11905071"
[6] "rs6080013" "rs6074978"
> imputeNotDir <- gwsSnpsImp[!(gwsSnpsImp %in% gwsSnps5k)]
> imputeNotDir
[1] "rs553378" "rs6047425" "rs6035871" "rs6051434" "rs6076466" "rs2326056"
> inBoth <- gwsSnps5k[gwsSnps5k %in% gwsSnpsImp]
> inBoth
[1] "rs566570" "rs6039167" "rs6037443" "rs7261762" "rs554362" "rs7273309"
[7] "rs7267882" "rs8123328" "rs7265788" "rs2325971" "rs873711" "rs6037425"
> summary(gtdata(df5k[,directNotImp]))
Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2
rs2104741 20 787070 + G A 120 1 0.004166667
rs7263171 20 855655 + C T 120 1 0.004166667
rs6110342 20 1458950 + T G 120 1 0.004166667
rs6105340 20 1493635 + G A 120 1 0.004166667
rs11905071 20 1557640 + T C 120 1 0.004166667
rs6080013 20 1590861 + G A 120 1 0.004166667
rs6074978 20 1595216 + T C 120 1 0.004166667
P.11 P.12 P.22 Pexact Fmax Plrt
rs2104741 119 1 0 1 -0.0041841 0.948425
rs7263171 119 1 0 1 -0.0041841 0.948425
rs6110342 119 1 0 1 -0.0041841 0.948425
rs6105340 119 1 0 1 -0.0041841 0.948425
rs11905071 119 1 0 1 -0.0041841 0.948425
rs6080013 119 1 0 1 -0.0041841 0.948425
rs6074978 119 1 0 1 -0.0041841 0.948425
> mlinfo[which(mlinfo$SNP %in% directNotImp),]
SNP Al1 Al2 Freq1 MAF Quality Rsq
871 rs2104741 G A 0.9562 0.0438 0.9145 0.0679
1027 rs7263171 C T 0.8639 0.1361 0.9246 0.7898
1978 rs6110342 T G 0.9160 0.0840 0.8472 0.1967
2028 rs6105340 G A 0.8790 0.1210 0.7849 0.2076
2064 rs11905071 T C 0.8846 0.1154 0.7989 0.1890
2111 rs6080013 G A 0.9577 0.0423 0.9342 0.3869
2118 rs6074978 T C 0.9576 0.0424 0.9340 0.3851
> summary(gtdata(df5k[,imputeNotDir]))
Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11
rs553378 20 849882 + G A 120 1 0.02500000 114
rs6047425 20 2141014 + A G 120 1 0.07083333 104
rs6035871 20 2143364 + A G 120 1 0.07083333 104
rs6051434 20 2781237 + T G 120 1 0.47500000 37
rs6076466 20 2982048 + T C 120 1 0.22916667 71
9.4. ANSWERS TO EXERCISES 217

rs2326056 20 2985607 + C A 120 1 0.22500000 72

P.12 P.22 Pexact Fmax Plrt
rs553378 6 0 1.0000000 -0.025641026 0.6948707
rs6047425 15 1 0.4561335 0.050382485 0.6092978
rs6035871 15 1 0.4561335 0.050382485 0.6092978
rs6051434 52 31 0.1473878 0.131161236 0.1502658
rs6076466 43 6 1.0000000 -0.014250614 0.8754486
rs2326056 42 6 1.0000000 -0.003584229 0.9686478
> mlinfo[which(mlinfo$SNP %in% imputeNotDir),]
SNP Al1 Al2 Freq1 MAF Quality Rsq
1017 rs553378 G A 0.9987 0.0013 0.9974 0.0005
2969 rs6047425 A G 0.9388 0.0612 0.8964 0.2573
2973 rs6035871 A G 0.9387 0.0613 0.8962 0.2582
3884 rs6051434 G T 0.5186 0.4814 0.8427 0.7732
4027 rs6076466 T C 0.7040 0.2960 0.6570 0.3365
4032 rs2326056 C A 0.7338 0.2662 0.6422 0.2403
> summary(gtdata(df5k[,inBoth]))
Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11
rs566570 20 2965113 + T C 120 1 0.495833333 33
rs6039167 20 846271 + G A 120 1 0.012500000 117
rs6037443 20 2909408 + G A 120 1 0.491666667 35
rs7261762 20 853448 + A G 120 1 0.008333333 118
rs554362 20 850002 + A G 120 1 0.012500000 117
rs7273309 20 853154 + C T 120 1 0.004166667 119
rs7267882 20 853785 + G A 120 1 0.004166667 119
rs8123328 20 855045 + G T 120 1 0.004166667 119
rs7265788 20 855426 + A G 120 1 0.004166667 119
rs2325971 20 2774477 + T C 120 1 0.470833333 38
rs873711 20 2775468 + A G 120 1 0.470833333 38
rs6037425 20 2785379 + G C 120 1 0.470833333 38
P.12 P.22 Pexact Fmax Plrt
rs566570 55 32 0.3648758 0.083269672 0.3614010
rs6039167 3 0 1.0000000 -0.012658228 0.8454919
rs6037443 52 33 0.1467972 0.133092526 0.1442654
rs7261762 2 0 1.0000000 -0.008403361 0.8968501
rs554362 3 0 1.0000000 -0.012658228 0.8454919
rs7273309 1 0 1.0000000 -0.004184100 0.9484250
rs7267882 1 0 1.0000000 -0.004184100 0.9484250
rs8123328 1 0 1.0000000 -0.004184100 0.9484250
rs7265788 1 0 1.0000000 -0.004184100 0.9484250
rs2325971 51 31 0.1032427 0.147097763 0.1065638
rs873711 51 31 0.1032427 0.147097763 0.1065638
rs6037425 51 31 0.1032427 0.147097763 0.1065638
> mlinfo[which(mlinfo$SNP %in% inBoth),]
SNP Al1 Al2 Freq1 MAF Quality Rsq
1009 rs6039167 G A 0.9875 0.0125 0.9995 0.9800
1018 rs554362 A G 0.6496 0.3504 0.5748 0.2481
1022 rs7273309 C T 0.9985 0.0015 0.9970 0.1896
218 CHAPTER 9. ANALYSIS OF IMPUTED DATA: AN EXAMPLE

1023 rs7261762 A G 0.9913 0.0087 0.9991 0.9472

1024 rs7267882 G A 0.9911 0.0089 0.9986 0.9230
1025 rs8123328 G T 0.9902 0.0098 0.9964 0.8126
1026 rs7265788 A G 0.9886 0.0114 0.9929 0.6884
3877 rs2325971 T C 0.5811 0.4189 0.8526 0.7807
3880 rs873711 A G 0.5888 0.4112 0.8597 0.7934
3886 rs6037425 G C 0.5871 0.4129 0.8651 0.8034
3958 rs6037443 A G 0.4975 0.4975 0.8918 0.8340
4009 rs566570 C T 0.5307 0.4693 0.8814 0.8074
Chapter 10

Meta-analysis of GWA
scans

10.1 Standard meta-analysis methods

Imagine you are interested in the effect of a certain polymorphism onto a par-
ticular disease. After scanning literature, you find some studies that implicate
certain allele as significantly increasing the risk of the disease, but you will typ-
ically find also that other studies were inconclusive (no significant association),
and that even some of the studies implicated the same allele as ”protective”.
Your gut feeling may be that the allele is indeed the risk one, because you feel
that the studies contradicting to this hypothesis were based on small number of
subjects; however, how do you quantify this feeling? In this situation you need
to perform meta-analysis of available data to come up with the joint effect size
estimate and P-value, as based on all available data.
Let us first consider a situation when you are interested in the effect of the
allele on a quantitative phenotype, expressed as a coefficient of regression of the
trait onto the number of this allele in the genotype. Under a favorable scenario,
from every individual study you would know the estimate of this regression
coefficient, and the standard error of the estimate (or, equivalently, the P −value
or the test statistics value for association).
One of approaches frequently used in meta-analysis of the data coming from
a number of independent studies is the inverse variance method. In essence,
this method is equivalent to combining likelihoods coming from separate studies,
using quadratic approximation. Denote coefficients of regression estimated in
N studies as βi , and associated squared standard errors of the estimates as s2i
where i ∈ 1, 2, ..., N . Note that the regression coefficient should be reported on
the same scale, e.g. centimeters, meters, or using observations reported on the
standard normal scale. Define weights for individual studies as
1
wi =
s2i
Then the pooled estimate of the regression coefficient is
PN
w i βi
β = Pi=1 N
i=1 wi

219
220 CHAPTER 10. META-ANALYSIS OF GWA SCANS

As you can see, the weights have straightforward interpretation: the bigger
the weight of the study (meaning the small is the standard error in the study),
the larger is the contribution from this study onto the pooled estimate.
The standard error of the pooled estimate is computed as
1
s 2 = PN
i=1 wi

and the χ2 -test for association is computed in standard manner as

P 2
N
β 2 i=1 w i βi
T2 = 2 = PN
s i=1 wi

or, alternatively, the Z-test is

PN
β i=1 wi βi
Z= = qP
s N
i=1 wi

Let us try to do meta-analysis using the inverse variance pooling method.

Imagine we have information from four different studies reporting effect and the
standard error of the same allele:

Table 10.1: Estimated regression coefficients from four studies

Study n β sβ χ2
1 225 0.16 0.07 5.224
2 560 0.091 0.042 4.694
3 437 0.072 0.048 2.25
4 89 -0.03 0.12 0.062
Total 1311 ? ? ?

Let us try to access the joint significance of the association using these data.
First, let us define a vector of regression coefficients and squared standard errors:
> beta <- c(0.16,0.091,0.072,-0.03)
> s <- c(0.07,0.042,0.048,0.12)
> s2 <- s*s
> s2

[1] 0.004900 0.001764 0.002304 0.014400

Compute the weight for individual studies as

> w <- 1/s2
> w

[1] 204.08163 566.89342 434.02778 69.44444

Estimate pooled regression coefficient as

> pbeta <- sum(w*beta)/sum(w)
> pbeta
10.1. STANDARD META-ANALYSIS METHODS 221

[1] 0.08898527

and pooled squared standard error as

> ps2 <- 1/sum(w)
> ps2

[1] 0.0007846539

To access significance of association in meta-analysis, let us compute χ2 test

statistics and the P − value with
> pchi2 <- pbeta*pbeta/ps2
> pchi2

[1] 10.09155

> ppvalue <- 1. - pchisq(pchi2,1)

> ppvalue

[1] 0.001489504

We conclude that there is a significant association in meta-analysis.

There is an important effect which should be considered when doing meta-
analysis of published data. Given numerous polymorphisms available in human
genome, a particular polymorphism usually becomes a focus of interest only
when it was shown to be significantly associated in some study which reports it.
Put it other way around: only when a significant association was detected and
reported, more studies are likely to be performed on the same polymorphism.
This first report, however, is not guaranteed to demonstrate a true association:
it may well report a false-positive, or, even in presence of association, it is likely
to over-estimate the effect of the polymorphism. Thus there is a positive bias
in literature reports, and this bias is particularly strong for the first report, a
phenomenon frequently referenced to as ”champion’s” or ”winner’s curse”.
The observations we have just considered are quite typical in that the first
study, where the association was originally discovered, reports the biggest effect
and most significant effect, while the follow-up studies suggest smaller effect.
Therefore, when you meta-analyse data from publications it is always good
idea to exclude the first report (in case it is positive – and it is always positive!)
and check if significant association is still observed. Let us try to do that:
> beta <- beta[2:4]
> s2 <- s2[2:4]
> w <- w[2:4]
> pbeta <- sum(w*beta)/sum(w)
> pbeta

[1] 0.07544522

> ps2 <- 1/sum(w)

> ps2

[1] 0.0009342602
222 CHAPTER 10. META-ANALYSIS OF GWA SCANS

> pchi2 <- pbeta*pbeta/ps2

> pchi2

[1] 6.092501

> ppvalue <- 1. - pchisq(pchi2,1)

> ppvalue

[1] 0.01357568

Indeed, when the first ”champion” report is excluded, the overall evidence is
decreased and results become less significant, though still pointing to the same
direction.
When binary traits are studied, and results are reported as Odds Ratios with
P − values, it is also possible to apply inverse variance method. For this, you
need to transform your Odds Ratios using natural logarithm, and, on this scale,
estimate the standard error. Generic inverse variance pooling may be applied
to the data transformed this way; the final results are back-transformed onto
Odds Ratio scale using exponentiation.
Let us consider a simple example. Let Odds Ratios and χ2 test statistics
values coming from four studies of a binary phenotype are as following: θ1 = 1.5
(χ2 = 5.1), θ2 = 1.3 (χ2 = 2.2), θ3 = 0.9 (χ2 = 0.5), θ4 = 1.2 (χ2 = 3.1).
Let us first transform the Odds Ratio to the logarithmic scale with
> or <- c(1.5,1.3,0.9,1.2)
> lnor <- log(or)
> lnor

[1] 0.4054651 0.2623643 -0.1053605 0.1823216

To compute standard errors from known χ2 values, one can use simple relation

β2
χ2 =
s2
and thus
β2
s2 =
χ2
Thus to compute the square standard errors corresponding to the log-Odds
Ratio, we can use
> chi2or <- c(5.1,2.2,0.5,3.1)
> s2lnor <- lnor*lnor/chi2or
> s2lnor

[1] 0.03223568 0.03128864 0.02220168 0.01072295

Now we can combine log-Odds Ratios and corresponding standard errors

using inverse variance method:
> w <- 1/s2lnor
> plnor <- sum(w*lnor)/sum(w)
> plnor
10.1. STANDARD META-ANALYSIS METHODS 223

[1] 0.1650462
> ps2 <- 1/sum(w)
> ps2
[1] 0.004968165
> pchi2 <- plnor*plnor/ps2
> pchi2
[1] 5.482958
> ppval <- 1.-pchisq(pchi2,1)
> ppval
[1] 0.01920274
And the corresponding estimate of pooled Odds Ratio is
> exp(plnor)
[1] 1.179448
and 95% confidence interval is
> exp(plnor-1.96*sqrt(ps2))
[1] 1.02726
> exp(plnor+1.96*sqrt(ps2))
[1] 1.354181
Some times, effects are reported on different scales, and/or there may be
suspect that these effects are not translatable across studies because of the
differences in experimental design or for some other reasons. In this case, it may
be better to poll the data without use of the effect estimate in exact manner,
based only on the sign of association and its significance. This can be done by
pooling Z-score values. Z-score refers to the test statistics, which has standard
normal distribution under the null and can be derived e.g. by dividing estimate
of the regression coefficient onto its standard error:
βi
Zi =
si
The Z-score pooling methods can be derived from the inverse variance pool-
ing by exploiting
√ the fact that generally standard error of the estimate is pro-
portional to 1/ n, where n is the number of observations used for estimation.
Therefore individual scores are assigned weights which are proportional to the
square root of number of independent observations used in individual study,
√
wi = ni . The pooled Z-score statistics is computed as
PN
i=0 wi Zi
Z= P q
N 2
i=0 wi

We can now repeat the analysis of our first data set using Z-score pooling
method. First, our data from table 10.1 are
224 CHAPTER 10. META-ANALYSIS OF GWA SCANS

> n <- c(225,560,437,89)

> beta <- c(0.16,0.091,0.072,-0.03)
> s <- c(0.07,0.042,0.048,0.12)
The Z-scores and weights are are:
> z <- beta/s
> z

[1] 2.285714 2.166667 1.500000 -0.250000

> w <- sqrt(n)

> w

[1] 15.000000 23.664319 20.904545 9.433981

The pooled estimate of Z-score is

> pz <- sum(w*z)/sqrt(sum(w^2))
> pz

[1] 3.163875

and corresponding P − value is

> 1.-pchisq(pz*pz,1)

[1] 0.001556839

which is almost the same P − value we have obtained previously using the
inverse variance method. Note, however, that now we do not know the ”pooled”
estimate of the regression coefficient.
Other important aspects of meta-analysis, such as heterogeneity, and a wide
range of methods different from the inverse variance and Z-score based methods
are not covered here, and we refer the reader to more epidemiologically-oriented
literature for a better review.

10.2 Exercise: meta-analysis of literature data

In this exercise, you will perform meta-analysis of results collected from liter-
ature. These results resemble these obtained for association analysis between
Pro12Ala polymorphism of the PPAR-GenABEL-packagemma gene and type 2
diabetes. The data collected from literature are presented in the table 10.2.
As you can see, only the original study report significant association, while
other four are insignificant and one point in opposite direction.
Answer the following questions:

Exercise 8 Perform meta-analysis of the data presented in table 10.2. Which

allele is the risk one? Is this risk significant? What is pooled Odds Ratio and
95% confidence interval? Do analysis using at least two methods. Which method
is better (best) in this situation? Why?

Exercise 9 Perform meta-analsys excluding the original report (study 1). Is

there still significant association between Pro12Ala and diabetes?
10.3. REPORTING GWA RESULTS FOR FUTURE META-ANALYSIS 225

Table 10.2: Summary of six studies of association between T2D and Pro12Ala
polymorphism of the PPAR-GenABEL-packagemma gene. n: number of sub-
jects; effective allele: the allele for which the OR was estimated.
Study Effective allele n ORE P − value
1 Ala 221 0.67 0.013
2 Pro 306 0.93 0.60
3 Pro 71 1.08 0.84
4 Ala 164 0.83 0.40
5 Pro 242 1.22 0.25
6 Pro 471 1.23 0.07

10.3 Reporting GWA results for future meta-

analysis
In this section, we will discuss specifics of GWA analysis when meta-analysis
is aimed at later stage. In order to perform meta-analysis at later stage, using
either inverse variance or Z-score based method, you generally need to report
only effect estimates, standard errors of The estimates (or, equivalently, P −
values or test statistics values), and number of observations used for estimation.
It is also clear that it is crucial to know for which allele the effect is reported,
and this is the point where meta-analysis of genetic data may be very confus-
ing. Generally, one may think that reporting what couple of nucleotide bases
correspond to the polymorphism under the study and defining what allele was
used as reference in the regression model may be enough. This, however, is not
true for certain class of polymorphisms and may be a source of great confusion.
Consider a DNA molecule; as you know it is made of two complementary
strands (forward or ”+” and reverse or ”-”). As you may guess, depending
on the strand, what is an ”A/G” polymorphism when reported on ”+” strand
becomes ”T/C” polymorphism on the ”-” strand (using complementarity rule
A¡-¿T and G¡-¿C). This is not a big problem for most of the polymorphism
classes, because if say you know that for a first study β1 is reported for the ”G”
allele of the ”A/G” polymorphism and in the second study β2 is the estimate
of the effect ”C” allele of the same polymorphism, but coded as ”T/C” (thus
other strand), you can easily spot that and say the ”C” is the same as ”G” in
this situation, and pool the two betas straightforwardly.
However, for two types of polymorphisms, ”A/T” and ”G/C”, where you can
not get away without knowing what the strand was: what is reported as the
effect of ”T” in ”A/T” polymorphism in study one; and seemingly corresponding
effect of ”T” in ”A/T” polymorphism in study two may be apparently reports for
two opposite alleles, if strands used for reporting in two studies were different.
The story may become even more complex, because the forward/reverse
orientation depends on the genomic build1 .
Thus if you want to pool your results with the results of others, there are
quite a few SNP characteristics which are absolutely crucial to report, namely,
the nucleotide bases describing the polymorphism, with indication which one
was used as reference, and which one as ”effective”, strand, genomic build,
1 There is alternative, top/bottom, strand designation, which does not depend on genomic

build. However, it is not always used.

226 CHAPTER 10. META-ANALYSIS OF GWA SCANS

and only than the effect estimate, standard error of the effect, and number of
observations used to do estimation.
Other characteristics which are also recommended for reporting because they
describe the quality characteristics of the sample and/or provide redundant in-
formation, which is good for double checks. Such characteristics include: fre-
quency of the effective or reference allele, call rate, P-value for Hardy-Weinberg
equlibrium and may be some parameter describing what is the direction of de-
viation from HWE (e.g. Fmax ). When reporting results for imputed SNPs,
more quality control characteristics should be included, suh as average maximal
posterior probability, R2 , etc.
Let us start with arranging two data sets which could then be used for meta-
analysis. Basically, we will use cleaned data from the GWA exercise you did in
section 5 (”Genome-wide association analysis”, page 103), and split that is two
parts.
If you did not do this yet, start R and load GenABEL-package library, which
you will need it to work with GWA-data

> library(GenABEL)

Load the data with

> load("[Link]")

and then split it in two parts:

> nids(data2)

[1] 124

> mdta1 <- data2[1:40,]

> mdta2 <- data2[41:nids(data2),]

We will analyse body mass index. If you pooling results of analysis of stud-
ies which are designed in approximately the same manner, you may think of
reporting the effect estimates on the same scale and use of the inverse variance
method for meta-analysis.
However, in meta-analysis of multiple data sets different individual studies
are likely to assess different population, will use different designs, measure dif-
ferent covariates, and so on. Therefore you should think of some standardisation
of the outcome variable (or apply Z-score method).
Therefore for the purpose of future meta-analysis, it becomes conventional
to analyse pre-adjusted data which are scaled to Standard Normal (mean of zero
and variance of unity). Note that this argument applies only to meta-analysis –
you may and should report effects on the original scale (e.g. in centimeters and
grams) in analysis of individual studies, in order to have better interpretability.

Moreover, in meta-analysis you heavily rely on the large numbers approxima-

tion when estimating P −values; while for individual study you can always apply
empirical, e.g. permutation-based, procedures to derive the correct P − value
whatever is the distribution of the trait, in meta-analysis the Normality assump-
tion becomes crucial, and you do not want few outliers spoiling your results by
screwing up P − values. Therefore some transformation improving normality is
10.3. REPORTING GWA RESULTS FOR FUTURE META-ANALYSIS 227

desirable. Note that transformation to Standard Normal does not improve the
fit to normality; to do that other transformation should be applied. Probably
the most famous transformations are log- and square root ones, then one may
think of Box-Cox transformation. At the same time there is a transformation,
called a Rank Transformation ro Normality which guarantees perfect fit to Nor-
mal in absence of heavy ties2 . We advocate the use of Rank Transformation to
Normal for meta-analysis purposes.
GenABEL-package implements the ztransform function for the purposes of
Z-transformation. This function takes two (actually three – see help for details)
arguments: formula (or just the variable name) and data. ztransform function
will perform (generalised) linear regression using the specified formula, and will
transform the residuals from analysis onto Z-scale by subtracting the mean and
division by the standard deviation.
Let us consider what this function does practically. Let us first transform
BMI from the first set without using covariates:
> zbmi0 <- ztransform(bmi,mdta1)
The histogram of the transformed variable and scatter-plot of raw against trans-
formed BMI is given at figure 10.1, column 1. Note that the fit to Normality
is not improved by this transformation; with the original BMI, the Shapiro test
for deviation from normality gives
> [Link](phdata(mdta1)$bmi)

Shapiro-Wilk normality test

data: phdata(mdta1)$bmi
W = 0.9328, p-value = 0.0199

with identical results from the transformed variable:

> [Link](zbmi0)

Shapiro-Wilk normality test

data: zbmi0
W = 0.9328, p-value = 0.0199

This is quite natural: as you can note from scatter-plot in column 1 of figure
10.1, only the centering and the spread of the scales are different for X (original
BMI) and Y (x0), otherwise there is an exact linear correspondence between the
two.
We can also do transformation using sex and age-adjusted residuals with
> zbmi1 <- ztransform(bmi~sex+age,mdta1)
The scatter-plot of raw against transformed BMI is given at figure 10.1, column
2. Note that this transformation may slightly change the fit to Normal, which
happens because we factor out the effects of sex and age:
> [Link](zbmi1)
2 Ties are generated by the subjects with exactly the same trait values
228 CHAPTER 10. META-ANALYSIS OF GWA SCANS

Shapiro-Wilk normality test

data: zbmi1
W = 0.9263, p-value = 0.01224

From the scatter-plot in column 2 of figure 10.1, it is quite clear what happens:
the residuals from linear regression are not corresponding to the original BMI
in exact linear manner.
A similar function, which performs rank-transformation to normality, is
named rntransform. For example if we want to adjust BMI for sex and age
and rank-transform the residuals to Normal, we can use
> rnbmi1 <- rntransform(bmi~sex+age,mdta1)
This transformation, however, indeed improves the fit to Normal:
> [Link](rnbmi1)

Shapiro-Wilk normality test

data: rnbmi1
W = 0.999, p-value = 1

In essence, the P − value of 1 means perfect fit to Normal – and this is what
should have occurred when this transformation is used on the data without ties.
Perfectly Normal distribution of the transformed trait may be enjoyed at the
histogram presented at column 3 of figure 10.1.
Let us analyse Rank-Normal-transformed, sex and age-adjusted BMI in the
two data sets, using qtscore function. Analysis of the first study is done with
> qts1 <- qtscore(rnbmi1,mdta1)
and analysis of the second study is done with
> zbmi2 <- ztransform(bmi~sex+age,mdta2)
> qts2 <- qtscore(zbmi2,mdta2)
The analysis looks very simple – is not it? However, the real difficulty did
not start yet: now we need to extract coding, reference allele, strand, etc. –
otherwise we can not do right meta-analysis later on!
Let us assume that you want to summarise the GW results from additive 1
d.f. test using following variables (as, e.g., requested by consortium):
• name: SNP name
• chromosome: chromosome number
• position: physical position of the SNP
• refallele: reference allele
• codedallele: coded (effect) allele
• strand: strand
• refallfreq: frequency of the reference allele
10.3. REPORTING GWA RESULTS FOR FUTURE META-ANALYSIS 229

Histogram of zbmi0 Histogram of zbmi1 Histogram of rnbmi1

10
8

8
8
6

6
Frequency

Frequency

Frequency
6
4

4
4
2

2
2
0

0
−2 0 1 2 3 −1 0 1 2 3 −2 0 1 2

zbmi0 zbmi1 rnbmi1

● ● ●
3

2
3

●
●
●
● ●
● ●
2

1
●
2

●
● ●● ●
●●
●●
● rnbmi1 ●
zbmi0

zbmi1

● ● ●●
1

● ●●●
1

0
● ●● ●
● ●
●● ●
●
●
●●
● ●●
●
●●
● ●● ●
●
●●
● ●
● ● ●● ●●
0

● ●
0

●
−1
● ●● ●
●
● ●
●
● ● ●
●
●
● ●●
● ●
● ●
●
●● ●
●
● ●●●●
−1

−1

●● ●
●
−2

● ●●
●● ●
● ●

20 30 40 50 60 20 30 40 50 60 20 30 40 50 60

phdata(mdta1)$bmi phdata(mdta1)$bmi phdata(mdta1)$bmi

Figure 10.1: Histogram of transformed BMI and scatter-plots of the raw BMI
against transformed BMI. Column 1: Z-transformation without covariates. Col-
umn 2: Z-transformation with adjustment for age and sex. Column 3: Rank-
transformation to normality, after adjustment for sex and age.

• n: number of people with data available for this SNP test

• beta: estimate of the effect of the allele
• se beta: standard error of the effect estimate
• p: P-value for the test
• p corr: corrected P-value (we will use Genomic Control)
• call: SNP call rate
• phwe: P -value from the exact test for HWE
Let us look what we get as an output from qtscore analysis:
> results(qts1)[1:2,]

Chromosome Position Strand A1 A2 N effB se_effB chi2.1df

rs1646456 1 653 + C G 40 0.1768955 0.2560635 0.4772414
rs4435802 1 5291 + C A 40 -0.4064900 0.4095768 0.9849838
P1df effAB effBB chi2.2df P2df Pc1df
rs1646456 0.4896745 0.09703712 0.4909062 0.6323627 0.7289272 0.4896745
rs4435802 0.3209715 -0.40649004 NA 0.9849838 0.3209715 0.3209715
230 CHAPTER 10. META-ANALYSIS OF GWA SCANS

You can see that most infromation is already present in the output, though
called using names which are different from these requested. However, we miss
reference allele frequency, SNP call rate and P -value from the Hardy-Weinberg
equilibrium test. These however can be computed using the summary function:
> # for data part 1:
> sum1 <- summary(gtdata(mdta1))
> sum1[1:2,]

Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11 P.12

rs1646456 1 653 + C G 40 1 0.3375 16 21
rs4435802 1 5291 + C A 40 1 0.0875 33 7
P.22 Pexact Fmax Plrt
rs1646456 3 0.4775382 -0.17400419 0.2616425
rs4435802 0 1.0000000 -0.09589041 0.4122629

> # ... and for data part 2:

> sum2 <- summary(gtdata(mdta2))
> sum2[1:2,]

Chromosome Position Strand A1 A2 NoMeasured CallRate Q.2 P.11

rs1646456 1 653 + C G 83 0.9880952 0.33734940 36
rs4435802 1 5291 + C A 82 0.9761905 0.07926829 69
P.12 P.22 Pexact Fmax Plrt
rs1646456 38 9 1 -0.02402597 0.826397
rs4435802 13 0 1 -0.08609272 0.289791

Note, however, that we now got frequency of the effective (or coded) allele,
not the frequency of the reference allele! The quantity we need can be easily
computed, though:
> # for data part 1
> refallfreq1 <- 1 - sum1[,"Q.2"]
> # ... and for data part 2
> refallfreq2 <- 1 - sum2[,"Q.2"]
At this moment we can arrange the required data frame:
> mdf1 <- [Link](name=snpnames(qts1),chromosome=chromosome(qts1),
+ position=map(qts1),refallele=refallele(qts1),
+ codedallele=effallele(qts1),strand = strand(qts1),
+ refallelefreq = refallfreq1,n=qts1[,"N"],
+ beta=qts1[,"effB"],
+ se_beta=qts1[,"se_effB"],p=qts1[,"P1df"],
+ p_corr=qts1[,"Pc1df"],call = sum1[,"CallRate"],
+ phwe = sum1[,"Pexact"], stringsAsFactors = FALSE)
note the last argument – stringsAsFactors = FALSE. I suggest you use that
by default when constructing a new data frame – unless you are sure that you
can work out your way later on with strings saved as factors.
Let us inspect the first 5 raws of the resulting output:
> mdf1[1:5,]
10.3. REPORTING GWA RESULTS FOR FUTURE META-ANALYSIS 231

name chromosome position refallele codedallele strand

rs1646456 rs1646456 1 653 C G +
rs4435802 rs4435802 1 5291 C A +
rs946364 rs946364 1 8533 T C -
rs299251 rs299251 1 10737 A G +
rs2456488 rs2456488 1 11779 G C +
refallelefreq n beta se_beta p p_corr call
rs1646456 0.6625000 40 0.1768955 0.2560635 0.4896745 0.4896745 1.000
rs4435802 0.9125000 40 -0.4064900 0.4095768 0.3209715 0.3209715 1.000
rs946364 0.7236842 38 0.2944870 0.2752734 0.2847102 0.2847102 0.950
rs299251 0.9615385 39 0.4958353 0.5989959 0.4077965 0.4077965 0.975
rs2456488 0.6625000 40 0.1145684 0.2271545 0.6140062 0.6140062 1.000
phwe
rs1646456 0.4775382
rs4435802 1.0000000
rs946364 0.6911168
rs299251 1.0000000
rs2456488 0.7320709

However, it is not recommended that you perform above-described report-

ing actions unless you develop your own format. In case if you plan to use
MetABEL-package for meta-analysis, you best use formetascore function, which
basically performs operations similar to described, and reports results in format
compatible with MetABEL-package.
Thus, if you plan to use MetABEL-package for meta-analysis, required tables
can be generated with single command:

> mdf1 <- formetascore(bmi~sex+age,mdta1,transform=rntransform, verbosity = 2 )

You can see that results are the same as previously:

> mdf1[1:5,]

name chromosome position strand allele1 allele2 build

rs1646456 rs1646456 1 653 + C G unknown
rs4435802 rs4435802 1 5291 + C A unknown
rs946364 rs946364 1 8533 - T C unknown
rs299251 rs299251 1 10737 + A G unknown
rs2456488 rs2456488 1 11779 + G C unknown
effallele effallelefreq n beta sebeta p pgc
rs1646456 G 0.33750000 40 0.1768955 0.2560635 0.4896745 0.4896745
rs4435802 A 0.08750000 40 -0.4064900 0.4095768 0.3209715 0.3209715
rs946364 C 0.27631579 38 0.2944870 0.2752734 0.2847102 0.2847102
rs299251 G 0.03846154 39 0.4958353 0.5989959 0.4077965 0.4077965
rs2456488 C 0.33750000 40 0.1145684 0.2271545 0.6140062 0.6140062
[Link] [Link] pexhwe call
rs1646456 1 NA 0.4775382 1.000
rs4435802 1 NA 1.0000000 1.000
rs946364 1 NA 0.6911168 0.950
rs299251 1 NA 1.0000000 0.975
rs2456488 1 NA 0.7320709 1.000
232 CHAPTER 10. META-ANALYSIS OF GWA SCANS

To write all the data to a file, we can use standard R [Link] function:
> [Link](mdf1,file="RData/[Link]",[Link]=F)
Similar analysis is applied to the second data set:
> mdf2 <- formetascore(bmi~sex+age,mdta2,transform=rntransform, verbosity = 2 )
We can inspect the first five lines of the output with
> mdf2[1:5,]
name chromosome position strand allele1 allele2 build
rs1646456 rs1646456 1 653 + C G unknown
rs4435802 rs4435802 1 5291 + C A unknown
rs946364 rs946364 1 8533 - T C unknown
rs299251 rs299251 1 10737 + A G unknown
rs2456488 rs2456488 1 11779 + G C unknown
effallele effallelefreq n beta sebeta p
rs1646456 G 0.32926829 82 -0.04879397 0.1663650 0.76929696
rs4435802 A 0.08024691 81 0.37724197 0.2984226 0.20618701
rs946364 C 0.25903614 83 -0.14414880 0.1790329 0.42073156
rs299251 G 0.04216867 83 -0.69920378 0.3919648 0.07444912
rs2456488 C 0.34146341 82 -0.23105805 0.1520352 0.12856957
pgc [Link] [Link] pexhwe call
rs1646456 0.77676571 1.070017 0.0009650323 0.8038996 0.9879518
rs4435802 0.22168453 1.070017 0.0009650323 1.0000000 0.9759036
rs946364 0.43635427 1.070017 0.0009650323 1.0000000 1.0000000
rs299251 0.08461892 1.070017 0.0009650323 1.0000000 1.0000000
rs2456488 0.14177789 1.070017 0.0009650323 0.4685397 0.9879518
Let us write the data to a file:
> [Link](mdf2,file="RData/[Link]",[Link]=F)
Finally let us analyse and save results for another data set, ge03d2c:
> data(ge03d2c)
> mdf3 <- formetascore(bmi~sex+age,ge03d2c,transform=rntransform, verbosity = 2 )
> [Link](mdf3,file="RData/[Link]",[Link]=F)

10.4 Meta-analysis with MetABEL-package

Now we are ready to meta-analyse GWA data coming from three studies. For
this we will need to use MetABEL-package package, implementing simple meta-
analysis functions for GWA data. Start with loading the package:
> library(MetABEL)
We will first meta-analyse the three studies using the data frames generated
in previous section, mdf1, mdf2 and mdf3. For this we will use the core function
of MetABEL-package, [Link]. This function takes four arguments:
two data frames with results from individual studies, and two arguments sup-
plying the study names. Pooling of multiple studies is possible by sequential
application of this function.
Let us pool two first data frames:
10.4. META-ANALYSIS WITH METABEL-PACKAGE 233

> pooled <- [Link](mdf1,mdf2,name.x="Part1",name.y="Part2")

analysing ...
Lambda Part1 = 0.9499903
Lambda Part2 = 1.098705
Corrected Lambda Part1 = 0.9499903
Corrected Lambda Part2 = 1
Lambda POOLED data = 1.023276
... DONE

The pooled data frame contains results of meta-analysis and essential details
of the original studies:
> pooled[1:5,]

name strand allele1 allele2 effallele chromosome position n npops

1 rs100616 - G C C 1 1911712 122 2
2 rs1006497 + T G G 1 2658810 122 2
3 rs1011580 + A G G 3 10048771 121 2
4 rs1011953 + A G G 2 6464510 123 2
5 rs1013473 + A T T 1 4487262 123 2
beta sebeta effallelefreq call pexhwe obetaPart1
1 0.11443814 0.1850338 0.1270492 0.9919020 4.64447636 0.08819818
2 -0.07946221 0.1954998 0.1434426 0.9920082 0.16813927 -0.25425751
3 -0.06801448 0.1397522 0.5041322 0.9837773 3.15056880 -0.06755996
4 0.07539928 0.1372443 0.3292683 1.0000000 0.10373542 0.18982068
5 0.30393338 0.1334451 0.5121951 1.0000000 0.04821391 0.31522388
obetaPart2 osePart1 osePart2 chi2 p
1 0.1230224623 0.3726827 0.2131624 0.3825070 0.5362646
2 -0.0001117686 0.3498751 0.2357343 0.1652071 0.6844070
3 -0.0683575796 0.2130800 0.1851327 0.2368565 0.6264858
4 0.0055254619 0.2228918 0.1741797 0.3018185 0.5827446
5 0.2978861467 0.2259542 0.1653644 5.1874252 0.0227509

If one needs to pool more studies, this data frame should be used as the first
argument of the [Link], and name.x argument should take special
value ”POOLED”:
> pooled <- [Link](pooled,mdf3,name.x="POOLED",name.y="mdf3")

NA for betas in both populations

18 SNPs removed
analysing ...
Lambda mdf3 = 1.128096
Corrected Lambda mdf3 = 1
Lambda POOLED data = 1.340905
... DONE

> pooled[1:5,]

name strand allele1 allele2 effallele chromosome position n npops

1 rs1000475 + T C C X 13721802 91 1
234 CHAPTER 10. META-ANALYSIS OF GWA SCANS

2 rs1000909 - A G G 2 8531681 190 1

3 rs1006092 - T G G X 13527448 190 1
4 rs100616 - G C C 1 1911712 310 3
5 rs1006497 + T G G 1 2658810 315 3
beta sebeta effallelefreq call pexhwe obetaPart1
1 -0.05521349 0.11185255 0.5824176 0.4690722 5.9671619 NA
2 -0.08504871 0.13218866 0.8157895 0.9793814 2.8521518 NA
3 -0.03712065 0.08523891 0.5078947 0.9793814 6.8378385 NA
4 0.11824774 0.12085884 0.1241935 0.9780568 6.5922479 0.08819818
5 -0.02366946 0.11439872 0.1603175 0.9937465 0.2313494 -0.25425751
obetaPart2 obetamdf3 osePart1 osePart2 osemdf3 chi2
1 NA -0.055213494 NA NA 0.11185255 0.24366808
2 NA -0.085048708 NA NA 0.13218866 0.41394922
3 NA -0.037120652 NA NA 0.08523891 0.18965112
4 0.1230224623 0.121082405 0.3726827 0.2131624 0.15961076 0.95725763
5 -0.0001117686 0.005382408 0.3498751 0.2357343 0.14107317 0.04280893
p
1 0.6215693
2 0.5199718
3 0.6632071
4 0.3278788
5 0.8360855

This procedure may become quite laborious if multiple studies are to be

pooled. In this case, it is possible to run meta-analysis using data provided in
files, by applying function [Link]. As the first argument, this func-
tion takes the path to the directory where the files with results of individual
studies are stored. It is assumed that the file names are made of two parts:
population/study name and an extension. Thus the second argument of the
[Link] function is the vector with names of studies, and the third one
provides extension. Other arguments, ”maf”, ”call” and ”phwe” provide the
threshold for QC filtering of SNPs in individual studies.
The function does not return any value, but rather creates a new file named
POOLEDextens, where ”extens” is the argument supplied to the function, in the
source directory. To run analysis on the three files in the directory ”RData” we
can use
> [Link](dir="RData",pops=c("part1","part2","part3"),extens=".[Link]"

Population part1 , reading RData/[Link] done

Dimesions after filters are 3482 18
population part2, reading RData/[Link] done
Dimesions after filters are 3530 18
analysing ...
Lambda part1 = 0.9493588
Lambda part2 = 1.095788
Corrected Lambda part1 = 0.9493588
Corrected Lambda part2 = 1
Lambda POOLED data = 1.024196
... DONE
Dimesions after pooling are 3533 20
10.5. ANSWERS TO THE EXERCISE 235

population part3, reading RData/[Link] done

Dimesions after filters are 7444 18
analysing ...
Lambda part3 = 1.12461
Corrected Lambda part3 = 1
Lambda POOLED data = 1.345936
... DONE
Dimesions after pooling are 7444 22
$[Link]
[1] "part1" "part2" "part3"

extra arguments regulate the SNP exclusion criteria: maf=0.01 tells to exclude
SNPs with minor allele frequency less then 0.5%, call=0.95 tells to drop SNPs
with call rate less than 95%, and phwe=1.e-8 instructs to exclude SNPs with
HWE P -value ≤ 10−8 .
Now we can read and inspect the results of meta-analysis with:

> poolf <- [Link]("RData/[Link]",strings=F)

> poolf[1:5,]

name strand allele1 allele2 effallele chromosome position n npops

1 rs1000909 - A G G 2 8531681 190 1
2 rs1006092 - T G G X 13527448 190 1
3 rs100616 - G C C 1 1911712 310 3
4 rs1006497 + T G G 1 2658810 315 3
5 rs1010481 + A C C 2 8409087 190 1
beta sebeta effallelefreq call pexhwe obetapart1
1 -0.08504871 0.13198426 0.8157895 0.9793814 2.8521518 NA
2 -0.03712065 0.08510711 0.5078947 0.9793814 6.8378385 NA
3 0.11825684 0.12070003 0.1241935 0.9780568 6.5922479 0.08819818
4 -0.02359567 0.11424656 0.1603175 0.9937465 0.2313494 -0.25425751
5 -0.03241475 0.12358721 0.2657895 0.9793814 0.0000000 NA
obetapart2 obetapart3 osepart1 osepart2 osepart3 chi2
1 NA -0.085048708 NA NA 0.13198426 0.41523235
2 NA -0.037120652 NA NA 0.08510711 0.19023899
3 0.1230224623 0.121082405 0.3726827 0.2128792 0.15936396 0.95992605
4 -0.0001117686 0.005382408 0.3498751 0.2354211 0.14085503 0.04265582
5 NA -0.032414750 NA NA 0.12358721 0.06879205
p
1 0.5193256
2 0.6627178
3 0.3272055
4 0.8363747
5 0.7931037

10.5 Answers to the exercise

Perform meta-analysis of the data presented in table 10.2. Which allele is the
risk one? Is this risk significant? What is pooled Odds Ratio and 95% confidence
236 CHAPTER 10. META-ANALYSIS OF GWA SCANS

interval? Do analysis using at least two methods. Which method is better (best)
in this situation? Why?
We first need to unify Odds Ratios by using the same effective allele. Let that
be the ”risk” allele, as may be guessed from a glance to the data, namely ”Pro”.
When the effects are reported for the other, ”Ala” allele, the corresponding ORs
for the ”Pro” allele can be found using simple relation ORP ro = 1/ORAla .
Thus, the vector of Odds Ratios for ”Pro” allele is
> [Link] <- c(1./.67,0.93,1.08,1./.83,1.22,1.23)
> [Link]

[1] 1.492537 0.930000 1.080000 1.204819 1.220000 1.230000

The corresponding P − values are

> p <- c(0.013,0.6,0.84,0.40,0.25,0.07)
Let us find log-ORs
> [Link] <- log([Link])
> [Link]

[1] 0.40047757 -0.07257069 0.07696104 0.18632958 0.19885086 0.20701417

Corresponding squared standard errors are

> s2 <- [Link]*[Link]/qchisq(1-p,1)
> s2

[1] 0.02599764 0.01915121 0.14531060 0.04901514 0.02988102 0.01305349

and weights are

> w <- 1/s2
> w

[1] 38.465034 52.216009 6.881811 20.401860 33.466060 76.607876

Thus the pooled estimate of log-OR is

> [Link] <- sum(w*[Link])/sum(w)
> [Link]

[1] 0.1686548

and the standard error is

> p.s <- 1/sqrt(sum(w))
> p.s

[1] 0.06622101

Thus the pooled estimate of Odds Ratio from association between type 2
diabetes and ”Pro” allele is
> exp([Link])
10.5. ANSWERS TO THE EXERCISE 237

[1] 1.183711
and the 95% confidence interval is
> exp([Link]-1.96*p.s)
[1] 1.039627
> exp([Link]+1.96*p.s)
[1] 1.347765
The χ2 test for association and corresponding P − value are
> p.chi2 <- ([Link]/p.s)^2
> p.chi2
[1] 6.486429
> [Link] <- 1-pchisq(p.chi2,1)
> [Link]
[1] 0.01087011
Z-score pooling though may be more appropriate method for such differen-
tially designed studies (e.g. control groups are very different). To get Z-score
pooling working, we need first find Z-scores from P-values
> p <- c(0.013,0.6,0.84,0.40,0.25,0.07)
> z <- sqrt(qchisq(1-p,1))
> z
[1] 2.4837693 0.5244005 0.2018935 0.8416212 1.1503494 1.8119107
and assign the right sign (let ”+” is for the risk effect of ”Pro”).
> effsig <- c(1,-1,1,1,1,1)
> z <- z*effsig
> z
[1] 2.4837693 -0.5244005 0.2018935 0.8416212 1.1503494 1.8119107
Now, we need to assign weights to the studies as
> n <- c(221,306,71,164,242,471)
> w <- sqrt(n)
ane the pooled estimate of Z and corresponding P − value are
> zpoo <- sum(w*z)/sqrt(sum(w^2))
> zpoo
[1] 2.537333
> 1-pchisq(zpoo*zpoo,1)
[1] 0.01117008
As you can see the results are almost identical to the previous obtained with
inverse variance pooling.
238 CHAPTER 10. META-ANALYSIS OF GWA SCANS

10.5.1 Exercise 9:
Perform meta-analsys excluding the original report (study 1). Is there still
significant association between Pro12Ala and diabetes?
The answer to this exercise can be obtained in exactly the same manner, as
for the previous one, limiting our consideration to the last five studies.
Thus, the vector of Odds Ratios for ”Pro” allele is
> [Link] <- c(0.93,1.08,1./.83,1.22,1.23)
> [Link]

[1] 0.930000 1.080000 1.204819 1.220000 1.230000

The corresponding P − values are

> p <- c(0.6,0.84,0.40,0.25,0.07)
Let us find log-ORs
> [Link] <- log([Link])
> [Link]

[1] -0.07257069 0.07696104 0.18632958 0.19885086 0.20701417

Corresponding squared standard errors are

> s2 <- [Link]*[Link]/qchisq(1-p,1)
> s2

[1] 0.01915121 0.14531060 0.04901514 0.02988102 0.01305349

and weights are

> w <- 1/s2
> w

[1] 52.216009 6.881811 20.401860 33.466060 76.607876

Thus the pooled estimate of log-OR is

> [Link] <- sum(w*[Link])/sum(w)
> [Link]

[1] 0.1216172

and the standard error is

> p.s <- 1/sqrt(sum(w))
> p.s

[1] 0.07262917

Thus the pooled estimate of Odds Ratio from association between type 2
diabetes and ”Ala” allele is
> exp([Link])
10.5. ANSWERS TO THE EXERCISE 239

[1] 1.129322
and the 95% confidence interval is
> exp([Link]-1.96*p.s)
[1] 0.9794776
> exp([Link]+1.96*p.s)
[1] 1.30209
The χ2 test for association and corresponding P − value are
> p.chi2 <- ([Link]/p.s)^2
> p.chi2
[1] 2.803937
> [Link] <- 1-pchisq(p.chi2,1)
> [Link]
[1] 0.09403318
Z-score pooling though may be more appropriate method for such differen-
tially designed studies (e.g. control groups are very different). To get Z-score
pooling working, we need first find Z-scores from P-values
> p <- c(0.6,0.84,0.40,0.25,0.07)
> z <- sqrt(qchisq(1-p,1))
> z
[1] 0.5244005 0.2018935 0.8416212 1.1503494 1.8119107
and assign the right sign (let ”+” is for the risk effect of ”Pro”).
> effsig <- c(-1,1,1,1,1)
> z <- z*effsig
> z
[1] -0.5244005 0.2018935 0.8416212 1.1503494 1.8119107
Now, we need to assign weights to the studies as
> n <- c(306,71,164,242,471)
> w <- sqrt(n)
ane the pooled estimate of Z and corresponding P − value are
> zpoo <- sum(w*z)/sqrt(sum(w^2))
> zpoo
[1] 1.709151
> 1-pchisq(zpoo*zpoo,1)
[1] 0.08742297
As you can see the results are almost identical to the previous obtained with
inverse variance pooling.
240 CHAPTER 10. META-ANALYSIS OF GWA SCANS
Chapter 11

Analysis of selected region

Small data set ’srdta’, which is part of the GenABEL-package library, will be
used in this section. Start R and load the GenABEL-package and the data with
> library(GenABEL)
> data(srdta)

11.1 Exploring linkage disequilibrium

See help for r2fast.

11.2 Haplotype analysis

Use
> gtforld <- [Link](srdta[,1:5])

to convert part of your SNPs to [Link] format.

You can also use interface function to do sliding widow analysis
> h2 <- [Link]("qt1~CRSNP",srdta,snps=c(1:5))

11.3 Analysis of interactions

See help for [Link].2D and [Link].2D

241
242 CHAPTER 11. ANALYSIS OF SELECTED REGION
Appendix A

Importing data to
GenABEL-package

This section is outdated. By far the most used way is through the
TPED
As described in section 4.1, the GenABEL-package [Link]-class consists
of phenotypic data frame and an object of [Link]-class, which contains all
genetic data. To import data to GenABEL-package, you need to prepare two
files: one containing the phenotypic, and the other containing genotypic data.
The phenotype file relates study subject IDs with values of covariates and
outcomes. In the phenotypic data file, the first line gives a description (variable
name) of the data contained in a particular column; the names should better
be unique, otherwise R will change them.
The first column of the phenotype file must contain the subjects’ unique
ID, named ”id”. The IDs listed here, and in the genotypic data file, must be
the same. It is recommended that the id names are given in quotation marks
(see example below), which will save you a possible troubles with e.g. leading
zeros.
There also should also be a column named ”sex” and giving sex information
(0 = female, 1 = male). Other columns in the file should contain phenotypic
information.
Missing values should be coded with ”NA”; binary traits should have values
0 or 1.
All subjects present in the genotypic files must be listed in the phenotypic
file as well, because sex information provided by the phenotypic file is an essen-
tial part of the genotypic QC procedure.
An example of few first lines of a phenotype file is as follows:

id sex age bt1 qt qt1

"cd289982" 0 30.33 NA NA 3.93
"cd325285" 0 36.514 1 0.49 3.61
"cd357273" 1 37.811 0 1.65 5.30
"cd872422" 1 20.393 0 1.95 4.07
"cd1005389" 1 28.21 1 0.35 3.90

This file tells us that, for example, person 325286 is female (0 in second

243
244 APPENDIX A. IMPORTING DATA TO GENABEL-PACKAGE

column), and she has ”1” (usually this means a ”case”) value for the trait
”bt1”, so on. Person 289982 has measurements only for sex, age and qt1, while
the other measurements are missing (NA, Not Available).
If you need to add phenotypes to an already created [Link]-class, you
can use the [Link] function. This function allows you to add variables
contained in some data frame to the existing data@phdata object. The data
frame to be added should contain an ”id” variable, identical to that existing in
the object, and should not contain any other variables with names identical
to those that already exist.
The second file you need should contains genotypic data. As described in sec-
tion 4.1 (”General description of [Link]-class”, page 79), GenABEL-package
[Link]-class contains different types of information. For every SNP, informa-
tion on map position, chromosome, and strand should be provided. For every
person, every SNP genotype should be provided. GenABEL-package provides a
number of function to convert these data from different formats to the inter-
nal GenABEL-package raw format. We will first consider our preferred format,
which we informally call ”Illumina”-like.

A.1 Converting from preferred format

We will consider use of [Link] procedure; details of other pro-
cedures are given later. Note that what we call ”illumina” format is not really a
proprietary format from that company, it is just one of the possible text output
format from the Illumina BeadStudio; similar formats are accepted/generated
by HapMap and Affymetrix.
The file of the ”Illumina” format contains SNPs in rows and IDs in columns.
The first line is a ”header”, containing column names. The first three columns
should contain information on SNP name, chromosome, and position. There is
an optional (though highly recommended!) fourth column, containing strand
information (acceptable codes: ”+”, ”-”, ”u”, the last stands for ”unknown”).
After that column, each of the residual ones corresponds to an individual, with
ID as the column name. Genotypes should be presented by two consecutive
characters (no separator).
An example of few first lines of the ”illumina” genotypic file is as follows:

name chr pos strand "cd289982" "cd325285" "cd357273" "cd872422" "cd1005389"

rs1001 1 1235 + AA AG AG AA GG
rs6679 9 2344 + GT GG GG TG GG
rs2401 22 3455 + AA CC CC CC AC
rs123 X 32535 - TT GT TT TT TT
rs6679 XY 2344 - GT GG GG TG GG
rs876 Y 23556 + 00 00 TT GG TT
mitoA1 mt 24245 - AA CC 00 00 00

It is clear that is not quite conventional Illumina file – because in BeadStudio

the alleles are reported using the ”top” strand; rather, this is an Affymetrix
or HapMap-type of a file. Anyways, this file contains all required genotypic
information, and this file format is the preferred one for import. Assume that
the file with the genotypic data is called ”[Link]”, and is stored in the directory
”RData”.
A.1. CONVERTING FROM PREFERRED FORMAT 245

First, start R and load GenABEL-package:

> library(GenABEL)
You can convert the data to GenABEL-package raw format by

> [Link](inf="RData/[Link]",
+ out="RData/[Link]",
+ strand="file")

Reading genotypes from file 'RData/[Link]' ...

Writing to file 'RData/[Link]' ...
... done.

Here is the content of the converted file ”[Link]” – internal raw data
representation:
Note the option strand="file" – it is telling GenABEL-package that strand
information is provided in the file.
At this moment, you can load the data into GenABEL-package. Assume
that the phenotypic file described above is called ”[Link]” and the converted
genotypic file in the raw GenABEL-package format is called ”[Link]”. You
can load the data using the command
> df <- [Link](phe="RData/[Link]",
+ gen="RData/[Link]",
+ force=TRUE)

ids loaded...
marker names loaded...
chromosome data loaded...
map data loaded...
allele coding data loaded...
strand data loaded...
genotype data loaded...
[Link] object created...
assignment of [Link] object FORCED; X-errors were not checked!

The option ”force=TRUE” tells that GenABEL-package should load the data
even if it fins sex errors.
You can inspect the loaded data; let us first look into phenotypic data by by
> df@phdata

id sex age bt1 qt qt1

cd289982 cd289982 0 30.330 NA NA 3.93
cd325285 cd325285 0 36.514 1 0.49 3.61
cd357273 cd357273 1 37.811 0 1.65 5.30
cd872422 cd872422 1 20.393 0 1.95 4.07
cd1005389 cd1005389 1 28.210 1 0.35 3.90

. . . than check that the genotypes have imported right:

> g0 <- [Link](df@gtdata)
> g0
246 APPENDIX A. IMPORTING DATA TO GENABEL-PACKAGE

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

cd289982 "A/A" "G/T" "A/A" "A/A" "T/T" "G/T" NA
cd325285 "A/G" "G/G" "C/C" "C/C" "T/G" "G/G" NA
cd357273 "A/G" "G/G" "C/C" NA "T/T" "G/G" "T/T"
cd872422 "A/A" "G/T" "C/C" NA "T/T" "G/T" "G/G"
cd1005389 "G/G" "G/G" "C/A" NA "T/T" "G/G" "T/T"

> [Link](df@gtdata@strand)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"+" "+" "+" "-" "-" "-" "+"

> [Link](df@gtdata@coding)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"AG" "GT" "CA" "CA" "TG" "GT" "TG"

In a real Illumina file, a coding on the TOP strand is supplied. Then, the
file will normally look like
name chr pos "cd289982" "cd325285" "cd357273" "cd872422" "cd1005389"
rs1001 1 1235 AA AG AG AA GG
rs6679 9 2344 GT GG GG TG GG
rs2401 22 3455 AA CC CC CC AC
rs123 X 32535 TT GT TT TT TT
rs6679 XY 2344 GT GG GG TG GG
rs876 Y 23556 00 00 TT GG TT
mitoA1 mt 24245 AA CC 00 00 00
and the conversion command will be
> [Link](inf="RData/[Link]",
+ out="RData/[Link]",
+ strand="+")

Reading genotypes from file 'RData/[Link]' ...

Writing to file 'RData/[Link]' ...
... done.

In this particular data set, after conversion, the ”+” strand will actually mean
not ”forward”, but TOP – something you should remember for this particular
data. The resulting file will look like this:
You can load the data with
> df <- [Link](phe="RData/[Link]",
+ gen="RData/[Link]",
+ force=TRUE)

ids loaded...
marker names loaded...
chromosome data loaded...
map data loaded...
allele coding data loaded...
A.2. CONVERTING PLINK TPED FILES 247

strand data loaded...

genotype data loaded...
[Link] object created...
assignment of [Link] object FORCED; X-errors were not checked!

Obviously, the ”strand” is always ”+” (here it means TOP):

> g1 <- [Link](df@gtdata)

> g1

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

cd289982 "A/A" "G/T" "A/A" "A/A" "T/T" "G/T" NA
cd325285 "A/G" "G/G" "C/C" "C/C" "T/G" "G/G" NA
cd357273 "A/G" "G/G" "C/C" NA "T/T" "G/G" "T/T"
cd872422 "A/A" "G/T" "C/C" NA "T/T" "G/T" "G/G"
cd1005389 "G/G" "G/G" "C/A" NA "T/T" "G/G" "T/T"

> [Link](df@gtdata@strand)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"+" "+" "+" "+" "+" "+" "+"

> [Link](df@gtdata@coding)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"AG" "GT" "CA" "CA" "TG" "GT" "TG"

We can see that the genotypes are identical to ones we imported previously,
as should be the case:

> g0 == g1

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

cd289982 TRUE TRUE TRUE TRUE TRUE TRUE NA
cd325285 TRUE TRUE TRUE TRUE TRUE TRUE NA
cd357273 TRUE TRUE TRUE NA TRUE TRUE TRUE
cd872422 TRUE TRUE TRUE NA TRUE TRUE TRUE
cd1005389 TRUE TRUE TRUE NA TRUE TRUE TRUE

A.2 Converting PLINK tped files

PLINK TPED (transposed-pedigree) format bears notable similarity to what
we call ”Illumina” format, with few exceptions: (1) there is no header line
giving field names (and therefore IDs are stored in a separate file) (2) the first
column gives chromosome, second – SNP name, third genetic map (usually kept
as zeroes), the fourth – physical position, and, starting with the fifth column,
genotypic data are listed, (3) finally, within a genotypes, alleles are separated
with a space. In TPED format, the data we already worked with would look
like
248 APPENDIX A. IMPORTING DATA TO GENABEL-PACKAGE

1 rs1001 0 1235 A A A G A G A A G G
9 rs6679 0 2344 G T G G G G T G G G
22 rs2401 0 3455 A A C C C C C C A C
X rs123 0 32535 T T G T T T T T T T
XY rs6679 0 2344 G T G G G G T G G G
Y rs876 0 23556 0 0 0 0 T T G G T T
mt mitoA1 0 24245 A A C C 0 0 0 0 0 0
Obviously, a separate file is needed to keep correspondence between geno-
types and IDs. This file emulated standard pedigree file without a header line.
The file, conventionally called a TFAM-file, should contain six columns, cor-
responding to pedigree ID, ID, father, mother, sex, and affection. Only the
second column is used by GenABEL-package – please make sure you use unique
IDs. Consequently, it does not matter what pedigree ID, father/mother, sex, or
affection status you assign in the file – the real information is coming from the
phenotypic data file. The TFAM file for our data will look like this:
1 cd289982 0 0 1 0
1 cd325285 0 0 1 0
1 cd357273 0 0 1 0
1 cd872422 0 0 1 0
1 cd1005389 0 0 1 0
You can convert the data from PLINK TPED format to the GenABEL-package
format using command [Link]:
> [Link](tped="RData/[Link]",
+ tfam="RData/[Link]",
+ out="RData/[Link]",
+ strand="+")

Reading individual ids from file 'RData/[Link]' ...

... done. Read 5 individual ids from file 'RData/[Link]'
Reading genotypes from file 'RData/[Link]' ...
...done. Read 7 SNPs from file 'RData/[Link]'
Writing to file 'RData/[Link]' ...
... done.

and load the data with

> df <- [Link](phe="RData/[Link]",
+ gen="RData/[Link]",
+ force=TRUE)

ids loaded...
marker names loaded...
chromosome data loaded...
map data loaded...
allele coding data loaded...
strand data loaded...
genotype data loaded...
[Link] object created...
assignment of [Link] object FORCED; X-errors were not checked!
A.3. CONVERTING LINKAGE-LIKE FILES 249

Obviously, the ”strand” is always ”+” (meaning TOP):

> g1 <- [Link](df@gtdata)
> g1

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

cd289982 "A/A" "G/T" "A/A" "A/A" "T/T" "G/T" NA
cd325285 "A/G" "G/G" "C/C" "C/C" "T/G" "G/G" NA
cd357273 "A/G" "G/G" "C/C" NA "T/T" "G/G" "T/T"
cd872422 "A/A" "G/T" "C/C" NA "T/T" "G/T" "G/G"
cd1005389 "G/G" "G/G" "C/A" NA "T/T" "G/G" "T/T"

> [Link](df@gtdata@strand)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"+" "+" "+" "+" "+" "+" "+"

> [Link](df@gtdata@coding)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"AG" "GT" "CA" "CA" "TG" "GT" "TG"

We can see that the genotypes are identical to ones we imported previously,
as should be the case:
> g0 == g1

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

cd289982 TRUE TRUE TRUE TRUE TRUE TRUE NA
cd325285 TRUE TRUE TRUE TRUE TRUE TRUE NA
cd357273 TRUE TRUE TRUE NA TRUE TRUE TRUE
cd872422 TRUE TRUE TRUE NA TRUE TRUE TRUE
cd1005389 TRUE TRUE TRUE NA TRUE TRUE TRUE

A.3 Converting linkage-like files

Linkage-like files, also known as pre-makeped files, or pedigree files, represent a
historic format which dates back to the time when only few markers could be
typed – thus the number of subjects was usually greater than the number of
markers. In that situation, it was natural and obvious to keep IDs in rows and
markers in columns. In the first six columns, standard linkage-like file would
contain pedigree ID, ID, father’s ID, mother’s ID, sex (coded as 1 = male and
2 = female), and affection status (0 = unknown, 1 = unaffected, 2 = affected).
In the following columns, genotypic information is provided. Alleles of the same
genotype could be separated by a space, or by a slash (”/”). Thus the data we
are working with could be presented as
1 cd289982 0 0 1 0 A A G T A A T T G T 0 0 A A
1 cd325285 0 0 1 0 A G G G C C G T G G 0 0 C C
1 cd357273 0 0 1 0 A G G G C C T T G G T T 0 0
1 cd872422 0 0 1 0 A A T G C C T T T G G G 0 0
1 cd1005389 0 0 1 0 G G G G A C T T G G T T 0 0
250 APPENDIX A. IMPORTING DATA TO GENABEL-PACKAGE

As you can see, this file misses header line, and information what are the SNP
names, position, etc. should be provided in a separate MAP-file. GenABEL-package
accepts map in Merlin format, and an extended format. A map in Merlin format
consist of header line, giving column names, and three columns with chromo-
some, name and position information, for example:

chr name pos

1 rs1001 1235
9 rs6679 2344
22 rs2401 3455
X rs123 32535
XY rs6679 2344
Y rs876 23556
mt mitoA1 24245

The data can be converted to the internal GenABEL-package format with

> [Link](pedfile="RData/[Link]",
+ mapfile="RData/[Link]",
+ out="RData/[Link]",
+ strand="+")

Reading map from file 'RData/[Link]' ...

... done. Read positions of 7 markers from file 'RData/[Link]'
Reading genotypes from file 'RData/[Link]' ...
...done. Read information for 5 people from file 'RData/[Link]'
Analysing marker information ...
Writing to file 'RData/[Link]' ...
... done.

and loaded with

> df <- [Link](phe="RData/[Link]",

+ gen="RData/[Link]",
+ force=TRUE)

ids loaded...
marker names loaded...
chromosome data loaded...
map data loaded...
allele coding data loaded...
strand data loaded...
genotype data loaded...
[Link] object created...
assignment of [Link] object FORCED; X-errors were not checked!

We can inspect the genotypic data and check that conversion results are
identical to previous runs with

> g1 <- [Link](df@gtdata)

> g1
A.3. CONVERTING LINKAGE-LIKE FILES 251

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

cd289982 "A/A" "G/T" "A/A" "A/A" "T/T" "G/T" NA
cd325285 "A/G" "G/G" "C/C" "C/C" "T/G" "G/G" NA
cd357273 "A/G" "G/G" "C/C" NA "T/T" "G/G" "T/T"
cd872422 "A/A" "G/T" "C/C" NA "T/T" "G/T" "G/G"
cd1005389 "G/G" "G/G" "C/A" NA "T/T" "G/G" "T/T"

> [Link](df@gtdata@strand)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"+" "+" "+" "+" "+" "+" "+"

> [Link](df@gtdata@coding)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"AG" "GT" "CA" "CA" "TG" "GT" "TG"

> g0 == g1

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

cd289982 TRUE TRUE TRUE TRUE TRUE TRUE NA
cd325285 TRUE TRUE TRUE TRUE TRUE TRUE NA
cd357273 TRUE TRUE TRUE NA TRUE TRUE TRUE
cd872422 TRUE TRUE TRUE NA TRUE TRUE TRUE
cd1005389 TRUE TRUE TRUE NA TRUE TRUE TRUE

If you are willing to import strand information, you can make use of the
extended map format. In this format the strand information is added to the
map-file:
chr name pos strand coding
1 rs1001 1235 + AG
9 rs6679 2344 + TG
22 rs2401 3455 + AC
X rs123 32535 - GT
XY rs6679 2344 - GT
Y rs876 23556 + GT
mt mitoA1 24245 - AC
The data can be converted to the internal GenABEL-package format with
> [Link](pedfile="RData/[Link]",
+ mapfile="RData/[Link]",
+ out="RData/[Link]",
+ strand="file")

Reading map from file 'RData/[Link]' ...

... done. Read positions of 7 markers from file 'RData/[Link]'
Reading genotypes from file 'RData/[Link]' ...
...done. Read information for 5 people from file 'RData/[Link]'
Analysing marker information ...
Writing to file 'RData/[Link]' ...
... done.
252 APPENDIX A. IMPORTING DATA TO GENABEL-PACKAGE

Note that option strand==file was used to specify that the extended map
format should be used. The data can be loaded with

> df <- [Link](phe="RData/[Link]",

+ gen="RData/[Link]",
+ force=TRUE)

ids loaded...
marker names loaded...
chromosome data loaded...
map data loaded...
allele coding data loaded...
strand data loaded...
genotype data loaded...
[Link] object created...
assignment of [Link] object FORCED; X-errors were not checked!

We can inspect the genotypic data and check that conversion results are
identical to previous runs with

> g1 <- [Link](df@gtdata)

> g1

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

cd289982 "A/A" "G/T" "A/A" "A/A" "T/T" "G/T" NA
cd325285 "A/G" "G/G" "C/C" "C/C" "T/G" "G/G" NA
cd357273 "A/G" "G/G" "C/C" NA "T/T" "G/G" "T/T"
cd872422 "A/A" "G/T" "C/C" NA "T/T" "G/T" "G/G"
cd1005389 "G/G" "G/G" "C/A" NA "T/T" "G/G" "T/T"

> [Link](df@gtdata@strand)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"+" "+" "+" "-" "-" "-" "+"

> [Link](df@gtdata@coding)

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

"AG" "GT" "CA" "CA" "TG" "GT" "TG"

> g0 == g1

rs1001 rs6679 rs2401 mitoA1 rs123 rs6679 rs876

cd289982 TRUE TRUE TRUE TRUE TRUE TRUE NA
cd325285 TRUE TRUE TRUE TRUE TRUE TRUE NA
cd357273 TRUE TRUE TRUE NA TRUE TRUE TRUE
cd872422 TRUE TRUE TRUE NA TRUE TRUE TRUE
cd1005389 TRUE TRUE TRUE NA TRUE TRUE TRUE
A.4. CONVERTING FROM MACH FORMAT 253

A.4 Converting from MACH format

The data from MACH format can be converted by using [Link].
This function actually calls [Link] in specific format. MACH soft-
ware is widely used for SNP imputations. For our needs we consider two files
produced by MACH: pedigree file with (the imputed) genotypic data, and info-
file, containing information about quality of imputations for particular SNP.
SEE HELP FOR [Link] for further details.

A.5 Converting from text format

254 APPENDIX A. IMPORTING DATA TO GENABEL-PACKAGE
Appendix B

GenABEL internals

B.1 Internal structure of [Link]-class

Start R and load GenABEL-package library using command

> library(GenABEL)

After that, load the data with the command

> data(srdta)
The object you have loaded, srdta, belongs to the [Link] class. This is
a special class developed to facilitate GWA analysis.
In GWA analysis, different types of data are used. These include the pheno-
typic and genotypic data on the study participants and chromosome and location
of every SNP. For every SNP, it is desirable to know the details of coding (what
are alleles? – A, T, G, C? – and what is the strand – ’+’ or ’-’, ’top’ or ’bot’ ?
– this coding is for).
One could attempt to store all phenotypes and genotypes together in a single
table, using, e.g. one row per study subject; than the columns will correspond
to study phenotypes and SNPs. For a typical GWA data set, this would lead
to a table of few thousands rows and few hundreds of thousands of columns.
Such a format is generated when one downloads HapMap data for a region. To
store GWA data in such tables internally, within R, proves to be inefficient. In
GenABEL-package, special data class, [Link]-class is used to store GWA
data. The structure of this data class is shown at the figure B.1.
An object of some class has ”slots” which may contain actual data or objects
of other classes. The information stored at a particular slot of an object can
be accessed by command object@slot.
At the first level, a [Link]-class object has slot phdata, which con-
tains all phenotypic information in a data frame ([Link]-class object).
The rows of this data frame correspond to study subjects, and the columns
correspond to the variables. There are two default variables, which are always
present in phdata. The first of these is ”id”, which contains study subject iden-
tification code. This identification code can be arbitrary character, but every
person must be coded with an unique ID. The second default variable is ”sex”,
where males are coded with ones (”1”) and females are coded with zero (”0”).

255
256 APPENDIX B. GENABEL INTERNALS

object
[Link]-class
All GWA data
object@phdata
[Link]-class
Phenotypic data

object@gtdata@nsnps
integer
# of SNPs in study
object@gtdata@nids
integer
# of people in study object@gtdata@snpnames
object@gtdata vector of character
[Link]-class IDs of study SNPs
object@gtdata@idnames All genetic data
vector of character
IDs of study participants
object@gtdata@chromosome
vector of character
object@gtdata@male Chromosome label (1, 2, ... X)
vector of integer
Sex (1=male, 0=female)
object@gtdata@map
object@gtdata@coding vector of double
[Link]-class SNPs map positions
SNP allele coding (”AG”, “AC”, ...)

object@gtdata@gtps object@gtdata@strand
[Link]-class [Link]-class
Genotypic data in compressed format SNP allele strand (”+”, “-”)

Figure B.1: Structure of [Link]-class. In every box, first line contains the
object and slot names, second line describes the class of this object, and third
line describes what information is contained.
B.1. INTERNAL STRUCTURE OF [Link]-CLASS 257

It is important to understand that this data frame is not supposed to be directly

modified by the user. In particular, it is extremely important to remember that
one should not directly add subjects to the table, change the values of ”id” and
”sex”, and change the order of subjects in phdata unless this one is really un-
derstands the way GenABEL-package works. One also should not run such data
manipulation functions as merge, cbind and rbind – exactly because they may
change the number of subjects or interfere with the order. On the other hand,
it is OK to add more variables to the data frame through direct computations,
for example, if one wishes to add computed body mass index, it is OK to run
the command like
obj@phdata$bmi <- obj@phdata$weight/((obj@phdata$height)^ 2)
To add many variables to phdata, special GenABEL function [Link]
should be used.
The other slot of an object of [Link]-class is slot gtdata, which con-
tains all GWA genetic information in an object of class [Link] class (figure
B.1). This class, in turn, has slots nids, containing the number of study sub-
jects, idnames, containing all ID names of these subjects, nsnps, containing
the number of SNPs typed, snpnames, containing the SNP names, chromosome,
containing the name of the chromosome the SNPs belong to and slot map with
map position of SNPs, and slot male, containing the sex code for the subjects
(1=male, 0=female). The latter is identical to the ”sex” variable contained in
the phdata, but is duplicated here because many operations with purely genetic
data, in particular these concerning analysis of sex chromosomes, depend on the
sex. The strand information is presented in the slot strand. GenABEL-package
codes strand as ”+” (forward), ”-” (reverse) or ”u” (unknown). Of cause, if you
prefer top/bottom coding, this information may be stored in the same form –
you will just need to remember that ”+” corresponds to e.g. ”top”, and ”-”
to ”bottom” strand. The allelic coding is presented in slot coding. Coding
for every allele is presented with a pair of characters, for example ”AG”. Thus,
for such polymorphism, you may expect ”AA”, ”AG” and ”GG” genotypes to
be found in population. The order (that is ”AG” vs ”GA”) is important –
the first allele reported is the one which will be used as a reference in associa-
tion analysis, and thus the effects are reported for the second allele. To avoid
memory overheads, the strand and coding information is internally stored as
[Link]-class and [Link]-class. Information can be converted to
human-readable format using [Link] function.
If, for example, you would like to know, how many SNPs were included in
the study (slot nsnps of the slot gtdata of srdta), you need to run command

> srdta@gtdata@nsnps

[1] 833

Thus, 833 SNPs were typed in the study. You can access information stored in
any slot in this manner.
You may want to read the general GenABEL-package man page using help(GenABEL).
To see help on [Link]-class, you can use help("[Link]-class") (mind
the quotation marks!).
258 APPENDIX B. GENABEL INTERNALS

Summary:

• An object of some class has ”slots” which may contain actual data or
objects of other classes. The information stored at a particular slot of
an object can be accessed by command object@slot.
• GenABEL-package uses special data class, [Link]-class, to store
GWA data.
Bibliography

Amin, N., C. M. van Duijn, and Y. S. Aulchenko, 2007 A genomic back-

ground based method for association analysis in related individuals. PLoS
One 2: e1274.

Astle, W. and D. J. Balding, (in press) Population structure and cryptic

relatedness in genetic association studies. Statistical Science .

Aulchenko, Y. S., D.-J. de Koning, and C. Haley, 2007a Genomewide

rapid association using mixed model and regression: a fast and simple method
for genomewide pedigree-based quantitative trait loci association analysis.
Genetics 177: 577–585.

Aulchenko, Y. S., S. Ripke, A. Isaacs, and C. M. van Duijn, 2007b

Genabel: an r library for genome-wide association analysis. Bioinformatics
23: 1294–1296.

Aulchenko, Y. S. and M. V. Struchalin, submitted Probabel package for

genome-wide association analysis of imputed data .

Bacanu, S. A., B. Devlin, and K. Roeder, 2000 The power of genomic

control. Am J Hum Genet 66: 1933–1944.

Bacanu, S.-A., B. Devlin, and K. Roeder, 2002 Association studies for

quantitative traits in structured populations. Genet Epidemiol 22: 78–93.

Boerwinkle, E., R. Chakraborty, and C. F. Sing, 1986 The use of mea-

sured genotype information in the analysis of quantitative phenotypes in man.
i. models and analytical methods. Ann Hum Genet 50: 181–194.

Campbell, C. D., E. L. Ogburn, K. L. Lunetta, H. N. Lyon, M. L.

Freedman, L. C. Groop, D. Altshuler, K. G. Ardlie, and J. N.
Hirschhorn, 2005 Demonstrating stratification in a european american pop-
ulation. Nat Genet 37: 868–872.

Chen, W.-M. and G. R. Abecasis, 2007 Family-based association tests for

genomewide association scans. Am J Hum Genet 81: 913–926.

Devlin, B. and K. Roeder, 1999 Genomic control for association studies.

Biometrics 55: 997–1004.

Devlin, B., K. Roeder, and L. Wasserman, 2001 Genomic control, a new

approach to genetic-based association studies. Theor Popul Biol 60: 155–166.

259
260 BIBLIOGRAPHY

Fisher, R. A., 1918 The correlation between relatives on the supposition of

mendelian inheritance. Trans R Soc Edinb 52: 399–433.

Freedman, M. L., D. Reich, K. L. Penney, G. J. McDonald, A. A.

Mignault, N. Patterson, S. B. Gabriel, E. J. Topol, J. W. Smoller,
C. N. Pato, M. T. Pato, T. L. Petryshen, L. N. Kolonel, E. S. Lan-
der, P. Sklar, B. Henderson, J. N. Hirschhorn, and D. Altshuler,
2004 Assessing the impact of population stratification on genetic association
studies. Nat Genet 36: 388–393.

Nelis, M., T. Esko, R. Mgi, F. Zimprich, A. Zimprich, D. Toncheva,

S. Karachanak, T. Piskckov, I. Balasck, L. Peltonen, E. Jakkula,
K. Rehnstrm, M. Lathrop, S. Heath, P. Galan, S. Schreiber,
T. Meitinger, A. Pfeufer, H.-E. Wichmann, B. Melegh, N. Polgr,
D. Toniolo, P. Gasparini, P. D’Adamo, J. Klovins, L. Nikitina-
Zake, V. Kucinskas, J. Kasnauskiene, J. Lubinski, T. Debniak,
S. Limborska, A. Khrunin, X. Estivill, R. Rabionet, S. Marsal,
A. Juli, S. E. Antonarakis, S. Deutsch, C. Borel, H. Attar,
M. Gagnebin, M. Macek, M. Krawczak, M. Remm, and A. Metspalu,
2009 Genetic structure of europeans: a view from the north-east. PLoS One
4: e5472.

Novembre, J. and M. Stephens, 2008 Interpreting principal component anal-

yses of spatial population genetic variation. Nat Genet 40: 646–649.

Pardo, L. M., I. MacKay, B. Oostra, C. M. van Duijn, and Y. S.

Aulchenko, 2005 The effect of genetic drift in a young genetically isolated
population. Ann Hum Genet 69: 288–295.

Patterson, N., A. L. Price, and D. Reich, 2006 Population structure and

eigenanalysis. PLoS Genet 2: e190.

Price, A. L., N. J. Patterson, R. M. Plenge, M. E. Weinblatt, N. A.

Shadick, and D. Reich, 2006 Principal components analysis corrects for
stratification in genome-wide association studies. Nat Genet 38: 904–909.

Rudan, I., N. Smolej-Narancic, H. Campbell, A. Carothers,

A. Wright, B. Janicijevic, and P. Rudan, 2003 Inbreeding and the ge-
netic complexity of human hypertension. Genetics 163: 1011–1021.

Ruiz-Narvez, E., 2005 Is the ala12 variant of the pparg gene an ”unthrifty
allele”? J Med Genet 42: 547–550.

Sasieni, P. D., 1997 From genotypes to genes: doubling the sample size. Bio-
metrics 53: 1253–1261.

Sladek, R., G. Rocheleau, J. Rung, C. Dina, L. Shen, D. Serre,

P. Boutin, D. Vincent, A. Belisle, S. Hadjadj, B. Balkau,
B. Heude, G. Charpentier, T. J. Hudson, A. Montpetit, A. V.
Pshezhetsky, M. Prentki, B. I. Posner, D. J. Balding, D. Meyre,
C. Polychronakos, and P. Froguel, 2007 A genome-wide association
study identifies novel risk loci for type 2 diabetes. Nature 445: 881–885.
BIBLIOGRAPHY 261

Wahlund, S., 1928 Zusammensetzung von population und korrelationserschei-

nung vom standpunkt der vererbungslehre aus betrachtet. Hereditas 11: 65–
106.

Yu, J., G. Pressoir, W. H. Briggs, I. V. Bi, M. Yamasaki, J. F. Doe-

bley, M. D. McMullen, B. S. Gaut, D. M. Nielsen, J. B. Holland,
S. Kresovich, and E. S. Buckler, 2006 A unified mixed-model method
for association mapping that accounts for multiple levels of relatedness. Nat
Genet 38: 203–208.

Zaykin, D. V., Z. Meng, and M. G. Ehm, 2006 Contrasting linkage-

disequilibrium patterns between cases and controls as a novel association-
mapping method. Am J Hum Genet 78: 737–746.
Zheng, G., B. Freidlin, Z. Li, and J. L. Gastwirth, 2005 Genomic control
for association studies under various genetic models. Biometrics 61: 186–192.