Chapter 1
Introduction
1
1. What is non-parametric method?
Non-parametric methods are methods which analyze data
samples (test hypotheses, find estimates etc) without
making any rigid assumptions about the type of
distribution from which the data is obtained.
- Most of the classical methods such as confidence intervals
for the mean, t-tests, and least squares regression operate
on the assumption that the data comes from a normal
distribution
- Non- parametric methods attempt to get away from putting
too many assumptions on the model.
2
Example 1: ANOVA
Yijk i j ijk
To perform F-test, you need to assume ijk ~ N ( 0, 2
).
Example 2: Regression
Y X
For statistical inferences, you need to assume ~ N (0, 2 ).
The normal assumption will no longer be required if you
employ nonparametric/distribution-free methods.
3
2. Non-parametric vs distribution-free
• The terms non-parametric and distribution-free are not
synonymous.
– Nonparametric methods were developed to be used in cases when
the researcher knows nothing about the parameters of the variable
of interest in the population. They do not rely on the estimation of
parameters (such as the mean or the standard deviation) describing
the distribution of the variable of interest in the population.
– A distribution-free procedure is one which makes no assumptions
about the precise form of the population.
• Popular usage, however, has equated the terms and they
will be interchangeable throughout this course.
4
3. Classical vs distribution-free methods
• Aim of a distribution-free /nonparametric method is to be
able to extract all the relevant information in the data
without having rely on a distributional assumption.
• In classical methods, we have model assumption (with
distribution).
Example:
Guess a model: Y X
~ N (0, 2 )
Based on this model, we perform analysis.
You may not have confidence about the conclusion because you
have made a guess on the model.
5
• What is driving this?
Model or data?
– Classical methods place the emphasis on the model.
– Distribution-free methods place the emphasis on data.
6
• In terms of estimation, model based methods in general
provide low variance but high bias ( if the model is wrong)
estimates. Data based methods provide high variance but
low bias estimates.
• Trade off between these two types of methods.
• If you use a classical method with the wrong assumptions,
are the conclusions valid?
– Your conclusions will be valid provided that the
assumptions are “close” to the truth.
7
Example: t-test
Two i.i.d. samples: {Xi , i=1, …n} {Yj, j=1,…m}
Assume X i ~ N ( 1 , 2 ), Y j ~ N ( 2 , )
2
Test H0:1 2 vs H1: 1 2
X n Ym
t ~ t n m 2
1 1
S2( )
n m
t t n/m2 2 .
We reject H0 if
- Can control
/2
Type I error
P( t t n m 2 | H 0 true)
- The test is /most
2
powerful
P ( t t n m 2 | H1 true)is as large as possible
8
This result is good, but what will happen if the assumptions are
not satisfied?
SupposeE ( X i ) 1 , E (Y j ) 2 , Var ( X i ) Var (Y j ) 2 .
If the normal assumption is not satisfied,
then P( t t nm
/2
2 H 0 is true )
.
Probably acceptable ( ) if n and m are large or
distributions of Xi and Yj are “close” to normal.
If W is a statistic with the distribution free property,
P(reject H0 based on W | H0 is true) = for any distribution.
9
Classical methods are popular because
1. Elegant theory
2. Easy to implement
- not so important nowadays because
computers can now handle very large data set
10
Reasons nonparametric fell out of favor:
1. Always taking the ranks in throwing away information.
2. Emphasis is placed on testing hypothesis, estimation is
harder.
3. Difficult to implement in complex problems (eg.
multivariate problems).
4. P( reject H0 | H0 is true) = for all distribution is too
restrictive because the test is then valid for any arbitrary
distribution:
A more sensible procedure:
P( reject H0 | H0 is true) = for all “sensible” distributions.
11
Reasons nonparametric becomes popular:
1. Require a few assumptions
2. Easy to understand
3. Theories have been established
4. Theoretical investigations suggest that using
ranks will not lose efficiency
5. Because of the rapid development of computers,
methods are easier to apply nowadays.
12
Remarks:
1. If you believe your model is correct, use
classical methods.
2. If your model is “approximately correct” and
the classical method you employed is robust,
then it is acceptable to employ classical
methods.
3. If you have no confidence on your model, use
nonparametric methods.
4. If data come in the form of ranks, then may
need to use nonparametric methods that are
based on ranks.
13
4. Ranks
When there are n observations:
Smallest observation: Rank 1,
2nd smallest observation: Rank 2,
The largest observation: Rank n.
Example:
Data: 7.3 2.1 6.9 11.3
Rank: 3 1 2 4
Whatever the distribution is, same set of ranks.
Note: Monotonic transformation of the observations
preserves the ranks. 14
Ties
Theoretically, if distribution is continuous, then there will
have no tied observations.
In practice, ties occur.
Use average of possible ranks when there are ties.
Data: 4.1 2.2 4.1 5.6
Rank: 2.5 1 2.5 4
2.5 is the average of (rank) 2 and (rank) 3.
15
5. Resource for self-learning
- What is distribution free test ?
[Link]
(a) Click “Overview” under “Contents”
A very brief description of distribution free test
(b) Click “Nonparametric tests” under “Other Sites”, by
Bruce Weaver
- A brief description of the test procedures that we will
introduce in this course
- A brief overview of nonparametric statistics:
Click “Nonparametric Statistics” in:
[Link]
- A very useful on-line statistics reference book
- read the section “When to use which method”.
16