P8130: Biostatistical Methods I
Methods of Inference: Two-Sample Means
Instructor: Vahe Khachadourian, MD, MPH, PhD
Methods of Inference: Two-Sample Means
This module focuses on inferential methods concerning two-sample
means (independent and paired samples).
You will learn how to:
• Construct a confidence interval to estimate the population mean
difference.
• Conduct a hypothesis test to compare the two population means.
• Conduct a hypothesis test for the equality of variances.
Two-Sample Inference
• Thus far we were only interested in one parameter from a single
population (one sample)
• What if we want to compare parameters in more than one population?
• In a two-sample setting, we want to make inferences about the
parameters of two populations (both values unknown)
• First, we need to decide if the two samples are Independent or Paired
The Paired t-test
Example: You are testing a new blood pressure medication on a sample of n randomly
selected patients that came for regular consults at CUMC in 2020. Systolic blood
pressure (SBP) measurements were recorded at the first visit and 6 months later for all
subjects.
Data structure: paired samples
Visit 1 Visit 2 Differences
𝑋!! → 𝑋!" 𝑑! = 𝑋!! − 𝑋!"
𝑋"! → 𝑋"" 𝑑" = 𝑋"! − 𝑋""
⋮ ⋮ ⋮
𝑋#! → 𝑋#" 𝑑# = 𝑋#! − 𝑋#"
The Paired t-test
Assumptions:
Visit 1 Visit 2 Differences
SBP measurements are normally distributed with
mean 𝜇! and 𝜇", for visit 1 and 2, respectively. 𝑋!! → 𝑋!" 𝑑! = 𝑋!! − 𝑋!"
𝑋"! → 𝑋"" 𝑑" = 𝑋"! − 𝑋""
If follows that the differences 𝑑! ~𝑁 ∆, 𝜎"# , 𝑖 = 1,2, … , 𝑛.
⋮ ⋮
𝑋#! → 𝑋#" 𝑑# = 𝑋#! − 𝑋#"
The problem reduces to one-sample t-test based
on the differences 𝑑$ (variance unknown), where:
"
𝑑̅ = ∑#$%! 𝑑$ /𝑛 and 𝑠& = ∑#$%! ̅
𝑑$ − 𝑑 /(𝑛 − 1), n represents # of pairs
Two-Sided Paired t-test
Test for the mean of the differences with unknown variance:
𝐻! : 𝜇" − 𝜇# = 0 𝑜𝑟 Δ = 0
vs.
𝐻" : 𝜇" − 𝜇# ≠ 0 𝑜𝑟 Δ ≠ 0
With a pre-specified significance level α, compute the test statistic:
𝑑̅ − 0
𝑡=
𝑠$ / 𝑛
Reject 𝐻! : if 𝑡 > 𝑡%&","&(/#
Fail to reject 𝐻! : if |𝑡| ≤ 𝑡%&","&(/#
𝑡$%&,&%(/# is called the critical value and it can be found in tables or calculated using software.
Confidence Interval for Δ
A 100 1 − 𝛼 % confidence interval for the true mean difference (Δ) of two
paired samples is given by:
𝑠$ 𝑠$
𝑑̅ − 𝑡%&","&(/# ̅
≤ ∆≤ 𝑑 + 𝑡%&","&(/#
𝑛 𝑛
Where:
𝑑̅ is the point estimate of the mean difference;
*!
%
is the estimated standard error of the differences;
𝑡%&","&(/# is the percentile of the t-distribution with (n-1) degrees of freedom.
The Paired t-test: Example
Test if a new diet has an effect on lowering the Subject Before After Difference
LDL cholesterol levels. Use data from 12 randomly (X1) (X2) (After-Before)
selected subjects (time interval: 3 months).
1 201 200 -1
2 231 236 5
Assumption?
3 221 216 -5
4 260 233 -27
The observed LDL differences constitute a random
5 228 224 -4
sample from a normally distributed population of
6 237 216 -21
differences.
7 326 296 -30
8 235 195 -40
9 240 207 -33
10 267 247 -20
11 284 210 -74
12 201 209 8
Paired t-test: Example
Does new diet have an effect on lowering the LDL cholesterol levels? α=0.05
Paired t-test: Example
Construct a one-sided 95% CI for the mean LDL cholesterol levels:
Two-Sample Test for Independent Samples
• Consider that our two samples are independent (no relation/correlation
between the data points from the two groups) and that they are normally
distributed:
𝑋" ~𝑁(𝜇" , 𝜎"# ) and 𝑋# ~𝑁 𝜇# , 𝜎##
1. We can assume that the underlying variances of the two samples are equal:
𝜎"# =𝜎## =𝜎 #
2. We can assume that the underlying variances of the two samples are unequal:
𝜎"# ≠ 𝜎##
Two-Sample Test: Equal Variances
• If we know the two population variances, then:
𝜎"# 𝜎##
𝑋" − 𝑋# ~𝑁 𝜇" − 𝜇# , +
𝑛" 𝑛#
• Because 𝜎"# = 𝜎## = 𝜎 # , it follows that:
𝜎# 𝜎#
𝑋" − 𝑋# ~𝑁 𝜇" − 𝜇# , +
𝑛" 𝑛#
• Thus, for testing 𝐻! : 𝜇" = 𝜇# can use the test statistic:
+"&+#
𝑧= " "
~𝑁(0,1) under 𝐻!
, # $ -$
" #
Two-Sample Independent t-test: Equal Variances
• However, most of the times we do not know the common variance and we have
to find one estimator using the sample variances 𝑠"# and 𝑠## .
• The pooled estimate of the variance from two independent samples is given by:
# #
#
𝑛" − 1 𝑠" + (𝑛# − 1)𝑠#
𝑠 = (∗)
𝑛" + 𝑛# − 2
• Explain the denominator.
Two-Sample Independent t-test: Equal Variances
Given the hypotheses: 𝐻! : 𝜇" = 𝜇# vs 𝐻" : 𝜇" ≠ 𝜇# (two-sided)
With significance level α pre-specified, compute the test statistic:
+"&+#
𝑡= " "
, ~𝑡%"-%#&# under 𝐻!
* -
$" $#
Where 𝑠 is called the pooled sample standard deviation and is given in (*).
Reject 𝐻!: if 𝑡 > 𝑡"! #"" $%,'$(/%
Fail to reject 𝐻!: if |𝑡| ≤ 𝑡"! #"" $%,'$(/%
P-value = 2×P(𝑡"! #"" $% < 𝑡|𝐻! ), if 𝑡 < 0
= 2×P(𝑡"! #"" $% ≥ 𝑡|𝐻! ), if 𝑡 ≥ 0
Two-Sample Independent t-test: Unequal Variances
• If we know the two population variances, then:
𝜎"# 𝜎##
𝑋" − 𝑋# ~𝑁 𝜇" − 𝜇# , +
𝑛" 𝑛#
• For testing 𝐻! : 𝜇" = 𝜇# we can use the z-test statistic:
+"&+#
𝑧= ~𝑁(0,1) under 𝐻!
%# #
"-%#
$" $#
• Or (again) use 𝑠"# and 𝑠## to estimate the two population variances (if these
values are unknown).
Two-Sample Independent t-test: Unequal Variances
Given the hypotheses: 𝐻!: 𝜇' = 𝜇% vs 𝐻': 𝜇' ≠ 𝜇% (two-sided)
With significance level α pre-specified, compute the test statistic:
*! $*"
𝑡= ~𝑡+%% under 𝐻!
#" "
! # #"
$! $"
Compute the approximating degrees of freedom 𝑑, , and round down to the nearest
integer 𝑑,, :
#" # " "
!# "
$! $"
𝑑, = " " "
(∗∗)
#"
! #
$!
/("! $')# $" /("" $')
"
Two-Sample Independent t-test: Unequal Variances
Given the hypotheses: 𝐻!: 𝜇' = 𝜇% vs 𝐻': 𝜇' ≠ 𝜇% (two-sided)
With significance level α pre-specified, compute the test statistic:
*! $*"
𝑡= ~𝑡+%% under 𝐻!
#" "
! # #"
$! $"
Reject 𝐻!: if 𝑡 > 𝑡+%% ,'$(/%
Fail to reject 𝐻!: if |𝑡| ≤ 𝑡+%% ,'$(/%
Where 𝑑,, represents the nearest integer of 𝑑, (degrees of freedom) given in (∗∗).
Confidence Interval for Two Independent Samples
Two-sided 100 1 − 𝛼 % confidence interval (unknown variances):
• Two-independent samples (equal variance): 𝜇' − 𝜇% 𝜎'% = 𝜎%% = 𝜎 % :
(𝑋& − 𝑋' − 𝑡(! )(" *',&*,/' % 𝑠 % 1/𝑛& + 1/𝑛' , 𝑋' − 𝑋% + 𝑡"! #"" $%,'$(/% J 𝑠 J 1/𝑛' + 1/𝑛%)
• Two-independent samples (unequal variance): 𝜇' − 𝜇% 𝜎'% ≠ 𝜎%% :
(𝑋& − 𝑋' − 𝑡.## ,&*,/' 𝑠&' /𝑛& +𝑠'' /𝑛' , 𝑋& − 𝑋' + 𝑡.## ,&*,/' 𝑠&' /𝑛& +𝑠'' /𝑛' )
Where 𝑑 ** represents the nearest integer of 𝑑 * (degrees of freedom) given in (∗∗).
Test for Equality of Variances
Assume two independent random samples from 𝑁(𝜇', 𝜎'%) and 𝑁(𝜇%, 𝜎%%).
• Testing the equality of variances implies testing the hypotheses:
𝐻!: 𝜎'% = 𝜎%% vs 𝐻': 𝜎'% ≠ 𝜎%%
• With significance level α pre-specified, compute the test statistic:
𝑠'%
𝐹 = % ~ 𝐹"! $',"" $'
𝑠%
• The test statistic follows an F distribution with 𝑛' − 1 and 𝑛% − 1 degrees of freedom.
Test for Equality of Variances
Testing the hypotheses:
𝐻!: 𝜎'% = 𝜎%% vs 𝐻': 𝜎'% ≠ 𝜎%%
With significance level α pre-specified, compute the test statistic:
/!"
𝐹 = /" ~ 𝐹"! $',"" $' under 𝐻!
"
Reject 𝐻!: if 𝐹 > 𝐹"! $',"" $','$(/% or 𝐹 < 𝐹"! $',"" $',(/%
Fail to reject 𝐻!: if 𝐹"! $',"" $',(/% ≤ 𝐹 ≤ 𝐹"! $',"" $','$(/%
P-value = 2×P(𝐹"! $',"" $' > 𝐹|𝐻!), if 𝐹 ≥ 1
= 2×P 𝐹"! $',"" $' < 𝐹|𝐻! , if 𝐹 < 1
Two-Samples Independent t-test
Example: To assess the impact of oral contraceptive (OC) use on bone mineral density (BMD),
researchers in Canada carried out a study comparing BMD for women who had used oral contraceptives
(OC) for at least 3 months to BMD for women who had never used OC. Summary values for BMD are
given below:
n 𝑋0 s
Non OC users 10 1.08 0.16
OC users 10 1.00 0.14
a) Assuming that the BMD is normally distributed in both OC and non-OC users, is there enough
evidence to conclude that BMD levels differ between the two groups?
b) Construct a 95% CI for the population mean difference.
Two-Samples Independent t-test
Step 1: Test for equality of the variances.
Two-Samples Independent t-test
Step 2: Test for equality of the means.
Readings
• Rosner, Fundamentals of Biostatistics 8th edition:
• Chapter 8: Sections: 8.1 – 8.7