Inference for two Population Means

Bret Hanlon and Bret Larget

Department of Statistics
University of WisconsinMadison
October 27November 1, 2011
Two Population Means 1 / 65
Case Study
Case Study
Example 19.2 on page 545 of the text describes an experiment with pseu-
doscorpions of the species Cordylochernes scorpiodes which live in the trop-
ics. The females typically mate multiple times, even though mating once
provides sucient sperm for fertilization. Researchers conducted an exper-
iment to examine a possible evolutionary explanation for this behavior. If
there is genetic incompatibility between some pairs, a female may be more
fertile by having multiple partners. In the experiment, females were ran-
domly assigned to one of two treatment groups. In one group, a female was
mated with the same male twice; in the other group, two dierent males
were mated to the same female. The response variable is the number of
successful broods. This variable is a small integer, ranging from 0 in some
cases to a maximum of observed value of 7.
Two Population Means Case Study 2 / 65
Data
The data set is small enough that we can display it.
The mean of the 20 values in the Same treatment group is 2.2.
The mean of the 16 values in the Dierent treatment group is 3.6.
Group Number of Successful Broods
Same 0 0 0 0 1 1 2 2 2 2 2 2 3 3 3 3 4 4 4 6
Dierent 0 1 2 2 2 3 3 4 4 4 4 4 6 6 6 7
Two Population Means Case Study 3 / 65
Questions
Do female pseudoscorpions have more successful broods on average,
when they have multiple partners?
The real biological question of interest involves the pseudoscorpions in
their natural environment.
The experimental setting seeks to explore the question in a controlled
setting.
In the experiment, the sample means are 2.2 and 3.6; what does this
imply about the populations?
Here, there is a single biological population of pseudoscorpions that
can be thought of as two statistical populations on the basis of
assignment to experimental treatment groups.
Two Population Means Case Study 4 / 65
A statistical model
A statistical model for the experiment is that there are two probability
distributions for the number of successful broods, one for each
treatment group.
The specic distributions are not specied, but each is summarized
my a mean or expected value, say
1
for the Same treatment group
and
2
for the Dierent treatment group.
The null hypothesis is
1
=
2
; the biologically interesting alternative
here is
1
<
2
.
How can we test this hypothesis?
For the sample estimates, 2.2 < 3.6, but what can we infer about the
populations?
Two Population Means Case Study 5 / 65
The Big Picture
We have two populations with means
1
and
2
.
We have independent samples from each sample means x
1
and x
2
of
sizes n
1
and n
2
respectively.
We want to test H
0
:
1
=
2
versus the alternative H
A
:
1
<
2
with
few assumptions about the distributions in the populations.
The key idea of a randomization test is to consider the null
distribution of the dierence in sample means for all possible random
samples assuming that the randomization is independent of the
observed data.
Two Population Means The Big Picture 6 / 65
Example
In the example, there are 20 females in the Same treatment group
and 16 in the Dierent treatment group.
The observed dierence in sample means is 2.2 3.625 = 1.425
(without roundo).
What if the number of successful broods each female pseudoscorpion
had was independent of the assignment to treatment group?
If this were the case, we could compare the observed dierence in
sample means to the null distribution of dierences in sample means
for all other possible results of the randomization.
There are
_
36
20
_
= 7307872110 possible ways to randomly separate 36
individuals into groups of size 20 and 16.
Instead of nding exactly how many of these 7 billion+ possible
randomization results have dierences in sample means at least as
extreme as the observed 1.425, we can use the computer to
simulate the randomization process and estimate the p-value.
Before beginning, we will examine methods to graph the data.
Two Population Means The Big Picture 7 / 65
Types of Graphs
When comparing two independent samples, we can use extensions of
the types of graphs for single samples:

density plots;

histograms;

box-and-whisker plots;

dot plots.
As this example data set is small, a dot plot is best because there is
no compelling reason to summarize the data.
Points should be jittered so equal values are not directly on top of one
another.
Two Population Means Graphics 8 / 65
Dot plot of the data
b
r
o
o
d
s
0
2
4
6
Different Same
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
GG
G
G
G
G
G
G
G
G
Two Population Means Graphics 9 / 65
Comments on the Graphics
There is a lot of overlap between the samples.
It would be dicult to place an individual in one group or the other
on the basis of the number of successful broods.
But the centers of the distributions appear to be a bit dierent with
generally larger values for the Dierent group on average.
Two Population Means Graphics 10 / 65
Components of a Randomization Test
Randomization Tests
1
State hypotheses;
2
Select and calculate a test statistic;
3
Use simulation to nd the null distribution of the test statistic;
4
Compare the value of the actual test statistic to its null distribution
to compute a p-value;
5
Summarize the results in the context of the problem.
Two Population Means Randomization Tests 11 / 65
State Hypotheses
Hypotheses are statements about populations;
Here we are assuming that the pseudoscorpions in the sample may be
treated as if they were randomly sampled from the population of
these pseudoscorpions in the wild;
In words, the hypotheses are:
H
0
: There would be no dierence in the mean number of
successful broods for each experimental condition
among all female pseudoscorpions in the population.
H
A
: The experimental condition with dierent partners
produces a larger mean number of successful broods
than the experimental condition with the same partner
mating twice.
Two Population Means Randomization Tests 12 / 65
State Hypotheses (cont.)
In symbols, letting
1
and
2
represent the mean number of
successful broods in the population for the Same and Dierent
groups, respectively, the hypotheses are:
H
0
:
1
=
2
H
A
:
1
<
2
One could also test the alternative hypotheses H
A
:
1
=
2
or
H
A
:
1
>
2
if appropriate for the setting.
Two Population Means Randomization Tests 13 / 65
Select a Test Statistic
The dierence in sample means is the natural test statistic for a
hypothesis that compares population means.
As we are determining the null distribution by simulation, there is no
need to standardize the test statistic so it can be compared to some
well-known benchmark distribution (such as standard normal,
chi-square, or t).
For the observed data, x
1
= 44/20 = 2.2 and x
2
= 58/16 = 3.625
and the dierence is x
1
x
2
= 1.425.
Two Population Means Randomization Tests 14 / 65
Compute the Null Distribution
Conceptually, we take a random sample of 20 without replacement
from the 36, compute its mean and the mean of the 16 remaining
values, and take the dierence.
Repeat this process very many times and see how may dierences are
1.425 or smaller.
The proportion of such values is the p-value.
Two Population Means Randomization Tests 15 / 65
Graph of Null Distribution
Difference in Sample Means
D
e
n
s
i
t
y
0.0
0.2
0.4
0.6
2 1 0 1 2 3
Two Population Means Randomization Tests 16 / 65
P-value
It is evident from the graph that the observed dierence is fairly
unusual relative to the sampling distribution.
The p-value is the actual proportion of sampled randomizations with
a dierence at least as extreme as that observed.
For the 100,000 randomly selected randomizations, the p-value is
estmated to be 0.0093.
A dierent simulation would estimate this dierently, but not by too
much.
The jaggedness of the preceding graph is caused by lots of ties in the
randomization distribution of the dierence between sample means.
Two Population Means Randomization Tests 17 / 65
Conclusions
The p-value is fairly small.
In the context of the problem, we can say this.
There is strong evidence that female pseudoscorpions have
fewer successful broods when they mate with only one
partner than when they mate with two partners under the
given experimental conditions (two independent sample
randomization test, p = 0.009, n
1
= 20, n
2
= 16). This
result is consistent with the evolutionary explanation that
the behavior of having multiple partners as seen in nature
may overcome the possibility of genetic incompatibility
among some partners.
Two Population Means Randomization Tests 18 / 65
Summary
Randomization tests are useful for comparing population means (or
other population characteristics).
The method simply considers the test statistic under the null
hypothesis of independence between the randomization and the
response variable of interest.
As simulation determines the null distribution, there is no need to
scale the test statistic so that it is comparable to a standard
benchmark null distribution.
Randomization tests are only practical with a computer.
We see that the shape of the null distribution is symmetric and
bell-shaped, which suggests that an approximation with a standard
distribution may be accurate.
Later in these notes we will reexamine this data using a t-test.
Two Population Means Randomization Tests 19 / 65
Two Dierent Designs
There are two standard designs to compare two treatment groups;
1
In a paired design, there is a single sample of pairs of observations and
each treatment is applied to each sampled unit.
Example
In a sample of people, scores from vision tests are recorded separately for each
eye. There is interest in comparing scores between right and left eyes.
2
In a two-independent-sample design, there are two separate samples,
and all elements of one sample get one treatment, all elements of the
other sample get the other treatment.
Example
In a dairy cattle study with two dierent feed supplements, cattle are randomly
separated into two groups and each group is given one of two dietary
supplements. Average daily milk yield is compared between the groups.
Two Population Means Comparison of Means 20 / 65
Butterfat Study
This study and the accompaning data is from a former Statistics 571
student.
Case Study
The butterfat content in milk is an important factor in determining its eco-
nomic value and in how it is processed to form dairy products such as cheese,
ice cream, and butter. In an experiment, a company is interested in com-
paring the performances of two dierent labs which measure the butterfat
content of milk. Two separate samples were collected from 107 loads of
milk, and one sample from each load was sent to one of two labs. Butterfat
content changes based on the identity of the cows, the time of milking, the
time since the last milking, and other factors, so the percentage butterfat
can be expected to vary from load to load, but should be consistent for
samples taken from the same load as each load is properly agitated before
sampling to promote mixing throughout the load.
How should this data be examined to compare the performances of the labs?
Two Population Means Case Studies 21 / 65
Horned Lizards
Case Study
The horned lizard Phrynosoma mcalli has horns it uses for protection. Re-
searchers tested a hypothesis that longer horns are more protective then
shorter horns. A predator of these lizards is the loggerhead shrike, a bird
that impales the lizards on thorns or barbed wire. Researchers compared the
horn lengths of 30 skewered lizards with 154 horned lizards that were living.
The average length of the skewered lizards was 21.99 mm and the average
length of the living lizards was 24.28 mm. Is this evidence that longer horns
are more protective?
Two Population Means Case Studies 22 / 65
Graphs
Density plots, histograms, dot plots, and box-and-whisker plots are all
useful for graphical comparisons between samples.
For paired samples, graphs of dierences are also useful.
It is very important to graph data and look for patterns or outliers
before carrying out statistical inferences.
Two Population Means Graphics 23 / 65
Butterfat Study
The butterfat percentage data is from a paired data design.
There is a single sample of size n = 107.
Each sample unit (a load) is measured twice (the two butterfat
measurements from the two labs).
Paired data design data is analyzed by:

taking dierences within each pair (same order);

analyzing the single sample of dierences using one-sample methods.

Two Population Means Application 24 / 65
Scatter plot
For paired data, a scatter plot of one measurement versus the other
helps show the relationship and identify potential outliers.
Denition
An outlier is a single observation that sticks out in a plot as unusual and
not part of the general trend.
There are multiple ways to deal with outliers in an analysis.
Two Population Means Application Graphics 25 / 65
Scatter plot
Percentage butterfat from Lab 1
P
e
r
c
e
n
t
a
g
e

b
u
t
t
e
r
f
a
t

f
r
o
m

L
a
b

2
3.0
3.2
3.4
3.6
3.0 3.5 4.0
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
Two Population Means Application Graphics 26 / 65
Dot plot
Difference in Percentage butterfat (Lab 1Lab2)
0.2 0.0 0.2 0.4 0.6 0.8
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
Two Population Means Application Graphics 27 / 65
Box-and-Whisker Plot
Difference in Percentage butterfat (Lab 1Lab2)
0.2 0.0 0.2 0.4 0.6 0.8
G
Two Population Means Application Graphics 28 / 65
Density Plot
Difference in Percentage butterfat (Lab 1Lab2)
D
e
n
s
i
t
y
0
5
10
15
0.0 0.5
Two Population Means Application Graphics 29 / 65
Histogram
Difference in Percentage butterfat (Lab 1Lab2)
P
e
r
c
e
n
t

o
f

T
o
t
a
l
0
10
20
30
40
0.0 0.5
Two Population Means Application Graphics 30 / 65
Observations
There are four individual measurements that are outliers.
There are several possible explanations:
1
data may be recorded incorrectly;
2
the load may not have been agitated properly before sampling;
3
one or both labs occasionally makes a poor measurement;
With additional background information, we understand that the
second explanation is most plausible, as some of the individuals that
do the work of taking samples fail to properly agitate the milk, and as
cream rises to the top, there is the possibility that two separate
samples from the same load might dier considerably in percentage of
butterfat.
As these observations are likely not telling us about the performance
of the laboratories, but about how a small part of the data is
collected, and our desired inference is about the laboratories, it is
reasonable to discard the outliers.
Two Population Means Application Graphics 31 / 65
What to do with outliers
For the purposes of this example in lecture, we will discard the four
outliers from further analysis, as explained on the previous slide.
For 103 of the 107 data points, dierences between measurements are
no more than 0.07 percent.
For the other loads, the dierences are at least twice as large, ranging
from 0.15 to 0.87.
If this was my data, I would seek explanations in each case for the
discrepancy: (was the data recorded improperly, is there a note to
cast suspicion on a particular measurement)
Generally, one should not discard data unless there is a good reason.
If I had more information that indicated that the observed
measurements were genuinely accurate and just unusual, I would not
discard the data.
This example highlights the importance of graphing data before
conducting an analysis for inference.
Two Population Means Application Graphics 32 / 65
Dot plot (after outliers removed)
After Removal of Outliers
Difference in Percentage butterfat (Lab 1Lab2)
0.06 0.04 0.02 0.00 0.02 0.04
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
Two Population Means Application Graphics 33 / 65
Histogram (after outliers removed)
After Removal of Outliers
Difference in Percentage butterfat (Lab 1Lab2)
P
e
r
c
e
n
t

o
f

T
o
t
a
l
0
5
10
15
20
0.05 0.00 0.05
Two Population Means Application Graphics 34 / 65
Estimation for Paired Designs
For paired designs, inference is on the dierences in measurements.
The data is considered to be a single sample of dierences.
Condence intervals use the single sample methods we have seen
before;
1
a t distribution method; or
2
the bootstrap
Here we will use the t distribution as the sample size is large enough
to overcome even extreme nonnormality and the (remaining) data has
a fairly symmetric distribution.
Two Population Means Application Estimation 35 / 65
Condence Intervals from Paired Designs
Condence Interval for
D
=
1

2
A P% condence interval for
D
=
1

2
has the form

D t

n
< <

D + t

n
where t

is the critical value such that the area between t

and t

under
a t-density with n 1 degrees of freedom is P/100, where n is the sample
size.
Two Population Means Application Estimation 36 / 65
Chalkboard example
Example
Here are summary statistics:

The sample size is 103.

The mean dierence (Lab 1 - Lab 2) is -0.0037.

The standard deviation of the dierences (Lab 1 - Lab 2) is 0.0223.

Find a 95% condence interval using the formula on the board.

0.008 <
D
< 7e 04
Two Population Means Application Estimation 37 / 65
Compare to R Solution
> t.test(lab1, lab2, paired = T)
Paired t-test
data: lab1 and lab2
t = -1.6811, df = 102, p-value = 0.0958
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
-0.0080422274 0.0006635867
sample estimates:
mean of the differences
-0.003689320
Two Population Means Application Estimation 38 / 65
Interpretation
After excluding observations with unexplained unusually large
dierences, we are 95% condent that the mean dierence (Lab1
- Lab 2) in butterfat determination is between 0.008 and 0.001
in the common situation where there is an absence of large
discrepancies in the measurements.
Two Population Means Application Estimation 39 / 65
Comparison with outliers included
An analysis with the outliers included would be appropriate if it were
determined that the outliers were, in fact, genuinely correct.
Notice that the interval still includes 0 (consistent with no dierences
in the labs), but that the condence interval is much wider.
Paired t-test
data: bfat[lab == "Lab1"] and bfat[lab == "Lab2"]
t = 0.1522, df = 106, p-value = 0.8793
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
-0.01686179 0.01966553
sample estimates:
mean of the differences
0.001401869
Two Population Means Application Estimation 40 / 65
Hypothesis Tests
We can also formally test the null hypothesis that the mean dierence
in measurements in the two labs is zero.
H
0
:
1

2
= 0
H
A
:
1

2
= 0
Paired t test
If dierences D
1
, D
2
, . . . , D
n
are normally distributed, then
T =

D d
0
s/

n
t(n 1)
where d
0
(usually 0) is the mean dierence in the null hypothesis, s is the
sample standard deviation of dierences, and n is the sample size.
Two Population Means Application Hypothesis Tests 41 / 65
Chalkboard example
Example
Here are summary statistics:

The sample size is 103.

The mean dierence (Lab 1 - Lab 2) is -0.0037.

The standard deviation of the dierences (Lab 1 - Lab 2) is 0.0223.

Test the hypothesis of no dierence between the labs.

T = 1.68, p = 0.0958, df = 102
Two Population Means Application Hypothesis Tests 42 / 65
Compare to R Solution
> t.test(lab1, lab2, paired = T)
Paired t-test
data: lab1 and lab2
t = -1.6811, df = 102, p-value = 0.0958
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
-0.0080422274 0.0006635867
sample estimates:
mean of the differences
-0.003689320
Two Population Means Application Hypothesis Tests 43 / 65
Two Independent Samples
The lizard example is not paired.
There is no reason to match a specic individual in one sample with
an individual in the other.
In fact, the two sample sizes are not even equal.
Inference methods and graphs are dierent for independent samples.
Two Population Means Two Independent Samples 44 / 65
Graphics
Separate graphics are produced for each separate sample.
The graphs use a layout to make comparisons between them easy.
Often, such graphs are side by side or on top of one another.
It is best to use common scales on axes.
Two Population Means Two Independent Samples Graphics 45 / 65
Histograms
Horn Length (mm)
P
e
r
c
e
n
t

o
f

T
o
t
a
l
0
10
20
30
15 20 25 30
Killed
0
10
20
30
Living
Two Population Means Two Independent Samples Graphics 46 / 65
Dot Plots
Horn Length (mm)
15 20 25 30
G
G G G
G
G
G
G G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
G G
G
G
G G
Killed
G
G G G
G
G G G
G
G
G
G
G
G
G G
G G
G G
G
G G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G G
G
G
G
G G
G
G
G
G G
G
G
G
G
G
G
G G
G
G G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
Living
Two Population Means Two Independent Samples Graphics 47 / 65
Density Plots
Horn Length (mm)
D
e
n
s
i
t
y
0.00
0.05
0.10
0.15
10 15 20 25 30
Killed Living
Two Population Means Two Independent Samples Graphics 48 / 65
Box-and-Whisker Plots
Horn Length (mm)
15 20 25 30
G
Killed
G
Living
Two Population Means Two Independent Samples Graphics 49 / 65
Violin Plots
H
o
r
n

L
e
n
g
t
h

(
m
m
)
15
20
25
30
Killed Living
Two Population Means Two Independent Samples Graphics 50 / 65
Remarks
While there are some extreme values, they do not stick out far from
the overall pattern of the data.
Both distributions look to be not strongly skewed.
Inferences based on t distributions will be accurate, even with
moderate nonnormality in the underlying populations.
Two Population Means Two Independent Samples Graphics 51 / 65
t Distribution for Independent Samples
t Distribution for Independent Samples
If two independent samples X
1
, . . . , X
n
1
and Y
1
, . . . , Y
n
2
are each normally
distributed with respective means
1
and
2
and if they share a common
variance
2
, then the test statistic
T =
(

X

Y) (
1

2
)
s
p
_
1
n
1
+
1
n
2
t(n
1
+ n
2
2)
where
s
p
=

(n
1
1)s
2
1
+ (n
2
1)s
2
2
(n
1
1) + (n
2
1)
is the pooled sample standard deviation and s
1
and s
2
are the respective
single sample standard deviations.
Note that s
2
p
, the pooled variance, is a weighted average of the
sample variances, weighted by the degrees of freedom.
Two Population Means Two Independent Samples t Distribution 52 / 65
Derivation
Using linearity properties of expectation and variances of independent
random variables, it is straightforward to show that if X
1
, . . . , X
n
1
and
Y
1
, . . . , Y
n
2
are independent samples with E(X
i
) =
1
, Var(X
i
) =
2
1
,
E(Y
i
) =
2
, and Var(Y
i
) =
2
2
, then
E(

X

Y) =
1

2
Var(

X

Y) =

2
1
n
1
+

2
2
n
2
Furthermore, if both distributions are normal, the distribution of the
dierence in sample means is also normal.
Under the additional assumption that
1
=
2
,
Var(

X

Y) =
2
_
1
n
1
+
1
n
2
_
Two Population Means Two Independent Samples t Distribution 53 / 65
Standard Error
The standard error for the dierence in sample means is
SE(

X

Y) =

2
1
n
1
+

2
2
n
2
Under the assumption that
1
=
2
, this is estimated as

SE(

X

Y) = s
p
_
1
n
1
+
1
n
2
Without this assumption, the estimate is

SE(

X

Y) =

s
2
1
n
1
+
s
2
2
n
2
Welchs t-test uses

SE for the T statistic; in this case the distribution
is only approximate and the degrees of freedom is approximated with
a messier formula (see page 304 in the textbook for details).
Welchs approach is the default in the R function t.test().
Two Population Means Two Independent Samples t Distribution 54 / 65
Chalkboard example
Example
Here are summary statistics for the lizard data:

The sample sizes are 154 for the living lizards and 30 for the
unfortunate skewered ones.

The sample means are 24.28 for the living lizards and 21.99 for the
killed ones.

The standard deviations are respectively 2.63 and 2.71.

Find a 95% condence interval for the dierence in mean horn length
for the two groups.

Test the hypothesis test of no dierence in mean horn size.

Two Population Means Two Independent Samples Estimation and Testing 55 / 65
Comparison with R
Use t.test() with argument var.equal=T for the conventional
two-sample test.
> t.test(hornLiving, hornKilled, var.equal = T)
Two Sample t-test
data: hornLiving and hornKilled
t = 4.3494, df = 182, p-value = 2.27e-05
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
1.253602 3.335402
sample estimates:
mean of x mean of y
24.28117 21.98667
Two Population Means Two Independent Samples Estimation and Testing 56 / 65
Comparison with R (Welchs Test)
Use t.test() with default arguments for Welchs two-sample test
which does not assume equal variances.
Note that the approximate degrees of freedom is always at least as big
as the smaller of n
1
1 and n
2
1 and no larger than n
1
+ n
2
2.
> t.test(hornLiving, hornKilled)
Welch Two Sample t-test
data: hornLiving and hornKilled
t = 4.2634, df = 40.372, p-value = 0.0001178
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
1.207092 3.381912
sample estimates:
mean of x mean of y
24.28117 21.98667
Two Population Means Two Independent Samples Estimation and Testing 57 / 65
Interpretation of Condence Interval
We are 95% condent that the mean length of the horns is
between 1.21 and 3.38 mm longer in surviving lizards than in
lizards killed by the loggerhead shrike in the study population
where the data was collected. This is consistent with the
biological hypothesis that larger horns oer more protection.
There are other possible interpretations, as other variables may
be confounded with horn length. Perhaps lizards with longer
horns are also older, heavier, faster, or something else that is the
real cause of the greater protection.
Two Population Means Two Independent Samples Estimation and Testing 58 / 65
A one-sided t-test
> t.test(hornLiving, hornKilled, alternative = "greater")
Welch Two Sample t-test
data: hornLiving and hornKilled
t = 4.2634, df = 40.372, p-value = 5.889e-05
alternative hypothesis: true difference in means is greater than 0
95 percent confidence interval:
1.388469 Inf
sample estimates:
mean of x mean of y
24.28117 21.98667
Two Population Means Two Independent Samples Estimation and Testing 59 / 65
Interpretation of Hypothesis Test
There is strong evidence that the mean horn length in living
lizards is greater than that in lizards killed by the loggerhead
shrike (p < 0.0001, t = 4.26, Welchs two-sample independent
t-test, n
1
= 154, n
2
= 30). This is consistent with the biological
hypothesis that longer horns oer greater protection, but is
likewise consistent with explanations of protection oered by
other characteristics associated with lizards with longer horns.
Two Population Means Two Independent Samples Estimation and Testing 60 / 65
Case Study
Example
Recall the pseudoscorpions example we used for the permutation test.
In one group, a female was mated with the same male twice; in the other
group, two dierent males were mated to the same female. The response
variable is the number of successful broods. This variable is a small integer,
ranging from 0 in some cases to a maximum of observed value of 7.
For the 100,000 randomly selected randomizations, the p-value is
estmated to be 0.0093.
Two Population Means Application 61 / 65
Comparison to Welchs t test
Compare the corresponding p-values.
There is also evidence that the mean number of successful broods is
smaller in the Same treatment group.
The p-values dier.
> t.test(same, different, alternative = "less")
Welch Two Sample t-test
data: same and different
t = -2.3424, df = 28.883, p-value = 0.01313
alternative hypothesis: true difference in means is less than 0
95 percent confidence interval:
-Inf -0.3911856
sample estimates:
mean of x mean of y
2.200 3.625
Two Population Means Application 62 / 65
Cautions and Concerns
Always plot data before doing inferences.
Handle potential outliers with care.
Use the paired or two-sample t methods when appropriate; paired
when the design is a single sample with each unit measured twice,
independent when the samples are independent of one another.
t methods are robust to nonnormality when sample sizes are large
enough.
Inferences to populations depends on random sampling; be prepared
to argue when a sample can be thought of as representative despite
nonrandom sampling.
The two-sample independent t methods assume equal variances in the
samples, but are robust to minor dierences.
Randomization or permutation methods are alternatives to t tests.
Two Population Means Cautions and Concerns 63 / 65
What You Should Know
You should know:
how to use randomization/permutation methods or t methods for
paired and independent sample problems;
how to determine of paired or independent sample methods are
appropriate;
how to graph data in various ways to display dierences in
distributions;
how to examine graphs of data to informally assess method
assumptions;
how to interpret inferences in context.
Two Population Means What you should know 64 / 65
Extensions
We have moved from one to two populations: what about three or
more?
ANOVA (Analysis of Variance) is on the way.
Two Population Means Extensions 65 / 65