Hypothesis Testing

Often, we need to choose between competing views or hypotheses concerning a

population parameter(s). We usu identify the status quo as the null hypothesis,
H0, and the competing hypothesis, Ha, as the alternative hypothesis. Once we
have our hypotheses, we collect and analyze sample data to determine whether it
is “sufficiently” (we will be more precise on this notion soon!) consistent with the
null hypothesis. If it is, we fail to reject the null hypothesis. Otherwise, we reject
the null hypothesis. As we will see, the null hypothesis (like a criminal defendant
assumed to be innocent until proven guilty in the US) requires lots of proof before
it is rejected.
Example 1: Does a New Drug Improve Cancer Survival Rates?
Assume that the current drug used to cure pancreatic cancer results in a 10%, 5
year survival rate. Let p = fraction of pancreatic cancer patients taking a new drug
that survive 5 years. Then our hypotheses are

H0: p <=.10 Ha: p >.10.

Therefore, the new drug is assumed to not be an improvement unless we receive
strong evidence to support the view that the drug improves survival.
To determine whether we should accept or reject or accept the null hypothesis
we would take a sample of patients given the new drug and look at phat = fraction
of patients in sample that survive at least 5 years. If phat<=.10 it is clear we
should accept the null hypothesis, but what if phat = .12 or phat = .15?
Note: In this example our alternative hypothesis specifies that the population
parameter is greater than the values specified in the null hypothesis. Such an
alternative hypothesis is called an upper one-sided alternative hypothesis.

1|Page
Example 2: Is a congressional district poorer than average?
The average US family income in 2015 was $79,263. You are interested in knowing whether your
congressional district has a lower average income than the US overall. Define µ = average family
income in your congressional district. Then, our hypotheses are:

H0: µ = $79,263 or µ >= $79,263; Ha: µ < $79,263.

In this example, our alternative hypothesis specifies that the population parameter is smaller
than the values specified in the null hypothesis. Such an alternative hypothesis is called a lower
one-sided alternative hypothesis.

Our null hypothesis is that our district is not different than the rest of the US.

We would now take a simple random sample of families in our district, and calculate the sample
mean 𝑥̅ . If 𝑥̅ = $80,000, should we fail to reject the null hypothesis? Again, it may seem like we
should, but we’re dealing with sample data, which contains error. Futher, if 𝑥̅ = $75,000 or 𝑥̅ =
$72,000, should we reject the null hypothesis or fail to reject it?

Example 3: Do stock and bond annual returns have equal volatility?

Often, we want to know if it is reasonable to assume that two populations have equal variance.
When looking at annual investment returns, the standard deviation of annual percentage returns
is referred to as volatility. In this situation, our hypotheses are:

H0: Annual Variance Stock Returns = Annual Variance on Bond Returns.

Ha: Annual Variance Stock Returns ≠ Annual Variance on Bond Returns.

In this example, our alternative hypothesis does not specify a particular direction for the
deviation of variances from equality. Therefore, the alternative hypothesis is called a two-sided
alternative hypothesis.

We could now look at, say, the last 10 years of annual returns on stocks and bonds. If the sample
variance of the annual percentage returns on stocks and bonds are relatively close, we would fail
to reject H0, but if the sample variance of the annual percentage returns on stocks and bonds
differ greatly, we would reject the null hypothesis.

Should I Use a One-Tailed or Two-Tailed Test?

Some statisticians believe you should always use a two-tailed test because a priori you have no
idea of the direction in which deviations from the null hypothesis will occur. Most statisticians
feel that if a deviation from the null hypothesis in either direction is of interest, then a two-tailed
alternative hypothesis should be used, while if a deviation from the null hypothesis is of interest
2|Page
in only one direction, then a one-tailed alternative hypothesis should be used. It is much easier
to reject the null hypothesis with a one-tailed test than it is with a two-tailed test.

Types of Errors in Hypothesis Testing

To determine whether to reject or fail to reject the null hypothesis, we look at a sample statistic.
We determine a set of values for the sample statistic (called the critical region) that result in the
rejection of the null hypothesis. There are two types of errors that can be made in hypothesis
testing:

Type I Error: Reject H0 BUT H0 true. We let α = probability of making a Type I Error. α is often
called the level of significance of the test.

Type II Error: Fail to reject H0 BUT H0 not true. We define β = Probability of making a Type II
Error.

In US criminal trials, the defendant is innocent until proven guilty. In this situation if we define
H0: defendant innocent and Ha: defendant guilty, then a Type I error corresponds to convicting
an innocent defendant while a Type II error corresponds to allowing a guilty person to go free.
Since a 12-0 vote is needed for conviction, it is clear that the US judicial system considers a Type
I Error to be costlier than a Type II Error.

In a similar fashion, our approach to hypothesis testing will be to set a small probability α
(usually 0.05) of making a Type I Error, and then choose a critical region that minimizes the
probability of making a Type II Error.

Type I and Type II Error for Example 1

In Example 1 a Type 1 error results when we reject p<=.10 when in reality p<=.10.
This corresponds to the risk of concluding the new drug is an improvement when
it is not.
A Type 2 error results when we accept p<=.10 when actually, p>.10. This
corresponds to the risk of concluding the new drug is not more effective when the
drug is actually more effective than the old drug.

3|Page
Using the One Sample Z Test to test Hypotheses about Population Mean
when n>=30 or Population is Normal and σ is known
If the sample size n is >=30, then according to the Central Limit Theorem, 𝑥̅ will follow a normal
random variable, even if the population is non-normal. In this situation, we will assume that the
sample standard deviation, s, closely approximates the population standard deviation, σ. Then,
the following table summarizes the critical region for upper one-tailed, lower one-tailed, and
two-tailed hypotheses concerning µ. Because the critical regions are based on the standard
normal, the tests are referred to as the One Sample Z Tests.

Critical Regions for One Sample Z-Test

Hypotheses Critical Region
H0: µ <= µ0 𝑥̅ >= µ0 + |zα|σ/√𝑛
Ha: µ > µ0
H0: µ >= µ0 𝑥̅ <= µ0 + zασ/√𝑛
Ha: µ < µ0
H0: µ = µ0 ̅ − µ0| >= |zα/2|σ/√𝑛
|𝑥
Ha: µ ≠ µ0
In the most common case where α = 0.05, z.05 = -1.645 and z.025 = -1.96, and the previous
formulas for the critical regions become:

Hypotheses Critical Region

H0: µ <= µ0 𝑥̅ >= µ0 + 1.645σ/√𝑛
Ha: µ > µ0
H0: µ >= µ0 𝑥̅ <= µ0 - 1.645σ/√𝑛
Ha: µ < µ0
H0: µ = µ0 ̅ − µ0| >= 1.96σ/√𝑛
|𝑥
Ha: µ ≠ µ0

Z-Test Example
Passing the HISTEP test is required for graduation in the state of Fredonia. The average state
score on the test is 75. A random sample of 49 students at Carver High School have 𝑥̅ = 79 and
s = 15. For α = 0.05, would you conclude that Cooley High Students perform differently than the
typical state student?

4|Page
There is no reason to believe that Cooley High is better or worse than the state, so we will use a
two-tailed test:

H0: µ = 75; Ha: µ ≠ 75.

|79-75| >= 1.96*15/√49.
|79-75| = 4.
1.96*15/√49 = 4.2.

In other words, to be significant at α = 0.05, the difference must be greater than 4.2. It is not;
thus we fail to reject the null hypothesis, and conclude that the average Cooley High School
score does not differ from the state average.

Now, suppose we have invested resources to improve test scores at Cooley High School. Then,
we might be interested in seeing if Cooley High School students performed better than the
state. In this case, we would want to conduct an upper one-sided alternative hypothesis test:

H0: µ = 75; Ha: µ > 75.

79 >= 75 + (1.645*15)/√49 = 78.525
In other words, to be significant at α = 0.05, the result of the right part of the equation must be
less than or equal to 79. In this case, we reject the null hypothesis, and conclude that our efforts
have resulted in significant improvement.

The astute reader should realize that for a .05 level of significance our data resulted in rejection
of Ho for a one-tailed test and failure to reject the Ho for a two-tailed test. This example
illustrates that for the same level of significance can result in different outcomes based on the
type of hypothesis test being evaluated. This is why many statisticians always recommend a two-
tailed alternative, because you have made a stronger case for rejecting H0.

Probability Values (P-Values) and Hypothesis Testing

The level of significance chosen is rather arbitrary. For that reason, most statisticians use the
concept of probability values (p-values) to report the outcome of a hypothesis test. The p-value
for a hypothesis test is the smallest value of α for which the data indicates rejection of H0. Thus,
to reject the null hypothesis the p-value must be <= α. If it is > α, we fail to reject the H0.

The p-value may also be interpreted as the probability of observing a value of the test statistic at
least as extreme as the observed value of the test statistic if H0 is true. In other words, it’s the
probability of incorrectly rejecting the null hypothesis.

If we let 𝑿
̅ represent the random variable for the sample mean under H0 and x be the observed
value of 𝑥̅ , then the p-value for the one sample Z-test is computed as follows:

5|Page
P-Values for One Sample Z-Test
Hypotheses P-Value
H0: µ <= µ0 Prob(𝑿
̅ >= x)
Ha: µ > µ0
H0: µ >= µ0 Prob(𝑿
̅ <= x)
Ha: µ < µ0
H0: µ = µ0 Prob(|𝑿
̅ − µ0 | ≥ 𝑥)
Ha: µ ≠ µ0

All probabilities are computed under the assumption that H0 is true.

In our Cooley High School example, the p-value for the two-tailed test is:

2*Prob(|𝑿
̅ -75|>=4) = 2*Prob(𝑿
̅ >=79),
which can be computed as the following in Excel:
2*(1-NORM.DIST(79,75,15/√49), True) = 2*(0.030974) = 0.061948.

Because our p-value of 0.06 > our alpha value of .05, we fail to reject H0.

The p-value for a one-tailed test is:

Prob(𝑿
̅ >=79) = 0.030974.

Because our p-value of 0.03 < than our alpha value of .05, we reject H0.

One Sample Hypothesis for Mean: Small Sample, Normal Population,

Variance Unknown
Suppose we are interested in testing a hypothesis about the mean of a normal population where
the population variance is unknown, and the sample size n is <30. Then, it can be shown that
(𝑥̅ - µ)/(s/√n) follows a t-distribution with n-1 degrees of freedom. Here s = sample standard
deviation.

Like the standard normal distribution, the t-distribution has a density symmetric around 0. As
shown below, the t-distribution has fatter tails than the standard normal density, and as n
increases the t-distribution approaches the standard normal density.

6|Page
 T and Normal Densities
0.5

0.4

0.3

0.2

0.1

0
-4 -3 -2 -1 0 1 2 3 4

Normal T 5 df T 15 df T 30 df

As shown in the Hypothesis Testing.xlsx spreadsheet, T random variable worksheet shows the
percentiles of the t-distribution and how they can be computed using the T.INV function. We let
t(α,n-1) represent the α percentile of a t-distribution with n-1 degrees of freedom. Below we find,
for example, t(.025,28) = -2.04841.

D E F G H
6 2.5 %ile 28 df -2.04841 =T.INV(0.025,28)
7 97.5%ile 28 df 2.048407 =T.INV(0.975,28)
8 0.5% ile 13 df -3.01228 =T.INV(0.005,13)
9 99.5%ile 13 df 3.012276 =T.INV(0.995,13)

Basically, one sample t-tests look just like one sample z-tests with s replacing σ and the t
percentiles replacing the z percentiles.

Critical Region for One Sample t-tests

Hypotheses Critical Region
H0: µ <= µ0 𝑥̅ >= µ0 + |t(α,n-1)|s/√𝑛
Ha: µ > µ0
H0: µ >= µ0 𝑥̅ <= µ0 + t(α,n-1)s/√𝑛
Ha: µ < µ0
H0: µ = µ0 ̅ − µ0| >= | t(α/2,n-1|s/√𝑛
|𝑥
Ha: µ ≠ µ0

If we let Tn-1 stand for a t-distribution with n-1 df and t represent the observed value of (𝑥̅ -
µ)/(s/√n), then the p-values for a one sample t-test may be computed as follows:

7|Page
P-Values for One Sample t-Test
Hypotheses P-Value
H0: µ <= µ0 Prob(Tn-1 >= t)
Ha: µ > µ0
H0: µ >= µ0 Prob(Tn-1 <= t)
Ha: µ < µ0
H0: µ = µ0 2*Prob(Tn-1 >=|t|)
Ha: µ ≠ µ0

As shown below, in a fashion similar to the NORM.DIST function, the T.DIST function in Excel can
be used to compute probabilities for the t-distribution.

D E F
12 Prob T10>=2 0.036694 =1-T.DIST(2,10,TRUE)
13 Prob T10<=-2 0.036694 =T.DIST(-2,10,TRUE)

Example 5
Passing the HISTEP test is required for graduation in the state of Fredonia. The average state
score on the test is 75. A random sample of 25 students at Carver High School has 𝑥̅ = 81 and s
= 15. For α= 0.05, would you conclude that Cooley High School students perform differently
than the typical state student?

We use a two-tailed test because we have no a priori view about whether Cooley High School
students will perform better or worse than the typical state student.

H0: µ= 75, Ha: µ ≠ 75.

Using the function T.INV(0.025,24) We find t(..025,24) = -2.06.
We reject H0 if |81-75| >= 2.06*15/√25 = 6.18.
Because 6 is not >= 6.18, we fail to reject H0.

The p-value for this test is 2*Prob(T24 >= (81-75)/(15/√25))

= 2*Prob(T14 >= 2).
Prob(T14 >= 2) may be computed with the formula = 1-T.DIST(2,14,TRUE) in Excel, which returns
0.033. Therefore, the p-value for this test is 2*0.033 = 0.066. Because .066 is > 0.05, we fail to
reject H0.

8|Page
Single Sample Test about Population Proportion
Often a population proportion is unknown. The “One Sample Proportion” worksheet in
Hypothesis Testing.xlsx spreadsheet uses the BINOM.DIST.RANGE function to calculate the p-
value for both one-tailed and two-tailed alternative hypotheses.

A B C D E F G H I J K
1 Testing a Proportion Player makes 300 of 400 Free Throws?
2 Has she improved from being a 70% foul shooter?
3 trials 400 Ho: p<=p0
4 successes 300 Ha: p>p0 H0: p<=0.70 Ha: P>0.70
5 Pzero 0.7
6 Righttailedpvalue 0.0155 =BINOM.DIST.RANGE(trials,Pzero,successes,trials)
7
8
9 Ho: p>=p0
10 Ha: p<p0
11
12 Lefttailedpvalue 0.9884 =BINOM.DIST.RANGE(trials,Pzero,0,successes)
13
14 Ho: p=p0
15 Ha: p≠p0
16
17 Twotailedpvalue 0.0311 =2*MIN(Lefttailedpvalue,Righttailedpvalue)
18
19 Reject H0 if pvalue<=α

To use this worksheet, simply enter the number of trials in cell B3, the number of successes in
B4, and in cell B5 enter the value of the population proportion (p0) assumed in the null
hypothesis. Then, the p-value for a right-tailed test can be found in cell E6, the p-value for a left-
tailed test can be found in cell E12, and the p-value for a two-tailed test can be found in cell E17.

Example 6
An NBA player has made 70% of his foul shots in the past. The owner has hired a free throw
coach to improve his free throw shooting. So far this year, the player has shot 400 free throws
and made 300. At the .05 level of significance, can you conclude that the coach has succeeded in
improving the player’s free throw shooting?

Because we are only interested in improvement we use an upper one-sided test.

H0: p = 0.70; Ha: p > .0.70. Here p = the probability that the player makes a free throw.

If we enter these values into the “One Sample Proportion” worksheet, we see the p-value is
0.016, which is <= 0.05, so we reject the null hypothesis and conclude the coach has improved

9|Page
the player’s free throw shooting. In this situation, the p-value is simply the probability of a result
as extreme as making 300 of 400 free throws, or the probability that a 70% free throw shooter
will make >= 300 free throws in 400 attempts.

Testing Hypotheses about Equality of Variances

Often, we want to test if two populations have identical variances. The “T Test Equal Variance”
worksheet in the Hypothesis Testing.xlsx spreadsheet contains a template to test for α= 0.05
equality of variance between two populations. The test assumes the two populations are normal
random variables.

Simply enter the sample size and sample variance for the two populations in D3:D6. D11 and
D12 give a 95% confidence interval for the ratio of the population variances.

If the 95% confidence interval contains 1, you fail to reject the null hypothesis that the
populations have equal variances; if the 95% confidence interval does not contain 1 reject the
hypothesis of equal variances.

In this worksheet, we have the grades on the final exam of 14 statistics students who took a
hybrid statistics class (mostly online) and the final exam grades of 18 students who took the
same final exam but took the course in person with the same instructor as the hybrid group. In
D3:D6, we entered the sample size and sample variance for each population. We are 95% sure
that the ratio of the population variances is between 0.68 and 5.66. Since this interval includes 1,
we conclude the variance of the scores for the two classes are identical.

A B C D
1 TESTING IF VARIANCES OF TWO POPULATIONS ARE EQUAL
2
3 SAMPLEVAR1 33.91758242
4 SAMPLESIZE1 14
5 SAMPLEVAR2 18.01633987
6 SAMPLESIZE2 18
7 SVAR1OVERSVAR2 1.882601164
8 LOWERCI 0.358900271
9 UPPERCI 3.003895725
10
11 LOWERLIMIT 0.675666068
12 UPPERLIMIT 5.65513759
13
14 Variances Equal

10 | P a g e
Testing the Difference Between Two Population Means
There are four important hypothesis tests that can be used to evaluate the difference in two
population means. Within Excel, the “Data Analysis Add-In” makes ii simple to perform each of
these tests:

11 | P a g e
 Situation Name of Test
Large sample size (n >= 30) from each z-Test Two Sample for Means
population and samples are independent
Small sample size (n < 30) for at least one t-Test Two Sample Assuming Equal Variances
population, populations are normal, variances
unknown but equal, and the samples are
independent
Small sample size (n < 30) for at least one t-Test Two Sample Assuming Unequal
population, populations are normal, variances Variances
unknown but unequal, and the samples are
independent
The two populations are normal and the t-Test Paired Two Sample for Means
observations from the two populations can
be paired in a natural fashion

z-test Two Samples for Means

Suppose we have a large sample size of at least 30 from two populations, and the samples from
the two populations are independent (that is, the values in the sample from the first population
have no effect on the values in the sample from the second population). Let µi = unknown mean
for population i. Then, the z-Test Two Sample for Means test from the Data Analysis Add-In can
be used to test:

H0: µ1 = µ2 against a one tailed or two-tailed alternative.

For example, in the “Two Sample Z test” worksheet of the Hypothesis Testing.xlsx workbook,
we are given the starting salaries (in thousands of dollars) for 227 marketing and 211 finance
graduates of a leading MBA program. We want to conduct a two-tailed test to determine if the
average starting salaries for marketing and finance majors are equal.

12 | P a g e
 B C D E F G H I J K
1
2 variance 131.6628591 144.021
3
4 Marketing Finance
5 118 105
6 110 90
7 106 101 z-Test: Two Sample for Means
8 94 130
9 91 124 Marketing Finance
10 102 104 Mean 98.64758 109.1896
11 96 129 Known Variance 131.66 144.02
12 116 110 Observations 227 211
13 106 110 Hypothesized Mean Difference 0
14 117 126 z -9.38203
15 90 116 P(Z<=z) one-tail 0
16 113 97 z Critical one-tail 1.644854
17 112 123 P(Z<=z) two-tail 0
18 109 124 z Critical two-tail 1.959964
19 106 115
20 114 100 H0: Mean Marketing=Mean Finance
21 92 130 Ha: Mean Marketing ≠Mean Finance
22 99 93
23 105 97 P-Value =0 so reject null hypothesis
24 81 93 and conclude significant difference
25 82 110 between Average salary of marketing and finance majors

After selecting the Data Analysis Add-in from the Data ribbon, we select the “z-Test Two
Samples for Means.” We assume α = 0.05. After computing the sample variance (with the VAR
function) for each population (in cells D2 and E2) fill in the dialog box as shown below:

13 | P a g e
We find a p-value of 0, so for any alpha, we would reject the hypothesis that average salaries for
marketing and finance majors are equal. Clearly, finance majors have significantly larger salaries.

t-Test Two Sample Assuming Equal Variances

Suppose we want to compare the means of two normal populations which have unknown but
equal variances. If we take samples from the two populations that are independent and at least
one of the sample sizes is <30, then we can use the t-Test Two Sample Assuming Equal
Variances to compare the population means. First, of course, we should test the hypothesis that
the two populations have Equal Variances.

In the “Test Equal Variance” worksheet of Hypothesis Testing.xlsx workbook, we are given the
final exam grades for 14 students who took statistics in a hybrid (mostly online) format and final
exam grades of 18 students who took the course in a traditional classroom format. Does
performance of the students in the two classes differ significantly? Our hypotheses are:

H0: µHybrid = µInclass; Ha: µHybrid ≠ µInclass

Our test requires both populations be normal. A quick eyeball test for normality is to compute
the skewness and kurtosis of a sample. If both the skewness (computed with SKEW function) and
kurtosis (computed with KURT function) of a sample are between -1 and +1 the assumption of
normality is almost surely justified. From cells F2:G3 the assumption of normal populations
appears reasonable.

14 | P a g e
 A B C D E F G
1 TESTING IF VARIANCES OF TWO POPULATIONS ARE EQUAL
2 SKEW -0.50544 0.144793
3 SAMPLEVAR1 33.91758242 KURT 0.103766 -0.61404
4 SAMPLESIZE1 14 Hybrid In Person
5 SAMPLEVAR2 18.01633987 87 88
6 SAMPLESIZE2 18 94 96
7 SVAR1OVERSVAR2 1.882601164 86 84
8 LOWERCI 0.358900271 89 82
9 UPPERCI 3.003895725 74 81
10 84 85
11 LOWERLIMIT 0.675666068 85 90
12 UPPERLIMIT 5.65513759 85 90
13 92 89
14 Variances Equal 90 95
15 77 88
16 Test H0: MeanHybrid=Mean In Person 82 89
17 Test Ha: MeanHybrid≠Mean In Person 94 93
18 84 85
19 Test H0: VarianceHybrid=Variance In Person 87
20 Test Ha: VarianceHybrid≠Variance In Person 83
21 Accept H0 88
22 84
23 Now test mean difference using equal variance t test
24 t-Test: Two-Sample Assuming Equal Variances

In cells B3 and B5, we use the VAR function to determine the sample variance for each data set.
After entering the sample sizes in B4 and B6, we find from D11 and D12 that we are 95% sure
the ratio of the population variances is between .68 and 5.65. This interval includes one, so the
assumption of equal variances is justified. Then, from the Data ribbon, we choose Data Analysis
and select the “t-Test Two Sample Assuming Equal Variances” and fill in the dialog box as
follows:

15 | P a g e
The results of the hypothesis test are:

D E F G
24 t-Test: Two-Sample Assuming Equal Variances
25

26 Hybrid In Person
27 Mean 85.92857 87.61111
28 Variance 33.91758 18.01634
29 Observations 14 18
30 Pooled Variance 24.90688
31 Hypothesized Mean Difference0
32 df 30
33 t Stat -0.94609
34 P(T<=t) one-tail 0.175831
35 t Critical one-tail 1.697261
36 P(T<=t) two-tail 0.351663
37 t Critical two-tail 2.042272

The p-value for a two-tailed test is .35. Because 0.35 >= .0.05, we fail to reject the null
hypothesis that the mean performance of students is equivalent regardless of class delivery.

t-Test Two Sample Assuming Unequal Variances

Suppose we want to compare the means of two normal populations which have unknown but
unequal variances. If we take samples from the two populations that are independent and at
least one of the sample sizes is < 30, then we might need to use the t-Test Two Sample
16 | P a g e
Assuming Unequal Variances to compare the population means. First, of course, we should test
the hypothesis that the two populations have unequal variances.

In the “T Test Unequal Variance” worksheet of Hypothesis Testing.xlsx spreadsheet, we illustrate

how to use the t-Test Two Sample Assuming Equal Variances analysis in Excel.
C D E F G
1 TESTING IF VARIANCES OF TWO POPULATIONS ARE EQUAL
2 SKEW -0.5968 0.484351
3 SAMPLEVAR1 11.0289756 KURT -0.21044 -0.22528
4 SAMPLESIZE1 14 Placebo Drug
5 SAMPLEVAR2 1.119337276 2.907128 10.41907
6 SAMPLESIZE2 18 4.40015 7.99349
7 SVAR1OVERSVAR2 9.853129914 5.492934 8.61935
8 LOWERCI 0.358900271 4.55906 9.152468
9 UPPERCI 3.003895725 7.550103 8.599242
10 -2.77164 8.922706
11 LOWERLIMIT 3.536290994 -3.85383 9.382386
12 UPPERLIMIT 29.59777483 1.075804 9.259688
13 3.053792 9.446886
14 Variances Not Equal 3.256407 10.71418
15 2.996793 11.12724
16 Test H0: MeanPlacebo=Mean Drug -2.98564 9.476701
17 Test Ha: MeanPlacebo<MeanDrug 1.396865 9.680459
18 2.843022 8.138261
19 Test H0: VariancePlacebo=VarianceDrug 10.3785
20 Test Ha: VariancePlacebo≠VarianceDrug 8.954938
21 Reject H0 11.99119
22 10.30647
23 Now test mean difference using Unequal variance t test

You are conducting a study to determine if a new anti-cholesterol drug is more effective than
the placebo in reducing cholesterol. Fourteen patients were given a placebo, and eighteen
patients were given a new anti-cholesterol drug. Your data indicates the change (reduction) in
cholesterol for each patient from month 1 to month 2. We are trying to determine if the mean
change cholesterol for the patients receiving the drug is bigger than that of those receiving the
placebo.

First, we confirm the assumption of normality using SKEW and KURT functions. As you can see
by the results in F2:G3, the assumption of normal populations appears reasonable.

17 | P a g e
Next, we determine if the variances are equal. In cells D3 and D5, we compute the sample
variances with the VAR function, and in cells D4 and D6, we note the sample sizes. From cells
D11 and D12, we are 95% sure the ratio of the population variances is between 3.54 and 29.56.
This interval does not include one, so we conclude the population variances are unequal.

To test H0: µPlacebo=µDrug and Ha: µDrug>µPlacebo, we select Data Analysis from the Data ribbon, and
choose “t-Test Two Sample Assuming Unequal Variances.” Then, we fill in the dialog box as
shown below:

The results of the hypothesis test are:

D E F G
26 t-Test: Two-Sample Assuming Unequal Variances
27
28 Placebo Drug
29 Mean 2.137211 9.586846
30 Variance 11.02898 1.119337
31 Observations 14 18
32 Hypothesized Mean Difference0
33 df 15
34 t Stat -8.08041
35 P(T<=t) one-tail 3.81E-07
36 t Critical one-tail 1.75305
37 P(T<=t) two-tail 7.61E-07
38 t Critical two-tail 2.13145

18 | P a g e
The p-value for a one tailed test is 4 in 10 million, a number that is much less than 0.05, so we
reject the null hypothesis that the placebo and drug are equally effective at reducing cholesterol,
and conclude that the new drug is more effective than the placebo at reducing cholesterol.

t-Test Paired Two Sample for Means

Often observations from two populations can be paired in a meaningful way. In such situations,
the t-Test Paired Two Sample for Means (often called Matched Pairs) can be used. Both
populations need to be normal random variables. For example:

Goal Design
To test if a drug reduces cholesterol Pick ten pairs of two people who are matched
on age, weight and cholesterol. We flip a coin
to randomly choose which member of each
pair receives the drug and which receives the
placebo.
To test if a new type of insulation reduces Pick ten pairs of two houses that had the
heating bills same heating bill last winter. Flip a coin to
choose house gets the new type of insulation;
the other retains its old insulation.
To test if cross training (not just swimming) Pick 15 pairs of swimmers who had identical
improves a swimmer’s time best times in their event. Flip a coin to
determine which in the pair starts cross
training.

In each of these situations, we are blocking the effect of a variable on the response and
focusing on the difference between the treatment variable. Blocking the effect of non-
treatment variables makes it easier to isolate the effect of the treatment variable.

Blocking Variable Treatment Variable

Physical characteristics of patients Difference between drug and placebo
Size and design of home Difference between new and old insulation
Swimmer’s ability Difference between cross training and just in
water training

In the “Matched Pairs” worksheet, we use the t-Test Paired Two Sample for Means to determine
whether a new type of insulation reduces heating bills.

19 | P a g e
The homes in each row had identical winter heating bills in 2016. Before the winter of 2017, a
coin is flipped for each pair of homes. If the coin is heads the first home in the pair is given a
new type of insulation while the other home keeps its old insulation. If the coin is tails the first
home keeps its current insulation, and the second home gets the new insulation. This
randomization blocks out random differences in homes and allows us to be more confident that
any differences in heating bills that we observe are based on the type of insulation in the home
rather than something else.

The numerical data in rows 6-15 of this worksheet indicate the change in the monthly winter
heating bill during the 2017 winter. For example, the first home given new insulation saw their
heating bill increase by an average of $23 per month, and the first home that kept their current
insulation saw a $34 per month reduction in their heating bill insulation. We wish to test Ho: µNew
= µOld against Ha: µNew < µOld.

Here µNew = mean reduction in monthly winter 2017 heating bill for homes with new insulation
and µold = mean reduction in monthly winter 2017 heating bill for homes with old insulation.

As with the previous tests, the first step is to check of the skewness and kurtosis for each sample
to ensure they are consistent with normality. Based on the values in H3:I4, we can be confident
that the normality assumption is not violated.

20 | P a g e
 G H I
3 skewness 0.43421808 -0.197214711
4 kurtosis -0.057517024 -2.107287845
5 Observation Old Insulation New Insulation
6 1 -34 23
7 2 6 16
8 3 31 -28
9 4 10 29
10 5 -2 30
11 6 -12 -72
12 7 49 -46
13 8 -15 -55
14 9 -45 21
15 10 -17 -61
16
17
18 t-Test: Paired Two Sample for Means
19
20 Old Insulation New Insulation
21 Mean -2.9 -14.3
22 Variance 806.3222222 1750.233333
23 Observations 10 10
24 Pearson Correlation -0.206862333
25 Hypothesized Mean Difference 0
26 df 9
27 t Stat 0.652971466
28 P(T<=t) one-tail 0.265049676
29 t Critical one-tail 1.833112933
30 P(T<=t) two-tail 0.530099352
31 t Critical two-tail 2.262157163

To do this analysis, from the Data ribbon, we select Data Analysis, choose “t-Test Paired Two
Sample for Means” and fill in the dialog box as shown below:

21 | P a g e
We obtain a p-value for a one-tailed test of 0.265, so for α= 0.05, we would fail to reject H0, and
conclude the new insulation is ineffective.

Chi Square Test for Independence

The Chi-Square test for Independence is used to determine if there is a significant relationship
between two categorical variables. For example, the distribution of eye color by gender in an
Iowa State statistics class is summarized in the following contingency table:

Eye Color

Gender Blue Brown Green Hazel

Female 370 352 198 187

Male 359 290 110 169
Total 729 642 308 356
The question is whether eye color is independent of or dependent on gender. As shown in the
“Chi-Square” worksheet, we find the percentage of eye color by gender is:

22 | P a g e
 K L M N O P
11 Eye Color
Gender Blue Brown Green Hazel Total
12
13 Female 33.42% 31.80% 17.89% 16.89% 100.00%
14 Male 38.69% 31.25% 11.85% 18.21% 100.00%
15 Total 35.82% 31.55% 15.14% 17.49% 100.00%

If eye color was completely independent of gender, then eye color would not depend on
gender, and we would find that:
• 35.82% of each gender with blue eyes.
• 31.55% of each gender with brown eyes
• 15.14% of each gender with green eyes.
• 17.49% of each gender with hazel eyes.
(These percentages are the marginal probabilities for each eye color in the table.)

Are the observed discrepancies from the expected percentages simply occurring by chance? If
so, gender and eye color are independent. In this situation, our hypotheses are:

H0: Gender and Eye Color are Independent

Ha: Eye Color depends on Gender

If a contingency table has R rows and C columns, then the relevant test statistic follows a χ2
random variable with (R-1) * (C-1) degrees of freedom. In this example, R = 2 and C = 4, giving
us 3 degrees of freedom.

To compute the relevant test statistic, r, we define:

• N = Total number of observations
• Oij = Observed number of observations in row i and column j of the contingency table.
• Eij =N * (proportion of observations in row i)*(proportion of observations in column j); or
the row marginal probability multiplied by the column marginal probability multiplied by
the total number of observations. Eij is simply the expected number of observations in
the row i column j cell if the row and column categories are independent.

Then, for each cell, compute (Oij – Eij)2/Eij. In other words, observed value – expected value
squared divided by the expected value.

Summing up this quantity for each cell yields the χ2 statistic. If there is perfect independence
between the row and column categories, then for each cell Eij =Oij, (the number of expected

23 | P a g e
observations equals the number observed), and the observed χ2 statistic would equal 0.
Therefore, a large χ2 statistic would result in rejection of the null hypothesis.

If the level of significance is α, then we reject the null hypothesis when the test statistic is >= the
(1-α) %ile of the Chi Square random variable with the appropriate degrees of freedom. This
value for the Chi-Square test can be found using the CHISQ.INV(1-α,degrees of freedom)
function in Excell. The cutoffs for 1,2,3, and 4 degrees of freedom for α=0.05 are as follows:

D E F
14
15 df Cutoff
16 1 3.841459 =CHISQ.INV(0.95,D16)
17 2 5.991465 =CHISQ.INV(0.95,D17)
18 3 7.814728 =CHISQ.INV(0.95,D18)
19 4 9.487729 =CHISQ.INV(0.95,D19)

In our example, we have R = 2 and C = 4, so we have (2-1)*(4-1) = 3 degrees of freedom and the
cutoff point for rejecting the null hypothesis of independence for α= 0.05 is 7.81.

The total number of observations (2035) is computed in cell I17.

To calculate the Chi Square, you must first calculate the expected values. Obtain these values, Eij,
by copying the formula from L19 to L19:O22. The formula in L19 = ($P7/Total)*(L$9/Total)*Total,
which is the observed marginal value for the associated row divided by the total multiplied by
the marginal value for the associated column divided by the total; this product is then multiplied
by the total. For example, the expected value for females with blue eyes is calculated like this:

Expected Number of Greened Eyed Females = (1107/2035)*(729*2035)*2035 = 396.56.

Note the row and column totals (L21:O21 and P19:P20) of each Eij match the observed data.

K L M N O P
17 Eye Color
Gender Blue Brown Green Hazel Total
18
19 Female 396.56 349.24 167.55 193.66 1107.00
20 Male 332.44 292.76 140.45 162.34 928.00
21 Total 729.00 642.00 308.00 356.00

24 | P a g e
Copying the formula ((L7-L19)^2/L19) from L25 to L25:O26 computes (Oij – Eij)2/Eij for each cell.

H I J K L M N O
22 Test Statistic
23 Eye Color
Gender Blue Brown Green Hazel
24
25 Female 1.78 0.02 5.54 0.23
26 Male 2.12 0.03 6.60 0.27
27 Chi Square Total
28 16.59

Summing those values, the Chi Square statistic is 16.59 which is >= cutoff of 7.81, so we reject
the hypothesis that eye color and gender are independent. We can find the p-value for the test
using the CHISQ.DIST.RT(test statistic value, degrees of freedom) in cell H29. RT in this formula
stands for “right tailed probability” which is the probability that we evaluate in Chi Square
analysis.

H I J K
29
30 0.000858 =CHISQ.DIST.RT(H28,3)

In this case, the p-value of 0.000858 is <= 0.05, which is consistent with our rejection of H0.

25 | P a g e