Basics of Hypothesis Testing

1


Hypothesis testing is a handy way to carry out some statistical inference about the
population based on sample information. First, we need a sample . . . we generally
assume that sample has a normal distribution, with common mean and variance, in order
to do the majority of our testing.
2
Put mathematically, were hoping that X ~N[,
2
].

We then use our sample mean to make statements about the population mean .
Depending on what type of statement were making, we use either a one-tailed or a two-
tailed test. Both tests require the calculation of a t statistic.

The T Statistic

The first piece of information we need is a T statistic for our data, which is given by

T =
x
s
n


where x is our sample mean, is the hypothesized population mean, s is our estimated
standard deviation, and n is our sample size. If you were looking at a table of summary
statistics, the denominator could be derived in two ways; either divide the estimated
standard deviation by the root of the sample size, or just take use standard error statistic.
3


We also need a critical t value, which is calculated independent of our data (minus the
fact that degrees of freedom are related to n). It is determined by the distribution of a t
statistic, and will be based off only our significance level and our degrees of freedom.
4

Back in the day, the critical t value used to be found using tables in the back of stat
books. For this class, we often use Excel to get the critical t. To do so, the formula is:

=TINV(, df) or =TINV((2*, df))

where is the significance level were going to use for our test. The formula on the left
applies for 2-tailed tests, while the formula on the right applies for 1-tailed tests.

Aside: Midterm


1
This handout is a brief review of material covered in lecture and the course notes. It is my production, and
not that of Dr. Cameron. It should not be considered as a substitute for or modification of any course
material. In any case of discrepancy, Dr. Camerons material supersedes any notes made here.
2
This assumption makes the t test the most powerful test on the population mean.
3
The standard error measures the variability in the mean of our sample. Since the mean is an average, it is
always going to be less variable than any one particular random draw. The larger the number of draws
taken from a normal distribution, the less variable the mean is going to be, as data will naturally begin to
cluster around a particular value. Thats why we use the root of n to scale down our error.
4
In this case, our degrees of freedom is given by (n-1), because we have lost one degree of freedom when
we estimated the variable s.
Make sure youre familiar with this formulation. Consider this information provided on
one of Dr. Camerons past midterms:

TINV(0.01,50) 2.67779
TINV(0.10,50) 1.64787

You will be asked to do hypothesis testing, and youll need to know which applicable
critical t to use. For example, if he asked you to do a two-tailed test with a significance
level of 0.01 and 50 degrees of freedom, the first one (2.678) would be appropriate. If he
asked you to do a one-tailed test with a significance level of 0.05, youd want the second
value (1.648).

After we get our critical t value, we compare it to the T calculated using our data. If were
doing a two-tailed test, we reject the null whenever the absolute value of our T is
greater than the critical t. If were doing a one-tailed test, our rejection region depends
on which test were using. Well go into more detail on that in a second.

Significance Levels

Our is chosen by deciding how certain we want to be in our testing. We could be 100
percent certain about our inference if we were general enough, but then the inference
would be pretty useless. Instead, we allow ourselves a little bit of an error cushion in
exchange for being allowed to make some more specific statements. The amount of error
we want to allow is our significance level.

Say we choose a significance level of 0.05. That means that if our data tells us to reject
our null hypothesis, 95% of the time wed be correct to do so, but 5% of the time we will
reject when we shouldnt. A significance level of 0.01 means that 99% of the time wed
be correct to reject the null, while 1% of the time wed be wrong to do so. Similar logic
applies with any specified significance level.

A little intuition using the two-tailed case

Like any other random variable, the T value calculated from our data has a particular
distribution. As the T value gets larger and larger in absolute value, there is a lower and
lower probability of seeing such values if the null hypothesis is true. When we run a
hypothesis test, we see if the T statistic falls toward the center or toward the tails.
Whenever it falls too far out in the tails (too far being determined by our critical t
statistic), we say there is a low probability that we would see that value if it our null were
true, and reject the null.

There two reasons that we could find we should reject our null. First, it could be that our
null hypothesis is incorrect, which is supported by our sample data. Second, it could be
that our null hypothesis is correct, but our sample data is way off base (i.e. we got a
strange sample for one reason or another).

When we use a significance level of 0.05, were setting it up so that 95% of the time, if
our data tells us to reject the null, it is because in fact the null hypothesis is incorrect.
That other 5% of the time, were rejecting when we shouldnt, because our sample just
happens to be strange and is leading us to the wrong conclusion about the population.

The P Value

P values are another way of deciding if we should or shouldnt reject the null hypothesis.
The formula in Excel is

=TDIST(T, df, tails)

where T is equal to the t statistic calculated using our data, df is again our degrees of
freedom, and tails is telling Excel if were doing a 1 or 2-tailed test.

The p value tells us what the probability of incorrectly rejecting the null hypothesis
given a particular value for T. For example, a p value of 0.023 would mean that given
our T stat, we will be incorrectly rejecting the null 2.3% of the time. A p value of 0.359
would mean that given our T, we would be wrong to reject the null 58.2% of the time.

Whenever the p value is LESS than the significance level at which were testing, we
reject the null. Say we were testing at a 5% significance level. Than the p value of 2.3%
would mean that we reject our null (what with 2.3 <5) whereas the 35.9% would mean
we fail to reject (as 35.9 >5).

Two-tailed vs. One-Tailed Tests

Told you wed come back to this. There are two differences worth noting between two-
tailed and one-tailed tests. The first is what we consider to be our rejection region in the t
distribution. For a two-tailed test, we reject if our T statistic is too extreme in either
direction. In other words, we reject if the absolute value of our T is greater than the
critical t. For a one-tailed test, our rejection region depends on our null hypothesis.
5


If our null hypothesis is that is less than or equal to
0
, then the rejection
region is in the upper tail (i.e. a very positive T)
If our null hypothesis is that is greater than or equal to
0
, then the rejection
region is in the lower tail (i.e. a very negative T)

The other difference is how we formulate our question. For a two-tailed test, the null is
always that is equal to some value, where the alternative is that is not. For a one-
tailed test, the null hypothesis is determined by what we are trying to test. As a general
rule, we try to make our hypothesis to be tested the alternative (H
A
), because it is harder
to reject the null than to accept. Said another way:


5
A good general overview of hypothesis testing, including null formulation and rejection regions, is on
page 54 of the class reader.
If the hypothesis to be tested is that is greater than
0
, then we make the null
hypothesis that is less than or equal to
0
.
If the hypothesis to be tested is that is less than
0
, then we make the null
hypothesis is less than or equal to
0
.