06.06.2013 Views

Theory of Statistics - George Mason University

Theory of Statistics - George Mason University

Theory of Statistics - George Mason University

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

7.8 Confidence Sets 537<br />

Monte Carlo simulation would be an appropriate way to study this situation.<br />

The data could be obtained by a process that involves a stopping<br />

rule, and the tests could be performed in a manner that ignores the process.<br />

Whether or not the test is valid could be assessed by evaluating the p-value<br />

under the null hypothesis.<br />

You are asked to explore this issue in Exercise 7.5.<br />

Evidence Supporting Hypotheses<br />

Example 7.4 is a good illustration <strong>of</strong> the Neyman-Pearson solution to a simple<br />

hypothesis testing problem. We might look at this problem in a slightly different<br />

context, however, as suggested on page 314. This particular example, in<br />

fact, is used in Royall (1997) to show how the data actually provide evidence<br />

that might contradict our decision based on the Neyman-Pearson hypothesis<br />

testing approach.<br />

Example 7.13 Critique <strong>of</strong> the hypothesis test <strong>of</strong> Example 7.4<br />

Let’s consider the Bernoulli distribution <strong>of</strong> Example 7.4 and the two hypotheses<br />

regarding the parameter π, H0 : π = 1/4 and H1 : π = 3/4. The test is<br />

based on x, the total number <strong>of</strong> 1s in n trials. When n = 30, the Neyman-<br />

Pearson testing procedure at the level α = 0.05 rejects H0 in favor <strong>of</strong> H1 if<br />

x ≥ 13.<br />

Looking at the problem <strong>of</strong> choosing H0 or H1 based on the evidence in the<br />

data, however, we might ask what evidence x = 13 provides. The likelihood<br />

ratio in equation 7.16 is<br />

L(P1, x)<br />

= 1/81,<br />

L(P0, x)<br />

which would seem to be rather compelling evidence in favor <strong>of</strong> H0 over H1.<br />

An additional problem with the test in Example 7.4 is the manner in which<br />

a decision is made if x = 12. In order to achieve an exact size <strong>of</strong> α = 0.05,<br />

a randomization procedure that does not depend on the evidence <strong>of</strong> the data<br />

is required. This kind <strong>of</strong> randomization procedure does not seem to be a<br />

reasonable way to make a statistical decision.<br />

The alternative approach to hypothesis testing involves a comparison <strong>of</strong><br />

the evidence from the data in favor <strong>of</strong> each <strong>of</strong> the competing hypotheses.<br />

This approach is similar to the use <strong>of</strong> the Bayes factor discussed in Section<br />

4.5.3.<br />

7.8 Confidence Sets<br />

For statistical confidence sets, the basic problem is to use a random sample<br />

X from an unknown distribution P to determine a random subfamily A(X)<br />

<strong>of</strong> a given family <strong>of</strong> distributions P such that<br />

<strong>Theory</strong> <strong>of</strong> <strong>Statistics</strong> c○2000–2013 James E. Gentle

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!