School Seminars and Colloquia

Why do users of classical statistical procedures continue to use them even though P-values are like a Birmingham screwdriver?

Statistics Seminar

by Geoff Robinson


Institution: CSIRO
Date: Tue 31st January 2012
Time: 1:00 PM
Location: Room 213, Richard Berry Building, University of Melbourne

Abstract: First, I will review the case against P-values and speculate
about why it has been largely ignored. The only novel result is that
for tests between simple hypotheses, any statistical analysis
procedure which is not conservative relative to the likelihood ratio
can be criticised by constructing a relevant betting procedure. This
betting procedure appears to offer a profit margin to a person quoting
Fisher-Neyman-Pearson confidence levels but it has positive
expectation under both hypotheses to the person criticising those
confidence levels.

Following on from the argument that P-values are not a sensible
measure of the strength of evidence for the simple situation of
testing between two simple hypotheses, I will argue that Wald's
sequential probability ratio test should be modified, that tests
between simple null hypotheses and compound alternative hypotheses
should quote a likelihood ratio rather than a P-level as the measure
of statistical significance, and that for pure hypothesis testing
where there is no defined alternative the tail probability should be
divided by an estimate of Mills' ratio before quoting the result as a
measure of statistical significance. Finally, I will discuss
practical consequences of this alternative approach for the
interpretation of two epidemiological trials.

For More Information: contact: Mihee Lee. email: miheel@unimelb.edu.au