 |
 |
Greetings ! I'd appreciate help in determining sample size for planned analyses using McNemar's test. It seems to be addressed in only one way in the literature, which generally assumes that the study null hypothesis is the same as the test's null hypothesis, i.e. that p1.=p.1. But, in practice, the study hypothesis might more reasonably be p1. ne p.1. If so, should this affect sample size calculations ? I have not been able to find an answer to this. For example, if looking at the change in voting preferences for Al Gore or GW Bush, after a televised viewing of how each smokes a cigar, it would seem reasonable that the study and test null hypothesis should be that the proportion voting for Al Gore remained unchanged by the experience. Standard sample size determinations might apply. But what if, for example, someone has proposed that the development of diabetes in an individual is related to that individual having brown hair ? Analysis using McNemar's test might seem usable: we are looking at paired attributes within individuals. But here, should not the burden of proof be on proving the association ? So should not the null hypothesis be that there is no relationship ? A p-value from McNemar's test here has less "significance" than the same p-value from the application above, given that one might expect the association to be less likely. So maybe it would be best to analyse in some other way (any suggestions ?) If you were to proceed with McNemar's test in this situation, and wanted to work out sample size, should you reverse "significance" and "power" (because the study's null hypothesis is reversed relative to the test's ?). I'd welcome any feedback on this. Best Wishes, Martin Holt [log in to unmask] Medical Statistician Derby City General Hospital %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%