Unidirectional multiple comparisons of Poisson rates.

This paper presents an exact significance test procedure for unidirectional multiple comparisons of Poisson variables. It is sufficiently general for application to situations where the 'denominator', which might be the amount of follow-up time in cohort studies or the expected number of events in studies that use standardized rates, differs from one group to the other, and where a group-specific alternative hypothesis is of interest. We use unidirectional Z statistics, of the form proposed by Dunnett, along with the conditional distribution needed in computing the exact attained significance levels, to address three distinct unidirectional hypotheses. In one case, we compare one exposed with several reference groups, while in the other two cases, we compare several exposed groups with one reference group. We also provide a multiple stage testing approach. Data from studies in cancer and trauma epidemiology illustrate the procedures.