Clinical Trials

Clinical Trials: A Practical Guide ■❚❙❘Table 1. Different types of multiplicity and possible strategies for adjusting for multiplicity.Type of multiplicityMultiple endpointsMultiple treatmentsSubgroup analysesInterim analysisPossible strategiesSpecify primary and secondary endpoints in advanceDefine a summary statistic as an endpointPredefine meaningful composite endpointsChoose a multivariate global test statisticUse Bonferroni correctionPredefine priorities or use a global test statistic to test for any differencePredefine a limited number of subgroups in the protocol to be analyzedwith interaction testsUse Bonferroni correction if uncritical testing of many subgroupsPredefine P-values for efficacy that statistically account for the frequencyof interim analysis, with more stringent P-values for early analysesThere were visible merits in using a multivariate approach in this trial, because allfive primary outcomes by themselves and as a whole gave a credible and consistentanswer in terms of an overall direction, indicating a benefit of treatment.However, in situations where trials have uninformative or inconsistent results,such a multivariate statistical approach with a possible borderline global resultmight give controversial answers that have to be interpreted with caution, therebytaking account of the clinical setting of the trial. In certain situations suchparadoxes may be natural, eg, if death is prevented then nonfatal events mayincrease in frequency, such as disease flair-ups during the extended survivalperiod [15].ConclusionThe basic principles of statistical testing apply to a single test of a single nullhypothesis. However, the problem of multiplicity or multiple statisticalcomparisons is a common one in clinical research, arising when severalcomparisons are undertaken from which to draw conclusions. Repeated ormultiple testing of different outcome variables or many different treatmentswithin the same trial will tend to increase the likelihood of finding a statisticallysignificant difference by chance alone, inflating the overall Type I error, which canundermine the validity of the statistical analyses unless accounted for.There are several strategies with which to deal with the problem of multipletesting – either at the design or at the analysis stage. A table of summary points isprovided (see Table 1). The simplest advice would be to design the trial by337