The paper by Moyà e [1] provides useful discussion for some important statistical issues concerning the possibly complicated ways in which multiple comparisons across primary and secondary endpoints can a ect the results from clinical trials. How to balance tolerable in ation of type I error against a more extensive structure for evaluating success or failure of a clinical trial is the principal question. As suggested by Moyà e, a reasonable way to address this question is to use a larger experimentwise signiÿcance level ( E = 0:10) for a prespeciÿed set of primary and secondary endpoints with the traditional P = 0:05 (two-sided) maintained for the primary endpoint. Through such a structure for conÿrmatory inference, the reporting of success or failure for the results of a clinical trial would include both primary and secondary endpoints, and Moyà e illustrates a convenient notation for this purpose [1]. Such reporting would also agree with the suggestions in Davis [2].
A very important point in the paper by Moyà e is that a priori speciÿcation of the secondary endpoints in a structure that controls the experimentwise signiÿcance level E is necessary for enabling conÿrmatory inferences concerning their results. Moreover, this point applies nearly as strongly to studies with positive results for primary endpoints as those with negative results. The relevant considerations for a situation without such structure for E are that favourable results for secondary endpoints are only interpretable as descriptively supportive when the primary endpoint is positive, whereas they have an exploratory hypothesis generating nature when the primary endpoint is negative. The only way for making conÿrmatory inference possible for a secondary endpoint is with its inclusion in a formal statistical testing procedure, and such a method for conÿrmatory inference must have a reasonable level of control for the experimentwise signiÿcance level (for example, E = 0:10) in order to be convincing.
An important property for a planned testing procedure for secondary endpoints is usefully high power (for example, ¿0:70) for potentially realistic alternatives [3][4][5]. This property is important because it restricts the scope of secondary endpoints for conÿrmatory inference to a possibly small set (for example, 1 to 5 members) for which high power might have some documentation through computations or simulations with respect to information from previous studies. In view of this underlying consideration for a priori speciÿcation for secondary endpoints, a natural question is why not have a secondary endpoint as primary if prior information suggested higher power for it. In cases where such a secondary endpoint has substantial clinical relevance, having it become primary and having a primary endpoint with weaker or more uncertain power become secondary can be a useful strategy [3; 4]. However, secondary endpoints typically have less clinical relevance than a primary endpoint, and whether they have higher power does not have a role