Parametric effect size estimates from sipuleucel-T randomized trials.

Abstract

75 Background: Median survival differences are frequently used as measures of effect size because they are easily understood. Other effect size estimates can more fully represent the whole of survival distributions, however, and are less subject to local variations. The sipuleucel-T D9901 (Small, JCO 2006) and IMPACT (Kantoff, NEJM 2010) trials provide 4.5 and 4.1 month estimated median survival differences, respectively, and hazard ratio estimates of 0.586 and 0.752 (not stratified or adjusted). These estimates are based on non-parametric (median) or semi-parametric (hazard ratio) methods. Parametric statistical models are used to illustrate alternative effect size estimates. Methods: Parametric models used are based on the Weibull distribution and a modification of the Weibull distribution allowing for delay of effect. Delay of effect is often conjectured as being responsible for delayed separation of Kaplan-Meier estimates in immunotherapy evaluations (Hoos, JNCI 2010), and both trials exhibit some degree of delayed separation. Parametric models specify the general shape of the hazard functions. Results: The following table repeats estimates given above and shows parametric modeling results. For D9901 the median difference estimate from the Weibull model suggest a much larger effect, and the median difference and the hazard ratio estimates using the delayed effect Weibull model are both indicative of more effect. For IMPACT the median difference and hazard ratio estimates from the Weibull model indicate more effect, but not as dramatic as for D9901. The delayed effect modified Weibull model for IMPACT does not support the conjecture of delayed effect. Conclusions: Statistical modeling illustrates alternative methods of obtaining effect size estimates. For the sipuleucel-T trials there is a suggestion of more effect than obtained from non-parametric or semi-parametric methods. [Table: see text]