Development of endpoints, primary and secondary endpoints, in clinical trials involves consideration of many aspects. Some aspects are whether the endpoints can provide adequate assessments of clinical events (e.g., mortality, morbidity), patient symptoms, measures of function (e.g., ability to walk or exercise), whether the trial can be reasonably powered to study the investigational therapy, whether the endpoints can reasonably well provide public health measures. Recently there seems to be an increasing trend toward adding more components to composite endpoints or multi-component endpoints. In addition, more and more attention has been drawn to the so-called “reasonably likely surrogate” that is a biomarker considered to be able to reasonably likely predict clinical outcomes. Given such greater complexities, handling errors associated with statistical inferences becomes increasing unclear. This talk is planned to focus on what statistical errors to control, measures of substantial evidence, whether other aspects (e.g., doses, heterogeneity of diseases) need to play roles in error control.

Typical analyses of clinical trials involve intervention comparisons for each efficacy and safety outcome. Outcome-specific effects are estimated and marginal effects are potentially combined in benefit:risk analyses. It is widely believed that such analyses provide comprehensive information regarding the intervention effects on patients. However such approaches do not incorporate associations between outcomes of interest, suffer from competing risk challenges when interpreting outcome-specific results, do not recognize the cumulative nature of multiple outcomes on individual patients, and since efficacy and safety analyses are often conducted using different analysis populations, the population to which such benefit:risk analyses apply, is unclear.

The Desirability of Outcome Ranking (DOOR) methodology is a paradigm that resolves these challenges. The DOOR methodology allows us to more effectively evaluate and select treatment strategies by providing a more informative way to compare the patient-centric benefits and risks of intervention alternatives. In this talk, we discuss statistical methods for design, analysis and monitoring of clinical trials with DOOR.

In randomized controlled clinical trials where time-to-event outcomes are the primary endpoint, hazard ratios are commonly employed to quantify the treatment effect. However, when faced with non-proportional hazards, the hazard ratio is not well defined, necessitating alternative pproaches to quantify treatment effects. Such non-proportional hazards might arise, for example, from treatments with delayed effects or when the treatment effect varies across subgroups. In scenarios with non-proportional hazards, a single parameter might be inadequate to capture the differences in survival functions. For instance, in cases where the survival functions intersect, several characteristics of the survival distribution must be considered to determine the more desirable survival function. In this study, we evaluate trials with multiple primary endpoints corresponding to various characteristics of the survival functions. This encompasses the differences and ratios of milestone survival probabilities, differences in the quantiles of the survival distribution, differences in restricted mean survival times, and the average hazard ratio. By employing the counting process representation of survival functions, we show that the parameter estimates are asymptotically multivariate normal and derive their correlations. To account for multiple comparisons, we introduce multiple testing procedures and simultaneous confidence intervals that consider the correlation between different test statistics and also incorporate the logrank test. Through simulations, we assess the finite sample type I error rate and power of the proposed methods and describe the R package 'nph' implementing the procedure.

Reference:

R Ristl, H Götte, A Schüler, M Posch, F König. Simultaneous inference procedures for the comparison of multiple characteristics of two survival functions, 2023 (submitted)