Estimands in Randomized Clinical Trials with Complex Life History Processes

Abstract

Clinical trials in oncology, cardiovascular disease and many other settings are dealing with complex outcomes involving multiple endpoints, competing or semi-competing risks, loss to follow-up and cointerventions related to the management and care of patients. These all complicate the design, analysis and interpretation of randomized trials. In such settings, traditional analyses of the time to some event are not sufficient for assessing new treatments. This thesis discusses the issues involved in the specification of estimands within a comprehensive multistate model framework. Intensity-based multistate models are used to (a) conceptualize the event-generating process involving primary and post-randomization outcomes, to (b) define and interpret causal estimands based on observable marginal features of the process, and to (c) conduct secondary analyses of marginal treatment effects. For a broad range of disease process settings, we investigate how marginal estimands depend on the full intensity-based process. Using large sample theory, factors influencing the limiting values of estimators of treatment effect in generalized linear models for marginal process features are studied. Rejection rates of a variety of hypothesis tests based on marginal regression models are also examined in terms of the true intensity-based process; based on these findings robustness properties are established. Such numerical investigations give insights into the interpretation and use of marginal estimands in randomized trials. We discuss in detail estimands based on cumulative incidence function regression for semi-competing risks processes and mean function regression for processes involving recurrent and terminal events. Specification of utilities for different disease-related outcomes, rescue interventions and other post-randomization events facilitate synthesis of information on complex disease processes, enabling simple causal treatment comparisons. Derivations are provided of an infinitesimal jackknife variance estimator for utility-based estimands to facilitate robust methods for causal inference of randomized trials.