Defining and Computing True Clinical Trial-relevant Causal Estimands in Simulation Studies
Simulation studies are widely used in pharmaceutical research to evaluate operating characteristics of statistical methods proposed for regulatory analyses. Central to such studies is the ability to compare against a known “truth” implied by assumed the data-generating mechanism. In practice, however, this advantage is often limited than assumed, as the target is not a simple function of data-generating parameters. This challenge is well illustrated by oncology trials in which patients randomized to control are allowed to switch to the experimental treatment after disease progression. Regulatory analyses in such settings often targets questions concerning what the treatment effect would have been had switching not occurred, however this is not a simple number that can be derived explicitly from parameters of the data-generating model.
Simulation studies are thus often used to try to determine the “truth” up to some reasonable level of precision. However, when the relationship between the type of treatment effect, the assumed data-generating mechanism, and the numerical procedure used to compute the “truth” are not explicit, simulation results may conflate approximation error with estimator bias, leading to potentially misleading conclusions.
This project will develop more precise approaches for determining “truth” in the context of longitudinal data.
Erica Moodie
Core Clinical Sciences
Mathematics
Professional, scientific and technical services
McGill University
Accelerate
