All Times EDT
Keywords: Causal inference, collapsibility, conditional estimand, logic respect, marginal estimand
ICH E9 R1 encourages understanding of ‘treatment effect’ with clarity: How the outcome of treatment compares to what would have happened to the same patients under alternative treatment (i.e., had they not received the treatment, or had they received a different treatment). While both conditional and marginal estimands allow for a causal interpretation in the sense of ICH E9 R1, their distinction may be neglected and this may lead to interpreting a conditional estimand as a marginal one, and vice versa. The distinction between conditional and marginal estimands is closely related to the collapsibility or the logic respecting property. In this talk, we assess the connection between the two definitions. For binary and time-to-Event endpoints, commonly used estimators may not satisfy the two definitions and be misinterpreted, e.g., using a conditional estimator for a marginal estimand. In addition, we also discuss standardization as a general tool to estimate marginal estimands with covariate adjustments.