Primary Submission Category: Randomized Designs and Analyses
Toward a Causal Framework for Crossover Trials: From Estimands to Estimation
Authors: Richard Liu, Mats Julius Stensrud, Michele Santacatterina,
Presenting Author: Richard Liu*
Crossover trials are commonly used in clinical research to improve efficiency in small sample settings, as each participant serves as their own control across treatment periods. Despite their widespread use, however, a principled framework for defining, identifying, and estimating causal effects in this design remains lacking. In particular, causal estimands are often poorly defined, and the identification assumptions required are rarely made explicit. As a result, it is often unclear which estimand is being targeted and under what conditions commonly used estimation procedures achieve efficiency gains or introduce bias. A key example of this ambiguity lies in the definition and identification of carryover effects and washout periods, which are inconsistently handled across both study designs and analytical approaches. This methodological gap has been highlighted in recent FDA guidance advocating for the use of a formal estimand framework. In this work, we leverage tools from causal inference and modern statistical methodology to address these challenges. We propose a set of causal models tailored to the crossover design, clarify the corresponding causal estimands and their identification assumptions, and examine when canonical statistical methods yield unbiased and efficient estimates. In addition, we also introduce doubly robust estimators to further enhance robustness while preserving efficiency. Further, we formalize the notion and the role of carryover effects.