Jessica Young: Estimands versus estimators in studies with causal goals and repeatedly measured outcomes truncated by death

Speaker: Jessica Young

 

Institution: Harvard Medical School

 

Title: Estimands versus estimators in studies with causal goals and repeatedly measured outcomes truncated by death

 

Abstract: Researchers often express interest in estimating causal effects of (possibly time-varying) treatment strategies on the mean of an outcome that is undefined after an individual dies.  For example, the Medications and Weight Gain in PCORnet (MedWeight) was an observational study leveraging electronic health records (EHR) that aimed to estimate the effect of initiating (and subsequently adhering to) different medications for the same indication on future weight change 6 months post treatment initiation in an adult, clinical population where some individuals died in that 6 month period. In this case, the standard notion of an average total causal effect in the original study population is itself undefined.  Truncation by death is a ubiquitous problem, even in randomized trials, that cannot be solved with any particular choice of estimator (statistic).  Rather, truncation by death creates the more fundamental challenge of defining an alternative effect notion to the total causal effect that actually answers the underlying question motivating the investigator in the first place. At the same time, an outcome like weight change is not only informatively measured, but also sparsely measured in an EHR at any specifically chosen follow-up time (e.g. 6 months post-baseline) such that estimators which can time-smooth in available repeated outcomes beyond that chosen time are desirable for improved precision.  In this talk, we consider a notion of causal effect for truncation by death settings that best aligns with the stated motivation behind the MedWeight study.  We emphasize the central importance of the substantive choice of outcome time in reasoning about the relevance of this choice of effect notion and the viability of causal assumptions to identify it in a longitudinal study like MedWeight.  We also pose two options for a scalable IPW estimator of this effect notion for a “causally reasoned” outcome time choice that can borrow across repeated outcomes measures at other times for improved precision.  We consider practical tradeoffs of each these time-smoothing approaches.  Both  are implemented in the R package smoothedIPW: https://r-packages.io/packages/smoothedIPW

 

Bio: Jessica Young is an Associate Professor and Biostatistician in the Harvard Medical School and Harvard Pilgrim Health Care Institute.  She holds a secondary appointment in the Department of Epidemiology at the Harvard T.H. Chan School of Public Health. Her research focuses on the development and practical application of statistical methods for the causal effects of time-varying treatment strategies in observational studies and randomized studies with real-world complications such as treatment nonadherence and loss to follow-up. She also has particular interest in the problem of making valid and meaningful causal inferences in the face of competing or truncation events: terminal events, such as death, that render the study outcome of interest impossible or undefined.