From estimand selection to adaptive and robust estimation for causal inference: applications in clinical epidemiology

Abstract

In causal inference research, subtle differences in the choice of estimand can have major implications for identifiability and interpretation. When observational data is used to estimate the chosen causal effect, an adaptive and robust approach is necessary for confounding adjustment. This thesis considers the issues of estimand selection and data adaptive estimation in clinical epidemiology through a series of data examples. In Chapter 1, we estimate the effect on mortality of treatment strategies which impose early versus delayed intubation in the critical care setting. We demonstrate how unrealistic treatment strategy definitions in prior analyses could have led to positivity violations, making a causal interpretation challenging. We show how selection of more realistic treatment strategies, and an explicit (target trial) approach to estimating causal effects, can dramatically affect inferences. In Chapter 2, we consider time varying treatment strategies which allow transient periods of non-adherence. Such strategies are of interest in many studies that aim to compare the safety and/or efficacy of different pharmacological treatments indicated for the same disease diagnosis. We formally define treatment strategies which allow transient non-adherence via the inclusion of `grace periods'. Using causal graphs and identification results, we show how the particular rule chosen to determine how treatment should be administered during a grace period impacts not only the interpretation of the causal effect, but also whether identifiability assumptions hold. These issues are illustrated through a data example in which we estimate the effect (on hypertension related outcomes) of taking a thiazide medication versus an angiotensin-converting enzyme inhibitor medication under various grace period treatment strategies. In Chapters 1 and 2, we estimate causal effects using augmented inverse probability weighting, a robust estimator which permits the use of data adaptive (i.e. machine learning) methods for nuisance parameter estimation. Under standard no confounding assumptions, so called doubly robust machine learning estimators can be nearly unbiased under mild complexity reducing assumptions. In practice, analysts cannot be certain that such assumptions hold. In Chapter 3, we present methods, based on higher order influence functions, which allow analysts to falsify their complexity reducing assumptions. These methods are applied to an analysis which estimates the effect of laparoscopic versus open surgery for the treatment of colon cancer. Our higher order influence function based estimator identifies substantial bias when less flexible nuisance parameter estimators are used to compute the causal effect.