LAUSR

D iscussions around the validity of treatment effect estimates in comparative effectiveness research (CER) commonly focus on general concepts of bias and, specifically, whether assumptions related to confounding are satisfied. Although necessary, confounding is only one of the considerations required for selecting estimators and interpreting estimates when using observational data for CER. It is important to acknowledge that average treatment effect estimates generated from observational data reflect the average treatment effect for alternative distinct and limited sets of patients in a study population depending on the data available, the estimator used, and the circumstances that beget treatment choice and treatment effect heterogeneity. This commentary provides an overview of the treatment effect estimates produced by estimators commonly used in CER and discusses the fundamental assumptions required for these alternative estimators to produce estimates that can inform different CER research objectives. It is assumed throughout that the conditions required for each estimator to produce unbiased and consistent estimates are satisfied; for example, all discussion surrounding risk-adjustment methods assumes there are no unmeasured confounders. It is well-known that linear 2-stage least squares instrumental variable (IV) estimators generate consistent estimates of a local average treatment effect (LATE) in a sample. An estimate of a LATE represents the average treatment effect for those patients whose treatment choices were sensitive to the instrument(s) in the model. This subset is commonly referred to as the “marginal” patients in the health economics literature and is often theorized to include those patients for whom the benefits and costs associated with treatment, and therefore the treatment choice, are less certain. It is also known—although perhaps not as well known—that standard risk adjustment (RA) estimators (eg, ordinary least squares, generalized linear models) yield estimates of the average treatment effect for the treated patients in the sample (ATT). Making inferences on treatment effects for any other patient groups or for the study population as a whole always requires additional assumptions grounded in theory linking treatment choice, outcome, and treatment effectiveness across patients in the real world.