Theory meets practice: a commentary on VanderWeele’s ‘principles of confounder selection’

Abstract

When I teach graduate students in pharmacoepidemiology they are well-trained by having taken multiple courses in epidemiologic methods and causal inference. I observe that many feel paralyzed when confronted with real data realizing that such data do not come with tags saying whether variables are common causes of the exposure and outcome or whether they are instrumental variables or colliders. How will they connect the concepts, rules, and exemptions they have learned studying causal inference to the reality of data? Tyler VanderWeele is to be applauded for having compassion with us who spend less time contemplating DAGs and still want to do non-experimental studies that lend themselves to causal conclusions. His pragmatic recommendations are actionable for a broad range of applications yet founded in principled considerations. I tried to put them to a test. I am a pharmacoepidemiologist and study the safety and effectiveness of medications as they are used in routine care without the help of randomized treatment assignment. Like in other areas of epidemiology, causal conclusions are paramount as we apply our insights to patients and influence treatment decisions. It is implicit in Dr. VanderWeele’s commentary that we have a clear grip on temporality and can distinguish between pre-exposure variables and those patient characteristics measured after exposure started. As trivial as this may sound one is surprised how often this is confused, particularly when working with secondary analyses of already collected data [1]. Such mix-up occurs more frequently in case–control sampling designs nested in cohorts of patients identified within longitudinal patient databases. Staying with a cohort analysis and focusing on new users of medications with a defined inception point and covariate assessment before cohort entry substantially reduces the chance for such self-inflicted errors [2]. How do we identify the causes of exposure? Textbooks recommend “prior knowledge” or “a review of the literature” among others. Studying treatment guidelines and conversations with colleagues (better outside of major academic medical centers with their idiosyncrasies) help us to identify potential predictors of treatment choice. While this is a good starting point, predictors for treatment choice often vary substantially depending on health care systems, provider organizations and individual providers. Are we as investigators always doing a most rigorous job in identifying all potential predictors of treatment choice? Investigators studying hypoglycemic events among users of the oral anti-diabetic drugs glyburide versus glipizide inadvertently oversaw that pregnancy and gestational diabetes are strong predictors for choosing one over the other in some healthcare systems. These covariates were not pre-specified by this highly experienced investigator team [3]. Causes of the disease outcome should be identifiable by reading medical textbooks and most causes of disease should be universal. However, in a given population the strength or the observability of such risk factors of a health outcome will vary. Based on our prior knowledge, how sure are we that we have identified all risk factors of the disease and have captured them sufficiently? In addition to these generic issues comes the complication that we usually work with secondary electronic data that were collected by operating a healthcare system, including administrative claims data and electronic health records [4]. This means we are not in control of what, when, and how to measure covariate information and we often rely on proxy variables of the constructs that we are really after. In order to identify diabetes, we may rely on a diagnostic code for diabetes. Was that code checked because a glucose test was performed to rule out diabetes? Should we therefore rather * Sebastian Schneeweiss [email protected]