Challenges in and Educational Strategies for Teaching Causal Inference in the Health Sciences

Abstract

Many areas of clinical research, such as comparative effectiveness research and patient-centered outcomes research, strongly depend on making causal inferences from observational data. Further, these topic areas also utilize pragmatic trials and quasi-experimental designs, where consistent estimation of causal effects is also more challenging than traditional randomized controlled trials, and/or involves distinct approaches for intention-to-treat versus as-treated or per-protocol effects. While substantial literature exists on associated designs and analysis strategy, the corresponding methods are complex and not always taught in formal training, even within graduate statistics or biostatistics programs. Therefore, a critical need exists for accessible educational resources and the expansion of relevant courses and training programs. Regarding that goal, however, significant debate exists on whether these advanced methods should even be taught at all to non-statisticians, and/or researchers with more limited statistical training (e.g. a fundamental course and some background in regression). This talk proposes some possible perspectives to effectively address these concerns, while still avoiding the result of "knowing enough to be dangerous". The presenter has some related links at www.landsittellab.pitt.edu. This work was supported by AHRQ grant R25HS023185, PCORI contract R-IMC-1306-03827, and supplemental funding from the NIH/NLM grant 5 T15 LM007059-32.