Title: Causal variance decompositions for polytomous and functional exposures
Abstract: Institutional/provider comparisons in healthcare can be considered as causal inference problems, answering questions of the type "What kind of care/outcome a patient would be expected to have if treated in a given hospital (in contrast to a reference hospital or a benchmark level of care)?". Since the treating hospital is considered here as a multi-category polytomous exposure, there will be a large number of such contrasts, motivating the question of how much of the overall variation in the patient outcomes is attributable to between-hospital differences in performance. For this purpose, we consider causal variance decompositions. In this talk, I will briefly review causal inference techniques used for institutional comparisons, including direct and indirect standardization. I will discuss the causal interpretation of a variance decomposition that splits the observed variation in the outcome to that due to between-hospital differences, due to patient case-mix, and residual variance. I will extend these ideas to hierarchical exposures (e.g. surgeons within hospitals), and causal mediation analysis. The methods are illustrated with applications to Ontario administrative data on cancer care. I will also discuss adapting the variance decomposition approach to functional exposures in a radiation treatment context, which requires dimension reduction techniques.
Download the flyer here.