Primary Submission Category: Causal Fairness, and Bias/Discrimination
Fairness-Constrained Individualized Treatment Rules and Medication Decisions for Opioid Use Disorder
Authors: Safiya Sirota, Rachael Ross, Kara Rudolph, Daniel Malinsky,
Presenting Author: Safiya Sirota*
There are multiple evidence-based medications that are commonly used for the treatment of opioid use disorder (OUD). It is not well-established which drug most effectively prevents negative health outcomes, given a patient’s individual characteristics. One may endeavor to estimate an individualized treatment rule (ITR) from observational data to address this question. However, observational datasets may encode social or structural biases (e.g., provider bias) that reflect disadvantages faced by marginalized groups. Imposing fairness constraints may help mitigate these pre-existing biases in the training data. We outline a framework for estimating an optimal fair ITR, using a causal path-specific approach to fairness. Our procedure enables estimation of the constrained ITR’s value and valid quantification of uncertainty in the value. We demonstrate the methodology with simulated data and an application to Medicaid data on medication decisions for patients with OUD. The estimation of a valid confidence interval around the fair ITR value provides a means of comparing alternative candidate rules, including the optimal unconstrained ITR. Assessing differences in expected outcomes under candidate ITRs is crucial to knowledgeably balance the dual objectives of curbing retrospective bias and suggesting treatment regimes that perform well.