Events

Colloquium

Causal Inference Beyond Common Support

Speaker: Harsh Parikh (Yale) Assistant Professor of Biostatistics Yale University Wednesday, January 28, 2026 11:30AM - 1:00PM Lunch at 11:30am in 1307 Talk 12:00-1:00pm in 1327 Location: Yale Institute for Foundations of Data Science & Webcast, 219 Prospect Street, 13th Floor, New Haven, CT 06511 and via Webcast: https://yale.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=6b5e69ff-cace-48a0-8f48-b3ca013e1936

Abstract: Generalizing causal findings from randomized controlled trials (RCTs) to target populations is a core challenge in modern statistics—one where violations of the positivity assumption force an uncomfortable choice between imprecise estimates and unverifiable parametric assumptions. This talk presents two complementary methodological contributions that reframe how we diagnose and address these violations. First, we introduce the Rashomon Set of Optimal Trees (ROOT), a novel functional optimization framework that leverages the multiplicity of near-optimal decision trees to interpretably characterize underrepresented subgroups. ROOT identifies covariate regions contributing most to estimator variance, enabling transparent communication of where reliable inference ends—transforming a nuisance problem into actionable diagnostics. Second, we develop a unified regularization framework for extrapolation that replaces hard non-negativity constraints on estimator weights with a tunable soft penalty. This formulation reveals a novel “bias-bias-variance” tradeoff—explicitly balancing covariate imbalance bias, model misspecification bias, and variance—yielding a continuum of estimators with provable properties for navigating extrapolation risk. We demonstrate both methods through a case study generalizing the START trial (comparing opioid agonist therapies) to a national treatment population. Together, these tools offer statisticians a principled workflow: first characterize the boundaries of credible inference, then carefully extend them through regularization when necessary.

Speaker Bio: My research focuses on developing (interpretable) causal inference approaches for aiding decisions in high-stakes complex scenarios. My collaborators and I have used my research to address challenges in healthcare, public health, and social sciences. Decision-making in these critical domains is fraught with difficulties stemming from, but not limited to, the intricate interplay of factors, including the heterogeneity of causal effects across subpopulations, the substantial costs associated with suboptimal decisions, and the inherent complexities in the available data, all of which complicate the assessment of risk-benefit trade-offs. In pursuit of more effective solutions, my work is centered around the development of causal inference methodologies that are:

Accurate: to ensure accurate estimation of heterogeneous causal effects, even in scenarios with data limitations, offering decision-makers a reliable foundation upon which to base their choices.

Trustworthy: to empower domain experts to comprehend the inner workings of the causal inference process. This not only enables experts to validate the underlying assumptions but also guarantees patients’ safety.

Domain-conscious: to bridge the research-to-practice gap and yield solutions that are readily implementable. I leverage the context and domain knowledge to tailor solutions specific to a subject matter.

Add To: Google Calendar | Outlook | iCal File

• Colloquium