Ehsan Karim

He develops causal inference and machine learning methods that turn large health administrative databases into reliable real-world evidence. Causal Inference Methods: Develops and evaluates methods for estimating treatment effects from non-randomized data — including targeted maximum likelihood estimation (TMLE) with single and double cross-fitting, marginal structural models, propensity score approaches, and target trial emulation. His work tackles time-dependent confounding, treatment-confounder feedback, and immortal time bias in longitudinal studies, and benchmarks estimators with plasmode simulations built from real health data. Machine Learning & AI for Health: Investigates when machine learning genuinely improves causal estimation — from Super Learner weight estimation to deep-learning propensity score architectures such as Dragonnet and autoencoders. The goal is rigorous, verifiable use of AI in health research rather than black-box prediction. Real-World Evidence & Pharmacoepidemiology: Advances the methodology of real-world evidence generation, including high-dimensional propensity score (hdPS) techniques and their machine learning and doubly robust extensions for residual confounding control. This program also addresses non-adherence and per-protocol effects in pragmatic clinical trials. He is co-principal investigator of a CIHR-funded program of target trial emulations evaluating prescribing practices for opioid use disorder with linked British Columbia administrative data. Health Administrative Data: Builds statistical strategies for the scale and messiness of population-level administrative databases — proxy adjustment, missing data, survival prediction, and disease-specific comorbidity indices — so routinely collected health data can support trustworthy inference. Multiple Sclerosis Applications: Applies these methods where they matter: comparative effectiveness of MS therapies, an MS-specific Comorbidity Summary Index (MSCSI), and a risk prediction algorithm for the MS prodrome to enable earlier detection. Supported by CIHR and MS Canada. Survey & Longitudinal Methods: Develops guidance for analyzing nationally representative complex surveys (Canadian, US, and international) and longitudinal cohorts, including propensity score analysis and design-based, survey-weighted TMLE in the complex survey context, and small-sample inference for stepped-wedge designs.

[2026] Prescription time-distribution matching and immortal time bias: promise, failure modes, and time-aware alternatives. Karim ME. Current Opinion in Epidemiology and Public Health 2026 (advance online) (Invited review) [2026] LASSO-based survival prediction modeling with multiply imputed data: a case study in tuberculosis mortality prediction. Hossain MB, Sadatsafavi M, Johnston JC, Wong H, Cook VJ, Karim ME. The American Statistician 80(1), 77–88, 2026 (Prediction) [2025] High-dimensional propensity score and its machine learning extensions in residual confounding control. Karim ME. The American Statistician 79(1), 72–90, 2025 (Methods) [2025] Towards robust causal inference in epidemiologic research: employing double cross-fit TMLE in right heart catheterization data. Mondol MH, Kari