Overview
James M. Robins is Professor of Epidemiology and Biostatistics at the Harvard T.H. Chan School of Public Health. He is one of the most influential methodologists in causal inference, having independently developed a potential outcomes framework for noncompliance in clinical trials (Robins 1989) in parallel with the econometric and statistical traditions. His contributions span causal graph theory, structural nested models, marginal structural models, G-computation, and inverse probability of censoring weighting — methods that have become standard in epidemiology and are increasingly used in economics and statistics.
Key Contributions / Features
- Robins (1989) — "The Analysis of Randomized and Nonrandomized AIDS Treatment Trials Using a New Approach to Causal Inference in Longitudinal Studies." In the context of AIDS clinical trial noncompliance, independently assumed the same conditions as AIR (exclusion restriction, monotonicity, random assignment) and analyzed the ATE parameter, computed bounds, and performed sensitivity analysis — predating the AIR framework.
- Structural Nested Mean Models (SNMM) — Robins' Assumption 6 (1989) identifies ATE as the IV estimand without monotonicity and without constant treatment effects; requires only that average treatment effects are equal for treated and untreated within each arm.
- G-computation algorithm — Formula for identifying E[Y(d)] under covariate-adjusted compliance: El E[Y|Z=d, D=d, L=l] pr[L=l|Z=d]; the foundational result for causal inference with time-varying treatments.
- IPCW estimators — Inverse probability of censoring weighted estimators for E[Y(d)] under conditional independence of compliance, extending G-computation to continuous covariates.
- Robins-Manski bounds — Bounds on ATE without monotonicity, under weak and strong randomization assumptions; sharpened by Balke and Pearl (1993) for certain data distributions.
- Robins and Greenland (1996) — Comment on AIR (1996); argued ATE is of greater public health interest than LATE or ITT, derived bounds and SNMM-based identification, and showed LATE breaks down in bioequivalence trials.
Related
Sources