Efficient Inverse Probability-Weighted Estimation of Longitudinal Causal Effects in Quasi-Experimental Studies

• Activity Number: 56 - Causal Inference
• Type: Contributed
• Date/Time: Sunday, August 8, 2021 : 3:30 PM to 5:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Nima Hejazi* and Mark Van Der Laan and Sathya Anand and Wenjing Zheng
• Companies: University of California, Berkeley and University of California and Netflix and Netflix
• Keywords: inverse probability weighting; machine learning; semiparametric efficiency; influence function; undersmoothing; nonparametric causal inference

Abstract

Inverse propensity weighting (IPW) estimates causal effects by re-weighting observations to balance confounder distributions across treatment groups, thus adjusting for selection bias. IPW can handle complex longitudinal data, with censoring, time-to-event outcomes, and multiple time-point interventions. Valid inference for effect estimates relies upon consistent estimation of the re-weighting mechanisms. While machine learning (ML) algorithms may seem aptly suited, most ML algorithms lack the necessary convergence rate guarantees for valid statistical inference. Moreover, traditional IPW estimators can suffer from bias and be inefficient. To resolve these issues, we propose an efficient, nonparametric IPW estimator of the causal effects of multiple time-point interventions in longitudinal settings. Our estimator leverages the highly adaptive lasso algorithm to flexibly estimate re-weighting mechanisms at rates suitably fast for valid inference. The proposed IPW estimator is consistent after a targeted undersmoothing correction, with variance converging to the semiparametric efficiency bound. We illustrate our approach in numerical experiments and introduce open source software.