Lars van der Laan

About me

I am a postdoctoral scholar in the Department of Management Science and Engineering at Stanford University and a Stanford Causal Science Center Fellow, working with Vasilis Syrgkanis and supported by Netflix Research. I completed my Ph.D. in Statistics at the University of Washington in June 2026, advised by Marco Carone and Alex Luedtke, during which I was supported by a Netflix Graduate Research Fellowship. I also collaborate with Nathan Kallus and Aurélien Bibaut at Netflix.

My research focuses on causal inference and semiparametric statistics, particularly how to use machine learning for valid and efficient statistical inference. I have developed methods for automatic debiasing of smooth functionals of nonparametric M-estimands, efficient estimation and calibration of heterogeneous treatment effects, calibrated debiased machine learning for doubly robust inference, and adaptive inference after data-driven model selection. My work also addresses a broad range of causal inference problems, including long-term causal inference through instrumental variables with surrogate outcomes and dynamic modeling in reinforcement learning, as well as data fusion to improve efficiency in randomized trials.

I also work on dynamic decision-making and reinforcement learning, with a focus on learning theory and statistical inference. In off-policy evaluation and offline reinforcement learning, I study the estimation, calibration, and model selection of occupancy ratios and value functions under function approximation and distribution shift, with an emphasis on avoiding restrictive Bellman completeness assumptions. I also study debiased estimation and efficient inference in off-policy evaluation, inverse reinforcement learning, and dynamic discrete choice models.

Calibration is a recurring theme across my research. I study how to improve existing machine-learning estimates through post-hoc calibration, often without retraining the underlying models. My work has shown how calibration can improve predictions of heterogeneous treatment effects and long-term policy values in offline reinforcement learning; improve inverse probability weighting estimators; increase statistical efficiency in prediction-powered inference and randomized trials; and weaken nuisance-estimation requirements for doubly robust inference. I have also developed calibration-based methods for conformal prediction and distribution-free uncertainty quantification.

CV · Google Scholar · Publications · Email