Non-parametric Causal Inference in Dynamic Thresholding Designs
Talkat Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026) · 2026-01-12..01-18·
summary
Argues that threshold rules in time-series settings identify a dynamic marginal policy effect (long-run benefit per extra treated person) and gives a concrete estimator: an aggregated, weighted local-linear-regression estimator that regresses realized discounted future rewards (and separately future actions) near the cutoff. Derives the target via a version of the policy-gradient theorem and Q-functions, shows a simulator where the long-run effect is −4.2 vs the one-step RD −1, and demonstrates consistent estimation with standard bandwidth choices and bootstrap CIs.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:30Regression discontinuity in dynamic systems
- 8:24Regression discontinuity design basics
- 11:29Prediction to intervention mindset
- 13:18Dynamic setting and policy value
- 16:42Assumptions for dynamic RD
- 19:21Estimating dynamic RD parameter
- 22:02Preventative care example
- 25:32Summary
- 27:49Q&A
related talks
15mExploring the math of thresholds
12mCausal Inference And Data-Driven Decision-Making
44mIdentification and Estimation of Dynamic Structural Models with Unobserved Choices
30mCounterfactual Inference For Sequential Experiment Design
01h 11mStatistical Limits of Causal Inference
33mSimple Yet Efficient Estimators For Network Causal Inference...
01h 08mAn Introduction to Potential Outcomes, DAGs and Single-World Intervention Graphs
47mPrivate targeting strategies