Non-parametric Causal Inference in Dynamic Thresholding Designs

Learn non-parametric causal inference methods for dynamic thresholding designs in this 30-minute conference talk from the Simons Institute. Explore how to estimate causal effects in settings where treatment assignment depends on crossing predetermined thresholds over time, using the example of prediabetes diagnosis based on fasting blood sugar levels. Discover why traditional regression-discontinuity analysis fails in dynamic systems where patients can move between treatment states across multiple visits, and understand how temporal dynamics complicate causal inference. Master the concept of dynamic marginal policy effects at treatment thresholds and examine reduced-form characterizations that remain valid in these complex settings. Gain practical skills in local-linear-regression approaches for estimation and inference of these estimands, and review numerical experiments demonstrating the effectiveness of these methods. Understand the broader applications of policy-gradient methods for causal inference in observational studies, particularly in healthcare and social systems where treatment decisions evolve over time based on continuously monitored biomarkers or behavioral indicators.