Bringing Light to the Threshold: Identification of Multi-Score Regression Discontinuity Effects with Application to LED Manufacturing

Oliver Schacht, Philipp Alexander Schwarz, Sven Klaassen, Johannes Oberpriller, Martin Spindler

Keywords: Causal Machine Learning, Regression Discontinuity Design, Treatment Effect Identification, Causal Inference Application

TL;DR: We derive new identification results for the Multi-Score RDD framework and apply ML-adjusted RDD estimators to manufacturing data accordingly.

Abstract: The regression discontinuity design (RDD) is a widely used framework for threshold-based causal effect estimation in causal inference. Recent extensions incorporating machine learning (ML) adjustments have made RDD an appealing approach for researchers utilizing causal ML toolkits. However, many real-world applications, such as production systems, involve multiple decision criteria and logically connected thresholds, necessitating more sophisticated identification strategies, which are not clearly addressed in the recent literature. We derive a novel identification result for the complier effect in the multi-score RDD (MRD) setting by extending unit behavior types to multiple dimensions. Further, we show that under mild assumptions, this identification result does not depend on subsets of units with constant response. We apply our findings to simulated and real-world data from opto-electronic semiconductor manufacturing, employing estimators that adjust for covariates through machine learning. Our results offer insights into enhancing current production policies by optimizing the cutoff points, demonstrating the applicability of MRD in a manufacturing context.

Primary Area: causal reasoning

Submission Number: 17425