Analyzing Heterogeneous Treatment Effects Using Panel Data Models

• Applicant: Dr. Christian Gische
• Subject Area: Personality Psychology, Clinical and Medical Psychology, Methodology
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 554620660

Over the past few decades, formal methods have been proposed that allow the learning of cause-effect relationships from observational data under certain conditions. The most prominent causal models include graph-based causal models and potential outcome causal models. Among observational designs, panel designs are of particular interest for causal inference since they adhere to the temporal order of cause-effect relationships and allow for inferences to be made on person-specific treatment effects. The literature in the behavioral sciences and psychology provides important initial findings with respect to three central challenges to causal inference from panel data: unobserved common causes, unobserved person heterogeneity, and the distinction of within- and between-person causal effects. However, most of these findings are restricted to linear panel data models with additive unobserved heterogeneity. While linear and additive models are a natural starting point for the development of quantitative research methods, they often provide poor approximations to the complex mechanisms underlying human behavior. For instance, linear and additive models do not capture situations in which individuals' responses to treatment differ in the mean of the outcome distribution. This project aims at developing causal models that adequately incorporate a variety of complex causal mechanisms that are ubiquitous in human behavior. To this end, I analyze the potential for causal inference based on nonlinear nonadditive panel data models. These models permit researchers to analyze unobserved effect heterogeneity, cross-level interactions, and moderated mediation. The overarching objective of this project is to provide researchers with adaptable tools to infer individual-specific causal effects in the presence of unobserved confounders and unobserved heterogeneity. To achieve this objective, I will proceed in accordance with a roadmap comprising six work packages. I commence by defining several causal estimands of frequent interest (Work Package 1), establish causal identification (Work Package 2), and derive efficient estimators of causal effects (Work Package 3), which will be implemented in a user-friendly R package entitled causalSEM (Work Package 4). To facilitate the accessibility of the proposed methods to researchers with diverse levels of prior training in quantitative methods, I will provide an illustrative nontechnical discussion of the causal assumptions underlying the model (Work Package 5) and an empirical application written in an accessible and reproducible style (Work Package 6). To ensure the reproducibility and transparency of the results, I will publish them in international peer-reviewed journals that ensure free online access.

• DFG Programme: Research Grants
• Co-Investigator: Professor Dr. Manuel Voelkle