Novel Methods for Extending Causal Inference to a Target Population

This dissertation presents two methods to address positivity violations in extending causal inference from study samples to broader target populations, enhancing the robustness and applicability of study findings to real-world settings. Positivity violations occur when selection bias, due to restrictive inclusion criteria or practical constraints, limits the representativeness of the study sample. This mismatch weakens external validity, making it difficult to generalize findings beyond the study population. In real-world applications, such violations can lead to unreliable or unidentifiable estimates. As evidence-based decisions increasingly rely on data, developing robust methods to account for these challenges is crucial. The first part of the dissertation addresses positivity violations for binary treatments by proposing a unified framework that categorizes the target population into unrepresented, underrepresented, and well-represented groups based on the overlap with the study sample. This framework identifies unrepresented subpopulations, where the average treatment effect (ATE) is unidentifiable, and underrepresented subpopulations, where estimation is inefficient. Inference for the well-represented group is conducted using a weighting estimator, while a sensitivity analysis assesses the impact of unrepresented and underrepresented groups on the overall population-level ATE. The second part develops a semiparametric framework for generalizing causal effects under multiple treatment arms, addressing limitations of existing methods that rely on parametric assumptions for binary treatments. This work introduces the transportability score, combining the propensity and sampling scores, to improve estimation under covariate shift. To address positivity violations, a smooth inclusion weight is introduced to restrict inference to the well-represented subpopulation. The framework incorporates flexible machine learning tools to enable valid and efficient inference, even under model misspecification, improving generalizability in complex real-world settings. Together, these methodologies form a comprehensive toolkit for enhancing causal inference under positivity violations. They improve the validity and applicability of study findings, with significant implications for evidence-based interventions that bridge the gap between research and practical impact.

Author dc:creator

• Jun Lu (36049)

• Causal Inference
• Generalizability and Transportability
• External Validity

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.30424993.v1

OAI identifier oai:identifier

oai:figshare.com:article/30424993