Modern advances in accounting for unmeasured confounding via sensitivity analysis Abhinandan Dalal

Format:

Book

Thesis/Dissertation

Author/Creator:

Dalal, Abhinandan, author.

Contributor:

University of Pennsylvania. Statistics and Data Science., degree granting institution.

Language:

English

Subjects (All):

Statistics.

0463.

0272.

Local Subjects:

Statistics.

0463.

0272.

Genre:

Academic theses

Physical Description:

1 online resource (215 pages)

Contained In:

Dissertations Abstracts International 87-12A

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2026

Language Note:

English

Summary:

The assumption of no latent confounding is a cornerstone of causal inference, yet it is often violated in observational studies, posing a significant challenge to the validity of causal conclusions. Sensitivity analysis provides a framework for assessing how causal claims may be affected in the presence of hidden confounders. By asking how strong the influence of an unmeasured confounder would need to be to negate a causal conclusion, sensitivity analysis offers a quantitative way to evaluate the credibility of conclusions drawn from observational data. This dissertation aims to advance the use of sensitivity analysis when drawing causal conclusions from non-experimental data by examining its implications for the design, execution, interpretation, and analysis of observational studies. First, we propose improvements to observational study design, emphasizing the selection of robust outcomes to mitigate the statistical burden associated with multiple comparisons. Specifically, we develop a procedure to evaluate an outcome's resilience to latent confounding, prioritizing those that remain stable even when the assumption of no unmeasured confounding is relaxed. Next, we extend the execution of sensitivity analysis methods to sequential settings, ensuring anytime-valid coverage guarantees even under continuous data monitoring. This allows researchers to stop data collection when inference is deemed adequate, providing a principled way to optimize resources when there are diminishing returns to collecting additional data. Then, we examine when and how the interpretation of sensitivity analyses can be sharpened using additional side information. When latent confounders cannot be directly adjusted for and are instead controlled using proxy variables, we show that strong associations between exposure and measured covariates can imply amplified sensitivity to residual confounding. These results highlight the importance of multicollinearity when interpreting sensitivity analyses based on proxy adjustment. Finally, we discuss the interpretation and analysis of sensitivity models for continuous exposures. We distinguish between global and targeted sensitivity analysis, highlighting their geometric interpretations and implications for compatibility with auxiliary information. We derive flexible, rate-robust estimators for the corresponding partial identification bounds and construct confidence intervals for these bounds, providing practical tools for sensitivity analysis with continuous treatments

Notes:

Source: Dissertations Abstracts International, Volume: 87-12, Section: A.

Advisors: Small, Dylan

Ph.D. University of Pennsylvania 2026

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798247973942

Access Restriction:

Restricted for use by site license