Addressing Covariate Missingness in Generalization of Randomized Controlled Trials: A Comparison of Methods Jimin Oh

Format:

Book

Thesis/Dissertation

Author/Creator:

Oh, Jimin, author.

Contributor:

University of Pennsylvania. Education., degree granting institution.

Language:

English

Subjects (All):

0443.

0458.

0463.

0515.

Local Subjects:

0443.

0458.

0463.

0515.

Physical Description:

1 electronic resource (124 pages)

Contained In:

Dissertations Abstracts International 87-07B

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2025

Language Note:

English

Summary:

Randomized controlled trials (RCTs) are increasingly used to estimate causal effects in education, yet their findings are often difficult to generalize when key moderators of treatment effects are unavailable in population data. This dissertation studies how to handle missing moderator covariates when estimating the population average treatment effect (PATE). I compare two primary strategies - (i) proxy augmentation of population frames with variables correlated to latent moderators and (ii) imputation of missing moderators-together with two supplementary tools: (iii) BART with missingness incorporated attributes (BART-MIA) and (iv) separating set covariate selection via Markov random fields. Results show that simple methods can perform competitively under realistic data constraints. Proxy quality is more consequential than quantity: a few strongly correlated proxies reduce bias and variance, whereas adding many weak proxies inflates error and degrades overlap. Multiple imputation improves precision when missing‐at‐random is plausible but is not uniformly superior to complete‐case analysis. Flexible outcome modeling (TMLE, BART) performs best when moderators are informative and alignment between sample and population is moderate, while weighting estimators remain more robust under limited support. The SimCalc analysis corroborates these patterns, yielding similar PATE estimates across strategies but highlighting trade‐offs between stability and bias reduction. The findings suggest that, when population moderators are unavailable, pragmatic combinations of strong proxies, moderate imputation, and diagnostic checks can yield reliable generalization without resorting to highly complex estimators

Notes:

Advisors: Chan, Wendy Committee members: Boruch, Robert F.; Maynard, Rebecca A.

Source: Dissertations Abstracts International, Volume: 87-07, Section: B.

Ph.D. University of Pennsylvania 2025

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798276005041

Access Restriction:

Restricted for use by site license