ORCID

https://orcid.org/0000-0003-0939-8062

Date of Award

Fall 2026

Language

English

Embargo Period

10-20-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

College/School/Department

College of Integrated Health Sciences

Program

Biostatistics

First Advisor

Eric Rose

Committee Members

Melissa Tracy, Edward Valachovic, Hazem Al-Mofleh

Keywords

Dynamic Treatment Regimes, Q-learning, Missing Data, Multiple imputation, Sequential multiple assignment randomized trails, Opioid use disorder

Subject Categories

Biostatistics

Abstract

Opioid use disorder (OUD) remains a major public health challenge, and identifying treatment strategies tailored to individual patients over time is an important methodological and clinical problem. Dynamic treatment regimes (DTRs) provide a framework for such individualized strategies, but conventional estimation methods are often presented under complete-data assumptions. Missingness, especially dropout in multistage studies, can introduce bias and affect estimated treatment rules. Comparatively little work has examined how missing data should be handled in sequential multiple assignment randomized trials (SMARTs), where the timing and ordering of variables are central to estimation.

This dissertation examines how different missing-data approaches affect DTR estimation in a two-stage setting, combining a simulation study and an empirical application based on the Prescription Opioid Addiction Treatment Study (POATS). The methods compared are complete case analysis (CCA), standard multiple imputation (MI), a time-ordered sequential imputation approach motivated by Shortreed and colleagues (Shortreed MI), and a hybrid random forest multiple imputation method (RF-MI), all within a common Q-learning framework.

The simulation study evaluated three parameter settings, four missingness scenarios, and multiple levels of missingness using coefficient bias, mean squared error, Monte Carlo variability, stage-specific treatment-rule agreement, regime value, and regret. Under the primary missing-at-random (MAR) scenario, Shortreed MI was most consistently closest to the full-data benchmark for Stage 1 coefficient recovery and treatment-rule agreement. CCA showed the largest degradation as missingness increased, while standard MI and RFMI generally showed intermediate performance, with no consistent overall advantage for RF-MI. The empirical application identified substantial phase-specific dropout, Phase 2 non-entry, design-based structural absence, and additional incompleteness among participants who entered Phase 2. In the empirical Q-learning analysis, the imputation-based methods retained the full Stage 1 sample and the full Phase 2 analysis sample and produced less negative estimated value summaries than CCA. Differences in fitted treatment recommendations were larger at Stage 1 than at Stage 2. Adaptive m-out-of-n bootstrap confidence intervals were used to summarize uncertainty in the estimated Q-learning values in the presence of potential nonregularity.

This dissertation demonstrates that missing-data handling is an important component of DTR estimation rather than a separate preprocessing step. Under the primary MAR scenario examined, the time-ordered imputation approach performed most favorably among the evaluated missing-data methods for several Stage 1 estimation and decision-rule measures. More broadly, performance depended on the missingness mechanism, treatment effect structure, and the degree to which the imputation procedure respected the sequential structure of the treatment process.

License

This work is licensed under the University at Albany Standard Author Agreement.

Download

Included in

Biostatistics Commons

Share

COinS