Abstract
The aim of causal effect estimation is to find the true impact of a treatment or exposure. Observational data is employed in social sciences to estimate causal effect but is susceptible to self-selection and unobserved confounding biases. Covariates included in analysis should strive to address these biases. This research focuses on investigating covariate selection approaches––common cause criterion (CC), Disjunctive Cause Criterion (DCC), Modified Disjunctive Cause Criterion (MDCC), and modified cause criterion (MCC)––in linear regression (LR) and propensity score methods (PSM) causal effect estimation in the presence of unmeasured confounding. Realistic social science scenarios such as––inclusion of proxy variables with varying degrees of strength, misidentification of the unmeasured covariate as a confounder, small sample sizes, and measurement error in proxy covariates—were investigated. For LR and PSM, five causal effect estimation models were built using different covariate selection approaches and compared on three performance metrics––bias, coverage, and empirical SE. Results showed that in the presence of an unmeasured confounder, the causal effect estimate is biased. Study 1 results indicate that MDCC approach resulted in more consistent and efficient causal effect estimates in the presence of unmeasured confounders. Studies 2a and 2b indicate that the MDCC approach is robust to the unobserved variable being a confounder and can be employed even if the unmeasured covariate is not a confounder without adversely impacting the performance measures. Studies 3 and 4 showed including a proxy of the unmeasured confounder, even a weak proxy (r ~ 0.20) or one with measurement error, results in an improvement in the consistency of the causal effect estimate and in the efficiency of the causal effect estimator. As the correlation between the proxy covariate and the unmeasured confounder gets smaller the causal effect estimator becomes less efficient and the causal effect estimate becomes less consistent.
Notes
If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu
Graduation Date
2022
Semester
Summer
Advisor
Hahs-Vaughn, Debbie
Degree
Doctor of Philosophy (Ph.D.)
College
College of Community Innovation and Education
Department
Learning Sciences and Educational Research
Degree Program
Education; Methodology, Measurement and Analysis
Identifier
CFE0009235; DP0026839
URL
https://purls.library.ucf.edu/go/DP0026839
Language
English
Release Date
August 2022
Length of Campus-only Access
None
Access Status
Doctoral Dissertation (Open Access)
STARS Citation
Nair, Uday, "Investigating Covariate Selection Criteria: To Draw Causal Inferences from Observational Data in the Presence of Unmeasured Covariates Using Regression and Propensity Score Methods" (2022). Electronic Theses and Dissertations, 2020-2023. 1264.
https://stars.library.ucf.edu/etd2020/1264