DOI
https://doi.org/10.25772/XE00-1894
Defense Date
2008
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Education
First Advisor
James H. McMillan
Abstract
Purpose: Disease management (DM) programs are typically evaluated using study designs that are susceptible to selection bias and other internal validity threats because participants are often allowed to self-select into the programs. As a result, DM evaluation results are usually biased because researchers are unable to control for preexisting differences between the DM participants and non-participants. Linden and Adams (2006) offer an instrumental variables (IV) regression procedure as a means of deriving unbiased estimates of DM program effectiveness. However, IV regression relies upon the existence of one or more variables (or instruments) that produce considerable variation in the program participation variable, but have no direct effect on the outcome variable. Linden and Adams argue that participant three-digit zip codes meet these criteria and can be used as instruments in IV regression.
Methods: To test the feasibility of their IV regression procedure, a series of ordinary least squares (OLS) and instrumental variables (IV) regression models were used to evaluate the effects of a high intensity Medicaid diabetes DM program on annual diabetes-related costs, emergency department visits, and hospital days. Program participation was the endogenous variable and age, gender, and propensity scores were the exogenous variable. Standard statistical tests were performed to assess the quality and validity of the IV regression models and zip code instruments.
Results: The study found that using propensity scores as covariates in the regression models appeared to offer a viable means of controlling for potential overt biases. However, the statistical tests performed to assess the quality and validity of the IV regression procedure using recipient three-digit zip codes as instruments indicated that it may not be appropriate due to various issues such as multicollinearity, lack of significant differences between the IV and OLS regression models, and weak instrument bias.
Conclusions: While the results of the present study do not support the use of participant three-digit zip codes as instruments in IV regression, the quality of the results obtained using this procedure may depend on the specific sample that is used in the analysis. Researchers may thus still wish to consider this procedure when evaluating DM programs because different samples may yield different results.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
9-13-2016