St. James Ballroom
Using Penalized Regression to Identify Subgroups of Clients Who Benefit from a Diabetes Intervention (303800)
Michele Heisler, University of Michigan
Edith C. Kieffer, University of Michigan
Barbara Mendez Campos, University of Michigan
Gretchen Piatt, University of Michigan
*Brandy R. Sinco, University of Michigan
Michael S. Spencer, University of Michigan
Keywords: penalized regression, variable selection
Background: T-tests and linear models indicated that clients aged 55+ in a diabetes program had a significant intervention effect on blood sugar change, while those under 55 did not. Age was categorized into 18-54 and 55+ because the linear model fit better with age dichotomized and statistical power to detect differences for older persons.
Objective and Methods: Use penalized regression with LASSO and Bayesian Information Criteria to check whether other demographic variables impact the intervention effect. Set the outcome as blood sugar drop from baseline to follow-up, with covariates of baseline value, treatment group, age group, gender, education, employment, and interactions between all demographic variables and treatment group. Confirm results with Akaike Information Criteria and elastic net.
Results. LASSO and elastic net indicated that information criteria were minimized with baseline value, treatment group, and the interaction between treatment group and age group in the model for blood sugar change. Results were consistent with AIC and BIC.
Conclusion. Penalized regression is a useful technique to confirm demographic sub-groups who benefit from a health intervention.