Beta-Binomial Regression and Bimodal Utilization

Objective To illustrate how the analysis of bimodal U‐shaped distributed utilization can be modeled with beta‐binomial regression, which is rarely used in health services research. Data Sources/Study Setting Veterans Affairs (VA) administrative data and Medicare claims in 2001–2004 for 11,123 Medicare‐eligible VA primary care users in 2000. Study Design We compared means and distributions of VA reliance (the proportion of all VA/Medicare primary care visits occurring in VA) predicted from beta‐binomial, binomial, and ordinary least‐squares (OLS) models. Principal Findings Beta‐binomial model fits the bimodal distribution of VA reliance better than binomial and OLS models due to the nondependence on normality and the greater flexibility in shape parameters. Conclusions Increased awareness of beta‐binomial regression may help analysts apply appropriate methods to outcomes with bimodal or U‐shaped distributions.