Quantile Regression Analysis of the Effect of Health Maintenance Organization Enrollment on Medical Expenditures
Jeroan J. Allison, University of Massachusetts Medical School 
Arlene S. Ash, University of Massachusetts Medical School 
*Lisa M. Lines, University of Massachusetts Medical School 

Keywords: quantile regression, HMO, health expenditures, MEPS

BACKGROUND: Since the early 1970’s, health maintenance organizations (HMOs) have been used with mixed results to constrain ever-increasing health expenditures. The effects of HMO enrollment may vary by level of patient expenditure, but this has not been previously investigated. Quantile regression (QR), which models the relationship between an exposure and conditional quantiles of an outcome given a set of covariates, may be useful. The Medical Expenditure Panel Survey (MEPS) is one of few publicly available sources of detailed population-based data on expenditures by HMO enrollees. METHODS: We used 2008 data from MEPS (all ages, n=31,262); the main outcome variable was total expenditures and the main exposure, HMO enrollment, regardless of payor. Covariates included sociodemographic and health status characteristics. All analyses were conducted in SAS 9.2 using weights, cluster, and strata variables to account for the complex survey design. Multivariable QR models of total expenditure at four different quantiles were fitted and compared to a multivariable OLS model. A sensitivity analysis removed uninsured respondents from the non-HMO cohort. RESULTS: In 2008, 23.4% of all respondents were in an HMO (27.0% of all insured). HMO enrollees were older, better educated, more likely to be white or Asian, married or widowed/divorced/separated, from the Northeast or West, and urban. HMO enrollees were less likely to have low English proficiency, be low income, and have no usual source of care. The cohorts did not differ in self-reported health status, activity limitation, or number of children with special healthcare needs. Mean total expenditure (±SE) was $3,665 ± $96 in the non-HMO cohort, vs. $4,127 ± $141 in the HMO cohort. The OLS model found that HMO enrollment reduced total expenditures by $543 (P=.001). The QR models found a monotonically increasing trend of greater savings with greater cost expenditures. At the 25th percentile, savings were $16; at the 50th, $40; at the 80th, $102; and at the 95th, $752. Results from the sensitivity analysis were similar. CONCLUSIONS: QR modeling in a nationally representative sample suggests that HMOs are cost-saving at all quantiles of total expenditure, with larger cost containment effects at the higher ends of the distribution. QR methods revealed that most of the cost savings identified by OLS for the average patient were driven by patients with higher costs.