Random Coefficients Models for Subgroup Differences in Surveys of Health Care Quality
Keywords: CAHPS , healthcare quality surveys, random coefficients , Medicare
In healthcare quality surveys, we use linear models to adjust for effects on measures of patient characteristics not under the control of the healthcare unit (health plan, etc.). Differences in coefficients of casemix variables across units suggest that units treat patients with distinct characteristics differently in ways that are to some extent controlled by the unit. We fitted random-coefficients models to data from CAHPS(R) surveys of elderly beneficiaries in over 200 Medicare health plans. The intercept and coefficients of 3 patient characteristics (age, health status, and education) were modeled for 6 quality measures. In a Fay-Herriott-type model, level 1 used design-based estimates of sampling covariances. The correlation matrix of the random coefficients was nearly separable, the Kronecker product of matrices for associations across predictors and across outcome measures. We determined which coefficients varied most substantially and their relationships to selective enrollment. 3-level models assessed regional variation and intertemporal stability of quality patterns.