East Coast Ballroom
Design and analysis considerations for adjusted comparative quality surveys (307891)*Alan M. Zaslavsky, Harvard Medical School
Keywords: quality, weighting, regression adjustment
Survey measures of health care quality are often reported comparatively, using ranks, or quantile ranges, or comparisons to specific reference groups, and "controlling" for the effects of characteristics not due to the units' activities. These inferences simulate potential results if cohorts of patients with identical sample covariate distributions had been treated at each of the units. Sample design and analysis for these surveys with objectives of these types differ from studies that assess each unit separately.
Where controlled comparisons and descriptions of units are important, we compare weighting estimators (nonparametric direct standardization) and regression adjustments. The weighting estimators are valid under relatively weak assumptions and clearly relate to the hypothetical replicated assessment cohorts, but incur a readily estimated variance penalty with unbalanced samples, and may be incalculable when the samples do not have common support in covariate space. Regression can implement more extreme adjust adjustments, but requires stronger statistical assumptions. We calculate and compare the additional variance corresponding to violations of these assumptions.