Abstract:
|
The proportion of cancer patients who receive guideline-concordant care varies by hospital or geographic region in the United States. Disparity studies often seek to identify the source of variation for each quality measure to guide quality improvement efforts. As quality measures are binary outcomes, mixed-effects logistic regression models are often used with hospitals (or regions) included as random-effects. Intra-class correlation (ICC) will then be calculated to quantify the degree of similarity of patients in the same hospital. However, the ICC does not describe how much of the total variance on cluster-level is explained by the random-effects. Characterizing cluster-level variance is important when the goal is to determine whether quality improvement efforts should target specific regions or hospitals. In this study, we developed a method to decompose the total variation on cluster-level into four attributions: 1) random variation, 2) variation in fixed-effects, 3) variation in random-effects, and 4) unexplained; and we applied this method to data from stage I-III breast cancer patients identified in the Surveillance, Epidemiology, and End Results-Medicare database.
|