Decomposing Variations on Cluster Level for Binary Outcomes in Application to Cancer Care Disparity Studies

To develop a method to decompose the observed variance of binary outcomes (proportions) aggregated by regional clusters to determine targets for quality improvement efforts to reduce regional variations.

Data from the 2018 linkage of the Surveillance, Epidemiology, and End Results-Medicare database.

We developed a method to decompose the observed regional-level variance into four attributions: random, patients' characteristics, regional cluster, and unexplained. To demonstrate the efficacy of the method, we conducted a series of numerical studies. We applied this method to our cohort to analyze endocrine therapy receipt 3-5 years after diagnosis, using health service area (HSA) as the regional cluster.

Our cohort included Stages I-III breast cancer patients diagnosed at ages 66-79 between 2007 and 2013 who received cancer surgery and were enrolled in Medicare Parts A and B.

After decomposition, 39% of the total variation was explained by HSAs, which was higher than that in some other breast cancer measures, such as the proportion of Stage I at diagnosis (4%), previously reported. This suggests geospatial efforts have a great potential to address the regional variation regarding this measure.

Our variance decomposition method provides direct information about attributable variance in the proportions at a cluster level. This technique can help in the identification of intervention targets to improve regional variations in the quality of care and clinical outcomes.