The long-term goal of this research is to develop, implement, and apply innovative statistical methodology for the analysis of longitudinal studies involving clustered binary or counted data in cancer treatment, prevention, and health services research. Clustered longitudinal studies play an increasingly important role in cancer research. Current studies in our group address issues of how differences in local health system resources influence the delivery of care and how greater capacity and increased utilization affect health outcomes. The studies share key statistical features in that each involves the measurement of repeated binary or counted outcomes on individual cancer patients clustered within hospital referral regions (HRRs); and the primary focus is in describing overall variations in outcomes across HRRs and in determining the influence of HRR-level covariates, such as physician supply and practice patterns, on these variations. While statistical methodology is currently available to analyze such data, analyses are hampered by their technical and computational complexity. Specifically, we will extend our computationally simple, robust, efficient two-stage method of estimation previously used for modeling linear and nonlinear growth curves to large clustered longitudinal studies involving binary or counted data; develop regression diagnostic methods for the assessment of quality of fit and sensitivity to outliers in clustered longitudinal regression models; and apply the new approaches to the study of geographic variations in surveillance patterns following potentially curative resection for colorectal cancer among U.S. Medicare beneficiaries. These analyses will examine how surveillance strategies for colorectal cancer patients vary across geographic regions and how local availability of resources, the specialty mix of physicians in the region, and patient attributes interact to influence downstream decisions. The methods we propose will circumvent many of the above technical difficulties and would not require either special software or high performance computer workstations.
