Retrospective Hierarchical Modeling of Multiple Case-Control Studies

• Activity Number: 225
• Type: Invited
• Date/Time: Tuesday, August 5, 2003 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Stat. Sciences
• Author(s): Peter Mueller*+
• Companies: University of Texas M.D. Anderson Cancer Center
• Keywords: hierarchical models ; mixture modeling ; probit regression

Abstract

Motivated by the comprehensive risk predictions needed in medical decision modeling and patient counseling, we propose an approach for analyzing case-control studies of risk factors. Our approach extends traditional methods in several directions, including: (i) Combination of several related case-control studies; (ii) Combination of case-control and prospective studies; (iii) Full posterior and predictive inference in case-control studies including both categorical and quantitative risk factors; and (iv) Use of a flexible nonlinear regression approach, replacing the linear log odds ratio traditionally used. The main methodological tools used in achieving these extensions are: Direct retrospective modeling, rather than derivation of the retrospective likelihood implied by an assumed prospective model; Flexible mixture modeling for the retrospective model, giving a general nonlinear regression for the prospective likelihood; Hierarchical modeling, allowing for heterogeneity in the different data sources. We present simple results that highlight the relationship with existing approaches.