Combining Information from Multiple Sources in the Analysis of a Non-Equivalent Control Group Design

In studies of whether hospital or health-center interventions can improve screening rates for mammography and Pap smears in Los Angeles County, the availability of data from multiple sources makes it possible to combine information in an effort to improve the estimation of intervention effects. Primary sources of information, namely computerized databases that record screening outcomes and some covariates on a routine basis, are supplemented by medical chart reviews that provide additional, sometimes conflicting, assessments of screening outcomes along with additional covariates. Available data can be classified in a large contingency table where, because medical charts were not reviewed for all individuals, some cases can only be classified into a certain margin as opposed to a specific cell. This paper outlines a multiple imputation approach to facilitate data analysis using the framework of Schafer (1991, 1995), which involves drawing imputations from a multinomial distribution with cell probabilities estimated from a loglinear model fitted to the incomplete contingency table. Because of the sparseness of the contingency table, a cavalier choice of a convenient prior distribution can be problematic. The completed data are then analyzed using the method of propensity score sub classification (Rosenbaum and Rubin 1984) to reflect differences in the patient populations at different hospitals or health centers.