Statistical methods and strategies for working with large data bases.

Abstract

This article describes the statistical methods and strategies to be used in establishing the linkages between processes and structures of care with risk-adjusted outcomes in a large multicenter Veterans Affairs cooperative study in health services of patients undergoing cardiac surgery. The statistical analyses consist of test involving nine specific hypotheses related to the effect of processes and structures of care on risk-adjusted outcomes. From the statistical point of view, the major obstacles of this study are the need for data reduction and imputation of missing data. The former obstacle is addressed through the use of data-reduction techniques, such as principal components and cluster of variables. The latter is addressed through the use of classic and new techniques for imputation of missing data, such as MISSGEN, principal components for qualitative data, and the expectation and maximization algorithm. Data reduction and imputation of missing data are done with clinically derived variable groups called "dimensions" or "subdimensions." The effect of processes and structures of care is assessed by a two-step process. First, outcomes are modeled using only patient risk factors. The selection of risk factors in the modeling process is discussed in detail. Second, these risk-adjusted outcomes are modeled using one of the nine process or structure subhypotheses. The relationship of the processes and structures of care dimensions and/or subdimensions that are linked to risk-adjusted outcomes are identified.