Missing Data in Regression Models for Non-Commensurate Multiple Outcomes

Abstract

Biomedical research often involves the measurement of multiple outcomes in different scales (continuous, binary and ordinal). A common approach for the analysis of such data is to ignore the potential correlation among the outcomes and model each outcome separately. This can lead not only to loss of efficiency but also to biased estimates in the presence of missing data. We address the problem of missing data in the context of multiple non-commensurate outcomes. The consequences of missing data when using likelihood and quasi-likelihood methods are described, and an extension of these methods to the situation of missing observations in the outcomes is proposed. Two real data examples illustrate the methodology.