NOVEL EXTENSION AND APPLICATION OF REGRESSION MODELS FOR CONTINUOUS AND ORDINAL OUTCOMES

Abstract

The linear regression model is widely used to assess exposure effects on continuous and ordinal outcomes, where the observed ordinal scores are assumed to be good proxies for the underlying continuous quantities of interest. This fully parametric approach is susceptible to model misspecifications, and in studies of ordinal outcomes concerns has been raised regarding the appropriateness of the proxy assumption. This thesis proposes a seamless analytical workflow that facilitates robust inference on continuous and ordinal outcomes by applying and extending existing methods that model the ordering of outcomes. Our proposed workflow assesses the presence and direction of exposure effects using a semiparametric approach based on the ordered outcomes that requires less restrictive assumptions. When model assumptions (e.g., normality and linearity, and the proxy assumption for ordinal outcomes) are adequate, our proposed workflow provides a similar measure of association as the linear regression model that is well-received in health-related research.