Douglas Simpson, University of Illinois

Some Developments in Ordinal Regression Analysis with Applications to Correlated Data

In social science research, questionnaires often use Likert scale items. In medical research, a researcher may grade tumors on an ordinal scale, or a patient may rate level of discomfort on a 4 point scale. In toxicology, severity of toxicological effect may be scored as 0="no effect", 1="mild effect", 2="severe effect". This talk will discuss ordinal regression analysis of data from such studies. Ordinal regression analysis is used to model the conditional probability distribution of the ordinal score as a function of covariates such as exposure, social status, treatment, and so on. A general class of models will be introduced along with unified approach to model fitting. Then two approaches to the analysis of correlated ordinal data will be discussed: mixed effect modeling using latent variables, and marginal modeling, which is derived from the theory of generalized estimating equations.