Coding of Categorical and Continuous Predictors

Abstract

When developing a prediction model, an important consideration is how we code the predictors. Raw data from a study are often not in a form appropriate for entering in regression models and must first be inspected and managed before the statistical analysis starts. As in any data analysis, we will usually start with obtaining an impression of the data under study, such as the occurrence of missing values and the distribution of predictors and outcome. Descriptive analyses, such as frequency tables and graphical displays, are useful to this aim. We will consider various issues in coding of unordered and ordered categorical predictors. For continuous predictors, we specifically discuss how we can limit the influence of outliers and interpret regression coefficients.