Finding the Appropriate Scale for a Continuous Covariate: a modeling issue for many linear models.

Larry Lessner, NYS Dept of Health

Finding the appropriate scale for a continuous covariate is a step in basic model building for normal linear regression, logistic regression, Poisson regression, the general linear model, and the Cox model for survival data. Finding the appropriate scale simply means finding a transformation of the continuous variable that significantly improves the model. For example polynomial terms are usually considered in normal regression for nonlinear alternatives. Such models can be strongly influenced by extreme values. Splines maybe appropriate but require computationally intensive modeling. The work of Royston and Altman (1994) suggests fractional polynomials as a simple alternative. The implementation of this technique in SAS will be discussed.