Pharmaceutical Statistics Using SAS

3.4. Variable Selection

The data used to developed computational models can often be characterized by a few observations and a large number of measured variables, some of which are highly correlated. Traditional approaches to modeling these data are principle component regression (PCR) and partial least squares (PLS) (Frank and Friedman, 1993). The factors obtained from PCR and PLS are usually not interpretable so if the goal is to develop an interpretable model, these are not as useful. By considering the loadings given to each variable, these methods can be used to prune the variables, though the criteria for removing variables can be inconsistent between data sets.

Other popular variable selection procedures are tied to model selection and ...