Composite indices based on partial least squares

Abstract

In this paper, we compare Principal Component Analysis (PCA) and Partial Least Squares (PLS) methods to generate weights for composite indices. In this context we also consider various treatments of non-metric variables when constructing such composite indices. Using simulation studies we find that dummy coding for non-metric variables yields satisfactory performance compared to more sophisticated statistical procedures. In our applications we illustrate how PLS can generate weights that differ substantially from those obtained with PCA, increasing the composite indices' predictive performance for the outcome variable considered.