Using the Decision Trees to Improve Variable Selection for Building Composite Indicators

The established variable selection methods for building composite indicators have strong limitations with respect to the results obtained. Some of them focus on getting an index structure with a high alpha reliability and/or a high percentage of the total data variance explained. These methods are likely to omit variables with strong explanatory power, and lead to an unsatisfactory classification of countries. Decision trees can also be used in selecting variables that are the most relevant for building composite indicators. An example of variable selection for building a composite indicator, which compares results using Cronbach's coefficient alpha, factor analysis, and decision trees, shows that the latter method yields comparable, or better results. Using cluster analysis on the selected variables, we show that the decision tree variable shortlist has better discrimination power than those obtained with the other methods, even in the presence of outliers and missing values.