Bigger is Not Better: Seeking Parsimony in Canonical Correlation Analysis via Variable Deletion Strategies

Abstract

This paper illustrates the value of applying the law of parsimony to canonical correlation analysis (CCA) solutions. The primary purpose of parsimony is that the more parsimonious the solution, the more replicable the model will be. The ultimate goal is to estimate an equal or reasonable amount of variance with the smallest variable set possible. A real-world data set is used that is composed of 287 sixth-grade students who were administered a geometry content knowledge test with three levels and a spatial visualization test as criterion variables, and a mathematics attitude survey with six subscales as predictor variables. Three different deletion methods are delineated in the paper that will assist the researcher in deleting predictor or criterion variables to obtain a more parsimonious canonical solution.