No abstract

In statistical modeling, it is crucial to have consistent variables that are the most relevant to the outcome variable(s) of interest in the model. With the increasing richness of data from multiple sources, the size of the pool of potential variables is escalating. Some variables, however, could provide redundant information, add noise to the estimation, or waste the degrees of freedom in the model. Therefore, variable selection is needed as a parsimonious process that aims to identify a minimal set of covariates for maximum predictive power. This study illustrated the variable selection methods considered and used in the small area estimation (SAE) modeling of measures related to the proficiency of adult competency that were constructed using survey data collected in the first cycle of the PIAAC. The developed variable selection process consisted of two phases: phase 1 identified a small set of variables that were consistently highly correlated with the outcomes through methods such as correlation matrix and multivariate LASSO analysis; phase 2 utilized a k-fold cross-validation process to select a final set of variables to be used in the final SAE models.