Modern Prediction Algorithms for Factor Analyses: Comparing Accuracy-Efficiency Trade-Offs

Abstract

The overarching aim of this dissertation is to investigate the performance of new factor modeling techniques that are exploratory in nature and show promise for determining the number of factors, as well as variable-factor matching methods, in the absence of solid theory or in the presence of small samples relative to the number of variables to be included in the modeling process. Although these methods do show promise for data in which variables have true cross-loadings (i.e., are not unidimensional), thus far the prior research has not studied the performance of these algorithms with different levels of measurement reliability in conjunction with variables that have true loadings of zero. As such, I use simulation methods to investigate the utility of extreme gradient boosting for recovering the accurate number of factors across a range of conditions (as compared to other traditional factor retention criteria), as well as to investigate the accuracy of regSEM for recovering true variable loadings in addition to the accuracy of factor-variable matching. Last, I compare three popular R packages using specific “case studies” of datasets with poor vs. good measurement quality on ease of use, elapsed time, and loading accuracy to provide practical recommendations. Limitations and future research are discussed.