Evaluating underlying factor structures using novel machine learning algorithms: An empirical and simulation study

Abstract

The scale development paradigm was created to improve the measurement of latent constructs. Although several statistical techniques have been successfully integrated into the overall process, identifying factor patterns and validating constructs using smaller datasets with different correlational structures remain a concern. This paper presents heatmapping and bootstrapping cluster analysis (HMBCA), a novel machine-learning based diagnostic workflow, as a new tool to aid in strengthening the process. A substantive example on the overall organizational knowledge acquisition behaviors demonstrates that the bootstrapping cluster simulation approach provided promising results regarding the factor structure as measured by the Approximately Unbiased (AU) p-values under the following conditions: when factor correlations are weaker or moderate, with simulated data containing smaller samples. The study suggests that researchers may leverage bootstrapping cluster simulations to validate constructs through both visual inspection and probability estimates when faced with constraints such as a small sample size.