A psychometric data science approach to study latent variables: a case of class quality and student satisfaction

Abstract

Data science follows a data-driven approach focused on non-hypothesised pattern discovery from data in an automatic manner (or semi-automatic). Social science areas frequently address latent variables, whose study is mainly guided by a deductive paradigm and requires a confirmatory scope. Both areas, data science and latent variable research, have ample development, but their combination is not mature. The objective is to develop a practical and reproducible framework for data science projects from a psychometric perspective for the study of latent variables. The framework is practical and integrative, and consists of nine steps, which guide from the construct definition, discovery and confirmed patterns (structural equation modelling and considering goodness-of-fit, reliability, validity, equity and subjects’ segmentation) to the use of patterns. This last step includes three sub-steps: one for structural hypotheses contrast and two computational approaches (supervised: machine learning; and non-supervised: principal component analysis). Additionally, the framework is efficient and accessible due to its semi-automation using the R language. The framework is applied to study the relationship between class quality and student satisfaction and is consistent with the reproducible research paradigm, widely used in computational areas and recently demanded by social science.