Examining Analytic Practices in Latent Dirichlet Allocation Within Psychological Science: Scoping Review

Topic modeling approaches allow researchers to analyze and represent written texts. One of the commonly used approaches in psychology is latent Dirichlet allocation (LDA), which is used for rapidly synthesizing patterns of text within "big data," but outputs can be sensitive to decisions made during the analytic pipeline and may not be suitable for certain scenarios such as short texts, and we highlight resources for alternative approaches. This review focuses on the complex analytical practices specific to LDA, which existing practical guides for training LDA models have not addressed.

This scoping review used key analytical steps (data selection, data preprocessing, and data analysis) as a framework to understand the methodological approaches being used in psychology research using LDA.

A total of 4 psychology and health databases were searched. Studies were included if they used LDA to analyze written words and focused on a psychological construct or issue. The data charting processes were constructed and employed based on common data selection, preprocessing, and data analysis steps.

A total of 68 studies were included. These studies explored a range of research areas and mostly sourced their data from social media platforms. Although some studies reported on preprocessing and data analysis steps taken, most studies did not provide sufficient detail for reproducibility. Furthermore, the debate surrounding the necessity of certain preprocessing and data analysis steps is revealed.

Our findings highlight the growing use of LDA in psychological science. However, there is a need to improve analytical reporting standards and identify comprehensive and evidence-based best practice recommendations. To work toward this, we developed an LDA Preferred Reporting Checklist that will allow for consistent documentation of LDA analytic decisions and reproducible research outcomes.