A Systematic Review of Latent Class Analysis in Social Work Research

Purpose:Latent class analysis (LCA) and mixture modeling are methods to estimate latent, heterogeneous groups within a population. These person-centered statistical methods use a set of observed indicators to estimate the number and nature of unobserved groups of individuals within a sample. LCA models provide a valuable tool for social work researchers to examine the possibility of diverse samples within a population each with unique characteristics. The purpose of the current systematic review is to examine the use and quality of LCA and related mixture modeling studies within social work literature.

Methods:Using the Social Work Abstracts, PsycArticles, PsycInfo, MEDLINE, CINAHL Complete, Criminal Justice Abstracts, ERIC, and Health Source databases, a search of 84 social work journals with the terms (“latent” AND “class or profile or cluster or structure”) was conducted to identify published LCA and mixed modeling studies. The search yielded 449 articles that were uploaded into Covidence, a web-based systematic review software, for screening and review. Three reviewers independently assessed each study for use of LCA or mixture model in social work journals. A total of 402 studies were removed during title and abstract screening. Full-text review excluded an additional 16 studies, to arrive at a final sample of 31 studies that meet inclusion criteria.

Results: The 31 studies were published between 2004 and 2016 with a majority published after 2012 (n=18, 58.0%). Social work academic faculty or researchers were listed as first or second authors on a majority of studies (n=27, 87.1%). The Journal of Social Service Research published the most LCA/mixed-model studies with 3 (9.7%). Mplus (n=19, 61.3%) was the most commonly used statistical software package. Samples sizes varied greatly, ranging from 199 to 1,002,122 (median=532). LCA was used in most studies (n=17, 54.8%) followed by latent profile analysis (LPA; n=6, 19.4%), and latent class growth analysis (n=4, 13.0%). Studies used an average of 3.13 tests of model fit (SD=1.25), the most common of which were Bayesian information criterion (n=28, 90.3%) and entropy scores (n=18, 58.1%). Of the studies presenting entropy scores (ability of the model to separate individuals distinctly into classes based on the probability of membership for each class), the average entropy scores were .908 of a possible 1.0 (SD=.061; scores above .80 considered acceptable model fit).

Conclusions: This systematic review demonstrates the usefulness and growing popularity of LCA and related statistical procedures in social work research. Important findings from the review include the number of different types of LCA models used, their increased usage in social work research in the last four years, and the variation in tests of model fit. Social work studies used both LCA and LPA with cross-sectional data and latent class growth modeling, an extension of LCA able to assess heterogeneity in change within longitudinal data. The multiple tests of model fit across studies demonstrated high quality of the LCA modeling. As familiarity with these methods grow, social work researchers are encouraged to employ these person-centered statistical methods to explore possible heterogeneity within cross-sectional and longitudinal data.