Using Latent Class Analysis to Model Preference Heterogeneity in Health: A Systematic Review.

BACKGROUND: Latent class analysis (LCA) has been increasingly used to explore preference heterogeneity, but the literature has not been systematically explored and hence best practices are not understood.

OBJECTIVE: We sought to document all applications of LCA in the stated-preference literature in health and to inform future studies by identifying current norms in published applications.

METHODS: We conducted a systematic review of the MEDLINE, EMBASE, EconLit, Web of Science, and PsycINFO databases. We included stated-preference studies that used LCA to explore preference heterogeneity in healthcare or public health. Two co-authors independently evaluated titles, abstracts, and full-text articles. Abstracted key outcomes included segmentation methods, preference elicitation methods, number of attributes and levels, sample size, model selection criteria, number of classes reported, and hypotheses tests. Study data quality and validity were assessed with the Purpose, Respondents, Explanation, Findings, and Significance (PREFS) quality checklist.

RESULTS: We identified 2560 titles, 99 of which met the inclusion criteria for the review. Two-thirds of the studies focused on the preferences of patients and the general population. In total, 80% of the studies used discrete choice experiments. Studies used between three and 20 attributes, most commonly four to six. Sample size in LCAs ranged from 47 to 2068, with one-third between 100 and 300. Over 90% of the studies used latent class logit models for segmentation. Bayesian information criterion (BIC), Akaike information criterion (AIC), and log-likelihood (LL) were commonly used for model selection, and class size and interpretability were also considered in some studies. About 80% of studies reported two to three classes. The number of classes reported was not correlated with any study characteristics or study population characteristics (p > 0.05). Only 30% of the studies reported using statistical tests to detect significant variations in preferences between classes. Less than half of the studies reported that individual characteristics were included in the segmentation models, and 30% reported that post-estimation analyses were conducted to examine class characteristics. While a higher percentage of studies discussed clinical implications of the segmentation results, an increasing number of studies proposed policy recommendations based on segmentation results since 2010.

CONCLUSIONS: LCA is increasingly used to study preference heterogeneity in health and support decision-making. However, there is little consensus on best practices as its application in health is relatively new. With an increasing demand to study preference heterogeneity, guidance is needed to improve the quality of applications of segmentation methods in health to support policy development and clinical practice.