A Restricted Latent Class Model with Polytomous Attributes and Respondent-Level Covariates

We present an exploratory restricted latent class model where response data is for a single time point, polytomous, and differing across items, and where latent classes reflect a multi-attribute state where each attribute is ordinal. Our model extends previous work to allow for correlation of the attributes through a multivariate probit specification and to allow for respondent-specific covariates. We demonstrate that the model recovers parameters well in a variety of realistic scenarios, and apply the model to the analysis of a particular dataset designed to diagnose depression. The application demonstrates the utility of the model in identifying the latent structure of depression beyond single-factor approaches which have been used in the past.

💡 Research Summary

This paper introduces an exploratory Restricted Latent Class Model (RLCM) designed for single‑time‑point, polytomous response data where each latent class is defined by a vector of ordinal attributes. The authors extend earlier RLCM work in three major ways: (1) they allow the attributes to be polytomous rather than binary, (2) they model correlation among attributes using a multivariate probit specification, and (3) they incorporate respondent‑level covariates that influence the latent attribute profile.

The measurement component is a cumulative probit model. For each item j with Mj categories, the probability that a respondent’s answer Ynj is at most m is linked to a linear predictor d n βj via the standard normal CDF, where d n is a “cumulative coding” design vector derived from the respondent’s latent attribute vector α n. This coding creates a binary expansion for each attribute level and then takes a tensor product across attributes, yielding a design matrix of dimension H that can be tuned by the analyst. A monotonicity condition is imposed: if one latent attribute profile dominates another component‑wise, the corresponding response probabilities must be non‑decreasing. This constraint translates into linear inequalities on the β‑coefficients, ensuring interpretability of the attribute‑response relationship.

The structural component treats the discrete attribute vector α n as a discretized version of a continuous latent vector α* n. The continuous vector follows a K‑dimensional multivariate normal distribution with mean X n λ (X n contains respondent covariates and an intercept) and correlation matrix R. Threshold vectors γ k partition each dimension of α* n into L ordered categories, producing the observed ordinal attributes. The authors place left‑truncated exponential priors on the thresholds to guarantee proper posterior sampling even when no respondent falls into the highest category.

A fully Bayesian framework is constructed. The β‑coefficients receive a spike‑and‑slab prior à la Kuo‑Mallick, combined with an indicator function that enforces the monotonicity constraints. The inclusion indicators δ j follow Bernoulli(ω) with a Beta hyperprior on ω, enabling automatic variable selection for item‑attribute effects. The regression matrix λ has a normal prior with covariance I D ⊗ R, preserving conjugacy with the multivariate probit. The correlation matrix R receives a prior analogous to the LKJ distribution, ensuring positive‑definiteness while remaining weakly informative.

Inference proceeds via data augmentation. For each observed categorical response Ynj, an auxiliary continuous latent variable Y* nj is introduced, yielding a normal likelihood conditional on d n βj. Similarly, α* n is augmented to facilitate Gibbs updates of λ and R. The authors employ parameter expansion techniques to improve mixing for the high‑dimensional β, δ, ω, γ, λ, and R blocks. The resulting MCMC algorithm cycles through: (i) sampling Y* given current β and thresholds, (ii) updating β and δ under monotonicity constraints, (iii) drawing γ thresholds from truncated exponential or uniform conditionals, (iv) sampling α* and α via the multivariate probit, and (v) updating λ and R using standard conjugate steps.

Two simulation studies evaluate performance under varying conditions: (a) strong versus weak attribute correlation, (b) presence versus absence of covariate effects, and (c) balanced versus highly imbalanced category frequencies. Across all scenarios the proposed model accurately recovers the true parameters, outperforms binary‑attribute RLCMs in classification accuracy, and yields posterior predictive checks that closely match the data‑generating process.

The methodology is applied to a real‑world depression‑diagnosis questionnaire. The instrument contains 12 items with 4–5 point Likert scales. Covariates include age, gender, and a socioeconomic index. Model selection using posterior predictive checks and WAIC identifies three latent attributes (cognitive, affective, somatic) each with four ordered levels. The fitted model uncovers three distinct latent subtypes of depression: (1) severe affective‑cognitive distress, (2) predominantly somatic symptom profile, and (3) mild or subclinical presentation. Covariate effects reveal that younger respondents are more likely to belong to the severe subtype, while lower socioeconomic status increases the probability of the somatic subtype. The estimated correlation matrix shows a strong positive association between cognitive and affective attributes in the severe group, but weaker links in the other groups, highlighting the model’s ability to capture nuanced multidimensional structures that single‑factor approaches miss.

In summary, the paper contributes: (i) a flexible RLCM for polytomous attributes, (ii) a multivariate probit latent structure that permits unrestricted attribute correlation, (iii) integration of respondent‑level covariates, (iv) a novel left‑truncated exponential prior for thresholds that ensures computational stability, (v) an efficient MCMC scheme with parameter expansion, and (vi) empirical validation through simulation and a substantive mental‑health application. Limitations include computational intensity for very large samples and sensitivity to prior specifications for the correlation matrix. Future work may explore variational inference, sparsity‑inducing priors for R, or extensions to longitudinal data.