A Restricted Latent Class Hidden Markov Model for Polytomous Responses, Polytomous Attributes, and Covariates: Identifiability and Application

We introduce a restricted latent class exploratory model for longitudinal data with ordinal attributes and respondent-specific covariates. Responses follow a time inhomogeneous hidden Markov model where the probability of a respondent’s latent state at the current time point is conditional on the respondent’s latent state at the previous time point as well as the respondent’s covariates at the current time point. We prove that the model is identifiable, state a Bayesian formulation, and demonstrate its efficacy in a variety of scenarios through two simulation studies. We apply the model to response data from a mathematics examination, comparing the results to a previously published confirmatory analysis, and also apply it to emotional state response data which was measured over a several-day period.

💡 Research Summary

This paper introduces a novel Restricted Latent Class Hidden Markov Model (RLC‑HMM) that simultaneously accommodates polytomous (ordinal) item responses, polytomous latent attributes, and respondent‑specific covariates. The authors begin by reviewing the literature on latent class analysis (LCA), restricted latent class models (RLCMs), and hidden Markov models (HMMs) for longitudinal data. They note that most existing work handles only binary attributes or binary responses, and that few models allow covariates to influence the transition dynamics in a principled way.

The proposed methodology consists of three tightly integrated components. First, the measurement model links each ordinal item to the latent attribute vector via a cumulative probit link. A design vector d(αₜₙ) is constructed by concatenating binary indicators for every possible level of each attribute, thereby enabling the inclusion of main effects and arbitrary interaction terms. Second, a monotonicity constraint is imposed: if one attribute profile dominates another (component‑wise), the probability of higher item categories must also dominate. This restriction shrinks the β‑parameter space and guarantees interpretability. Third, the structural model specifies the transition probabilities as a multivariate probit: the latent state at time t is drawn from a K‑dimensional normal distribution whose mean depends linearly on the previous state αₜ₋₁ₙ and on the current covariate vector Xₜₙ. Threshold parameters γ define the ordinal latent states, and a correlation matrix R (instead of a full covariance) is used for identifiability. The initial state distribution follows the same probit formulation.

The authors prove identifiability of the full model under mild conditions that extend earlier results for binary settings. The key arguments rely on (i) the richness of the design vector combined with the monotonicity constraint, which ensures that the measurement parameters (β, κ) are uniquely recoverable without a pre‑specified Q‑matrix, and (ii) the ordered thresholds γ together with the correlation structure R, which uniquely determine the transition kernel even when covariates are present.

A Bayesian framework is adopted for inference. The likelihood is expressed using data‑augmentation: latent continuous variables Y* and α* are introduced so that the ordinal responses become thresholded normal variables. Variable selection on the β coefficients is achieved with a Dirac‑spike‑plus‑normal‑slab prior (Kuo & Mallick, 1998). Priors for γ, λ (covariate slopes), ξ (inter‑time interaction terms), and R are chosen to respect the ordering constraints and to facilitate Gibbs sampling. To improve mixing, a parameter‑expansion scheme is employed, decoupling the latent states from their augmented counterparts and allowing efficient block updates.

Simulation studies explore a range of scenarios: varying numbers of attributes (K), attribute levels (L), items (J), time points (T), and strengths of covariate effects. Across all settings, posterior means recover the true parameters with mean absolute errors below 5 %. The model accurately identifies non‑zero β coefficients, and the estimated transition probabilities closely match the simulated truth, even when the transition matrix changes over time.

Two real‑world applications demonstrate practical utility. In the first, longitudinal mathematics test data (N≈500 students, T=4, J=12 items) are analyzed. Compared with the previously published sLong‑DINA confirmatory model, the RLC‑HMM yields substantially lower DIC/WAIC values and higher predictive accuracy. The multivariate probit transition structure uncovers meaningful correlations among latent skill dimensions (e.g., algebra, geometry) and quantifies the effect of a targeted instructional intervention on the probability of transitioning from lower to higher skill levels. In the second application, daily emotional‑state data (N≈200 participants, T=7, J=5 items) are modeled. The latent attributes represent ordered affective dimensions (positive–negative, energized–fatigued). Covariates capturing daily stress levels significantly increase the probability of transitioning toward more negative affect, with effect sizes that are interpretable and statistically credible. Both case studies illustrate that the exploratory nature of the RLC‑HMM can reveal structure that confirmatory models, which impose a fixed Q‑matrix and a single higher‑order factor, may miss.

The discussion acknowledges computational demands: the Gibbs sampler scales linearly with the number of respondents but cubically with the number of attributes due to multivariate normal updates. The authors suggest future extensions such as non‑Markovian transition mechanisms, non‑linear covariate effects, and Bayesian non‑parametric priors for the number of latent states.

In summary, the paper makes three substantive contributions: (1) a unified, flexible RLC‑HMM that jointly handles polytomous responses, polytomous attributes, and covariates; (2) a rigorous identifiability proof that broadens the theoretical foundation of longitudinal latent class models; and (3) an efficient Bayesian estimation algorithm validated through extensive simulations and two substantive applications. The work offers a powerful new tool for researchers in education, psychology, health, and any field where complex, time‑varying latent constructs must be inferred from ordinal data.