Goodness-of-Fit Tests for Latent Class Models with Ordinal Categorical Data

Abstract: Ordinal categorical data are widely collected in psychology, education, and other social sciences, appearing commonly in questionnaires, assessments, and surveys. Latent class models provide a flexible framework for uncovering unobserved heterogeneity by grouping individuals into homogeneous classes based on their response patterns. A fundamental challenge in applying these models is determining the number of latent classes, which is unknown and must be inferred from data. In this paper, we propose one test statistic for this problem. The test statistic centers the largest singular value of a normalized residual matrix by a simple sample-size adjustment. Under the null hypothesis that the candidate number of latent classes is correct, its upper bound converges to zero in probability. Under an under-fitted alternative, the statistic itself exceeds a fixed positive constant with probability approaching one. This sharp dichotomous behavior of the test statistic yields two sequential testing algorithms that consistently estimate the true number of latent classes. Extensive experimental studies confirm the theoretical findings and demonstrate their accuracy and reliability in determining the number of latent classes.

• The paper introduces a spectral test statistic that distinguishes correct and underfitted models by analyzing the largest singular value of a normalized residual matrix.
• It presents two sequential algorithms—GoF-LCM and RGoF-LCM—that consistently recover the true number of latent classes under mild growth conditions.
• Empirical studies confirm the method's robustness, outperforming traditional EM-based and naive spectral approaches even in high-dimensional, weak signal settings.

Goodness-of-Fit Tests for Latent Class Models with Ordinal Categorical Data

Introduction and Motivation

Latent Class Models (LCMs) for ordinal categorical data are essential in disciplines where discrete, naturally ordered response data are analyzed, such as psychometrics, education, and social sciences. LCMs discern latent population heterogeneity by categorizing individuals into homogeneous latent classes based on their response vectors. A core statistical challenge is consistently and efficiently determining the true number of latent classes ($K$), as model misspecification introduces significant inferential risks: overestimating $K$ leads to overfitting and spurious structure, while underestimating masks heterogeneity.

Classic approaches for identifying $K$—AIC, BIC, penalized likelihood, and likelihood ratio tests—are heavily reliant on the EM algorithm, raising serious computational and local optimum concerns, particularly as the dimensionality grows. Moreover, in high-dimensional regimes ($J$ large, comparable to $N$), none of these methods has strong theoretical guarantees; their large-sample consistency results often depend on regularity conditions that can fail for LCMs and other mixture models.

Spectral methods, based on analyzing the low-rank structure of the data matrix, have proven computationally efficient for binary response LCMs, but do not accommodate ordered categorical (polytomous/ordinal) settings, nor do they provide formal hypothesis tests for the model's adequacy.

Main Contributions

This work provides a comprehensive theoretical and empirical investigation into spectral goodness-of-fit (GoF) testing for LCMs with ordinal categorical responses, specifically targeting consistent estimation of $K$. The central innovation is a test statistic based on the largest singular value of a normalized residual matrix (corrected for sample size), which has sharply dichotomous asymptotic properties under the null and underfitted alternatives. The paper advances the field in several technically significant ways:

1. Test Statistic Construction and Theory: A practical test statistic, $T_{K_0}$, is constructed by centering the largest singular value of the normalized residual matrix by $1 + \sqrt{J/N}$. Under $H_0: K = K_0$ (correct model fit), the statistic’s upper bound converges to zero in probability. Under any underfitted alternative ($K_0 < K$), the statistic with high probability exceeds a fixed positive constant. This behavior is established via matrix concentration inequalities and perturbation bounds governing error propagation from parameter estimation.
2. Sequential Testing Algorithms:

Two consistent, computationally efficient sequential strategies, GoF-LCM and ratio-based RGoF-LCM, are yielded by the statistic’s dichotomy: - GoF-LCM sequentially increases $K_0$, stopping when $T_{K_0}$ falls below a vanishing threshold. - RGoF-LCM stops at the $K_0$ where the ratio $|T_{K_0-1}/T_{K_0}|$ explodes, leveraging the strictly localized extremum in the ratio sequence at the correct $K$.

1. Estimation Consistency: Both sequential procedures are proven to strongly consistently recover the true $K$ under mild, practically verifiable constraints on class balance, separation, and growth conditions for $N, J, K$.
2. Algorithmic Methodology: SC-LCM, a straightforward spectral clustering method using the leading left singular vectors of $R$ for $k$-means partitioning, is justified as a computational primitive for class assignment. Its consistency is proved for growing $N,J,K$.
3. Empirical Verification: Thorough simulation studies account for strong, moderate, and weak signal regimes, and also investigate large $J$, showing that the proposed approach significantly outperforms spectral thresholding and regularized likelihood competitors, even when theoretical regularity (e.g., $J = o(N)$) is violated.

Theoretical Results: Test Statistic Properties

The normalized residual matrix is designed to mimic the behavior of a noise matrix under $H_0$. Under mild boundedness and separation, strong matrix concentration yields that the spectral norm of the normalized residuals obeys a Tracy–Widom-type upper bound, specifically,

$\|R^*\| \leq 1 + \sqrt{J/N} + o_P(1)$

under the null. When $K_0 < K$, the normalized residuals necessarily retain deterministic signal due to underfitting, ensuring the statistic exceeds a positive threshold with probability tending to one. This spectral dichotomy facilitates sequential testing with asymptotically vanishing type I and II errors.

Critically, for growing $N, J$, the separation condition (a fraction $c_1$ of items sharply distinguish any two classes) ensures the residual signal remains detectable under underfitting, and class size lower bounds scale estimation error control.

Algorithmic Procedures and Consistency

Two estimation algorithms are proposed:

• GoF-LCM: Increments $K_0$; stopping rule is $T_{K_0} < \tau_N$ (with $\tau_N \to 0$ at a specified rate). Consistency is proved by showing that when $K_0 = K$ the statistic drops below threshold, and for all $K_0 < K$ the statistic remains far above threshold with high probability.
• RGoF-LCM: Sequentially examines the ratios $r_{K_0} = |T_{K_0-1}/T_{K_0}|$; the estimate is the minimal $K_0$ where $r_{K_0}$ exceeds a diverging sequence.

Both algorithms are shown to be consistent under natural growth conditions ($K^2 \log(N+J)/N\to 0$, $JK\log(JK)/N\to 0$), mild class balance, and identifiable separation.

Empirical Studies and Results

Extensive simulations demonstrate:

• The test statistic has mean near zero and shrinking variance under the true $K$, but diverges under underfitting.
• Both GoF-LCM and RGoF-LCM achieve near-perfect estimation accuracy over all parameter regimes; the ratio method provides greater finite-sample robustness in weak signal and large-$K$ regimes.
• Accuracy of both methods is insensitive to choices of threshold parameter within modest intervals, with GoF-LCM robust for $\tau_N = N^{-\epsilon}$, $0 < \epsilon < 0.5$, and RGoF-LCM robust for $\gamma_N = a \log N$, $0.5 \leq a \leq 4.5$.
• The methods perform well even for $J > N$, suggesting practical robustness well beyond the scope of theoretical guarantees.
• On real Big Five personality test data, RGoF-LCM identifies interpretable latent grouping in the absence of well-specified theoretical growth conditions, confirming empirical adaptability.

Comparison to plain spectral thresholding (e.g., "Spec" method of [lyu2025spectral]) reveals catastrophic breakdown of naive singular value methods under weak signal, where the proposed strategies maintain stability.

Practical and Theoretical Implications

The results yield a statistically rigorous, computationally scalable approach to determining the number of latent classes in high-dimensional ordinal categorical data, circumventing the weaknesses of EM-based model selection intractability and fragility. The sharp dichotomy of the test statistic, together with sequential testing, provides a confidence mechanism absent from point-estimate thresholding and information criteria.

The method is immediately extensible for large-scale surveys and psychometric analyses, where computational feasibility is paramount, and presents a unified framework compatible with recent advances in categorical mixture models.

Empirically, the approach is robust to moderate threshold misspecification, extreme class size imbalance, and even to situations violating theoretical regularity ($J \gg N$).

Future Directions

Several promising avenues are highlighted for further development:

• Extension to richer response distributions (e.g., Poisson, multinomial exponential family models).
• Incorporation of mixed membership or degree heterogeneity (e.g., Grade-of-Membership models, degree-corrected class models).
• Procedures for hierarchical or dynamic (longitudinal) latent structure.
• Strong theoretical relaxation of class size and separation requirements, harnessing recent advances in spectral analysis and singular vector perturbation.

Conclusion

This paper presents a spectral goodness-of-fit paradigm for sequential estimation of the number of latent classes in high-dimensional ordinal categorical data, offering strong theoretical guarantees, computational scalability, and broad empirical robustness. The sharp dichotomous behavior of the statistic enables both formal hypothesis testing and practical diagnostic use, marking a significant advance in statistical latent variable methodology and offering a unifying platform for future work on increasingly complex latent structure models.

Reference:

Huan Qing, "Goodness-of-Fit Tests for Latent Class Models with Ordinal Categorical Data" (2602.21572).