Latent t Models for Ordinal Data: Likelihood Inference beyond the Multivariate Normal

The multivariate normal distribution is commonly regarded as the gold standard latent variable model for correlated ordinal data. In this work, we consider the multivariate Student’s t distribution as the underlying latent variable, which introduces additional flexibility through the degrees-of-freedom parameter while preserving the threshold-based construction of ordinal responses. The resulting model, however, poses substantial theoretical and computational challenges for likelihood-based inference. These challenges arise both from the discretization mechanism—requiring repeated evaluation of multivariate t copula probabilities over threshold-defined regions—and from the presence of the degrees-of-freedom parameter itself, which complicates optimization and inference. We discuss how this framework can improve the modeling of dependence in ordinal data and provide practical guidance for conducting maximum-likelihood estimation in this setting.