A Poisson Factor Mixture Model for the Analysis of Linguistic Competence in Italian University Students' Writing

Public debate on the alleged decline of language skills among younger generations often focuses on university students, the most highly educated segment of the population. Rather than addressing the ill posed question of linguistic decline, this paper examines how formal written Italian is currently used by university students and whether systematic patterns of competence and heterogeneity can be identified. The analysis is based on data from the UniversITA project, which collected formal texts written by a large and nationally representative sample of Italian university students. Texts were annotated for linguistically motivated features covering orthography, lexicon, syntax, morphosyntax, coherence, register, and sentence structure, yielding low frequency multivariate count data. To analyse these data, we propose a novel model-based clustering approach based on a Poisson factor mixture model that accounts for dependence among linguistic features and unobserved population heterogeneity. The results identify two correlated dimensions of writing competence, interpretable as communicative competence and linguistic grammatical competence. When educational and socio demographic information is incorporated, distinct student profiles emerge that are associated with field of study and educational background. These findings provide quantitative evidence on contemporary writing and offer insights relevant for language education and higher education policy.

💡 Research Summary

The paper investigates contemporary written Italian among university students by analysing a large, nationally representative sample of 2,137 second‑year students who each produced a 250‑500‑word formal essay. Each essay was manually annotated for seven linguistically motivated error types—orthography, register, marked sentences, lexicon, morphosyntax, coherence, and syntax—resulting in low‑frequency multivariate count data. In addition, a detailed socio‑demographic questionnaire (58 items) supplied information on gender, family background, high‑school diploma type, field of study, and geographic region, enabling the exploration of relationships between linguistic performance and student characteristics.

To model these sparse counts while accounting for inter‑feature dependence and unobserved heterogeneity, the authors propose a Poisson factor mixture model (PFMM). The PFMM is a mixture of generalized linear latent variable models (GLLVMs) for count data. For each individual, the p‑dimensional count vector y (p = 7) is linked to a q‑dimensional latent Gaussian vector z (q < p) through a log‑link: log ω(z) = λ₀ + Λz, where Λ is a p × q factor loading matrix and ω(z) are the Poisson rates. The latent vector z follows a finite mixture of k multivariate normals with means μ_i and covariances Σ_i; a categorical allocation variable s indicates component membership. This hierarchical structure simultaneously reduces dimensionality (via the low‑dimensional latent space) and induces clustering in the observed count space.

Identifiability is ensured by (i) centering the mixture means (∑π_i μ_i = 0) and scaling the mixture covariance to the identity, (ii) fixing the upper‑triangular part of Λ to zero (Jöreskog constraint), and (iii) setting the first intercept to zero. The Ledermann condition is applied to guarantee that the chosen q does not exceed the maximum admissible dimensionality given p.

Parameter estimation proceeds via a generalized Expectation–Maximization (EM) algorithm. The E‑step approximates the intractable expectations over z and s using Gauss‑Hermite quadrature; the M‑step updates mixing proportions, component means, covariances, and the loading matrix. λ₀ and Λ are refined with a Newton–Raphson routine, yielding a fully iterative EM scheme.

A comprehensive simulation study demonstrates that the PFMM accurately recovers the true number of clusters, latent dimensionality, and loading structure across a range of scenarios, outperforming competing Poisson‑lognormal mixture models that require one latent per observed variable.

Applying the PFMM to the UniversITA data, the authors identify two substantive latent dimensions. The first dimension loads heavily on register, coherence, and lexicon errors and is interpreted as “communicative competence.” The second dimension loads on orthography, morphosyntax, and syntax errors, representing “grammatical competence.” Both dimensions are positively correlated, suggesting that students who are strong in one tend to be strong in the other, yet the model still distinguishes distinct profiles.

Model‑based clustering (optimal k determined by BIC/ICL) yields four to five student clusters. Cluster characteristics align with academic and educational background: (1) students in humanities/social sciences exhibit high grammatical competence but relatively lower communicative competence, reflecting frequent register/lexicon deviations; (2) students in STEM fields show high scores on both dimensions, indicating overall strong writing; (3) students with a language‑oriented high‑school diploma score higher on communicative competence; (4) students from non‑Italian‑speaking families or with weaker socioeconomic status tend to have lower scores on both dimensions. These findings illustrate how the latent scores can be mapped onto concrete socio‑educational variables.

The authors discuss practical implications: (i) curriculum designers can tailor writing instruction to target the specific deficits of each cluster (e.g., register training for humanities students); (ii) policymakers can monitor the impact of secondary‑school curricula on university‑level writing; (iii) the latent scores provide individualized diagnostic feedback for students and instructors.

In conclusion, the Poisson factor mixture model offers a parsimonious yet powerful framework for analysing low‑frequency multivariate count data in linguistic research. It simultaneously achieves dimensionality reduction, captures inter‑feature dependence, and uncovers meaningful subpopulations. Future work may extend the model to longitudinal data, incorporate automatically extracted textual features, or explore Bayesian estimation to further enhance flexibility.