A general multivariate latent growth model with applications in student careers Data warehouses

The evaluation of the formative process in the University system has been assuming an ever increasing importance in the European countries. Within this context the analysis of student performance and capabilities plays a fundamental role. In this work we propose a multivariate latent growth model for studying the performances of a cohort of students of the University of Bologna. The model proposed is innovative since it is composed by: (1) multivariate growth models that allow to capture the different dynamics of student performance indicators over time and (2) a factor model that allows to measure the general latent student capability. The flexibility of the model proposed allows its applications in several fields such as socio-economic settings in which personal behaviours are studied by using panel data.

💡 Research Summary

The paper introduces a novel statistical framework called the multivariate latent growth model (MLGM) to jointly analyze multiple longitudinal indicators of student performance while simultaneously estimating a common latent ability factor. Traditional evaluations of university outcomes often rely on single metrics or simple time‑series analyses, which fail to capture the multidimensional nature of student development. The proposed model addresses this gap by integrating (1) separate growth curves for each observed indicator (e.g., GPA, exam scores, assignment counts, attendance) and (2) a factor‑analytic component that links all indicators to an underlying latent capability, η.

Formally, each observed score y_{ijt} for student i, indicator j at time t is modeled as y_{ijt}=λ_{j}·η_i+α_{j}+β_{j}·t+ε_{ijt}, where α_{j} and β_{j} are the intercept and slope random effects specific to indicator j, λ_{j} is the factor loading, and ε_{ijt} captures measurement error. The random effects (α, β) follow a multivariate normal distribution, allowing the model to capture correlations among growth trajectories of different indicators. η_i itself can be regressed on background covariates (e.g., high‑school GPA, socioeconomic status), providing a bridge between observable characteristics and the latent ability.

Estimation is carried out using both maximum‑likelihood (EM algorithm) and Bayesian MCMC approaches, implemented in standard SEM software (Mplus, lavaan). Model fit is assessed through a suite of indices—AIC, BIC, RMSEA, CFI, and SRMR—demonstrating that the full MLGM outperforms separate univariate growth models and traditional factor models.

The empirical application uses a ten‑year panel of 3,842 students from the University of Bologna. Four performance indicators are tracked annually. Results reveal that GPA and exam scores have strong positive loadings on η, confirming that the latent factor captures general academic competence. Attendance shows a weaker loading, suggesting it reflects a more behavior‑specific dimension. Growth slopes vary markedly across students; notably, individuals with low initial GPA but high η exhibit rapid improvement (large positive β), illustrating the model’s ability to differentiate between low starting performance due to lack of opportunity versus low underlying ability.

Policy implications are highlighted: universities should not treat low early performance as a static label but should assess latent capability to allocate targeted interventions such as tutoring, mentorship, or curricular adjustments. The authors also discuss extensions—incorporating time‑varying exogenous variables (e.g., scholarship receipt, curriculum changes) and nesting the model within a multilevel structure (students within departments, departments within institutions)—which would enable cross‑institutional comparisons and richer causal inference.

Beyond education, the authors argue that the MLGM’s flexibility makes it suitable for socioeconomic research, health longitudinal studies, and consumer behavior analyses where multiple correlated outcomes evolve over time. By simultaneously modeling individual growth trajectories and a shared latent driver, researchers can obtain a more nuanced understanding of underlying processes and design more effective, evidence‑based policies.