Identify model identification conditions for jointly time-varying components in longitudinal peer grading

Investigate the identification and develop statistically sound methodology for a longitudinal latent variable model for peer grading data that simultaneously allows (i) time-varying examinee true scores (e.g., via latent growth curves), (ii) time-varying rater characteristics (bias and reliability), and (iii) time-varying assessment difficulty parameters, and ascertain conditions under which such a model is identifiable and estimable in realistic settings.

Background

The paper proposes a Bayesian latent variable framework for peer grading that captures dependencies induced by students serving as both examinees and graders. In the discussion, the authors note that a natural extension is to model growth in student ability over multiple assessments using latent growth curve models.

However, they point out that in practice not only examinee ability but also rater characteristics and assessment difficulty may change over time. Simultaneously modeling all of these time-varying components may lead to model identification issues, which they explicitly defer for future investigation.