What it’s about

In this paper, we propose auto-regressive item response theory models for progressive matrices, allowing each item response to depend not only on latent ability and item parameters, but also on prior responses in sequence.

Abstract

Standard measurement models typically assume local independence among item responses conditional on latent traits. This article argues that, in sequential cognitive tasks such as progressive matrices, responses may show meaningful lagged dependencies that should be modeled directly. Building on prior auto-regressive modeling work, the paper introduces AR-IRT extensions of binary IRT models and applies them to a publicly available progressive matrices dataset. Results indicate that a lag-1 auto-regressive 2PL model improves model fit relative to a traditional 2PL, while producing more conservative discrimination estimates and standard errors. The authors discuss implications, limitations, and extensions such as multiple-lag and variable-lag specifications.

How to access the paper

Go to the journal page for this article here.

Download the PDF here.