Multi-Dimensional Item Response Theory and the Force Concept Inventory

TL;DR

This study applies Multi-Dimensional Item Response Theory to analyze the Force Concept Inventory, revealing its factor structure, testing theoretical models, and demonstrating the superiority of an optimal model over the original.

Contribution

It introduces the use of MIRT for FCI analysis, providing a statistically rigorous method to identify and test the underlying conceptual factors.

Findings

Optimal 9-factor solution identified with MIRT

Reduced item set yielded a 6-factor structure with limited relation to mechanics

Theoretical model constrained by expert solutions fit the data better than the original FCI model

Abstract

Research on the test structure of the Force Concept Inventory (FCI) has largely been performed with exploratory methods such as factor analysis and cluster analysis. Multi-Dimensional Item Response Theory (MIRT) provides an alternative to traditional Exploratory Factor Analysis which allows statistical testing to identify the optimal number of factors. Application of MIRT to a sample of $N=4,716$ FCI post-tests identified a 9-factor solution as optimal. Additional analysis showed that a substantial part of the identified factor structure resulted from the practice of using problem blocks and from pairs of similar questions. Applying MIRT to a reduced set of FCI items removing blocked items and repeated items produced a 6-factor solution; however, the factors had little relation the general structure of Newtonian mechanics. A theoretical model of the FCI was constructed from expert solutions and fit to the FCI by constraining the MIRT parameter matrix to the theoretical model. Variations on the theoretical model were then explored to identify an optimal model. The optimal model supported the differentiation of Newton's 1st and 2nd law; of one-dimensional and three-dimensional kinematics; and of the principle of the addition of forces from Newton's 2nd law. The model suggested by the authors of the FCI was also fit; the optimal MIRT model was statistically superior.