Matrix Decomposition Perspective for Accuracy Assessment of Item Response Theory
This work provides a matrix decomposition perspective for evaluating item response theory accuracy, which is incremental for researchers in psychometrics and educational assessment.
The paper tackles the problem of assessing the accuracy of reconstructed item response matrices in item response theory by comparing them to low-rank approximations from matrix decomposition methods, finding that the reconstructed matrix's closeness to the observed matrix falls between approximations using different ranks.
The item response theory obtains the estimates and their confidence intervals for parameters of abilities of examinees and difficulties of problems by using the observed item response matrix consisting of 0/1 value elements. Many papers discuss the performance of the estimates. However, this paper does not. Using the maximum likelihood estimates, we can reconstruct the estimated item response matrix. Then we can assess the accuracy of this reconstructed matrix to the observed response matrix from the matrix decomposition perspective. That is, this paper focuses on the performance of the reconstructed response matrix. To compare the performance of the item response theory with others, we provided the two kinds of low rank response matrix by approximating the observed response matrix; one is the matrix via the singular value decomposition method when the response matrix is a complete matrix, and the other is the matrix via the matrix decomposition method when the response matrix is an incomplete matrix. We have, firstly, found that the performance of the singular value decomposition method and the matrix decomposition method is almost the same when the response matrix is a complete matrix. Here, the performance is measured by the closeness between the two matrices using the root mean squared errors and the accuracy. Secondary, we have seen that the closeness of the reconstructed matrix obtained from the item response theory to the observed matrix is located between the two approximated low rank response matrices obtained from the matrix decomposition method of k= and k=2 to the observed matrix, where k indicates the first k columns use in the decomposed matrices.