Nonparametric diagnostic classification analysis for testlet based tests
- UNCG Author/Contributor (non-UNCG co-authors, if there are any, appear on document)
- Shuying Sha (Creator)
- Institution
- The University of North Carolina at Greensboro (UNCG )
- Web Site: http://library.uncg.edu/
- Advisor
- Robert Henson
Abstract: Diagnostic classification Diagnostic Classification Models (DCMs) are multidimensional confirmatory latent class models that can classify individuals into different classes based on their attribute mastery profiles. While DCMs represent the more prevalent parametric approach to diagnostic classification analysis, the Hamming distance method, a newly developed nonparametric diagnostic classification method, is quite promising in that it does not require fitting a statistical model and is less demanding on sample size. However, both parametric and nonparametric approach have assumptions of local item independency, which is often violated by testlet based tests. This study proposed a conditional-correlation based nonparametric approach to assess testlet effect and a set of testlet Hamming distance methods to account for the testlet effects in classification analyses. Simulation studies were conducted to evaluate the proposed methods. In the conditional-correlation approach, the testlet effects were computed as the average item-pair correlations within the same testlet by conditioning on attribute profiles. The inverse of the testlet effect was then used in testlet Hamming distance method to weight the Hamming distances for that particular testlet. Simulation studies were conducted to evaluate the proposed methods in conditions with varying sample size, testlet effect size, testlet size, balance of testlet size, and balance of testlet effect size. Although the conditional-correlation based approach often underestimated true testlet effect sizes, it was still able to detect the relative size of different testlet effects. The developed testlet Hamming distance methods seem to be an improvement over the estimation methods that ignore testlet effects because they provided slightly higher classification accuracy where large testlet effects were present. In addition, Hamming distance method and maximum likelihood estimation are robust to local item dependency caused by low to moderate testlet effects. Recommendations for practitioners and study limitations were provided.
Nonparametric diagnostic classification analysis for testlet based tests
PDF (Portable Document Format)
1424 KB
Created on 8/1/2016
Views: 1533
Additional Information
- Publication
- Dissertation
- Language: English
- Date: 2016
- Keywords
- Conditional correlation, Diagnostic classification modeling, Local item dependency, Nonparametric classification analysis, Testlet
- Subjects
- Educational tests and measurements
- Categories (Philosophy)
- Classification