基于分部评分模型思路的多级评分认知诊断模型开发

基于分部评分模型的思路, 本文提出了一般化的分部评分认知诊断模型(General Partial Credit Diagnostic Model, GPCDM), 与国际上已有的基于分部评分模型思路的多级评分模型GDM (von Davier, 2008)和PC-DINA (de la Torre, 2012)相比, GPCDM的Q矩阵定义更加灵活, 项目参数的约束条件更少。Monte Carlo实验研究表明, GPCDM模型的参数估计精度指标RMSE介于0.015, 0.043], 表明估计精度尚可; TIMSS (2007)实证数据应用研究表明, 与GDM和PC-DINA模型相比, GPCDM与该数据的拟合度更好, 并且使用GPCDM分析该数据的诊断效果也更优。总之, 本研究提供了一种约束条件更少、功能更为强大的多级评分认知诊断模型。

Development of a Generalized Cognitive Diagnosis Model for polytomous responses based on Partial Credit Model

Currently, a large number of cognitive diagnosis models (CDMs) have been proposed to satisfy the demands of the cognitively diagnostic assessment. However, most existing CDMs are only suitable for dichotomously scored items. In practice, there are lager polytomously-score items/data in educational and psychological tests. Therefore, it is very necessary to develop CDMs for polytomous data. Under the item response theory (IRT) framework, the polytomous models can be divided into three categories: (i) the cumulative probability (or graded-response) models, (ii) continuation ratios (or sequential) models, and (iii) the adjacent-category (or partial-credit) models. At present, several efforts have been made to develop polytomous partial-credit CDMs, including the general diagnostic model (GDM; von Davier, 2008) and the partial credit DINA (PC-DINA; de la Torre, 2012) model. However, the existing polytomous partial-credit CDMs need to be improved in the following aspects: (1) These CDMs do not consider the relationship between attributes and response categories by assuming that all response categories of an item measure the same attributes. This may result in loss of diagnostic information, because different response categories could measure different attributes; (2) More importantly, the PC-DINA is based on reduced DINA model. Therefore, the current polytomous CDMs are established under strong assumptions and do not have the advantages of general cognitive diagnosis model.The current article proposes a general partial credit diagnostic model (GPCDM) for polytomous responses with less restrictive assumptions. Item parameters of the proposed models can be estimated using the marginal maximum likelihood estimation approach via Expectation Maximization (MMLE/EM) algorithm.Study 1 aims to examine (1) whether the EM algorithm can accurately estimate the parameters of the proposed models, and (2) whether using item level Q-matrix (referred to as the Item-Q) to analyze data generated by category level Q-matrix (referred to as the Cat-Q) will reduce the accuracy of parameter estimation. Results showed that when using Cat-Q fitting data, the maximum RMSE was less than 0.05. When the number of attributes was equal to 5 or 7, the minimum pattern match rate (PMR) was 0.9 and 0.8, respectively. These results indicated that item and person parameters could be recovered accurately based on the proposed estimation algorithm. In addition, the results also showed that when Item-Q is used to fit the data generated by Cat-Q, the estimation accuracy of both the item and person parameters could be reduced. Therefore, it is suggested that when constructing the polytomously-scored items for cognitively diagnostic assessment, the item writer should try to identify the association between attributes and categories. In the process, more diagnostic information may be extracted, which in turn helps improve the diagnostic accuracy.The purpose of Study 2 is to apply the proposed model to the TIMSS (2007) fourth-grade mathematics assessment test to demonstrate its application and feasibility and compare with the exiting GDM and PC-DINA model. The results showed that compared with GDM and PC-DINA models, the new model had a better model fit of test-level, higher attribute reliability and better diagnostic effect.