多级评分的多维计算机化自适应测验选题策略研究
发布时间：2018-10-19 13:45
【摘要】：多维计算机化自适应测验(MCAT)将计算机化自适应测验(CAT)与多维项目反应理论(MIRT)相结合,以尽可能多得获取被试的多维特质信息为目的,在保证测验准确高效的同时,还具有从被试反应中获取诊断信息的潜力。多级评分项目因能提供更多信息并可测量更复杂的能力和技能而被广泛应用。然而,目前大多数MCAT算法技术是以M3PL或M2PL模型为条件的,这些算法与技术可能并不适用于多级评分模型。本文的目的是探讨将MCAT中的选题策略拓展到PMCAT中,并开发出新的选题策略。本研究还进一步探索了测验维度数、维度间的相关大小如何影响PMCAT的准确性和安全性。一些常用的MCAT选题策略——包括基于FI的D-优化、A-优化、E-优化及其贝叶斯版本;基于传统KL信息量的KI方法、后验期望KL信息方法(KB);以及基于后验分布间KL距离的KLP方法、互信息(MUI)方法和连续熵方法(CEM)都被拓展以适合于多级评分的项目。通过将CEM算法中后验概率的计算公式中预设的固定的先验概率替换为随着测验不断更新的当前后验概率,对原有CEM方法进行了改进。然后展开了两项Monte Carlo模拟研究:一是验证了PMCAT的可行性(研究二),并比较了各种选题策略间的表现;二是进一步探索了能力维度数(p=2 and 5)以及维度间相关大小(r=0,0.2,0.5 and 0.8)这两个因素对估计精度及项目曝光率的影响(研究三)。本研究使用的多级评分模型为多维等级反应模型(MGREM),选择EAP为测验进行中的潜在特质估计方法,测验终止条件设置为定长。模拟试验表明,拓展的PMCAT选题策略基本合理、可行,本文开发的新选题策略(MCEM)整体表现最好。研究发现:(1)大多数选题策略的估计误差随着维度数的增加变大,而由2维到5维时,KI方法的估计精度提升了;(2)维度间相关只在中等强度以上时才对选题策略的估计精度有影响,KI方法的估计误差随着维度间相关增加显著下降;(3)维度数量越多,维度间的相关越高,项目曝光率越低。特别是A-优化方法,在2维时曝光率最高,在5维时其曝光率下降到最低。多级评分项目广泛应用于李克特式评分的心理测量量表和成就测验中。采用多级评分项目的MCAT具有广阔的应用前景。对于PMCAT选题策略的拓展可供理论研究和实际应用参考。
[Abstract]:The Multidimensional computerized Adaptive Test (MCAT) combines the computerized Adaptive Test (CAT) with the Multidimensional item response Theory (MIRT) in order to obtain as much multidimensional trait information as possible, while ensuring the accuracy and efficiency of the test. It also has the potential to obtain diagnostic information from subjects' reactions. Multilevel scoring is widely used because of its ability to provide more information and to measure more complex competencies and skills. However, most of the current MCAT algorithms are based on M3PL or M2PL models, and these algorithms and techniques may not be suitable for multilevel scoring models. The purpose of this paper is to explore how to extend the topic selection strategy in MCAT to PMCAT and develop a new topic selection strategy. The study also explores how the number of dimensions and the correlation between dimensions affect the accuracy and security of PMCAT. Some common MCAT selection strategies include D- optimization, A- optimization, E- optimization and Bayesian version based on FI, KI method based on traditional KL information content, (KB); method of posterior expectation KL information and KLP method based on KL distance between posteriori distributions. Both the mutual information (MUI) method and the continuous entropy method, (CEM), are extended to suit multilevel scoring items. By replacing the fixed prior probability of the posteriori probability in the calculation formula of CEM algorithm with the current posteriori probability which is updated continuously with the test, the original CEM method is improved. Then two Monte Carlo simulation studies are carried out: one is to verify the feasibility of PMCAT (study 2), and to compare the performance of various topics; The second is to further explore the influence of two factors, the capability dimension (pt2 and 5) and the correlation between dimensions (r 0. 2 0. 2 0. 5 and 0. 8) on the estimation accuracy and item exposure (study 3). The multi-level rating model used in this study is a multi-dimensional rating response model (MGREM),). EAP is chosen as the potential trait estimation method in the test, and the test termination condition is set to a fixed length. The simulation results show that the extended PMCAT selection strategy is reasonable and feasible, and the new topic selection strategy (MCEM) developed in this paper is the best overall performance. The results show that: (1) the estimation errors of most of the selection strategies increase with the increase of the number of dimensions. From 2 to 5 dimensions, the estimation accuracy of KI method is improved. (2) Interdimensional correlation only affects the estimation accuracy of the selection strategy when the correlation is more than moderate intensity. The estimation error of KI method decreases significantly with the increase of interdimensional correlation. (3) the more dimension, the more dimension. The higher the correlation between dimensions, the lower the item exposure. Especially, the A- optimization method has the highest exposure at 2 D and the lowest exposure at 5 D. Multi-level scoring is widely used in Richter scale and achievement test. MCAT with multilevel scoring items has a broad application prospect. The development of PMCAT topic selection strategy can be used as a reference for theoretical research and practical application.
【学位授予单位】：江西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：B841
[Abstract]:The Multidimensional computerized Adaptive Test (MCAT) combines the computerized Adaptive Test (CAT) with the Multidimensional item response Theory (MIRT) in order to obtain as much multidimensional trait information as possible, while ensuring the accuracy and efficiency of the test. It also has the potential to obtain diagnostic information from subjects' reactions. Multilevel scoring is widely used because of its ability to provide more information and to measure more complex competencies and skills. However, most of the current MCAT algorithms are based on M3PL or M2PL models, and these algorithms and techniques may not be suitable for multilevel scoring models. The purpose of this paper is to explore how to extend the topic selection strategy in MCAT to PMCAT and develop a new topic selection strategy. The study also explores how the number of dimensions and the correlation between dimensions affect the accuracy and security of PMCAT. Some common MCAT selection strategies include D- optimization, A- optimization, E- optimization and Bayesian version based on FI, KI method based on traditional KL information content, (KB); method of posterior expectation KL information and KLP method based on KL distance between posteriori distributions. Both the mutual information (MUI) method and the continuous entropy method, (CEM), are extended to suit multilevel scoring items. By replacing the fixed prior probability of the posteriori probability in the calculation formula of CEM algorithm with the current posteriori probability which is updated continuously with the test, the original CEM method is improved. Then two Monte Carlo simulation studies are carried out: one is to verify the feasibility of PMCAT (study 2), and to compare the performance of various topics; The second is to further explore the influence of two factors, the capability dimension (pt2 and 5) and the correlation between dimensions (r 0. 2 0. 2 0. 5 and 0. 8) on the estimation accuracy and item exposure (study 3). The multi-level rating model used in this study is a multi-dimensional rating response model (MGREM),). EAP is chosen as the potential trait estimation method in the test, and the test termination condition is set to a fixed length. The simulation results show that the extended PMCAT selection strategy is reasonable and feasible, and the new topic selection strategy (MCEM) developed in this paper is the best overall performance. The results show that: (1) the estimation errors of most of the selection strategies increase with the increase of the number of dimensions. From 2 to 5 dimensions, the estimation accuracy of KI method is improved. (2) Interdimensional correlation only affects the estimation accuracy of the selection strategy when the correlation is more than moderate intensity. The estimation error of KI method decreases significantly with the increase of interdimensional correlation. (3) the more dimension, the more dimension. The higher the correlation between dimensions, the lower the item exposure. Especially, the A- optimization method has the highest exposure at 2 D and the lowest exposure at 5 D. Multi-level scoring is widely used in Richter scale and achievement test. MCAT with multilevel scoring items has a broad application prospect. The development of PMCAT topic selection strategy can be used as a reference for theoretical research and practical application.
【学位授予单位】：江西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：B841
【参考文献】
相关期刊论文 前10条
1 韩雨婷;涂冬波;王潇o，
本文编号：2281305
本文链接：https://www.wllwen.com/shekelunwen/xinlixingwei/2281305.html
