【作者】 罗照盛； 张飘；

【Author】 Zhang Piao;Luo Zhaosheng;School of Psychology, Jiangxi Normal University;

【机构】 江西师范大学心理学院；

【摘要】 等值是在教育测验、自适应测验、题库建设等大规模测验和评价中经常用到的测量技术。随着社会的发展,教育测验越来越追求测验的公平性,而测验公平则需建立在精度良好的等值技术之上。当把测验等值应用到现实环境中时,我们发现,受不同地区社会经济状况及教育发展水平差异的影响,不同被试群体的能力分布形态并不是那么地有规则,且会因受到各种各样的影响,在作答数据上产生无可避免的缺失。本研究旨在探讨不同条件下的不规则分布数据的缺失及缺失处理方法对等值精度的影响,并寻找适用于不同条件的等值精度最高的处理方法,以应用于对能力状态不规则分布的被试群体进行等值的现实情境。研究基于0、1计分题型的两参数逻辑斯蒂模型(2PLM),采用非等组锚测试(NEAT)的五因素完全随机设计,运用蒙特卡洛(MCMC)模拟缺失机制为完全随机缺失(MCAR)的不规则分布数据,比较对不同能力分布形态的目标组进行测验等值的结果,同时考虑了不同的样本容量、缺失比例,以及缺失数据有无平滑的预处理,并采用不同的缺失数据处理方法,包括删除法、插补法、极大似然法,模拟对缺失数据的处理。研究采用项目特征曲线法中经典的黑巴诺(HB)法对等值系数进行估计,并对不同的因素之间的交互作用进行分析。研究以等值系数平均绝对离差(ABSE)的大小作为等值精度的衡量标准,ABSE越小,等值精度越高。研究发现,基于MCAR机制下的处理方法对数据分布形态的假设比较敏感,不同处理方法的等值结果之间存在显著差异。本研究将进一步研究现实中常用的多级计分的等级反应模型(GRM)以及混合模型下的不规则分布数据的缺失及处理方法对等值精度的影响。更多还原

【Abstract】 Equating is a measurement technique that is often used in large-scale quizzes and evaluations such as educational tests, compute adaptive tests, and item bank construction. Along with the development of society, education test more and more pursuit of the test fairness, and test fair need to be established in the accuracy of the equivalent technology. When applying the test equating to the real world, we found that distribution of the capacity of different subjects is not so regular, because of the different socioeconomic status and differences in the level of educational development in different regions, and will be affected by various effects, resulting in missing phenomena in the response data inevitably. The aim of this study is to investigate the effect of the missing and methods of irregular distribute data on the precision of equating under different conditions, and to find the highest equating method for different conditions to apply to realistic situation to the subjects with the distribution of the irregular ability to equate. Based on the 2-parameter logistic model(2 PLM) of dichotomously-scored items, used a five-factor non-equal group anchor test(NEAT) completely random design, and the Monte Carlo(MCMC) is used to simulate the irregular distributed data, the missing mechanism for irregularly distributed data is Missing Completely at Random(MCAR). Compared with the results of the distribution of the capacity of different target groups, taking into account the different sample size, the missing ratio, and the pretreatment of the missing data,and use different missing data processing methods, including the deletion method, interpolation method, the maximum likelihood method, to simulate the processing of missing data. The equivalence coefficients are estimated by the classical Haebara method(HB) in the item characteristic curve method, and the interaction between different factors is analyzed. The size of the Absolute Error(ABSE) of the equivalent coefficient is taken as the measure of the precision of equating. The smaller the value, the higher the precision of equating is. This study found that the method based on the MCAR mechanism is more sensitive to the hypothesis of the distribution of the data, and there is a significant difference between the equivalent results of the different processing methods. This research will further study the effect of the missing and processing methods of irregular distributed data on the precision of equating in the reality of the commonly used the Graded Response Model(GRM) of polytomously-scored items and the mixed model conditions.更多还原