The Performance And Applied Research Of DetectingDIF And DAF Under Modified Higher-order DINA Model

Because of Differential Item Functioning directly related to the validity of the test questions and equity issues, in recent years, the Differential Item Functioning in gradually by scientists and educators psychometric attention. Our country has a vast number of ethnic, besides economic and cultural development is very uneven between regions, urban and rural areas vary greatly, unified examination in a wide range, will definitely lead to a lot of factors that affect the fairness and effectiveness of the exam. Therefore, the study of Differential Item Functioning is particularly important.This study attempts to detect Differential Item Functioning from a new perspective, and to split the source of Differential Item Functioning in two projects: architecture independent unfavorable(Differential Item Functioning, DIF) and the structurally related adverse(Differential Attribute Functioning, DAF) and explore the use of modified higher-order DINA model approach for performance testing to detect DIF and DAF, influencing factors and applications. This study consists of two sub-study components:One study using Monte Carlo simulation method probed the performance of modified higher order DINA model detecting DIF and DAF: the model parameter estimation has strong stability, high accuracy. When using the model to detect DIF and DAF for testing, and found: in 144 experimental conditions, with the exception of several experimental conditions, DIF and DAF make type I error detection probability was less than 0.05.The complexity of Q matrix and the sample size, the ability distribution between groups and discrimination parameter have a great impact on the model of statistical tests when detecting DAF and DIF. When the larger the sample size, the power is higher. DAF statistical tests for strength, when the Q matrix simple statistical test on the whole power is higher than 0.8. When more complex Q matrix, the higher of the attribute discrimination, the ability does not match the distribution of the statistical tests, the lower force. When the sample size is large with a high statistical test force guessing parameter is high, higher than 0.8 on the whole; the more complex Q matrix, the higher the attribute discrimination, the ability does not match the distribution of the statistical tests, the higher force. Parametric statistical tests on the \ force turnovers medium-low, the more complex Q matrix, the higher the attribute discrimination, the ability to match the distribution of the more, the more consistent DIF type, the higher the statistical test force turnovers parameters; but Q matrix and distribution when the matrix is simple, DAF detect and while and returned with a true model and the power of DIF, DAF was satisfactory.The second study using English reading comprehension test as an example to explore the performance of modified higher-orderDINA detecting DIF and DAF. The results show that: in the measured data, the model can effectively detect DIF items and the presence of DAF, DAF detection results but will be correctly calibrated Q matrix effects.In summary, the method which is based on modified higher-order DINA model to detect DIFand DAF is feasible, under this framework of cognitive diagnosis provides a natural framework of simultaneous detection of DAF and DIF: Under the same conditions as property master profile, according to differences between different groups in the project answer for DIF detection can ensure the fairness test, and can improve the validity of the tests. Under normal conditions the same capacity, depending on the performance of groups of subjects in the case of a property answer DAF detection can be supplied to the main strengths and weaknesses of the test group were tested on different attributes of a better understanding and awareness.