The parameter estimation of violating local independence － the comparison between BILOG-MG and the software of HLM

• Main Authors: Chang Shih-Yu, 張世諭
• Other Authors: Lin Yuan-Horng
• Format: Others
• Language: zh-TW
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/74051966503248177100

碩士 === 國立臺中教育大學 === 教育測驗統計研究所 === 96 === There are four important assumptions in item response theory. Item response models entail strong assumptions, and the benefits they offer would be accrue only when the assumptions hold. One important assumption made in item response theory is the assumption of local independence. In fact, the situations of the violation of local independence assumption may be produced among items by many conditions, such as testlet test or speed test. This lack of conditional independence, if ignored by applying item response models using the assumption of local independence, will result in the bias of the parameter estimation. Therefore, it is an important issue that reducing the bias of the parameter estimation when the items exist the violation of local independence assumption. This research utilize the method of the computer simulation to explore the effect of the estimating conditions of Rasch model and 1-P HGLLM when items exist the violation of local independence assumption. Three factors are considered in this simulation study. They are number of items, sample size and degree of local dependence. This study uses testlet response model to produce the simulate data, then estimates parameter by the software of BILOG-MG and HLM6.03. Finally, this research uses goodness-of-recovery to analyze the proficiency estimates of these two software. The main findings are as follows. 1.When the number of items is fixed, the estimation preciseness of HLM6.03 will be better than BILOG-MG. 2.When sample size is fixed, the estimation preciseness of HLM6.03 will be better than BILOG-MG on the condition of local dependence. 3.When the degree of local dependence is fixed, the robustness of HLM6.03 will be better than BILOG-MG.