The Detection and Effect Size Computation of Differential Item Functioning in Polytomous Items with Logistic Regression

• Main Authors: Yuk-Hing Lam, 林郁馨
• Other Authors: 王文中
• Format: Others
• Language: zh-TW
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/89659048742421457005

碩士 === 國立中正大學 === 臨床心理學研究所 === 95 === In past studies, it is advisable to examine effect size measure when using statistic tests. Recently, Zumbo (1999) proposed R-square as an effect size measure when using logistic regression to detect differential item functioning (DIF) in polytomous items. In this study, based on 鄭致寯(2004), two experiments were conducted to compare the performance in DIF detection of the logistic discriminant function analysis (LDFA; Miller & Spray, 1993) and ordinal logistic regression (OLR; Zumbo, 1999). Situation 1 focused on the partial credit model (PCM) and Situation 2 on the generalized partial credit model (GPCM). True positive and false positive are used to show the performance in DIF detection. Three computation of R-square were also computed to investigate effect size measures. Furthermore, a real data analysis was performed. The results show that both methods performed better when tests contained solely polytomous items than when tests contained both dichotomous and polytomous items. When tests contained the same percentages of DIF items, the false positive and true positive rates were increased as the DIF magnitude were increased. The false positive rates were inflated greatly when the percentages of DIF items or the mean latent trait differences between groups were increased. Besides, the OLR performed poorer on the false positive rates. The effect size measure of three computation of were larger when tests contained solely polytomous items than when tests contained both dichotomous and polytomous items. When tests contain same percentages of DIF items, the effect size measures were increased as the DIF magnitudes were increased, and were decreased as the mean latent trait differences between groups were increased. A new criteria for three computation of R-square is advised. In conclusion, the logistic regression performs better when tests contain solely polytomous items. When the mean latent trait difference between groups is large or tests contain both dichotomous and polytomous items, the LDFA is recommended. In addition, the OLR provides R-square to measure effect sizes of DIF.