| dc.description.abstract |
This study examines the following DIF procedures--Transformed Item Difficulty (TID), Full Chi-Square, Mantel-Haenszel chi-square, Mantel-Haenszel delta, Logistic Regression, SOS2, SOS4, and Lord's chi-square under three sample sizes, two test lengths, four cases of item discrimination arrangement, and three item difficulty levels. The study is in two parts: The first part examines the distributions of the indices under null (no bias) conditions. The second part deals with the power of the procedures to detect known bias in simulated test data. Agreements among procedures are also addressed. Lord's chi-square certainly appears to perform very well. Its detection rates were very good, and its percentiles were not affected by discrimination level or test length. In retrospect, one would like to know how well it might do at smaller sample sizes. When the tabled values were used, it performed equally well in detecting bias and improved in reducing false positive rates. Of the other indices, the Mantel-Haenszel and the logistic regression indices seemed the best. Camilli chi-square had a number of problems. Its tabled values were not at all useful for detection of bias. The TID was somewhat better but does not have a significance test associated with it. One would need to rely on baseline studies, if one were to use it. For uniform bias either Mantel-Haenszel chi-square or logistic regression would be recommended, while for nonuniform bias logistic regression would be appropriate. It is interesting to note that Lord's chi-square was effective for detecting either kinds of bias. We have been told that sample size is related to chi-square values. For each of the chi square indices the observed values were considerably lower than tabled values. Of course, these were conditions where no bias was present except that which might be randomly induced in data generation. Perhaps it is those instances where bias is truly present that larger sample sizes allow us to more easily identify biased items. Certainly the proportions of biased items detected was greater for large sample sizes for Camilli chi-square, Mantel-Haenszel chi-square, and logistic regression chi-squares. |
en |
