NotesFAQContact Us
Search Tips
ERIC Number: ED131091
Record Type: RIE
Publication Date: 1976-Apr
Pages: 11
Abstractor: N/A
Reference Count: 0
Test Bias in the Absence of a Criterion.
Echternacht, Gary
This paper proposes a method of transforming item p-values (the proportion answering a test item correctly) to what are termed "delta" values. First used by Conrad in l948, deltas are routine statistics computed in all analyses at Educational Testing Service. Using this approach one would conclude no test bias if differences in resulting deltas are constant from item to item for the groups. Under a null hypothesis of no test bias, the sample delta differences should be distributed normally with some unknown mean and unknown variance. If evidence can be gathered to the contrary, the null hypothesis can be rejected and bias concluded. To test the hypothesis of normality, one plots the differences in item deltas on normal probability paper. First, one orders the delta differences. Second, one pairs each difference with the value S/(m+1) where S is the rank of the delta difference and m the number of items. The purpose of S/(m+1) is to anchor the median difference to the 50th percentile. If the differences follow a normal distribution, the plotted points will be on a straight line. A statistical test has been developed by Lilliefors (1967) using the sample mean and variance as a basis for the straightness of the resulting line. That technique was adapted for use with probability paper and formed the crux of the technique. (Author/MV)
Descriptors: Item Analysis, Test Bias
Publication Type: Reports - Research
Education Level: N/A
Audience: N/A
Language: N/A
Sponsor: N/A
Authoring Institution: N/A
Note: Paper presented at the Annual Meeting of the American Educational Research Association (60th, San Francisco, California, April 19-23, 1976)