Winsteps

The more papers are published about DIF detection, the more squirrelly ("eccentric, cunningly unforthcoming or reticent" - yourdictionary.com) it becomes. There have been comments that the Winsteps and Facets DIF estimates (computed as interaction terms, after the measure main-effects have been estimated) may be errant. So what is the best method? According to the literature, no method shines. Logistic regression is promising because it provides a DIF size estimate for polytomies, but, at least for dichotomies, "The DIF effect size measures based on logistic regression, however, appeared to be insensitive to the specified DIF conditions." (Hidalgo & López-Pina, EPM 64, 6, 903-915 (2004)).

So what to do? Mantel-Haenszel (1959) is well established for significance and size of dichotomous DIF. Mantel (1963) provides a significance test for ordinal data. These will be included in the Winsteps DIF Table, using "thin" matching based on estimated measures, so missing data can be accommodated. Here is part of the Knox Cube Test output, showing that Winsteps estimates DIF of 1.61 logits where the MH estimate is 1.95:

+--------------------------------------------------------------------------------------------------------+
 KID     DIF   DIF   KID     DIF   DIF      DIF    JOINT                MantelHanzl TAP                
 CLASS MEASURE S.E.  CLASS MEASURE S.E.  CONTRAST  S.E.   t  d.f. Prob. Prob.  Size Number  Name       
--------------------------------------------------------------------------------------------------------
 F       2.85   .89  M       1.24   .70      1.61  1.13  1.42  32 .1639 .2049  1.95     13 1-4-3-2-4   
+--------------------------------------------------------------------------------------------------------+