While the technique of looking at the effect including group assignment has on the relationship between ability level and item responses is superior in our view, we recognize that there are other opinions of how best to assess confounding relationships. Therefore, DIFdetect enables several other assessment techniques.

One of these is based on the predictive ability of group assignment for determining item responses when controlled for ability level. In the models we have developed, this implies looking at the coefficient for the groupassignment variable in the third model:

          ologit itemresponse abilitylevel groupassignment (3).

The b coefficient corresponding to group assignment (i.e., b₂) is examined.

The same 1993 paper by Maldonado and Greenland recommends that if this approach is used for the decision that there is significant confounding in the relationship between exposure and outcome, a higher alpha level should be accepted. Specifically, they recommend 0.20 rather than the more conservative 0.05.

DIFdetect offers the ability to look at b₂ for detection of uniform DIF. It also enables the user to establish the alpha level accepted as evidence of significant confounding.

Analogous to the negative 2 log likelihood test used by DIFdetect for detection of non-uniform DIF, DIFdetect also provides for testing the difference in negative 2 log likelihood of models with and without the group assignment. Again a chi squared test with 1 degree of freedom is used to test the null hypothesis that adding group assignment does not improve the model. Again following Maldonado and Greenland's recommendation, the default for DIFdetect is an alpha level of 0.20 for this test, though you can choose whatever value of alpha you wish.