As mentioned earlier, uniform DIF is analogous to confounding. Thus, searching for uniform DIF is equivalent to taking sides in a long-standing debate within epidemiology as to how best to empirically look for confounding when the true relationships between variables is unknown (as is almost always the case). We were very much influenced in our thinking by a 1993 paper: Maldonado G, Greenland S, Simulation study of confounder-selection strategies. Am J Epidemiol 1993 Dec 1;138(11):923-36.

Because there is still healthy debate over the best technique to detect confounding, DIFdetect offers several options for the detection of uniform DIF. We do take a position on this issue: in our opinion, the best technique for empirical assessment of confounding is whether the inclusion of the groupassignment variable effects the relationship between abilitylevel and itemresponse. If inclusion of group assignment affects the strength of that relationship, then uniform DIF is present. If not, then uniform DIF is not present, and, as long as no non-uniform DIF was detected for the item, it is free from DIF of either sort.

Following the recommendation of Maldonado and Greenland, we use a 10% difference as our marker for the presence of uniform DIF. This is operationalized as follows:

First, models with and without group assignment are fit to the data. The model with the group assignment has been seen before (it was model 3 on the non-uniform DIF page) and we repeat it here:

ologit itemresponse abilitylevel groupassignment (3)

ologit itemresponse abilitylevel (4)

Then, the beta coefficients for itemresponse are compared. That coefficient from model 3 we will call b₁, and from model (4) b₁*. We then compute the absolute value of the ratio

       (5)

If the ratio from equation (5) is found to be larger than 0.10, we conclude that uniform DIF is present.