To elaborate on the third point above, the ordinal logistic regression decomposes the
variation in y∗ , the latent continuous variable defined in Equation (3), into "explained"
and "unexplained" components. As per the typical use of regression, this squared
multiple correlation then represents the proportion of variation in the dependent variable
captured by the regression and is defined as the regression sum of squares over the total
sum of squares. Therefore, the R-squared values arising from the application of ordinal
logistic regression are typical in magnitude to those found in behavioral and social
science research and Cohen (1992) and Kirk's (1996) guidelines may be useful in
interpretation.
Finally, although the ordinal logistic regression R-squared for measuring DIF in
either multicategory scores and binary scores is introduced here, it is founded on the
statistical theory for ordinal logistic regression, and the hierarchical sequential modeling
strategy implicit in the definition of DIF. Future research applying this technique to a
variety of measures will almost certainly result in a refinement of the approach and its
interpretation.
3 Note that not all R-squared measures for nominal binary responses can be
used in a hierarchical regression.