In the ordinal regression analyses, interview questions (items) are summed to create ordinal scores for depression. The summing of items does not necessarily optimize the information in the items. In contrast, IRT models a normally distributed latent liability based on the responses to the individual questionnaire items. This allows the “difficulty” (endorsement probability) and “discrimination” (extent to which items discriminate persons across the latent continuum) of each item to be taken into account. A review of the advantages of IRT over classical test theory is provided by Reise and Henson [2003]. One advantage is that IRT analyses model missingness of item responses by interpolating from the rest of the data. Therefore, individuals who were excluded from the ordinal regression analyses could be retained (Table II).
