Treatment resistant depression incidence estimates from studies of health insurance databases depend strongly on the details of the operating definition.

Background: Health services databases provide population-based data that have been used to describe the epidemiology and costs of treatment resistant depression (TRD). This retrospective cohort study estimated TRD incidence and, via sensitivity analyses, assessed the variation of TRD incidence within the range of implementation choices.

Methods: In three US databases widely used for observational studies, we defined TRD as failure of two medications as evidenced by their replacement or supplementation by other medications, and set maximum durations (caps) for how long a medication regimen could remain in use and still be eligible to fail.

Results: TRD incidence estimates varied approximately 2-fold between the two databases (CCAE, Medicaid) that described socioeconomically different non-elderly populations; for a given cap varied 2-fold to 4-fold within each database across the other implementation choices; and if the cap was also allowed to vary, varied 6-fold or 7-fold within each database.

Limitations: The main limitations were typical of studies from health services databases and included the lack of complete -rather than recent - medical histories, the limited amount of clinical information, and the assumption that medication dispensed was consumed as directed.

Conclusion: In retrospective cohort studies from health services databases, TRD incidence estimates vary widely depending on the implementation choices. Unless a firm basis for narrowing the range of these choices can be found, or a different analytic approach not dependent on such choices is adopted, TRD incidence and prevalence estimates from such databases will be difficult to compare or interpret.