LAUSR

Decades after the case-crossover (CCO) design was first described [1], recent studies of potential biases and alternative applications of CCO analyses using pharmaceutical and healthcare utilization databases are still yielding new insights [2, 3]. In this issue of Drug Safety, Burningham et al. [4] report their assessments of the performance of CCO methods for studying more chronic effects of drugs, similar to the work of Schuemie et al. [5] on studying longterm effects of accumulated exposures using the self-controlled case-series (SCCS) method. Based on their confidence that the Observational Medical Dataset Simulator version 2 (OSIM2) succeeds in producing a realistic testing platform from real MarketScan data, Burningham et al. infer that estimates of risk ratios from CCO analyses are generally closer to true risk ratios when effects of exposure are more transient and outcome onset is more abrupt than when CCO methods are used to study accumulative effects. This was an underlying assumption of the CCO design when first proposed [1], but such assumptions warrant testing. Based on my prior belief in virtues of narrow windows of observation in CCO analyses [6], I am persuaded that OSIM2 is capable of informative assessments of CCO assumptions. The inventors of OSIM2 incorporated between-person confounding in creating the simulated data for realistic methodologic studies of traditional cohort and case–control designs [7]. However, much, if not all, of that between-person confounding would be automatically eliminated from CCO analyses because cases serve as their own controls. Ironically, the main strength of OSIM2 as a creator of simulation data might thus be negated by the main strength of the CCO design. OSIM2 was not explicitly designed to produce realistic within-person confounding in the simulated data. Murray et al. [7] wrote ‘‘For each simulated distinct drug therapy, the simulation randomly draws ...the calendar duration from the start of the first exposure to the end of the last from the Drug Occurrence Count and Drug Duration Probability table. The simulation then randomly distributes the desired number of drug exposures throughout the simulated total drug duration’’. Such within-person random distribution of drug exposure times is pragmatic but problematic because it would not generate realistic patterns of use and magnitudes of within-person confounding. Indeed, in the paragraphs on limitations of the simulation study, they wrote ‘‘Data characteristics that are not explicitly described in the empiric distribution tables may not be accurately represented within the simulated datasets, such as co-prescription of drugs’’. Still, I agree strongly with the overall conclusion of Burningham et al.: ‘‘Biological plausibility, unique characteristics of the drug and condition pair under study, and one’s beliefs or assumptions about the distributional relationship between the drug and condition should play a considerable role in optimizing the CCO specifications’’. When they recommend that ‘‘Careful consideration must be given to the hypothesized drug exposure and outcome distribution’’, I would be more explicit: ‘‘careful This comment refers to the article available at doi:10.1007/s40264017-0540-3.