LAUSR

Schuemie et al1 demonstrate how the case-control design is vulnerable to between-person and within-person confounding, by replicating two case-control studies in healthcare databases: a study of isotretinoin on the risk of ulcerative colitis in the PharMetrics Patient-Centric Database2 and a study of dipeptidyl peptidase-4 (DPP-4) inhibitors on the risk of acute pancreatitis in the Taiwan's National Health Insurance Research Database.3 In summary, they consider that the case-control study is unnecessary in database studies, where all exposure and covariate data are available for the entire cohort, arguing a lack of reproducibility in such designs. Schuemie et al raise an interesting point of debate. Firstly, they highlight that, when focusing on case status, the case-control design hides the question of comparability between exposed and unexposed groups, whereas the cohort type with new-user designs4 makes this comparison explicit. They attempted to design comparative new-user cohort studies but found it impossible to identify appropriate start of follow-up for people not exposed to isotretinoin in the cohort study, an issue referred to as target trial emulation failure.5 Furthermore, Schuemie et al consider that confounding stems from lack of comparability between exposure groups, and not from the lack of comparability between case and control groups. Consequently, trying to adjust for confounding based on a case-control comparison may include causal intermediaries. That led them to report patient characteristics according to exposure status, rather than case status as is usually done. That is ingenious. However, in their example of Crockett's case-control study replication, exposure is defined over 12 months before the index date. Therefore, one should be careful when reporting potential confounders, especially those for which the codes are retrieved in the database over the same 12 months before the index date. This situation will automatically include causal intermediaries. It could be avoided by capturing codes only more than 12 months before the index date, either for exposed or unexposed subjects.6 From there, it is possible to adjust for these potential confounders, without fear of accidentally including intermediary variables. Moreover, they reported biased estimates with the use of negative controls, thus demonstrating that the case-control design is prone to confounding bias arising from differences between the exposed and unexposed groups. Based on this result, I find that case-control designs are severely rejected, especially when authors do not attempt to account for confounding in their case-control replication study. The case-control design has been used to reproduce results obtained from clinical trials (generally known to provide unbiased estimates of outcome risk associated with exposure) and found results close to those found in the initial trial.7 Indeed, such emulated trials with appropriate analyses are proposed as a valuable method in observational studies.8-10 Cleverly, Schuemie et al used a self-controlled design to demonstrate that within-person confounding is accounted for. However, as mentioned in their paper and already discussed elsewhere,11 the differences among estimated from case-control and self-controlled designs arise from inherent differences between the designs and answer questions that are slightly different (“why me?” versus “why now?”). Finally, I think case-control designs still can be of value in database studies, provided the use of appropriate study design and analyses. One can stress the importance of the use of sensitivity analyses12,13 to investigate discrepancies in designs, as well as appropriate reporting of methods and results, following the corresponding STROBE or RECORD reporting guidelines.14-16