Response: The risk of Stevens-Johnson syndrome and toxic epidermal necrolysis in new users of antiepileptic drugs: Comment on data sparsity.

Abstract

Dear Editor-in-Chief We have read the comment on our article with great interest, and we would like to thank Dr Safiri and Dr Ashrafi-Asgarabad for their interest in our study. However, we do not think that the suggested penalized odds ratios (ORs), which were quantified using data augmentation, accurately reflect the association between Stevens-Johnson syndrome (SJS) and aromatic antiepileptics. Drs Safiri and Ahran-Asgarabad deemed our presented ORs for the association between SJS and carbamazepine (OR 92.57, 95% confidence interval [CI] 19.89-∞), phenytoin (OR 49.96, 95% CI 10.13-∞), and lamotrigine (OR 26.90, 95% CI 4.88-∞) as too extreme and biased by sparse data, and instead presented penalized ORs (eg, OR 6.05, 95% CI 1.46-25.01 for phenytoin), which they quantified using the data augmentation method. Although Greenland et al emphasized that the method of data augmentation strongly relies on preexisting evidence, the authors do not report on the basis of which existing evidence their penalization priors were chosen or which 95% prior intervals were applied, which prevents us from accurately interpreting the suggested results.1, 2 SJS is a rare disease that is caused mainly by new use of a few specific drugs, of which aromatic antiepileptics by far bear the highest risk of triggering SJS. We previously calculated absolute risks of SJS/toxic epidermal necrolysis in new users of trigger drugs other than antiepileptics of 1-6 cases/100 000 new users (one study not published yet),3 whereas the absolute risk among new users of aromatic antiepileptics was 20-45 cases/100 000 new users.4 Thus, the expected relative risk estimates for SJS in association with new antiepileptic use can be expected to be very high, which was also suggested in the comprehensive hospital-based EuroSCAR case-control study, which reported ORs of 72 (95% CI 26-225) for carbamazepine and 26 (95% CI 7.8-90) for phenytoin.5 We agree that extremely high relative risks from observational studies need to be interpreted carefully. However, in the absence of any prior evidence suggesting lower ORs, using a penalization prior which artificially corrects ORs toward the null might provide a false sense of certainty. We therefore think that the more conservative approach we chose is the method of choice here, whereby the wide confidence intervals indicate the level of uncertainty due to small sample size. Furthermore, given the low numbers of exposed patients in our study population, which is an inherent problem when studying rare diseases like SJS, we took several precautions to avoid sparse data bias. First, we refrained from conducting multivariable adjustment of our ORs (confounding is not a major issue when studying SJS), but instead matched cases and controls on age, sex, and index date as suggested by Greenland et al and quantified the proportion of patients who were concomitantly exposed to other high-risk drugs. Furthermore, we conducted exact logistic regression whenever a zero cell was observed to avoid sparse data bias.2 In conclusion, we agree that data augmentation is a valid new method to avoid sparse data bias, but we do not necessarily agree that this method should have been used in our study. None of the authors have any conflict of interest to declare. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.