Reply to the Letter to the Editor “Toward a Finer‐Grained Understanding of Greater‐Occipital‐Nerve Block Duration in Migraine”

We would like to reply to the letter by Lan 1 about our paper “Greater Occipital Nerve Block as preventive treatment in migraine: a time-to-event analysis” 2, discussing criticisms in a methodological rather than counter-argumentative manner, hoping to be helpful to physicians and benchside scientists. One lesson I learned in statistics is: “While math starts with axioms, statistics starts with common sense”. In our view, this principle is particularly relevant in the present context. Although some of the points raised by Lan are reasonable, they seem more like a matter of stylistic refinement than a truly decisive factor. The main counterargument addressed in this letter appears to concern the potential risk of biased results due to a high proportion of censored data. However, this concern is unlikely to apply to our paper, given that only 15% of the measurements were censored. Moreover, Kaplan–Meier is considered reliable up to 50% of censored data 3. Instead, we would like to discuss the other two points raised by Lan. Lan suggested using penalized Cox instead of the usual Cox regression to minimize overfitting since we analyzed eleven predictive baseline variables with 75 individuals as outcome observations. It could be correct in principle, but it hides methodological shortcomings. While, according to the seminal article by Topliss 4, 65 observations are enough for analyzing up to 20 variables, we would rather like to focus on the logic behind what we did. Let's assume that there is an overfitting problem. Indeed, using too many variables with too few observations increases the chances of overfitting (i.e., when the model works too well), namely a type-I error. This means that some spurious associations are mistakenly taken for significant and kept in the model by chance. However, we didn't consider that scenario since we reported a list of negative results without any significant predictor of an effect. In this case, this means that none of the variables included had predictive value despite the fact of a possible significance augmentation, reinforcing the robustness of our negative results. Moreover, these negative results mirror the current literature and the clinical experience of any physicians performing GON block as migraine prophylaxis of a want in predictive factors. By contrast, if significant results had resulted, we would have taken into account the overfitting and used countermeasures (possibly also penalized Cox model) and/or discussed it as a limitation. Similarly, the suggestion of using an adaptive Bayesian technique to use the dose of steroid presents conceptual concerns. In brief, we pinpoint that there is no current consensus on the correct nature and the dosage for both local anesthetic and steroid in the procedure 5, 6, and this answer will derive from ad-hoc clinical trials. In conclusion, statistics should be regarded as a toolbox by scientist-physicians and researchers, who should select the most appropriate statistical test based on their knowledge of the study subject and the logic and properties of the statistical method, rather than on the latest update or a newly introduced model. A.V. was primarily responsible for the statistical analysis in our previous manuscript, to which this letter refers. All listed authors contributed to the final version of the manuscript and agreed to the final submitted version. The authors have nothing to report. The authors have nothing to report. The authors have nothing to report. The authors have nothing to report. The authors declare no conflicts of interest. Data sharing is not applicable to this article as no new data were created or analyzed in this study.