Re: SARS-CoV-2 vaccination and myocarditis or myopericarditis: population based cohort study
I appreciate the authors’ prompt reply and providing the reference (https://cran.r-project.org/web/packages/survival/vignettes/timedep.pdf). First, I’d like to point out that the authors could have made it clear that almost everyone in the study population contributes to the unvaccinated group in the article. Had they done that, and provided information about those who did not get the two vaccines, this conversation could be very short.
The reference the authors provided used an example of the advanced lung cancer data set to illustrate how to make an invalid analysis by generating random response and fit a Cox model. This is where the authors’ and my opinion differ. I feel this study made the same error as illustrated by that example. In this study, the responders, i.e. those who had an event in the unvaccinated period, were unlikely to be in the vaccinated groups. The selection bias is stronger for those with more severe or fatal outcomes like cardiac arrest or death (there were also more cases of this outcomes); this selection bias is also pointed out by other reviewers. The selection bias for participants with myocarditis or myopericarditis is probably more subtle as the number of the events is much smaller and the consequences of the events might not be as severe.
To address my concern about omission of sample size in the self-controlled case series (SCCS) analysis, the authors said, “in SCCS, only individuals who have experienced an event are included… The number of subjects (i.e., sample size) is the same as the number of events.” I thank them for their clarification. However, it is still not clear to me, of those 145 unvaccinated individuals, how many got BNT162b2 or mRNA-1273, and for each vaccine, how many had myocarditis or myopericarditis events. Similarly, for example, of those 43 subjects with events after receiving the BNT162b2 vaccine, how many had the events in their unvaccinated period. (Essentially, a table as in McNemar's test would be able to provide us a clear picture. Of course, in this table, the number of cases of no event before AND after vaccination will be zero by the definition of SCCS) . I think this is more relevant information to fit a conditional logistic regression. Had they provided those numbers, I could use the numbers to do a quick check to understand the model better.
The authors mentioned conditional logistic regression for SCCS studies. While I don’t have experience with this method for this kind of study, i think odds ratio, not rate ratio, is used in logistic regression. The authors seem to have used the two terms interchangeably. This may also have caused all the confusion. While for rare events in a (relatively) big population, the estimations of odds ratio and rate ratio could be similar, these are two different concepts and therefore should be used accordingly.
I want to point out the authors also mentioned about Firth's correction in Figure 2 and they made it clear that they used conditional logistic regression in their reply. As I understand, Firth logistic regression is mainly used for modelling rare events. I am not sure how to combine it with conditional logistic regression when the strata are pairs (https://en.wikipedia.org/wiki/Conditional_logistic_regression). I realize this becomes quite technical, but it is essential to understand exactly how the model was fitted.
In summary, I think this paper has raised more questions than it has answered.
Competing interests: No competing interests