Asking the right question is arguably the hardest thing to do in science, or at least in epidemiology. The question that you want to answer dictates the study design, the data that you collect and the type of analyses you are going to use. Often, especially in causal research, this means scrutinizing how you should frame your exposure/outcome relationship. After all, there needs to be positivity and consistency which you can only ensure through “the right research question”. Of note, the third assumption for causal inference i.e. exchangeability, conditional or not, is something you can pursue through study design and analyses. But there is a third part of an epidemiological research question that makes all the difference: the domain of the study, as is so elegantly displayed by the cartoon of Todays Random Medical News or the twitter hash-tag “#inmice“.
The domain is the type of individuals to which the answer has relevance. Often, the domain has a one-to-one relationship with the study population. This is not always the case, as sometimes the domain is broader than the study population at hand. A strong example is that you could use young male infants to have a good estimation of the genetic distribution of genotypes in a case-control study for venous thrombosis in middle-aged women. I am not saying that that case-control study has the best design, but there is a case to be made, especially if we can safely assume that the genotype distribution is not sex chromosome dependent or has shifted through the different generations.
The domain of the study is not only important if you want to know to whom the results of your study actually are relevant, but also if you want to compare the results of different studies. (as a side note, keep in mind the absolute risks of the outcome that come with the different domains: they highly affect how you should interpret the relative risks)
Sometimes, studies look like they fully contradict with each other. One study says yes, the other says no. What to conclude? Who knows! But are you sure both studies actually they answer the same question? Comparing the way the exposure and the outcome are measured in the two studies is one thing – an important thing at that – but it is not the only thing. You should also make sure that you take potential differences and similarities between the domains of the studies into account.
This brings us to the paper by KA and myself that just got published in the latest volume of RPTH. In fact, it is a commentary written after we have reviewed a paper by Folsom et al. that did a very thorough job at analyzing the role between migraine and venous thrombosis in the elderly. They convincingly show that there is no relationship, completely in apparent contrast to previous papers. So we asked ourselves: “Why did the study by Folsom et al report findings in apparent contrast to previous studies? “
There is, of course, the possibility f just chance. But next to this, we should consider that the analyses by Folsom look at the long term risk in an older population. The other papers looked at at a shorter term, and in a younger population in which migraine is most relevant as migraine often goes away with increasing age. KA and I argue that both studies might just be right, even though they are in apparent contradiction. Why should it not be possible to have a transient increase in thrombosis risk when migraines are most frequent and severe, and that there is no long term increase in risk in the elderly, an age when most migraineurs report less frequent and severe attacks?
The lesson of today: do not look only at the exposure of the outcome when you want to bring the evidence of two or more studies into one coherent theory. Look at the domain as well, as you might just dismiss an important piece of the puzzle.