LAUSR

I welcome this thought-provoking piece by Dr Remiro-Azócar. He is surely right in stressing that we need to think about what we are trying to estimate, that is to say estimands, when comparing approaches. If different quantities are targeted by estimates, there is little point in comparing them in terms of efficiency. There is much that he claims with which I am in agreement. In my discussion I shall stress points in which I am not in complete agreement. That I do so should not be taken as meaning that I am critical of the piece as a whole. The specific context in which he discusses estimands is that of indirect comparisons through network meta-analysis. It is not central to my discussion but I take this opportunity to make the point that the evidence based movement, starting in the 1990s, has reinvented, not always well, much statistical theory, that had already been developed 60 years earlier for incomplete blocks analysis.1,2 Similarly, some more recent work on prediction,3 a topic of relevance when using conditional estimates to construct marginal ones, seems to have been largely ignored by biostatisticians, perhaps because the context was agriculture rather than medicine. However, I do not consider that the theme of network meta-analysis is central to the issue. The fact that covariate distributions will vary from trial to trial and that this will complicate the task of estimation for a synthesis based on a network of trials was, of course, the subject of the piece by Phillippo et al4 that provides the background to Remiro-Azócar’s analysis. While not wishing to minimize this technical challenge, a central claim of his,