Should we do a meta-analysis of non-inferiority trials? Since non-inferiority trials are designed meticulously to achieve an effect size demonstrating no significant difference between treatments, so there may be certain essential differences in design of a non-inferiority trial from the design of a typical randomized trial intended to establish a difference between comparators. Also, since the effect sizes in individual trials are not significant (mostly), it is highly likely that pooled estimate after a meta-analysis may not be significant as well. Essentially, pooling non-inferiority data will also result in a non-inferiority estimate, so what is the point of pooling? I accept that I sound more like a skeptic, not evidence-based, but I seek your opinions on my doubts? I tried to search the web if there are any special considerations while conducting meta-analysis of non-inferiority trials, but could not find any (seems my doubts are groundless?)
 A: This is an interesting question... and some scholars have tried to address this issue in the past:
https://www.ncbi.nlm.nih.gov/pubmed/15702203
https://www.ncbi.nlm.nih.gov/pubmed/20837637
https://www.ncbi.nlm.nih.gov/pubmed/21113051
https://academic.oup.com/biostatistics/article/13/4/637/240843
My take is that meta-analysis is akin to a retrospective observational research, you can always pool non-inferiority, equivalence, or superiority studies, but only with a superiority aim. Indeed, you cannot usually stipulate you defined pre hoc any non-inferiority or equivalence threshold before your meta-analysis.
Very rarely and early on (eg in case you work for a pharmaceutical company and are planning a series of phase I, II and III trials) you could pre hoc design prospectively a non-inferiority meta-analysis, but I assume the credibility and impact of such work would be small.
In any case, the best way to exploit any meta-analysis is by using confidence intervals of effect estimates (within a classical or Bayesian framework). Accordingly, you can always post hoc estimate then distance of your actual effects from a given threshold of choice.
