Hi Pedro

I wouldn't recommend using it even for pairwise meta-analysis.

The method is adding studies that are just not there, and hypothetises on their effect and sample size. It is like making up data. Also the result of the method is based on the magnitude of the effect size. Additionally, the method places too much emphasis on statistical significance, instead of focusing on the actual effect and its uncertainty.

I think, though I have not read the original publication, the method was not developed as a method for detecting publication bias. It tells you how many null studies are needed to get a null pooled effect, not how many studies with null results went undetected.

Tests for small-study effects exist in NMA. For example the comparison-adjusted funnel plot that is available both in R (netmeta) and Stata (check the work of Chaimani). You should be careful that you now have many comparisons and you need to define the direction of small studies in all comparisons. For example, in AB comparisons, small studies may favor A but in AC comparisons they may favor C. In pairwise meta-analysis we assume that small studies always favor the experimental treatment.

Hi Pedro Lopez ,

As far as I know, either Rosenthal's, Orwin's, or Rosenberg's fail-safe tests are not currently recommended for meta-analysis (see Cochrane handbook here https://handbook-5-1.cochrane.org/chapter_10/10_4_4_3_fail_safe_n.htm). I would recommend performing publication bias assessment with funnel plot and Egger's and/or Begg's test. You can perform both of them by using metabias() command in R.

I hope this helps.

Best,