Network-metaanalysis query - Ranking analyses of "negative" outcome e.g. death and final summary statistics

[Dil Singh]

Dear all

I am working on learning network meta-analyses in Stata based on Ian White's paper in the Stata Journal (http://www.econ.uzh.ch/dam/jcr:75a2c...15.pdf#page=57) and Chaimani et al (http://journals.plos.org/plosone/art...l.pone.0076654).

However, I have two queries.

The first is if the data we have is looking at a "negative" outcome e.g. death, how should this be handled when using the "network rank" command to rank treatments from best to worst. In the case of death, I would imagine the "best" (namely the treatment associated with the highest number of outcomes i.e. deaths) would actually be the "worst"?

This is using:
. network rank max, line cumulative

The second query is whether the "rank" command and forest plots are the final outcomes for network meta-analyses. These are the limit of what are described in the above two papers. There does not appear to be a summary "odds ratio" as there is for pairwise meta-analyses, so wanted to check I was not missing anything here.

This is using:
.network forest, diamond eform

Many thanks in advance.

Regards, Dil Singh

[Federico Tedeschi]

The first is if the data we have is looking at a "negative" outcome e.g. death, how should this be handled when using the "network rank" command to rank treatments from best to worst. In the case of death, I would imagine the "best" (namely the treatment associated with the highest number of outcomes i.e. deaths) would actually be the "worst"?

This is using:
. network rank max, line cumulative

If "network rank max" orders treatments using the wrong direction, then "network rank min" should work.

The second query is whether the "rank" command and forest plots are the final outcomes for network meta-analyses. These are the limit of what are described in the above two papers. There does not appear to be a summary "odds ratio" as there is for pairwise meta-analyses, so wanted to check I was not missing anything here.

Yes, SUCRA doesn't inform either about point estimates of relative treatment effects, nor about their statistical evidence, and p-best even worse, since it may reward treatments whose effect estimate has the highest variance (see, for example, here: https://pubmed.ncbi.nlm.nih.gov/32819946/ and here: https://discovery.ucl.ac.uk/id/eprint/10150504/1/White_The%20complexity%20underlying%20treatment%20 rankings.%20How%20to%20use%20them%20and%20what%20t o%20look%20at_AOP.pdf).
And yes, forest plots just compare each treatment with a one chosen as reference. You can however build a forest plot for how many reference treatments you like, or a net league (see the "netleague" command of the "network graphs" package: https://www.stata-journal.com/articl...article=st0411), that could even allow you to show results on two outcomes simultaneously.