Methods issues that create statistical issues

[timdisher]

PDA is missing BPD for high dose oral ibuprofen which is the best treatment on the primary outcome. We can impute but hard to justify how. Bets case is multivariate extension but that is beyond scope of this paper and can be discussed as a limitation. Options I've thought of:

PDA

• Use mean SUCRA/OR for the treatment (would have to make sure ORs are in right direction)
  • Assumes treatment maintain it's average effectiveness across all outcomes.
  • Problematic for ORs since it could result in being wildly higher/lower than others for that outcome which will have a big influence on SMAA results.
• Use mean SUCRA/OR for outcome
  • Assumes treatment will be about average for the outcome, simpler.
  • Might underestimate benefit if it was the best on everything else but now is about average.
• Use SUCRA/OR for a physiologically similar outcome
  • Example would be to use SUCRA/OR from NEC when BPD is missing
  • Run into the same issues as average for the treatment in that average NEC OR could be ver different from average BPD OR across alternatives.

I am leaning towards option 2 unless @souvikiwk can think of an alternative?

Steroids

• There is just too many missing treatments here. I think it is best to just say there is a limit to how much we're willing to impute and this is it.

SUCRAs in general

Imputing ORs is one thing... imputing SUCRA is a little worse because it is based on ranking and thus sensitive to the number of alternatives that are missing. Probably best just to list this as a limitation/reason to prefer regular approach.

[timdisher]

Thinking if we are targeting clinical journal it will be cleaner to avoid sucras all together. We can just say they are inappropriate because they don't account for baseline prevalence and inflate differences in relative effects.

Can do a separate paper on that issue with just the PDA results, aim for Clin epi or something