Absolute or relative effects? Arm-based synthesis of trial data.

S Dias,A E Ades

doi:10.1002/jrsm.1184

Abstract

Absolute or relative effects? Arm-based synthesis of trial data.

Highlights

We congratulate Hwanhee Hong and colleagues on another fascinating paper (Hong et al, 2015a) arguing the case for arm-based models for meta-analysis
We are not sure whether proponents of “Arm Based” (AB)” models would agree, but to some extent the choice between armand contrast based models could be regarded as an empirical question: do we find in practice that relative effects, given that an appropriate scale has been chosen, are more stable than absolute effects? This is the kind of question that has been studied previously using L’Abbé plots (L’Abbe et al, 1987, Song, 1999)
We believe that AB models are thoroughly misguided, and a huge step back from the separation of absolute and relative effects that has been at the core of modern epidemiology and biostatistics for so long

Summary

Missing-ness of treatment arms in NMA

The key assumption in all the CB models is the exchangeability of the trial-specific treatment effects δi,XY across the entire ensemble of trials. In previous papers (Caldwell et al, 2005, Dias et al, 2011) we have said that this was equivalent to the arms being “missing at random” (MAR). By this we meant that missing-ness was unrelated to any factors that influenced the relative treatment effects δi,XY. The only requirement for CB models is that missing-ness is unrelated to the relative effects, which is related to the requirement that the trial-specific relative effects are exchangeable (Lu and Ades, 2009). Covariates can readily be added to AB models, but there appears to be no way of adding a covariate to a relative effect, or putting different covariates on relative and absolute effects

Simulation studies

Conclusions