Solved – Peto vs Mantel-Haenszel fixed effect odds ratios for meta-analysis

meta-analysisodds-ratio

We are writing a protocol for PROSPERO for a pairwise meta-analysis comparing comparing two cardiovascular treatments in terms of their effects on a dichotomous endpoint, such as death. We expect to include 5 o 6 homogeneous trials, eventually (thus most likely statistical inconsistency will be minimal, e.g. I-squared will be 0).

We are uncertain as to whether opt for the Peto method to compute odds ratios, or the Mantel-Haenszel fixed-effect method.

It appears the Peto method might be less robust, as it relies on more assumptions (eg Khera et al), but the Cochrane Collaboration considers both Peto and Mantel-Haenszel approaches as fine (see for instance section 9.4.4 Meta-analysis of dichotomous outcomes of the Cochrane Handbook for Systematic Reviews of Interventions).

Any suggestion of which method to prefer for the primary analysis in the protocol (leaving the other one for sensitivity analysis)?

Best Answer

For a start, using time-to-event methods on the individual patient data is the gold standard approach and one should try to get it. It is important to realise that assuming a binomial distribution means that the losses to follow-ups in the arms being compared need to follow the exact same distribution. If the number of losses to follow-up are similar and there is no suspicion of different time patterns in drop-outs, a binomial distribution may be reasonable and okay for detecting an effect. Conventional wisdom 10 or so years ago was that

  1. The Peto-Yussuf one-step odds ratio has a relatively good reputationfor rare events (with some potential problems for unbalanced randomization).
  2. For slightly less rare events, the Mantel-Haenszel fixed-effects odds-ratio (however, note: zero-cell corrections are a difficult topic, because you do not want to ignore trials with some zero cell, but a 0.5 correction biases the result too much towards no effect) and logistic regression (standard, exact or Firth penalized likelihood logistic regression) have also been reported to be relatively good methods.

This is, I believe what the Cochrane handbook bases its recommendations on. However, this is an area of ongoing research and papers are regularly appearing proposing new methods in particular with respect to trials with no (or very few events) e.g. this one last year. For a particular problem, I would have an in-depth discussion with people familiar with the disease area and statisticians specializing in meta-analysis to pick a reasonable method that should perform decently for the particular question.

Related Question