Is it better to meta-analyze number of events or adjusted summary data (i.e. aHR)?

Question

This question is not related to any example in particular - it is more like a "methodological" matter, which I feel may be useful to discuss.

If we are conducting a meta-analysis on the comparative efficacy of two treatment, we can do two different things:

• Meta-Analyze the number of events and total number of patients in each group
• Meta-Analyze the log-adjusted summary data that comes from the original studies included (i.e., adjusted Hazard Ratio, or adjusted Odds Ratio).

While the former method is more "straightforward", one can question about the opportunity to combine non-adjusted data from different studies - so using the adjusted HR/OR would be better.

On the other side, I already know that we cannot easily combine together logHR and logOR (they are not interchangeable). But what are the other caveats to this approach?

Answer 1

In general terms, if the included studies are all randomized controlled trials (RCTs), in keeping with most experts, I recommend to extract raw data and analyze them appropriately (eg with odds ratios [OR], risk ratios [RR], risk differences [RD], incidence risk ratios [IRR], and so forth).1 This ensures transparency and reproducibility, and let you apply many different models for sensitivity purposes. Adjusted effect estimates can however prove useful occasionally, for instance when pooling HR after landmark analyses.2

Conversely, if you are pooling also non-RCTs, then raw data are useful mainly for descriptive purposes (unless stemming from propensity-matched samples), and thus only OR or hazard ratios stemming from adjusted models are inferentially useful (but still potentially confounded by unmeasured variables).3