Robustness assessments are needed to reduce bias in meta-analyses that include zero-event randomized trials

Abstract

Objectives: Meta-analysis of randomized trials with binary data can use a variety of statistical methods. Zero-event trials may create analytic problems. We explored how different methods may impact inferences from meta-analyses containing zero-event trials.

Methods: Five levels of statistical methods are identified for meta-analysis with zero-event trials, leading to numerous data analyses. We used the binary outcomes from our Cochrane review of randomized trials of laparoscopic vs. small-incision cholecystectomy for patients with symptomatic cholecystolithiasis to illustrate the influence of statistical method on inference.

Results: In seven meta-analyses of seven outcomes from 15 trials, there were zero-event trials in 0 to 71.4% of the trials. We found inconsistency in significance in one of seven outcomes (14%; 95% confidence limit 0.4%-57.9%). There was also considerable variability in the confidence limits, the intervention-effect estimates, and heterogeneity for all outcomes.

Conclusions: The statistical method may influence the inference drawn from a meta-analysis that includes zero-event trials. Robustness assessments are needed to reduce bias in meta-analyses that include zero-event trials.