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.
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.
Authors: Jørn Wetterslev; Christian S Meyhoff; Lars N Jørgensen; Christian Gluud; Jane Lindschou; Lars S Rasmussen Journal: Cochrane Database Syst Rev Date: 2015-06-25
Authors: Lars H Lundstrøm; Christophe Hv Duez; Anders K Nørskov; Charlotte V Rosenstock; Jakob L Thomsen; Ann Merete Møller; Søren Strande; Jørn Wetterslev Journal: Cochrane Database Syst Rev Date: 2017-05-17
Authors: Luit Penninga; André Wettergren; Colin H Wilson; An-Wen Chan; Daniel A Steinbrüchel; Christian Gluud Journal: Cochrane Database Syst Rev Date: 2014-06-05