Effects of adjusting for censoring on meta-analyses of time-to-event outcomes.

BACKGROUND Systematic reviews of published time-to-event outcomes commonly rely on calculating odds ratios (OR) at fixed points in time and where actual numbers at risk are not presented. These estimates are usually based on the total numbers included in the published analysis and take no account of censoring. We have assessed the impact of adjusting for censoring on weighting, estimates and statistical heterogeneity of meta-analyses in cancer. METHODS Meta-analyses of survival data for five meta-analyses of published trials in cancer were conducted. The OR and associated statistics were calculated based on unadjusted total numbers of participants and events. These were compared with calculations that first adjusted the numbers at risk for censoring using a simple model. RESULTS Pooled OR were changed in 17/24 cases. On average, there was a 2.6% difference between the adjusted and unadjusted OR. Confidence intervals were frequently wider for the adjusted OR. Adjusting also reduced weighting of individual trials with immature follow-up. In 18/24 cases, adjusting reduced statistical heterogeneity and affected the associated P-values. CONCLUSIONS Reviewers conducting meta-analyses of published time-to-event data where actual numbers at risk are not available should adjust the numbers at risk, estimated from total numbers analysed, to account for immature data and censoring.