Article type
Year
Abstract
Background:
The presence of publication bias is often verified in meta-analyses by visual inspection of the funnel plot. Statistical tests may estimate the association between the reported effect size and their standard error (Egger's test), total sample size (Macaskill's test) or inverse of the total sample size (Peter's test). Although these tests have been evaluated for pooling odds ratios, their application may be less appropriate for survival data where censoring influences statistical significance (and thus selective reporting) of the hazard ratio.
Methods:
We propose and evaluate two new publication bias tests for survival data that are based on the total number of events (Test-1) and the total survival time (Test-2). We compare their performance to existing tests in an extensive simulation study where more than 20,000,000 meta-analyses were generated. Here, we varied the true hazard ratio (HR = 0.20-1.00), the number of available studies (N = 10-100), the censoring proportion (cp = 0.10-0.50) and the scale of the hazard distribution. Furthermore, we used a set of predefined reflecting meta-analyses of randomized clinical trials in the medical literature.
Results and Conclusions:
When treatment groups are balanced, simulation results demonstrate that type-1 errors are problematic for Egger's test (averaging from 0.110 for N = 10 to 0.195 for N = 100), but consistently good (around 0.10) for Peter's test and Test-1. The power of all tests was low; for example Test-1 yielded power from 0.112 (for N = 10) to 0.208 (for N = 100). Finally, we compare and discuss the performance of Peter's test and Test-1 in imbalanced treatment groups, and make recommendations for practice.
The presence of publication bias is often verified in meta-analyses by visual inspection of the funnel plot. Statistical tests may estimate the association between the reported effect size and their standard error (Egger's test), total sample size (Macaskill's test) or inverse of the total sample size (Peter's test). Although these tests have been evaluated for pooling odds ratios, their application may be less appropriate for survival data where censoring influences statistical significance (and thus selective reporting) of the hazard ratio.
Methods:
We propose and evaluate two new publication bias tests for survival data that are based on the total number of events (Test-1) and the total survival time (Test-2). We compare their performance to existing tests in an extensive simulation study where more than 20,000,000 meta-analyses were generated. Here, we varied the true hazard ratio (HR = 0.20-1.00), the number of available studies (N = 10-100), the censoring proportion (cp = 0.10-0.50) and the scale of the hazard distribution. Furthermore, we used a set of predefined reflecting meta-analyses of randomized clinical trials in the medical literature.
Results and Conclusions:
When treatment groups are balanced, simulation results demonstrate that type-1 errors are problematic for Egger's test (averaging from 0.110 for N = 10 to 0.195 for N = 100), but consistently good (around 0.10) for Peter's test and Test-1. The power of all tests was low; for example Test-1 yielded power from 0.112 (for N = 10) to 0.208 (for N = 100). Finally, we compare and discuss the performance of Peter's test and Test-1 in imbalanced treatment groups, and make recommendations for practice.