Feasibility and reliability of using hazard ratios in meta-analyses of published time-to-event data

Objectives

Meta-analyses of published time-to-event outcomes usually rely on estimating odds ratios (ORs), yet the best statistic to use is the hazard ratio (HR). Meta-analyses presenting HRs are mostly those of individual patient data (IPD). However, HRs given in trial reports can be used directly or, if sufficient summary statistical information or Kaplan-Meier curves are presented, then HRs can be estimated indirectly (Parmar et al.). The latter works well for simulated trial data (D'amico et al.), but it isn't clear how these methods perform in the real world. We aimed to assess retrospectively, the feasibility of generating HRs from published data, for existing IPD meta-analyses in cancer and then, compare meta-analyses of HRs derived from published data (PD) with meta-analyses of HRs from IPD.

Methods

Initially, a meta-analyses of post-operative radiotherapy for lung cancer (PORT) and of chemotherapy for soft tissue sarcoma (STS) were selected. All trials included in these projects were screened to determine whether HRs could be obtained from the PD. For each project, where feasible, HR estimates were obtained and a meta-analysis of the trials carried out. Then, IPD meta-analyses of these same sets of trials were performed. The results of the meta-analyses of PD, the equivalent IPD meta-analyses (of the same trials) and the full IPD meta-analysis (of all identified trials) were compared.

Results

Nine trials had been identified for the PORT project, but a meta-analysis of HRs from PD was possible for only 5. Fourteen trials had been identified for the STS project, but a meta-analysis of HRs from PD was possible for only 8. Trials could not be included because they were unpublished or, published with insufficient detail, as part of a combined analysis or across 2 publications. For PORT, the pooled HR from PD and equivalent HR from IPD were very similar, although somewhat different to the full IPD meta-analysis, in terms of confidence in the estimate. By contrast, the results of the STS meta-analysis of PD were quite different from both the equivalent and full IPD meta-analysis.

Conclusions

Where events happen quickly and so publications report quite mature data and patient exclusion is modest, e.g. the PORT comparison, then HRs from PD and IPD were very similar, although the PD was less convincing. However, where events happen over a prolonged period, such that publications report immature data, and many patients are excluded, e.g the STS comparison, then the HR of PD was a poorer approximation of both its IPD equivalent and the full IPD analysis. Furthermore, meta-analyses of PD that use HRs, cannot circumvent the other potential biases associated with this data (publication, patient exclusion, selective reporting and follow-up bias). Further comparisons of HRs derived from published versus individual patient data will be presented.

Have something to say? Post a comment on this article!