OBJECTIVES : Hazard ratios (HR) are considered the most appropriate statistic for meta-analysis of time dependent outcomes, but HRs may not be provided in published literature. We evaluated whether it is feasible to use the Guyot Algorithm (recommended by the Cochrane handbook for this purpose) to approximate the HR from Kaplan-Meier curves (KM) under a variety of conditions.

METHODS : We simulated data for two treatment groups with 300 subjects in each group. HRs were varied from 0.5 to 2.0, censoring rates were varied from 50% to 97%, and the number of time points with N at risk (Nrisk) was varied from 0 to 5. KM survival estimates and Nrisk were exported from each simulated data set and used as input for the Guyot algorithm. The HR was calculated in both the simulated and the reconstructed individual patient data using Cox-Proportional Hazards. The difference between the simulated and reconstructed data was calculated to quantify the accuracy of the Guyot algorithm and was summarized using the inter-quartile range (IQR). The algorithm was executed and validated in R and SAS.

RESULTS : Varying the proportion censored, the difference between actual and reconstructed HR was IQR [-.05 to .13]. The algorithm could not be used for 97% censoring rate due to insufficient number of events. Varying HRs, the IQR was [-.04 to .03]. When no information for Nrisk was given, the IQR of the difference was [-.07 to .07]. With ≥1 time point available, the IQR of the difference was [-.02 to .02].

CONCLUSIONS : The Guyot Algorithm can be implemented in SAS or R. It performs well across a range of hazard ratios and for all but the most extreme censoring distributions. The accuracy is greatly improved when ≥1 Nrisk is provided.

Reference

Guyot et al. BMC Med Res Methodol. 2012;12:9.