OBJECTIVES

: Meta-analyses involving time-dependent outcomes can be challenging, particularly when the statistical presentation of outcomes of interest are heterogeneous and thus difficult to pool. This methodology illustrates a hierarchical decision tree for calculating and combining hazard ratios (HR) to maximize the data included.

METHODS

The hierarchy was based on the number of assumptions required. Method 1 entailed using the reported HR, when available. Method 2 involved estimating the HR using equations outlined by Parmar {Parmar 1998} from the reported number of events, cohort totals and log-rank p-value. Method 3 used the Guyot {Guyot 2012} algorithm to reconstruct individual patient data from Kaplan-Meier survival curves and then applied the Parmar formulae. Method 4 approximated the number of events assuming no censoring from the Kaplan-Meier proportions and then applied Parmar. The resultant HRs were pooled for overall survival (OS), disease/progression/recurrence-free survival (DFS), and recurrence.

RESULTS

: A meta-analysis of robotic, laparoscopic and open surgery comparative papers was performed to assess long-term cancer outcomes across four surgical procedures. Long-term outcomes were reported in 198 articles. Use of Methods 2-4 increased the available HRs from 112 (Method 1) to 300 HRs (Methods 1-4), facilitating an increase of 168% in the included outcome assessments. Outside of the hierarchy, when multiple methods could be used, the resulting HRs were found to be consistent.

CONCLUSIONS

: A hierarchical approach to extracting data for time-dependent outcomes allows for the largest possible inclusion of results and improves the quality of meta-analyses.

References:

Parmar et al. Stat Med. 1998: 17(24): 2815-2834.

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