OBJECTIVES: In the absence of OS data from RCTs, predicted significance of OS benefit plays a vital role for regulatory approval of anticancer agents and clinical guideline developments. We proposed LoR for a direct prediction of the significance of OS from SEs and demonstrated the statistical utility of our approach over indirect prediction via linear regression (LiR).

METHODS: In LoR, predicted outcome was the odds of observing a significant OS whereas in LiR it was the OS hazard ratio (HR). Significance of OS was defined as a binary outcome using the estimated upper bound of the 95% CI of its HR. Independent variables were the HRs of SEs (or their natural logarithms). Both models weighted RCTs with their corresponding sample sizes. Leave-one-out-cross-validation was employed to compare the predictive performances of both methods in a case study consisting of previously published 35 instances from 30 distinct meta-analyses in 13 different cancers. SEs were progression-free survival or its analogues in 22 instances, disease-free survival in 12 instances and metastasis-free survival in 1 instance. Across all instances, there were 556 RCTs.

RESULTS: LiR and LoR correctly predicted the significance of OS in 64% and 80% of all RCTs in the case study, respectively. Predictive accuracy of LoR was higher (lower) than LiR in 26 (4) meta-analyses. Among the 8 meta-analyses with <10 RCTs, LiR and LoR correctly predicted the significance of OS in 53% and 72% of a total of 72 RCTs, respectively. Among the 15 meta-analyses with >15 RCTs, LiR and LoR correctly predicted the significance of OS in 64% and 81% of a total of 341 RCTs, respectively.

CONCLUSIONS: LoR can serve as a robust and accurate alternative to LiR in predicting the significance of OS particularly for meta-analyses including large number of RCTs with a balanced sample of OS HRs with potential outliers.