Background: Surrogate endpoints for overall survival (OS) in advanced non-small cell lung cancer (NSCLC), such as overall response rate (ORR) and progression-free survival, are prone to bias due to crossover and unbalanced post-progression therapy or inconsistent response assessment criteria. This study aimed to evaluate the surrogacy of duration of response (DoR) for OS in phase III trials adjusted for crossover, unbalanced post-progression treatment, insufficient information and inconsistent response criteria. Methods: The analysis was based on systematic literature review and data extraction. The relationship between absolute differences in median DoR and in median OS was assessed using the correlation coefficient (R). Additionally, the relationship of the combination of ORR and DoR, which may be a better surrogate for OS because it captures both the frequency and duration of response, with OS was evaluated. The bias arising from different definitions of ORR and DoR was addressed in a subset analysis on RECIST-based trials and WHO-based trials. Further stratification by reported definition of DoR, either from start of treatment (randomization) or onset of response, was performed. Results: ORR, DoR, and OS values were reported in 20 trials (8,382 patients). The correlation coefficient of DoR with OS was 0.356 (95% CI: 0.000-0.690). 8 trials defined response according to RECIST criteria and 11 trials defined DoR from the first documented response. In these subsets, the correlation was 0.630 (95% CI:0.000–0.924) and 0.572 (95% CI:0.000–0.873), respectively. The correlation coefficient of the combination of DoR and ORR with OS was 0.790 (95% CI: 0.535–0.913). In the RECIST-based trials and trials defining DoR from first documented response, the correlation coefficients of the combination of ORR and DoR with OS were 0.742 (95% CI: 0.079–0.950) and 0.887 (95% CI: 0.614–0.971), respectively. Conclusions: Evaluation of DoR as a surrogate for OS should take into consideration both response assessment criteria and ORR. The adjustment with ORR gives a better estimate of the treatment effect and can be used jointly with DoR as a surrogate endpoint to predict OS benefit.