And variance from the Kaplan eier curves.6. Statistical analysisAnalyses were performed in intention-to-treat (ITT) population. We first tested the statistical heterogeneity between trials (meaningful differences between studies) using the chi-squared 22948146 Q-test based on the fixed-effect model. The clinical trials were considered heterogeneous when the P value of the chi-squared Qtest was less than 0.10, or when I2 was greater than 50 . When the analyses showed heterogeneity between different clinical trials, a random effect model was applied to accommodate theheterogeneity [20]. The pooled odds ratios for response rate (ORORR), HRs for PFS and OS (HRPFS or HROS) were calculated. We decided to present three primary measures 
to show the treatment effect from different angles because PFS and OS can better describe the efficacy of a targeted drug than response rate. In addition, it is not uncommon to detect discrepancy between a clear benefit in PFS and a vague benefit in OS for lung cancer patients [21?3]. Furthermore, we estimated and tested the difference of treatment effect between bevacizumab combined with chemotherapy and other targeted drugs using the meta-regression model. The crude and riskadjusted 95 confidence interval were reported when the models included/excluded patient characteristics. To demonstrate whether the progression free survival was associated with stable disease (SD) or objective response rate (ORR) to the medication, or both, we performed the additional analysis of logarithm transformed outcomes (HRPFS) against use of bevacizumab and ORORR, controlling for patient characteristics (median age, mean ECOG performance score) and study design (chemotherapy type for theThe Efficacy of Bevacizumab for Advanced NSCLCFigure 4. Response rate, PFS, OS of Bevacizumab versus other targeted drugs in EGFR untested NSCLC patients. doi:10.1371/journal.pone.0062038.gcontrol group). Similarly, logarithm transformed HROS was modeled against HRPFS and bevacizumab. In addition to the above tests, we performed imputation study to test the influence of each individual study using the leave-one-out strategy [20]. Finally, we performed the Ergocalciferol site funnel plot as well as Begg’s and POR8 cost Egger’s tests to examine potential publication bias. We performed subgroup analysis in this study based on patient treatment status using the meta-regression models. Chemothera?py-naive patients were defined as those with no prior chemotherapy and no previous treatment with EGFR-targeted 15755315 drugs or monoclonal antibodies. Previously-treated patients were defined as patients progressed or recurred after at least one previous chemotherapy regimen. All the analyses were performed using STATA 11.0. The study was written according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [24].ResultsThe flowchart of our study is shown in Figure1. From 1,329 published papers and abstract that we found, 967 were excluded from this study based on our inclusion/exclusion criteria. In addition, 309 articles were 
further excluded if they were already review papers or comments. Among the 53 articles that were left from the above exclusion criteria, five articles were excluded since they were duplicate reports. Finally, 15 additional articles were excluded since they did not report outcomes relevant to our study. Our final sample included 15,650 patients collected from 30 randomized clinical trials. Among the 30 multi-center randomized clinical trials [9.And variance from the Kaplan eier curves.6. Statistical analysisAnalyses were performed in intention-to-treat (ITT) population. We first tested the statistical heterogeneity between trials (meaningful differences between studies) using the chi-squared 22948146 Q-test based on the fixed-effect model. The clinical trials were considered heterogeneous when the P value of the chi-squared Qtest was less than 0.10, or when I2 was greater than 50 . When the analyses showed heterogeneity between different clinical trials, a random effect model was applied to accommodate theheterogeneity [20]. The pooled odds ratios for response rate (ORORR), HRs for PFS and OS (HRPFS or HROS) were calculated. We decided to present three primary measures to show the treatment effect from different angles because PFS and OS can better describe the efficacy of a targeted drug than response rate. In addition, it is not uncommon to detect discrepancy between a clear benefit in PFS and a vague benefit in OS for lung cancer patients [21?3]. Furthermore, we estimated and tested the difference of treatment effect between bevacizumab combined with chemotherapy and other targeted drugs using the meta-regression model. The crude and riskadjusted 95 confidence interval were reported when the models included/excluded patient characteristics. To demonstrate whether the progression free survival was associated with stable disease (SD) or objective response rate (ORR) to the medication, or both, we performed the additional analysis of logarithm transformed outcomes (HRPFS) against use of bevacizumab and ORORR, controlling for patient characteristics (median age, mean ECOG performance score) and study design (chemotherapy type for theThe Efficacy of Bevacizumab for Advanced NSCLCFigure 4. Response rate, PFS, OS of Bevacizumab versus other targeted drugs in EGFR untested NSCLC patients. doi:10.1371/journal.pone.0062038.gcontrol group). Similarly, logarithm transformed HROS was modeled against HRPFS and bevacizumab. In addition to the above tests, we performed imputation study to test the influence of each individual study using the leave-one-out strategy [20]. Finally, we performed the funnel plot as well as Begg’s and Egger’s tests to examine potential publication bias. We performed subgroup analysis in this study based on patient treatment status using the meta-regression models. Chemothera?py-naive patients were defined as those with no prior chemotherapy and no previous treatment with EGFR-targeted 15755315 drugs or monoclonal antibodies. Previously-treated patients were defined as patients progressed or recurred after at least one previous chemotherapy regimen. All the analyses were performed using STATA 11.0. The study was written according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [24].ResultsThe flowchart of our study is shown in Figure1. From 1,329 published papers and abstract that we found, 967 were excluded from this study based on our inclusion/exclusion criteria. In addition, 309 articles were further excluded if they were already review papers or comments. Among the 53 articles that were left from the above exclusion criteria, five articles were excluded since they were duplicate reports. Finally, 15 additional articles were excluded since they did not report outcomes relevant to our study. Our final sample included 15,650 patients collected from 30 randomized clinical trials. Among the 30 multi-center randomized clinical trials [9.
