Radiomics is referred to as quantitative image biomarkers for medical image analysis. Recently, Welch et al. and Traverso, Kazmierski et al. showed that radiomics' predictive power might be caused by its correlation with tumor volume. In this work, we investigated how tumor delineation affects the radiomic values and their prognostic power.Data. Three hundred sixty-one patient CT images from the Lung1 dataset [https://xnat.bmia.nl] and 210 patient images from the Lung2 dataset were used in the study. The gross tumor volume (GTV) ROI of each patient was eroded by applying eight times a kernel of size one. The GLCM, GLSZM, and GLRLM features were extracted from the original and the eroded images.The feature-volume relations were assessed by calculating the Spearman correlation for each patient using the eroded ROIs. Based on mean(|ρSpearman|) < 0.3 criteria, we selected five features. We built two models to stratify the patient into two survival groups and generated their survival curves with the Kaplan-Meyer estimator: 1) the first model was fit with features calculated from the original ROIs; 2) the second model was fit with features from the eroded ROIs. We used the log-rank test to evaluate the significance of the stratification, calculating their P-values.The stratification was significant in Lung1 and Lung2 datasets, with a P-value of 0.02 and 0.05 using the features extracted from original ROIs, however there were multiple intersection make the P-values inaccurate. Instead, the stratification was significant in Lung1 and Lung2 with a P-value of 0.03 and less of 0.005 using features extracted from eroded ROIs.Radiomic features extracted from routinely delineated GTVs can be biased by the tumor-air edge in CT. Our results show that features extracted from eroded GTVs can achieve better results than the features from the original GTVs. In addition, five features used in our analysis were selected to be unbiased by tumor volume.