Net survival is the survival that would be observed if the only possible underlying cause of death was the disease under study. It can be estimated with either cause-specific or relative survival data settings, if the informative censoring is properly considered. However, net survival estimators are prone to specific biases related to the data setting itself. We examined which data setting was the most robust against violation of key assumptions (erroneous cause of death and inappropriate life tables). We identified 4285 women in the Geneva Cancer Registry, diagnosed with breast, colorectal, lung cancer and melanoma between 1981 and 1991 and estimated net survival up to 20 years using cause-specific and relative survival settings. We used weights to tackle informative censoring in both settings and performed sensitivity analyses to evaluate the impact of misclassification of cause of death in the cause-specific setting or of using inappropriate life tables on net survival estimates in the relative survival setting. For all the four cancers, net survival was highest when using the cause-specific setting and the absolute difference between the two estimators increased with time since diagnosis. The sensitivity analysis showed that (i) the use of different life tables did not compromise net survival estimation in the relative survival setting, whereas (ii) a small level of misclassification for the cause of death led to a large change in the net survival estimate in the cause-specific setting. The relative survival setting was more robust to the above assumptions violations and is therefore recommended for estimation of net survival.