The recent article by Dr Bansal and colleagues addressed a major clinical concern. Using the Surveillance Epidemiology and End Results (SEER) Tumor Registry (1988-2005), survival was compared between women with stage IB-IIA cervical cancers treated with either primary radiotherapy or radical hysterectomy. Women treated with radical hysterectomy as compared to primary radiation had a 59% reduction in cancer-specific mortality rate (hazard ratio [HR], 0.41; 95% confidence interval [CI], 0.35–0.50). Large databases like the SEER Tumor Registry and the Medicare database are widely available for population-based cancer outcome studies. Advantages of these data include the excellent external validity and the facility to study populations that usually are not enrolled in clinical trials like minorities, the elderly, and individuals with coomorbidities. However, survival analysis among individuals treated with different cancer therapies using observational data may pose major biases. Selection bias are likely to distort the results and affect the validity of these studies. Control for confounding factors available in large databases may improve the validity, but unmeasured factors can persist. A recent study found that men with localized prostate cancer with active treatment (either radical prostatectomy or radiation) had significantly lower prostate cancer-specific mortality (HR, 0.64; 95% CI, 0.57–0.73) compared with patients on observation. However, these patients also had lower mortality from all other causes (HR, 0.68; 95% CI, 0.65–0.71). The mortality benefit associated with active treatment from other causes, such as diabetes or pneumonia, was similar to the benefit found for prostate cancer. The same study also found that men with locally advanced prostate cancer treated with radiation and adjuvant androgen therapy as compared to individuals treated with only radiation had a higher cancer-specific mortality rate (HR, 1.63; 95% CI, 1.32–2.01), contradicting previous clinical trial data. To understand potential selection bias and implausible results, at a minimum the study should examine the treatment effects on mortality from cancer vs other causes. Additional statistical analyses including competing risk analysis and instrumental variable analysis should be conducted to control unmeasured confounders. It is likely that the performance of these analyses may elucidate whether biases do or do not have strong effects on the findings reported by Dr Bansal and colleagues. We acknowledge that observational data turn out to be the best source of cancer therapy outcome studies of large populations; however, the results should be interpreted with caution knowing the presence of potential biases and the likelihood for improbable results.