Matching has become a popular approach to estimate average treatment effect. However, matching cannot control for unobserved bias. Using Rosenbaum bounding approach, we aim to show how strongly unmeasured variables must influence the selection process to undermine the implication of matching analysis. The Surveillance, Epidemiology, and End Results (SEER) Data is used for the analysis. The SEER-Medicare Database is created by linking Medicare identifiers to SEER patients aged 65+ and all claims collected including hospital, physician and clinic. For each patient, their hospital of care and associated hospital volume is computed. Patients in the high and low volume hospitals are matched in terms of demographic and clinical characteristics. Treatment costs are compared, Rosenbaum bonds estimated and Mantel and Haenszel test statistics is calculated to provide evidence on the degree to which any significance results hinge on unconfoundedness assumption. A volume cohort was constructed consisting of 19,375 female SEER-Medicare patients, aged 65+, suffering an in situ and/or invasive breast cancer during 2003-2005 with surgical treatment performed at 567 hospitals. After the matching, samples were similar in terms of race, comorbidity and adjuvant therapies. Under the assumption of no hidden bias, costs were lower of the high volume hospitals (p=0.000). Results were insensitive to a bias that would double the odds of being treated at high volume hospitals but sensitive to a bias that would triple the odds. Rosenbaum bonds provide evidence on sensitivity of the estimated results with respect to deviations from propensity score matching assumptions.