OBJECTIVES: In cancer research using claims data, identifying metastases is often essential yet difficult.  The objective of this study was to examine the validity of algorithms identifying metastatic breast (BC), lung (LC), or colorectal (CRC) cancer in healthcare claims data. METHODS:  A proprietary clinical cancer database containing physician-reported clinical data on patients with BC, LC, or CRC between January 1, 2007 and March 31, 2010, was linked to claims data. Inclusion required health plan enrollment > 3 months prior to the initial clinical cancer diagnosis date. Un-validated claims algorithms from previous research were identified.   A generic metastatic algorithm with all metastatic ICD-9 codes and tumor-specific variations of the algorithm were assessed for validity. The algorithms’ validity versus the clinically reported metastases was tested using sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). RESULTS: Of 14,480 patients in the database, 4631 BC (mean age 53.6 yr), 2449 LC (mean age 62.9 yr), and 2058 CRC patients (mean age 58.3 yr) met inclusion criteria. Metastases at diagnosis were recorded in 8.0% (371) of BC, 49.2% (1204) of LC, and 25.7% (528) of CRC patients. The tumor specific algorithm for identifying metastatic BC had 53.2% sensitivity and 98.6% specificity; PPV and NPV were 77.6 and 95.8.  The lung-specific algorithm had 55.2% sensitivity and 85.3% specificity; PPV and NPV were 81.0 and 62.6. Similarly, the CRC-specific algorithm had 59.4% sensitivity, 89.8% specificity, with PPV 72.9 and NPV 82.7.  The generic algorithm had lower specificity and higher sensitivity for all 3 cancers and a significantly lower PPV for breast cancer. CONCLUSIONS:  Specificity, but not sensitivity, was high for all tumor-specific algorithms. Although not tested, better sensitivity might be gained by including chemotherapy in the algorithms for some tumor types.