Guidance on the choice of data sources for cost weights in cost-effectiveness analyses in the literature is scarce with few empirical examples of cost weights derived from different approaches. OBJECTIVES: This paper describes the large differences that can be found in cost weights derived from different data sources and analytic approaches. We highlight the implications of the choice of costing data source and the magnitude of error that could be introduced by assuming data parity in direct comparisons of different studies. Though our examples focus on HIV disease treatment, the issues we explore are generic to cost estimation for all conditions. METHODS: We compared annual HIV disease treatment cost weights from two data sets, the HIV Cost and Services Utilization Study (HCSUS) survey and South Carolina Medicaid data. Cost weights for hospital and nursing home stays and emergency room and physician visits were included in the analyses. RESULTS: The cost weights for Medicaid patients with AIDS were $15,033 in HCSUS and $10,281 in SC Medicaid data. Within the HCSUS data set, cost weights also varied greatly depending on whether the patient was covered by Medicaid ($15,033), Medicare ($8,487), private ($4,200), or other public insurance ($19,347). Further analysis of the HCSUS data found that the mean cost estimates for patients were larger when viral load (VL) data was missing ($8,413) than when it was included ($5,362). This difference was especially pronounced at higher CD4 cell count ranges. For example, the mean cost for patients with CD4 200-499 was $3,829 when VL was reported and $6,458 when it was missing, a difference of 67%. CONCLUSION: As this research illustrates, there may be large differences in cost weights that are derived from different data sources and analytic approaches. Accordingly, they should be given careful consideration when determining the external validity of economic studies.