Uncertainties about the age of base-flow discharge can have serious implications for the management of degraded environmental systems where subsurface pathways, and the ongoing release of pollutants that accumulated in the subsurface during past decades, dominate the water quality signal. Numerical groundwater models may be used to estimate groundwater return times and base-flow ages and thus predict the time required for stakeholders to see the results of improved agricultural management practices. However, the uncertainty inherent in the relationship between (i) the observations of atmospherically-derived tracers that are required to calibrate such models and (ii) the predictions of system age that the observations inform have not been investigated. For example, few if any studies have assessed the uncertainty of numerically-simulated system ages or evaluated the uncertainty reductions that may result from the expense of collecting additional subsurface tracer data. In this study we combine numerical flow and transport modeling of atmospherically-derived tracers with prediction uncertainty methods to accomplish four objectives. First, we show the relative importance of head, discharge, and tracer information for characterizing response times in a uniquely data rich catchment that includes 266 age-tracer measurements (SF₆, CFCs, and ³H) in addition to long term monitoring of water levels and stream discharge. Second, we calculate uncertainty intervals for model-simulated base-flow ages using both linear and non-linear methods, and find that the prediction sensitivity vector used by linear first-order second-moment methods results in much larger uncertainties than non-linear Monte Carlo methods operating on the same parameter uncertainty. Third, by combining prediction uncertainty analysis with multiple models of the system, we show that data-worth calculations and monitoring network design are sensitive to variations in the amount of water leaving the system via stream discharge and irrigation withdrawals. Finally, we demonstrate a novel model-averaged computation of potential data worth that can account for these uncertainties in model structure.