A model of recreation demand is developed to determine the role of water levels in determining participation at and frequency of trips taken to various federal reservoirs and rivers in the Columbia River Basin. Contingent behavior data are required to break the near-perfect multicollinearities among water levels at some waters. We combine demand data for each survey respondent at different levels of time aggregation (summer months, rest of year, and annual), and our empirical models accommodate the natural heteroskedasticity that results. Our empirical results show it to be quite important to control carefully for survey nonresponse bias.