A nonlinear expectations model of the term structure with time-varying risk premia is specified, estimated, and tested, allowing for conditional heteroskedasticity and serial correlation in the disturbance terms. The model is subsequently linearized by means of a stationary, lognormal distributional assumption, and a constant risk premium results. Relative to the nonlinear model, the loglinear model performs poorly. Further investigation suggests the importance of taking both a time-varying risk premium and conditional heteroskedasticity in disturbance terms into account when formulating and estimating the term structure. Copyright 1989 by Ohio State University Press.