This paper introduces a new specification of technical inefficiency in panel data models. First, the overall technical inefficiency is decomposed into a persistent component and a residual component. Second, a multistep procedure is used to estimate the parameters of the production function as well as persistent and residual technical inefficiency. The advantage of this multistep procedure is that the parameter estimates are robust to distributional assumptions on the error components. Distributional assumptions are required in the final stage to estimate the residual component of technical inefficiency. The model is used to examine technical efficiency in Swedish dairy farms during the period 1976 to 1988.