Estimating production technology for policy analysis: trading off precision and heterogeneity

This paper develops an approach to select models that can make the best use of limited micro‐level data sets to estimate production function parameters. Since production is often the core of the agricultural and environment policy analyses, we evaluate the models using criteria that reflect the objectives of policy analysis. We argue that policy production models should optimize the precision of policy response predictions, but also incorporate sufficient heterogeneity to allow policy makers to consider the distributional consequences of policies. Hence we develop a series of quantitative metrics of both precision and heterogeneity to compare model performance. Our approach consists of two steps. We first combine the method of generalized maximum entropy and data envelopment analysis and simultaneously estimate the production frontier and technical inefficiency parameters. With a set of household level data, we estimate production models at three different levels. The province‐level model restricts the production technology parameters to be the same for all households. The county‐level models allow production technology parameters to vary by county but restrict them to be equal across communities within the same county. The community‐level models allow production technology parameters to vary by community. In the second step, we use the disaggregated information gain, percentage absolute prediction error and the Theil’s U statistic to evaluate these models.