|
Abstract:
|
The USDA's National Agricultural Statistics Service (NASS) produces monthly and annual yield forecasts for major crops at state and regional levels. To support the forecasting program, several monthly surveys are conducted during the growing season. In addition to the surveys, administrative data are also available for crops such as upland-cotton, for which production data are collected biweekly from cotton gins. The Research and Development Division at NASS has developed Bayesian hierarchical models that combine data from these surveys and several covariates, including weather data, to produce yield forecasts for each state where the crop is grown and for the region that comprises major crop producing states. The Bayesian approach is extended to adjust for heterogeneities in yield, production, weather and other factors within a state, by partitioning the state into more homogeneous sub-areas and then incorporating data from these sub-areas into the model. Alternative approaches to covariate selection are discussed; and estimation of measures of uncertainty associated with administrative data are considered. Performances of alternative models are compared.
|
