Model-Based Crop Yield Forecasting; Covariate Selection and Related Issues

• Activity Number: 138 - Modeling Applications for Backcasting, Nowcasting and Forecasting
• Type: Contributed
• Date/Time: Monday, July 29, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Habtamu Benecha* and Luca Sartore and Nathan Cruze
• Companies: NASS/USDA and National Institute of Statistical Sciences and USDA National Agricultural Statistics Service
• Keywords: Variable selection; Yield forecasting; Composite estimation; Survey sampling; Bayesian hierarchical model

Abstract

USDA’s National Agricultural Statistics Service (NASS) publishes hundreds of reports every year. Such publications include monthly and annual yield forecasts and estimates for major crops. To produce the forecasts, several surveys are conducted during the growing season. In recent years, NASS has been applying Bayesian hierarchical models to combine summaries from multiple surveys, administrative data and several covariates to produce a single estimate for a state or a region that comprises major crop producing states; the model estimates supplement NASS’s yield forecasting program. The influences of covariates on forecasted yield generally decrease through the growing season, but model covariates play an important role in early season forecasting. Currently, covariates being considered in the models are selected based on expert knowledge of crop development and growth dynamics. In this paper, formal variable-selection approaches are considered for the identification of the best covariates. The best sets of covariates are then compared using information criteria and other distances between model-based early-season forecasts and official estimates.