Model-Based County Level Crop Estimates Incorporating Auxiliary Sources of Information

Abstract

Summary In 2011, the US Department of Agriculture's National Agricultural Statistics Service started the complete implementation of the County Agricultural Production Survey (CAPS). The CAPS is an annual survey to provide accurate county level acreage and production estimates of approved federal and state crop commodities. The current top down method of producing official county level estimates that satisfy the county–district–state benchmarking constraint is an expert assessment incorporating multiple sources of information. We propose a model‐based method that combines the CAPS acreage data with auxiliary data and improves county level survey estimation, while providing measures of uncertainty for the county level acreage estimates. Auxiliary sources of information include remote sensing data, weather data and planted acreage administrative data from other US agencies. A hierarchical Bayesian subarea level model is proposed and implemented, with an additional hierarchical level for the sampling variances. County level, model‐based acreage estimates have lower coefficients of variation than the corresponding county level survey acreage estimates. Top down benchmarking methods are investigated and the final acreage estimates satisfy the county–district–state benchmarking constraint.

Statistical Challenges in Combining Survey and Auxiliary Data to Produce Official Statistics

“…A detailed discussion of classic benchmarking adjustments is given in Rao and Molina (2015). Studies on different benchmarking adjustments to crop acreage prediction are discussed in Erciulescu et al (2019). In this section, we illustrate a benchmarking adjustment applied to the model predictions constructed under the different data availability cases, so that the county-level predictions aggregate to the district-level predictions and the district-level predictions aggregate to the prepublished state-level value.…”

“…Raking provides a suitable benchmarking adjustment to ensure consistency of substate predictions with state targets. For this study, we use the extension of the classic ratio adjustment given in Erciulescu et al (2019), and we apply the constraint at the (MCMC) iteration level. This type of benchmarking adjustment is not adopted as part of the prior information or the model, but it facilitates its application to the set of in-sample and not-insample counties, in a small amount of time.…”

“…In this article, we integrate all the data to identify the set of counties with planting activity for a specific crop (or the prediction space), in a given crop season, and to construct a covariate with good predictive power and observations available for all the counties in the prediction space. Challenges in multistage, nationwide prediction for counties with sample sizes as small as zero (not the case in Erciulescu et al 2019) are addressed using hierarchical Bayes subarea-level models.…”

“…Modeling strategies are developed to deal with incomplete data and benchmarking methods are implemented to overcome the challenge of attaining consistency among predictions at nested levels of aggregation. Whereas Erciulescu et al (2019) developed and compared models for direct estimates scaled by the sample sizes, with a hierarchy for sampling variances and different benchmarking methods, here we adopted the model for the direct estimates, with fixed sampling variances and the ratio adjustment, as a practical method with good performance that allows for prediction for subareas with sample sizes of just one or even zero where suggested by auxiliary information. This outcome was not possible under the model specification pursued in Erciulescu et al (2019).…”

“…Whereas Erciulescu et al (2019) developed and compared models for direct estimates scaled by the sample sizes, with a hierarchy for sampling variances and different benchmarking methods, here we adopted the model for the direct estimates, with fixed sampling variances and the ratio adjustment, as a practical method with good performance that allows for prediction for subareas with sample sizes of just one or even zero where suggested by auxiliary information. This outcome was not possible under the model specification pursued in Erciulescu et al (2019). Moreover, due to the extended prediction space, the possible over-adjustment due to benchmarking is less of a concern for NASS than it was for Erciulescu et al (2019).…”