Using machine learning algorithms to identify farms on the 2022 Census of Agriculture

Abstract

As is the case for many National Statistics Institutes, the United States Department of Agriculture’s (USDA’s) National Agricultural Statistics Service (NASS) has observed dwindling survey response rates, and the requests for more information at finer temporal and spatial scales have led to increased response burdens. Non-survey data are becoming increasingly abundant and accessible. Consequently, NASS is exploring the potential to complete some or all of a survey record using non-survey data, which would reduce respondent burden and potentially lead to increased response rates. In this paper, the focus is on a large set of records associated with potential farms, which are operations with undetermined farm status (farm/non-farm) and are referred to here as operations with unknown status (OUS). Although they usually have some agriculture, most OUS records are eventually classified as non-farms. Those OUS that are classified as farms tend to have higher proportions of producers from under-represented groups compared to other records. Determining the probability that an OUS record is a farm is an important step in the imputation process. The OUS records that responded to the 2017 U.S. Census of Agriculture were used to develop models to predict farm status using multiple data sources. Evaluated models include bootstrap random forest (RF), logistic regression (LR), neural network (NN), and support vector machine (SVM). Although the SVM had the best outcomes for three of the five metrics, the sensitivity for identifying farms was the lowest (13.8%). The NN model had a sensitivity of 80.5%, which was substantially higher than the other models, and its specificity of 45.3% was the lowest of all models. Because sensitivity was the primary metric of interest and the NN performed reasonably well on the other metrics, the NN was selected as the preferred model.