Using Machine Learning Algorithms and Impact Scores to Manage Cost, Burden, and Data Quality of the Agricultural Resource Management Survey

• Activity Number: 208 - Survey Estimation
• Type: Contributed
• Date/Time: Tuesday, August 4, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Government Statistics Section
• Author(s): Gavin Corral* and Tyler Wilson and Andrew Dau and Audra Zakzeski
• Companies: USDA NASS and USDA NASS and USDA NASS and USDA NASS
• Keywords: machine learning; propensity; impact; non response

Abstract

The United States Department of Agriculture’s (USDA’s) National Agriculture Statistics Service (NASS) and Economic Research Service implement the Agricultural Resource Management Survey (ARMS). The ARMS survey is designed to measure the well-being of farm establishments and consists of three parts. The questionnaire for the third and final part, known as ARMS III, is approximately 24 pages long and asks the respondents sensitive financial information. Consequently, data collection often requires expensive field enumeration and is burdensome on respondents. To address this problem, NASS has developed methods to decrease cost and burden, while mitigating bias of the survey estimates. This research explores the efficacy of using response propensity and impact scores to identify records associated with farms that are unlikely to respond and have low impact on the final estimates. The identified records will not be sent out for expensive field follow up. The effect of this approach on cost, burden, and bias is measured via simulations.