Online Program

Who is Responsible for the Bias? Using Proxy Data and Tree Modeling to Identify Likely Nonrespondents and Reduce Bias
*Morgan Sue Earp, Bureau of Labor Statistics 

Keywords: Nonresponse Bias; Propensity Scores; Classification Trees; Ensemble Trees; Logistic Regression; Incentives.

With increasing nonresponse rates, biased survey estimates are a growing concern. The USDA National Agricultural Statistics Service (NASS) conducts the annual Agricultural Resource Management Survey (ARMS). ARMS collects detailed economic data from a sample of US agricultural operations and suffers from relatively low response rates for a federal survey. Using Census of Agriculture data as a proxy for ARMS data, NASS has identified characteristics associated with nonresponse and thus subgroups in the population who are less likely to respond. By using proxy data that is related to ARMS key estimates, NASS is not only able to identify likely nonrespondents, but specifically those with the greatest impact on nonresponse bias. This information is being used to help field offices better manage nonresponse follow up, so NASS can allocate nonresponse follow-up funds toward operations where nonresponse will result in biased estimates.