Imputing for Missing Data in the ARMS Household Section: A Multivariate Imputation Approach

This study proposes a new method to impute for ordinal missing data found in the household section of the Agricultural Resource Management Survey (ARMS). We extend a multivariate imputation method known as Iterative Sequential Regression (ISR) and make use of cut points to transform these ordinal variables into continuous variables for imputation. The household section contains important economic information on the well-being of the farm operator’s household, asking respondents for information on off-farm income, household expenditures, and off-farm debt and assets. Currently, the USDA's Economic Research Service (ERS) uses conditional mean imputation in the household section, a method known to bias the variance of imputed variables downward and to distort multivariate relationships. The new transformation of these variables allows them to be jointly modeled with other ARMS variables using a Gaussian copula. A conditional linear model for imputation is then built using correlation analysis and economic theory. Finally, we discuss a Monte Carlo study which will randomly poke holes in the ARMS data to test the robustness of our proposed method. This will allow us to assess how well the adapted ISR imputation method works in comparison with two other missing data strategies, conditional mean imputation and a complete case analysis.