Model-Building and Computational Efficiency for Imputation in the Agricultural Resource Management Survey

• Activity Number: 612
• Type: Topic Contributed
• Date/Time: Thursday, August 4, 2011 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Michael Robbins*+
• Companies: National Institute of Statistical Sciences
• Keywords: Imputation ; ARMS ; Government Surveys ; MCMC ; SWEEP Operator ; Economic Modeling

Abstract

At least year's JSM in Vancouver we introduced an imputation procedure which samples imputations from a joint model that is constructed from a sequence of conditional regression models. This technique was labeled as iterative sequential regression (ISR) and was applied to data from the USDA's Agricultural Resource Management Survey (ARMS). This procedure is conducive for high dimensional problems, since it allows for flexible selection of a predictor function in each conditional model while maintaining a valid joint distribution. ISR will be used to jointly impute for 150+ ARMS variables, and in this talk we discuss the manner by which conditional models are selected for each of these variables. To elaborate, the models are created using a combination of economic expertise (which is used to group variables according to known economic relation and to select predictors for each grouping) and automated variable selection procedures. We also discuss numerical techniques which are used to determine coefficients for each conditional model while minimizing the number of necessary computations. These numerical techniques allow the procedure to be run in a reasonable timeframe.