All Times ET
Program is Subject to Change
TBD
Imputation Models for a Monthly Economic Indicator Survey with Seasonal Data and Low Unit Response (308031)
Nicole Czaplicki, U.S. Census BureauBrian Dumbacher, U.S. Census Bureau
*Stephen James Kaputa, US Census Bureau
Keywords: regression trees, regARIMA, hierarchical Bayesian regression models, advance survey
Currently, the Advance Monthly Retail Trade Survey (MARTS) has no form of automated imputation, relying instead on subjective analyst imputed values. Consequently, the U.S. Census Bureau is investigating methodological enhancements to the current procedures designed to minimize revisions between advance and preliminary estimates, focusing on alternative estimators and missing-data treatments (primarily imputation methods). However, large revisions between the two corresponding estimates is undesirable, especially when the revision reverses the direction of the month-to-month change. This paper demonstrates the use of three different imputation models: Bayesian hierarchical regression models, ARIMA time series models, and regression tree models. All imputation models make use of historic MRTS data and industry structure to impute missing values in the current reference month. We briefly walk through the proposed models, provide simulation study results, and conclude with a discussion of the strengths and benefits of each approach.