Multivariate Linear Mixed-Effects Models for Missing Data Applied to a Business Survey

• Activity Number: 633
• Type: Topic Contributed
• Date/Time: Thursday, August 8, 2013 : 8:30 AM to 10:20 AM
• Sponsor: Social Statistics Section
• Author(s): Joanna Fane Lineback*+ and Joseph Schafer
• Companies: U.S. Census Bureau and U.S. Census Bureau
• Keywords: multiple imputation ; business survey ; nonresponse adjustment ; adaptive design ; multivariate model

Abstract

In this paper we apply multivariate linear mixed-effects models for missing data (Schafer and Yucel 2002) to the Monthly Wholesale Trade Survey. The MWTS is a longitudinal survey of U.S. wholesale businesses that provides relative month-to-month change estimates of sales and total inventories. As such, it is important that nonresponse adjustments preserve relationships among variables. The current method for handling missing data is to impute via a ratio adjustment to prior month data. We test the feasibility of this model, and we examine the patterns of missingness among the data. We develop a mixed-effects imputation model for sales and inventories, create multiple imputations of missing values using a Markov chain Monte Carlo procedure, and compare the results to those of the current method. We discuss how this methodology could be explored for use in a real-time, adaptive design framework.