Imputation for Nonmonotone Past-Value-Dependent Nonresponse in Longitudinal Studies with Application to the Survey of Industrial Research and Development

• Activity Number: 221
• Type: Topic Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Government Statistics
• Author(s): Martin Klein*+ and Jun Shao
• Companies: U.S. Census Bureau and University of Wisconsin
• Keywords: Bootstrap ; Imputation model ; Intermittent missing ; Kernel Regression ; Linear regression ; Missing not at random

Abstract

We present an overview of new regression based imputation methods for longitudinal studies with nonmonotone past-value-dependent nonrespondents. In the case of nonmonotone nonresponse, the past-value-dependent nonresponse mechanism is nonignorable as defined by Little and Rubin (2002). The methods do not require any parametric model on the joint distribution of the study variables across time points or on the response mechanism. We explore the application and customization of these methods to the Survey of Industrial Research and Development (SIRD), a survey conducted jointly by the U.S. Census Bureau and the U.S. National Science Foundation (NSF). In the current imputation procedure used for the SIRD, total spending on research and development for a nonresponding company is imputed by the company's data from a previous year, after making an adjustment for industry growth. The proposed methods share similarities with the current procedure, while providing a framework for imputation grounded on the formal assumption of nonmonotone past-data-dependent nonresponse. We use the bootstrap to obtain variance estimates that account for uncertainty due to imputation.