Using Uncertainty Bounds for Regression Imputation in Statistical Matching

• Activity Number: 115
• Type: Topic Contributed
• Date/Time: Monday, August 1, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Survey Research Methods
• Author(s): Hans Kiesl*+ and Susanne Rässler
• Companies: Regensburg University of Applied Sciences and Otto-Friedrich-Universität Bamberg
• Keywords: statistical matching ; regression imputation ; uncertainty bounds

Abstract

Statistical matching (also called data fusion) tries to combine information from different data sets by matching on those variables that are common to both files. Algorithms like nearest neighbour or Mahalanobis distance matching are routinely applied, but it is well known that they implicitly assume conditional independence of those variables that have not been jointly observed (called specific variables). In this paper, we discuss how to quantify the amount of uncertainty in the matching process by calculating bounds on distribution parameters of the specific variables. Data fusion might be viewed as a missing data problem, and we propose a regression imputation algorithm that creates different matched data sets with different feasible correlation matrices. Since several recent studies have used propensity score matching for combining different data sets, we will also discuss why propensity score matching is appropriate for the estimation of average treatment effects in the context of Rubin's causal model (where we have to deal with a different conditional independence assumption) but should not be applied in the data fusion setting.