Missing Data in Record-Linked Data Sets: Comparing the Performance of Different Missing Data Techniques

• Activity Number: 242
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Gerhard Krug*+
• Companies: University of Erlangen-Nuremberg
• Keywords: missing data ; record linkage ; multiple imputation ; sample selection model

Abstract

Combining data from a survey with register data using record linkage (RL) can lead to missing data and potentially to biased estimates, if survey respondents have to consent to it. Missing data (MD) techniques can be used to correct for potential record linkage bias. Based upon a survey where participants were asked permission for RL the performance of different missing data techniques is compared. For respondents who refuse their permission I set their survey answers to missing, creating pseudo-missing data. To correct for potential bias, OLS Regression is performed using complete case analysis (MCAR), multiple imputation (MAR) and Heckman's sample selection model (MNAR), respectively. Their performance is compared to a benchmark regression that is based on the complete data set. Several missing data scenarios are compared. Results indicate that when RL-bias was small, all missing data techniques performed well. In contrast, when RL-bias was high, only multiple imputation was able to correct for the RL-bias, given that only independent variables had missing values. With high RL-bias and missing values in the dependent variable, none of the MD techniques eliminated the bias.