Investigating the Bias of Alternative Statistical Inference Methods in Mixed-Mode Surveys

• Activity Number: 606
• Type: Contributed
• Date/Time: Wednesday, August 7, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Survey Research Methods Section
• Author(s): Zeynep Tuba Suzer-Gurtekin*+ and Steven G. Heeringa and Richard Valliant
• Companies: ISR - University of Michigan and ISR - University of Michigan - Program in Survey Methodology and University of Michigan and University of Maryland
• Keywords: Mixed-mode surveys ; Mode effects ; Imputation ; Current Population Survey (CPS)

Abstract

Mixed-mode surveys combine different data collection modes to reduce non-observational survey errors under certain cost constraints. In this survey design, usually there is no control over who responds by which mode. As a result, data are obtained by a nonrandom mixture of survey administration modes. Without adjusting for this nonrandom mixture of modes, the standard method of estimation that combines responses from different modes has a bias that depends on both mode effects and the mix of respondents that choose each mode. Unless mode effects are zero, data should be adjusted for both nonresponse and nonrandom mixture of modes. We present alternative methods that account for both nonresponse and the nonrandom mixture of modes. Although in principle the separate mode effects are not estimable in a mixed-mode survey design, the alternative estimators do allow estimation of the difference in average mode effects. In addition, the bias properties of alternative methods can be better understood when compared to the standard estimation method. The alternative methods use models to impute each respondent's values for each counterfactual response mode-e.g. a telephone response value of in-person respondents. Combining the observed values with the imputations results in a "completed" data set for each mode. Alternative estimators are then used to combine these mode-specific "completed" data sets in an attempt to reduce bias associated with confounded and nonrandom influences of mode choice and mode effects. This paper presents some results for empirical comparisons of mean personal income and percent health insurance coverage based on the alternative methods and standard method. The public-use 2012 Current Population Survey (CPS) March data are used for empirical evaluations.