December 3, 2013
The usual approach to unit non-response bias detection and adjustment in social surveys has been post-stratification weights, or more recently, propensity score adjustment (PSA) based on auxiliary information. There exists a third approach, which is far less popular: using multiple imputed values for survey outcomes of each missing unit. Multiple imputation (MI) is suggested as a promising alternative to PSA since the latter is known to increase variance substantially without reducing bias if auxiliary variables are not associated with the survey outcome of interest. Given that most social surveys have multiple target variables, creating imputed data sets may address bias in survey outcomes with less variance inflation. We examined the performance of PSA and MI on mean estimates under various conditions using full-simulated data. To evaluate the performance of the methods, we report average bias, estimated standard errors, root mean squared errors and percent coverage of full sample 95% confidence intervals. In light of our results, we can more confidently say that MI is a viable alternative to PSA in unit nonresponse adjustment, compromising bias reduction, efficiency and coverage of the full sample confidence intervals for mean estimates.