Nonresponse and Measurement Error: Relationship, Relative Magnitude, and Correction

• Activity Number: 103
• Type: Invited
• Date/Time: Monday, August 1, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Survey Research Methods
• Author(s): Andy Peytchev*+
• Companies: RTI International
• Keywords: Nonresponse Bias ; Measurement Error ; Multiple Imputation ; MSE ; Total Survey Error ; Paradata

Abstract

Commonly employed weighting methods to address nonresponse generally lead to reduced precision, spurring tradeoffs between bias and variance. No corrections for measurement error are commonly incorporated, while variances of survey estimates are typically penalized when ample auxiliary information is available to correct for nonresponse. Multiple imputation can be used for unit nonresponse, item nonresponse, as well as measurement error, making use of frame data, paradata, and survey data with varying levels of missingness. Unlike nonresponse weighting adjustments, it can reduce variances, thus also reducing mean squared error of estimates (MSE). This approach was applied to data from NSFG cycle 5, which includes a rich sampling frame, paradata, and replicate measures less prone to measurement error. Based on multiply-imputed data, measurement error bias was almost three times larger than nonresponse bias. Although the same conclusion was reached through weighting, multiple imputation yielded substantially lower variance estimates and estimates of MSE.