• Presentation

Nonresponse Bias Correction in Telephone Surveys Using Census Geocoding: An Evaluation of Error Properties

Citation

Biemer, P. P., & Peytchev, A. A. (2010, May). Nonresponse Bias Correction in Telephone Surveys Using Census Geocoding: An Evaluation of Error Properties. Presented at AAPOR 2010, .

Abstract

The threat of nonresponse bias has been increasing with the precipitous decline of response rates in random-digit-dialed (RDD) telephone surveys. Often, researchers have only geographic information for sample cases. Census demographic information can be appended at varying levels of geographic aggregation for both respondents and nonrespondents in landline samples, in an attempt to correct for nonresponse bias. The effectiveness of this method depends on the error properties of the census geocoding (CG) process, yet it has not been thoroughly evaluated. At an extreme, error in the CG process can do more harm than good for survey estimates. More realistically, components of the process can be identified as more susceptible to error and improvements can then be made.

The CG process can be implemented in various ways, such as identifying census block groups for listed telephone numbers and census tracts for unlisted numbers. Matches to geographic areas may be erroneous because of incorrect addresses, and also from different postal and census geographic definitions. Although rates of some types of mismatches have been examined, their impact on estimates has not. Furthermore, demographic information in smaller geographic areas may not match the demographic sample composition, inducing higher variance in adjusted estimates without necessarily reducing bias even in demographic variables.

Using parameters from an RDD survey, we impose nonresponse on a face-to-face study that allows us to decompose the error in the CG process. This approach provides a gold standard for respondents and nonrespondents, as well as for listed and unlisted telephone numbers. Preliminary findings show that while for some variables incorrect matches of unlisted numbers contribute to the most to bias in the CG process, surprisingly, bias is also relatively large for correctly matched telephone numbers. We continue this research and aim to offer practical suggestions for the use of the CG method.