• Presentation

Coverage Error in Telephone Surveys: Bias in Point Estimates, Variances, Associations, and Total Error from Exclusion of the Cell Phone-Only Population

Citation

Peytchev, A. A., & Carley-Baxter, L. (2008, May). Coverage Error in Telephone Surveys: Bias in Point Estimates, Variances, Associations, and Total Error from Exclusion of the Cell Phone-Only Population. Presented at AAPOR 2008, New Orleans, LA.

Abstract

While landline telephone household surveys often draw inference about the U.S. population, a proportion with only cell phones is excluded. This proportion is substantial and increasing, providing potential for coverage bias. Improved understanding of the resulting coverage bias and the ability to adjust for it is needed.Studies have looked at bias in means and proportions, but undercoverage can affect other essential statistics. The precision of point estimates can be biased, leading to erroneous conclusions. In addition, research examining multivariate relationships will be further affected by bias in associations.Coverage bias is suspected as the cell-only population is different on demographic characteristics. These characteristics are commonly related to survey measures, creating conditions for bias. Yet bias in adjusted estimates occurs only when differences on survey variables remain within demographic groups.A national landline telephone survey was conducted, followed by a survey of adults with only cell phones. Respondents from the two frames were weighted first separately using census and survey data, to minimize the confounding effect of nonresponse on estimates of coverage bias. Differences between samples were found in estimates of not only means and proportions, but also variances and associations. Bias in some point estimates was reduced through poststratification, but became larger and in opposite direction for others; cell phone respondents were more likely to report victimization, but conditional on demographic characteristics, were less likely to report it. Different uses of survey data can be affected by omitting the cell-only population, while reliance on postsurvey adjustments can be misleading.