Using Data on Interruptions in Telephone Service to Reduce Nontelephone Bias in a Random-digit-dialing Survey

Telephone surveys are subject to coverage bias from noncoverage of nontelephone households. Though the percentage of households not having telephone service is small nationally, it can vary substantially by geographic area and by socioeconomic factors. For example, lack of telephone service is more common among low-income households than in other income groups, so low-income households may not be adequately represented in a telephone survey (Thornberry and Massey 1988). Postsurvey weighting can reduce this bias. Keeter (1995), using a panel constructed from the 1992-93 Current Population Survey (CPS), observed that at any given time telephone households include households that were recently a part of the nontelephone population. Generally, these households have had an interruption in telephone service. By comparing the characteristics of these households with those without telephones, he showed that it is possible to use data from households with interruptions in telephone service to adjust for noncoverage of nontelephone households. Brick et al. (1996) suggested a method of adjusting survey estimates to reduce bias from noncoverage, by using data on interruptions in telephone service. This paper describes a method to adjust for noncoverage of nontelephone households. This method assumes that the population number of individuals in telephone households is known (or can be estimated from a survey or from alternative sources) and uses the survey results to estimate the weighted proportion of individuals in telephone households with an interruption in service. It then forms a weighted average of the estimates from the non-interruption and interruption parts of the sample, using the interruption estimate for both the interruption and the nontelephone parts of the population. We use the proposed method to adjust the estimates of vaccination coverage rates for children between 19 and 35 months of age from the National Immunization Survey (NIS), and we compare the adjusted estimates with the current estimates obtained through the usual poststratification methods. Truncation at an upper limit reduces the impact of large weights resulting from this adjustment. We provide the mathematical theory and empirical results from the 1997 National Immunization Provider Record Check Study (NIPRCS), which indicate that this interruption adjustment method substantially reduces the nontelephone associated bias in estimated childhood vaccination rates. We compare our results to those obtained using poststratified weights without explicit adjustment for noncoverage and to a currently used noncoverage adjustment.