Internet Surveys: Can Statistical Adjustments Eliminate Coverage Bias?

Abstract

The Internet is an attractive mode of data collection to survey researchers due to cost savings and timeliness in comparison with other modes. However, survey estimates are subject to coverage bias if sampled persons with Internet access are systematically different from those without Internet access who were excluded from the survey. Statistical adjustments, either through weighting or modeling methods, can minimize or even eliminate bias due to noncoverage. In the current paper, we examine the coverage bias associated with conducting a hypothetical Internet survey on a frame of persons obtained through a random-digit-dial sample. We compare estimates collected during telephone interviews from households with and without Internet access using data from the 2003 Michigan Behavioral Risk Factor Surveillance System in the United States. A total of 25 binary variables (e.g., the percent of adults who have asthma or who are classified as being obese) and four count variables (e.g., the number of alcoholic drinks consumed per month) were analyzed for this study in addition to eight demographic characteristics. Weights based on the general regression estimator are computed such that the coverage bias is reduced to undetectable levels for most of the health outcomes analyzed from the Michigan survey. Though not definitive, the analysis results suggest that statistical adjustments can reduce, if not eliminate, coverage bias in the situation we study.