• A comparison of the results of a five-day survey employing the Pew Research Center’s usual methodology (with a 25 percent response rate) with results from a more rigorous survey conducted over a much longer field period and achieving a higher response rate of 50 percent found that, in 77 out of 84 comparisons, the two surveys yielded results that were statistically indistinguishable. Among the seven items that manifested significant differences between the two surveys, the differences in proportions of people giving a particular answer ranged from 4 percentage points to 8 percentage points (Keeter et al., 2006).

• An analysis of data from the ATUS—the sample for which is drawn from the CPS respondents—together with data from the CPS Volunteering Supplement was undertaken to demonstrate the effects of survey nonresponse on estimates of volunteering activity and its correlates (Abraham et al., 2009). The authors found that estimates of volunteering in the United States varied greatly from survey to survey and did not show the decline over time common to other measures of social capital. They ascribed this anomaly to social processes that determine survey participation, finding that people who do volunteer work respond to surveys at higher rates than those who do not do volunteer work. As a result, surveys with lower response rates will usually have higher proportions of volunteers. The result of the decline in response rates over time likely has been an increasing overrepresentation of volunteers. Furthermore, the difference shows up within demographic and other subgroups, so conventional statistical adjustments for nonresponse cannot correct the resulting bias.

• In a study of nonresponse bias for the 2005 National Household Education Survey (NHES), Roth et al. (2006) found evidence of a potential bias in the survey on adult education, in which females were more likely to respond than males were. The problem was resolved by a weighting class adjustment that used sex in forming the weighting classes.

• In another study of nonresponse bias in the 2007 NHES (Van de Kerckhove et al., 2009), results indicated undercoverage-related biases for some of the estimates from the school readiness survey. However, nonresponse bias was not a significant problem in the NHES data after weighting.

Recommendation 2-5: Research is needed on the theoretical limits of what nonresponse adjustments can achieve, given low correlations with survey variables, measurement errors, missing data, and other problems with the covariates.

Nonresponse can also affect the variance of any statistic, reducing confidence in univariate statistics, and it can also bias estimates of bivariate and multivariate associations, which could bias results from substantive



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement