class of indicators include a comparison of respondents and nonrespondents on auxiliary variables, variance functions of nonresponse or post-stratification weights, and goodness-of-fit statistics of propensity models (Schouten et al., 2009). They utilize more data and are reported at the survey level. But, like response rates, they are mainly useful only in the context of a specific survey, whereas nonresponse bias is a statistics-level problem. As a result, they are not commonly used.

Representativity indicators (R-indicators) are an example of a new metric for assessing the effects of nonresponse.1 R-indicators attempt to measure the variability of response propensities (van der Grijn et al., 2006; Cobben and Schouten, 2007). They do this by measuring “the similarity between the response to a survey and the sample or the population under investigation” (van der Grijn et al., 2006, p. 1). A survey that exhibits less variability in response propensities is likely to exhibit a better match between the characteristics of the respondents and the population they are meant to represent for the variables that the model uses to estimate the propensities.

R-indicators can be monitored during data collection to permit survey managers to direct effort to cases with lower response propensities and, in so doing, to reduce the variability among subgroup response rates. The indicator can be monitored during survey collection because the response propensities can be calculated with complete information available on the frame. In order for the R-indices to be completely comparable across surveys, they need to be estimated with the same variables. Wagner (2008) suggests that this would be facilitated if all surveys had a common set of frame data. However, there is no proof that such indices would apply across all types of surveys, or even across all relevant estimates within a given survey. More research that uses these indicators in survey settings is needed.

An example of the use of the R-index was provided by John Dixon of the Bureau of Labor Statistics in his presentation to the panel (Dixon, 2011). As shown in Figure 2-1, he charted the R-index for the Current Population Survey (top line) with the response rate for the survey (bottom line). He noted that, at 95 percent confidence intervals, the R-index is somewhat flatter than the response rate, which suggests that, for this survey, response propensities indicate a good match between the characteristics of the respondents and the population they are meant to represent.

Estimate-level indicators are indicators that use a response indicator, frame and paradata, and survey variables. They require an explicit model for each variable, and the model is usually estimated from the observed data and relies on the assumption that the missing data are missing at random (MAR). Wagner reported that there is very little research into non-MAR

______________________________________

1Additional content on R-indicators is provided by Wagner (2011).



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