variables are the same for respondents as for all those sampled, we call the response set perfectly balanced” (p. 4). The B-indicator can be used to measure how effective techniques for improving the balance of the sample, such as responsive design (see Chapter 4), have been.
In addition to the R-indicators and B-indicators, other indicators could be imagined, such as a measure based on the variance of the weights. Such indicators are promising, but, as Wagner reminded the committee, a research agenda on alternative indicators of bias should include research on the behavior of different measures in different settings, the bounds on nonresponse bias under different assumptions (especially non-MAR), how different indicators influence data collection strategies, and how to design or build better frames and paradata.
Recommendation 2-7: Research and development is needed on new indicators for the impact of nonresponse, including application of the alternative indicators to real surveys in order to determine how well the indicators work.
This chapter describes a large and growing body of research into the characteristics of nonresponse bias and its relationship (or lack of relationship) to response rates. While encouraging, the work has gone forward in piecemeal fashion and has not been conducted under an umbrella of a comprehensive statistical theory of nonresponse bias.
In his presentation to the panel, panel member Michael Brick suggested that a more comprehensive statistical theory would enhance the understanding of such bias and aid in the development of adjustment techniques to deal with bias under different circumstances (Brick, 2011). A unifying theory would help ensure that comparisons of nonresponse bias in different situations would lead to the development of standard measures and approaches to the problem. In the next chapter, the need for a comprehensive theory will again be discussed, this time in the context of refining overall adjustments for nonresponse.