response propensity and the data quality indicators, potential common causal factors were examined. The researchers found that, in general, data quality, at least as they measured it, decreased for some variables as the probability of nonresponse increased. The study concluded that efforts to reduce nonresponse can lead to poorer quality data (Fricker and Tourangeau, 2010). Other work in this field is under way and may shed additional light on this important issue.
Recommendation 2-2: Research is needed on the impact of nonresponse reduction on other error sources, such as measurement error.
Recommendation 2-3: The research agenda should seek to quantify the role that nonresponse error plays as a component of total survey error.
Although the exact relationship between nonresponse and bias is not yet clear, it is still important to understand the effects of nonresponse bias because bias jeopardizes the accuracy of estimates derived from surveys and thus the ability of researchers to draw inferences about a general population from the sample. The interactions are complex because nonresponse exists at the item level as well as at the interview level, and item nonresponse contributes to bias at the item-statistic level so that bias is a function of both unit and item nonresponse.
Recommendation 2-4: Research is needed to test both unit and item nonresponse bias and to develop models of the relationship between rates and bias.
Examples cited in Peytchev (2009) and other sources (e.g., Groves, 2006; Groves and Peytcheva, 2008) tend to show that the effect of nonresponse bias on means and proportions can be substantial:
• Nonrespondents to the component of the National Health and Nutrition Examination Survey III that measured glucose intolerance and diabetes through eye photography were 59 percent more likely to report being in poor or fair health than respondents to the main survey (Khare et al., 1994).
• The Belgian National Health Interview Survey with a response rate of 61.4 percent obtained a 19 percent lower estimate for reporting poor health than did the Belgian census, which had a response rate of 96.5 percent (Lorant et al., 2007).