Effects of attrition and non-response in the health and retirement study

We study the effect of attrition and other forms of non-response on the representativity over time of the Health and Retirement Study (HRS) sample born 1931-1941; the sample was initially drawn in 1992. Although some baseline characteristics of respondents do appear correlated with non-response over time, the 2002 sample of respondents does not appear to suffer significantly from selection on observables, except for race and ethnicity; for these two observables, longitudinal weights based on the Current Population Survey (CPS) can be used and are provided with the data set. We attribute this lack of selection to the fact that attritors who differ most eventually come back to the survey in waves prior to 2002. Although this allows cross-sections to remain fairly representative in later waves, it suggests that longitudinal analysis should use the unbalanced sample rather than the balanced sample of those interviewed in all waves. Individuals who attrit but who are recruited back into the survey are very different from those who are permanent attritors to the HRS. Finally, we investigate the selective nature of the decision of respondents to grant HRS permission to access their Social Security records and of the non-response introduced by employers of pension policyholders not providing HRS with worker's Summary Plan Descriptions. We find that subsamples for which such information is available are selective on a number of dimensions, such as education and other socioeconomic status (SES) outcomes.