Purpose: Cohort studies often conduct periodic follow-up interviews (or waves) to determine disease incidence since the previous follow-up and to update measures of exposure and confounders. The common practice of excluding nonrespondents from standardized incidence ratio (SIR) analyses of these cohorts can bias the estimates of interest if nonrespondents and respondents differ on important characteristics related to outcomes of interest. We propose an analytic approach to reduce the impact of nonresponse in the analyses of SIRs.
Methods: Logistic regression models controlling baseline information are used to estimate the propensity, or the probability of response; the reciprocals of these propensities are used as weights in the analysis of risk. This is illustrated in the analysis of 15 years of follow-up of a cohort of US radiologic technologists after an initial interview to assess the risk at several cancer sites from occupational radiation exposure. We use information from the baseline survey and certification records to compute the propensity of responding to the second survey. SIRs are computed using Surveillance, Epidemiology, and End Results (SEER) cancer incidence rates. Variances of the SIRs are estimated by a jackknife method that accounts for additional variability resulting from estimation of the weights.
Results: We find that, in this application, weighting alters point estimates and confidence limits only to a small degree, thus providing reassurance that the results are robust to nonresponse. This indicates that results from the analyses excluding the missing data may be slightly biased and weighting helps in reducing the nonresponse bias.
Conclusion: This method is flexible, practical, easy to use with existing software, and is applicable to missing data from cohorts with baseline information on all subjects.