Abstract:
|
A common problem in statistical inference from surveys with multiple contacts is the bias induced by nonrespondents. An approach to bias reduction is to consider the relationship of missingness with covariate information obtained on all subjects in the survey population. We applied weighting methods, using response propensity models, to assess site-specific cancer risk from occupational radiation exposure among a cohort of US radiologic technologists that was surveyed twice over fifteen years. Standardized incidence ratios (SIR) were computed using US population cancer incidence rates (SEER). Variance was estimated by the jackknife method using random groups. When the number of observed cancers was less than fifty, modifications were made to the confidence limits computed using the Poisson distribution. Simulations were carried out to evaluate the coverage properties of this method. For most of the cancer sites, the SIRs reduced moderately (e.g., Colon Cancer: 1.11 to 1.06). The method described here is flexible, practical, easy to use with existing software, extendable to missing data on multiple survey waves, and can be applied to other cohorts to reduce nonresponse bias.
|