AN APPLICATION OF THE BOOTSTRAP VARIANCE ESTIMATION METHOD TO THE PARTICIPATION AND ACTIVITY LIMITATION SURVEY

The bootstrap method is increasingly used to estimate the variance of estimates obtained from complex survey designs. This method has been shown to work well for a wide range of estimators, including medians and quantiles, as well as smooth functions based on totals. In addition, the bootstrap can be less computer intensive than the jackknife method for surveys with a very large number of primary sampling units (PSUs). The sampling plan of the Canadian 2001 Participation and Activity Limitations Survey (PALS) is a stratified two-stage design in which PSUs are selected without replacement with probability proportional to size. The survey presents specific challenges to the use of the bootstrap method. For instance, the sampling fraction for PALS is relatively high in many strata, which causes the bootstrap method to overestimate the variance. What is the magnitude of this overestimation? Also, a logistic regression response propensity model is used for the nonresponse adjustment in PALS. Should a logistic regression model be fitted to each bootstrap sample? How does this method compare with maintaining fixed response classes over all bootstrap samples? This paper will address these issues. KEY WORDS: bootstrap, without replacement design, response propensity model, logistic regression