Bootstrap Variance Estimation For The National Population Health Survey

Abstract

The National Population Health Survey (NPHS) is a longitudinal survey on the health of Canadians, the determinants and correlates of health, and health care utilisation. A panel of respondents, selected in 1994/95, is being surveyed by Statistics Canada every two years for up to twenty years. NPHS data products include master files for use internally and by NPHS data-sharing partners, public-use microdata files (PUMFs), and dummy master files used by remote- access researchers to test analysis programs before sending them to the Agency for processing. In the first wave, all variance estimation had to be done internally, using a jackknife variance program, since detailed survey design information (i.e., stratum, cluster) is deemed confidential and cannot be disseminated. In the second wave, in 1996/97, NPHS became the first longitudinal survey at Statistics Canada to use the bootstrap for variance estimation. A major motivation for this change was the need to enable data users to do their own variance estimation using bootstrap weights provided with the PUMFs. We present the studies carried out to determine the optimum number of weights, to examine the possible use of the mean bootstrap, and to determine if the dissemination of these weights would cause a disclosure risk. We describe the bootstrap weight files, both longitudinal and cross-sectional, now being used for variance estimation by NPHS and its data-sharing partners, Health Canada and the provincial Ministries of Health. Finally, the remote access program is described, focussing on the dummy bootstrap weights created to allow researchers to write and test their own variance estimation programs before transmitting them to Statistics Canada for execution using the master files. (joint with Wesley Yung, Douglas Yeo and Harold Mantel )