Abstract:
|
The Canadian Census long form is a quinquennial large-scale sample survey for which millions of estimates on the Canadian population are published at various levels of geography. In 2016, to improve the analytical potential and the intelligibility of the published point estimates, Statistics Canada wants to be able to calculate a variance-based quality-indicator (QI) for each estimate. In addition, for the first time, analysts having access to microdata will be provided replicate weights enabling them to produce variance estimates on their own. This paper summarizes the development of a replication variance estimator that uses few replicates to be integrated into the existing dissemination systems. Emphasis will be put on the challenges of developing the variance estimator and the results of a Monte Carlo simulation supporting the choice of the method to be used. These challenges include the very large sample size along with the large sampling fraction, the need to calibrate the replicate weights and the numerous variance estimates being calculated for both smooth and non-smooth statistics in a limited timeframe while respecting confidentiality of the data provided.
|