Analysis

The Canadian Labour Force Survey uses the rotation panel design. Every month, one sixth of the sample rotates and five sixths remain. Hence, under this rotation scheme, once a rotation panel enters in the sample, it stays 6 months in the sample before it rotates out. Because of this design feature and the way of selecting the rotate-in panel, the estimates based on the panels in the same or different months are correlated. The correlation between two panel estimates is called the panel correlation. Three kinds of panel correlations are defined in this paper: (1) the correlation (denoted by \rho) between estimates for the same characteristic based on the same panel in different months; (2) the correlation (denoted by \gamma) between estimates of the same characteristic based on geographically neighboring panels in different months; (3) the correlation (denoted by \tau) between estimates of different characteristics based on the same panel in the same or different months. This paper describes a methodology for estimating these panel correlations and presents estimated correlations for selected variables using 1980-81 and 1985-87 data with some discussion.

Benchmarking is a method of improving estimates from a sub-annual survey with the help of corresponding estimates from an annual survey. For example, estimates of monthly retail sales might be improved using estimates from the annual survey. This article deals, first with the problem posed by the benchmarking of time series produced by economic surveys, and then reviews the most relevant methods for solving this problem. Next, two new statistical methods are proposed, based on a non-linear model for sub-annual data. The benchmarked estimates are then obtained by applying weighted least squares.

The problem of collapsing the imputation classes defined by a large number of cross-classifications of auxiliary variables is considered. A solution based on cluster analysis to reduce the number of levels of auxiliary variables to a reasonably small number of imputation classes is proposed. The motivation and solution of this general problem are illustrated by the imputation of age in the Hospital Morbidity System where auxiliary variables are sex and diagnosis.

The Canadian Labour Force Survey uses the rotation panel design. Every month, one sixth of the sample rotates and five sixths remain. Hence, under this rotation scheme, once a rotation panel enters in the sample, it stays 6 months in the sample before it rotates out. Because of this design feature and the way of selecting the rotate-in panel, the estimates based on the panels in the same or different months are correlated. The correlation between two panel estimates is called the panel correlation. Three kinds of panel correlations are defined in this paper: (1) the correlation (denoted by \rho) between estimates for the same characteristic based on the same panel in different months; (2) the correlation (denoted by \gamma) between estimates of the same characteristic based on geographically neighboring panels in different months; (3) the correlation (denoted by \tau) between estimates of different characteristics based on the same panel in the same or different months. This paper describes a methodology for estimating these panel correlations and presents estimated correlations for selected variables using 1980-81 and 1985-87 data with some discussion.

Benchmarking is a method of improving estimates from a sub-annual survey with the help of corresponding estimates from an annual survey. For example, estimates of monthly retail sales might be improved using estimates from the annual survey. This article deals, first with the problem posed by the benchmarking of time series produced by economic surveys, and then reviews the most relevant methods for solving this problem. Next, two new statistical methods are proposed, based on a non-linear model for sub-annual data. The benchmarked estimates are then obtained by applying weighted least squares.

The problem of collapsing the imputation classes defined by a large number of cross-classifications of auxiliary variables is considered. A solution based on cluster analysis to reduce the number of levels of auxiliary variables to a reasonably small number of imputation classes is proposed. The motivation and solution of this general problem are illustrated by the imputation of age in the Hospital Morbidity System where auxiliary variables are sex and diagnosis.