Conferences

For more than forty years, the sample frames for national household surveys carried out by INSEE (except the Labour Force Survey) have been built from the lists of dwellings established by the Census.

A new rotative Census has been taking place in France since January 2004. Instead of an exhaustive counting of the whole population, it consists of yearly surveys on a part of the territory based on samples of municipalities or addresses (called “rotation group”, five of them are built over a five years cycle), one of them being covered by the Census each year. This very new technique implies to redefine the sample designs of all household surveys, the Census becoming the first phase of the new sampling design.

One of the most important principles is to use as a frame for the surveys of a current year the lists of dwellings covered by the Census of the previous year. But, since most of the surveys are face to face, the dwellings which must be surveyed have to be concentrated in some areas and not spread all over the territory, in order to reduce the costs. It implies to build primary units (PU) and to employ a network of interviewers located not far from those PU.

The first difficult work was to build fixed PU, under constraints of size and conditionally to the Census sample of addresses or municipalities. An automatic process of construction of PU has been developed for that purpose.

A major methodological issue was then to establish the sample frame of the PU. A sample of 525 PU was drawn, using the balancing technique developed by Deville and Tillé. A methodological work of simulation has led to establish the final PU sample design.

The next step was to check if the annual sample frames formed by the censused parts of the drawn PU are representative of the whole country. As some results tended to show that some specific populations were underestimated or overestimated by the drawn PU, a solution of calibration of the PU was established. It was then checked that this calibration procedure leads to have a more representative Master Sample.

The number of dwellings to be drawn in each PU is then calculated in order to minimize the variations of the final weights of dwellings in the sample (derived from the different steps of sampling), so that final weights for all the dwellings drawn will be as close as possible, under additional constraints of minimum and maximum number of dwellings drawn in each PU. We give here some details about the algorithm and some results on the variation of the final weights. In each drawn PU, dwellings are then drawn with equal probabilities with a systematic random sampling procedure.