Économie et Statistique n° 483-484-485The overhauled Census: progress in methodology and contribution to knowledge

Economie et Statistique
Paru le : 28/04/2016
Consulter

Municipal estimates exploit data from the 2011 Family and Housing Survey and the census: a high-risk operation

Pascal Ardilly

In summer 2013, INSEE distributed estimates to more than 1,400 municipalities of the headcount for various sub-populations living within their territory: people in a Civil Solidarity Pact, people in a couple but not living together, grandparents, elderly people living alone and whose children live close by, etc. These municipal estimates were based on national data collected for the Family and Housing Survey, a very wide-ranging survey associated with the 2011 annual census. To ensure that data was of good quality, a modelling procedure was used, of the “small fields” type. The first stage involved the entire survey sample. It consisted in modelling the individual probabilities of belonging to the sub-population of interest. The explanatory variables used were binary variables available in the census: sex, age group, marital status, etc. They therefore define the population categories within which these probabilities of belonging to the sub-population under consideration will be considered as homogeneous and independent of the municipality of residence. Once these probabilities were evaluated, we obtained municipal estimates by multiplying them by the municipal headcount of the associated categories (supplied by the census). Finally, a calibration was carried out on national headcount for the sub-population under consideration taken from the survey. Because the estimates are based on large samples, this procedure reduces sampling variance considerably compared with an estimate using only information from the survey at municipal level. On the other hand, using a model generates bias, as municipal behaviour is assimilated to supra-municipal behaviour. The resulting error can be appreciated at an aggregated level and it can be seen that, on the disseminated variables, it is usually still perfectly acceptable.

Economie et Statistique
No 483-484-485
Paru le : 28/04/2016