Online Program

Adaptation of Winsorization to weight share method (308130)

Keywords: Business Statistics, Winsorization, Kokic and Bell, Weight share method, profiling, unit problem

The French Structural Business Statistics (SBS) production system, known as ESANE, has two main uses:

- Production of statistics based on the European SBS regulation; - Estimation of businesses’ contributions to GDP for the national accounts. The system is based on a mix of exhaustive administrative fiscal data and data obtained on a random sample of the business population. ESANE is currently changing to produce estimates based on enterprises (economic definition) and no longer on legal units (juridical definition).

To do so, we have to transform the random sample of legal units in a random sample of enterprises. Since an enterprise is a set of legal units, we use the weight share method to move from the legal units sample to the enterprises sample.

The use of the weight share method has an impact on several aspects of the estimation process, and the following study is focused on the adaptation of the treatment of influential values.

Until now, influential values were treated by winsorisation of the variable turnover with Kokic and Bell's thresholds. Those thresholds minimize the mean square error of the estimator of the total of the winsorized variable but are based on the fact that the sampling design is a stratified simple random sampling, which is no more the case after a weight sharing.

Since the conditions of application of the Kokic and Bell method are not verified, we had to find an adaptation of the method.

Four adaptation scenarios have been considered and compared based on a simulation study involving 50 000 replications of the survey sample and tax data.