PLS-Calibration: a new calibration method

摘要

Calibration is a technique that adjusts the initial sample on margins that are supposed known for the whole population. Its basic idea consists of using the auxiliary information not taken into account during the sampling design, in order to increase the precision of the Horvitz and Thompson estimator. However, the improvement provided by the calibration estimator depends on the choice of the auxiliary variables as well as their number. In fact, the variance of the calibration estimator may increase very much when a huge number of auxiliary variables are used or when there is a strong multicollinearity. For this reason, some solutions have been proposed in the literature: the Principal Component calibration and the Ridge calibration. Through this paper we propose a new technique named the Partial Least Squares calibration (PLS calibration) that allows to avoid the multiollinearity problem. To show the effectiveness of our method in comparison with the Ridge calibration and the PC calibration, we applied the three methods on a data provided by Marocmetrie a Morrocan company specialized on TV audience measurement.