Un Point de Depart pour les Methodes Iteratives de Regression. Application a la Regression Lineaire Robuste

The importance of the starting point for iterative robust regression is well known since the ordinary least square solution often used doesn't always give good results. We suggest a new method called “regression by the sectors” (RPS) as a starting point to an iterative windsorisation of the original data (ITER). We used artificial samples to compare RPS and other quick techniques of fit: regression by median, Hinich-Talwar's method, L₁ norm minimization. Besides RPS-ITER algorithm is tested (examples and simulations) in relation to algorithms using different starting points associated to ITER. The results show the superiority of RPS-ITER.