Discriminant Analysis Versus Random Forests on Qualitative Data : Contingent Valuation Method Applied to the Seine Estuary Wetlands
Résumé
Contingent valuation method is a technique used to quantify the value of non-market resources
such as environmental features. A Contingent Valuation survey was carried out to establish the
Willingness To Pay (W.T.P.) of local populations for the conservation of the Seine Estuary
Wetlands, an important and threatened ecological area in Northern France. Our objective was to
build a model that predicts the WTP using known samples of the same survey. As the predictors
were mostly qualitative variables (nominal or ordinal), we applied a procedure to transform them
into quantitative variables. We analyzed the multiple correspondences of the predictors i.e. of the
correspondences of the complete binary table. The p selected explanatory variables X1 , X2 ,…, Xp
were replaced by the co-ordinates on q factorial axes (with q ≤ p ) with weighting to preserve the
importance of the components.
Two methods of classification were implemented: Discriminant Analysis and Random Forests, with
the aim of comparing their performances in terms of classification. Our results for this survey show
that Discriminant Analysis gives better predictions than Random Forests.
This deserves mention because in previous comparative studies (with methods other than
Discriminant Analysis), Random Forests had consistently been shown to be superior to the other
methods. These results need further investigation.