Analyse comparative des méthodes de classifications : l'exemple du bien-être au travail

Many researches in human relation management, marketing or strategy try to identify profiles. Their shared aim is to study links between concepts. This keen interest for profiles creation takes into account samples' heterogeneity. This person-centered approach succeeds in conceptual clarification and managerial advice which better fits with the context than the ones which result from average estimation. To succeed, researchers can use several classification methods. The "traditional" ones are midpoint, median, scale center split, non-hierarchical method as K-means clusters, hierarchical methods or combined method. More recently, a new generation appears from American studies : mixture models. The first section of this paper is dedicated to a presentation and a critical study of these "traditional" and new generation "methods". Their limits and mixture models' advantages are described. Indeed, Latent Profile Analysis and Factor Mixture Analysis come with a large number of statistical indicators which define the profiles' number. The arbitrary part in this key step of the decision process is reduced but not totally deleted. To empirically illustrate the mixture models' comparative advantages, we used a multidimensional concept which comes from the organizational behavior field : Well-being at work. We used the EPBET scale (positive scale of well-being at work) on a large sample of French employees. This scale defines well-being at work starting from four dimensions : management, colleagues, environment and time. This example shows the dependency of the profiles' number and people allocation in these profiles on the traditional method which has been used. We find four to sixteen profiles according to the method. Thanks to the indicators, the mixture model methods show a stable classification. This statistical way is really useful to increase conceptual knowledge and to resolve decision stake in management science.