Séminaire Octogone : "Statistical tests for strategy analysis"

Publié le 8 septembre 2011 Mis à jour le 1 mars 2012
le 28 septembre 2011
14h - 16h
MdR, salle D30
Maartje Raijmarkers (Université d'Amsterdam)

In many domains in cognitive development, children show large individual differences, that is large variation in performance within age group. If a theory predicts for a specific domain that there are categorical individual differences, we would like to test this prediction instead of assuming it in the analysis of the data. Moreover, in analyzing developmental data, if there exist categorical individual differences within age group, we would like to analyze performance in terms of different types of behavior. Averaging performance within age group could seriously obscure results. For example, according to Kendler et al. during development children show different types of trial-and-error learning, associative learning versus hypothesis testing strategies, as opposed to gradually becoming more efficient in hypothesis testing (cf. Schmittmann, Visser & Raijmakers, 2006). In perceptual classification, as another example, it was hypothesized that children develop from holistic to analytic processing, as opposed to becoming more consistent in analytic processing (cf. Raijmakers, Jansen & vanderMaas, 2004). Also for the case of children's intuitive ideas about physics and astronomy, for example, it was hypothesized that individual differences are characterized by different types of performances, i.e. children have different types of mental models, as opposed to increasingly accurate knowledge (cf. Straatemeier et al., 2004; Raijmakers, van Es & van Schijndel, in prep.). We have been developing a methodology to test for the existence of different types of behavior based on categorical latent variable techniques, such as latent class analysis. With these techniques we can test instead of assume the presence of different types of behavior. Moreover, we are able to reveal unexpected types of performances (including guessing). I will illustrate the importance of using of statistical techniques instead of using pattern matching techniques for classification. For the case of mental models I will show how different domains result in different conclusions concerning the mental models versus fragmented knowledge discussion.