A comparative evaluation of dissimilarity-based and model-based clustering in science education research: the case of children's mental models of the Earth. (24th April 2020)

• Record Type: Journal Article
• Authors: Stamovlasis, Dimitrios; Vaiopoulou, Julie; Papageorgiou, George
• Is Part Of: International journal of data analysis techniques and strategies. Volume 12:Number 3(2020)
• Page Start: 247
• Page End: 261
• Publication Date: 2020-04-24
• Subjects: mental model -- latent class analysis -- partitioning around medoids -- dissimilarity-based clustering -- coherent mental model hypothesis -- fragmented knowledge hypothesis -- science education -- model-based clustering; Electronic data processing -- Periodicals; Database searching -- Periodicals; 005
• Languages: English
• ISSNs: 1755-8050

Abstract

In the present work, two different classification methods, a dissimilarity-based clustering approach (DBC) and the model-based latent class analysis (LCA), were used to analyse responses to a questionnaire designed to measure children's mental representation of the Earth. It contributes to an ongoing debate in cognitive psychology and science education research between two antagonistic theories on the nature of children's knowledge, that is, the coherent versus fragmented knowledge hypothesis. Methodology-wise the problem concerns the classification of response patterns into distinct clusters, which correspond to specific hypothesised mental models. DBC employs the partitioning around medoids (PAM) approach and selects the final cluster solution based on average silhouette width, cluster stability and interpretability. LCA, a model-based clustering method achieves a taxonomy by employing the conditional probabilities of responses. Initially, a brief presentation and comparison of the two methods is provided, while issues on clustering philosophies are discussed. Both PAM and LCA attained to detect merely the cluster which corresponds to the coherent scientific model and an artificial segment added on purpose in the empirical data. The two methods, despite the obvious deviations in cluster-membership assignment, finally provide sound findings as far as hypotheses tested, by converging to identical conclusions.