Stacked Domain Learning for Interdisciplinary Theory Development
In recent decades, there has been increasing attention to the benefits of integrating predictive models into psychological research and how they can be used for theory development (Yarkoni & Westfall, 2017; Shmueli, 2010). By shifting the focus from model fit to predictive accuracy, theories can be formed that are more generalizable to new samples and therefore better used in applied contexts where real-world problems require methods that ensure reliable predictions. This is among other benefits such as finding undiscovered patterns and generating new hypotheses or comparing competing theories in terms of their generalizability (for a comprehensive discussion, see Yarkoni & Westfall, 2017; Cranmer & Desmarais, 2017).
In this project, the aim is to develop a new predictive modeling framework called Stacked Domain Learning that can be used for theory development in interdisciplinary research where data is collected from multiple research domains. The goal is to combine domain-specific theories based on predictive accuracy to develop an interdisciplinary theory and form hyperdomains. To this end, a combination of stacked generalization (Wolpert, 1992; Breiman, 1996) and regularization is used. In a multi-step process, cross-validated predictions are first generated for each domain based on domain-specific theories and methodology, and then used in a meta-model that applies regularization techniques to evaluate predictive accuracy and select which of the domains are relevant and to what extent. This methodological approach builds on earlier work by van Loon et al. (2020a, 2020b, 2022), who developed the procedure for analyzing multiview data based on regularized GLMs.
By exploring a framework that seeks to leverage the use of different statistical learning techniques to contribute to theory development in interdisciplinary research contexts, the project fits well within the IOPS program.
Prof. dr. H.M. Huizenga
Prof. dr. M.J. de Rooij
Dr. M. Fokkema
1 October 2023 – 31 September 2027