A Machine Learning Approach to Well-Being in Late Childhood and Early Adolescence: The Children’s Worlds Data Case

Abstract

Explaining what leads to higher or lower levels of subjective well-being (SWB) in childhood and adolescence is one of the cornerstones within this field of studies, since it can lead to the development of more focused preventive and promotion actions. Although many indicators of SWB have been identified, selecting one over the other to obtain a reasonably short list poses a challenge, given that models are particularly sensitive to the indicators considered.Two Machine Learning (ML) algorithms, one based on Extreme Gradient Boosting and Random Forest and the other on Lineal Regression, were applied to 77 indicators included in the 3rd wave of the Children’s Worlds project and then compared. ExtremeGradient Boosting outperforms the other two, while Lineal Regression outperforms Random Forest. Moreover, the Extreme Gradient Boosting algorithm was used to compare models for each of the 35 participating countries with that of the pooled sample on the basis of responses from 93,349 children and adolescents collected through a representative sampling and belonging to the 10 and 12-year-olds age groups. Large differences were detected by country with regard to the importance of these 77 indicators in explaining the scores for the five-item-version of the CWSWBS5 (Children’s Worlds Subjective Well-Being Scale). The process followed highlights the greater capacity of some ML techniques in providing models with higher explanatory power and less error, and in more clearly differentiating between the contributions of the different indicators to explain children’s and adolescents’ SWB. This finding is useful when it comes to designing shorter but more reliable questionnaires (a selection of 29 indicators were used in this case)