Life Course Prediction for Labor Market Outcomes: Lessons from Register-Based Machine Learning
2026 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]
Machine learning has begun making its way into the methodological toolkit for the social sciences, but only very recently, with many questions still unanswered. Combining this toolkit with population register data offers a promising perspective into researching the human life course and the complexities within. This dissertation explores how Swedish and Finnish population registers combined with decision tree-based machine learning algorithms can be leveraged for labor market research, in particular for studying the predictability of long-term unemployment in early adulthood and lifetime disposable income. The theoretical foundations on this work lie at the intersection of the life course framework and how different forms of capital accumulate and transfer intergenerationally. Machine learning offers a useful tool to jointly model multiple domains in life, and how they together form trajectories towards these labor market outcomes.
Studies 1 and 2 concern long-term unemployment in early adulthood in the context of Finland and leverage data from the full birth cohort of 1987. The predictors for both studies are chosen based on theories and previous empirical literature on unemployment, and operationalized longitudinally where possible, some following the ages related to school transitions. Study 1 focuses on the extent to which long-term unemployment at ages 25–30 can be predicted, as well as the differences in predictability across the outcome groups and socioeconomic background. The results show that when compared to logistic regression, all the tree-based models fare better overall, but their performance is mostly due to correctly predicting the outcome class employed, while incorrectly predicting most of the unemployed. Regarding the socioeconomic origin, the low socioeconomic status (SES) group is more difficult to predict overall than the high-SES group, but within the low-SES group, the correct predictions for the unemployed increase. The findings highlight our limited understanding of the processes leading to unemployment, as well as the importance of learning more about sub-group predictability. Study 2 unpacks these results for the best performing model, Gradient Booster, by exploring the combinations that predict someone as: A) employed or unemployed, and B) correctly or incorrectly employed or unemployed. The aim is to further evaluate which processes and theories are the model’s blind spots, and the method used is Principal Component Analysis. For the employed group, parental unemployment was the prominent component steering the model wrong, while for the unemployed it was the combination of family cohesion, mental wellbeing and intergenerational transmission of education.
Study 3 focuses on lifetime disposable income in the context of Sweden, leveraging population registers for the birth cohorts of 1967–1971. This study explores the extent to which it is possible to predict a person’s income class based on their early life conditions and further unpacks this to answer which characteristics each income class shares, and which sub-groups end up in the same income class, through different combinations of characteristics. The predictors are based on theories and empirical literature, not only focusing on socioeconomic origins, but also geographic origins, family structure, and gender. The analysis of predictability compares logistic regression to two tree-based algorithms, as well as a composite model where a different model for each income group is chosen based on highest performance. The results find the machine learning models to outperform multinomial logistic regression, with the composite model being the best. The middle-income class has the highest predictive accuracy, while the lowest income class has the lowest. These predictions are then unpacked using Shapley Additive Explanations, which are further analyzed using Hierarchical Agglomerative Clustering. The results reveal heterogeneity both across and within income classes, with complex realities at the bottom of the income distribution, and a particularly streamlined path at the top.
This dissertation contributes to the field of life course prediction by exploring the predictability of the two labor market outcomes and highlights the importance of exploring the predictability of sub-groups. Furthermore, it presents ways in which population register data can be analyzed with machine learning methods to reveal more details about the human life course.
Place, publisher, year, edition, pages
Stockholm: Department of Sociology, Stockholm University , 2026. , p. 47
Series
Dissertation series / Stockholm University Demography Unit, ISSN 1404-2304 ; 34
Keywords [en]
life course, machine learning, labor market, unemployment, income, population registers, register-based, life course prediction
National Category
Sociology (Excluding Social Work, Social Anthropology, Demography and Criminology) Demography
Research subject
Sociological Demography
Identifiers
URN: urn:nbn:se:su:diva-260353ISBN: 978-91-8107-753-7 (print)ISBN: 978-91-8107-754-4 (electronic)OAI: oai:DiVA.org:su-260353DiVA, id: diva2:2106243
Public defence
2026-11-19, hörsal 5, hus B, Universitetsvägen 10 B and online via Zoom, public link is available at the department website, Stockholm, 13:00 (English)
Opponent
Supervisors
2026-10-272026-09-292026-10-15Bibliographically approved
List of papers