Abstract
Do machine learning algorithms perform better than statistical survival analysis when predicting retirement decisions? This exploratory article addresses the question by constructing a pseudo-panel with retirement data from the Survey of Health, Ageing, and Retirement in Europe (SHARE). The analysis consists of two methodological steps prompted by the nature of the data. First, a discrete Cox survival model of transitions to retirement with time-dependent covariates is compared to a Cox model without time-dependent covariates and a survival random forest. Second, the best performing model (Cox with time-dependent covariates) is compared to random forests adapted to time-dependent covariates by means of simulations. The results from the analysis do not clearly favor a single method; whereas machine learning algorithms have a stronger predictive power, the variables they use in their predictions do not necessarily display causal relationships with the outcome variable. Therefore, the two methods should be seen as complements rather than substitutes. In addition, simulations shed a new light on the role of some variables—such as education and health—in retirement decisions. This amounts to both substantive and methodological contributions to the literature on the modeling of retirement.
Highlights
Do machine learning algorithms offer added value vis-à-vis traditional statistical methods for the prediction of retirement outcomes? This article compares several modeling techniques for retirement patterns in the light of this question
The research agenda proposed by Fisher et al [1] identifies several possible avenues for future research: the incorporation of the influence of variables such as country, gender, socioeconomic status, and health status—and their interaction—on retirement; the critique of current theories; the exploitation of longitudinal data; the construction of a common language for the discipline; interdisciplinary and holistic work; and the incorporation of selection bias and heterogeneity in the analysis
As Scharn et al [2] point out, more research is needed on the differential impact of pension determinants on men and women, the effects of ongoing changes in the labor market, the interaction of different domains, and the impact of policies
Summary
Do machine learning algorithms offer added value vis-à-vis traditional statistical methods for the prediction of retirement outcomes? This article compares several modeling techniques for retirement patterns in the light of this question. Numerous theoretical and empirical models have been used to assess its likelihood and determinants. The widespread availability of longitudinal data—whether from administrative databases or surveys—has opened up new research avenues. These data have been used for both basic and applied research. The research agenda proposed by Fisher et al [1] identifies several possible avenues for future research: the incorporation of the influence of variables such as country, gender, socioeconomic status, and health status—and their interaction—on retirement; the critique of current theories; the exploitation of longitudinal data; the construction of a common language for the discipline; interdisciplinary and holistic work; and the incorporation of selection bias and heterogeneity in the analysis. As Fisher et al [1] suggest, most of these issues can best be assessed by relying on complex databases, rich in the number of variables and observations
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
