Fertility prediction challenge, Episode I: does survey data beat population registries?

Social sciences uncovered many factors associated with fertility outcomes but rarely assessed their predictive ability. Benchmarking predictive ability can give us insight into which factors are most important and how well we can explain fertility behavior, and also drive scientific progress. However, prediction benchmarks in social sciences are still rare. 

We conducted a pilot fertility prediction benchmark at SICSS-ODISSEI. Seven teams competed to predict having a(nother) child within the next three years (2020-2022) based on data up to and including 2019. The first phase was based on the survey data from the LISS, and the second phase – on administrative data collected by Statistics Netherlands (CBS). 

For both datasets, the predictive ability is low: the best F1 score is 0.59 for LISS and 0.54 for the CBS data. The best models are only able to identify half of positive cases. In the case of LISS, the most important variable is fertility intentions, followed by other factors related to views and behavior (division of childcare labor, political views, frequency of participant's contact with the mother) and several socio-demographic variables (urban place of residence, marriage status, age, dwelling type, cohabitation, having children). This demonstrates the importance of time-dependent fertility intentions for predicting the timing of children.

The pilot's results show modest predictability of having a new child in the next three years, with theory-identified factors being important but not very predictive. To further test fertility theories, advanced methods such as neural networks and transfer learning should be used that can leverage huge longitudinal datasets and combine the strengths of survey data and administrative data.