Predicting fertility outcomes with networks

People’s networks are considered an important factor in explaining whether people want and have children. Existing research has rarely quantified how well we can predict fertility outcomes, and the importance of networks relative to other factors is entirely unknown. Here we use unique data from a representative sample of Dutch women reporting on over 18000 relationships. We use several machine learning techniques, ranging in complexity from LASSO regression to Graph Neural Networks, to examine how well we can predict five different fertility preference variables. Top models accounted for 5 to 45% of the out-of- sample variation in the different outcomes. Tree-based methods (e.g., XGBoost) performed best but only by a small margin. Graph Neural Networks required least preprocessing of the data and led to high predictive ability scores and thus presents a useful tool for personal network analyses. We discuss how the differences in predictive ability across different techniques are useful in understanding what traditional models may miss. We further discuss to what extent our results provide support for different mechanisms of social influence, and conclude that particularly those people in the network desiring children or those choosing to be childfree are important and understudied.

Summary

     Predicting fertility outcomes with networks

     Dutch Day Demography, Utrecht

     Click here for website

     Download materials here

Description

People’s networks are considered an important factor in explaining whether people want and have children. Existing research has rarely quantified how well we can predict fertility outcomes, and the importance of networks relative to other factors is entirely unknown. Here we use unique data from a representative sample of Dutch women reporting on over 18000 relationships. We use several machine learning techniques, ranging in complexity from LASSO regression to Graph Neural Networks, to examine how well we can predict five different fertility preference variables. Top models accounted for 5 to 45% of the out-of- sample variation in the different outcomes. Tree-based methods (e.g., XGBoost) performed best but only by a small margin. Graph Neural Networks required least preprocessing of the data and led to high predictive ability scores and thus presents a useful tool for personal network analyses. We discuss how the differences in predictive ability across different techniques are useful in understanding what traditional models may miss. We further discuss to what extent our results provide support for different mechanisms of social influence, and conclude that particularly those people in the network desiring children or those choosing to be childfree are important and understudied.