📄 social_science

Who Will Have a Child? Predicting Three-Year Childbearing Outcomes Using Machine Learning

This study utilizes machine learning models, specifically XGBoost enhanced by SMOTE for class imbalance and domain-informed feature selection for high dimensionality, to successfully predict individual three-year childbearing outcomes in the Netherlands using PreFer survey data.

Original authors: Li Ya Chang, Mojtaba Rostami Kandroodi

Published 2026-09-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Li Ya Chang, Mojtaba Rostami Kandroodi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every society faces a fundamental question of planning: how many children will be born in the coming years? This is not merely a matter of curiosity for demographers; it is a practical necessity for governments and communities. Knowing the likely number of new families helps schools plan for classrooms, hospitals prepare for maternity wards, and cities design housing for growing populations. For decades, researchers have tried to forecast these numbers by looking at broad trends, such as the average age of first-time parents or the overall economic climate. However, predicting whether a specific individual will have a child within a short window of time, such as the next three years, has remained a difficult puzzle. The factors influencing this decision are vast and deeply personal, ranging from financial stability and relationship satisfaction to career goals and the simple desire to start a family.

A team of researchers at Tilburg University in the Netherlands recently tackled this challenge by turning to a powerful set of tools known as machine learning. Instead of relying on a single formula to guess the future, they trained computer programs to recognize complex patterns in the lives of thousands of people. Their goal was to see if these digital models could accurately predict which individuals, aged between 18 and 45, would welcome a new child between 2021 and 2023. The researchers worked with a massive collection of survey data that contained over 30,000 different pieces of information about each person, from their birth year and education level to how they divide household chores and how satisfied they are with their partnership. The central difficulty they faced was that the data was heavily skewed; far more people in the study did not have a new child than did, making it easy for a computer to guess "no" for everyone and still appear accurate, while missing the actual parents entirely.

To solve this, the researchers first had to simplify the overwhelming amount of information. They began by using their knowledge of social science to identify which of the 30,000 variables were most likely to matter, narrowing the field down to a more manageable list. They then used a computer program to test which of these factors were truly the strongest predictors, eventually settling on a focused set of just 69 key variables. This process was crucial because trying to use all 30,000 details at once would have confused the model, much like trying to find a single specific book in a library where every shelf is filled with thousands of unrelated titles. By focusing on the most relevant signals, such as age, partnership status, and financial situation, they allowed the computer to see the forest rather than getting lost in the trees.

The team then tested three different types of machine learning models to see which one could best identify the future parents. They compared a basic decision tree, which makes predictions by asking a series of yes-or-no questions, against two more advanced methods that combine many such trees to reach a smarter conclusion. One of these advanced methods, known as extreme gradient boosting, worked by building a sequence of models where each new one learned from the mistakes of the one before it. The researchers also had to address the imbalance in their data, where the group of people having a new child was much smaller than the group having none. To fix this, they used a technique that created synthetic examples of the smaller group, effectively teaching the computer to pay closer attention to the rare event of a new birth without simply copying the same data over and over.

The results showed that the most sophisticated model, the one that learned from its own errors, performed the best. When tested on data it had never seen before, this model correctly identified the future parents with a high degree of accuracy, significantly outperforming the simpler methods. The use of the technique to balance the data proved essential; without it, the model struggled to spot the minority group of new parents. The study found that by carefully selecting the right questions to ask and using a model that could learn from its mistakes, it is possible to forecast short-term childbearing outcomes with surprising precision. While the researchers noted that their findings are based on a specific group of people in the Netherlands and would need further testing on other populations, the work demonstrates that machine learning can be a valuable tool for understanding the complex human decisions behind family growth.

Technical Summary: Predicting Three-Year Childbearing Outcomes Using Machine Learning

Problem Statement
This study addresses the challenge of accurately forecasting short-term fertility outcomes at the individual level, a task critical for public policy and infrastructure planning but historically difficult to predict. Specifically, the research focuses on predicting whether individuals in the Netherlands (aged 18–45) will have a child within a three-year window (2021–2023). The study operates within the framework of the "Predicting Fertility in the Netherlands" (PreFer) data challenge. The dataset presents two significant methodological hurdles:

High Dimensionality: The dataset contains over 30,000 variables (features) derived from longitudinal survey data (LISS Core Study and Background surveys) but only 1,382 observations, creating a "small $n$ , large $p$ " scenario prone to overfitting and the curse of dimensionality.
Class Imbalance: The target variable is binary (having a new child vs. not), with a substantial skew toward the negative class (individuals without a new child), which can bias standard classification algorithms toward the majority class.

Methodology
The authors employed a structured machine learning pipeline utilizing secondary analysis of existing survey data. The methodology proceeded through the following stages:

Data Preparation: The dataset was split into a training set (approx. 70%, 987 instances) and a holdout set (approx. 30%, 395 instances) by the challenge organizers. For internal model development, the training set was further stratified into training (80%) and test (20%) subsets to ensure representative class distribution.
Feature Selection: To mitigate the high dimensionality, the authors first curated an initial set of 325 features based on established social science literature regarding fertility determinants (e.g., age, partnership status, income, work-life balance). These features were then evaluated using feature importance scores from a Random Forest classifier. Through cross-validation, the feature set was iteratively reduced, identifying an optimal subset of 69 features that maximized predictive performance while minimizing complexity.
Handling Class Imbalance: The Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training set. Unlike random oversampling, SMOTE generates synthetic minority class instances by interpolating between existing samples. This balanced the training data (creating an equal number of instances for both classes) without simply duplicating data, aiming to reduce overfitting risks.
Model Selection and Training: Three tree-based algorithms were evaluated: Decision Tree (DT), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). These were chosen for their proven efficacy on tabular data. Hyperparameters were tuned using GridSearchCV with 5-fold cross-validation on the training set.
Evaluation Metrics: Given the imbalanced nature of the problem, the primary performance metric was the F1-score (the harmonic mean of precision and recall), aligning with the PreFer challenge objectives. Accuracy, precision, and recall were also reported for comprehensive analysis.

Key Results

Feature Reduction Impact: Reducing the feature set from the initial 325 (and the original 30,000+) down to 69 variables significantly improved model performance. The submission using 69 features achieved the highest F1-score among the authors' five challenge entries.
Model Performance: XGBoost consistently outperformed both Decision Trees and Random Forests. On the holdout set, XGBoost achieved an F1-score of 0.73, compared to 0.68 for RF and 0.66 for DT. This performance ranked 2nd among the 40 participating teams in the challenge.
Impact of SMOTE: The application of SMOTE improved the F1-score on the holdout set from 0.68 to 0.73. This improvement was driven primarily by an increase in recall (from 0.59 to 0.69), indicating the model became better at identifying the minority class (individuals having a child). This gain came with a slight trade-off in precision, a common characteristic of oversampling techniques.
Confusion Matrix Analysis: The use of SMOTE reduced false negatives (missed childbearing events) from 36 to 27 on the holdout set, confirming its utility in capturing rare events in imbalanced datasets.

Significance and Claims
The paper claims that careful feature selection, appropriate model choice, and the treatment of class imbalance are critical for improving the prediction of childbearing outcomes from high-dimensional survey data. The study demonstrates that:

Feature Selection is Vital: Reducing a massive feature space to a concise, relevant set (69 predictors) enhances predictive power and reduces overfitting risks in demographic forecasting.
XGBoost is Superior: Among the tested tree-based ensemble methods, XGBoost provided the most robust generalization to unseen data for this specific task.
SMOTE Effectiveness: Synthetic oversampling effectively improves the identification of minority class instances in fertility prediction, directly addressing the challenge of class imbalance.

The authors maintain a modest stance, acknowledging that while the model performed well on the holdout set, the small sample size relative to the feature count remains a limitation. They note that external validation on independent populations is necessary to fully assess the generalizability of these findings, and the study does not claim to have solved the broader problem of fertility prediction but rather offers a validated methodological approach for this specific context.

Technical Summary: Predicting Three-Year Childbearing Outcomes Using Machine Learning

More like this