Abstract
<title>Abstract</title> <p>What strategies improve machine learning performance in tabular social science datasets? The Predicting Fertility (PreFer) data challenge invited dozens of researchers around the world to predict births and adoptions using Dutch survey data. Our model produced the most accurate predictions in the competition. This paper describes two strategies behind the model, in which we engineered the training data to create effectively “more” data. First, we outline a strategy we call “time-shift data augmentation,” which artificially increases the sample size by repurposing data across time periods. Second, we demonstrate “partner linkage,” which increases the number of features by linking data from each focal person with data from their spouse or cohabiting partner. Our findings provide a case study of how these two data engineering strategies improve predictive performance. Both strategies can be implemented on a wide range of social data, including longitudinal surveys and administrative records, and may improve predictive performance in a variety of settings.</p>