Scientific Presentations

LabTalks­@DCD

Department of Digital and Computational Demography
Max Planck Institute for Demographic Research (MPIDR), Rostock, Germany, July 08, 2025

03:00 PM: Talk with Elizaveta Sivak - Model interpretation for life course research

Abstract

A prediction-focused, data-driven approach in the social sciences offers a new, potentially powerful tool for exploratory research. By identifying the most predictive model – the one that best captures patterns in the data – and interpreting it, we can generate data-driven hypotheses about causal mechanisms underlying various life outcomes and discover new predictors for further investigation. However, what if different models, all with similarly strong predictive performance, offer different ‘explanations’? This phenomenon is known as the Rashomon effect: multiple models can have the same performance despite using different predictors and uncovering different relationships in the data. In light of the Rashomon effect, there is a growing argument that these relationships should be studied based on a set of good-performing models rather than just one. In the workshop, we will provide an example and general guidelines for drawing more robust data-driven insights about potential relationships in the data based on a set of models. We will explain how we are studying predictors of fertility outcomes based on the top-performing models submitted to the PreFer data challenge and discuss the challenges in aggregating information from the best-performing models. We will also review the ways to establish the best set of models for a given predictive task and how to aggregate information from them.

About

Elizaveta Sivak is a PhD candidate at the Department of Sociology, University of Groningen, where she studies the predictability of fertility. Her scientific interests include the predictability of life outcomes, especially related to parenting and childhood, and how predictive approaches can contribute to a better understanding of social phenomena