Publication Alert: Pre-Trained Nonresponse Prediction in Panel Surveys with Machine Learning

Pre-trained Models for Panel Survey Nonresponse Prediction

In a recent study published in Survey Research Methods, Christoph Kern of the SODA lab, together with John Collins from the University of Mannheim, demonstrate that machine learning models trained on one panel study can effectively predict nonresponse in entirely different panel studies. This "cross-training" approach allows new panels to leverage existing data rather than waiting to accumulate their own training data.

The researchers tested their approach on five German panel surveys with diverse designs, showing that nonresponse indicators—particularly response history and demographics—are so consistent across contexts that models achieve strong predictive performance (AUROC 0.75-0.85) even when transported between studies. This finding has immediate practical implications for newly launched surveys, which can now identify at-risk participants and implement targeted interventions during critical early waves when attrition is typically highest.

The work reveals that fundamental processes driving panel nonresponse are remarkably consistent across different survey contexts, opening new possibilities for how longitudinal studies manage participant retention from their inception.

For more information: https://ojs.ub.uni-konstanz.de/srm/article/view/8473