Abstract

Machine learning (ML) is increasingly explored as a method in survey methodology to improve data collection by predicting response behaviour. This paper examines whether ML can predict high respondent burden and survey break-off in the Norwegian Labour Force Survey (LFS) using paradata, demographic variables, and expert-defined questionnaire indicators.

Initial results show limited predictive performance across both telephone and web modes, with low precision and F1-scores. A key challenge is strong class imbalance, combined with potential measurement issues in both target variables and predictors. The modelling approach is therefore extended by addressing class imbalance and refining features and target definitions to assess potential improvements.

In addition, the Total Machine Learning Error (TMLE) framework is applied to assess potential sources of error.

The analysis also introduces response burden as a central dimension by comparing predicted and observed response burden. The relationship between these measures is examined to assess how well burden can be captured and predicted. Break-off and response burden are analysed in parallel, as examining both outcomes within the same survey provides a more holistic understanding of response behaviour and helps inform feature selection and modelling strategies.

The paper highlights both the potential and the limitations of ML for predicting response behaviour in complex survey environments and provides practical insights for respondent-centred survey design.