Longitudinal Nonresponse Prediction with Time Series Machine Learning

Abstract Panel surveys are an important tool for social science researchers, but nonresponse in any panel wave can significantly reduce data quality. Panel managers then attempt to identify participants who may be at risk of not participating using predictive models to target interventions before data collection through adaptive designs. Previous research has shown that these predictions can be improved by accounting for a sample member’s behavior in past waves. These past behaviors are often operationalized through rolling average variables that aggregate information over the past two, three, or all waves, such as each participant’s nonresponse rate. However, it is possible that this approach is too simple. In this paper, we evaluate models that account for more nuanced temporal dependency, namely recurrent neural networks (RNNs) and feature-, interval-, and kernel-based time series classification techniques. We compare these novel techniques’ performances to more traditional logistic regression and tree-based models in predicting future panel survey nonresponse. We apply these algorithms to predict nonresponse in the GESIS Panel, a large-scale, probability-based German longitudinal study, for surveys conducted between 2013 and 2021. Our findings show that RNNs perform similar to tree-based approaches, but the RNNs do not require the analyst to create rolling average variables. More complex feature-, interval-, and kernel-based techniques are not more effective at classifying future respondents and nonrespondents than RNNs or traditional logistic regression or tree-based methods. We find that predicting nonresponse of newly recruited participants is a more difficult task, and basic RNN models and penalized logistic regression performed best in this situation. We conclude that RNNs may be better at classifying future response propensity than traditional logistic regression and tree-based approaches when the association between time-varying characteristics and survey participation is complex but did not do so in the current analysis when a traditional rolling averages approach yielded comparable results.