Research question/goal:

The project investigated the use of prediction models and innovative treatments to build adaptive survey designs (ASD) for reducing attrition and nonresponse bias in panel surveys. Specifically, the project paired advances in prediction modelling from the field of machine learning, including the use of neural network-based methods tailored to longitudinal data, with treatments that aimed to increase survey enjoyment (interesting survey topic) and reduce response burden (shorter surveys), next to the common differential incentives approach. A core component of the project work was the planning and implementation of a survey experiment in the GESIS Panel, where these treatments were administrated in the data collection wave of August 2023. A further focus was on assessing the transferability of the developed methodology to other panel study contexts.

The results showed that time series machine learning methodology such as recurrent neural networks (RNNs) and long short-term memory (LSTM) networks are a promising approach to predict nonresponse in panel surveys in complex, data-rich settings with considerable temporal dependencies. Pre-training nonresponse prediction models with data from a different panel proved to be a feasible option, especially if this panel shares design features with the target panel used for model deployment. Furthermore, the findings showed that an interesting survey content and higher cash incentives can effectively reduce nonresponse. The project also showed how both components—machine learning-based nonresponse prediction and different types of treatments—can be combined and studied in realistic simulations that can help panel managers to compare the likely implications of different ASD regimes.