Prediction-based Adaptive Designs for Panel Surveys

Research question/goal: 

Despite its promising potential to reduce attrition and biases, the use of adaptive survey designs in panel studies is lacking in two critical areas: (1) in predicting nonresponse and thus creating appropriate strata and (2) in the treatments that are administered in practice. This project will pair the implementation and testing of innovative prediction methodology from the field of machine learning with innovative treatments that can be assigned to likely nonrespondents. Prediction models will be trained and evaluated in a longitudinal framework that is tailored to identify panellists at risk of nonparticipation in a given (new) panel wave. The predicted risk scores of the most accurate model allow us to test the effectiveness of different treatments. Specifically, this project will compare the usage of innovative treatments in adaptive survey designs that aim to increase survey enjoyment to the more common differential incentives approach. Testing these strategies on a common ground will add to previous research on adaptive designs, which has been inconclusive about which approach is best suited to stimulate respondents’ participation and engagement. Furthermore, the treatments will be compared and evaluated not only with respect to their effects on participation but also with regard to other, potentially unintended consequences for data quality in the long run. In addition, the transferability of the developed methodology to other panel studies will be investigated.

Current stage: 

During the first year of the project, we developed and implemented the prediction models that form the basis for the adaptive design of the GESIS Panel. One focus was on the comparison of standard machine learning approaches and prediction techniques that take the time series nature of panel data into account. A paper that summarises these model comparisons has been submitted for publication and is currently under review.

Fact sheet

2022 to 2025
Data Sources: 
Panel survey data
Geographic Space: