a) Relevance & Research Question:
Swiss based market and social research institute LINK and the University of Applied Sciences Northwestern Switzerland collaborated on a two year project to develop methods to measure and reduce the bias of surveys, accomplished with the LINK web-panel, the largest Swiss panel with random selection.
b) Methods & Data:
The bias shall be reduced by applying a weighting scheme based on propensity scores. The success depends mainly on the choice of covariates.
To determine which variables serve best to model the response propensity we have conducted an experimental survey. In a CATI survey, people were asked questions concerning psychological characteristics (Big Five Scale), about their set of values (Schwartz Scale), about the perceived costs and benefits from surveys and their internet behavior. Simultaneously, the CATI served as a recruitment interview for the web-panel.
After checking which of the willing CATI respondentsactually did react to their first online survey invitation, we developed a model to predict the response behavior. This model allowed us to boil down the set of possible explaining variables, to a few. These variables are the basis for calculating the propensity scores.
To allow a thorough study of the resulting estimators and their variances we conduct Monte-Carlo-simulations. Observed missing value mechanisms have been implemented and modified to compare estimations with propensity-score weighted data with the universe.
c) Results:
The experimental survey has been conducted so far and could be analyzed. The response behavior has been modeled and some crucial variables have been identified. We can show the explanatory power of the proposed scales. The next step is to start the simulations. First results are expected by the end of the year.
d) Added Value:
Our project is a comprehensive study about applying propensity scores as a method to reduce bias in a well established web panel. We are able to implement easy to use methods in every day web-panel research.