Relevance and research question:
Self-selected survey participants typically differ from the general population with regard to demographic variables that are important for voting behaviour. For example, elderly persons and persons with a low educational level are usually underrepresented. To forecast the result of elections based on such non-probability samples, it is therefore necessary to statistically adjust for selection bias. We investigated whether the validity of an election forecast can be improved using propensity scores which rely on demographic characteristics to model the propensity to participate in web surveys.
Methods and data:
In the week prior to the 2009 German federal election to the Bundestag/Lower House of the Parliament, we surveyed a large self-selected sample of potential voters. Recording several demographic variables that are presumably relevant for political preferences, we also asked the participants to indicate their voting intention in the upcoming election. We then computed predictions for the outcome of the election using propensity score adjustments based on official Mikrozensus data. We compared these predictions with the outcome of the election, and also tested whether propensity adjusted Web-based convenience samples allow for a prediction that is competitive with probability sample-based benchmark estimates.
Results:
We found that propensity score adjustments were able to considerably improve the predictive validity of an Internet-based survey of a non-random sample of potential voters, leading to a much better forecast of the election outcome. Our results thus document the usefulness of propensity scores to improve the validity of election forecasts. However, we also noted several limitations of the approach.
Added value:
Propensity score adjustments are increasingly being used to control for selection bias, but there are few opportunities to validate this approach against external criteria. Forecasting the outcome of an election is a rare but important exception, because adjusted and unadjusted predictions can be directly compared to an election’s official result.