What Voting Advice Applications can teach us about Voters and Elections

VAA sites generate rich sources of data, potentially allowing researchers to address a range of questions related to voting and political attitudes and behaviour. However, the representativeness of VAA samples is questioned because they are opt-in rather than samples selected through randomly generated process. In this paper, we make a case for using VAA generated data to analyse voting intentions. We find that with proper statistical adjustment, non-representative samples can be used to forecast election outcomes, offering similar results to national representative samples. Our analysis proceeds as follows: first, using multilevel regression and postratification, we demonstrate how to weight voluntary data obtained via a VAA to create a ‘representative’ sample. Second, to assess this technique, we employ opt-in sample from 2014 EUvox and official statistics data, compared against exit polls. Our results show that VAA samples are particularly useful in understanding vote preferences amongst supporters of small parties.