Abstract
Can Facebook engagement tell us who will win an election? I compare parties' shares of interactions on their Facebook pages in the 30 days before 31 national elections in 15 European countries between 2011 and 2020 with their vote, their final polls and their previous result, for 172 party-elections. As a forecast on its own, engagement was poor: its mean absolute error was 9.6 points, against 1.8 points for the final polls and 5.8 points for simply repeating the previous election. It did worse than the polls in every one of the 31 elections, and it overstated populist parties by about ten points. Engagement did carry some information: on its own it explained about a third of the variance in the change in parties' votes since the previous election. But almost all of that was already in the polls. Where a party's share of engagement ran ahead of its poll share, it beat its poll only slightly and not significantly, and adjusting the polls using engagement made forecasts worse in a test on elections left out of the estimation. In the largest poll misses engagement often pointed the same way as the error, but these cases come from a few elections with few parties covered, and an apparent link for populist parties disappears without Slovakia's 2016 election. On this evidence, engagement is a measure of attention, not a poll, and it should not be read as one.
Cite
Hotham, T. (2026). Engagement Is Not a Poll: Facebook and 31 European Elections, 2011 to 2020. Preprint, Zenodo. https://doi.org/10.5281/zenodo.23396654