Forecasting presidential elections: What about within-mode effects?

Recent elections have shown that polls are not equal in terms of modes of administration (Durand et Johnson, 2021). In the U.S. 2020 election, 51 pollsters published 222 national polls from September 1st to election day. The polls that used mixed-mode or multiple sampling sources produced significantly better estimates of the results than polls using sole opt-in panels. Besides, these polls did not trace similar trends in support for Biden. We compare this situation with the French presidential election where ten pollsters published 126 polls from January 1st to April 8. Almost all the pollsters used opt-in panels, often the same source. One pollster used a more probabilistic method to sample voters. His estimates were significantly different and tended to be higher for the more extreme candidates – on the right and on the left alike. The presentation concludes on the necessity to monitor emerging developments in electoral polls.