Heading into election day last November, the polls showed Joe Biden with a substantial lead in the race for president. In some surveys, he was ahead by double digits. When it was all said and done, 43'000 voters in Wisconsin, Georgia and Arizona made the difference. Yes, the dynamics of the electoral college are to be considered here. Nationwide polls are to be taken with caution as the U.S. presidential election is in fact not a nationwide election. This being said, the blunder is nevertheless remarkable. In 2016, polls were rather off, and the consensus was that the weighing of the raw data was the problem. In particular, the vote of rural, lower educated white voters was underestimated. Herding is another problem. No pollster wants to be the outlier presenting different data, so they all copy each other's flawed turnout model. By the time of the mid-term elections in 2018, everybody assumed that had been corrected. Yet, the polls were off again quite substantially in some places. There was barely a public survey showing Ron de Santis ahead in the race for governor of Florida, for example. Guess who is serving as governor there now. After the midterm election, everybody again thought that pollsters learned the lesson heading into the 2020 cycle, and we know what happened. Looking at it more closely, the polls were in fact rather accurate with respect to Joe Biden. Donald Trump, on the other hand, outperformed polls substantially: by more than 4% in Florida and Wisconsin, by 7% in Ohio.

A group of Democratic pollsters have joined forces to analyze what went wrong. I think that it is a good idea that they carried out this effort, but their results are rather disappointing. They write that Republican low-propensity voters were much more likely to end up voting then Democratic low-propensity voters. Trump voters were probably less likely to even participate in a poll, and if they did, they were more likely to hide their real feelings. So, it all comes back to the "shy Trump voter", a hypothesis that was already discussed before the election. I also think that Democrats must come to terms with the reality that the minority vote is not as monolithic as they think it is or want it to be. Moving forward, I don't think that more "accurate" weighing of data and more sophisticated models are necessarily the solution. Obviously, weighing is not getting the job done. I would be much more interested to see polls with bigger sample sizes with no weighing whatsoever, for example. And I also think that good public opinion research has as much to do with listening to voters than with statistics. Qualitative data could bring light into the blind spots.