Election 2020 Forecasts Predict A Comfortable Biden Victory, BUT….
Election polling is considerably less precise than other sciences. Yet election forecasting is surprisingly accurate if you understand how polls work and the information contained in them.
Polling and election forecasts in early October 2020 suggest a comfortable- but not a landslide- victory for the Democratic Party’s nominee, Joe Biden, over the incumbent Republican, Donald Trump. If you want to guess what will happen on November 3rd, this is a pretty reasonable guess. But an educated guess is still a guess and sometimes guesses are wrong.
Election forecasts based on polling can be difficult for a number of reasons.
All polling includes a margin of error, meaning that the actual preferences measured in the poll might be different than the preferences identified by the poll. This occurs because the poll is just a sample of the larger population being measured. Polls that use appropriate political science methods can identify this margin of error. However, people often ignore this when reading or interpreting polls.
For example, consider a poll that shows Candidate A ahead of Candidate B by a margin of 50% to 48% with a +/-3% margin of error. If Candidate B were to win the election by a margin of 49% to 47%, the poll was not wrong. This outcome was within the margin predicted by the poll. A person claiming that this poll was inaccurate is not actually critiquing the poll but rather demonstrating their inability to understand it.
Polling in U.S. presidential elections is further complicated by the fact that most organizations produce national polls. Yet, the outcome of a presidential election is determined by the results in individual states. There is not a single presidential election in the United States, but more than 51 individual contests. (56 to be precise. This includes the winner-take-all outcomes of the 50 states, Washington D.C., and the 5 individual congressional districts in Nebraska and Maine.) Conducting accurate polls in individual states and congressional districts is time-consuming and expensive. Most polling organizations forego this and rely on national polls and past election results to estimate individual state outcomes.
This leads to problems, as was the case in 2016 when individual states deviate from national trends more than expected. Michigan, Wisconsin, and Pennsylvania were considered to be “safer” states for Hillary Clinton than other swing states like Florida, Ohio, Virginia, Nevada, and North Carolina. Neither the Clinton campaign nor the national polling organizations paid as much attention to these states, partially based on the belief that Hillary Clinton would have to lose the national popular vote in order for these states to swing sharply enough for her to lose them. As long as she was leading in the popular vote, this was unlikely to happen. But unlikely events do happen.
Sampling also creates challenges for scientific polling. To get a reasonably accurate poll, one does not have to interview everyone. For an electorate the size of the United States, a sample of 2000 voters is more than enough. Interviewing more than this will not increase the accuracy of the poll by a statistically significant margin. But to translate the interview sample into a meaningful snapshot of the electorate, the polling organization must have a good idea how the sample interviewed for the poll compares to the population of actual voters. Current demographic information, past election results, and stated voting intentions typically provide enough information to do this in a reasonably accurate way.
Reasonably accurate, however, can still be inaccurate if the inaccuracies are skewed in a particular direction. For example, a college education is a fairly good indicator of a person’s likelihood of voting. More educated people vote more frequently than less educated ones. In 2016, most polls weighted the responses of college-educated whites more than whites without a college degree because past elections suggest that they are more likely to vote.
2016 presented a unique challenge, however, in that college-educated whites that had voted Republican in the past expressed an unusually high propensity to switch to Hillary Clinton. Meanwhile, whites without a college degree were mobilized by the Trump campaign to a degree that polling organizations did not account for. Most polls gave more weight to college-educated whites and less weight to whites without a college degree. When Republican-leaning voters with a college degree sat out the election in greater numbers than anticipated and whites without a college education turned out more than expected, actual election outcomes shifted slightly from polling predictions.
What does this all mean for the 2020 election? One can presume that pollsters have learned from the 2016 election. Polling has generally become more accurate over time. However, polling is still an imprecise science and surprises are possible. Biden will likely win on November 3rd. But he might not.