Does Trump Really have a 30% Chance of Winning?
Over the past few election cycles, the 538 blog has developed a reputation as the premiere election forecasting website. Like nearly every forecaster, they got the 2016 outcome wrong, but they did give Trump about a 3 in 10 chance of winning going into election day, higher odds than nearly everyone else gave him.
This year, their updated model again, as of Sep 9, gives Trump a 27% chance of winning in the electoral college. This is true even though Biden maintains a 7.6% lead in their weighted polling average.
How is it possible that Trump would retain such a realistic chance of retaining the Presidency even at such a significant deficit in the polls?
To answer this question, I dug into the weeds of the 538 model’s methodology. Before getting into the details, let me give you the headline: the primary reason that Trump’s chances appear significant is the possibility for changes in the polls between now and election day. (As opposed to either major polling error or his winning the electoral college despite heavily losing the popular vote.)
If the polls remained as they are today on Nov 3, the model would probably give Biden about a 90% chance of winning.
The gap between 73% and 90% can be explained by a number of factors, some of which, in my opinion, are overstated in the 538 model.
How Their Model Works
The 538 model is built on a state by state basis. If you compare the current polling average to the forecast, you can see that, in six out of seven key swing states, the forecast anticipates Biden’s lead to shrink. Here’s a look at the differences in the seven states most likely to tip the election:
Florida. Biden leads by 2.8, but is forecast to win by 2.0
Wisconsin. Biden leads by 7.0, but is forecast to win by 5.8
Arizona. Biden leads by 5.1, but is forecast to win by 3.2
Michigan. Biden leads by 7.4, but is forecast to win by 7.7
Pennsylvania. Biden leads by 5.0, but is forecast to win by 4.3
North Carolina. Biden leads by 1.7, but is forecast to win by 0.8
Minnesota. Biden leads by 6.1, but is forecast to win by 5.7
There are a few factors built into their model which cause the forecast to predict a narrowing of the margin. I’ll evaluate them one by one.
Historical partisanship. The model gives a weighting to how a state voted in 2016 (as well as a smaller weighting to how it voted in 2012). Since all the above states except for Minnesota voted for Trump in 2016, the model assumes that the opinion in that state will, to some extent, revert to its historical norm. This would seem like a legitimate reason to expect these states to tighten, except that it doesn’t factor in the more recent data from the 2018 mid-term election. For example, while Pennsylvania voted for Trump in 2016, in 2018, Democrats beat Republicans 55%-45% in total House of Representatives votes.
Incumbency advantage. There is an incumbency bonus baked in for Trump. This is apparently based on the record of incumbents winning more often than they lose: 14 of 19 since 1900. But the data on incumbents with approval ratings where Trump has consistently hovered (41-44%) tells a totally different story. Only Harry Truman in 1948 was re-elected with an approval rating as low as Trump’s. It seems to me that Trump’s incumbency is a disadvantage for him, compared with 2016, when people could not be sure what type of president he would be.
Economic Fundamentals. The economic fundamentals incorporated into the 538 model include six equally weighted variables. Three are pretty strong: income, inflation, and the stock market. Three are pretty weak: jobs, spending, and manufacturing. As a result, the economic fundamentals are seen as neutral. Neutral fundamentals tug the forecast to the center, reducing Biden’s expected performance. Given the unprecedented nature of this year’s economy, all historical analysis of the effect of economic fundamentals on voting seems of little value. I think people will see in the economy what they want to see: Trump partisans will remember a pre-Covid economy and focus on our returning to normal. Biden partisans will point to widespread continuing unemployment. The idea that people who currently oppose Trump will switch to supporting him over the next two months because of strong economic fundamentals is pretty farfetched. I can’t see this as a reason to expect the race will tighten.
Uncertainty. The model attempts to assess the level of uncertainty in the race. The theory is that higher uncertainty implies greater likelihood of shifting polls over the next two months. To measure uncertainty, it computes a number of factors:
- The number of undecided voters
- Sum of undecided voters and 3rd party voters
- Polarization of votes in the U.S. House of Representatives
- Volatility of the polling average
- Amount of national polling
- Size of the gap between the polling forecast and the fundamentals forecast
- Economic volatility
- Amount of big news (measured by headline size in the New York Times)
The first five points all point to very low uncertainty. The final two point to very high uncertainty. As a result, the model considers uncertainty to be roughly average compared to prior elections.
Here is where I find the biggest flaw in the 538 model’s methodology. Economic volatility and substantial big news do not need to be considered separately, as they should already be “baked in” to the calculation through their impact on the polling average. We would expect high economic volatility and major news events to swing the polls.
Moreover, this calculation assumes a continuation of prior trends, which is not necessarily a given. In fact, a good case could be made that after a period of unprecedented economic volatility, the next two months will likely show less volatility than the previous six.
In any case, the fact that polling averages have been very stable, even in the face of unprecedented economic volatility and boatloads of big news, suggests that people have made up their minds and are unlikely to change them. If the enormous swings in GDP and employment and the fluctuations in COVID-19 cases and the unrest related to Black Lives Matter protests have barely moved the polling averages, what could move them?
Given that each of the various components of the 538 model which suggest the race will tighten appear to be overstated, Trump’s odds of winning are probably a fair amount less than 3 out of 10.