Projecting the Presidential Election
What if we’re all wrong? What if we’re all wrong about how we analyze and project the Presidential Election? If one watches political analysis on TV or the internet, the most common reference used in analyzing the future is polls.
Analysts would discuss the results of the poll and what it portends for the candidates running. Some more mathematical-inclined individuals may make a statistical model to project the election. However, despite using other variables, the most prolific models still are based on polls or heavily impacted by polls.
In theory, using polls makes sense. After all, the idea of a poll is to garner enough information to give a snapshot of the population by testing a small sample. By basing their results on a normal population curve and generating a random sample, the goal is to find results wherein the pollster can confidently predict the atmosphere of the election at the time of polling.
While confidence can change per the pollster, most if not all modern polls release data wherein they are 95% confident of the result. Meaning that 95% of the time, the results within the poll (including its margin of error) match the population as a whole. If the polls were 95% accurate, then using the polls as a reference would make sense.
Unfortunately, polls are not this accurate. Before we criticize the results of polling, it is important to mention the difficulties in polling. Because the biggest challenge to accuracy is not the pollsters themselves but rather a much more challenging atmosphere.
In the 1980s, a pollster could reasonably expect an 80% response rate when making their calls. As long as the phone numbers were random, then the results should follow the normal population curve and the natural law of statistics (Zukin 2015).
If that same pollster, were to make phone calls a decade ago, they could expect a response rate of 8% (Zukin 2015). Since then, the number has actually declined even further (Kennedy and Hartig 2019).
If most of the population is unwilling to answer a survey, then garnering accurate results would be challenging. Add on top of that the possibility that some groups, whether demographically or politically, are more or less likely to respond to surveys; therefore, this creates a scenario where pollsters have a much harder time judging the population.
To correct for a skewed sample, they may weight groups to fit better with the population as a whole or to the projected population that will vote in November. But, this poses a serious issue. What if their weighting is wrong?
In the 2020 election, per Real Clear Politics, there were 12 “toss-up” states. These states were Florida, Georgia, North Carolina, Pennsylvania, Ohio, Michigan, Wisconsin, Iowa, Minnesota, Texas, Arizona, and Nevada. Out of the 66 state polls used within their average, only 32 polls correctly predicted the election results.
(For this article, the definition of correctly predicting is the actual election result margin was within the margin of error of the poll. For example, Monmouth had a poll for Georgia where they predicted Biden would win by a 4% margin with a Margin-of-Error of 4.4%. Since Biden won the state by .3%, this poll was considered accurate.)
So, less than half the polls were accurate and only half the states per their average fell within 2%. In some states, the polling missed completely. For Texas, only one of four polls was accurate. For both Iowa and Michigan, only one poll out of five was accurate. But the worst state was Ohio; out of four polls, none of them were accurate.
On the other hand, Minnesota, Nevada, and Pennsylvania had two-thirds of theirs which were accurate; Minnesota and Nevada had three polls apiece and Pennsylvania had six. Georgia had five polls out of six that were accurate and North Carolina had seven out of eight polls that were accurate.
But even in the “accurate” states, there seem to be some discrepancies. Georgia was predicted to be won by Trump by 1% per the average, which may be considered a material error.
However, Nevada and Pennsylvania had a difference of 0% from their average margin compared to real-life results. But, looker closer at the data, this appears to be more a happy coincidence than an actual achievement. In Nevada’s three polls, two polls had Biden winning; with margins of 6% and 2%. A third poll had Trump winning by 1%.
Combined, their average matched the election result of 2.4%. However, this was only the case because two of the polls were counterweights for each other. While Pennsylvania was less drastic than Nevada, the accuracy was not due to consistency but rather the errors from the pollsters countering each other. In fact, when there seems to be a degree of consistency, the pollsters are likely to be wrong as they are correct.
Again, this is not to demean pollsters but rather how those who analyze politics use them. While polls work well looking at trends, they do not work well as the political snapshot of the population. This is not just my judgment but rather using statistics itself. If polls should be accurate 95% of the time but are achieving less than a 50% success rate, then the pollsters are either unable to calculate the normal population curve or unable to generate a sample to accurately calculate the normal population curve.
While we just discussed the state polls, polls looking at the USA as a whole were as inaccurate as their state counterparts. This in effect, should dissuade us from using polls as if they hold the keys to the election results. This should also dissuade us from using polls within statistical models to project elections.
Instead, we will argue that a new election simulator is needed. We have created a new projection model based on history and statistics. In the 2024 election, there will be two indisputable facts. There will be a national shift and each state will have its own distinct shift.
These two shifts will determine the results of each individual state and its electors. While we do not know how each state will shift, we do know how each state shifted in the past. Using this data, we can create the standard deviation of how much a state can shift, then design a model that gives each state its own normal population curve, and then based on the national shift from their 2020 result determine the odds of each candidate winning each state.
To be safe, we run this model through 1,000 “elections” to determine who is the favorite to win the election. The only variable left is the national shift.
While polls should not be used in a model to create an outcome or be used as gospel when analyzing the future, they could give a hint on how the electorate is feeling. Are they satisfied with the state of the nation? Are they optimistic or pessimistic about the future?
Do they have greater support of one candidate/party or is the support split? Using quantitative and qualitative variables, we can create a range of what the national shift may be in 2024. This would effectively give us the State means and update the presidential odds for 2024. While polls should not be used as the foundation of the political analysis, they could be used here after the model is created.
Instead of being the map where half the time the trails will get the traveler lost, it should be used as a lighthouse. A lighthouse will not necessarily give the sailor an exact distance on how far they are from land, but rather an idea of how close the land may be.
There will be many analysts projecting the results in 2024. There will be some analysts who have created their own electoral models. But be warned, any analyst or model incorporating polls will have a high risk of being led astray. Having a model simply rely on history and statistics instead will lead to a statistical model free from this error with a greater chance to generate better odds than most, if not all, mainstream models in use today. Particularly, when we get results on election night.
I have also working on a model that calculates the odds when the votes come in on election night. If interested in learning more or keeping up with the odds on election night, please follow me on Twitter at @m_guglielmello.
