November 24, 2020
OK the polls were generally off again. They underestimated the Trump vote even after all the changes recommended by independent experts since 2016. In the last post, I showed that Trump voters were not candid about who they were voting for on election day, a form of social desirability effect. “The Return of the Shy Voter?”
As I was analyzing polling data before the Presidential election, I was having some doubts about the polling data. A little voice in my head kept saying “are you sure these numbers are right?”
At the time, I was trying to predict the election outcome in key swing states like Florida. As you may recall, the average of polls for Florida gave Biden a statewide lead of 1.2%. Making a prediction for Florida on that statistic would be like betting on a horse with one fast time but who never won a race. So I called the state for Trump. That’s how scientific it was.
I have always been convinced there had to be a better way to predict an election other than based on polls only. But my previous attempts to find that method have always led to failure.
But that has never deterred me from continuing my gallant quest to discover the magic variable that would solve the election prediction problem. Until now…
I have always been interested in the “favorability rating,” which supposedly measures how voters feel about the candidate. The question is simple: “Do you have a favorable or unfavorable opinion of Donald Trump?”
As you can tell, this question doesn’t measure anything specific like the common Job Approval rating, which measures the incumbent’s performance in office. The favorable rating is considered a popularity measure, but I’m now convinced that it is more than that.
There is a general lack of studies on what the rating means due to a lack of analysis on the subject. It was generally believed that they correlated with job approval ratings, but when Bill Clinton’s affair became public his favorability declined significantly but his job approval rose.
The academic literature has only a few studies about the impact favorability rating except to suggest that it is more than just another way to measure an incumbent’s job approval.
But I found an article on the fivethirtyeight web site published in 2012 entitled “Do Romney’s Favorability Ratings Matter?” https://fivethirtyeight.com/features/do-romneys-favorability-ratings-matter/
This brief study shows that in the late stages of a Presidential campaign “the candidate with the stronger favorable rating in the late stages has won every election since 1980, and there has been an almost-perfect correspondence between the margin of the favorability gap and his margin of victory or defeat in the popular vote.” As important as this statement is, they posted no data to confirm this assumption. But this was my first clue that the favorability rating was more than just a popularity variable.
Some of you know my fondness for job approval ratings and why they matter for the incumbent. But the biggest issue with this rating is it doesn’t apply to non-incumbents who are challenging the President. Consequently, we cannot measure the difference between the two, which could tell us how the voter would likely choose on election day.
But that is not the case with the favorability ratings. It applies to both the incumbent and the challenger, allowing us to make a significant comparison on how voters feel about candidates. At least that is my hypothesis.
But there is one important problem. Most pollsters don’t ask this question at the state level. At least until now. There is a new resource for state by state surveys from a relatively new company called Civiqs.
Their methodology is an opt-in survey panel covering every state in the country (https://civiqs.com/). One of their continuing questions is the favorability rating. Each day the data is uploaded to their public site.
Using this data, I started compiling the favorability rating for Trump and Biden for each state as shown in Table 1.The date for these ratings was October 25, 2020, some nine days prior to the election .
|STATE FAVORABILITY RATING||TRUMP||BIDEN|
|South Dakota ?||62||35|
|AVG. FAVORAABILITY RATING||49.782||47.054|
If you look at Florida, you will see that the favorability ratings between Trump and Biden were quite close. There is just 3.3% difference between the two, in Trump’s favor. As we all know now, the Florida contest was close. In fact, when the dust cleared Trump defeated Biden by only 3% (51%-48%) in Florida.
At first, I didn’t see the similarity between the actual vote and the candidate’s favorability rating. When I took a closer look, I noticed that there was a 3.3% difference between Trump and Biden’s favorability ratings taken some 9 days before the election. Just three-tenths of a percent difference than the final results in Florida.
As interesting as that result seemed, I was not ready to say Eureka just yet. But it gave me the incentive I needed to compile by hand all the favorability ratings for all states on that same date, as shown below in Table 2 below.
A negative value indicates that Biden had a negative rating because I subtracted Trump’s percent from Biden’s in every case. For example, Alabama shows the favorability rating difference was -26.1%. That means Trump had 26.1% more favorable ratings than Biden in Alabama. In California, Biden had a positive 31.6%, which indicates he had 31% more favorable ratings than Trump in California.
As a comparison, I have added in the third column the final vote percent difference between Biden and Trump.
As an example, the final election vote for Biden in Maryland was 63.3%. For Trump it was 34.9%. If you subtract 34.9% from 63.3% your get 28.4% (Trump vote % – Biden vote %). The final vote difference on election day was exactly 28.4%!
|STATE||BIDEN/TRUMP.FAV.DIFF||FINAL VOTE PERCENT DIFFERENCE|
Out of the 50 states, the largest error rate between Trump and Biden’s final vote difference and their favorability difference was in Hawaii, with 4.6% difference. For all 50 states, the average error difference was only .04%.
In political science, we use regression models to determine the relationship between variables which also produces a scatterplot of the results. In the Chart below, this scatterplot shows how tight the “fit” is for the models estimates.
Each small circle represents a state and the red line the linear relationship between the Biden / Trump favorability difference and the final vote percent difference. From a statistical viewpoint, the relationship between the two measures is almost perfect (.994 R square).
In other words, we can now predict the final vote percent difference between two candidates by using only the favorability ratings prior to the election. More importantly, the error rate is only .04%.
Does this method apply to the national vote? The national favorability rating for Biden on October 25th was 45% and for Trump 42%, a 3% difference. The final national vote for Biden was 51.1% and for Trump 47.2%, a 3.9% difference. The final vote difference was 3.0%, an error of nine-tenths of one percent.
So why has no one discovered this by now? There a couple of possible reasons. First, there is very little research on the favorability ratings and what they mean to voters. Secondly, until recently, no polling firms have consistently used these ratings at the state level. Finally, with the exception of Nate Silver’s fivethiryeight review, no one else has seen any relationship between the ratings and the final election results.
The bad news is that Civiqs data does not go back to 2016, so we can’t compare the ratings to the actual vote in that Presidential election.
Who knows, in the future we might be able to cancel the election altogether and just rely on this system instead. I’ll have to check with my pocket Constitution to see if that is legal. Be safe…