Home / General / Approval ratings and polls

Approval ratings and polls


This is a guest post from LGM commenter Leeward Mountain


Approval and favorability of incumbents and challengers (henceforth mostly referred to as “approval” for shorthand), respectively, is an interesting  complement to traditional matchup polling in projecting the performance of candidates. It  should be taken into account in observing the upcoming election. I will contend that it is  likely to be fairly predictive, even more than regular polling could be. Here is a review of the past 8 elections. Polling figures were sought as near to Election Day as feasible.

A rough rubric from my understanding of how analysts tend to look at polling results: Above +/-3 is bad, 3 is mediocre, 2 is OK, 1.5 is good, 1 is great, 0.5 is perfect.

From the data, I calculated average margins of error for each metric, which reflects their difference against the results of the presidential elections across the entire time period from 1992-2020. I separated approval and favorability polling because the former is inherently applied to incumbents, so it may have different behavior. But in general, favorability polling of incumbent presidents is not very far from their contemporaneous approval rating. I do not intend to analyze that spread here, as it is an overcomplication that didn’t appear to contribute to predictiveness. But for reference, Obama’s 2012 RCP favorability at E-Day was 50.3%, slightly closer than his 50% approval, and Trump’s in 2020 was 41.8%, significantly worse than his approval rating. Where there were variants in the data, the averages use both. The averages sum absolute values. 

Approval polling average margin of error: 2.7 %
3.6 8.1 3 1.2 1.2 1.1 2.2 0.9
Favorability polling average margin of error: 5.9% 
8 9.3 6.6 10.1 4.6 1.3 8.1 4.1 11.3 6.6 1.2 2.2 6.4 8.6 0.2
Matchup polling average margin of error on between-candidates spread: 2.3% 
6.5 2.5 1.5 2.4 0.9 0.2 0.4 3.2 1.2 1.9 2.7 3.9
Matchup polling average margin of error per-candidate: 1.9% 
0. 6 2.8 0.3 0.4 1.9 1.7 0.7 1.8 0.9 2.9 2.7 0.8 1.2 2.3 0.9 1.4 2.5 2.5 4.3 2.8 3.4  0.1 0.5

Some observations: 

1. When both candidates are majority-popular, traditional matchup polling is much more reliable, as barring a landslide both candidates naturally fail to match their approval/favorability in their vote share, and one may lose more than the other. Approval is not zero-sum, but an election vote is (under most systems). Thus the 1996 through 2008 elections are best fit by matchup polling,  though in 2004 approval/favorability was already good, maybe because both candidates were borderline.

2. When both candidates are similar in popularity to one another, many voters who like – or, pointedly, dislike – both will be put in the position of a decisive demographic, and approval polling may be at its most useful. Such conditions prevailed from 2012-2020, and 2012 & 2020 were indeed elections where approval polling compared well to traditional match-up. 2016 was different. First, it should be acknowledged, as some have pointed out before, that the matchup polling was not nearly as bad in 2016 as the immediate and lasting narrative tells us. It was neither great nor terrible by candidate, and the estimation of the popular vote spread was fine. But approval polling was once again trash in 2016, and I believe it’s because both candidates were so unpopular. The reverse of the traditional popularity of both major-party candidates still demands a cross-pressured choice from voters (“lesser of two evils”).

3. The polling overall was definitely a bigger miss in 2020, as many noted, but it remains underappreciated just how perfectly it captured Biden’s support level (this is often visible in state-level polling as well), while underestimating Trump’s perhaps even more than it did in 2016. I continue to think this has a lot to do with the accounting of the “Undecided” vote in 2020 polling; Trump and Republican polling often-to-usually looks very accurate if you simply assume Undecideds were overwhelmingly Republican voters. But that’s a story for another day, or another person. The relevant thing is that approval polling nailed Biden just as well as matchup polling

4. What made 1992 unique is that many unhappy voters, instead of apportioning themselves between the two major-party candidates, chose to vote with their votes and support a third party (Perot). And yet, Clinton didn’t receive votes anywhere near what his favorability might suggest, despite ranking  higher in that regard than Perot. This election may be an important, or at least relevant,  reference point for 2024.
4.a. As a tangent on 1992, the polling at the end of October according to Wiki was nearly perfect, as at 43-36 Clinton-Bush it almost exactly captured the actual two-party share of 43-37.5. But the polls in the final days seem to have sharply overblown Clinton, even as they got even more accurate on Bush. The polling overestimated Clinton again in 1996. Numerically, these misses harm the averages at or above the level of the 2020 underestimation of Trump/Republicans.

5. RCP has a pretty good record with its simple averages. I’m not sure if 538 ever had public poll tracking for 2012, or if it was taken down. Since RCP seems to have the best historical track record, and for the sake of consistency, I will consolidate further calculations preferring its data.

2004+ RCP approval polling average margin of error: 1.1%
2004+ RCP favorability polling average margin of error: 4.2% 
2004+ RCP matchup polling average margin of error on between-candidates spread: 1.7% 
2004+ RCP matchup polling average margin of error per-candidate: 1.5% 

So my theory is that approval polling should be most relevant when both candidates are somewhere in the 40s and not too distant from one another. Anything more makes it mathematically irrelevant to project a non-zero sum measure onto a zero-sum outcome. Anything less, and approval polling fails to capture either lesser-evil calculations or the viability of third party/protest options (or, of course, abstention). That would explain why it is so close to the true results in 2004, 2012, and 2020, but not at all in other cases.

Since the conditions of apparent predictiveness for approval/favorability polling have only appeared in more recent elections, it makes sense to recalculate the preceding margins of error once more with only those elections taken into account, viz. 2004, 2012, 2020. These were the competitive elections where neither candidate was, so to speak, overly loved or hated. In other words, I am doubly streamlining the dataset, first by choosing a single, aggregate source that has consistent good performance, or relatively so, and second by isolating the elections which it should be (and is) best fit for.  

2004/12/20 RCP approval polling average margin of error: 1.1%
2004/12/20 RCP favorability polling average margin of error: 1.2% 
2004/12/20 RCP matchup polling average margin of error on between-candidates spread: 2.3% 
2004/12/20 RCP matchup polling average margin of error per-candidate: 1.5% 

How do we know this exercise isn’t just overfitting to events that have no bearing on the future? To summarize the observations straightforwardly, my theory accounts for the recent history of US presidential elections well, could be tested against elections before 1992 given the appropriate data, and shows that both approval and matchup polling, but certainly the former, were and can be robust predictors of a candidate’s support on Election Day. Moreover, in the 2020 election, approval polling was a particularly-excellent indicator for both candidates, and performed better than for any previous election in the set. And even the much-maligned matchup polling was half-perfect.

There will be no past election that could bear more similarities  to the upcoming election than the previous one. There is not evidence that presidential polling has categorically degraded or transformed since 2020. Approval/favorability polling under Biden has not evinced any anomalies or wild swings, and  has often visibly tracked objective developments in the macroeconomy, such as general or fuel inflation, or complementary polling such as on “direction of the country.” By no means then do we have reason to believe that its own validity has degraded in some invisible way since 2020 such that it’s good news for Democrats.

But of course, it’s not the eve of the election. Note that the 2024-current figures in the table should not be treated as any more than contextual. Any modeling taking into account the themes discussed here should ‘plug in’  updated data over time. I wouldn’t expect peak accuracy of any type of polling until sometime in October, at least. That is why every time I discuss election polling, I emphasize that what we’re seeing is merely what Biden and Democrats have to beat. Biden being unpopular is an indirect measure of his political support, and if he wants more political support, he must become more popular, by some action or by Ariana Grande technical knockout. IF Biden is not more popular in the runup to the election than he is now, he will be wiped out. It’s that simple. Keep your eyes peeled. 

But there is yet another wrinkle, one I touched upon earlier, that could scramble the entire theory. That is, third-party performance. If both candidates are unpopular enough, the theory admits this could energize third-party movements to the extent that many voters who dislike both major-party candidates refuse to choose between them. In the 2024 race, RFK Jr. has appeared as someone who could potentially take up the mantle of Perot in 1992, destroying any predictive power of approval polling and throwing the election into chaos until well into November.

Why should we take Kennedy into account above all? According to the 538 tracker, Kennedy’s support has been quite consistent at about 10% for months. RCP is less favorable, with the RFK baseline slowly slipping from ~13% over time, starting from late 2023, but it still hovers at around 10%.

A general feature in favor of alertness to third-party performance is that the third-party vote share in 2020 – the high-water mark of recent American civic engagement – was 1.9%, which is higher than it was in 2004, 2008, or 2012, despite those elections and their candidates not rising to the level of distinctiveness and stakes that many would attribute to a potential reelection of Donald Trump. Given the profoundly-expressed sum of fatigue with both Biden and Trump, archetyped as the Sick-Old-Mens of the Sick Old Man of the Americas, and both their current approval ratings being significantly lower than they were in 2020 (those ratings are almost exactly parallel to the 2016 race, which had 5.7% third party vote), it withstands scrutiny to imagine that third party voting will at least reach appreciable levels this time.

But a case against Kennedy’s relevance is that he will eventually wither on his vine as the elections draw nearer, due to the reasoning that whether or not the political environment is ripe for double-digit poaching from the Volksparteien Kennedy is not the man for the hour, not like Perot was. Kennedy’s favorability will only decline, and with it his polling stature, in this view.

 I can neither refute nor confirm the poles of this debate; all I can do is highlight that it’s plausible that dismissiveness fails to capture this moment, as well as describe the consequences of that.

I will again appeal to testament as a bottom line. TLDR: Polls still work and 2024 could easily be a cycle suited to my electoral theory. If it is, and Biden is still this unpopular, outright and on net, in October, he will lose for sure. If Kennedy is still threatening double digits in October, my theory is inapplicable as presented and approval polling will be useless as a guide.

I just hope this analysis will prove some useful guidance through the height of the 2024 campaign season. 

  • Facebook
  • Twitter
  • Linkedin
This div height required for enabling the sticky sidebar
Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views : Ad Clicks : Ad Views :