Post-Mortem of the 2024 Forecast
Helmut Norpoth
The Primary Model gave Democratic nominee Vice President Kamala Harris a 75% chance to defeat the Republican nominee former President Donald Trump in the 2024 presidential election. She was predicted to garner 52.1 percent of the major-party popular vote and 315 electoral votes. Instead Trump was elected President with 312 electoral votes to 226 for Harris; by latest count, Trump is also leading Harris by more than two million in the popular-vote count. What went wrong with the Primary Model?
For starters, Harris did not compete in primaries in 2024. She was awarded the Democratic nomination without a contest after President Biden withdrew from the race on July 21. Some might argue that her lack of primary performance should have forced the Primary Model to suspend operations in 2024, but this was too drastic an option to choose.
Two other solutions instead were considered for making a forecast with Harris in the race. One was the way statistical analysis commonly handles missing data. That is, by filling the void with a neutral value, which would not bias the outcome one way or another. The other was to follow the precedent established by previous elections where a sitting president decided to withdraw from the race during the election year.
This happened in 1952 and 1968, when first Harry Truman and then Lyndon Johnson dropped out of the presidential race after primary contests were held. In both cases the model algorithm assigned the primary performance of the departed presidents to the ultimate Democratic nominees. With that adjustment, the model got it right, albeit retroactively, both times (1952 and 1968). With that assurance the power of precedent prevailed and Harris was assigned Biden’s primary score.
Some would call it a stretch, nonetheless. Was it realistic to expect Harris to perform so well in 2024 primaries if there had been an open race? Didn’t she flame out as a presidential candidate in 2020 before a single vote was cast in primaries or caucuses that year? Indeed. What worked in her favor in 2024 was that the sitting president made a full-throated endorsement of her candidacy in the same breath that announced his departure from the race. What is more, within days Harris was able to nail down support for her candidacy from enough delegates to clinch the nomination at the upcoming national convention. Not a single challenger threw their hat in the ring. There is no denying that at that moment Harris enjoyed as much support as Biden mustered in his primary campaign. Perhaps even more so since she did not carry the age-related liabilities of the sitting president. The decision to use Biden’s primary score as a proxy for Harris was not unreasonable.
As it was, that score was based on the outcomes of early primaries (New Hampshire and South Carolina). President Joe Biden won both contests in commanding fashion, brushing off a challenge from a little-known congressman from Minnesota, Dean Phillips. Meanwhile former President Donald Trump faced a more organized challenge from a better-known rival in Republican primaries, Nikki Haley. By comparison, Biden racked up the better primary score. Transferring this advantage to Harris after Biden withdrew from the race made her the favorite in November. What also benefited Harris in the general election was the working of an electoral cycle that favors the White House party after one term. The combination of those two factors predicted a Harris victory over Trump in 2024 with 75-percent certainty, the same degree as had been forecast in March for Biden.
Just in case you are interested: What would the alternative considered but not chosen have forecast? That option entailed making the missing-data correction, which would have assigned Harris a neutral primary score. As it turns out, that would have given Trump a 58-percent chance of winning the 2024 election with 50.9 percent of the popular vote and 282 electoral votes to 256 for Harris. Woulda, coulda, shoulda, an opportunity missed. Obedience to precedent be damned!
Before fully exploring why the Harris forecast was a miss, a brief review of the model may be in order primarymodel.com, along with a probe of previous misses.
The core premise of the Primary Model is that winning primaries predicts electoral victory in November. Presidential elections going back as far as 1912 are used to estimate the weight of primary performance. It was in 1912 that presidential primaries were introduced. That year the candidate who won his party’s primary vote, Woodrow Wilson, went on to defeat the candidate who lost his party’s primary vote, William Howard Taft. As a rule, the candidate with the stronger primary performance wins against the candidate with the weaker primary performance.
For elections prior to 1952 all primaries were included. Beginning in 1952, only the New Hampshire Primary has been used, as a rule. South Carolina has been added for elections since 2008. Both Obama then and Hillary Clinton in 2016 enjoyed strong support in a large and most loyal Democratic constituency, African-Americans, who are few in numbers in New Hampshire.
In 2024, Joe Biden won 63.9 percent of the primary vote in New Hampshire, as a write-in candidate no less, followed by 96.2 percent in South Carolina. Meanwhile, on the Republican side, Donald Trump scored 54.3 percent in New Hampshire and 59.8 percent in South Carolina. Bottom line: the Democrat had the stronger primary performance. With Harris as the Democratic nominee after Biden’s withdrawal, she was now favored to be elected president.
In addition to primaries, the Primary Model also takes account of an electoral cycle that has been shown by the author of the model to be operating for nearly 200 years. Though not quite with the precision of a metronome, the presidential vote swings quite regularly from one party to the other.
A simple way to illustrate the operation of this cycle is by way of the snapshot of elections since 1960. After one term in the White House the incumbent party has won the next contest almost all the time; when it has held the White House for two or more terms, it has lost almost all of them.
Prior to the 2024 election, the PRIMARY MODEL, with slight modifications, has picked the winner of all but two of the presidential elections since it was introduced in 1996. For elections from 1912 to 1992, the PRIMARY MODEL gets it right, albeit retroactively, every time except in 1960.
Without much doubt the two recent misses (2000 and 2020) occurred under extraordinary circumstances. The 2000 case, with Gore predicted to win, featured a protracted recount of disputed ballots in the pivotal state of Florida. It was not settled until the U.S. Supreme Court, in a controversial 5-4 ruling (Bush v. Gore), handed the state to Bush. What is more, an oddly designed ballot in Palm Beach County, the infamous Butterfly Ballot, produced thousands of invalid votes. This cost Al Gore enough votes to deprive him of victory in Florida and thus in the nation. https://websites.umich.edu/~wmebane/mebane.pop2004.pdf
The other case the model got wrong was the 2020 election. This time a once-in-a-century combination of election-year surprises wreaked havoc on a forecast posted as early as March 2, which gave Donald Trump a 91% chance of winning re-election. The first surprise was the outbreak of the coronavirus pandemic. Next came an economic downturn of a scale not seen since the Great Depression; that was triggered by the lockdown aimed at keeping people safe. Then, with the same goal in mind, came an unprecedented expansion of voting by mail. Add to that the killing of George Floyd, which sparked a wave of racial unrest not seen since the 1960’s. This cascade of unexpected election-year events constituted what is known as an “October Surprise,” the bane for forecasters and strategists alike.
In the end, those surprises spelled bad news for Trump, the president in office, whom the Primary Model happened to forecast as the winner. The Fox News exit poll showed voters to be gloomy about the Covid Pandemic. (https://www.foxnews.com/elections/2020/general-results/voter-analysis).
A majority felt it was “not at all” under control, and their votes favored Biden over Trump by a better than 5-1 ratio.
The assessment of the economy fared no better. A majority rated its condition as “poor” or “not so good.” Their votes favored Biden over Trump by a lopsided ratio of 3-1.
Add to it widespread concern over racism in policing in the wake of the George Floyd killing. A large group of voters close to a majority considered racism in policing a very serious problem and wound up voting 5-1 for Biden over Trump.
Finally, the easing of rules for voting by mail in the wake of the coronavirus pandemic led to a vast expansion of the group of voters who chose that option in 2020: 4 of 10 availed themselves of this option to cast their vote this time. It turned out, as widely expected and feared by the Trump campaign, that Biden garnered the lion share of votes cast by mail (66% to 32% for Trump).
Mail ballots along with drop boxes, according to my post-election research, helped Biden flip three battleground states (Michigan, Pennsylvania and Wisconsin) though this was not enough to deprive Trump of victory. https://doi.org/10.1016/j.electstud.2023.102693
Perhaps the only way the Primary Model could have tried to handle the election-year surprises in real time was to update the forecast after the last primary was held in 2020. That forecast would have given Donald Trump a 69% chance of winning the election, with 304 electoral votes to 234 for Biden. A good bit closer than the March forecast, but still a miss.
Back to the future with the 2024 election.
In 2024, to be sure, there was no Covid Pandemic with all its consequences or a protracted recount of disputed ballots. There was, of course, the rare event of a sitting president withdrawing from the race in mid-year, following his disastrous performance in the debate with Trump. The swift elevation of Vice President Harris to Democratic nominee sparked a surge of popular support in polls for the new Democratic ticket. This was, of course, what notable Democrats had been hoping for as they pressured Biden to step aside.
The Primary Model, however, did not treat the Biden-Harris switch as a game changer: it had forecast a Biden victory before and now forecast a Harris victory since Biden’s primary record was transferred to her. If that forecast was to fail, as it did in the end, perhaps that was because Biden’s primary record proved to be a poor predictor.
A major selling point for using primaries as a predictor of general elections is that for sitting presidents support in primaries closely tracks with their job approval. Gallup polls make it possible to chart the approval ratings of presidents as far back as FDR (https://news.gallup.com/interactives/185273/presidential-job-approval-center.aspx).
Those with high approval marks at the start of an election year, say above 50 percent, tend to breeze through primaries almost unscathed. Those with low approval tend to invite challenges that wind up depressing their primary vote. In 1964, for example, President Lyndon Johnson won close to 100 percent of the primary vote while commanding an approval rating above 70 percent. He ended up winning the general election in a landslide. By contrast, in 1952, President Truman lost the primary battle while mustering an approval rating in the 20s. Any wonder he soon withdrew from the presidential race that year? Rare is the case where a president with above-par approval performs below-par in primaries (1980). The explanation is quite simple, however. Jimmy Carter’s approval rating was briefly inflated by the Iranian hostage crisis. Once that rally subsided his approval returned to a level that closely aligned with his primary score, foreshadowing defeat in November.
To come to the main point of this exercise, in 2024, Joe Biden won 80 percent of the primary vote with an approval rating of barely 40 percent (and disapproval close to 60). A glaring mismatch, with one side screaming victory in November and the other side defeat. Glaring but not unprecedented. Misery loves company, as they say. The cases of 1996, 2002, and 2020 cluster nearby. Hence Clinton, Obama and Trump in their quest for re-election were all under water in approval early on while racking up big primary scores. Both Clinton and Obama, however, managed to boost their approval during the election year until they crossed the 50-percent mark and headed for victory. In so doing, they vindicate the forecast of the Primary Model based on the early primaries. What about Biden as well as his replacement in 2024?
Unlike his two Democratic predecessors, President Joe Biden proved unable during the election year to boost his approval rating enough to cross the winning mark (50-percent). In fact, in the last poll before his withdrawal the Gallup Poll recorded his lowest score during his presidency (36%). His replacement as the Democratic standard bearer, Kamala Harris, fared little better. Her final rating barely reached 45%.
Poll averages compiled by 538 https://projects.fivethirtyeight.com/polls/approval/kamala-harris/ confirm this pattern. Thus, whether it was Biden or Harris, the Democratic primary record vastly overstates the expected level of support based on their approval ratings. So much so that the forecast of the Primary Model gets in wrong in 2024.
Before contemplating possible lessons from this for future use of the model, let’s take note of another outlier in the primary-approval chart.
It is located quite close to the 2024 case along with 1992 and 2012: 2020. The sitting president that year, Donald Trump, also racked up impressive primary victories while mustering below-par approval at the beginning of the election year. The latter clearly being a bad omen for November. Just as it turned out for Biden in 2024. The question, then: Was Trump able, like Clinton and Obama before, to move his approval into positive territory, thereby vindicating the model, or was he unable to so and would be end up like Biden/Harris in 2024?
The answer, according to Gallup polling, is unmistakable. Trump fell short of breaking into positive approval territory. Surprisingly perhaps the onset of the Covid Pandemic boosted his approval in April-May, hinting at a rally effect quite common in responses of the public to international crises. The Covid rally, however, proved too weak and fleeting to carry Trump to victory in the election. By June his approval dropped below the 40-percent mark. Even though it recovered somewhat in the remainder of the year it never reached 50 percent, At the end, more Americans disapproved than approved of the job Trump was doing as president. This was at odds with his impressive primary performance, which led the model to forecast an easy victory for Trump. Hard to admit but it was not just a cascade of “October Surprises” that made the forecast miss the target but a flaw of the model. Namely when primary performance is out of line with presidential approval. The solution? An adjustment of the key predictor of the model for such approval. The search is on.