Can we trust the polls this year?

The Democrats’ election to lose?

The upcoming midterm elections in the United States will be critical to the path of the Trump agenda and likely beyond. Republican majorities are narrow in the House and the Senate, and both historical precedent, and polling bode ill for their continued control.¹[1] Midterm elections generally favor the party not in the White House and national polls are daunting: Democrats average an 8-percentage-point margin in generic-ballot polls for the House, while eight currently Republican Senate seats rated as “toss-ups” (versus two Democrat-held seats).²[2] The increasing polling advantage for Democrats is shifting the consensus view from one of a likely swing of the House to a potential landslide victory in both chambers.³[3] Yet, this is no ordinary midterm election, and persistent polling errors in recent elections cast doubt on the reliability of polls as forecasts. In a companion piece that I will publish at Thematic Markets later this week, I present my comprehensive analysis of polling errors, and their likely electoral and market impact. Here, I’ll shed light on what causes polling bias and explain why it may again be an issue for pollsters and election watchers.

Polling errors for president

Unless, like Rip Van Winkle, you’ve been asleep for 20 years, you’re probably aware that national polling in the US has been objectively bad in the last three presidential elections, with President Trump pulling off a surprise Electoral College win in 2016, coming far closer to victory than polls predicted in 2020, and unexpectedly winning the popular vote with coattails for Republicans in 2024. But, as shown in the table below, polls predicted the intervening midterm elections in 2018 and 2022 nearly perfectly. As a result, some have speculated that polling bias only infects elections in which Donald Trump is on the ballot.⁴[4] But that hypothesis is challenged by the persistent undercounting of populist voters across Western democracies, which suggests bias instead reflects populism itself being on the ballot.⁵[5]

Election year/type Average polling error in favor of Democrats

2016 presidential +3.1 pp

2018 midterm +0.1 pp

2020 presidential +4.6 pp

2022 midterm −0.6 pp

2024 presidential +2.7 pp

Errors versus bias

It is important to distinguish between errors and bias in polls. A polling error is the absolute margin by which a poll (or an average of polls) misses the actual election result. In the above table, polling errors in presidential elections, ranging from 2.7 to 4.6 percentage points, are an order of magnitude larger than those in the midterm elections (0.1 and 0.6 percentage points). But critically, the errors in presidential elections were not random: they were all positive, i.e., overestimating Democrats’ votes. That is polling bias: errors consistently favoring the same party. The magnitude of polling errors in the 2016, 2020 and 2024 elections was not abnormally large compared to earlier elections — indeed, it is fairer to say that the 2018 and 2022 midterms had unusually small errors — but unlike prior US elections, they consistently overestimated Democrats’s support.⁶[6] The question is whether this year’s midterm elections will look like those of 2018 and 2022, or will be infected by the bias that has wreaked havoc with election forecasting in recent presidential elections.

The source of bias

To answer that question, you need to understand the source of polling bias in the last three US presidential elections and numerous non-US elections. The cause, as identified by Democratic political data scientist David Shor, is remarkably simple yet nearly impossible to correct: trust. Conventional wisdom holds that the populist divide is between the educated on one side and the less educated on the other. Yet, Mr. Shor shows that both college-educated and non-college-educated white male voters strongly preferred Hillary Clinton to Donald Trump in the 2016 election if — critically — if they had high levels of social trust (trust in others). Conversely, among low-trust individuals, even advanced-degree holders preferred Mr. Trump to Ms. Clinton.⁷[7] The conventional wisdom of an educational divide stems from less-educated citizens generally having lower trust than their more-educated peers. (It probably helps that this provides a convenient narrative for high-trust, educated voters to dismiss their populist opponents as “uninformed.”)

Correcting for bias

Pollsters always contend with bias in their initial samples but have developed robust techniques to correct for many sources of bias. To illustrate, imagine the difference in response rates between 19-year-old men and 75-year-old widows. Pollsters usually get too many older widows waiting by the phone (or online) and not enough young men who “can’t be bothered.” If young men have different political preferences from older women, this induces bias in the poll. But pollsters correct for input bias by “reweighting” survey participants. Suppose a pollster gets 10 responses from men aged 18 to 29 and 300 responses from women over 65 in a survey of 1,000 registered voters, but the former account for 5% of likely voters and the latter for 15%. The pollster “upweights” the young men’s responses by multiplying them by five (to equal 50 responses, or 5%) and “downweights” the older women by halving their weight (to equal 150 responses, or 15%). In practice, the process is much more complicated as sex, race, religion, income, urban or rural residence, and education are all associated with differences in political preferences. This is why pollsters ask respondents about a range of demographic characteristics to facilitate a reweighting that reduces sample bias.

Hard to correct

But social trust is a surprisingly difficult source of bias to remove from polls, hence its persistence across recent presidential elections. Low-trust individuals generally do not answer phone calls from strangers or participate in online polls. Even if they did, they’re unlikely to answer accurately the question: “Do you trust others?” As a result, pollsters get too few low-trust respondents and have no accurate way to identify and appropriately upweight them. This has long been a problem for pollsters attempting to measuring societal preferences, but didn’t affect election polling since low-trust voters also tended not vote. But that began to change in the mid-2010s. Populist voters, as Mr. Shor found, tend to be low trust and not respond to polls. But as demonstrated by the persistent polling bias across countries — US presidential elections in 2016-24, Brexit in the UK, Italian parliamentary elections in 2013, 2018 and 2022, the Dutch 2023 Parliamentary election, and recent German state elections — populists do vote when populism is on the ballot.

The case of 2014

Despite the manifestation of both in the 2013 Italian and 2014 European Parliament elections, few noticed of the rising trends of populism and associated polling errors until 2016. Even then, it was seen as a “Trump effect” in the US that wasn’t present in midterm elections. But the 2014 midterm elections call those presumptions into question. Republican House and Senate candidates, on average, outperformed polls by 6.8 and 5.3 percentage points, respectively, that year.⁸[8] Those large errors could have been random. But a closer look at the Republican primary contests, resulting general-election platform, and some of the surprise winners suggests 2014 may have reflected an early wave of populism. The Republican primaries that year were filled with populist challengers to establishment candidates. While most challengers were defeated, establishment figures were forced to adopt their populist agenda — a more confrontational approach to expanding federal spending (especially Obamacare) and to immigration — in the general election.⁹[9] Further, among a wave of newly elected Republican senators were those later strongly associated with President Trump’s agenda: Tom Cotton of Arkansas, Joni Ernst of Iowa, and David Perdue of Georgia, whose support was underestimated in pre-election polling by 9.4, 6.6, and 6.1 percentage points, respectively.¹⁰[10]

No populist choices?

In contrast, it is hard to argue that populists could get excited about the Republican candidates on offer in either 2018 or 2022. In 2018, while President Trump — as now — actively campaigned as though he were ”on the ticket”¹¹[11] and sought to influence which Republican candidates were chosen in open races, he was far less successful in shaping the Republican candidate field. Indeed, a majority of nonincumbent Republican candidates that year made no mention of President Trump and identified themselves with the establishment wing of the party in their campaigns.¹²[12] In the 2022 midterm elections, then-former President Trump was more successful in recruiting aligned candidates to Republican races, but with the stigma of the 6 January 2021 Capitol Hill riots still hanging over him, many sought to distance themselves in the general election.¹³[13]

Is 2014 or 2018 this year’s model?

Thus, the absence of polling bias in the midterm elections of 2018 and 2022 may reflect a lack of populist choices in contrast to 2014. If so, whether polling bias affects this year’s midterm elections may hinge on whether populist voters believe populism is on the ballot. President Trump has actively sought to convince them that it is. As he did in 2018, he has campaigned across the country, exhorting his faithful to see the election as a referendum on his agenda and to vote as though he were on the ballot. But, in contrast to 2018, he has had far more success in shaping Republican nominations, including unseating popular establishment figures in favor of aligned populist candidates.¹⁴[14] Using his substantial campaign war chest as a carrot, he has also encouraged Republican candidates to campaign on his record as much as their own,¹⁵[15] and has devoted significant resources to a voter-turnout effort.¹⁶[16]

Will it make a difference?

The ultimate question is whether polling bias will make a difference this year. After successfully calling the 2016, 2020 and 2024 elections — I didn’t have a contrarian view on the 2018 or 2022 midterm elections — that is the question my institutional clients at Thematic Markets most frequently ask. Unfortunately, if you are not a paying Thematic Markets subscriber, you’ll have to wait until after the election to find out the answer. If, however, you wish to see my full analysis of the 2026 US midterm elections, including assessment of polling bias, forecasts for House and Senate control, and expected effects on the economy and markets before the election, subscribe to Thematic Markets today and you can read it later this week.