The Democratic lead on the House generic ballot is now as large as it was in 2018
The Democratic Party may be in disarray, but Democratic voters are not
In the last few weeks, Democrats have significantly improved their standing in generic ballot polls. On May 1, 2026, according to our average of polls here at FiftyPlusOne, the Democratic party had a lead of around 4.5 percentage points in the generic ballot; by May 26, that lead had expanded to almost 6 percentage points.1
For the first time this cycle, Democrats now lead the generic ballot by slightly more than they did at the same point in the 2018 election cycle, and by far more than the Republicans led at the same time in 2022. And generally, in most cycles in recent history, the party out of power grows their lead during the final months of the campaign (note that the Dobbs decision to overturn Roe v. Wade 137 days prior to the election disrupted this pattern somewhat in 2022, though Republicans gained about 3 points in the final month of that campaign).
As we often do here at 50+1, we went diving into the survey crosstabs to figure out why the Democrats have been gaining ground in the last few months. It turns out that the recent shift in the race is being driven, at least in part, by how partisans plan to vote in November (rather than how political independents say they’ll vote — which is also boosting Democrats today).
Democrats are benefiting from rising loyalty from their voters — and falling loyalty among the GOP’s
In particular, if we consider the percent of Democrats and Republicans that say they plan to vote for a candidate for Congress representing their own party, we see a fairly similar looking trend as in the topline average.
Today, nearly 96 percent of Democrats say they plan to vote for the Democratic candidate in November, and that number has been growing in recent months; in July 2025, it was 94 percent. Among Republicans, however, the percent who plan to vote for a candidate of their own party has spent the cycle hovering around 91 percent, and despite a brief surge in April has been falling since around the beginning of the month, from a cycle high of 91.9 percent on May 5 to 90.6 percent on May 26, near the cycle low of 90.4 percent.
Of course, a drop of 1.2 percentage points isn’t really all that much, in the grand scheme of things. It’s the kind of shift that may be coincidental, and could quite possibly revert as time goes on. But small changes also add up; Democrats in recent elections are about 45 percent of all voters, so increasing party loyalty by, e.g., 2 percent contributes 2 * 0.45 = 0.9 points to their share of the vote — or nearly 2 points on margin. Movement by 1.5 points in the other direction among Republicans corresponds to another 1-1.5 points for Democrats (2-3 points on margin).
Putting this data into historical context also shows some marked differences between 2026 and the last few midterm cycles. This is where our story gets pretty interesting.
When we compare 2026 polling to the last two midterm cycles, we can clearly see that 2026 Democrats are the most loyal party, by far, in recent midterm history. As of today, a larger percentage of Democrats say they plan to vote for a candidate of their own party than either Democrats or Republicans have ever said in any of the last three midterm cycles. And from this point forward, the parties tend to grow their loyalty, as perhaps reluctant partisans “come home” as the vote gets nearer — so conditions may get even more favorable for the opposition party. On the other hand, with numbers this high, it’s possible that the Democratic party is close to being maxed out with its own voters.
Republicans, on the other hand, look more like a typical midterm cycle, if a bit on the high side. Party loyalty across all three cycles generally hovers in the mid 80s to low 90s, with partisans really consolidating around the candidate of their own party in the final 100 or so days of the cycle, which so far seems to be about where 2026 Republicans stand. And Republicans this cycle are much more loyal to their party than Republicans in 2018 were, during the first Trump presidency. One warning sign for the GOP may be the general flat or negative trend for the GOP today, where loyalty in previous cycles had tended to begin trending up by now.
Republicans are also more likely to defect in 2026 than in-party voters in 2018, 2022
In addition to this slightly higher-than-usual loyalty rate among Republicans, we see another trend in this cycle’s data: a slightly higher-than-usual defection rate among Republicans, too. (This is possible because there is a reduction in the percentage of respondents saying they “don’t know” how they’d vote, or that they wouldn’t vote at all.)
Currently, nearly 6 percent of Republicans say they plan to vote for a Democratic candidate for the House in November — about 2 points higher than the average number of partisans who planned to defect at this point in the previous two cycles — while Democratic defections are historically low. As of May 26, the percentage of Democratic defectors is lower than at any point in either of the last two cycles among any party.
In the previous two midterms, we see the defection rate generally highest among voters whose party holds the presidency; as of now, 2026 exceeds both 2018 and 2022 in defection among presidential co-partisans (though not by very much), but it hasn’t been consistently higher than those previous midterms this cycle. As with party loyalty, defections tend to grow in the last 100 days of the cycle, especially among the party that controls the presidency. In 2018 and 2022, polls among the party in power showed an average of 6 percent defection on the day of the election; the GOP today is nearly already at that point.
In addition, unlike in previous cycles, the higher number of Republicans planning to defect from their party than Democrats has been consistent throughout the 2026 polling. In both 2018 and 2022, by this time in the cycle, the number of partisan defections on each side had more or less begun to even out. But this year, the difference has stayed more or less the same, around 3 percent, throughout the campaign. Republican voters are not “coming home” yet.
Changes in party identification could be a part of increasing party loyalty
Of course, there are some important caveats to all of this data. The first will be familiar: anytime we are using poll crosstabs to do analysis of subsegments of the electorate, you should consider our estimates as more uncertain than when we’re dealing with toplines. Crosstabs are noisy, polling is imperfect, and you should look askance at small differences in single polls (or even several!). However, broad trends generally are worth noting, such as the especially high Democratic loyalty in 2026, and the especially low Republican loyalty in 2018.
But there’s even more to be thinking about here, which is how exactly we define “Democrat” and “Republican.” Different pollsters will do this in different ways: by asking voters about their party, by using partisan voter registration (where available), by some combination of these data, or by modeling party ID in cases where it isn’t known. Online panels might store a party ID for their respondents, which they update from time to time by asking the respondent to fill out their demographic data again.
It’s also worth acknowledging that these are not necessarily the same people across elections. Because pollsters measure party identification by asking respondents which party they feel closer to, parties can lose (or gain) identifiers over time — like when their party leader is very unpopular. There’s possibly a significant contingent of voters that would have called themselves (or been registered as) Republicans in 2018 that would not claim that label in 2026, and vice versa. What looks to be an unusually high loyalty rate among 2026 Republicans, compared to previous years’ loyalty among members of the president’s party, could just be a difference in the types of people that call themselves Republicans.
The numbers among Democrats, both in defections and loyalty, also raise plenty of questions about whether the types of voters that identify as Democrats have shifted over the course of the last 8 years. As voters increasingly express dissatisfaction with the party, it’s possible that only the most dedicated partisans are still identifying themselves (or registering) as Democrats, among whom you would expect to see extremely high loyalty and low defections. Indeed, according to Gallup, identification as either a Republican or a Democrat has been on the decline, while identification as an independent recently hit a new high of 45 percent.
Fundamentally, what it all boils down to is this: If people who called themselves Republicans in early 2025 no longer approve of Donald Trump, they may also no longer want to call themselves Republicans. This would leave behind a smaller group of Republican identifiers that were fiercely loyal to the president — while increasing the pool of anti-Trump Democratic and independent voters. Likewise, if people who called themselves Democrats in late 2024 are angry with what they see as mismanagement of the presidential election and its aftermath by party elites, they may no longer want to call themselves Democrats, and those that remain would have much higher party loyalty than those that left.
Despite the many proclamations of doom for the Democratic Party (mostly stemming from the party’s low favorability rating), it appears that folks that currently identify as Democrats are much more united than partisans have been in the recent past. And Republicans, paradoxically, appear both a bit more loyal and a bit less loyal than in previous midterm cycles, perhaps a hint of the underlying fault lines in the party that have yet to truly become an earthquake.
Paid subscribers to 50+1 get access to premium analysis, plus sortable tables and complete data access on our polling website. If you want to follow the 2026 cycle with the best data at your fingertips, become a paid subscriber.
Methodology
To estimate demographic subgroup trends for the generic congressional ballot, we use an exponentially weighted moving average (EWMA) that adjusts polls for their sample size, recency, subgroup noise, population and pollster effects, and movement in the national generic ballot since each poll was conducted.
We calculate this average in six steps. For any given day, we look at the polls conducted for a subgroup on or before that day, and do the following:
First, each poll’s Democratic and Republican vote shares are de-trended by subtracting the corresponding national topline estimate on the poll’s end date, isolating each subgroup’s deviation (aka residual) from the national environment on that day. If the national average for the Democrats is 45% and the subgroup is 50%, for example, we record the subgroup residual as +5.
Second, we calculate a EWMA with a decay rate of 0.96 (roughly a 17-day half-life) on the de-trended values from all polls for this subgroup. Then we add back the national topline to produce trend-adjusted subgroup estimates. This ensures our average accounts both for movement in the national trend and a subgroup’s overall lean toward one party.
Third, we compute house effects for each pollster in each subgroup based on how much their polls differ from the average. Each firm’s house effect is equal to the average difference between their results and the trendline-adjusted national average on the day the pollster released their results. We then shrink the house effect toward zero to account for noisy residuals, via the formula
adjusted = raw_effect * n / (n + 5), where n = the number of polls from a pollster among that subgroup. So a pollster with 1 poll keeps 1/6 (~17%) of their estimated bias. With 5 polls it’s 50%, with 10 polls it’s 67%, etc. This ensures that when we have relatively few results from a given pollster, we aren’t overconfident in their house effects.Fourth, we subtract each pollster’s house effect from their results, and then calculate a new aggregate estimate that adjusts for house effects and movement in the national generic ballot average. We do this by repeating steps 1 and 2 above, but using the house-effect-adjusted poll results instead of the raw results. Fifth, we estimate “population” effects for polls of likely voters, registered voters, and all adults to control for systematic deviations in results due to the types of voters pollsters sample. In horse race contexts (like the generic ballot), we want out average to capture vote intention among likely voters so we can make characterizations about the electorate and likely political outcomes. This adjustment works similarly to the house effect adjustment explained in steps 3 and 4 above, except we group polls by the population each surveyed instead of the survey firm responsible for each survey. Once we have estimated residuals for each group, we subtract those values for polls of all adults and registered voters to estimate what they would have shown if they had been conducted among the LV population. We run those polls through our average a final time, and that gives us a final EWMA for each party in demographic group!
The final step is to blend this EWMA with a model-based prediction of each party’s vote share that is used to smooth out phantom swings in the average that result from noisy data. The modeled prior for each subgroup is simply the national topline plus that group’s average historical offset from the national topline — giving us a stable estimate of where the group “should” be on any given day based on the overall national environment.
The amount of weight assigned to the modeled prior vs EWMA depends on the cumulative information provided by the polls for each subgroup: groups with many recent, large-sample polls that consistently agree with each other will be driven almost entirely by the EWMA, while groups with few, small, or noisy polls will lean heavily on the modeled prior. This effective weight also incorporates a reliability score for each subgroup, estimated by comparing the variance of that group’s de-trended poll results to the median variance across all groups. Subgroups whose polls are noisy and inconsistent accumulate effective weight more slowly, keeping them closer to the modeled prior even when poll counts are similar to more stable groups.
The final average we report for each group is thus equal to an average of the EWMA for that group and the modeled estimate, using the following formulas:w = eff_weight / (eff_weight + 20)
and
final = w * EWMA + (1 - w) * model_prior
Where eff_weight is the cumulative sum of decayed poll weights (each poll’s sqrt(sample_size) * reliability, decayed by 0.96 per day since the poll was conducted). So at eff_weight = 20, it’s a 50/50 blend. At eff_weight = 60, it’s 75% EWMA. A brand new group with one small poll might have eff_weight around 5, giving ~20% EWMA and ~80% prior.
Footnotes
The polling average at FiftyPlusOne switched from estimates among registered voters to estimates among likely voters on May 25, 2026. This article exclusively uses estimates among likely voters, which may differ from what appear on our official tracker. Data in this article is as of 5:00 p.m. Eastern on May 26, 2026.