Polling Average Methodology [2026 Update]
How they work, plus a few changes in preparation for the 2026 election cycle
US politics are all about public opinion, and there’s no better way to capture that than a well-designed poll. Unfortunately, polls can be prone to error, whether it be due to limited sample size or poor methodology.1 Averaging a large number of surveys minimizes both sources of error.
1. Collection
We try to include as many polls as possible by referencing different sources. This involves doing online searches, following pollsters on social media, and checking our lists against those compiled by Wikipedia, RacetotheWH, RealClearPolitics, the New York Times, Silver Bulletin, and DecisionDeskHQ.
If pollsters release different versions of a poll, we use the likely voter version if the poll is about an election. For approval, we use the broadest population.2 We also prefer the version that pushes undecided voters. For polls that include minor third party candidates3, we use the version with major candidates until October. That’s because these minor candidates tend to decline in vote share, and including them caused slight extra movement in our historic backtesting.
2. Weighting
We assign each poll five weights:
Recency: as polls age, their weight decays exponentially. For election averages, weighting is less aggressive early on.
What changed: our models now calculate averages on two settings—one slow, one fast—and put more weight on the fast one as Election Day approaches.
Sample size: larger samples are favored in all of our averages. We use the same system as the now defunct 538 averages, weighting polls by the square root of their sample size, with no additional weight for very large polls.4
What changed: we reduced the weights of larger polls in our election averages. Empirically, large samples don’t seem to help much, possibly because accuracy is more dependent on post-sampling adjustments and turnout modeling. All samples of over 1500 are given no extra weight in our election averages. For our approval averages, which are not meant to predict elections, the limit is raised to 10000 (designed to catch mass panels).
Population polled: elections are determined by those most likely to vote, so our state/district-level averages favor likely and registered voter polls (weighted 1 and 0.9, respectively), followed by adult polls (0.7).
What changed: this is a crude way to adjust averages toward the preferred population, and is no longer used in our national polling averages. That’s because national polls are more common, giving us far more data about differences between populations.
Pollster quality: all of our current and recent averages are weighted using Silver Bulletin ratings, while historic averages use the old 538 version of those ratings.5 We deduct 5% from a pollster’s weight for each Silver Bulletin/538 grade drop (A+ weighted 1, A 0.95, and so on). Ungraded pollsters get the median weight of 0.65.
Lastly, we limit the weight of pollsters that release surveys frequently to prevent any one firm from dominating the average. This “flood penalty” weight uses the following formula:
weight = 1 / polls released in 3-week windowAll five weights are then multiplied together to determine a poll's overall weight.
3. Prior adjustments
A few factors can tell if us a poll is “biased”—not in terms of accuracy, since we can’t know that ahead of time—but in terms of consistently deviating from other polls:
Population polled: whether a poll uses an adult or registered/likely voter screen. To handle polls of different populations, we look at pollsters who release results for different populations in the same poll. Polls are then adjusted towards likely voters for our election averages, and all adults for presidential approval.
What changed: because polls of multiple populations are rare, the adjustment was previously made using polling and a modeled prior. That blurred the line between polling average and full-blown election model, so we’re removing the prior to keep our averages polls-only.
National adjustment: polls at the state/district level are adjusted for shifts in the national environment (as measured by the generic ballot in congressional races). This is based on the trend line adjustment pioneered by 538.
Pollster-specific partisan lean: we estimate each pollster’s historic lean relative to other pollsters based on data from Nate Silver/538.
4. Final average
Pollsters have “house effects”—something in their methodology that produces consistently different results from other firms—not captured by the above adjustments. This can have a big impact on unadjusted averages.
For example, if more Democratic-leaning firms start releasing polls, an unadjusted average will shift towards Democrats even if public opinion hasn’t actually changed. Alternatively, let’s say state A is being polled by lots of Republican-leaning firms, while state B is primarily polled by Democratic-leaning firms. As a result, B will artificially look worse for Republicans than A.
We address this by tracking how individual pollsters differ from the simple average calculated in steps 1-3 (excluding the pollster in question), then adjusting each pollster using its average difference. We do this for both a pollster’s lean (i.e producing better results for one party/candidate or different approval for the president) and the share of undecideds. The new average uses these adjusted values, with slightly less aggressive weighting than the simple average to avoid “double-counting” weights.
There are cases where deviating from the average isn’t evidence of bias. For example, a poll may be capturing a real shift prior polls hadn’t, or it might just be deviating due to statistical noise. That’s why our averages take a “skeptical” approach to calculating house effects: if a pollster hasn’t released many polls, or if there aren’t enough other pollsters to compare to, the estimate is reverted towards the prior (the historic lean discussed in step 3). We also make two cycles of calculations to avoid smoothing out genuine shifts in the polls.
For our Senate poll averages, where data is a lot sparser, we set a prior using historic data and house effects calculated from national (generic ballot) polls. The prior is updated using two types of calculations: one for a pollster’s house effect across all Senate races, and one for its effect in individual races.
Check out some of our averages below!
Methodology is the main source of polling error, but sample size-related error is the more commonly reported “margin of error.” When a pollster says that their poll has a certain margin of error, that doesn’t include error due to methodology (question wording, failing to reach a certain demographic, etc.)
All adults are the broadest, followed by registered, then likely voters.
Those who are consistently polling in the single digits.
This is based on the statistical formula for the margin of error.
These ratings give more weight to pollsters with track records of accuracy, transparent methods, and less tendency to “herd” (release results similar to other firms).