Mapping polling errors in US elections, 2016–24

In my previous post, I mapped polling error through the 2022 elections. That post went up before the 2024 election, so its maps could tell us about past misses but not the one we were about to get. This update adds the 2024 results and revisits the same question: How much does polling accuracy vary from state to state?

The broad answer is that 2024 was a better year for polling than the previous two presidential elections. But polls still tended to overestimate Democrats. Those statements are compatible. Polls can get closer to the result while continuing to miss in the same direction, and a relatively small miss can matter quite a bit in a close election.

There is also a practical change. The old analysis depended on FiveThirtyEight, which has since closed (Dunbar, 2025). For this update, I’m using the Silver Bulletin pollster-ratings archive. It provides both polling estimates and election results, including 2024. I’ve rebuilt the earlier years from that source as well, so this is a fresh comparison using one dataset throughout.

Error and bias

As before, I’m distinguishing the size of a polling miss from its direction. Error is the absolute difference between the margin in a poll and the margin in the election result. Bias also tells us which party’s margin the poll overestimated. I label that direction with D or R: +3.0 D means overestimating the Democratic margin by 3 percentage points, while +3.0 R means overestimating the Republican margin by the same amount. A poll showing a Democrat ahead by 2 points in a race the Republican wins by 1 has an error of 3 points and a bias of +3.0 D.

Here is the descriptive starting point for the polls included in this analysis. Each poll record receives equal weight in this table. These are averages across the sampled statewide contests, not national popular-vote polling errors or the errors of a polling average.

All errors are in percentage points. Blue D labels indicate overestimating Democrats; red R labels indicate overestimating Republicans. Odd-year contests are grouped with the following even-year cycle; a runoff retains the source’s election-cycle assignment.

The contrast between presidential and midterm cycles is hard to miss. The 2016, 2020, and 2024 cycles all show an average overestimate of Democrats. The 2018 and 2022 cycles are much closer to zero. But the average absolute error in 2024 is also appreciably smaller than in 2016 or 2020. Calling all three presidential cycles the same kind of polling failure would obscure that improvement. The mix of states, offices, and pollsters also changes across cycles, so this table alone cannot tell us how much of the difference reflects better polling methods.

This general pattern is consistent with the AAPOR task force’s evaluation. It found improved accuracy in 2024 alongside another average overestimate of Democratic performance (AAPOR, 2025, pp. 4–5). Its figures differ from mine because it uses a different collection of polls and a final two-week window. The numbers above are calculations from the Silver Bulletin data under the inclusion rules described below.

The updated maps

The maps use a multilevel model to adjust for pollster differences, partisan sponsorship or affiliation, sample size, and time until the election. Think of each estimate as the expected miss for a nonpartisan poll with a sample of 800, fielded on election day, with the pollster-specific effects set to zero. This is a standardized comparison across places and cycles. It is not the observed error of a particular polling average.

Choose a cycle and switch between bias and absolute error. Blue indicates overestimating Democrats; red indicates overestimating Republicans. The absolute-error view uses a separate scale because it has no partisan direction. By default, gray states have fewer than five qualifying poll records in that cycle. Gray does not mean zero error. There are 23 states meeting that minimum in 2024, but 37 with at least one poll. The minimum-poll selector lets you include those sparse states. Iowa, for example, has three records, so its estimate appears when you lower the minimum to three or one. The cutoff affects the display, not which polls enter the model.

Select a state, or hover over it, to see the estimate, its 95% credible interval, and the amount of data behind it. The interval describes uncertainty about the model’s mean, not the range in which a future poll must fall. Estimates based on very few polls depend more on the model’s partial pooling, so I would interpret them accordingly.

Open the full-size maps or download the state estimates.

The multi-cycle option averages each state’s cycle estimates that meet the selected minimum, weighted by its number of poll records in each cycle. It can help summarize a state’s recent experience, but it also combines different kinds of elections and gives more influence to cycles with more polls. I would look at the individual years before treating the average as a reliable correction for the next election.

Where 2024 missed

For a closer look at the presidential election, the table below isolates presidential polls in seven closely contested states. These are simple averages of individual-poll errors in the final three weeks. They are intentionally separate from the map estimates, which pool presidential, Senate, and gubernatorial races and apply the model’s adjustments.

All seven show an average Democratic overstatement, but the size varies. It ranges from about 0.9 points in Georgia to 3.5 points in Arizona. Michigan also stands out relative to Pennsylvania and Wisconsin. A blanket adjustment of the same size in every state would miss some of this variation.

There is a limit to what we can infer from that pattern. These states did not all have the same pollsters, number of polls, or field dates. The map’s model helps account for some of those differences, but it cannot identify why respondents were missing from a survey or why a particular adjustment failed. Three presidential cycles with Democratic overestimation give us reason to take the possibility seriously. They do not establish a fixed bias that can simply be subtracted from every future poll.

How this works

Data

I use the February 2025 release of Silver Bulletin’s raw pollster-ratings data, preserved on its archive page (Silver & McKown-Dawson, 2025/2026). A newer release includes 2025 elections, but this post stops with the completed 2024 cycle. A local snapshot keeps the analysis reproducible if the download changes or disappears.

The analysis includes 3,008 poll records for presidential, Senate, and gubernatorial contests in the 50 states, from the 2016 through 2024 cycles. All have a poll date within 21 days of their election, using the archive’s polldate field. National polls, primaries, House races, the District of Columbia, territories, and presidential results in individual congressional districts are excluded. I retain only contests whose two listed candidates are a Democrat and a Republican. This excludes, for example, the 2024 Senate contests involving Angus King in Maine and Dan Osborn in Nebraska; it avoids treating an independent candidate’s error as Democratic bias.

The same restrictions apply to the error and bias analyses. The source has already made decisions about which surveys and versions to include, so these are records in that archive, not an exhaustive inventory of every public polling release. I use the supplied candidate margins, orient them as Democratic minus Republican, and calculate error and bias myself. I also correct three Monmouth records labeled 2024 whose contest identifiers and election dates place them in 2022; the replication files document those corrections. Some source results are rounded, so the displayed estimates use one decimal place.

The shorter window matters for geographic coverage. The previous post described a three-week window, but its code actually included polls up to 61 days before the election. This update consistently uses the final 21 days supplied by the replacement archive. In the previous 538 file, restricting the 2020 cycle to that window reduces the number of states with at least five eligible polls from 43 to 33. Silver Bulletin also gives 33 for 2020. For 2022, both sources give 23 states at that threshold, the same as this update’s 2024 count. The historical estimates should therefore be compared within this update rather than spliced onto the previous maps.

There is also evidence of a real change in where polling took place. In the replacement archive, presidential polling alone falls from 488 records covering all 50 states in 2020 to 315 records covering 35 states in 2024. AAPOR (2025, p. 43) independently reports that 15 states had no publicly released presidential polls in the final two weeks of 2024, and another 11 had only one or two. In 2016 and 2020, it found at least one in every state during the same two-week window. Although its window and contest coverage differ from mine, this helps explain why even a comprehensive late-election archive would leave many states sparsely represented.

Model

I retain the joint multilevel model from the previous update, estimating signed bias and absolute error together with correlated residuals. Each outcome has adjustments for poll partisanship, log sample size, and days until the election. The grouping terms allow differences across pollsters, pollsters within cycles, states, and states within cycles, along with cycle-specific differences by poll partisanship. Partial pooling helps keep estimates for places with limited data from depending too heavily on a few unusual polls.

mvbf(
  bias ~ partisan + log_n + days +
    (1 | pollster_rating_id/year) + (1 | location/year) + (partisan | year),
  error ~ partisan + log_n + days +
    (1 | pollster_rating_id/year) + (1 | location/year) + (partisan | year),
  rescor = TRUE
)

Here, log_n is the log of sample size divided by 800. Predictions set it and days to zero, set partisanship to nonpartisan, and retain the state and cycle effects while omitting pollster effects. I use weakly regularizing priors and four sampling chains. The maximum R-hat is 1.006, with 0 divergent transitions. The analysis code and source snapshot specify the priors and other fitting details.

There are several limitations. The Gaussian model for absolute error is an approximation to a nonnegative, asymmetric outcome. Its intervals can extend below zero in sparse states, although negative absolute error is not physically possible. The model does not separately estimate effects for each office or account for every dependence between polls of the same race. Its credible intervals are conditional on these modeling choices. States with five polls can still have substantial uncertainty, and states with no qualifying data are not evidence of accurate polling. Most of all, these estimates describe past polling performance. Applying them to a future election would require the additional assumption that the relevant biases will recur.

Sources and downloads

American Association for Public Opinion Research. (2025). Task force on 2024 pre-election polling: An evaluation of the 2024 general election polls. Report.

Dunbar, M. (2025, March 5). Political poll news site 538 to close amid larger shuttering across ABC and Disney. The Guardian. Article.

Silver, N., & McKown-Dawson, E. (2026, January 14). Silver Bulletin pollster ratings 2025 archive [Data originally released February 17, 2025]. Archive and original workbook.

My derived data are available as included poll records, descriptive cycle summaries, state estimates with credible intervals, and presidential, Senate, and gubernatorial summaries. Poll data are used with attribution to Silver Bulletin under the permission stated on its archive page.