Paper Seats: What a Poll Can and Cannot Tell You

With 1,000 interviews, a poll cannot know how many seats a party will win: from sampling chance alone, there are about thirty seats between the best and the worst plausible result for the conservative People's Party (PP). Even so, almost four in five polls published since 2023 give seats as an exact figure. We explain what a poll really measures, what it does not, why polls resemble each other more than they should, and why it pays to rely on how people have actually voted in the past.

23 October 2026 · Miguel · Leer en español

Every week, polls are published that split the 350 seats of Spain’s Congress of Deputies down to the last seat. They make convenient headlines and they are easy to share. The problem is that a poll cannot produce those figures. This is not an opinion: it is arithmetic. This article explains what a poll can measure, what it cannot, and what we do so as not to depend on it more than is reasonable.

What a poll measures: votes, with a margin

A poll of 1,000 interviews estimates each party’s vote share with a margin of error. For a party at around 30%, the 95% interval is about ±2.8 points if the sample were perfectly random. It is not: polls are weighted by sex, age, region and recalled vote, and that reduces the effective sample. With a design effect of 1.6, the one we use in our aggregator, the margin rises to ±3.6 points. For a party at 10%, from ±1.9 to ±2.4.

With 3,000 interviews the margin falls, but less than people tend to think: to halve the error you have to quadruple the sample. A party at 30% still has a margin of ±2.1 points.

The law requires every published poll to come with its technical details, including the sample size and the margin of error (article 69 of the Ley Orgánica del Régimen Electoral General, LOREG, Spain’s electoral law). But that margin is usually given for the whole sample and for the most favorable case, not for each party, and almost never for seats.

From votes to seats: the model supplies them, not the poll

Seats are not allocated from the national vote but in 52 constituencies (the 50 provinces plus Ceuta and Melilla). To know who wins the last seat in Soria you have to know how Soria votes. And a national poll has hardly any interviews in each province.

If a sample of 1,000 interviews is spread in proportion to the electorate, Soria gets 2, Teruel 3, Segovia 3 and Ávila 4. Ten constituencies have fewer than 5 interviews; twenty, fewer than 10; and thirty-six, fewer than 20. Madrid has 139 and Barcelona 113. With 3,000 interviews, eleven constituencies are still below 20.

With two interviews you cannot estimate anything. So the seats that accompany a poll do not come from the answers in that province but from a model: almost always, each province’s previous result shifted by the national change the poll measures. It is a reasonable method, but the seats belong to the model, not to the poll.

How much seats swing from chance alone

To measure it we ran an experiment. We took the 2023 result as the truth, when the PP won 137 seats, the Socialists (PSOE) 121 (including the Catalan Socialists, PSC), the far-right Vox 33 and Sumar, the left alliance, 31. We simulated 2,000 perfect national polls: no bias, no in-house adjustments, and using the same projection method as almost all pollsters. The only source of error is the chance of picking 1,000 or 3,000 people.

With 1,000 interviews, the PP ends up between 118 and 149 seats in 90% of polls, and the PSOE between 110 and 142. Only one in twenty gets the PP’s exact figure right. With 3,000 interviews, the interval narrows to 126–144 for the PP and 115–133 for the PSOE, still about twenty seats wide.

And this is the best possible case. Real polls add non-response error, weighting error and projection error: the people who answer are not the same as those who do not, recalled vote is unreliable, and the vote does not change equally in every province.

Exact figures for thirty seats of uncertainty

Since the 2023 general election we have collected 413 polls from private pollsters that give seats for the PP. In 321 of them, 78%, the figure is exact, with no range. When there is a range, its median width is two seats. A two-seat range is about nine times narrower than the one left by pure sampling chance with 3,000 interviews.

Giving seats is not wrong in itself: it is useful to know what a vote estimate would mean. What misleads is presenting them as if they were measured data, with a precision the poll does not have.

Polls that look too much alike

There is another signal you can check. If each poll were an independent sample, their results should differ from one another by at least what sampling chance dictates. We compared the actual dispersion of polls in the last three weeks of each campaign with what would be expected from their sample sizes.

In 8 of the 12 cases, the polls resembled each other more than sampling allowed. With the PSOE it is very pronounced: its variance was between 17% and 29% of the expected one in 2016, in November 2019 and in 2023, and under half this October. It may be because they all weight the same way, because they share assumptions, or because they adjust toward what the others publish. With this data the cause cannot be known, but the result is the same: many polls agreeing is not independent evidence that they are right.

The 2023 case is illustrative. The polls agreed on the PSOE far more than chance allowed, at around 28%, and they all missed in the same direction: the PSOE got 31.7%, more than three points above the average of the final week (we covered it here). With the PP the opposite happened: in April 2019 and in 2023, the pollsters disagreed with each other more than chance explains, a sign of very different methods.

A poll is also a message

Polls do not just measure: they also influence. In an experiment with a national sample, David Rothschild and Neil Malhotra found that showing people how much support a policy had changed their opinion of it (Rothschild and Malhotra, 2014, Research & Politics). And in the French overseas territories, which voted when the exit polls from mainland France were already known, knowing that data reduced turnout by about 11 points and increased the vote for the expected winner (Morton, Müller, Page and Torgler, 2015, European Economic Review).

That is why the law bans publishing polls in the five days before the vote, and why it matters how the ones published earlier are presented: an exact seat figure conveys a certainty the poll does not have, and the reader has no way of knowing it.

What does work: looking at how people actually vote

Our forecast uses polls, but as one piece among others, not as the result. The starting point is what people actually did:

• The real vote in about 36,000 census sections in the seven general elections from 2008 to 2023, with their income, age and composition. That tells us how the vote of each neighborhood and each town moves when the national vote changes, instead of assuming they all move the same way.
• Who actually votes. People overestimate their likelihood of going to vote, and not all equally. We calibrated it by comparing what they answered in the CIS polls (the state polling institute) with what they did on 23J (the 23 July 2023 election): those who did not vote in 2023 overestimate the most. For each province’s turnout, using its past behavior leaves an error of about 2 points, against 4 or 5.5 for using the national average.
• What each pollster tends to get wrong. We measure each pollster’s historical deviation from the result from 2015 to 2023 and discount it, with more or less weight depending on how stable it has been.
• The error they all share. The final polling average has missed in each of the last five general elections, and with the PSOE always in the same direction: too low. That common error does not disappear when you average, so we build it into the forecast instead of ignoring it.

With all that we run 4,000 simulations of the election, province by province, and publish probabilities and intervals, not isolated figures. Today, the PP wins between 112 and 155 seats in 95% of our simulations. It is a wide interval because the uncertainty is real, and saying so is part of the forecast.

What to look for in a poll

• The sample and the fieldwork dates, which the law requires to be published. If it does not have them, there is no way to assess it.
• The vote, not the seats. Seats depend on each pollster’s model; the vote, on the sample.
• The margin for each party, not the overall one. For a party at 30% with 1,000 interviews, about ±3.6 points.
• The average of many polls rather than a single one, knowing that the average can also fail if they all fail together.
• The ranges. If a poll gives seats without a range, information is missing; if the range is two or three seats, it is narrower than the sample allows.

How we calculated it

• Margin of error. 95% interval of a proportion, 1.96 × √(p(1 − p) × design effect / n), with a design effect of 1.6.
• Interviews per province. Sample spread in proportion to the 2023 resident electorate.
• Simulation. Official 2023 result by constituency (votes from the BOE, Spain’s official gazette). In each of the 2,000 simulated polls, the national vote of each party and the blank vote is drawn from a multinomial with n / 1.6 interviews; each province’s vote is multiplied by the ratio between the poll and the real national result, and the 2023 seats are allocated with D’Hondt and the 3% threshold.
• Ranges. Polls from private pollsters compiled from Wikipedia since the 2023 general election, excluding the CIS.
• Dispersion. Polls from private pollsters with fieldwork ending in the three weeks before each election (in 2026, October 1 to 21), one per pollster and date. Expected variance: mean of p(1 − p) / n across polls, times 1.6; with no published size, 1,000. A trend during those weeks would increase the observed dispersion, so the comparison favors the polls.
• Code. articulo_escanos_papel.py, in the altnalysis repository.