Spain's Public Pollster Under Review: Where the CIS Misses, Why, and How to Read It

We reviewed the nine elections since 2023 for which the CIS ran a pre-election survey. Its estimate overstated the Socialists (PSOE) by 2.3 points on average and fell short for the conservative PP (1.9) and Vox (1.4). The skew does not come from its modeling: it is already in the raw answers of its sample. Once adjusted, the CIS is still a useful source, and that is how we use it.

21 October 2026 · Miguel · Leer en español

Spain’s Center for Sociological Research (CIS), the government-funded polling institute, runs the country’s largest regular survey: about 4,000 interviews a month, with raw data that anyone can download for free. It is also the most contested. To see how much it can be trusted, we did the simplest thing possible: compare what it said before each election with what voters actually did.

Since 2023, the CIS has run a pre-election survey before nine elections: the 2023 general election; the 2024 regional elections in Galicia, the Basque Country and Catalonia and the 2024 European election; Extremadura’s regional election in 2025; and those of Aragon, Castile-León and Andalusia in 2026. For each one we measured two things: the error of the estimate it published and the error of its respondents’ raw answers.

Four parties, nine elections

The pattern repeats with few exceptions:

• PSOE: almost always too high. The CIS overstated it in six of nine elections, by 2.3 points on average. The biggest misses came in Galicia 2024 (6.3 points) and Extremadura 2025 (5.8).
• PP: almost always too low. It fell short in seven of nine, by 1.9 points on average; in the 2024 European election, by 5.2.
• Vox: too low in eight of nine, by 1.4 points on average, and increasingly so: in the three regional elections of 2026, by 2.7 to 3.5 points.
• The left of the PSOE: too high, by 1.5 points on average in the five elections where it can be measured properly, and by 4.1 points in the 2023 general election, when the PSOE estimate was nearly spot on.

A two-point miss is not a disaster for a poll, and no survey gets every election right. What stands out is the direction: always the same way. Random errors would cancel out across elections; these do not.

The problem is the sample, not the modeling

The usual suspect is the CIS’s “cooking,” the step in which it turns raw answers into an estimate. We checked using the raw data: we calculated each party’s vote directly from what respondents say, with no modeling at all, in four different ways.

Raw stated intention (what people say they will vote, taken at face value) already overstates the PSOE by 1.7 points and understates the PP by 2.2. The published estimate adds about half a point to the PSOE, but most of the skew is there before any modeling: in who agrees to answer the survey.

Fixing it is not trivial. The standard method used by private pollsters, adjusting the sample so that recalled past vote matches the previous result, gives smaller average errors but very erratic ones: in the 2023 general election it would have understated the PSOE by 5.6 points and overstated Vox by 4, and in the 2024 European election it would have understated the PP by 9 points.

A clue: what TV its sample watches

In its survey before the 2023 general election, the CIS asked respondents which TV channel they mainly followed the campaign on. Comparing the split among the five national channels with the actual audience of their newscasts reveals a clear mismatch.

The CIS sample had too many viewers of public broadcaster La 1 and of laSexta, and too few of Telecinco and Cuatro. If the sample is weighted so that its channel split matches the audience, the skew in raw stated intention shrinks a lot.

The PP gap falls from 3.7 to 1.9 points and the PSOE gap from 4.8 to 3.2; Vox and Sumar end up nearly spot on. It is a single election, and the channel question is not exactly the same as audience share, so it is not conclusive proof. But it is a clue to where the problem lies: in the profile of the people who agree to answer.

How we use it at altnalysis

We never use the CIS’s published estimate. We use its raw data in two ways:

• To see where votes come from and where they are going. We weight the sample to official population figures from Spain’s statistics office (INE), use a model to fill in the past vote of respondents who do not reveal it and the vote of undecided respondents, and derive the flows of voters between parties. In the 2023 pre-election survey, this method cut the average error for the four main parties from 2.9 to 1.7 points.

• As one more poll in our polling average. Its raw intention, weighted by each respondent’s likelihood to vote and adjusted by its average skew across these nine elections, goes in alongside the other pollsters, and the model estimates its own “house effect,” as it does for any other firm.

2026: the CIS, raw and adjusted

In the September survey, raw intention put the PSOE at 31.4% and the PP at 24.8%. Adjusted for its historical skew, they come to 29.4% and 27.1%. Even so, they remain far from what other polls show, as captured in our forecast: PP 31.7% and PSOE 26.1%. In other words, in 2026 the CIS is further off than usual; our polling average detects it and assigns the CIS a house effect of roughly 4 to 5 points against the PP and 2 to 3 in favor of the PSOE. That matches what happened in this year’s three regional elections, where its raw stated intention overstated the PSOE by 2.4 points on average.

What we don’t know

• Part of the error may be late shifts. Pre-election surveys are conducted three to six weeks before election day.
• There is only one general election in the series. The other eight are regional or European elections, with different turnout and different parties.
• Party groupings are approximate. In four elections the left of the PSOE cannot be measured properly because an estimate or a result is missing for one of its parties, so we left them out.
• We are not judging anyone’s intentions. We measure deviations; the possible causes (who answers, how questions are asked, when) cannot be fully separated with these data.

How we calculated it

• Data. Estimates published by the CIS and raw data from its pre-election surveys 3411, 3437, 3448, 3453, 3458, 3538, 3543, 3545 and 3558, compared with official results. 2026 monthly surveys: studies 3540 to 3577.
• Averages. Average deviation in points; the model works on a logistic scale with a random-effects model that separates the average skew from the variation between elections.
• TV. Raking of the sample to the June 2023 newscast audience share of the five national channels, on top of the population weighting.
• Code. sesgo_cis.py, backtest_m4.py, backtest_medios.py, cis_rebase.py and articulo_cis.py, in the altnalysis repository; details in docs/sesgo_cis.md.