The figures 30.5, 17 and 5.0 per cent are point estimates, not exact vote counts. A methodologically correct calculation can still be read too strongly in a headline, especially for a one-point change or for a party on the five-per-cent edge. A nationwide Chamber of Deputies model, moreover, does not describe the result in a particular municipality or Senate constituency.
What we know: The August wave of Trendy Česka used fieldwork on 3–21 August 2026, a CATI/CAWI mix, 1,200 respondents and 991 respondents included in the election model. Czech Television published point estimates and a range of statistical error of ±1.0 to ±3.3 percentage points according to the size of support.[1]
What we do not know: From the public article we do not know every parameter needed for a precise interval of the difference between the May and August waves — in particular the covariance, the detailed design effect and the procedure used for a given pair of estimates. The change therefore cannot be tested by a simple comparison with the error of a single wave.[19]
1. A point that lies inside measurement error
Imagine a Sunday meeting in a television newsroom. The editor-in-chief wants ANO’s number, because the graphic must have a winner. The statistician wants an interval, because without it the point is only decoration. The political reporter wants a story about a shift towards STAN, because without a story the discussion will not open. The party spokesperson wants “the opponent’s collapse”, because a campaign lives off movement. The sociologist wants to warn that 991 people are not ten million. The viewer wants to know “who is ahead”.
At first glance this is a dispute about courage. In fact each person is describing a different problem. The editor-in-chief compares intelligibility. The statistician sampling error. The reporter a narrative. The spokesperson a fight over the frame. The sociologist the population from which inference is drawn. The viewer a desire for a ranking.
Each role in the newsroom pursues a different aim, but a point estimate without its uncertainty can create a conclusion that the data do not support. The phrase “mood of the nation” here conceals the selection of respondents, weighting, the wording of the question, the list of parties, declared turnout and the fieldwork period.
First it is necessary to determine whether we are comparing point estimates, their uncertainty, the composition of the party list, or actual votes. Each of these quantities answers a different question.
A point is not a mood. A mood is not turnout. Turnout is not a mandate. And a headline is not a measurement.
— Jiný Kontext
2. From the respondent to the number on television
When 30.5% appears on screen, it is easy to form the impression that “this is how the Czech Republic votes”. The figure, however, arises in a chain. Each layer can raise, lower or change the kind of error. Two estimates with the same logo therefore need not measure the same thing, and the same estimate can behave differently according to whether we read the point, the interval, or the headline.
Six layers of a single percentage
- Population: adults in the Czech Republic, not “the nation in its soul”.
- Sample: who was reached, who refused, who entered the model.
- Question and list: a stand-alone party, or a coalition.
- Weighting: certainty of turnout, quotas, corrections.
- Measurement error: the published range of statistical error and the information the agency gives about how it was constructed.
- Publication frame: graphic, headline, commentary, sharing.
Kantar CZ’s Trendy Česka project for Czech Television is one of the longest-running surveys of this type in the country. That matters, and it is not immunity from the layers. The official description of the August 2026 wave states fieldwork from 3 to 21 August, CATI and CAWI techniques, a representative sample of 1,200 respondents by sex, age, region, education and place of residence, and the entry of 991 respondents into the election model — those who do not rule out voting. In this wave the agency states the model’s statistical error in the range ±1.0 percentage point for parties with a low share to ±3.3 percentage points for parties with a high share.[1]
Distinguishing the layers is not a methodological ornament. It determines where to turn in a dispute. If the point moves by one percentage point for a party with a high share, the appetite of the headline may have failed, not the population. If SPD is measured without partners and compared with a coalition result at the ballot box, the list in the question has failed. If a mayor is read from the model, the unit has failed. The sentence “the poll got it wrong” is the start of a diagnosis, not its end.
Illustrative scheme: The same point estimate can, in a different interval, with a different question and a different headline, produce a different conclusion. Comparing a point with the ballot box without the layers is the same error as comparing a thermometer with a month-ahead forecast.
3. How costly an error in the headline is
The ranking of parties is only part of the information; the meaning of an error depends on its impact. A typo in an infobox can be corrected quickly. A headline about a “fall below five per cent” is far more serious if the published error range crosses both sides of the threshold and a later correction no longer reaches the same audience.
ESOMAR and WAPOR, in their guideline on published polls, require that the public be given enough information to judge quality: who measured, when, whom, with what question, how large the sample was and what the uncertainty is.[2] The ICC/ESOMAR code and the Czech SIMAR standards translate the same demand into the operating language of member agencies.[3][4] It does not follow that every graphic is a fraud. It follows that publication without an interval fails in a different discipline from the calculation.
| Type of failure | What it looks like | How to test it | What limits the harm |
|---|---|---|---|
| A point instead of uncertainty | “It fell by a point” | Is the uncertainty of the difference between both waves published? | The point and the uncertainty in the graphic |
| The threshold as certainty | “Six parties in the Chamber” | Does the interval cross 5%? | Ban a binary reading at the edge |
| Coalition artefact | Comparing a party with a coalition at the ballot box | The same list in the question? | A description of what was merged |
| Non-response | The willing speak, the tired stay silent | Who refused, whom the weight does not catch | Humility towards the unknown remainder |
| A swapped unit | A Chamber model as a municipality | A different law, a different contest | Separate the type of election |
That also changes the editorial question. We do not ask only which estimate hit the average. We ask which way of publishing will fail so that the reader still sees it. An analytical programme can tolerate ten nuances. A news ticker cannot tolerate a single certainty where the agency itself draws a range of several points.
The quality of publication depends not only on the model, but also on the visibility of uncertainty and on the strength of the conclusion in the headline. A clear methodology and cautious wording are more useful for public debate than a sharp number without explanation.
4. What can be read from the published error
Polls are needed because they replace part of impression with a systematic estimate. The problem arises at the moment when one number starts to represent a party’s exact support. For the August 2026 wave Kantar published a range of statistical error of ±1.0 to ±3.3 percentage points.[1] This information belongs with the point estimate, not only in a methodological note.
For a party with a high share, according to the published description, the upper bound of this range applies. ANO’s point estimate of 30.5% can thus, after simply subtracting and adding the published error of one wave, be written as an indicative band of 27.2 to 33.8%. This is not a formal confidence interval: the public article does not state a confidence level, nor whether the published error fully includes weighting and sample design. Czech Television also reported a month-on-month drop of one percentage point.[1] The statistical significance of this difference cannot be judged by comparing one point with the error of the August estimate. We need the standard error of the difference of both waves and information about their dependence; with independent samples the errors combine, with an overlap of respondents covariance enters the calculation.[19] The published figures therefore do not by themselves suffice for a strong headline about change, nor for the claim that support has not changed.
For a party with a low share the agency states ±1.0 percentage point. The Motorists’ point estimate of 5.0% is therefore not the sentence “a sixth party in the Chamber”. Simply subtracting and adding the published error gives only an indicative band of 4.0 to 6.0%, not a documented formal confidence interval; the band reaches both sides of the electoral threshold.[1][5] Act No. 247/1995 Coll. sets a five-per-cent threshold for a party’s entry into scrutiny; the model does not vote the threshold.[5]
Five facts without which a point is only a promotional number
- The size of the survey sample and the size of the model subsample.
- The fieldwork period and technique (telephone, online, mix).
- The wording of the question and whether parties stood separately or in a coalition.
- The error interval at the given size of share, not “some plus or minus”.
- What the wave does not measure: the ballot box, a municipality, the Senate, tomorrow’s affair.
For the purchase of political calm, therefore, a small honest interval matters more than someone else’s overall ranking. Thirty genuine readings with an interval say more than a thousand shares of a point.
5. Compare constructions, not party mascots
The market of models can be described without a false podium of winners. The following map does not say who hit the voter’s soul. It shows around what work the individual constructions build an estimate, and which failure the reader should look for on purpose.
| Construction | Main working position | What to check when reading | Typical reason for use |
|---|---|---|---|
| Kantar / Czech Television, Trendy Česka | A hypothetical Chamber, CATI/CAWI mix | The wave’s interval, the model’s n, the party list | A long series, a public methodology |
| Another agency model | A different sample, a different question, a different weight | Whether points can be compared across agencies | A second estimate, not a clash of logos |
| Aggregation of several waves | An average of maps, not an average of souls | The weight of n, the age of the fieldwork | Damped noise of a single wave |
| The result at the ballot box | Actual votes in the given type of election | A different law, different turnout | The only verdict on seats |
| A municipal / Senate map | Thousands of contests, constituencies, a second round | A ban on reading from the nationwide model | A different unit |
The table does not rate agencies as a whole. It separates different outputs so that the reader does not mistake a well-calculated model for an unduly certain headline, nor a single wave for a confirmed trend.
Interpretive scheme: The columns of the table are not an accounting statement of support. They separate the estimate, the interval, the threshold and the ballot box.
6. Anatomy of a wave: what Kantar actually calculated in August 2026
On the August wave Czech Television reported point estimates and, at the same time, a frame of error. Honest reading holds both in one sentence, not the point in the headline and the interval in a note.[1]
| Subject | Wave point estimate | Error Kantar states for the size of share | Indicative band from simply subtracting/adding the published error | What a headline can bear |
|---|---|---|---|---|
| ANO | 30.5% | ±3.3 p.p. (high share) | 27.2–33.8% | A stable first band, not “a fall of a point” |
| STAN | 17% | in the wave’s range of ±1.0 to ±3.3 p.p. | a band of several points around 17 | Direction can be watched in a series, not in a single point |
| ODS | 15.5% | the same range of the wave | a band around 15.5 | “Stagnation” is often noise |
| Pirates | 8.5% | the same range of the wave | a band around 8.5 | A one-point move is weak evidence |
| SPD | 6.5% | closer to the lower bound of the range | a band around 6.5 | Comparison with a coalition at the ballot box separately |
| Motorists | 5.0% | ±1.0 p.p. (low share) | 4.0–6.0% | The indicative band reaches both sides of the threshold |
On ANO, Czech Television reported a drop of one point against May and of four points against the election result; on STAN a rise of 1.5 points against May and of six against the election.[1] Comparison with the last ballot box and with the previous research wave are two different things: turnout, the list and the method of measurement all change. Without the uncertainty of the difference, a one-point month-on-month move cannot itself be labelled a statistically demonstrated change. A series of more comparable waves can support a trend, but this article does not determine one from a single pair.[19]
Estimated turnout in the same wave was 62.5% of citizens aged over 18; Kantar added that since March the indicator has not changed markedly and that interest usually rises only before an election is held.[1] This is a declaration in a questionnaire, not turnout measured at the ballot box.
What needs to be tested
The first test is statistical: do we know the uncertainty of the difference between the two waves being compared, or only the error of each point separately? The second is about the list: did the parties stand in the question as they last stood at the ballot box? The third is temporal: which days were used for fieldwork and which event fell into them. The fourth is about the unit: does the text speak of the Chamber, or already of a municipality?
For an individual who wants one broad orientation, the Trendy Česka series may be decisive. For a newsroom the sentence “we have Kantar” is not enough. It needs to know n, the interval, the question and the boundary beyond which a person must approve the headline.
7. A coalition is not the sum of parties. A party is not a coalition
A second source of crooked reading is the form of the list. If the model tests parties separately — SPD without partners, or ODS, TOP 09 and KDU-ČSL apart — it measures a different thing from the one a pre-election coalition assembled at the ballot box.[1][6] That is not evidence of an outflow. It is a change of object.
Comparing a separately measured party with a past coalition result is a comparison of two accounting axes. Support may genuinely have fallen. It may have fallen only because the adding-up on the ballot paper broke apart. Without the same list in the question this cannot be decided.
Kostelecký, on a Czech example, describes a prediction model that combines a nationwide representative survey with the spatial distribution of support from previous elections, and explicitly addresses the conditions that affect the quality of the estimate.[6] Linek and Lyons analyse the long-term stability and change of party support in the years 1990–2010; their work is not evidence that a particular one-point move in the August 2026 wave represents a real change.[7] It does not follow that agencies “cannot count”. It follows that a model answers a particular question by means of particular data and assumptions.
A pleasant style of graphic can conceal the same old problem: a smooth curve from two incommensurable points. The quality of the visual lowers the cost of reading. At the same time it can raise the cost of detecting a methodological error, because the line does not look suspicious.
8. Five per cent is a threshold in law. It is not a certainty of the graphic
Headlines like to announce that six parties would enter the Chamber. For a party with a point estimate of 5.0% and a published error of ±1.0 percentage point, a simple calculation gives only an indicative band of 4.0–6.0%, which reaches both sides of the clause. It is not a formal confidence interval, because the public source does not describe its full construction.[1][5]
That means something more modest than “a sixth party”. It means that the wave does not know on which side of the legal threshold the party would finish if the election were held on the days of fieldwork. A binary break has a large impact on who would assemble a majority. Presenting the point 5.0 as a completed entry is therefore a crude simplification of the probabilistic nature of the data — and Kantar itself acknowledges that nature in the methodological sentence on error.
The law on elections to Parliament sets the clause as a condition of scrutiny, not as a property of a preference on the telephone.[5] A model can estimate who might miss the threshold. It cannot abolish it or confirm it. A newsroom that turns a point into a seat is buying a certainty that a sample of 991 people at the edge does not have.
9. Who picks up the phone is not the same as who comes to vote
Even the best model runs into willingness to respond. Groves and colleagues in Survey Methodology describe non-response as a structural feature, not as an isolated defect of a wave: more willing, more educated and more politically engaged people enter a survey differently from the tired and the unreachable. Weights try to correct this shift. They do not remove it without remainder.[8]
The AAPOR Transparency Initiative and SIMAR rest on the demand that enough methodology be published for it to be possible to judge who is missing.[4][9] It does not follow that the 1,200 people reached are a fraud. It follows that inference from 991 model respondents to millions of those entitled to vote is an estimate, not a census.
Imagine an illustration, not a report. In July a fictitious Movement X has 22.4% in the model at n around 950. In August 21.1% at n around 980. In the table a minus of 1.3 points lights up. The headline writes of a crisis. The intervals of both waves overlap so widely that support in the population need not have moved. The narrative arose from noise. The names and figures of this illustration are not from Kantar; they are only meant to show the same arithmetic that the August wave itself, for large shares, acknowledges with a range of ±3.3.[1]
10. Declared turnout is a promise. Turnout at the ballot box is something else
The August model estimated turnout at 62.5%.[1] The Czech Statistical Office, at actual elections, measures something else: how many of those entitled came to the ballot boxes in the given type of election, in the given round, in the given weather of the campaign.[10] A declaration in August and turnout in October are not two versions of the same column.
The difference is not cosmetic. The result is decided by who comes on Saturday afternoon, what mobilisation there is in the last two days, and how many people with a fickle preference stay at home. Linek and Lyons describe this fickleness as part of Czech voting behaviour.[7] Kantar itself, on turnout in August 2026, noted that it has not changed markedly since March and that a rise in interest usually comes only before an election.[1] That is an honest reservation. A headline about “certain turnout” would erase it.
Two similar columns can have a different data regime. The same percentage in the model and at the ballot box can arise from different people. A reading pilot must record which regime is being cited. Otherwise the newsroom tests the comfort of the graphic and by mistake issues it as a property of voters.
11. Uncertainty does not vanish after publication
Before an interval was published, the reader had to sense that the number was not exact. With an interval the check is something more complicated: whether the headline cites the interval, or overrides it. A headline is navigation towards meaning. It is not automatic proof.
A single report can combine three true figures into a false conclusion. First: ANO’s point is 30.5. Second: against May that is one point less. Third: in open questions voters mention promises. Each figure may be in the wave, but the synthesis “the nation has turned away” is wider than the data. At other times a newsroom cites a high-quality secondary comment even though a primary description of n, the fieldwork date and the error is available.[1]
A pleasant headline can conceal the same old interval
A well-cadenced sentence about “losing one’s breath” sounds like a mark of understanding. It is not one automatically. In news work it is necessary to check in particular whether the system has not mistaken a point for an interval, a stand-alone party for a coalition, the Chamber for a municipality. The quality of wording lowers the cost of reading. At the same time it can raise the cost of detecting a factual error, because the sentence does not look suspicious.
A four-step check of the headline
- Existence: the number is in the wave and has a fieldwork date.
- Authority: the primary source is the agency and the client, not a copy of a copy.
- Support: for a claim about change, the uncertainty of the difference between waves is documented; the error of one wave is not enough.
- Scope: the population, type of election, coalition and unit fit.
The most reliable working procedure therefore separates the table from the word. First a row is created: estimate, n, interval, question. Only then the headline. If the interval cannot be opened, a claim is not promoted to a fact merely because it declines well.
12. One model is not a market. Aggregation changes a different error
Nate Silver in The Signal and the Noise shows that a single forecast can be noisy, while a set of forecasts with honest uncertainty can damp the noise — and that even then it fails when the model asks the wrong thing.[11] An aggregation of Czech waves is not Silver’s American exchange. It is only a reminder that a clash of agency logos on a single day measures differences of construction more than the voter’s soul.
The report of an independent inquiry into the failure of British pre-election estimates in 2015 (the Royal Statistical Society and the related inquiry) showed that honest agencies can err in the same direction when sampling, weighting or willingness to respond fails, not when they “calculate the average badly”.[12] This is an audit of one particular election cycle in another country, not evidence that Czech models have no value. It does show, however, that a precise percentage can precisely measure a defective condition.
The Czech SIMAR standards and the international ESOMAR/WAPOR guideline therefore insist on transparency of publication, not on a promise of infallibility.[2][4] Act No. 234/2025 Coll. on election campaigns further addresses, among other things, campaign rules and related transparency of political advertising in Czech law; election laws have long restricted the publication of pre-election polls in the immediate vicinity of voting. That is a legal frame for dissemination, not a correction of the interval.[13]
13. The price of a point is the price of a raw material, not the price of a conclusion
The market uses several languages at once. The agency pays for fieldwork. Television pays for intelligibility. A party pays for a campaign according to movement. The reader pays with attention. A direct comparison of “who fell by a tenth” without knowledge of the interval is therefore misleading.
A cheap point can be economically excellent for orientation in the rough ranking of large camps whose intervals do not overlap. The same point can be expensive for a task on the five-per-cent edge that a newsroom repeats as a seat. An expensive paragraph on methodology can be pointless for a sporting ranking. It can be cheaper for a decision with a high cost of error if it reduces the number of unsuccessful narratives.
The true price of an accepted conclusion
Fieldwork + weighting + interval + human reading + verification of the question + correction of the headline + the price of a false movement + interruption of debate.
A conclusion is cheap only once it has passed an interval check and can be used without embarrassment the next day.
The costs also include fluctuation. If the same shift fits inside the error four times out of five, the average price of a “correct story” is not the price of one graphic. If the model, at the peak of a campaign, measures the affect of a week, the cost is the inertia of the headline. If campaign teams are tightly bound to a single number, the cost is also an inability to see another map.
It is therefore useful to measure the price on the accepted reading, not the price on a point. Accepted means that the reading met criteria set in advance: the interval, the question, the unit, and did not cause a later correction that nobody reads any more.
14. What to read for which question
Reasonable reading does not start with a party brand, but with the type of question and the control mechanism. The following recommendations are not a ranking of agencies. They are a signpost to a first reading. In each category it makes sense to choose two sources and compare them over the same period and with the same interval.
| Need | First group of candidates | Decisive test | Unacceptable shortcut |
|---|---|---|---|
| A rough ranking of large camps | A public wave with n and an interval | Do the intervals overlap? | Choose by a single tenth |
| The edge of the clause | Estimate + lower/upper bound of error | Does it cross 5%? | “It is in the Chamber” from a point of 5.0 |
| A trend | A series of waves, not one month | A shift outside the error repeatedly | One point against the previous wave |
| A coalition | The same list as at the ballot box | Party vs. bloc | An artefact as an outflow |
| A municipality and the Senate | Local data, constituencies, a second round | A different law | Reading a mayor from Kantar |
| Turnout | A declaration vs. the CZSO series at the ballot box | A different definition | Mistaking a promise for arrival |
For Czech use a separate language test is necessary. The English “margin of error” translates badly here when all that remains is “plus or minus three” for every party alike. In August 2026 Kantar differentiated the range by the size of share.[1] A test set of headlines should include the five-per-cent edge, a coalition, a municipality and a text in which the correct answer is “the interval is not enough”.
Some newsrooms eventually need more layers. One graphic with an interval, a second paragraph on the question, a third ban on reading the Senate. That is not a failure of intelligibility. It is the same decision as using a different tool for sport and a different one for the courts. The condition is that the division is visible and that people know where the final caution arises.
15. A short test that says more than a hundred headlines
A reading can be chosen without a month-long seminar and without an impulsive “ANO is falling”. A short, disciplined checklist is enough. Its aim is not to prove that the wave can do something impressive. The aim is to find situations in which the reading fails, and to see whether the newsroom can catch them.
Three tests of a single number
The first test concerns the difference: is its own uncertainty available for the claimed shift, and are the waves comparable? The second is about origin: is it a primary wave, or a copy of a copy? The third is contradictory: is there an equally honest figure — another wave, another question, turnout at the ballot box — that calls the conclusion into question? A faster headline is not better if it later cannot be traced back to the measurement from which it arose.
Before publishing a graphic
- Write the size of the survey and the size of the model.
- Write the interval for a high and for a low share.
- Separate the point that may stand in the table from the sentence that may stand in the headline.
- Mark whether the parties stood separately.
- Forbid reading a municipality and the Senate from the model if the model does not measure them.
The conditions must be comparable. The same week, the same question, the same type of election. If one text may turn a point into history and another must carry an interval, we are not testing the estimate. We are testing the courage to be imprecise — and we must name it as such.
The last output is not only the name of the winning party in the graphic. It is an operating rule: what the wave is used for, what it is not, which shift may enter the headline, what must be approved by an editor who knows the interval, and when the reading is opened again. Models, tariffs and campaigns change. Without a revision date even a good series gradually turns into an editorial superstition.
16. A Chamber model does not measure October in a municipality
The last, often cheapest swap is the unit. In August 2026 Kantar measured a hypothetical Chamber of Deputies.[1] October 2026 is about municipal councils and a third of the Senate. These are different laws, different constituencies, different turnout, different names. Chytilek and colleagues in Volební systémy describe that a majority two-round Senate and a proportional Chamber are not two versions of the same ranking.[14] Kostelecký and colleagues, on municipal elections, show the strength of local lists that have no row in the nationwide model.[15]
It does not follow that a nationwide mood has no effect on a municipality. It follows that the effect has to be proved by the municipality, not by transferring 30.5% into the town hall. The interval applies here too: even if ANO’s point in the model moved outside ±3.3, it would still not be the result in a municipality where there is a single list. For registration towards October 2026 the Czech Statistical Office recorded 1,842 such municipalities.[16] The Chamber model is silent there. It is not bad. It is about something else.
Agresti, in a statistics textbook for the social sciences, recalls that a confidence interval is a statement about a procedure, not a seal on an individual percentage.[17] Šubrt, in a survey of public-opinion research, translates the same point into the Czech tradition of the field: opinion is ascertained, not proclaimed.[18] A reader who wants October needs a different map. A reader who wants the Chamber needs an interval. Both can be right. A shared headline can still be wrong.
17. A graphic can overshadow the method
A poll chart today is not only an overview of ranking. Newsrooms build headlines from it, campaign teams adjust communication, and viewers form an idea of a probable result.
The contest is therefore not only for the best estimate. It is a contest for political context. For the date of publication, for the graphic, for the habit of reading a point. The more comfortable the ecosystem becomes, the higher the exit price can be. Campaign teams stop distinguishing their own campaign from the product in which it is measured.
A regular series can show a longer-term movement better than a single point. Its publication, however, must keep the question, the sample size, the published uncertainty and the comparability of waves, so that short-term noise does not turn into a headline about an earthquake.
Responsible interpretation does not pretend to a precision that a sample survey does not have. It formulates a conclusion in the range that the given wave and the published methodology actually support.
At the moment when the model only ranks large camps, we are choosing a map. At the moment when we assemble a majority from a point on the five-per-cent edge and a crisis from a single point against May, we are also choosing a new structure of responsibility. And then the question is no longer: “Who is ahead?”
Let us ask: what uncertainty belongs with the point and with its change, who may turn an estimate into a headline, and who will correct it when the data cannot bear a strong conclusion?
— Jiný Kontext