Same Night, Same Election, Fourteen Different Seats

What the margin of error can’t explain — and why it matters who’s asking

Opinion polls are estimates. Nobody should expect two polls conducted at roughly the same time to produce exactly the same result — different people are questioned, and chance alone produces some variation. But how much variation is reasonable?

On 23 September, News 13 and Channel 14 published polls on the same evening. Both gave Gadi Eisenkot’s Yashar exactly 22 seats. Yet News 13 gave Likud 18 seats while Channel 14 gave it 32 — a difference of fourteen seats.

Taken alone, that comparison is striking, but it can also be misleading. There are 120 seats in the Knesset; if one party gains fourteen in one poll, those seats must come from somewhere else. The real comparison isn’t Likud 18 versus Likud 32, but the political map each poll produced — and those maps were very different.

News 13 placed the coalition bloc at 51 seats, the opposition at 52, the Arab parties at 13, and the Hendel-Zelekha list at four, with Bennett and the Democrats at 11 each, United Torah Judaism, Israel Beiteinu and the Joint List at eight each, Otzma Yehudit and Shas at seven, Religious Zionism–Zehut at six, Ofer Winter at five, and Ra’am at five. Channel 14 reported a right-wing bloc of 63 seats: Likud alone at 32, Shas 11, United Torah Judaism seven, Religious Zionism–Zehut seven, Yashar 22, Bennett’s party seven, and the Democrats nine.

So the discrepancy wasn’t fourteen Likud seats wandering mysteriously between polls. It was two substantially different electorates, and — more importantly — two different bloc balances: 51 seats for the coalition in one poll, 63 for the right-wing bloc in the other. That is the phenomenon that needs explaining.

A little statistics helps, but only a little. Sample size is conventionally denoted n. News 13 had an n of 1,503 and reported a sampling error of 2.5%; Channel 14‘s NEXT DATA poll had an n of 647.¹ A larger n generally makes an estimate more precise, because random sampling produces less fluctuation as the sample grows — statisticians express that uncertainty through the standard error, which tells us how much an estimate might move if we repeatedly drew different samples from the same population.

But a larger sample doesn’t guarantee accuracy, and doubling it doesn’t double reliability — the statistical gain shrinks as n grows. More importantly, sample size cannot cure systematic problems: an unrepresentative panel, inappropriate weighting, different turnout assumptions, nonresponse, or different treatment of undecided voters. That is why the familiar “margin of error” can’t explain everything. It addresses random sampling uncertainty. It says nothing about the other choices that go into producing a political poll.

That’s nearly all the statistics a reader needs. The more important question is what happens repeatedly.

The September discrepancy didn’t emerge from nowhere. On 30 August, Channel 14 gave Yashar 24 seats, Likud 31, and the right-wing bloc 63. The next day, a News 12 poll also gave Yashar 24 — but Likud only 23. Then, in the days immediately around 23 September, a News 12 poll on the 22nd put Likud at 20 and Yashar at 23; the next evening News 13 put Likud at 18 while Channel 14 put it at 32. The Israeli electorate had not transformed itself overnight. The same political moment was being measured and represented in markedly different ways, more than once, in the same direction.

On 14 September, Kan News and News 12 published polls on the same evening that were broadly similar: both put Likud at 20 seats, while Yashar stood at 24 in Kan and 23 in News 12. This matters because it shows that sharp discrepancies are not inevitable when different organizations poll the same electorate — contemporaneous polls can also produce closely aligned results.

Polling professionals call this a house effect: persistent differences between organizations driven by different panels, weighting rules, turnout estimates, or methods for allocating undecided voters. A house effect need not imply anything improper. But naming the phenomenon doesn’t explain it — and if one methodology repeatedly produces a markedly stronger result for one political camp than another, the public is entitled to know why.

That is where the distinction between model and bias matters. Every polling model contains assumptions: whom the sample is meant to represent, how underrepresented groups are weighted, whether reported past voting should correct the sample, how likely each respondent is to vote, how undecided voters are treated, how parties near the electoral threshold are handled. Different reasonable choices can produce different results — and can also create systematic bias in the statistical sense, pushing estimates in one direction without anyone deliberately manipulating anything.

There is also a more sensitive possibility that can’t simply be ruled out. Election polls are commissioned and published by media organizations operating in an intensely political environment. A poll showing a governing party at 18 seats creates one narrative; the same party at 32 on the same evening creates another. That doesn’t prove manipulation — the figures alone can’t establish motive — so political or editorial influence has to remain a hypothesis to be tested against evidence, not a conclusion drawn from an inconvenient result. But neither should “methodology” become a black box that shelters every persistent discrepancy from scrutiny. If methodology explains the differences, the pollsters should be able to show how.

Transparency is where that showing happens — and it is uneven. The News 12 poll of 22 September states its sampling and error figures in some detail. The Channel 14 report of 23 September states that NEXT DATA surveyed 647 adults that day and that Shlomo Filber analyzed the data — and gives substantially less detail about sampling, weighting, interviewing mode, or statistical uncertainty. That gap in disclosure doesn’t prove one poll is better than the other. It means a reader can examine one far more easily than the other.

That matters because election polls are never neutral in their effects, whatever the intentions behind them. They generate headlines. They create perceptions of momentum and collapse. They shape discussion of alliances and of parties hovering near the threshold. They can affect how people vote strategically.

A one- or two-seat difference between polls is unremarkable. A recurring gap of eight, ten, or fourteen seats in a major party — accompanied by entirely different bloc outcomes — is something else, and the numbers alone can’t tell us whether the cause is a statistical model, a methodological bias baked into that model, or editorial pressure shaping how the model is built. Very likely it is some mixture of the three, in proportions no reader can currently see.

Israeli voters shouldn’t have to decide which poll to believe based on which channel they watch. They should have enough information to understand why polls taken on the same evening can describe such different political realities. “Margin of error” is not the answer to that question. It’s where the question begins.

¹ In May 2026, Haaretz reported that NEXT DATA was registered the previous October by Netanel Siman-Tov, described as a representative of Channel 14’s ownership, and is held through a holding company controlled by the channel’s controlling owner, Itzik Mirilashvili. Israel’s Press Council subsequently said the arrangement — an undisclosed ownership link between the outlet publishing the polls and the company producing them — may constitute a breach of public trust and a possible violation of journalistic ethics rules. This article does not draw conclusions from that fact about the specific numbers discussed above; it is noted here as public-record context relevant to the transparency question raised in this piece.