Election polls get it wrong: The human biases behind the numbers - opinion

Hundreds of election polls, and hardly any two tell the same story. It is tempting to blame voter secrecy, biased polling, or whatever explanation happens to make the numbers feel more coherent. But are voters simply saying one thing and thinking another? Not quite.

There is the answer we give out loud, the “skip this question” button, the I haven’t decided yet, the choice we eventually make in the voting booth, and sometimes, simply, the answer that feels easier to give to a stranger asking how we plan to vote.

These may sound like small, insignificant behaviors. Put them inside an election poll, however, and they can become the difference between what the public says it will do and what it does. Before a poll turns into percentages, blocks and projected seats, it must pass through the least predictable variable in the equation: human beings.

There is something almost reassuring about an election poll. In the middle of a noisy, emotional and constantly shifting political reality, a table appears and puts everything in order: percentages, blocs, seats and margins. The numbers look so exact that it is easy to forget that before any one of them made it onto a chart, there was a person being asked a question. That is where the problem begins.

People are not neutral measuring instruments. We are influenced by our surroundings, by the way a question is phrased, by what feels comfortable to say, by how we want to be perceived, and even by how much patience we have left by the time we reach question 17.

Researchers describe these effects in terms of cognitive, behavioral, and methodological biases. But bias does not mean respondents are lying. Most of the time, they are not even aware that these mechanisms are at work.

One of the most basic problems is known as non-response bias. Sometimes the most important question in a poll is not what respondents said, but who never responded at all.

If the people who ignore a call, swipe away a message, or refuse to participate are systematically different from those who do answer, a sample can look impressively representative on paper while still missing something important about the public it is supposed to measure.

Then there is self-selection bias. Someone who chooses to open a political survey, spend time on it, and make it all the way to the end is not necessarily a random member of the population. They may be more politically engaged, have stronger opinions, or simply be the kind of person who is more willing to participate in surveys.

And even after we have found someone willing to answer, another question remains: how much thought did they put into those answers? Or, put differently, how thoroughly did the survey wear them down?

This is where satisficing comes in. It is the point at which we stop searching for the most accurate answer and settle for one that feels good enough to move on. Anyone who has completed a long questionnaire knows the feeling. The database records a valid response. The respondent may have recorded little more than fatigue and a desire to get to the next screen.

Alongside it sits another familiar concept, the spiral of silence. When people feel that their opinion is unpopular or unacceptable within their social environment, they do not necessarily change it. Sometimes they simply stop voicing it.

They may say, I haven’t decided yet, avoid the question, or offer a less controversial answer, only to reveal their actual preference in the privacy of the voting booth.

I know. By now this is beginning to sound like the study notes someone drops into the student WhatsApp group before an introduction to social sciences exam. The difference is that in an election year, these are not terms you memorize to pass a course. They describe the exact points at which a poll can begin drifting away from the reality it is trying to measure.

For an election poll to predict the ballot accurately, a surprising number of things must go right. It must reach the right person. That person has to agree to participate, think seriously about the question, feel comfortable enough to state an honest preference, and then, finally, still make the same choice weeks later behind the privacy screen.

At every stage, something can shift or twist.

So age, gender and other characteristics can be adjusted for, but the question remains: to what extent are people who agree to speak to pollsters systematically different from those who do not?

Research has also long suggested that the less socially acceptable people believe their political position to be, the less likely they may be to voice it openly. There is no conspiracy here, and no secret army of hidden voters.

Political preferences are simply expressed inside a social interaction, and social interactions carry consequences that some of us would rather avoid.

The effect can appear even when people are asked about behavior that has already happened. A 2024 study found that when respondents were told that researchers could verify whether they had actually voted, self-reported turnout fell by 5.7%.

Sometimes we are not only telling pollsters what we did. We are also telling them something about the kind of person we would like to believe we are.

But if these mechanisms repeatedly appear in research outside Israel, there is at least a reasonable basis for asking whether some of them might operate here as well.

I am not a prophet, nor do I pretend to be one. None of this proves that such distortions will necessarily occur, and it certainly does not tell us which political camp might benefit or lose from them. Human biases do not carry party membership.

They measure people at a particular moment, in a particular context, while asking them to say out loud what they believe they will do later. A poll meets us while we are talking to someone. The ballot meets us when no one is watching.