When a survey asks who you support, it receives one person's preference. When it asks what people in your social circle support, it may receive information about many people who were never sampled. In January 2022, Émile Servan-Schreiber highlighted a paper by philosopher and social epistemologist Kristoffer Ahlstrom-Vij arguing that this second question can create an implicit super-sample.
Servan-Schreiber recommended the research because it also offers a possible explanation for a familiar forecasting result: expectation questions and prediction markets can outperform polls even when their participants are not representative of the electorate. The paper's claim is not that representation stops mattering. It is that a respondent can carry information about more than one preference.
From a sample to a network
Traditional preference surveys estimate a population from the people selected into the sample. Social-circle surveys add a layer: each participant reports what they know about friends, relatives, colleagues or other contacts. Those contacts are not formal respondents, but their preferences can enter indirectly through the people who know them. That is the super-sample.
The attraction is clearest when a conventional sample is small or uneven. A sampled person may know how several non-sampled people think, including people who are difficult for a pollster to reach. Asking about the circle can therefore recover information that a question about the respondent alone would miss. It can also reduce the social cost of revealing a controversial preference: describing a friend's view may feel easier than admitting one's own.
The three objections the paper tested
Ahlstrom-Vij's simulations did not assume an ideal social world. They examined three reasons the approach might fail.
- Selection bias: some groups may be less likely to respond or may be absent from the recruitment channel.
- Egocentric bias: people may project their own preference onto others and overestimate how widely it is shared.
- Homophily: people tend to associate with others like themselves, creating social bubbles rather than miniature representative populations.
The simulations generated 500 populations of 1,000 people and repeatedly compared traditional and social-circle estimates across sample sizes and different levels of connectedness. Even one known neighbour could improve estimates in very small samples. The advantage became larger as each respondent contributed knowledge about more people, and it remained robust under severe simulated selection bias, egocentrism and homophily.
Simulation evidence is conditional: it shows what follows from the model's assumptions, not that every real survey will behave the same way. The paper also reported a survey experiment in which a short warning about egocentric bias reduced projection onto others. Ahlstrom-Vij described that result as preliminary, and the article did not present the prompt as a complete cure.
Why an expectation question can carry more information
The distinction between intention and expectation is small in wording but large in information. “Who will you vote for?” reports one intended vote. “Who do you expect will win?” invites a respondent to combine that intention with conversations, local observations, media, polls and beliefs about turnout. The answer can therefore summarize an informal network rather than only the speaker.
The paper cites research on US elections from 1952 to 2008 that found aggregated voter expectations more accurate than aggregated intentions. That earlier work estimated that one expectation response carried predictive information comparable to twenty intention responses. The number belongs to that historical setting; it should not be treated as a universal conversion rate.
The possible link to prediction markets
Prediction markets also ask for an expectation rather than a personal preference. A trader deciding whether a candidate is underpriced has to think about what everyone else will do, not just whom the trader wants to win. Ahlstrom-Vij proposes that this question may tap the same implicit super-sample as a social-circle survey.
That mechanism is a hypothesis, not the only explanation for market accuracy. Markets can also reward research, expose participants to changing prices, and give more influence to confident or successful traders. The paper notes that financial incentives alone are incomplete as an explanation, because studies have often found little or context-dependent accuracy difference between play-money and real-money markets.
What the research changes
The practical lesson is not to replace representative polling with one clever prompt. It is to notice that different questions collect different information. Preference questions estimate what respondents want. Social-circle questions estimate what respondents observe around them. Expectation questions ask them to integrate those observations into a view of what will happen.
For forecasting systems such as Hypermind, that distinction helps explain why a non-representative panel can still be informative about an election. Participants are not being used as a miniature electorate. They are being asked to estimate the electorate, update as evidence changes, and accept a score when the outcome arrives.
