Original research · Preregistered experiment

Does trust in an election change based on how the forecast is worded?

A controlled experiment with 1,012 people found that, when the same forecast is worded as "10% chance of losing" instead of "90% chance of winning," the voter whose candidate loses is far less likely to think the forecast was wrong, and a bit less likely to doubt the election's legitimacy.

Type
Working paper · preregistered
Method
Controlled experiment (RCT) 2×3
Sample
1,012 participants · Prolific (U.S.)
Authors
Del Carpio & Rodriguez-Paiva · 2024

It's a 2024 working paper, not yet peer-reviewed. We're publishing it transparently, with the preregistration, data, and materials open for anyone to review or replicate.

The question

Losing an unexpected election fuels distrust

When someone votes for the candidate who loses, they tend to distrust the process and the system more. This is called the winner-loser gap, and it's not a minor detail. It erodes perceived election legitimacy, and Toshkov and Mazepus (2022) found its consequences can even affect the health, happiness, and life satisfaction of those on the losing side.

Forecasts feed that tension. Communicating that a candidate has a "90% chance of winning" inflates the sense of certainty about the outcome. When the unlikely happens and the candidate loses, that surprise can turn into suspicion. As Mongrain (2023, as cited in the paper) notes, wrongly anticipating an outcome "tends to create a state of psychological discomfort, which might even breed denial."

Our question was direct. Can that distrust be reduced just by changing how the forecast is framed, without touching the number? That is, by communicating the same probability as "10% chance of losing" instead of "90% chance of winning."

What we did

An experiment that simulates defeat

We designed a controlled experiment, preregistered before collecting a single data point, and ran it with 1,012 people in the United States via Prolific.

1

We framed it

It was a 2×3 design. The forecast was presented as probability of winning or its equivalent probability of losing, with three levels of favoritism: clear favorite (90/10), moderate favorite (75/25), or slight lead (60/40).

2

We simulated the defeat

Each person read that the candidate they supported was the favorite according to the forecasts… and that they ultimately lost the election. Exactly the scenario that triggers distrust.

3

We measured

We measured two perceptions on a 1-to-7 agreement scale. Was the forecast wrong? Is the election's legitimacy doubtful? The higher the score, the more distrust.

The finding

The same probability, different trust

Level of distrust by how the forecast was framed
Average agreement (1-to-7 scale) with each statement, among those who saw the forecast framed as "probability of winning" versus "probability of losing." Lower = less distrust.
Framed as "probability of winning"…as "probability of losing"
Means by framing condition (Del Carpio & Rodriguez-Paiva, 2024). The vertical lines show the 95% confidence interval. The dashed line marks point 4 on the scale.
On the forecast, the effect is large. Agreement that "the forecast was wrong" drops from 6.14 to 3.71 on the 7-point scale just by changing the framing. It's a large effect (η² = 0.34), not a marginal one.
On legitimacy, the effect is real but small. Suspicion about the election's legitimacy drops from 3.25 to 2.81. Framing helps, but distrust of the forecast and distrust of the election don't move with the same force.
The insight

The number didn't change. "90% winning" and "10% losing" are identical; the only thing that changed was which word carries the focus, and that alone was enough to substantially shift how much the losing voter trusts the forecast.

Why it should work

It wasn't a blind bet. The design builds on prior findings: how a probability is worded shapes what people expect, beyond the number itself (Teigen & Brun, 1999). And across different contexts, negatively worded information and messages tend to be perceived as more truthful and trustworthy (Hilbig, 2009, 2012; Koch & Seeger, 2017). The paper reviews several more mechanisms; these two were the starting point.

The nuances

A clearer favorite, more doubt about the forecast if they lose

How lopsided the forecast was also matters. When there was a clear or moderate favorite, agreement that "the forecast was wrong" was higher than when the lead was slight; between the clear and moderate favorite conditions, the difference wasn't statistically significant.

It makes sense. The more a victory was anticipated, the more surprising the defeat, and the easier it is to think the forecast was wrong.

The bigger the favorite, the more the forecast is questioned if they lose
Agreement that "the forecast was wrong" (1-to-7 scale) by how lopsided the favorite was, averaged across both framing conditions.
Means by forecast favoritism level (Del Carpio & Rodriguez-Paiva, 2024). The vertical lines show the 95% confidence interval.

One of the study's attention checks asked people to recall the exact forecast number. Some people, instead of repeating the number they'd read, wrote its complement: if they'd been told "10% chance of losing," they answered "90% chance of winning."

131
of the 132 people who made that spontaneous conversion came from the group shown the forecast as "probability of losing." The reframing barely happened in the other direction. The paper doesn't report what share of that group this represents, but it's a clue that which word carries the focus matters even in how the number is remembered.
How we made it rigorous

The study was preregistered on OSF (hypotheses, design, and analysis plan fixed before collecting data, so results couldn't be cherry-picked after the fact). We calculated sample size with a power analysis and, out of 1,219 responses, kept 1,012 after strict attention-check filters. Of the two preregistered hypotheses, we confirmed that framing affects perception; we did not confirm that favoritism moderates that framing effect, since the interaction between both factors wasn't significant once covariates were included. The framing main effect holds even controlling for gender, age, education, and political affiliation, so it isn't explained by participants' profile.

In an exploratory, non-preregistered analysis, framing also reduced suspicion about the election's legitimacy, but only when there was a clear or moderate favorite, not when the lead was slight; since this came after the data, more evidence is needed to confirm it. We detail other limits of the study below.

It's a working paper and hasn't gone through peer review yet; that's why we're keeping the data and materials open.

The limits

What this study doesn't say

Being honest about limitations is part of the method. These are the ones the authors themselves acknowledge in the paper.

The sample is from the United States and isn't representative of the country. Participants were recruited in English via Prolific; the effect could differ in contexts with different partisan dynamics, such as Latin America.
Question order may have influenced the answers. Everyone answered first whether the forecast was wrong, then whether the election was illegitimate; the authors note that this fixed order could have affected the second answer.
The questions were worded negatively. Agreement was measured with "the forecast was wrong" and "the legitimacy is doubtful," not their positive-worded versions. Since the study is precisely about how framing affects perception, the authors themselves flag this as a limitation.
Only one candidate's probability was shown, not both. Real polls usually show both candidates' probabilities together (for example, "A has 10%, B has 90%"). The authors have already announced a follow-up study using that full format.
Why it matters

A low-cost word, an effect on trust

Communicating a forecast as "probability of losing" costs nothing and misleads no one, because it's exactly the same probability. But it can measurably reduce the distrust of whoever loses after an unexpected result.

For pollsters, media outlets, or forecasters, it's a concrete lever, a wording decision backed by experimental evidence that can soften the blow to trust in the system, especially in the single-probability format we tested here. It doesn't solve polarization, but it moves part of the problem at the cost of changing a sentence.

This study also shows how the lab works. We start with a question, turn it into a preregistered hypothesis, test it with a controlled experiment, and measure the effect. It's the same method we apply with our clients, here applied to a question of public interest.

Have a behavioral question worth measuring?

We design and run the experiment that answers it with evidence.