Why did surveys fail to reflect people’s choice? A dive into brainwaves has some insight: we are doing the wrong questions.

Disclaimer: Despite the results we share in this article, we at SEELE Neuroscience are not claiming we predicted or anticipated Trump’s victory. This is the report of a scientific study on the usefulness of understanding the underlying processes behind self-reporting a decision.

Georgian 80’s rock band R.E.M. had their “It’s the end of the world as we know it” hit in 1987, but instead of what the lyrics should suggest, we are far from feeling fine. Brexit surveys failed last June with margins far from all statistical errors and now, the 45th president of the United States wins with a margin absolutely no survey nearly forecasted. It is no news for anyone that we are doing something wrong. Is not a matter of sampling or even of statistical models, it is something simpler and more human: the willingness to express our decisions.

The “hidden vote”

Last year we performed for our customer Neuropolítika the now famous experiment where the same sample reported extremely different values between their declared vote intent for a candidate and what we call “neural vote” a fancy term for a brain association index based on EEG interhemispheric coherences. While our neural vote with a sample n=98 predicted a victory of the actual winning candidate with an error of 2.6%, what was more notable was that the same sample expressed a declared intent 14% bellow the actual voting results.


It was clear that certain part of the sample, whether the brainwaves said something, the verbal expression of the intent muted. Survey experts have called this part of the population that does have a vote intent but it is not willing on declaring it as the hidden or undisclosed voter.

With this background, last year we decided to measure and compare the declared perception of the most relevant candidates for the presidency of the United States, but in a very specific context: with Mexican population in our country. It is no secret that Trump’s campaign was heavily aimed to the south, especially under the script of building a great wall to avoid the incoming of more rapist, drug dealers and so. Mexican surveys claimed a generalized hate to Donald Trump’s speech while also a generalized acceptance of Hillary Clinton was on the rise. The perfect scenario to compare what a population says with what really think.

The experiment to dive into undeclared responses

So, when we talk about “neural vote”, what bogus claim can this be? Absolutely none, we promise. We are talking about our version of the very well-known Implicit Association Test but with an EEG component. The main problem with IAT is that it relies on reaction time, a variable that raises more questions than answers. What we do is use the paired-item model, where we present to each subject an item with a picture and a word or statement. The presentation of the item is for 5 seconds to give the subject to decide if the word describes the picture. If so, the subject must press a trigger button as fast as possible when the attention cross appears. The trigger button and the EEG system collect simultaneously the brain activity of the subject. The design looks more like this:


Under this setting, subjects are trained with items of explicitly-true associations, let us say, the picture of a cat and the word “Cat”, the picture of the sun and the word “sun”. Explicitly-false associations are alternated such as the picture of a Mexican flag and the word “Flag of Australia”, the picture of the moon and the word “chicken soup”.

Those familiar with BCI interfaces have already detected that what we are doing here is generate enough samples of brain activity when the subject needs to decide to push the trigger when faced to a explicitly-true or a explicitly-false association. The data what we use is the EEG power-spectrum hemispheric contralateral coherence for the bandwidth from 4hz to 12 hz segmented In ranges of 4hz: 4-8, 5-9, 6-10, etc.


Once we collect enough samples to identify patterns for both of the associations, we continued the trial with the pictures of Donald Trump and Hillary Clinton, each paired with different phrases the political experts suggest as useful statements to measure vote intent. These were: “I sympathize with”, “Represents me”, “I would never vote for”, “I would bet one month of my wage to his (her) triumph”, “I want him (her) to win the presidential election” and also each of the candidates name.

We performed the study last December 2015 with a sample of 98 politically active citizens such as militants, journalists, and consultants. The results were analyzed to establish an EEG association index, which is the proportion of similarity from 0 to 1 between each experimental item (such as the picture of Hillary Clinton with the phrase “Represents me”). One amazing finding was the predictive value of such EEG index when related to the proportion of sample that in fact pressed the trigger for each item:


The declared responses and the EEG index fitted a linear model with an adjusted R2=0.74, normality and constant variance test passed and a power with alpha 0.05:1. Each dot represents an experimental item.

What we learned about the Mexican “hate” to Donald Trump

Despite the wave of reactions, Internet memes and surveys reflecting a declared and factual hate to the new president of the United States, the results of our study revealed some disturbing realities. First, here are the results of the study:


The most evident insight is that Mexicans, back in December 2015, were aware of the high odds of Trump’s triumph, despite the very low sympathy detected. Moreover, Clinton’s results are least surprising: her higher EEG association index was with the phrase, “I sympathize with her” but was poorly associated with triumph perception.

With this results we really see a great opportunity to use neuroscientific tools to revisit the current way of exploring the decision making process. One way we are using this EEG association index, away from the “neural vote”, is to validate survey items before the implementation in large samples. In this way, we verify some key aspects of the items such as their sensibility to priming or likeliness of obtaining a skewed response.

About SEELE Neuroscience:

We are the leading lab in Latin America  specialized in translational neuroscience for the private sector with more than 10 years of expertise and six certified labs within the region. We do not have “proprietary methodologies” but translate the models and principles most accepted by the scientific community to answer everyday questions. We only use replicable and auditable methods; our tools are electroencephalography (EEG), Event Related Potentials (ERP) and Implicit Association Tests (IAT).  www.seele.education (in Spanish)