Research using neuroimaging to detect the emotional response of undecided voters has led to controversy among scientists. An op-ed article in the New York Times, written by the leader of one such study, argued that brain scans could help determine the voters' true feelings about candidates, eventually making pollsters obsolete. Dr. Geoffrey Aguirre discusses the flaws of this argument, the feasibility of this method to determine hidden preferences and the ethical issues inherent in the process.
By November 11, 2007, the Democratic and Republican presidential nominating contests were well under way. The Democratic candidates spoke that night at the Jefferson-Jackson fund-raising dinner in Iowa, and a second debate was approaching for the Republicans. With the first votes of the caucuses and primaries only weeks away, pollsters and pundits were working to divine the intentions of voters, particularly the coveted “swing” voters not committed to a candidate. Which Republican would appeal to women, closing the so-called “gender gap”? Was anyone truly undecided regarding Mrs. Clinton, a candidate who had been in the political spotlight for more than 15 years? That Sunday, the op-ed page of the New York Times promised insight into these central questions, in the surprising form of pictures of brain activity.
Neuroscientists from the University of California, Los Angeles, led by Marco Iacoboni, had used functional magnetic resonance imaging to measure the responses of undecided voters to the candidates. Their conclusions were startling in their depth and breadth. One Republican candidate, Fred Thompson, was found to evoke particularly strong feelings of empathy. Further, while some voters said that they disapproved of Hillary Clinton, their brain activity revealed that they had unacknowledged impulses to like her. The study had seemingly reached into the minds of voters and plucked out their hidden emotions and conflicts. Perhaps political talk-show hosts and Gallup pollsters would soon be unnecessary. Why analyze and poll when the feelings and intentions of voters could be read directly from their brains?
Instead of sparking a revolution in political science, however, the editorial provoked broad condemnation from the neuroscience community. Within days the New York Times had published a letter from 17 scientists who argued that the study was fundamentally flawed. At scientific meetings and on the discussion boards of Web sites the hue and cry continued. The prominent scientific journal Nature published a scathing editorial1 that lamented the absurdity of the study. After more than a decade of increasing publicity for brain-scanning results in the lay press, the Iacoboni editorial had provoked a backlash. Neuroimaging had jumped the shark.
For his part, Iacoboni defended his study. In an online letter, he argued that the approach he used in his study of voters is common to many cognitive neuroscience experiments. If all those previous studies were valid, he asked, was his study considered flawed simply because he had left the ivory tower to examine political candidates or reported his results in a newspaper? Iacoboni’s defense raises challenging questions for scientists and consumers of scientific studies. If his group’s undecided-voter editorial column is flawed, are there scientific studies that use comparable methods, published in respected, peer-reviewed journals, that are also absurd? What, exactly, was so wrong with his study given that it used modern neuroimaging techniques and analyses? Could there be valid studies of political topics that would either provide insight into political thought or be of value to a pollster or candidate? To address these questions, we must first understand how raw neuroimaging data can be transformed into a picture of brain activity that a researcher might interpret as showing latent sympathy for Hillary Clinton.
Brain Imaging Approaches
Magnetic resonance imaging (MRI) has been used for some decades to construct pictures of brain anatomy. Functional MRI (fMRI), developed in the 1990s, offers a measure of brain activity. For fMRI data to be collected, a participant lies on a table that is slid within a powerful magnet. The subject receives instructions and is presented with pictures and sounds during the scan. Meanwhile, weak radio waves are used to measure the effect that nerve cell activity has upon the magnetic field. The effect is indirect; local changes in brain activity induce a cascade of effects upon blood flow, upon oxygen, and in turn upon the iron atoms in hemoglobin molecules that ultimately warp the microscopic magnetic field. The procedure is extremely safe, painless, and it can be completed in about an hour. Nerve cell activity can be measured over the entire brain from second to second, and with millimeter resolution.
An image of brain activity is not available immediately after the scan. To create a picture, a researcher must first decide which two (or more) behavioral conditions are to be compared. This is an important, and generally unrecognized, aspect of neuroimaging studies. There is no brain picture “for” anxiety or memory. Instead, the experiment must compare the relative brain activity between two behavioral states, with the hope of isolating the mental operation of interest. To study anxiety, one might present the subject with pictures of snakes and guns and then at another time show pictures of puppies and flowers. The experimenter might conclude that a brain region, such as the amygdala, that shows a greater neural response to the snakes than the puppies is responding to the differential anxiety provoked by the stimuli. The colorful brain image simply shows where statistically greater activity was seen for one condition as compared to the other.
This approach to brain imaging, in which the experimenter tries to manipulate the mental state of a subject in order to then observe the evoked brain activity, is termed “forward inference.” Experiments like this dominated the application of neuroimaging for many years. The study of sensory processing has been particularly successful, in part because the mental states to be studied can be differentially evoked quite readily. For example, a brain region, “area MT,” has been identified that invariably responds when the subject sees something moving but does not respond to static pictures. Neuroimaging and forward inference have been used to study more-complex behavioral states as well, such as emotion, conflict resolution, sense of self and reward processing. Specific brain areas have been found that reliably increase their neural activity during these behaviors, although the link between a particular behavior and a brain region is more tenuous. First, it is challenging, and in some cases arguably impossible, to compare two complex behavioral states and leave behind the isolated mental concept of, for example, greed, or risk-taking. These behaviors are necessarily embedded in complex tasks and emotions and cannot be isolated by experimental design in the same way that visual motion may be. Second, the attempt to map a single behavior to a single brain region quickly breaks down past early sensory representation. The amygdala may consistently respond more strongly to anxiety-provoking stimuli, but it is also activated by positive stimuli (puppies and flowers) as compared to neutral pictures (toasters and trees). The state of affairs is even worse for areas of the frontal lobe, where dozens of different mental operations have been identified that might activate a given square centimeter of cortex. A related complication is that different subjects may have quite different behavioral or emotional responses to a particular experimental situation, foiling attempts to describe a consistent relationship between behavior and brain region for a population.
The application of neuroimaging to political questions does not involve “forward inference,” however. Political neuroimaging, along with the burgeoning fields of social, economic, and even marketing neuroscience, relies upon the opposite approach. Instead of determining the brain region associated with a particular behavioral state, a “reverse inference” study attempts to identify the behavioral state of subjects by observing their brain activity. Initially, studies of this kind examined basic sensory phenomena. The activity within the aforementioned area MT might be used to determine if a particular optical illusion induces a sense of motion in some people. Such a conclusion could be well supported. After dozens of “forward inference” studies, it has become quite clear that the perception of motion, and only motion, is always associated with activity in this patch of cortex. The reverse inference approach has also been used to probe more-complex behaviors. Activity within the insula when a subject is presented with recognizable lies has been taken as evidence that lies induce the same sense of disgust that rotten food does, as the latter has also been observed to activate the insula.
The Trouble with Reverse Inferences
The problem, of course, and the source of the widespread displeasure with Iacoboni’s newspaper article, is that these reverse inferences are only as good as the evidence that supports a unique mapping of a particular mental operation to a particular cortical region. And for many of the claims that Iacoboni makes, this evidence is not good at all. The presence of an amygdala response to pictures of Mitt Romney did not necessarily indicate anxiety regarding his becoming president, as positive emotions can activate this region as well. A further limitation is that the response to pictures of Mr. Romney was compared to (presumably) the neural response elicited by a blank screen. The amygdala response may have been not to Mr. Romney per se but to his attractive hair. Finally, even if we were to grant that amygdala responses indicate anxiety, and were specific to Mr. Romney himself, perhaps the subject was simply anxious because his favorite candidate, Mitt, was not doing well in the polls!
Further compounding these weaknesses is Iacoboni’s tendency to engage in what might be termed “neuromythology.” When presented with a picture of a brain with colorful activity, he has a tendency to spin a yarn to explain what he sees. The claim that voters who stated a dislike for Mrs. Clinton actually harbored latent kind feelings toward her was not even partially implied by the faulty logic of the study; rather, it was an explanation, made up from whole cloth, for the observation of cortical activity that implied “conflict.” This unfortunate tendency to treat neuroimaging data as a Rorschach blot is on full display in a recent article in the Atlantic in which the author, Jeffrey Goldberg, visits with Dr. Iacoboni and his associates who operate a “neuromarketing” company. The initially uncomfortable finding that Mr. Goldberg had a “positive, reward” response to a picture of Mahmoud Ahmadinejad leads to the tortured explanation that the author is actually imagining the happy day that the Iranian president is deposed. Equally bereft of logic is the explanation of how the equivalent responses of Mr. Goldberg’s brain to Hillary Clinton and his own wife actually signify two quite different behavioral states.
Does the preceding criticism suggest that a valid study of political behavior using neuroimaging is not possible? No. Instead, while there are pitfalls to be avoided, much might be learned regarding the behaviors and emotional states that people develop and deploy in evaluating political candidates. To be successful, such studies must compare carefully controlled states to isolate a behavior of interest and draw well-supported inferences regarding the activity seen. In fairness, Iacoboni and his colleagues have published an example of such a study2. Beyond simply being valid, however, there is an additional requirement that a neuroimaging study of political behavior be useful: it must provide an insight not available by simply asking a voter his or her opinion.
Imaging Versus Polling
For the most part, human behavior is readily available to be observed or queried. It would not come as a surprise to learn that voters who identify strongly with one party tend not to like candidates from the other party. Thus, it seems an unnecessarily roundabout way to learn this truth by measuring increased amygdala and insula responses to pictures of opposing candidates. Similarly, if you want to know how someone will vote for a candidate, you can generally just ask the person. The chief challenge for pollsters is obtaining a sample of responses that are representative of the population, a problem that would not be solved by neuroimaging. There is nothing automatically more informative about measuring neural activity as compared to directly observing behavior.
There are many circumstances, however, in which asking voters their opinions will not provide the entire story. In the face of an overt desire to mislead or a simple lack of introspection, neuroimaging of political behavior might provide insights not otherwise available. For example, a plausible study might examine the emotional response to political “spin.” Politicians frequently provide an unrealistically favorable description of events, omitting details that are inconvenient. While voters claim that they object to spin, they may nonetheless respond positively. Given previous studies that have identified patterns of brain responses for overt lies as compared to truths, what is the response to spin? Is spin treated as a lie, and how is this modulated by one’s political affiliation? There are certainly many other topics in the realm of political behavior that fall into this category and could eventually come under study.
We may also consider applications of neuroimaging techniques to assist polling in cases where voters are unwilling or unable to provide accurate responses. Obviously, a source of much uncertainty in polling results is “undecided voters.” Perhaps some proportion of voters really do have a strong preference but are insufficiently confident to share this with a pollster. Further, voters may consider one candidate to be the more socially acceptable choice to report to the pollster, although they intend to choose the other in the privacy of the voting booth. This is the “Bradley effect,” named for Tom Bradley, an African American former mayor of Los Angeles who lost his 1982 race for governor despite polling that showed him ahead of his white opponent.
Could neuroimaging be used to determine true voting preference in these cases? Perhaps, although not in any straightforward way. Simply presenting the candidates’ pictures and recording a response would not be enough. As we have considered, the presence of, for example, an amygdala response to one candidate cannot be taken as evidence that the voter will vote a certain way. Recently, techniques to analyze the pattern of neural responses across the entire brain have been developed. These “multi-voxel patterns” (MVPs) can be used to deduce a subject’s unstated intention in controlled settings. For example, if a subject is presented with two targets on a screen and told to choose one but not yet indicate which, the choice can be accurately read from the MVPs in advance of the response. It is possible that the pattern signature for responses for a given voter could be measured while the person is making a series of innocuous decisions. In the critical test, the subject would then be presented with pictures of the candidates, side by side. Although the voter would withhold an overt response, the implicit preference might be available in the distributed fMRI data.
Suppose that this were shown to be a valid way to measure implicit voter preference—would it be of practical value? Only a small number of subjects could ever be examined in this fashion, as the collection of such data is a time-consuming and expensive undertaking. Further, obtaining a representative sample would be very difficult, as older subjects, for example, generally find it hard to participate in an hour-long, uncomfortable neuroimaging scan. Finally, simple polling questions and adjustments are available to address these challenges. Undecided voters can be asked to indicate which way they “lean,” which predicts well how they will ultimately vote. The magnitude of the Bradley effect can be estimated by asking a voter if she thinks her friends and acquaintances would be hesitant to vote for a certain candidate, even if she professes to have no such qualms. Indeed, a recent paper in the journal Science3 has demonstrated that purely behavioral techniques can be used to accurately predict the way an undecided subject will eventually vote.
Therefore, it seems unlikely that neuroimaging techniques will have much impact upon the practice of politics. Ultimately, politicians and political operatives care about behavior—if and how a voter will vote—and not much about the underlying neural basis for these actions. Simple polling provides this information much more readily and inexpensively than neuroimaging could ever do. In contrast, neuroimaging may find a place in the study of political science, in which the underlying motivations and behavioral states of voters have become an area of increasing interest.
Neuroimaging Our Preferences Versus Our Preference for Neuroimages
We have considered that neuroimaging techniques may be able, in principle, to identify voter preference. While this ability may be desired by politicians, it may be rejected by the polity. The secrecy of an individual’s ballot is a cornerstone of modern democracy; if our voting preferences were known we could be subject to the threat of retribution by a government we voted against. Fortunately such an abuse of neuroimaging is unlikely. Given the size and noise of an fMRI scanner, no one could be scanned unknowingly. Moreover, an fMRI study requires tremendous subject cooperation, making these studies trivially easy to defeat.
While of little immediate risk, the possibility that neuroimaging might invade our political privacy has been of concern to ethicists who anticipate the impact of emerging neuroscience technologies. This attention is not inappropriate. It is almost certainly better for philosophers and ethicists to have their say before a technological revolution sweeps an unprepared society. I believe, however, that the attention and concern devoted to the possibility of a neuroimaging invasion of political privacy is somewhat misplaced. Greater and more immediate threats to privacy loom. In the same way that behavior in a laboratory setting or in a formal poll can accurately predict a voter’s preference, so can our routine, daily actions provide a window to our intentions. Knowledge of where we live, what we buy, how we travel, and who we know can be aggregated to provide information about our preferences. The possibility of this silent, creeping invasion of our privacy, advanced by profit-seeking corporations and terrorist-seeking government agencies, strikes me as far more menacing than the clanging of a seven-ton MRI scanner.
Instead of a threat to privacy, the principal risk is that misuse of neuroimaging will add further distraction and irrelevance to the political process. Although carefully designed neuroimaging studies might eventually provide valuable insights into political decision making, the slow, unglamorous grind of the scientific process will leave us time to be tempted by colorful pictures of the brain and stories of secret voter intention. The New York Times op-ed page is arguably the most influential two square feet of newsprint in American politics. The editorial column by Iacoboni and his colleagues stands as a testament not to the power of neuroimaging to make manifest our political preferences but to the manifest preference we all have for neuroimages.