Neuromarketing scans are very good at pinpointing exact points of reaction. Pradeep cited a study for an unnamed company that makes chips and salsa. They found that the moment a snacker lifts a salsa-covered chip to his mouth, before taking a bite, "that moment is extremely evocative for the brain. Your brain just goes nuts."

~ CNN Money

Neuroscience begot neuroeconomics, which begot neuromarketing. Neuroscience is the study of the brain and what it does. Neuroeconomics applies neuroscientific theories and methods to the study of economic behaviors such as buying, selling, trading, or investing. Neuromarketing seeks to use these advances with the goal to measure consumer preferences more deeply, more precisely, and more accurately than ever before. If they succeed, marketers will be able to tailor product design and delivery more effectively, presumably to the benefit of all.

Ariely and Berns (2010) provided a recent authoritative review of this emerging field. Their goal was “to distinguish the legitimate hopes from the marketing hype” (p. 284). Some issues can be both hope and hype, depending on whom you ask. Marketers may hope to gain access to hidden but true preferences that consumers themselves cannot or will not articulate. If so, marketing initiatives will face the stigma of manipulation (again) and ethical questions.

Although Ariely and Berns state that their first question is whether neuromarketing can reveal hidden preferences, they never provide an answer. This is reassuring. The results of neuroimaging need to be validated by other measures, which tend to be cognitive (self-reported preferences) or behavioral (revealed preferences). If what appears to be a preference at the neurological level does not agree with intended (e.g., “willingness to pay”) or actual consumption, we don't know if the neurological signal is a false positive or a truly hidden true preference that does not reveal itself anywhere else (and if it doesn’t, who cares?).

If preferences are psychological events of desiring and enjoying, patterns of activity in the medial orbitofrontal cortex (OFC), the prefrontal cortex (PFC), and elsewhere are their corresponding neural footprints. I explained in my previous post that knowing the former permits predicting the latter. This “forward inference” is captured by the conditional probability of a certain neural activity, A, given a certain mental event, M, or p(A|M). If there is a correlation between A and M, the so-called Bayes factor, p(A|M)/p(A|~M), is not 1. We learn that A occurs selectively when M (rather than non-M) is going on.

Ariely and Berns note that neuromarketers cannot make a living by predicting brain activity from preferences. They need to do the opposite. They need to predict preferences from neural activity, which means they need to make a reverse inference from A to M. As Poldrack (2006) before them, Ariely and Berns know that reverse inferences are problematic, but like Poldrack they hope that they are good enough.

Ariely table

To illustrate the potential of neuromarketing, they present data on activation (A vs. ~A) in the nucleus accumbens (one of the reward centers) from studies involving reward (R) and non-reward (~R) tasks. The table shows the four joint frequencies. The Bayes factor is healthy: p(A|R)/p(A|~R) = .40/.05 = 8 (or 9, depending on rounding). The reverse inference is also on solid ground: p(R|A)/p(R|~A) = .31/.03 = 10. Indeed, the reverse inference is stronger than the forward inference and this is so because of the low base rate of R (.05).

There is no inherent problem with reverse inference. The fallacy is to assume that p(R|A) is the same as p(A|R) without checking whether the base rates of R and A are the same. In Ariely and Berns’s data, they are almost identical, and so there is little difference between forward and reverse inference.

For their own reverse inference, however, Ariely and Berns ignore the base rate of R that they themselves had gleaned from the literature. Instead, they assume that p(R) = .5. (Why .5...?) Armed with this assumption, they find that the probability of R given A is .9. This, Ariely and Berns say amounts to “odds 9:1.” Presumably, this refers to the odds of rewards over no rewards. They continue that “this yields a Bayes factor of 9, which is considered moderate to strong evidence for a causal relationship” (p. 286).

Hold on a minute. If we are free to select any prior probability, we can generate any reverse inference at will. And besides, Bayes’s Theorem refers to statistical relationships, not causal relationships. If it did, we could conclude that activation in the nucleus accumbens caused participants to be in the conditions with rewarded-bearing tasks rather than in the control conditions.

The irony is that Ariely and Berns could have made a stronger reverse inference had they used the data at hand instead of an assumed base rate. What is critical in the end is the ratio of p(R|A) over p(R|~A), and this ratio is .31/.03 = 10 in the obtained data but .9/.39 = 2 if the base rate is set to the arbitrary .5.

In spite of these rather technical objections, a not fully acknowledged part of me shares the hopes of the neuromarketers. If things work out, entertainment and other consumer products will become available that we enjoy more than we know or want to admit. God bless the reverse inference.

Ariely, D., & Berns, G. S. (2010). Neuromarketing: the hope and hype of neuroimaging in business. Nature Reviews Neuroscience, 10, 284-292.

Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data? TRENDS in Cognitive Sciences, 10, 59-63.

You are reading

One Among Many

Bayes Odyssey

Where will the stat wars lead us?

The Kingdom of Bayes

The priors are low.

Interest of Conflict

Can you assert yourself without losing social capital?