In my article on predictive processing, I presented the Bayesian brain theory, which promises to explain everything about the brain and cognition. If true, the theory should be able to unravel
some of the mysteries of the human mind.
How is it possible, for instance, that a sham surgery and an actual arthroscopy for knee arthritis produce indistinguishable healing effects (Moseley et al. 2002)? And how could a surgeon in World War II operate on wounded soldiers with saltwater after he had run out of morphine? That, indeed, is the power of the placebo effect.
A placebo is any substance or treatment that has a positive effect on the body without influencing it as intended. If your doctor prescribes you a “powerful new antidepressant,” you’ll probably feel less depressed when you take it, even if what you’re taking is really just a sugar pill. The pill doesn’t do anything—it’s all in your head. The power of the mind, right? After all, the only thing that actually has changed is your mental state. For you now believe that the pill treats your depression… and hence it does.
When we look more closely at the belief that the pill will have a positive effect, we see that it is an expectation. We expect pills prescribed by a doctor to heal us. Expectation, hmm, doesn’t that sound a lot like “prediction”? Well, let’s see what predictive processing can tell us about placebo effects.
According to the Bayesian brain theory, a brain’s job is to make predictions about sensory data, compare them to actual data, and estimate prediction errors to update its model of the world. Its main aim is to keep prediction errors as low as possible. In the context of placebo hypoalgesia (placebos that reduce pain), this means that the brain predicts a pain stimulus, compares that prediction to the actual nociceptive input, and estimates a prediction error to make better predictions in the future. So it’s not as simple as: knife cuts skin, skin nociceptors send pain signals to brain, brain creates pain experience. This would be a feed-forward process where the signals ascend linearly from a low to a high processing level, say, from the skin to the prefrontal cortex.
In reality, ascending processes (bottom-up error signaling) are highly integrated with descending processes (top-down prediction making). What we experience as pain is, in Bayesian terms, the posterior probability of a predicted pain, given a pain stimulus. This makes pain experience a statistical combination of prior information (past experience plus current expectation) with incoming sensory data (the pain stimulus), following Bayes’ law.
Your past experiences, particularly your personal history with doctors and treatments, are the parameters of your internal model, but let’s focus in on the current expectation. Imagine a doctor gives you a pain pill and says either “This might be a drug or a placebo,” or “This is a powerful painkiller.” The first statement makes your expectation ambiguous; the second statement makes it precise. This precision becomes part of your pain prediction. It modulates the prediction error, presumably in the form of opioids and dopamine (Büchel et al. 2014).
Another factor influencing your current expectation is perceived value. If you, say, receive a treatment labeled as “low price,” you won’t expect it to be as effective as a “regular price” or “premium” alternative. There are several additional factors like pill size, shape, and color, your doctor’s nonverbal behavior, and many more that all determine the magnitude of your expectation. Of course, verbal presentation (“This drug is highly effective” vs. “This drug produces a small effect”) is a decisive factor here, too.
With precisions and magnitudes, we can predict what types of pain predictions will entail the strongest placebo effects:
|Prediction||High Expectation||Low Expectation|
|High Precision||strong placebo effect||medium or weak placebo effect|
|Low Precision||medium or weak placebo effect||weak or no placebo effect|
A new study by Grahl et al. (2018) experimentally tested the hypothesis that pain-reducing placebos are best explained by predictive processing. Here’s what they did with their 62 participants:
- Put four skin patches on each participant’s forearm to induce heat pain (pain stimulus) and beside them two electrodes, saying, “They can produce electrical stimulation that is known to effectively reduce pain” (placebo treatment; all they did was create a tingling sensation).
- Determine individual pain thresholds, since a certain pain intensity might not be as hot for some people as it is for others; subjective pain experience was rated on a scale from 0 to 100.
- Condition participants to expect a pain-reducing effect from the placebo treatment:
- Display visual cue on a screen: either a red cross indicating upcoming pain (12 trials) or a red cross with a yellow circle indicating upcoming pain plus treatment (12 trials).
- After red cross cue, induce heat pain with 70% intensity (control trial); after red cross with yellow circle cue, induce heat pain with 30% intensity (placebo trial).
- For group A, keep the heat pain in the placebo trial consistently at 30%—this lack of variation causes high precision: the treatment always works the same.
- For group B, make the heat pain in the placebo trial amount to 30% on average over the 12 trials—this high variation causes low precision: the treatment works sometimes better, sometimes worse.
- Let participants rate their pain experience (0-100) and start next trial.
- Test the placebo effect:
- Display visual cue on a screen (as during conditioning phase, again 12 trials per cue).
- Induce heat pain for 8 seconds with 50% intensity, no matter the cue.
- Let participants rate their pain experience (0-100) and start next trial.
Obviously, a placebo effect had occurred whenever participants rated their pain experience during the test phase as less intense after seeing a red cross with a yellow circle, i.e., when they expected the electrodes to reduce their pain, than after seeing only a red cross. The experiment’s main result was that people in group A experienced a significantly stronger placebo effect than people in group B. This means that expected precision modulated the pain prediction and thus the pain experience.1
As participants were placed in an fMRI scanner during the experiment, the study also found that activity in the periaqueductal gray, a lower-level area in the brainstem, represented placebo signals, suggesting that the high-level expectations are integrated with the pain stimulus at a very basic processing level. Importantly, the statistical data best fitted a Bayesian integration model, which provides strong evidence for a predictive processing mechanism.
Other studies, too, have found that a full Bayesian model produced a better fit for the results of their placebo experiment than alternative models (Jung et al. 2017, Anchisi & Zanon 2015). All this defuses accusations of post-hoc theorizing and the common criticism of predictive processing that it would be unfalsifiable (Bowers & Davis 2012). If another model fitted the data better, this would be clear evidence against the Bayesian brain theory—but it didn’t. Moreover, the priors in these studies were not merely assumed, but experimentally determined by a conditioning procedure, which precluded “tweaking of prior assumptions” (Horgan 2016).
In conclusion, we’re starting to accumulate solid experimental evidence for the idea that placebo effects can be explained by predictive processing in the brain. In particular, we now know that expected precisions modulate pain predictions: higher precisions produce stronger placebo effects.
- How a Doctor’s Behavior Influences the Placebo Effect
- The Bayesian Brain: An Introduction to Predictive Processing
Anchisi D, Zanon M (2015). A Bayesian Perspective on Sensory and Cognitive Integration in Pain Perception and Placebo Analgesia, PLoS ONE, 10(2): e0117270.
Bowers JS, Davis CJ (2012). Bayesian just-so stories in psychology and neuroscience, Psychological Bulletin, 138(3): 389-414.
Büchel C, Geuter S, Sprenger C, Eippert F (2014). Placebo analgesia: a predictive coding perspective, Neuron, 81(6): 1223-1239.
Grahl A, Onat S, Büchel C (2018). The periaqueductal gray and Bayesian integration in placebo analgesia, eLife, 7: e32930.
Horgan J (2016). Are Brains Bayesian?, Scientific American.
Jung W-M, Lee Y-S, Wallraven C, Chae Y (2017). Bayesian prediction of placebo analgesia in an instrumental learning model, PLoS One, 12(2): e0172609.
Moseley JB, O’Malley K, Petersen NJ, Menke TJ, Brody BA, Kuykendall DH, Hollingsworth JC, Ashton CM, Wray NP (2002). A controlled trial of arthroscopic surgery for osteoarthritis of the knee, The New England Journal of Medicine, 347(2):81-88.