Behavioural and neural evidence for self-reinforcing expectancy effects on pain

Abstract

Beliefs and expectations often persist despite evidence to the contrary. Here we examine two potential mechanisms underlying such 'self-reinforcing' expectancy effects in the pain domain: modulation of perception and biased learning. In two experiments, cues previously associated with symbolic representations of high or low temperatures preceded painful heat. We examined trial-to-trial dynamics in participants' expected pain, reported pain and brain activity. Subjective and neural pain responses assimilated towards cue-based expectations, and pain responses in turn predicted subsequent expectations, creating a positive dynamic feedback loop. Furthermore, we found evidence for a confirmation bias in learning: higher- and lower-than-expected pain triggered greater expectation updating for high- and low-pain cues, respectively. Individual differences in this bias were reflected in the updating of pain-anticipatory brain activity. Computational modelling provided converging evidence that expectations influence both perception and learning. Together, perceptual assimilation and biased learning promote self-reinforcing expectations, helping to explain why beliefs can be resistant to change.

Figure 1.

Experimental design and behavioural results. A. Cue-outcome pairings in the conditioning and test phase. Ten neutral-cue trials were included in Study 2 only. B. One test-phase trial in Study 2. Study 1 had slightly shorter interstimulus intervals. C. Pain rating as a function of stimulus temperature and cue type. Heat was applied to the inner forearm in Study 1 and to the, less sensitive, lower leg in Study 2, which explains the overall lower pain ratings in Study 2. Error bars indicate within-subject standard errors. Plots for Study 1 and Study 2 are based on data from 28 and 34 participants, respectively. D. Average expected (open circles) and experienced (filled circles) pain ratings as a function of cue type and trial. The difference between red and blue filled circles is the effect of cue type on pain ratings, which was robust in both studies, and did not disappear over time. The difference between red and blue open circles is the effect of cue type on pain expectations, which remained stronger than the effect on pain ratings throughout the test phase. Error bars indicate between-subject standard errors. Plots for Study 1 and Study 2 are based on data from 28 and 33 participants, respectively (one participant in Study 2 misunderstood the expected-pain rating procedure, and was excluded from all analyses and figures involving pain expectations).

Figure 2.

Cue effects on heat-evoked brain activity. A. Heat-evoked brain activity on high- vs. low-cue trials. Colored regions indicate stronger activity when heat was preceded by a high- than a low-pain cue. All colored regions were significant at q<0.05, FDR-corrected. For the purpose of display, we pruned the results using two additional, less conservative levels of voxel-wise threshold. B. Pain rating as a function of NPS response. We sorted each participant’s pain ratings into five bins according to their single-trial NPS response (both mean-centered), and plotted the group-average data for each bin. Error bars indicate within-subject standard errors. C. NPS response as a function of stimulus temperature and cue type. D. Across-subject correlation between the effects of cue type on pain rating and on the NPS response (r(32) = .46, p = .007, R² = .21, CI = .15 to .95). E. Average NPS response as a function of cue type and trial. Before plotting, the effect of temperature was regressed out and single-trial NPS responses were smoothed using locally weighted scatterplot smoothing. Vertical lines indicate the first trial of each scan run (heat was applied to a new skin site in each run). Note that the ten neutral-cue trials were evenly distributed amidst the low- and high-cue trials (one neutral-cue trial was presented during each series of 3 low- and 3 high-cue trials). Error bars indicate between-subject standard errors. All plots are based on data from 34 participants.

Figure 3.

Bidirectional effects of expectations and pain on one another. A. Dynamic feedback circuit through which expectations can become self-reinforcing. B. Pain rating as a function of expected pain and cue type. We sorted each participant’s pain ratings for the low- and high-cue trials into five bins, according to their trial-specific expected-pain ratings (both mean-centered), and plotted the group-average data for each bin. As there were fewer neutral-cue trials, we used three bins for the neutral-cue condition. Note that we binned the data for plotting purposes, but used single-trial measures in our statistical analyses. C. NPS response as a function of expected pain and cue type. D. Cue-based pain expectation as a function of cue type and the previous pain rating for that cue. E. Cue-based pain expectation as a function of cue type and the previous NPS response for that cue. All error bars indicate between-subject standard errors. All plots for Study 1 and 2 are based on data from 28 and 33 participants, respectively.

Figure 4.

Confirmation bias in expectation updating. A. Estimated learning rate as a function of prediction error sign and cue type in each study. PE = prediction error. The grey dots indicate individual participants, and the horizontal black lines are the mean values in each condition. The two individual data points at the top of the last condition (learning rates for aversive prediction errors on high-cue trials) in Study 2 fall outside the range of the figure; estimated learning rates for these points are 11 and 16. For Study 1, the first three conditions include data from 28 participants, and the last condition includes data from 26 participants. For Study 2, the first three plots include data from 33 participants, and the last plot includes data from 21 participants. Fewer participants contributed to the last condition because some participants never experienced aversive prediction errors on high-cue trials. B. Expectation updating as a function of signed prediction error magnitude and cue type in each study. Negative and positive prediction errors indicate lower- and higher-than-expected pain, respectively. Negative and positive expectation updates indicate decreases and increases in pain expectations, respectively. We sorted each participant’s low- and high-cue trials into five bins according to signed prediction error magnitude, and plotted the group-mean signed expectation updates for each bin. Lines show linear fits to unbinned single-trial data. Note that there was a significant main effect of cue type but no significant interaction between prediction error and cue type. Plots for Study 1 and 2 are based on data from 28 and 33 participants, respectively. Error bars indicate between-subject standard errors. C. Effect of cue type on expectation updating in each participant (first-level regression coefficients).

Figure 5.

Computational models capturing effects of cue-based expectations on pain and confirmation bias on expectation updating. A. Reinforcement learning model (Model 1). Perceptual inference within a trial combines expectations with noxious input to determine perceived pain. The γ parameter controls the relative impact of these two sources. Learning between trials involves updating the expectation for the current cue toward the current perceived pain. Experience-resistant expectations are modeled by assuming different learning rates, the α_c and a parameters, when the direction of prediction error is, respectively, consistent and inconsistent with the cue’s initial low or high pain association. B. Bayesian model (Model 2). Pain perception and expectation are products of Bayesian inference with respect to a generative model of the task environment that extends the classic Kalman filter. Note that arrows in this diagram indicate statistical dependencies in the subject’s generative model, not dynamics of the subject’s state of knowledge as in Figure 5a. Under the generative model, the mean threat level signaled by each cue (μ_c, index c suppressed in figure) drifts randomly from trial to trial, with step size determined by the σ_η parameter. The current threat level (or objectively correct pain level, π_t) on any trial t deviates randomly from μ_t, with standard deviation equal to the σ_ψ parameter. The noxious input (N_t) is a noisy indicator of π_t, with standard deviation equal to the σ_ε parameter. Inference within a trial (not shown) combines the current belief about with the observed value of N_t to estimate the current value of π_t; this estimate is the subject’s experienced level of pain (P_t). Inference across trials combines the beliefs about μ_t and π_t to estimate μ_t+1; this is the subject’s reported expectation the next time this cue is presented (E_t+1). Experience-resistant expectations are modeled by letting cue-pain associations drift in the directions of their initial values between trials, to a degree governed by the β parameter. Model 2 is formally nearly equivalent to Model 1, except that it assumes γ and α adapt from trial to trial to reflect the subject’s current level of uncertainty (Eqs. 3.1, 3.3, 3.6).

Figure 6.

Posterior distributions for the group-level means of the models’ parameters. A. Model 1: γ controls the impact of expectations vs. nociceptive input on pain (higher values cause a stronger weighting of expectations); α_c and α_i are learning rates for cue-consistent and cue-inconsistent prediction errors. The rightmost panels are joint density plots of ${\overline{\mathrm{\alpha }}}_c$ and ${\overline{\mathrm{\alpha }}}_i$ (dots are samples from the MCMC), showing that ${\overline{\mathrm{\alpha }}}_c$ is reliably greater than ${\overline{\mathrm{\alpha }}}_{\mathrm{i}}$. B. Model 2: β controls the drift of expectations toward (if β > 0) or away from (if β < 0) their initial values after each update; ${\mathrm{\sigma }}_{\mathrm{\Psi }}^2$ is the assumed variance in pain on any given trial; ${\mathrm{\sigma }}_{\mathrm{\eta }}^2$ is the assumed variance of the random walk process (random variation in pain across trials). ${\mathrm{\sigma }}_{\mathrm{\Psi }}^2$ and ${\mathrm{\sigma }}_{\mathrm{\eta }}^2$ are estimated relative to the variance of contributions to noxious input unrelated to pain (${\mathrm{\sigma }}_{\mathrm{\epsilon }}^2$, which was fixed at 1 to eliminate redundancy in model parameters).

Figure 7.

Confirmation bias in the updating of pain-anticipatory brain activity. A. We computed the change in anticipatory activity across successive trials in which the same cue was presented, separately for trials with higher- and lower-than-expected pain (aversive and appetitive prediction errors, respectively), and tested for effects of cue type. B. Individual differences in confirmation bias on estimated learning rate predict participants’ confirmation bias on anticipatory activity updating following aversive prediction errors. Yellow/red colors indicate positive across-subject correlations between increases in anticipatory activity following aversive prediction errors on high- vs. low-cue trials and estimated learning rate following aversive prediction errors on high- vs. low-cue trials. Blue colors indicate negative correlations. Cluster statistics can be found in Supplementary Table 2. Scatterplots illustrate the correlations in sensorimotor cortex and right striatum. The shading of the points is proportional to each observation’s weight in robust regression; white dots indicate outliers down-weighted by the algorithm. C. Individual differences in confirmation bias on estimated learning rate predict participants’ confirmation bias on anticipatory activity updating following appetitive prediction errors. Yellow/red colors indicate positive across-subject correlations between increases in anticipatory activity following appetitive prediction errors on low- vs. high-cue trials and estimated learning rate following appetitive prediction errors on low- vs. high-cue trials. Blue colors indicate negative correlations. All colored regions were significant at q < 0.05, FDR-corrected. For display purposes, we show the extent of results surrounding FDR-corrected peaks at p < .01 and p < .05 uncorrected. Cluster statistics can be found in Supplementary Table 3. Figures 7B and C are based on data from 21 participants.