How unrealistic optimism is maintained in the face of reality.

Unrealistic optimism is a pervasive human trait that influences domains ranging from personal relationships to politics and finance. How people maintain unrealistic optimism, despite frequently encountering information that challenges those biased beliefs, is unknown. We examined this question and found a marked asymmetry in belief updating. Participants updated their beliefs more in response to information that was better than expected than to information that was worse. This selectivity was mediated by a relative failure to code for errors that should reduce optimism. Distinct regions of the prefrontal cortex tracked estimation errors when those called for positive update, both in individuals who scored high and low on trait optimism. However, highly optimistic individuals exhibited reduced tracking of estimation errors that called for negative update in right inferior prefrontal gyrus. These findings indicate that optimism is tied to a selective update failure and diminished neural coding of undesirable information regarding the future.

People’s estimations of the future are often unrealistically optimistic[1-5]. A problem that has puzzled scientists for decades is why human optimism is so pervasive, when reality continuously confronts us with information that challenges these biased beliefs[5]. According to influential learning theories agents should adjust their expectations when faced with disconfirming information[6,7]. However, this normative account is challenged by findings showing that providing people with evidence that disconfirms their positive outlook often fails to engender realistic expectations. For example, highlighting previously unknown risk factors for diseases is surprisingly ineffective at altering peoples’ optimistic perception of their medical vulnerability[8,9]. Even experts show worrying optimistic biases. For instance, financial analysts expect improbably high profits[3] and family-law attorneys underestimate the negative consequences of divorce[2].

The wider societal importance of these errors derives from the fact that they reduce precautionary actions, such as practicing safe sex or saving for retirement[8,9]. On the upside, optimistic expectations can lower stress and anxiety, thereby promoting health and well-being[10-12]. While the existence of unrealistic optimism has been extensively documented[1-5, for review 13&14], the biological and computational principles that help to maintain optimistically biased predictions in the face of reality are unknown[5]. Importantly, such biases are not explained by theories assuming equal learning across outcome valence[6,7].

There is now a rich literature indicating that learning from information that disconfirms one’s expectations is mediated by regions involved in error processing and conflict detection, including anterior cingulate cortex, adjacent medio-frontal regions and lateral prefrontal regions[15-20]. Here, we tested whether a failure to alter optimistic predictions when presented with disconfirming data was mediated by differential error processing in response to information that was better, or worse, than expected.

We combined a novel learning paradigm with functional brain imaging. Our task enabled us to quantify how much participants adjusted their beliefs in response to new information. While optimism is defined both as overestimation of positive future events and underestimation of future negative events[1], we focused here on the latter as this is strongly related to a concern that people do not take necessary action to protect themselves against hazards[9]. We obtained fMRI (functional Magnetic Resonance Imaging) data while participants estimated their likelihood of experiencing 80 different types of adverse life events (e.g. Alzheimer’s disease, robbery etc. – see Supplementary List of Stimuli). After each trial, participants were presented with the average probability of that event occurring to an individual living in the same socio-cultural environment as the participants (Fig. 1a – paradigm). We then assessed whether participants used this information to update their predictions by subsequently asking them in a second session to provide estimates of their likelihood of encountering all 80 events.

Figure 1

Task design. (a) On each trial participants were presented with a short description of one of 80 adverse life events and asked to estimate how likely this event was to occur to them. They were then presented with the average probability of that event occurring to a person like themselves, living in the same socio-cultural environment. For each event an estimation error term was calculated as the difference between the participant’s estimation and the information provided. The second session was the same as the first session. For each event an update term was calculated as the difference between the participant’s first and second estimations. (b,c) Examples of trials for which the participant’s estimate was (b) higher or (c) lower than the average probability. Here, for illustration purposes, the blue and red frames denote the participant’s response (either an overestimation or underestimation, respectively). The blue and red filled boxes denote information that calls for an adjustment in an (b) optimistic (desirable) or (c) pessimistic (undesirable) direction.

RESULTS

Behaviour: selective updating

Our behavioural results revealed a striking asymmetry in belief updating. Participants learned to a greater extent from information offering an opportunity to adopt more optimistic expectations (hereafter “desirable information” Fig. 1b) than from information that challenged participants’ rosy outlook (hereafter “undesirable information” Fig. 1c). Specifically, participants were more likely to update their beliefs when the average probability of experiencing a negative life event was lower than the participants’ own probability estimate (mean absolute update from first to second session = 11.2), relative to a situation where the average probability was higher (update = 7.7, t (18) = 4.4, P < 0.001; Supplementary Fig. 1a). This robust effect was observed in 79% of the participants (Fig. 2a) and replicated across additional behavioural studies (Supplementary Behavioural Studies I & II). The difference was also significant when updates were calculated as the percentages of the initial estimate, (t (18) = 4.7, P < 0.001; Supplementary Fig. 1b).

Figure 2

Behaviorally observed bias. (a) After receiving (desirable) information that presented an opportunity to adopt a more optimistic outlook, participants updated their estimations to a greater extent than after receiving (undesirable) information that called for a more pessimistic estimate. This asymmetric update was observed in 15 out of 19 participants. For group means see Supplementary Fig. 1a. (b) Betas indicating the association between update and estimation errors on an individual basis showed that estimation errors predicted update to a greater extent when participants received desirable information than when they received undesirable information. This asymmetry is observed in all 19 participants. For group means see Supplementary Fig. 1d.

One obvious explanation for this pattern of findings is that a greater weighting given to desirable information simply reflects differential memory for desirable compared to undesirable information. This was not the case. After the scanning session, participants were asked to indicate the actual probability (as previously presented) of each event occurring to an average person in the same socio-cultural environment. Memory errors were calculated as the absolute difference between the actual probability previously presented and the participants’ recollection of that statistical number. Participants remembered information presented to them equally well irrespective of whether it was desirable or undesirable (t (18) = 0.75, P > 0.4; Supplementary Fig. 1c, also Supplementary Behavioural Studies I & II).

Furthermore, post-scanning questionnaire scores showed that any differential updating across valence was not explained by differences in emotional arousal, extent of negative valence, familiarity, or past experience with the adverse life event (Supplementary Results). Specifically, the difference in absolute update for events for which participants received desirable and undesirable information remained significant after entering all these scores as covariates (F(1, 13) = 9.7, P < 0.01). Thus, differential update could not be explained by differences in the degree of adversity of the events or familiarity/ past experience with the events.

Importantly, positively biased updating was not due to differences in the underlying true probabilities (base rates) of the events. The difference in update for events in which participants received desirable and undesirable information remained significant even after entering the true probabilities of the events as covariates (F (1,17) = 6.04, P < 0.05). In other words, how common or rare the occurrence of an event is had no bearing on selective updating. Neither were there differences in the number of trials (t (18) = 0.02, P > 0.9), the magnitude of the initial estimation errors (t (18) = 1.85, P > 0.05), or reaction times (t (18) = 1.04, P > 0.3) when participants received desirable or undesirable information. Finally, asymmetric updating was not explained by a differential processing of high and low percentages, as this was controlled for by asking participants to estimate their likelihood of encountering the adverse event on half of the trials, and to estimate their likelihood of not encountering the adverse event on the other half of the trials (see Supplementary Results).

Formal models suggest that learning is mediated by a prediction error signal that quantifies a difference between expectation and outcome[7,21,22]. We hypothesized that an analogous mechanism underpinned belief update in our task. We formulated the difference between participants’ initial estimations and the information provided in terms of estimation errors (i.e. estimation error = estimation − probability presented). Indeed, estimation errors predicted subsequent updates on an individual level (mean beta from individual linear regressions relating estimation errors to update = 0.53, P < 0.0005). However, the strength of this association was valence-dependent, being greater for information that offered an opportunity to adopt a more optimistic outlook (beta = 0.72), relative to information that called for a more pessimistic outlook (beta = 0.33, t (18) = 5.8, P < 0.0005; Fig. 2b and Supplementary Fig. 1d). Importantly, this difference remained significant after controlling for the true underlying probabilities of events by including them as a factor to the regression analysis (t (18) = 2.6, P < 0.02).

These behavioural findings suggest a likely computational principle that mediates unrealistic optimism. Specifically, they point to estimation errors as providing a learning signal whose impact depends on whether new information calls for an update in an optimistic or pessimistic direction. Consequently, we next examined our fMRI data to identify how BOLD signal tracks estimation errors in response to information that entails either a belief adjustment in an optimistic or in a pessimistic direction.

fMRI: tracking of estimation errors

We first identified voxels where activity tracked desirable or undesirable estimation errors. Trials were partitioned based on whether participants received desirable or undesirable information regarding the future after their first rating. We entered the absolute estimation error on each trial as a parametric regressor modulating the time point when participants were presented with information regarding the average probability of events (see Fig. 1). From this analysis we identified regions where BOLD signal correlated with estimation errors for either desirable or undesirable information on a trial by trial basis (P < 0.05, cluster level corrected after voxel-wise thresholding at P < 0.001).

BOLD signal correlated positively with desirable estimation errors in three distinct regions; left inferior frontal gyrus (IFG: peak voxels in Talairach coordinates: −58, 21, −1; k = 274; Fig. 3a), left and right medial frontal cortex / superior frontal gyrus (MFC/SFG: −10, 62, 28; k = 412; Fig. 3b) and right cerebellum (33,−79,−28; k = 185). In addition, BOLD signal correlated negatively with undesirable estimation errors in the right inferior frontal gyrus extending into precentral gyrus and insula (46, 12, 9 & 60, 10, 9; k = 190; Fig. 3c). There were no voxels where activity correlated negatively with desirable errors or positively with undesirable errors. In other words, brain activity increased when the average probability was more desirable than participants’ estimates (in left IFG, bilateral MFC/SFG and right cerebellum) and decreased when it was more undesirable (in right IFG). All of these regions have been previously implicated in error processing in a variety of tasks[15-20].

Figure 3

Brain activity tracking estimation errors. Regions where BOLD signal tracked participants’ estimation errors on a trial by trial basis in response to desirable information regarding future likelihoods included (a) the left IFG and (b) bilateral MFC/SFG; (P < 0.05, cluster level corrected). (c) BOLD signal tracking participants’ estimation errors in response to undesirable information was found in the right IFG/precentral gyrus (P < 0.05, cluster level corrected). (d) Parameter estimates of the parametric regressors in both the left IFG and bilateral MFC/SFG did not differ between individuals who scored high or low on trait optimism. In contrast, in the right IFG a stronger correlation between BOLD activity and undesirable errors was found for individuals who scored low on trait optimism relative to those who scored high. Error bars (s.e.m.). * = P < 0.05, two-tailed independent sample t-test.

As our behavioural findings tied estimation errors to update, it follows that activity that tracks estimation errors should also predict subsequent update. To test this we entered the level of subsequent update as a parametric regressor modulating the time point when participants were first presented with information regarding the average probability of adverse events. We then retrieved the parameter estimates for each participant and each condition in the four functional ROIs (regions of interest) identified above, averaged across all voxels. Indeed, activity in the right IFG predicted update in response to undesirable information (parameter estimates significantly greater than zero; t (18) = 2.2, P < 0.05). Activity in MFC/SFG (t (18) = 2.1, P < 0.05) and right cerebellum (t (18) = 2.4, P < 0.05) predicted update in response to desirable information (in the left IFG parameter estimates were not significantly different from zero).

Trait optimism, biased updating, and BOLD signal

Next, we examined whether the extent to which brain activity tracked estimation errors was related to optimism. We compared betas from the parametric analysis (relating BOLD signal to estimation errors) in each functional ROI (averaged across all voxels) of participants who scored high on trait optimism with those who scored low (according to the Life Orientation Test-Revised scale: LOT-R[23]). There was no difference between high and low optimists on betas tracking desirable errors in the left IFG (Fig. 3d), bilateral MFC/SFG (Fig. 3d) or cerebellum. Importantly, however, participants who scored high on trait optimism showed a weaker correlation of right IFG activity with undesirable errors relative to those who scored low (Fig. 3d). Specifically, betas indicating a correlation between BOLD signal and undesirable estimation errors in the right IFG showed a weaker inverse relation in high compared to low optimists (t (17) = 2.1, P < 0.05). This difference could not be explained by the magnitude of the initial undesirable estimation errors (P > 0.8) or memory for undesirable information (P > 0.5), as neither differed between groups.

The same finding was obtained when we considered all 4 ROIs together in a single linear regression with group membership (high or low optimism) as the dependent measure, controlling for differential scores (desirable minus undesirable) of memory, familiarity, past experience, vividness and arousal, and extent of initial estimation errors. A stepwise procedure revealed that the best model predicting trait optimism was one that included only betas from the right IFG tracking undesirable estimation errors (beta = 0.46, P < 0.05). This result suggests a specific relationship between trait optimism and a reduced neural coding of undesirable errors regarding the future.

To examine this further we determined whether the optimistic update bias in our task related to how well BOLD activity tracked undesirable errors. Across individuals BOLD signal tracking undesirable estimation errors predicted the observed update bias (desirable update minus undesirable update) in right IFG (partial correlation controlling for behavioural association between estimation errors and update r12.3 = 0.45, P = 0.05). Participants with the largest optimistic update bias failed to track undesirable errors (the correlation between BOLD signal in the right IFG and undesirable errors in these participants was close to zero; Fig. 4a). Participants who did not show selective updating had the strongest relationship between BOLD signal and undesirable estimation errors. Thus, these results suggest that an asymmetry in belief updating is driven by a reduced expression of an error signal in a region implicated in processing undesirable errors regarding the future.

Figure 4

Optimism and brain activity tracking undesirable estimation errors. (a) Across individuals, the extent of the update bias (difference in update in response to desirable errors minus undesirable errors) correlated with how strong the peak voxel in the right IFG ROI correlated with undesirable errors (r12 = 0.45). Participants showing the greatest optimistic bias in updating showed the weakest tracking of undesirable estimation errors. (d) Furthermore, participants who updated their estimates more in response to undesirable information showed a greater reduction in activity in the right IFG ROI the second time desirable information was presented, relative to the first (r = −0.47).

A final prediction arising from our data is that those participants, who updated their response the most in response to undesirable information, should show greater BOLD reduction the second time undesirable information is presented. The logic here is that a greater update should correspond to smaller estimation errors in response to repeated presentation of information and thus to less overall activity in regions tracking estimation errors to repeated information. To test this we examined the correlation, across participants, between suppression of activity in the right IFG (BOLD signal at second presentation of undesirable information minus BOLD signal at first presentation of undesirable information, averaged across all voxels in the ROI) and degree of update on undesirable trials. We found that participants who showed greater update in response to undesirable information had a greater reduction in activity at the second time-point when undesirable information was presented (r = −0.47, P < 0.05; Fig. 4b). For a description of differential activation during the period when participants thought about encountering the adverse events see Supplementary Table 1.

DISCUSSION

Our findings offer a mechanistic account of how unrealistic optimism persists in the face of challenging information. We showed that optimism was related to diminished coding of undesirable information about the future, in a region of the frontal cortex (right IFG) identified as sensitive to negative estimation errors. Participants scoring high on trait optimism were worse at tracking undesirable errors in this region, relative to those who scored low. In contrast, tracking of desirable information in regions processing desirable estimation errors (MFC/SFG, left IFG and cerebellum) did not differ between high and low optimists.

Reduced BOLD tracking of undesirable errors in right IFG predicted the extent to which participants selectively updated their beliefs using information that enforced optimism, while (relatively) dismissing information that contradicted it. Importantly, this effect was not explained by how well participants recalled the information presented to them. Neither did it reflect specific characteristics of the adverse events (e.g. familiarity, negativity, arousal, past experience, how rare or common the event is). Thus, unlike predictions from learning theory, where both positive and negative information are given equal weight[6,7], we showed a valence-dependent asymmetry in how estimation errors impacted on beliefs about one’s personal future.

Interestingly, our fMRI data revealed that error evoked activity differed in response to desirable and undesirable information regarding possible future outcomes. Segregated regions encoded error-related-activity in response to new information that called for optimistic or pessimistic adjustments. While the left IFG, left and right MFC/SFG and right cerebellum tracked desirable errors, the right IFG tracked undesirable errors. It is worth noting that previous studies have suggested hemispheric asymmetry in processing positive and negative information consistent with that found here in the IFG [24]. Furthermore, while BOLD signal in regions tracking positive estimation errors increased when the average probability was better than the participant’s estimate, in regions tracking negative estimation errors it decreased when the average probability was worse than expected. In other words, activity increased for a better than expected outcome and dipped for a worse than expected outcome – a pattern resembling that of neurons signalling prediction errors [21-22].

All regions identified in this study as coding estimation errors have previously been shown to track errors in different contexts, including errors due to incorrect responses[15], errors in expectations in the absence of action[16], reversal errors[17,18] and prediction errors that code differences between expectations and outcomes[19]. Note that those errors emerge when outcomes are experienced, and capture differences in expected and real magnitudes. Here, estimation errors captured the difference between participants’ predictions of the likelihood of a possible future outcome and information about the actual probability of experiencing these outcomes. This information was presented in an explicit form that engaged higher cognitive functions, and this is likely to explain the preferential engagement of cortical regions. Our findings suggest that these error signals are subject to motivational modulation (see also[20]). Specifically, a motivation to adopt the most rewarding (or least aversive) perspective on future outcomes is likely to modulate the impact of an error signal that subsequently influences update. As information regarding hypothetical outcomes is less constrained than actual experiences, the impact of such information may be more easily altered by motivation. It is possible that reduced coding of negative errors is restricted to such cases, and we do not know whether it will generalise to instances where outcomes are in fact experienced.

We previously described a positivity bias in the imagination of future life events, where participants imagined positive future events as closer in time and more vivid than negative events – a bias mediated by activity in rostral anterior cingulate cortex (ACC) and amygdala[25]. Whereas participants in our previous study showed an optimistic bias in unconstrained imagination, here we identified an optimistic learning bias when participants’ beliefs were challenged by new information. These results provide a powerful explanatory framework for how optimistic biases are maintained.

Underestimating susceptibility to negative events can serve an adaptive function by enhancing explorative behaviour and reducing stress and anxiety associated with negative expectations[11,12,26]. This is consistent with the observation that mild depression is related to a more accurate estimation of future outcomes, and severe depression to pessimistic expectations[27]. However, any advantage arising out of unrealistic optimism is likely to come at a cost. For example, an unrealistic assessment of financial risk is widely seen as contributing factor to the 2008 global economic collapse[28,29]. The current study suggests that this human propensity towards optimism is facilitated by the brain’s failure to code errors in estimation when those call for negative updates. This failure results in selective updating, which supports unrealistic optimism that is resistant to change. Dismissing undesirable errors in estimation renders us peculiarly susceptible to view the future through rose coloured glasses.