Bipolar disorder is estimated to affect approximately 3% of the general population (American Psychiatric Association, 2013). The disorder is characterized by successive periods of elation (mania) and depression interspersed with periods of euthymia, an asymptomatic phase in which patients are in clinical remission (Grande et al., 2016). Shorter periods of euthymia between relapses are associated with poorer functioning, increased odds of suicidality, disability, unemployment, and hospitalization (Peters et al., 2016). A major research goal is thus to identify predictors of upcoming relapse in order to facilitate timely treatment (Harrison et al., 2016). Yet, predicting relapse using existing clinical diagnostic tools or demographic information has proven largely ineffective in bipolar disorder (Bukh et al., 2016; de Dios et al., 2012; Vieta et al., 2018).

Here, we test the hypothesis that prior to the clinical manifestation of a relapse patients exhibit a specific learning pattern from self-relevant information, which makes them more vulnerable to relapse. In particular, it has been shown that healthy individuals tend to update their beliefs to a greater extent in response to unexpected positive information (e.g. learning that the likelihood of receiving a job offer is greater than expected) than negative information (e.g. learning it is lower than expected; Kappes et al., 2020; Kuzmanovic et al., 2016; Ma et al., 2016; Moutsiana et al., 2015; Sharot et al., 2011). Such asymmetry in valence-dependent belief updating leads to optimistic beliefs (Sharot et al., 2011). This is because a person who takes in positive information to a greater extent than negative information from their experiences in the world, uses such information to update their beliefs, which eventually leads to more positive expectations.

In contrast, individuals with depression do not show a positivity bias in learning (Garrett et al., 2014; Korn et al., 2014). These patients update their beliefs less in response to positive information (Korn et al., 2014) and more in response to negative information (Garrett et al., 2014), leading to pessimistic views that may exacerbate symptoms.

However, it is unknown whether a lack of positivity bias in learning precedes symptoms of affective disorders. If it does, then measuring learning biases could facilitate early diagnosis. Indeed, there is theoretical basis to believe that reduced learning from positive information and/or enhanced learning from negative information will occur before a relapse. In particular, a reduction in learning from positive information (or enhancement in response to negative information) will cause individuals to take in less desirable information from the world around them (or more undesirable information), which over time will tip beliefs toward a pessimistic direction. Once a patient generates a negative set of beliefs, deleterious symptoms may appear such as negative mood, sleep disturbance, suicidal thoughts etc.

Testing this possibility on Bipolar patients is especially compelling for three reasons. First, bipolar patients tend to relapse frequently, with a mean distance between episodes of 18 months on average (Radua et al., 2017). If we were to examine how patients update their beliefs during euthymia to predict how soon they would relapse, within only a few years most patients would have likely relapsed, which would allow us to test the reliability of those predictions. Second, whilst traditional diagnostic tools that measure symptom severity has proven useful in predicting relapse in unipolar depression (Dinga et al., 2018), these are largely ineffective in bipolar disorder (Bukh et al., 2016; de Dios et al., 2012). Third, by examining bipolar patients we could test whether valence-dependent belief updating predicts episodes of both polarities (that is mania and depression) differently or in the same manner. On one hand, the one-dimensional model suggests that mania and depression lie on opposite ends of the same mood spectrum. Hence manic or depressive symptoms arise only in the absence of the other (Bonsall et al., 2015; Bonsall et al., 2012; Chang and Chou, 2018; Daugherty et al., 2009; Eldar et al., 2018; Eldar et al., 2016; Eldar and Niv, 2015; Mason et al., 2017; Nana, 2009; Steinacher and Wright, 2013). We may thus expect that positivity bias in belief updating would predict manic episodes and vice versa for depressive episodes. Other models, however, conceptualize mania and depression as independent dimensions (Bystritsky et al., 2012; Cochran et al., 2018; Goldbeter, 2013; Goldbeter, 2011; Hadaeghi et al., 2015; Huber et al., 2000; Lopez, 2008). According to these models, perturbations can trigger episodes of either polarity. The degree to which patients update beliefs in response to positive and negative information may reflect such perturbations (or perhaps even cause them) predicting relapse of both polarities similarly.

To test whether a valence-dependent bias in belief updating, or the lack thereof, precedes relapse of bipolar episodes, we tested patients with bipolar disorder on the belief update task (Sharot et al., 2011) while they were in the euthymic stage. The patients were then followed up for 5 years and their time of relapse was recorded.

Our findings show that the extent to which bipolar patients updated their beliefs in response to positive information, relative to negative information, was predictive of when they would relapse. In particular, euthymic patients who updated their beliefs to a greater extent in response to positive information were more likely to remain in euthymia for a longer duration. In other words, biased processing of information in a manner that supports positive outlook was linked to a more favorable course of bipolar illness.

These results are important from a clinical perspective due to the difficulty in predicting the course of bipolar disorder (Vieta et al., 2018). Maintenance treatment in bipolar disorder relies on mood stabilizers and psychological programs aimed at increasing treatment compliance, educating the patients and identifying prodromal sub-syndromic symptoms, which provide clues for an upcoming episode (Oud et al., 2016; Scott et al., 2007). Despite their efficacy, these approaches are highly dependent on the physician’s ability to identify and promptly treat symptoms as they arise, to reduce the likelihood of full-blown episodes. Being able to predict a relapse could inform patients and clinicians to step up vigilance to recognize prodromal symptoms and intervene where necessary. The way patients update beliefs could be introduced in the future as a risk prediction tool for bipolar disorder.

Importantly, our results demonstrate that the absence of an optimistic update bias precedes the clinical manifestation of relapse. This finding is important for understanding the relationship between valence-dependent learning and mood. While a relationship between mood and valence-dependent learning has been previously demonstrated in depressed patients (Garrett et al., 2014; Korn et al., 2014), it has been unclear whether a change in mood likely trigger changes in learning or whether changes in learning likely trigger changes in mood. While our data is correlational, the temporal order of events in our study support the latter possibility. That is, we speculate that a reduction in learning in response to positive information relative to negative information over time, lead to a less optimistic perspective, which eventually provides a fruitful ground for clinical affective symptoms to manifest.

It is interesting that a reduction of a positivity bias in belief updating was predictive of both depressive and manic episodes. The results are consistent with multistable models that suggest that environmental or inner perturbation such as stressors could trigger either episode onset (Cochran et al., 2018; Cochran et al., 2017). The update bias may be an indicator of the existence of such stressors and their associated prodromal symptoms. While mood symptoms of full-blown episode are different for depression and mania, both polarities are preceded by many of the same subsyndromal a-specific symptoms. These symptoms, known as prodromes, represent the main therapeutic target when treating recurrences (Vieta et al., 2018) and include sleep disturbances, irritability, and mood instability (Andrade-González et al., 2020). It is thus possible that a pessimistic view triggers these common subsyndromal symptoms.

We emphasize that our task carefully controls for an individuals perception of how likely they are to experience the aversive events by including their first estimate of these likelihoods in the model. We also control for how often participants had experienced these events in the past and how familiar they are with the events from others experiencing them. Thus, it is unlikely that the relationship between the update bias and future time in euthymia can be explained by some measure of ‘disadvantage setting’ of the patients. Indeed, the association between the update bias and time in euthymia was not explained by differences in memory, attention, experience with the negative life events used as stimuli, the perceived negativity of those stimuli, familiarity with the stimuli, or patients’ initial predictions of how likely they were to experience the events in the future, as all those were controlled for in our analysis. The association remained strong also when clinical measures, demographics, and other questionnaires were controlled for, such as number of past episodes and trait optimism. Interestingly, no clinical measures were able to predict future time in euthymia, except for antidepressant medication (Hooshmand et al., 2019). Other medication, such as mood stabilizers, did not predict time in euthymia. However, because most of the participants (83.3%) were on mood stabilizers, it is impossible to reliably determine if indeed such medication is associated with time in euthymia and whether taking these medications affects the relationship between task performance and time in euthymia. We also note that the sample size, although in line with the a-priori power analysis, is relatively small for assessing those models with a large number of variables and future studies with a larger sample will be helpful to verify the robustness of the current results.

The fact that general measures of cognitive ability such as memory errors and overall update did not predict time in euthymia suggests that the effect is valence-specific and cannot be explained by a general cognitive impairment, as this would have been better captured by overall update rather than the difference in update between good news and bad news trials. However, it is possible that other valence-dependent measures of learning, such as ones that can be quantified in a reinforcement learning task (Lefebvre et al., 2017), may be good predictors of future time of euthymia. This remains to be tested in future studies.

In sum, we found that greater belief updating in response to positive information was predictive of how long euthymic bipolar patients will remain in euthymia. This finding could be informative for developing tools for patient assessment. Future studies should examine whether the same measure is predictive of the reoccurrence of other clinical disorders, most notably depression. Moreover, it would be informative to test at-risk patients to examine if the update bias can be used as a tool for early diagnosis of affective disorders.

All data generated and analysed during this study are included in the manuscript and supporting files. Specifically the complete dataset is available on Mendeley Data.

1. 1. Ossola P

   (2020) Mendeley Data

   Belief Updating in Euthymic Bipolar Disorder.

   https://doi.org/10.17632/crp8k3dgj5.2

1. 1. Altamura AC
   2. Buoli M
   3. Caldiroli A
   4. Caron L
   5. Cumerlato Melter C
   6. Dobrea C
   7. Cigliobianco M
   8. Zanelli Quarantini F

   (2015) Misdiagnosis, duration of untreated illness (DUI) and outcome in bipolar patients with psychotic symptoms: a naturalistic study

   Journal of Affective Disorders 182:70–75.

   https://doi.org/10.1016/j.jad.2015.04.024

   • PubMed
   • Google Scholar
2. Report

   1. American Psychiatric Association

   (2013)

   Diagnostic and Statistical Manual of Mental Disorders (DSM-5)

   Washington, DC, United States: American Psychiatric Association.

   • Google Scholar
3. 1. Andrade-González N
   2. Álvarez-Cadenas L
   3. Saiz-Ruiz J
   4. Lahera G

   (2020) Initial and relapse prodromes in adult patients with episodes of bipolar disorder: a systematic review

   European Psychiatry 63:e12.

   https://doi.org/10.1192/j.eurpsy.2019.18

   • PubMed
   • Google Scholar
4. 1. Beck AT
   2. Steer RA
   3. Brown GK

   (1996) Beck depression inventory-II

   APA PsycTests 78:490–498.

   https://doi.org/10.1037/t00742-000

   • Google Scholar
5. 1. Bonsall MB
   2. Wallace-Hadrill SMA
   3. Geddes JR
   4. Goodwin GM
   5. Holmes EA

   (2012) Nonlinear time-series approaches in characterizing mood stability and mood instability in bipolar disorder

   Proceedings of the Royal Society B: Biological Sciences 279:916–924.

   https://doi.org/10.1098/rspb.2011.1246

   • Google Scholar
6. 1. Bonsall MB
   2. Geddes JR
   3. Goodwin GM
   4. Holmes EA

   (2015) Bipolar disorder dynamics: affective instabilities, relaxation oscillations and noise

   Journal of the Royal Society Interface 12:20150670.

   https://doi.org/10.1098/rsif.2015.0670

   • Google Scholar
7. 1. Bukh JD
   2. Andersen PK
   3. Kessing LV

   (2016) Rates and predictors of remission, recurrence and conversion to bipolar disorder after the first lifetime episode of depression--a prospective 5-year follow-up study

   Psychological Medicine 46:1151–1161.

   https://doi.org/10.1017/S0033291715002676

   • PubMed
   • Google Scholar
8. 1. Bystritsky A
   2. Nierenberg AA
   3. Feusner JD
   4. Rabinovich M

   (2012) Computational non-linear dynamical psychiatry: a new methodological paradigm for diagnosis and course of illness

   Journal of Psychiatric Research 46:428–435.

   https://doi.org/10.1016/j.jpsychires.2011.10.013

   • PubMed
   • Google Scholar
9. 1. Chang S-S
   2. Chou T

   (2018) A dynamical bifurcation model of bipolar disorder based on learned expectation and asymmetry in mood sensitivity

   Computational Psychiatry 2:205–222.

   https://doi.org/10.1162/cpsy_a_00021

   • Google Scholar
10. 1. Chowdhury R
    2. Sharot T
    3. Wolfe T
    4. Düzel E
    5. Dolan RJ

    (2014) Optimistic update Bias increases in older age

    Psychological Medicine 44:2003–2012.

    https://doi.org/10.1017/S0033291713002602

    • PubMed
    • Google Scholar
11. Book

    1. Cochran AL
    2. Schultz A
    3. McInnis MG
    4. Forger DB

    (2017) A Comparison of Mathematical Models of Mood in Bipolar Disorder, In: Computational Neurology and Psychiatry

    Springer.

    https://doi.org/10.1007/978-3-319-49959-8_11

    • Google Scholar
12. 1. Cochran AL
    2. Schultz A
    3. McInnis MG
    4. Forger DB

    (2018) Testing frameworks for personalizing bipolar disorder

    Translational Psychiatry 8:4.

    https://doi.org/10.1038/s41398-017-0084-4

    • PubMed
    • Google Scholar
13. 1. Daugherty D
    2. Roque-Urrea T
    3. Urrea-Roque J
    4. Troyer J
    5. Wirkus S
    6. Porter MA

    (2009) Mathematical models of bipolar disorder

    Communications in Nonlinear Science and Numerical Simulation 14:2897–2908.

    https://doi.org/10.1016/j.cnsns.2008.10.027

    • Google Scholar
14. 1. de Dios C
    2. González-Pinto A
    3. Montes JM
    4. Goikolea JM
    5. Saiz-Ruiz J
    6. Prieto E
    7. Vieta E

    (2012) Predictors of recurrence in bipolar disorders in Spain (PREBIS study data)

    Journal of Affective Disorders 141:406–414.

    https://doi.org/10.1016/j.jad.2012.03.009

    • PubMed
    • Google Scholar
15. 1. Dinga R
    2. Marquand AF
    3. Veltman DJ
    4. Beekman ATF
    5. Schoevers RA
    6. van Hemert AM
    7. Penninx B
    8. Schmaal L

    (2018) Predicting the naturalistic course of depression from a wide range of clinical, psychological, and biological data: a machine learning approach

    Translational Psychiatry 8:241.

    https://doi.org/10.1038/s41398-018-0289-1

    • PubMed
    • Google Scholar
16. 1. Eldar E
    2. Rutledge RB
    3. Dolan RJ
    4. Niv Y

    (2016) Mood as representation of momentum

    Trends in Cognitive Sciences 20:15–24.

    https://doi.org/10.1016/j.tics.2015.07.010

    • PubMed
    • Google Scholar
17. 1. Eldar E
    2. Roth C
    3. Dayan P
    4. Dolan RJ

    (2018) Decodability of reward learning signals predicts mood fluctuations

    Current Biology 28:1433–1439.

    https://doi.org/10.1016/j.cub.2018.03.038

    • PubMed
    • Google Scholar
18. 1. Eldar E
    2. Niv Y

    (2015) Interaction between emotional state and learning underlies mood instability

    Nature Communications 6:6149.

    https://doi.org/10.1038/ncomms7149

    • PubMed
    • Google Scholar
19. 1. Faul F
    2. Erdfelder E
    3. Lang AG
    4. Buchner A

    (2007) G*power 3: a flexible statistical power analysis program for the social, behavioral, and biomedical sciences

    Behavior Research Methods 39:175–191.

    https://doi.org/10.3758/BF03193146

    • PubMed
    • Google Scholar
20. Report

    1. First MB
    2. Spitzer RL
    3. Gibbon M
    4. Williams JBW

    (2002)

    Structured Clinical Interview for DSM-IV-TR Axis I Disorders, Patient Edition (SCID-I/P), New York: Biometrics Research

    Psychiatric Institute.

    • Google Scholar
21. Book

    1. Fletcher D

    (2018) Why Model Averaging?, In: Model Averaging

    Springer.

    https://doi.org/10.1007/978-3-662-58541-2_1

    • Google Scholar
22. 1. Folstein MF
    2. Folstein SE
    3. McHugh PR

    (1975) "Mini-mental state". A practical method for grading the cognitive state of patients for the clinician

    Journal of Psychiatric Research 12:189–198.

    https://doi.org/10.1016/0022-3956(75)90026-6

    • PubMed
    • Google Scholar
23. 1. Freckleton RP

    (2011) Dealing with collinearity in behavioural and ecological data: model averaging and the problems of measurement error

    Behavioral Ecology and Sociobiology 65:91–101.

    https://doi.org/10.1007/s00265-010-1045-6

    • Google Scholar
24. 1. Garrett N
    2. Sharot T
    3. Faulkner P
    4. Korn CW
    5. Roiser JP
    6. Dolan RJ

    (2014) Losing the rose tinted glasses: neural substrates of unbiased belief updating in depression

    Frontiers in Human Neuroscience 8:639.

    https://doi.org/10.3389/fnhum.2014.00639

    • PubMed
    • Google Scholar
25. 1. Garrett N
    2. González-Garzón AM
    3. Foulkes L
    4. Levita L
    5. Sharot T

    (2018) Updating beliefs under perceived threat

    The Journal of Neuroscience 38:7901–7911.

    https://doi.org/10.1523/JNEUROSCI.0716-18.2018

    • PubMed
    • Google Scholar
26. 1. Garrett N
    2. Sharot T

    (2014) How robust is the optimistic update Bias for estimating self-risk and population base rates?

    PLOS ONE 9:e98848.

    https://doi.org/10.1371/journal.pone.0098848

    • PubMed
    • Google Scholar
27. 1. Garrett N
    2. Sharot T

    (2016) Optimistic update Bias holds firm: three tests of robustness following shah et al

    Consciousness and Cognition 50:12–22.

    https://doi.org/10.1016/j.concog.2016.10.013

    • PubMed
    • Google Scholar
28. 1. Goldbeter A

    (2011) A model for the dynamics of bipolar disorders

    Progress in Biophysics and Molecular Biology 105:119–127.

    https://doi.org/10.1016/j.pbiomolbio.2010.11.007

    • PubMed
    • Google Scholar
29. 1. Goldbeter A

    (2013) Origin of Cyclicity in Bipolar Disorders: A Computational Approach

    Pharmacopsychiatry 46:S44–S52.

    https://doi.org/10.1055/s-0033-1341502

    • Google Scholar
30. 1. Grande I
    2. Berk M
    3. Birmaher B
    4. Vieta E

    (2016) Bipolar disorder

    The Lancet 387:1561–1572.

    https://doi.org/10.1016/S0140-6736(15)00241-X

    • Google Scholar
31. 1. Hadaeghi F
    2. Hashemi Golpayegani MR
    3. Murray G

    (2015) Towards a complex system understanding of bipolar disorder: a map based model of a complex winnerless competition

    Journal of Theoretical Biology 376:74–81.

    https://doi.org/10.1016/j.jtbi.2015.02.020

    • PubMed
    • Google Scholar
32. 1. Harrison PJ
    2. Cipriani A
    3. Harmer CJ
    4. Nobre AC
    5. Saunders K
    6. Goodwin GM
    7. Geddes JR

    (2016) Innovative approaches to bipolar disorder and its treatment

    Annals of the New York Academy of Sciences 1366:76–89.

    https://doi.org/10.1111/nyas.13048

    • PubMed
    • Google Scholar
33. 1. Hooshmand F
    2. Do D
    3. Shah S
    4. Gershon A
    5. Park DY
    6. Yuen LD
    7. Dell'Osso B
    8. Wang PW
    9. Miller S
    10. Ketter TA

    (2019) Antidepressants have complex associations with longitudinal depressive burden in bipolar disorder

    Journal of Affective Disorders 246:836–842.

    https://doi.org/10.1016/j.jad.2018.12.074

    • PubMed
    • Google Scholar
34. 1. Huber MT
    2. Krieg J-C
    3. Braun HA
    4. Pei X
    5. Neiman A
    6. Moss F

    (2000) Noisy precursors of bifurcations in a neurodynamical model for disease states of mood disorders

    Neurocomputing 32-33:823–831.

    https://doi.org/10.1016/S0925-2312(00)00250-2

    • Google Scholar
35. 1. Kappes A
    2. Faber NS
    3. Kahane G
    4. Savulescu J
    5. Crockett MJ

    (2018) Concern for others leads to vicarious optimism

    Psychological Science 29:379–389.

    https://doi.org/10.1177/0956797617737129

    • PubMed
    • Google Scholar
36. 1. Kappes A
    2. Harvey AH
    3. Lohrenz T
    4. Montague PR
    5. Sharot T

    (2020) Confirmation bias in the utilization of others’ opinion strength

    Nature Neuroscience 23:130–137.

    https://doi.org/10.1038/s41593-019-0549-2

    • Google Scholar
37. 1. Korn CW
    2. Sharot T
    3. Walter H
    4. Heekeren HR
    5. Dolan RJ

    (2014) Depression is related to an absence of optimistically biased belief updating about future life events

    Psychological Medicine 44:579–592.

    https://doi.org/10.1017/S0033291713001074

    • Google Scholar
38. 1. Kuzmanovic B
    2. Jefferson A
    3. Vogeley K

    (2014) Self-specific optimism Bias in belief updating is associated with high trait optimism

    Journal of Behavioral Decision Making 293:281–293.

    https://doi.org/10.1002/bdm.1849

    • Google Scholar
39. 1. Kuzmanovic B
    2. Jefferson A
    3. Vogeley K

    (2016) The role of the neural reward circuitry in self-referential optimistic belief updates

    NeuroImage 133:151–162.

    https://doi.org/10.1016/j.neuroimage.2016.02.014

    • Google Scholar
40. 1. Lähteenvuo M
    2. Tanskanen A
    3. Taipale H
    4. Hoti F
    5. Vattulainen P
    6. Vieta E
    7. Tiihonen J

    (2018) Real-world effectiveness of pharmacologic treatments for the prevention of rehospitalization in a finnish nationwide cohort of patients with bipolar disorder

    JAMA Psychiatry 75:347–355.

    https://doi.org/10.1001/jamapsychiatry.2017.4711

    • PubMed
    • Google Scholar
41. 1. Lefebvre G
    2. Lebreton M
    3. Meyniel F
    4. Bourgeois-Gironde S
    5. Palminteri S

    (2017) Behavioural and neural characterization of optimistic reinforcement learning

    Nature Human Behaviour 1:67.

    https://doi.org/10.1038/s41562-017-0067

    • Google Scholar
42. Report

    1. Lopez A

    (2008)

    Markov Models for Longitudinal Course of Youth Bipolar Disorder

    University of Pittsburgh.

    • Google Scholar
43. 1. Ma Y
    2. Li S
    3. Wang C
    4. Liu Y
    5. Li W
    6. Yan X
    7. Chen Q
    8. Han S

    (2016) Distinct oxytocin effects on belief updating in response to desirable and undesirable feedback

    PNAS 113:9256–9261.

    https://doi.org/10.1073/pnas.1604285113

    • PubMed
    • Google Scholar
44. 1. Mason L
    2. Eldar E
    3. Rutledge RB

    (2017) Mood instability and reward Dysregulation-A neurocomputational model of bipolar disorder

    JAMA Psychiatry 74:1275.

    https://doi.org/10.1001/jamapsychiatry.2017.3163

    • PubMed
    • Google Scholar
45. 1. Moutsiana C
    2. Garrett N
    3. Clarke RC
    4. Lotto RB
    5. Blakemore SJ
    6. Sharot T

    (2013) Human development of the ability to learn from bad news

    PNAS 110:16396–16401.

    https://doi.org/10.1073/pnas.1305631110

    • PubMed
    • Google Scholar
46. 1. Moutsiana C
    2. Charpentier CJ
    3. Garrett N
    4. Cohen MX
    5. Sharot T

    (2015) Human Frontal-Subcortical circuit and asymmetric belief updating

    Journal of Neuroscience 35:14077–14085.

    https://doi.org/10.1523/JNEUROSCI.1120-15.2015

    • PubMed
    • Google Scholar
47. 1. Nana L

    (2009) Bifurcation analysis of parametrically excited bipolar disorder model

    Communications in Nonlinear Science and Numerical Simulation 14:351–360.

    https://doi.org/10.1016/j.cnsns.2007.10.002

    • Google Scholar
48. 1. Oud M
    2. Mayo-Wilson E
    3. Braidwood R
    4. Schulte P
    5. Jones SH
    6. Morriss R
    7. Kupka R
    8. Cuijpers P
    9. Kendall T

    (2016) Psychological interventions for adults with bipolar disorder: systematic review and meta-analysis

    British Journal of Psychiatry 208:213–222.

    https://doi.org/10.1192/bjp.bp.114.157123

    • Google Scholar
49. 1. Peters AT
    2. West AE
    3. Eisner L
    4. Baek JH
    5. Deckersbach T

    (2016) The burden of repeated mood episodes in bipolar I disorder: results from the national epidemiological survey on alcohol and related conditions (NESARC)

    The Journal of Nervous and Mental Disease 204:583–592.

    https://doi.org/10.1097/NMD.0000000000000425

    • Google Scholar
50. 1. Radua J
    2. Grunze H
    3. Amann BL

    (2017) Meta-Analysis of the risk of subsequent mood episodes in bipolar disorder

    Psychotherapy and Psychosomatics 86:90–98.

    https://doi.org/10.1159/000449417

    • PubMed
    • Google Scholar
51. 1. Scheier MF
    2. Carver CS
    3. Bridges MW

    (1994) Distinguishing optimism from neuroticism (and trait anxiety, self-mastery, and self-esteem): a reevaluation of the life orientation test

    Journal of Personality and Social Psychology 67:1063–1078.

    https://doi.org/10.1037/0022-3514.67.6.1063

    • PubMed
    • Google Scholar
52. 1. Scott J
    2. Colom F
    3. Vieta E

    (2007) A meta-analysis of relapse rates with adjunctive psychological therapies compared to usual psychiatric treatment for bipolar disorders

    The International Journal of Neuropsychopharmacology 10:123–129.

    https://doi.org/10.1017/S1461145706006900

    • PubMed
    • Google Scholar
53. 1. Sharot T
    2. Korn CW
    3. Dolan RJ

    (2011) How unrealistic optimism is maintained in the face of reality

    Nature Neuroscience 14:1475–1479.

    https://doi.org/10.1038/nn.2949

    • PubMed
    • Google Scholar
54. 1. Sharot T
    2. Guitart-Masip M
    3. Korn CW
    4. Chowdhury R
    5. Dolan RJ

    (2012a) How dopamine enhances an optimism Bias in humans

    Current Biology 22:1477–1481.

    https://doi.org/10.1016/j.cub.2012.05.053

    • PubMed
    • Google Scholar
55. 1. Sharot T
    2. Kanai R
    3. Marston D
    4. Korn CW
    5. Rees G
    6. Dolan RJ

    (2012b) Selectively altering belief formation in the human brain

    PNAS 109:17058–17062.

    https://doi.org/10.1073/pnas.1205828109

    • PubMed
    • Google Scholar
56. 1. Steinacher A
    2. Wright KA

    (2013) Relating the bipolar spectrum to dysregulation of behavioural activation: a perspective from dynamical modelling

    PLOS ONE 8:e63345.

    https://doi.org/10.1371/journal.pone.0063345

    • PubMed
    • Google Scholar
57. 1. Tondo L
    2. Vázquez GH
    3. Baldessarini RJ

    (2017) Depression and Mania in bipolar disorder

    Current Neuropharmacology 15:353–358.

    https://doi.org/10.2174/1570159X14666160606210811

    • PubMed
    • Google Scholar
58. 1. Tundo A
    2. Musetti L
    3. Benedetti A
    4. Massimetti E
    5. Pergentini I
    6. Cambiali E
    7. Dell'Osso L

    (2018) Predictors of recurrence during long-term treatment of bipolar I and II disorders. A 4 year prospective naturalistic study

    Journal of Affective Disorders 225:123–128.

    https://doi.org/10.1016/j.jad.2017.08.013

    • PubMed
    • Google Scholar
59. 1. Vieta E
    2. Berk M
    3. Schulze TG
    4. Carvalho AF
    5. Suppes T
    6. Calabrese JR
    7. Gao K
    8. Miskowiak KW
    9. Grande I

    (2018) Bipolar disorders

    Nature Reviews Disease Primers 4:18008.

    https://doi.org/10.1038/nrdp.2018.8

    • PubMed
    • Google Scholar

Acceptance summary:

This research highlights a predictor of mood states in individuals with bipolar disorder, above and beyond clinical predictors. Namely, a bias in belief updating when an initial prediction overstates prevalence (i.e. "good news" situation) relative to when an initial prediction understates an outcome (i.e. "bad news" situation) is predictive of future time spent euthymic in individuals with bipolar disorder. These findings could lead to novel therapeutic interventions for those with bipolar disorder.

Decision letter after peer review:

Thank you for submitting your article "Belief Updating in Bipolar Disorder Predicts Time of Recurrence" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by Shelly Flagel as a Reviewing Editor and Michael Frank as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Amy Louise Cochran (Reviewer #1); Catherine Harmer (Reviewer #3).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

The editors have judged that your manuscript is of interest, but as described below noted that additional analyses and significant revisions are required before it is published. Thus, we would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). First, because many researchers have temporarily lost access to the labs, we will give authors as much time as they need to submit revised manuscripts. We are also offering, if you choose, to post the manuscript to bioRxiv (if it is not already there) along with this decision letter and a formal designation that the manuscript is "in revision at eLife". Please let us know if you would like to pursue this option. (If your work is more suitable for medRxiv, you will need to post the preprint yourself, as the mechanisms for us to do so are still in development.)

Summary:

Ossala et al. seek to predict future time spent euthymic in individuals with bipolar disorder (BP) based on how they update their beliefs. To predict time spent euthymic, they emphasize one variable, in particular, the bias in average updates when their initial prediction overstates actual prevalence (which they call a "good news" situation) vs. when their initial prediction understates actual outcomes (which they call a "bad news" situation). The principal finding is that the bias in average updates is predictive of future time spent euthymic after controlling for a number of possible confounds including subjective biases, task-related variables, and clinical variables. The study design is thoughtful and interesting, and the principal finding appears to be exceedingly strong. If biases in updating is a robust predictor of future outcomes, above and beyond clinical predictors, then this variable should have utility in helping set expectations for individuals with BP as well as helping get at underlying cognitive deficits in BP.

Essential revisions:

Considering the small sample size (n = 36 outcomes) and the number of covariates (e.g., n = 24 in Model 3), a major concern is the possibility of an overinflation of p values for various reasons given below. Based on this concern, the authors need to address the following:

1) Did the authors perform a power analysis before study start? The small sample size should be raised as a limitation in the Discussion.

2) There may be multicollinearity among independent variables in the regression models that might lead to an overinflation of p values. Broadly, please address the issue of multicollinearity. Specifically, can you report univariate analyses relating update bias to time spent euthymic? And can you report (in the supplement) the correlation between independent variables in the regression model?

3) While cross-validation was used to show the prediction models would generalize to a larger population, it was not used to determine if the principal finding (i.e. relationship between update bias and time spent euthymic) would generalize. Can you use cross-validation to evaluate the generalizability of this finding, e.g., by comparing cross-validated prediction errors to a model that does not include update bias?

4) Based on the Materials and methods section, it appears that you are running a standard linear regression, but you mention that you construct general linear models (GLM). If the former, please simply state that. If the latter, can you provide more details in what exact analysis was performed? For example, using repeated measures could lead to lower p values, but would need to be described (though inappropriate for the given outcome which is only measured once per individual).

5) Please discuss the small sample size as a limitation in the Discussion. Assuming a standard linear regression was performed, there should be 5-10 outcomes per independent variable, but the presented analysis only had 2-3.

6) Given that the task has been analyzed elsewhere, please compare/contrast your analyses to analyses from prior papers? Among task-related variables (e.g., difference in number of good vs. bad trials) that were controlled for in the regression models, how many were used in prior papers?

7) Related to the last comment, in prior work using the task and in models of associative learning more broadly, it is common to analyze bias in learning rates (roughly the ratio between the update and the error) rather than the update bias. Can the authors comment on why they focused on bias in overall updates rather than bias in learning rates?

8) One might have expected that optimistic updates predict the transition only to manic and not depressives states. The authors should discuss this and other alternative hypotheses and whether their reported findings fit into any existing and coherent model of BP disorder.

9) The authors show that the belief update task can predict the time of the dysthymic transition. But they only have tested the belief update task. What if any task predicts the time of the dysthymic? On the other side, optimistic biases can be assessed in several ways. On one side of the spectrum there are standard personality measures, such as the LOT-R, on the other side of the spectrum there are reinforcement learning tasks, where the optimistic bias is frequently reported. Does any measure of optimism predict the time of the dysthymic transition? Or is it just the belief update task? In the Supplementary material it is mentioned that trait optimism (LOT-R) was assessed. Does this scale also predict the time spent in euthymia? Do the results hold if it is included in the GLMs?

10) Another important check that is lacking is whether the results hold if the number of previous dysthymic episodes is entered as a control. In such pathological states, it seems that crises become more and more frequent with time. A possible interpretation of the results (that cannot be ruled out presently) is that a higher number of previous dysthymic episodes makes a patient concomitantly less optimistic and more likely to have a shorter euthymic phase. A related question is whether the belief task is a better predictor compared just the number of previous dysthymic episodes (note that even if correlated this factor is not completely taken into account by including the patient age in the regression).

11) Medication effects were controlled for statistically by including 0 or 1 for different classes of medication. However, this is unlikely to be a perfect control as few participants are medication free (>80% on a mood stabiliser). This should be discussed in more detail.

12) The inclusion of the different clinical and demographic factors that might influence the association is appreciated. One might wonder if this could just be a general cognitive effect e.g. the more “impaired” cognition (or difference to normal) associated with a worst illness profile. The authors go some way towards assessing this by looking for example at memory for the new information presented in the task. However, this justifies a bit more attention i.e. did the authors include relapse history in the model? Was memory (as oppose to belief updating) a predictor by itself?

13) What is the likely mechanism underpinning the failure to update beliefs on the basis of positive information? Is it that despite knowing the population average, the participant believes they have more specific information to refute that a positive update applies to them e.g. burglaries happen to 20% of people. but for some its 30% because they live in an area where burglaries are more common. So, could it be that indirectly the task measures the disadvantaged setting of individuals more prone to relapse? Or is it in the mechanism by which this learning takes place?

Essential revisions:

Considering the small sample size (n = 36 outcomes) and the number of covariates (e.g., n = 24 in Model 3), a major concern is the possibility of an overinflation of p values for various reasons given below. Based on this concern, the authors need to address the following:

1) Did the authors perform a power analysis before study start? The small sample size should be raised as a limitation in the Discussion.

Sample size was based on a previous study (Korn et al., 2014), which found a negative correlation between the Update Bias and BDI scores of r = -0.50 (p < 0.001). The required sample size based on this was n=23 for an alpha of 0.05 and 80% power, beta=0.80 (Faul et al., 2007). We started with an N of 45 and following drop out over the five years ended with an N of 36. We now include this information in the Materials and methods section. We have also added a Discussion of the small sample size.

2) There may be multicollinearity among independent variables in the regression models that might lead to an overinflation of p values. Broadly, please address the issue of multicollinearity. Specifically, can you report univariate analyses relating update bias to time spent euthymic? And can you report (in the supplement) the correlation between independent variables in the regression model?

To test for multicollinearity, we calculated the Variance Inflation Factor (VIF) for each independent variable. The VIF quantifies the amount of multicollinearity in a set of regression variables. It is equal to the ratio of the overall model variance to the variance of a model that includes only that single independent variable. A high VIF indicates that the independent variable is highly collinear with the other variables in the model. A low VIF (less than 4) indicates that it is not. The VIF of our measures of interest – the update bias – was 1.93 for model 1 and 2.87 for model 2, which is low and not considered problematic. We now report this in the revised manuscript.

Indeed, the only independent variable the update bias was significantly correlated with was difference in memory errors. Taking away that variable from the model did not change the results (Update bias Beta = 0.544, p = 0.002 for model 1; Beta = 0.680, p = 0.001 for model 2). We now report the correlation between all independent variables in the regression models as requested (see Supplementary Material).

We also tested whether the effect of the update bias was contingent on entering a specific set of variables in the model by running “model averaging” (Freckleton, 2011). This approach involves first running every single combination of models given the independent variables. For example, running a model only with 2 of the variables, only 3, only 4 and so on. Each time with a different combination of independent variables. With N = 13 (Model 1) this involved running 4095 nested models. Then the betas of each variable are averaged across all models, weighting them on the model’s BIC (Bayesian Information Criterion). This revealed a significant effect of update bias in explaining “time in euthymia” (weighted estimates = 0.47, 95% CI = 0.34-0.60). The eight best fitting models (out of 4095), according to the lowest BIC score, all included the update bias. Running the same procedure for Model 2 will involve running 8,388,607 models and we lack the computational power needed for that exercise. However, given the significant effects of Model 1, the goal of Model 2 is simply to ensure that the results of Model 1, which includes less variables, are not explained away by additional clinical and demographic variables.

While multicollinearity is not driving the effects, our additional analysis reveals that there are differences between the two types of trials (good news trials and bad news trials) that introduce noise and can obscure the effects of update bias if not controlled for. In particular, testing if any of the task variables showed significant differences between desirable and undesirable trials revealed significant effects in all subjective ratings: past experience, familiarity, vividness, arousal, vividness and negativity (p-values < 0.02) but not in any of the other variables (that is no difference for memory errors, estimation errors, number of trials or RTs). Simply correlating update bias with time in euthymia without accounting for these significant differences reveals a trend (rho = 0.297 p = 0.078), and once the differences are accounted for by running a new model controlling only for these subjective measures the relationship between Update Bias and “time in euthymia” is clearer (Beta = 0.351; p = 0.047). The VIF of the update bias in this model is low (VIF = 1.098) as were the VIFs of all other variables (all VIFs < 4). We now report all this in the revised manuscript.

3) While cross-validation was used to show the prediction models would generalize to a larger population, it was not used to determine if the principal finding (i.e. relationship between update bias and time spent euthymic) would generalize. Can you use cross-validation to evaluate the generalizability of this finding, e.g., by comparing cross-validated prediction errors to a model that does not include update bias?

Thank you for suggesting this confirmatory analysis. Running the LOO-cv analysis using the model that does not include update bias showed that the predicted values and actual values were not correlated (r = 0.129; p = 0.459). Compare this to the same model which does include update bias, for which we find both a strong correlation between predicted values and actual values (r = 0.445, p = 0.007) and no significant difference between these two values (t(34) = 0.123; p = 0.901, paired sample t-test). The model’s fit (SSE, sum of squared errors) was better for the model including the Update Bias (SSE = 61.80) than the one that did not (SSE = 97.10). We have now added this analysis to the revised manuscript.

4) Based on the Materials and methods section, it appears that you are running a standard linear regression, but you mention that you construct general linear models (GLM). If the former, please simply state that. If the latter, can you provide more details in what exact analysis was performed? For example, using repeated measures could lead to lower p values, but would need to be described (though inappropriate for the given outcome which is only measured once per individual).

Thank you for noting this. We ran linear regressions not general linear models. This typo has now been corrected.

5) Please discuss the small sample size as a limitation in the Discussion. Assuming a standard linear regression was performed, there should be 5-10 outcomes per independent variable, but the presented analysis only had 2-3.

Following the reviewer’s suggestion, we have now added a discussion of the limitation of the small sample size. In addition, we now perform two additional analyses that help alleviate the concern of number of outcomes per independent variable. First, we rerun the regression analysis using bootstrapping which enables us to construct confidence intervals with a small sample. Performing bootstrapping on 10000 samples again showed that the relationship between the Update Bias was significantly related to time in Euthymia (Beta = 0.589, CI = 0.122-1.129). Second, we run model averaging (Fletcher, 2018). In particular, running 4095 nested models, most of which include a smaller number of independent variables. This again revealed a significant effect of update bias in predicting “time in euthymia” (Estimates weighted by log model evidence = 0.465, 95% CI = 0.34-0.60). Third, we now report the results of a new model with only 6 variables, which is 6 outcomes per independent variable. We now add these additional analyses to the revised manuscript.

6) Given that the task has been analyzed elsewhere, please compare/contrast your analyses to analyses from prior papers? Among task-related variables (e.g., difference in number of good vs. bad trials) that were controlled for in the regression models, how many were used in prior papers?

Our approach was to include all the control variables that were included in the previous papers adopting the same task in clinically depressed participants (Korn et al., 2014; Garrett et al., 2014). We now mention this in the Materials and methods. To our knowledge, seven past papers (Chowdury et al., 2013; Korn et al., 2014, 2016; Garrett et al., 2014, 2018; Kuzmanovic et al., 2017; Ma et al., 2016) include difference in number of good vs. bad trials as controls, and nine papers include RT (Sharot et al., 2011; Sharot et al., 2012b, Chowdury et al., 2013; Korn et al., 2014; Garrett et al., 2014; Moutsiana et al., 2013; Kuzmanovic et al., 2017, 2018; Ma et al., 2016). All papers from the Affective Brain Lab (N = 10) include memory and subjective ratings (i.e. familiarity, prior experience, vividness, emotional arousal and negativity) as controls, as do some papers not from the lab (e.g., Ma et al., 2016; Oganiana et al., 2018; Schonfelder et al., 2017). Thus, the controls included in the paper are in line with past studies.

7) Related to the last comment, in prior work using the task and in models of associative learning more broadly, it is common to analyze bias in learning rates (roughly the ratio between the update and the error) rather than the update bias. Can the authors comment on why they focused on bias in overall updates rather than bias in learning rates?

We focused on the update itself rather than the ratio between update and error for the following reasons. First, in the 20 papers published thus far with the update task 100% used update as a measure of interest (Sharot et al., 2011, 2012a, 2012b; Chowdury et al., 2013; Korn et al., 2014; Moutsiana et al., 2013, 2015; Garrett et al., 2014, 2017, 2018; Kappes et al., 2018; Oganian et al., 2018; Kuzmanovic et al., 2014, 2016, 2017, 2018, 2019; Ma et al., 2016; Schonfelder et al., 2017) and only 40% use the ratio between update and error as a measure of interest (Sharot et al., 2011; Kuzmanovic et al., 2017, 2018; Kappes et al., 2018; Ma et al.,l 2016). Moreover, in the three papers that use this task to examine affective disorders all three used update as a measure of interest (Garrett et al., 2014; Korn et al., 2014; Schonfelder et al., 2017) and only one used the ratio between update and error as a measure of interest. Thus, to be consistent with past research we focused on the update measure. Second, the reason that past studies, in particular those looking at patients (Garrett et al., 2014; Korn et al., 2014, 2016; Kuzmanovic et al., 2019), focus on update rather than the ratio between update and error is that such populations usually show a general impairment in the latter but not the former. This makes the ratio a noisy measure to use in such populations. Indeed, the average ratio between update and error for good news in our clinical sample was 0.63, below that seen in healthy populations (0.72 in the original study by Sharot et al., 2011; and ranging between 0.67 in Kappes et al., 2018 to 0.77 in Ma et al., 2016).

8) One might have expected that optimistic updates predict the transition only to manic and not depressives states. The authors should discuss this and other alternative hypotheses and whether their reported findings fit into any existing and coherent model of BP disorder.

Thank you for raising this question. Indeed, one hypothesis would be that greater update bias will be associated specifically with transition to a manic episode. This type of prediction will be compatible with the notion held by some that mood is a one-dimensional construct with depression and mania as opposites of the same continuum (Chang and Chou, 2018). An alternative hypothesis (which our data supports) is that greater update bias will be associated with episodes of either polarity. Such a hypothesis would be consistent with multi-stable models that suggest that environmental or inner perturbation such as stressors could trigger either episode onset (Cochran et al., 2017, 2018). The update bias may be an indicator of the existence of such stressors and their associated prodromal symptoms. These models consider mood a two-dimensional construct and suggest that mania and depression can vary independently and do not inhibit each other. We now detail these hypotheses in the revised manuscript.

9) The authors show that the belief update task can predict the time of the dysthymic transition. But they only have tested the belief update task. What if any task predicts the time of the dysthymic? On the other side, optimistic biases can be assessed in several ways. On one side of the spectrum there are standard personality measures, such as the LOT-R, on the other side of the spectrum there are reinforcement learning tasks, where the optimistic bias is frequently reported. Does any measure of optimism predict the time of the dysthymic transition? Or is it just the belief update task? In the Supplementary material it is mentioned that trait optimism (LOT-R) was assessed. Does this scale also predict the time spent in euthymia? Do the results hold if it is included in the GLMs?

Following the reviewer’s comment, we tested whether the LOTR was associated with time in euthymia. It was not (rho = 0.164, p = 0.340). Moreover, including the LOT-R in our model showed that the results hold. That is, update bias was still related to time in euthymia (beta = 0.740; p = 0.001) while LOTR was not (beta = 0.208, p = 0.148). In addition, we note that other measures such as memory bias (that is memory errors for positive trials minus for negative trials) also did not predict time in euthymia (beta = 0.13; p = 0.959). It may certainly be that other measures not tested here, such as a bias in learning in a reinforcement learning task (e.g., Lefebvre et al., 2017) may be related to time in euthymia. That will require future testing. However, our results show that not all tasks/measures predict time in euthymia. We now include this information in the revised manuscript.

10) Another important check that is lacking is whether the results hold if the number of previous dysthymic episodes is entered as a control. In such pathological states, it seems that crises become more and more frequent with time. A possible interpretation of the results (that cannot be ruled out presently) is that a higher number of previous dysthymic episodes makes a patient concomitantly less optimistic and more likely to have a shorter euthymic phase. A related question is whether the belief task is a better predictor compared just the number of previous dysthymic episodes (note that even if correlated this factor is not completely taken into account by including the patient age in the regression).

We thank the reviewer for prompting us to examine the effect of number of previous episodes. We extracted the number of previous episodes from the clinical records and found that it was not associated with time in euthymia (r = -0.013; p = 0.93) nor with the Update Bias (r = -0.107; p = 0.53). Adding this variable into the model did not change the results: update bias was still associated with time in euthymia (Beta = 0.066, p = 0.013). Thus, we conclude that update bias is a better predictor than number of previous dysthymic episodes in predicting future time in euthymia. We now add this information to the manuscript.

11) Medication effects were controlled for statistically by including 0 or 1 for different classes of medication. However, this is unlikely to be a perfect control as few participants are medication free (>80% on a mood stabiliser). This should be discussed in more detail.

It is true that given that most participants were on mood stabilisers it is impossible to reliably determine the effect of medication on the relationship between task performance and time in euthymia. This question could only be addressed if there are sufficient patients not on such medication. We now highlight this caveat in the Discussion.

12) The inclusion of the different clinical and demographic factors that might influence the association is appreciated. One might wonder if this could just be a general cognitive effect e.g. the more “impaired” cognition (or difference to normal) associated with a worst illness profile. The authors go some way towards assessing this by looking for example at memory for the new information presented in the task. However, this justifies a bit more attention i.e. did the authors include relapse history in the model? Was memory (as oppose to belief updating) a predictor by itself?

Following the reviewers’ suggestion, we looked at the association between the overall magnitude of memory errors and future time in euthymia and found no association (r = -0.008; p = 0.964). Neither was relapse history, calculated as the number of previous episodes, associated with future time in euthymia (r = -0.013; p = 0.939). Finally, we note that the association between update and time in euthymia is valence-dependent. Greater update in response to positive information is associated with longer time in euthymia but more update in response to negative information is associated with less time in euthymia. We thus conclude that the association between the update bias and time in euthymia is unlikely to be explained by a general cognitive impairment as this would have been better captured by overall update and overall memory performance. We now add this to the revised manuscript.

13) What is the likely mechanism underpinning the failure to update beliefs on the basis of positive information? Is it that despite knowing the population average, the participant believes they have more specific information to refute that a positive update applies to them e.g. burglaries happen to 20% of people. but for some its 30% because they live in an area where burglaries are more common. So, could it be that indirectly the task measures the disadvantaged setting of individuals more prone to relapse? Or is it in the mechanism by which this learning takes place?

We thank the reviewer for raising this question. Our task carefully controls for people’ perception of how likely they are to experience these aversive events by including their first estimate of these likelihoods in the model. We also have two additional measure of the “disadvantaged setting” of the participants; we ask them how often they had experienced these events in the past and how familiar they are with the events from others experiencing them. Including the average rating on either of these measures in our model did not change the main result: update bias was still associated with future time in euthymia (Beta = 0.806; p = 0.003) while these measures were not (past experience Beta = -0.022 p = 0.929; familiarity Beta = 0.177 p = 0.416). It is possible that the failure to update in response to positive information is due to the participant perceiving the positive information as irrelevant to them.

Author details

1. Paolo Ossola

   Psychiatry Unit, Department of Medicine and Surgery, Università di Parma, Parma, Italy

   Contribution

   Conceptualization, Resources, Data curation, Formal analysis, Investigation, Methodology, Writing - original draft, Project administration

   For correspondence

   paolo.ossola@unipr.it

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-0644-3158

2. Neil Garrett

   Department of Experimental Psychology, University of Oxford, Oxford, United Kingdom

   Contribution

   Conceptualization, Data curation, Software, Formal analysis, Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

3. Tali Sharot

   Affective Brain Lab, Department of Experimental Psychology, University College London, London, United Kingdom

   Contribution

   Conceptualization, Supervision, Validation, Methodology, Writing - original draft, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-8384-6292

4. Carlo Marchesi

   Psychiatry Unit, Department of Medicine and Surgery, Università di Parma, Parma, Italy

   Contribution

   Conceptualization, Resources, Data curation, Supervision, Funding acquisition, Methodology, Project administration, Writing - review and editing

   Competing interests

   No competing interests declared

• 3,750

  Page views

• 219

  Downloads

• 5

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.