When learning about and interacting with the world, individuals vary in the extent to which their beliefs and behaviours are influenced by the events they experience. Often this variation displays an affective gradient with some individuals being more influenced by positive and others by negative events. For example, many people display an optimism bias, updating their beliefs to a greater extent following positive than negative outcomes (Sharot and Garrett, 2016). The opposite effect, a tendency to be more influenced by negative events, has been argued to cause illnesses such as depression and anxiety (Mathews and MacLeod, 2005). However, relatively little work has explored why individuals might develop affective biases in the first place. This question is of particular importance as understanding the mechanisms which lead to the development of affective bias is an essential first step in the development of novel treatments designed to alter this process and thus reduce symptoms of depression and anxiety. One way of answering why individuals develop affective bias is to consider when affective biases might be the appropriate way to think about the world. In this study we draw on recent advances from the computational neuroscience of learning to investigate whether affective biases may be understood in terms of how informative an individual judges an event to be. Below we describe the conceptual framework of this proposal and then suggest how this may be used to account for the occurrence of affective biases.

Recent computational work has demonstrated that individuals’ expectations are influenced more by those events which carry more information; that is, those events which improve predictions of future outcomes to a greater degree (Behrens et al., 2007; Browning et al., 2015; MacKay, 2003; Nassar et al., 2012). One factor which influences how informative an event is is the changeability, or volatility, of the underlying association which is being learned. For example, imagine trying to learn what your colleagues think about your performance at work, based solely on their day-to-day feedback. One colleague seems to have a stable positive view of you, complimenting you on your work on 80% of the occasions you meet and never increasing or decreasing this frequency. In this case, each particular event (being complimented or not) provides little new information about what your colleague thinks about you, as you will always have an 80% chance of being complimented the next time you meet. In contrast, a second colleague’s appraisal of you seems to be more changeable, with periods when they think highly of you and compliment you regularly and others when they rarely compliment you at all. In this case each event provides more information; if you have recently been complimented by this colleague it is more likely that their opinion of you is currently high and they will compliment you the next time you meet (Figure 1B). When learning what your colleagues currently think about you, you should be more influenced by whether the second, more volatile, colleague compliments you or not, because this provides more useful information than the behaviour of the stable colleague.

Figure 1

Download asset Open asset

Within a reinforcement learning framework, the influence of events on one’s belief is captured by the learning rate parameter, with a higher learning rate reflecting a greater influence of more recently experienced events (Sutton and Barto, 1998). Humans adjust their learning rate precisely as described above, using a higher learning rate for events, such as those occurring in a volatile context, which they estimate to be more informative (Behrens et al., 2007; Browning et al., 2015; Nassar et al., 2012). The neural mechanism by which this modification of learning rate is achieved is thought to depend on activity of the central norepinepheric system (Yu and Dayan, 2005), with increased phasic activity of the system, which may be estimated using pupilometry (Joshi et al., 2016), reporting the occurrence of more informative events (Browning et al., 2015; Nassar et al., 2012) and acting to enhance the processing of these events (Aston-Jones and Cohen, 2005).

This computational framework provides an overarching logic for when an individual might develop an affective biases; individuals should bias their processing towards those affective events that they estimate to be most informative. As well as providing a novel reformulation of why affective biases may develop, this framework also suggests a potential novel method for modifying such biases; for example, if a higher estimate of the information content of negative relative to positive events leads to negative affective bias, interventions which redress these estimates should also reduce the negative bias.

However, a number of critical questions concerning this account remain outstanding. Firstly, no previous study has demonstrated that humans maintain separate estimates of the information content of positive and negative events. While a number of studies have examined the effect of volatility on learning (Behrens et al., 2008; Behrens et al., 2007; Browning et al., 2015; Nassar et al., 2012), they have all utilised only one type of outcome (i.e. rewards or punishments) and thus their results could be accounted for by learners maintaining an estimate of how volatile the general environment is and learning more rapidly to all outcomes in those environments they judge to be more volatile, rather than estimating the information content of specific outcomes. We tested whether these specific estimates were maintained using a novel learning task (Figure 1) in which participant choice led to both positive and negative outcomes, with the volatility of the outcomes (and therefore their information content) being independently manipulated in separate task blocks. Secondly, for estimates of the information content of positive and negative events to lead to the development of affective biases, the estimates themselves must be malleable. We assessed this malleability by testing whether the volatility manipulation described above altered participants’ estimated information content, as reflected by the learning rates they used. Lastly, while activity of the central NE system has been argued to represent estimates of volatility, it is not clear whether or how this system might multiplex separate representations of the volatility of different classes of event, such as the positive and negative outcomes examined here. We investigated this using pupilometry as a measure of NE activity while participants completed the task. We hypothesised that humans maintain separable estimates of the information content of positive and negative outcomes, that we could measure and manipulate these estimates using our task and that phasic NE activity yoked to a specific type of outcome would track the volatility of that outcome.

30 participants (see Table 1 for demographic information) completed a two option learning task in which, on every trial, one option would be associated with a monetary win and one with a loss (Figure 1). The win and loss outcomes occurred independently which required participants to learn separately which shape was associated with each outcome (Figure 1c). The information content of the outcomes was varied across the three blocks by altering the volatility of the stimulus-outcome associations (Figure 1b). We estimated separate learning rates for the positive and negative outcomes by fitting a computational model (see Materials and methods) to participant choice in each task block. This allowed us to test whether participants independently altered the learning rates they used for the win and loss outcomes in response to how informative that outcome was (i.e. its volatility). Pupilometry data was collected during the task as a measure of activity in the central NE system.

Do human learners maintain independent estimates of the information content of positive and negative outcomes?

As predicted, participants’ learning rates for positive and negative outcomes reflected the information content of the outcomes in the learning task (block volatility x parameter valence; F(1,28) =27.97, p<0.001; Figure 2). Specifically, learning rates were higher for win (F(1,28) =15.47, p=0.001) and loss (F(1,28) =18.02, p<0.001) outcomes when they were volatile (informative) than when they were stable (not informative). Similarly the learning rate for wins was higher than that for losses when wins were more volatile than losses (F(1,28) =26.02, p<0.001) and the learning rate for losses was higher than for wins when losses were more volatile (F(1,28) =6.74, p=0.015). These results demonstrate that participants maintain independent estimates of the information content of positive and negative outcomes and that it is possible to alter these estimates using a simple volatility manipulation. In contrast to the effects on learning rate there were no significant effects of the task on the inverse temperature parameter of the learning model (Figure 2b; F(1,28) =0.01, p=0.92) indicating that, as intended, the volatility manipulation specifically altered learning rate. See the Figure Supplements for Figure 2 for additional analysis of the behavioural results as well as an additional experiment in which the impact of expected uncertainty was assessed.

Figure 2 with 2 supplements see all

Download asset Open asset

Figure 2—source data 1

Fitted model parameters, questionnaire measures and mean pupil response to volatility for all participants.

https://doi.org/10.7554/eLife.27879.007

Download elife-27879-fig2-data1-v3.xlsx

Does activity of the central NE system, as Estimated by Pupil Dilation, Track the Volatility of Positive and Negative Outcomes?

Next, we investigated the extent to which central NE activity, as estimated using pupilometry, was related to the information content of positive and negative outcomes in the learning task. Consistent with the behavioural findings a significant interaction between block volatility and outcome valence was found for the degree to which participants’ pupils dilated in response to outcome receipt (Figure 3; F(1,27)=6.16; p=0.02). In other words, participants’ pupils dilated more on receipt of an outcome when that outcome was volatile (informative) relative to when it was stable (not informative). This effect was not further modified by the time bin following outcome (block volatility x outcome valence x time; F(5,135)=1.13, p=0.35). Analysing the positive and negative outcomes separately indicated that the effect of block volatility was significant for the loss outcomes (F(1,27)=10.46, p=0.003), but not for the win outcomes (F(1,27)=0.38, p=0.54). Indeed a direct statistical comparison of the size of the volatility effect between the positive and negative outcomes indicated a greater effect of volatility on the negative relative to positive outcomes (outcome volatility x valance; F(1,27)=4.34, p=0.047). This effect was seen on the background of a generally greater pupil dilation to receipt of a loss relative to a win (main effect of valence; F(1,27)=16.7, p<0.001).

Figure 3 with 2 supplements see all

Download asset Open asset

Are the behavioural and pupilometry measures capturing the same process?

As central NE activity is thought to mediate the effect of outcome information content on participant choice (Yu and Dayan, 2005), there should be a relationship between how much a participant’s pupils differentially dilate in response to an outcome during the informative and non-informative blocks and the degree to which that participant adjusts their learning rate between blocks for the same outcome. We tested this by assessing the correlation between the change in mean pupil response between blocks and the change in behaviourally estimated learning rates, separately for wins and losses. As can be seen (Figure 4) the change in pupil response to loss outcomes between blocks was significantly correlated with the change in loss learning rate (r(28)=0.5, p=0.009) but pupil response to win outcomes was not correlated with change in win learning rate (r(28)=-0.08, p=0.7). This correlation was significantly greater for losses than for wins (Fisher r-to-z transformation z = 2.27, p=0.02).

Figure 4 with 2 supplements see all

Download asset Open asset

Humans adapt the degree to which they are influenced by a positive and negative outcome in response to how informative they estimate those outcomes to be. These estimates produce a bias resulting in preferential learning, with a higher learning rate being used for the outcome which is most informative. These estimates are also malleable and thus may represent one route by which affective biases develop. A physiological measure of central NE activity was associated with this process, although this was only seen convincingly for loss outcomes.

Previous work has demonstrated that humans adapt their learning in response to subtle statistical aspects of the environment, such as employing an increased learning rate in volatile, or changeable, contexts (Behrens et al., 2007; Browning et al., 2015; Nassar et al., 2012). These previous findings could be explained by learners maintaining an estimate of how volatile the general environment is and learning more rapidly from all outcomes experienced in those environments they judge to be more volatile. However, the results of the current study, in which learning rates for positive and negative outcomes were seen to be altered independently, cannot be accounted for by a general estimate of environmental volatility. Rather, this behaviour requires the parallel representation of the estimated volatility of distinct outcomes which are then used to specifically tune the learning from that outcome. More generally these results suggest that human learners are able to maintain independent estimates of the information content of different events and use these estimates to rationally adjust their learning, in this case producing a valence dependent affective bias.

In the current study we investigated the link between the learning rate used by participants, which provides a behavioural index of how informative they estimate an outcome to be, and pupil dilation which has been shown to correlate with central norepinepheric activity (Joshi et al., 2016). Pupil dilation in response to outcome receipt differed as a function of the information content of the outcome, although this was only significant for losses. Specifically, when losses were informative, the difference in pupil dilation between trials in which a loss was received and when it was not received was greater than when the losses were not informative. This result is similar to previously reported findings of an increased pupil response to outcomes in a volatile context (Browning et al., 2015; Nassar et al., 2012), although these earlier studies reported a general increase in pupil dilation rather than a dilation conditioned on receipt of the outcome. A possible explanation of this difference is that, as discussed above, in order to complete the tasks used in previous studies, which involved only one class of outcome, only an estimate of the general environmental volatility is required, whereas to perform the current task a volatility signal dependent on the outcome presented is needed. In other words, the volatility signal found in the pupil data from the current study is of the form required for participants to accurately perform the task. This suggests a degree of flexibility of the pupillary volatility signal, in that it may reflect the general volatility of a learned association or the volatility of specific dimensions of more complex associations depending on task demands. It is not clear whether these general and specific volatility signals are produced by a single or separate neural systems, although it may be possible to address this question using a task in which the total volatility of all task outcomes is manipulated independently of the volatility of individual outcomes.

The effect of outcome volatility on pupil dilation in the current study was significantly greater for loss than win outcomes with the correlation between this signal and behaviour also being significantly greater for losses. This surprising result raises the possibility that central norepinepheric activity is particularly related to the information content of negative, as opposed to positive outcomes. However, while we are not aware of previous studies which have reported pupillary volatility signals to positive and negative outcomes from a single task as in the current study, previous work has reported the presence of pupillary volatility signals in reward only tasks (Nassar et al., 2012). This suggests that the norepinpeheric system does respond to the volatility of positive outcomes, but that this response is less pronounced than that for negative outcomes. One explanation for this may be that, as discussed above, the volatility signal in the current task modified pupillary response to outcome receipt vs. non-receipt. The receipt of a loss led to a significantly greater pupil dilation than that produced by a win (see Figure 3—figure supplement 1) and thus the volatility effect, which modifies the relative dilation observed when an outcome is received, may be less apparent for wins. Of course, this explanation leaves open the question as to why the pupillary response to similar magnitude losses and wins is asymmetric in the first place. A similar pupillary asymmetry has been reported in decision making tasks designed to assess behavioural loss aversion (Tversky and Kahneman, 1992; Yechiam and Telpaz, 2011). It may be, therefore that the greater pupillary response to both the occurrence of a loss and to loss volatility is related to the general overweighting of loss relative to win outcomes reported in the broader decision making literature. If correct, this would suggest that increasing the relative magnitude of the pupillary response to the receipt of a win relative to a loss (for example, by increasing the salience of wins by increasing their magnitude) would also increase the size of the pupillary response to win volatility.

The pupilometry measure included in the current study raises the possibility that estimated information content may be influenced by pharmacological as well as cognitive interventions. Pupil size is influenced by the activity of a number of central neurotransmitters including norepinephrine (Joshi et al., 2016) and previous work exploring the neural systems which control response to volatility have predicted a key role for NE (Yu and Dayan, 2005) suggesting it as an obvious pharmacological target. A single study has reported an effect of atomoxetine, a norepinephrine reuptake inhibitor, on learning in a volatile environment (Jepma et al., 2016) although no previous work has examined the effect of a pharmacological intervention on learning to positive vs. negative outcomes. It would be interesting to test whether a pharmacological manipulation of norepinepheric function was able to modify the outcome specific volatility effect demonstrated in this paper as such an effect may indicate a clinically useful interaction between pharmacological and cognitive interventions. A pharmacological approach could also be used to investigate related mechanistic questions such as the greater pupillary response to loss than win outcomes and the greater pupillary signal for loss outcome volatility discussed above. Specifically, a greater impact of a pharmacological manipulation on learning rates for losses than wins would provide experimental evidence for a preferential role for the NE system in estimates of the information content of losses.

The parallel representation of estimated information content of two distinct outcomes, provides a potential mechanism by which individuals may come to be generally more influenced by events of one class than another. This finding may be relevant to clinical questions. In the case of depression, patients have been shown to be more influenced by negative events, for example tending to remember more negative than positive events (Bradley et al., 1995), attend to negative more than positive events (Gotlib et al., 2004) and learn more from negative and less from positive outcomes (Eshel and Roiser, 2010). As the negative biases described above are believed to be causally related to symptoms of depression (Mathews and MacLeod, 2005), and interventions designed to alter negative biases can reduce symptoms (Browning et al., 2012; NICE, 2009), these results raise the possibility that novel interventions which target exstimated information content may act to alter negative affective biases and thus reduce symptoms of the illness. Of course, identifying potential targets for treatment and showing that they may be altered experimentally as done in this paper is only the first step in the development of new treatments. The next step, analogous to a phase 2a study in drug development (Ciociola et al., 2014), is to assess the initial efficacy of a potential intervention which engages the target in a clinical population. A study designed to do this is currently underway using the volatility manipulation described in this paper (study identifier NCT02913898).

While the results of the current study provide evidence that the information content of different events can be estimated in parallel during learning, the level of abstraction at which these estimates function is not clear. For example, does the task used in the current study alter the estimated information content of all positive and negative outcomes, or just those used in the task? This question is relevant to the potential clinical application of interventions which modify this estimate, as the affective biases associated with emotional disorders are seen across a wide variety of contexts (Mathews and MacLeod, 2005) and thus an intervention which modifies one particular instance of bias is unlikely to be useful therapeutically. It is clearly unlikely that completing a single block of a learning task, as done in the current study, will produce a broad and generalised alteration of the estimated information content of all outcomes. Rather, in order to test the degree to which alteration of estimated information content generalises it will be necessary to repeatedly expose participants to situations in which one class of outcome (e.g. positive) is more informative and then measure whether this alters learning performance in separate tasks and, ultimately, whether it impacts on clinical symptoms. The intervention used in the ongoing clinical study described above involves repeatedly completing the ‘positive volatile’ block from the current study over the course of two weeks (see; Browning et al., 2012 for a similar design) which will provide an initial assessment of this question.

The information content of an outcome is not solely a function of the volatility of its occurrence. Other factors, such as the strength of the association between a stimulus, or action, and the subsequent outcome, sometimes called the ‘expected uncertainty’ (Yu and Dayan, 2005) of the association, will also influence how informative the outcome is. Outcomes in the learning task reported in this paper vary in terms of both volatility and expected uncertainty, with both of these factors predicted to influence learning rate in the same direction (i.e. both factors should increase learning rate in the volatile blocks). An additional experiment (see Figure 2—figure supplement 2) in which volatility was kept constant but expected uncertainty varied found no effect on learning rate suggesting that the current findings were likely to be due to the effects of volatility rather than expected uncertainty. However, it would be interesting in future studies to explore whether it was possible to use manipulations of expected uncertainty, in the same way that volatility is used in this study, to induce a preference for positive over negative events. This may provide an alternative approach to engaging and altering expected information content than the volatility based effect reported here.

The current study demonstrates that human learners maintain separable estimates of the information content of distinct positive and negative outcomes and provides an initial proof of principle as to how these estimates may be modified. The study illustrates a little explored application of computational techniques in cognitive neuroscience; they may be used to identify potential novel treatment targets and by so doing spur the development of new and more effective treatments.

1. 1. Aston-Jones G
   2. Cohen JD

   (2005) An integrative theory of locus coeruleus-norepinephrine function: adaptive gain and optimal performance

   Annual Review of Neuroscience 28:403–450.

   https://doi.org/10.1146/annurev.neuro.28.061604.135709

   • PubMed
   • Google Scholar
2. 1. Behrens TE
   2. Hunt LT
   3. Woolrich MW
   4. Rushworth MF

   (2008) Associative learning of social value

   Nature 456:245–249.

   https://doi.org/10.1038/nature07538

   • PubMed
   • Google Scholar
3. 1. Behrens TE
   2. Woolrich MW
   3. Walton ME
   4. Rushworth MF

   (2007) Learning the value of information in an uncertain world

   Nature Neuroscience 10:1214–1221.

   https://doi.org/10.1038/nn1954

   • PubMed
   • Google Scholar
4. 1. Bradley BP
   2. Mogg K
   3. Williams R

   (1995) Implicit and explicit memory for emotion-congruent information in clinical depression and anxiety

   Behaviour Research and Therapy 33:755–770.

   https://doi.org/10.1016/0005-7967(95)00029-W

   • PubMed
   • Google Scholar
5. 1. Browning M
   2. Behrens TE
   3. Jocham G
   4. O'Reilly JX
   5. Bishop SJ

   (2015) Anxious individuals have difficulty learning the causal statistics of aversive environments

   Nature Neuroscience 18:590–596.

   https://doi.org/10.1038/nn.3961

   • PubMed
   • Google Scholar
6. 1. Browning M
   2. Holmes EA
   3. Charles M
   4. Cowen PJ
   5. Harmer CJ

   (2012) Using attentional bias modification as a cognitive vaccine against depression

   Biological Psychiatry 72:572–579.

   https://doi.org/10.1016/j.biopsych.2012.04.014

   • PubMed
   • Google Scholar
7. 1. Ciociola AA
   2. Cohen LB
   3. Kulkarni P
   4. FDA-Related Matters Committee of the American College of Gastroenterology

   (2014) How drugs are developed and approved by the FDA: current process and future directions

   The American Journal of Gastroenterology 109:620–623.

   https://doi.org/10.1038/ajg.2013.407

   • PubMed
   • Google Scholar
8. 1. Daw ND
   2. Gershman SJ
   3. Seymour B
   4. Dayan P
   5. Dolan RJ

   (2011) Model-based influences on humans' choices and striatal prediction errors

   Neuron 69:1204–1215.

   https://doi.org/10.1016/j.neuron.2011.02.027

   • PubMed
   • Google Scholar
9. 1. Eshel N
   2. Roiser JP

   (2010) Reward and punishment processing in depression

   Biological Psychiatry 68:118–124.

   https://doi.org/10.1016/j.biopsych.2010.01.027

   • PubMed
   • Google Scholar
10. 1. Gotlib IH
    2. Krasnoperova E
    3. Yue DN
    4. Joormann J

    (2004) Attentional biases for negative interpersonal stimuli in clinical depression

    Journal of Abnormal Psychology 113:127–135.

    https://doi.org/10.1037/0021-843X.113.1.121

    • PubMed
    • Google Scholar
11. 1. Jepma M
    2. Murphy PR
    3. Nassar MR
    4. Rangel-Gomez M
    5. Meeter M
    6. Nieuwenhuis S

    (2016) Catecholaminergic regulation of learning rate in a dynamic environment

    PLoS Computational Biology 12:e1005171.

    https://doi.org/10.1371/journal.pcbi.1005171

    • PubMed
    • Google Scholar
12. 1. Joshi S
    2. Li Y
    3. Kalwani RM
    4. Gold JI

    (2016) Relationships between pupil diameter and neuronal activity in the locus coeruleus, colliculi, and cingulate cortex

    Neuron 89:221–234.

    https://doi.org/10.1016/j.neuron.2015.11.028

    • PubMed
    • Google Scholar
13. Book

    1. MacKay DJ

    (2003)

    Information Theory, Inference and Learning Algorithms

    Cambridge: Cambridge University Press.

    • Google Scholar
14. 1. Mathews A
    2. MacLeod C

    (2005) Cognitive vulnerability to emotional disorders

    Annual Review of Clinical Psychology 1:167–195.

    https://doi.org/10.1146/annurev.clinpsy.1.102803.143916

    • PubMed
    • Google Scholar
15. 1. Nassar MR
    2. Rumsey KM
    3. Wilson RC
    4. Parikh K
    5. Heasly B
    6. Gold JI

    (2012) Rational regulation of learning dynamics by pupil-linked arousal systems

    Nature Neuroscience 15:1040–1046.

    https://doi.org/10.1038/nn.3130

    • PubMed
    • Google Scholar
16. Book

    1. NICE

    (2009)

    Treatment and management of depression in adults, including adults with a chronic physical health problem

    London: NICE.

    • Google Scholar
17. 1. Prelec D

    (1998) The probability weighting function

    Econometrica 66:497–527.

    https://doi.org/10.2307/2998573

    • Google Scholar
18. Book

    1. Rescorla RA
    2. Wagner AR

    (1972)

    A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement

    In: Black A. H, Prokasy W. F, editors. Classiacal Conditioning II: Current Research and Theory. New York: Appleton-Centuary-Crofts. pp. 64–99.

    • Google Scholar
19. 1. Rush AJ
    2. Trivedi MH
    3. Ibrahim HM
    4. Carmody TJ
    5. Arnow B
    6. Klein DN
    7. Markowitz JC
    8. Ninan PT
    9. Kornstein S
    10. Manber R
    11. Thase ME
    12. Kocsis JH
    13. Keller MB

    (2003) The 16-item quick inventory of depressive symptomatology (QIDS), clinician rating (QIDS-C), and self-report (QIDS-SR): a psychometric evaluation in patients with chronic major depression

    Biological Psychiatry 54:573–583.

    https://doi.org/10.1016/S0006-3223(02)01866-8

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

    (2016) Forming beliefs: why valence matters

    Trends in Cognitive Sciences 20:25–33.

    https://doi.org/10.1016/j.tics.2015.11.002

    • PubMed
    • Google Scholar
21. Book

    1. Spielberger CD
    2. Gorsuch RL
    3. Lushene RD

    (1983)

    Manual for the State-Trait Anxiety Inventory

    Palo Alto, CA: Consulting Psychologists Press.

    • Google Scholar
22. Book

    1. Sutton R
    2. Barto AG

    (1998)

    Reinforcement Learning

    Cambridge, Massachusetts: MIT Press.

    • Google Scholar
23. 1. Tversky A
    2. Kahneman D

    (1992) Advances in prospect theory: cumulative representation of uncertainty

    Journal of Risk and Uncertainty 5:297–323.

    https://doi.org/10.1007/BF00122574

    • Google Scholar
24. 1. Yechiam E
    2. Telpaz A

    (2011) To take risk is to face loss: a tonic pupillometry study

    Frontiers in Psychology 2:344.

    https://doi.org/10.3389/fpsyg.2011.00344

    • PubMed
    • Google Scholar
25. 1. Yu AJ
    2. Dayan P

    (2005) Uncertainty, neuromodulation, and attention

    Neuron 46:681–692.

    https://doi.org/10.1016/j.neuron.2005.04.026

    • PubMed
    • Google Scholar

Author details

1. Erdem Pulcu

   Department of Psychiatry, University of Oxford, Oxford, United Kingdom

   Contribution

   Formal analysis, Writing—original draft

   Competing interests

   No competing interests declared

2. Michael Browning

   1. Department of Psychiatry, University of Oxford, Oxford, United Kingdom
   2. Oxford Health NHS Foundation Trust, Oxford, United Kingdom

   Contribution

   Conceptualization, Formal analysis, Supervision, Funding acquisition, Writing—original draft

   For correspondence

   michael.browning@psych.ox.ac.uk

   Competing interests

   Received travel expenses from Lundbeck for attending conferences.

   "This ORCID iD identifies the author of this article:" 0000-0001-9108-3144

• 4,612

  Page views

• 600

  Downloads

• 49

  Citations

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