1
The confidence accuracy of pathological gamblers
its role in risky decision-making and sensitivity to monetary incentive
Author : S.K. Slagter, 10631496
University of Amsterdam; Graduate school of Psychology
supervision: Richard Ridderinkhof and Ruth van Holst
date: 28-07-2017
Highlights
Pathological gamblers show overconfidence, but this is similar to healthy controls Monetary incentive has no modularly effect on confidence accuracy, in PG and HC The less overconfidence, the more risk-seeking
Keywords
Pathological gambling Confidence accuracy VMPFC Risk attitude Monetary incentiveAbstract
Pathological gamblers (PG) are characterized by risky and disadvantageous decision-making, which partly relies on the functioning of the ventromedial prefrontal cortex (VMPFC). During decision-making a feeling of confidence arises, reflecting one’s estimation of choosing the correct decision. It is unknown whether PG have poor confidence accuracy (i.e. the mismatch between one’s confidence and one‘s actual performance) and whether this has a relationship with risky decision-making. Earlier research has suggested that the VMPFC plays an important role in subjective confidence. However, whether confidence is coded differently in the VMPFC of PG requires further exploration. To test confidence accuracy, PG and matched healthy controls performed a perceptual confidence task. This task was performed during a functional magnetic resonance imaging (fMRI) session, to explore the possible role of the VMPFC in subjective confidence. Furthermore, confidence accuracy was correlated to a measure of risky decision-making, assessed with the Choice at Risk Task.
2
1. Introduction
Even if gambling starts as a harmless entertainment, it has the power to become pathological in a minority of gamblers. Pathological gamblers (PG) experience a constant drive to gamble, despite the negative consequences on their quality of life, including their psychological well-being, social relationships and physical health (Loo et al., 2016; Mythily et al., 2011).
Pathological gambling is an addictive disorder as described in the new DSM-5 (American Psychiatric Organisation, 2013); this has made gambling disorder the first recognized behavioural addiction (Clark, 2014). Pathological gambling is characterized by risky and disadvantageous decision-making. Indeed, studies have consistently shown that PG pursue choosing potential high rewards with a low probability accompanied by high probabilities of punishments,over highly probable small rewards with a low probability for punishment (Brand et al., 2005; Goudriaan et al., 2006; Power et al., 2012; Kraplin et al., 2014; Lakey et al., 2007; Lawrence et al., 2009).
Role of confidence in decision-making
During decision-making a feeling of confidence about one’s decision will arise, derived from one’s subjective assessment of the probability that one’s action, statement, or choice is correct (Pouget et al., 2016; Yeung et al., 2012). This estimate, a meta-cognitive process, depends upon the available evidence (e.g. Kiani et al., 2014; Vickers & Packer, 1982;Kepecs et al., 2008). Because subjective confidence is often decisive for the choice we make, one can imagine that confidence accuracy is critical for decision-makers. Overestimating the probability of being correct (i.e. overconfidence) can be detrimental in situations involving decisions with significant consequences, e.g. monetary gains or losses. Not surprisingly, overconfidence is a feature that might be present in problematic gamblers and PG (Goodie et al., 2005). Such a miscalibration of confidence might explain why problematic gamblers gather and evaluate less information (e.g. evidence) before reaching a decision, hence they tolerate more uncertainty in their decisions than controls (Lawrence et al., 2009). In line with Goodie et al. (2005), Brevers et al. (2012) has suggested impaired meta-cognitive abilities in PG for a gambling task. More specifically, participants’ meta-cognitive ability was assessed by asking them to wager on their own decisions. They found that PG were inclined to wager high even when performing poorly on the Iowa Gambling decision-making task; this has
3 been interpreted as gamblers overestimating their performance. In a later study, Brevers et al. (2014) found a significant correlation between controls’ performance on a grammar task (i.e. based on learned grammar rules, subjects needed to indicate whether a string of letters obeyed the learned rules) and their confidence, which was absent for PG. This suggests a disconnection between performance and subjective confidence in PG, also for a
non-gambling task. However, research on confidence accuracy in pathological gamblers is sparse. Thus, the present study aims to examine whether confidence accuracy is impaired in PG. In situations where the outcome is unknown and risk is involved, confidence might be a key factor in decision-making. Interestingly, Murad et al. (2016) showed that risk aversion is associated with reduced confidence. Gamblers have shown risky behaviour during various decision-making tasks (e.g. Goudriaan et al., 2006; Power et al., 2012) and increased risk attitude has shown to be positively correlated with gambling severity (Ligneul et al., 2013). Because risky behaviour is highly prevalent in gamblers and knowledge about the role of (high) confidence in pathological gambling is lacking, this study will also investigate whether overconfidence is associated with the risky behaviour of PG. The result would provide insight into a possible factor mediating risky decision-making in PG and could potentially inform the development of new treatments for PG.
Neurobiology of confidence
Recent studies suggest that the neural encoding of confidence during decision-making centers on the ventromedial prefrontal cortex (VMPFC; De Martino et al., 2013; Lebreton et al., 2015). A previous study already indicated a link between this brain area and the ability to discriminate correct from incorrect decisions, based on one’s subjective confidence feeling (Fleming et al., 2010). However, exploring confidence coding in the brain is still in its infancy. In line with its potential role in confidence coding, the VMPFC has already shown to be involved in (risky) decision-making. Meta-analyses of functional MRI studies illustrate that the VMPFC represents value reappraisal of the choices during decision-making (Bartra et al., 2013; Levy & Glimcher, 2012). Patients with focal VMPFC damage show elevated betting tendencies and diminished sensitivity to the level of uncertainty during decision-making under risk (Clark et al., 2008) and ambiguity (Bechara et al., 2000). These patterns of
responding are very similar to the behaviour observed in PG (Zois et al., 2014); this opens the possibility that VMPFC dysfunction plays a role in the poor decision-making of PG.
4 However, the effect of aberrant VMPFC functioning on subjective confidence has not been explored yet. Since, poor confidence accuracy is expected to be present in PG and might contribute to risky decision-making, it would be useful to investigate whether PG show aberrant VMPFC activity, related to confidence coding.
Modulatory effects on confidence accuracy
Confidence accuracy seems to be modulated by factors such as mood (Koellinger et al., 2015), emotional arousal (Jönsson et al., 2005; Allen et al., 2016), desires (Giardini et al., 2008), worry (Massoni, 2014) and monetary incentive (Lebreton et al., 2017). Lebreton et al. (2017) found increased subjective confidence ratings when healthy subjects could potentially gain money compared to the condition wherein they could potentially lose money. It would be relevant to know if and how monetary incentive modulate the confidence accuracy of gamblers, since gambling is all about choices leading to monetary gains or losses. However, no study has investigated this in PG. PG have shown aberrant sensitivity for rewards and punishments compared to controls, which has been linked to gambling severity (De Ruiter et al., 2009; Kreussel et al., 2013; Gaher et al., 2015; Mackillop et al., 2014; Jiménez-Murcia et al., 2016). This leaves it still speculative if and how the modulatory effect of monetary incentive is present in PG, leading to this studies’ fourth question of interest.
The Present Study
We hypothesize that PG will have impaired confidence accuracy compared to healthy controls (HC), driven by overconfidence. Furthermore, we expect to find activity in the VMPFC during decision-making, that is positively linearly related to subjective confidence. Moreover, due to inaccurate subjective confidence in PG, we expect an aberrant relationship between VMPFC activity and subjective confidence in PG compared to HC.
Moreover, a positive correlation is expected between subjective confidence and risky behaviour; the more overconfidence, the more risk-seeking that person will be. PG will be more confident and risk-seeking than controls.
Whether confidence accuracy is affected by monetary incentive, is mostly exploratory. Based on the behavioural study of Lebreton et al. (2017), we expect that monetary incentive will affect HC is such a way that potential gains will increase overconfidence and potential losses will decrease overconfidence. However, we expect that this modulatory effect will be
5 present to a smaller extent or even absent in PG, because hyposensitivity for rewards and punishment have been found for this group, when playing a gambling task (Lole et al., 2015; Goudriaan et al., 2006; De Ruiter et al., 2009). However, Brevers et al.’s (2013) review shows that PG have a persistent preference toward high, immediate, and uncertain rewards despite experiencing larger losses. Based on this, one can hypothesize that PG’ subjective confidence will increase for potential gains (as in HC), but will be unaffected by choices leading to potential losses.
To test these hypotheses, PG and HC will perform a perceptual decision-making task, including subjective confidence ratings about their perceptual decisions. The reported confidence ratings (i.e. subjective confidence) will be matched to the overall performance on the task to measure confidence accuracy. The task will be performed in the MRI-scanner to explore whether the VMPFC is related to subjective confidence. Confidence ratings will be used as parametric modulator, to extract regions, whose blood oxygenation level dependent (BOLD) activation vary linearly with subjective confidence. Furthermore, this additional fMRI measure could indicate whether the confidence of PG is coded aberrant in the VMPFC. Subjects’ risky attitude will be measured during a risky decision-making task, using
computational modelling with a variational Bayesian approach (VBA). The relation between risk attitude and confidence accuracy will be tested using a correlational method.
2. Method
2.1 Sample
16 pathological gamblers and 30 healthy controls between 18 and 65 years participated in this study. Only gamblers who met the DSM-5 criteria for gambling disorder and who had gambled in the last 6 months were included in the study. PG were recruited from the Jellinek Addiction Treatment Center (Amsterdam) and HC were recruited through local
advertisement. PG and HC were matched on age and sex.
Based on the matching and the quality check of the behavioural data (see Methods) only 13 PG and 19 HC were included in the behavioural analysis. Furthermore, 4 PG were excluded based on the quality check of the fMRI data, resulting in 19 HC and 9 PG for the fMRI analysis. All PG scored above 8 on the problematic gambling severity index (PGSI), indicating severe gambling problems. Both groups had no (other) psychiatric disorders,
6 indicated by the mini international neuropsychiatric interview (MINI; Sheehan et al., 1998). The Fagerstrom Test for Nicotine Dependence (FTND; Heatherton et al., 1991) was used to quantify the nicotine dependence severity of the smoking participants. All participants retained at least 3 days from other drug use and psychotropic medication prior test session (according to self-report). All participants had normal or corrected to normal vision. Participants were excluded when having an IQ below 80, MRI-contraindications or an insufficient command of the Dutch language. All participants signed informed consent before participating in this study, which was approved by the Medical Ethics Committee of the Academic Medical Center Amsterdam. A detailed overview of the included sample its characteristics is presented in table 1.
Table 1. Sample characteristics
Clinical and psychometric characteristics for pathological gamblers and healthy controls. Groups were compared by mean with an independent-samples t-test.
Pathological gamblers Healthy controls
N (male/ female) 12/1 17/2
Age 34.36 (13.72) 39.16 (15.20)
Education level (N) Primary school 1 0
High school 6 3 MBO 7 6 HBO 0 6 WO 0 5 PGSI 17.38 (2.63) 0.0 (0.0) ** Cigarette use (n) 5 2 FTND 3.2 (2.168) 1.00(0.0)
Means (SD). SD, standard deviation; PGSI, Problem Gambling Severity Index; FTND, Fagerstrom Test for Nicotine Dependence.**, p < 0.0001.
2.2 Procedure
Behavioural and neurobiological outcome measurements were used to explore similarities and differences between PG and HC regarding confidence accuracy, confidence codingand risky attitude.
The MINI, PGSI, FTND questionnaire and demographics were questioned before the test day (used for screening), and at the start of the test day. After the questionnaires,
7 their confidence (on a scale from 50% to 100%) about their immediately preceding decision, the so called perceptual confidence task (Fig.1). Two sessions of the perceptual confidence task were performed in the MRI-scanner. Prior to the test-session, subjects performed two practice sessions, one outside the scanner and one inside the scanner, of approximately 3 minutes per session. This practice allowed subjects to get used to the fast representation of the stimuli and the button boxes, used inside the scanner.
After scanning, in case of a participant with gambling problems, an interview, regarding the quality of their compulsivity, was conducted. Lastly, risky attitude was measured with the Choice at Risk task behind a computer screen.
2.3 Perceptual confidence task
The perceptual confidence task was derived from the study of Fleming et al. (2010) and Lebreton et al. (2017), with only minor modifications regarding incentive and timing to make the task applicable for fMRI. The task was implemented using MATLAB® (MathWorks) and the COGENT toolbox (http://www.vislab.ucl.ac.uk/cogent.php). This task can reveal
individual differences in the ability to discriminate correct from incorrect choices (Fleming et al., 2010), and is sensitive to the modulatory effect of monetary incentive on subjective confidence, at least among HC (Lebreton et al., 2017).
2.3.1 task design
During this task participants viewed a Gabor patch (150 ms), displayed on both sides of the screen, and needed to judge which of the two patches had the highest contrast (self-paced; Fig. 1). After their choice, an incentive cue was presented for 900 ms, which varied among the conditions. In the gain condition the cue indicated that participants could potentially gain 100 points (framed in green), whereas in the loss-condition the cue indicated a
potentially loss of 100 points (framed in red). In the neutral condition the cue was 0 points (framed in grey), indicating there was nothing to gain or lose. The cue was presented after the choice to prevent performance from being affected by motivation. After the cue,
participants were asked to report their confidence about their choice, using a confidence bar ranging from 50 to 100%. Participants could move a cursor with the left and right button box to select their desired answer (self-paced). Participants were instructed that 50% meant the choice was just a guess, whereas 100% meant they were totally sure about the choice.
8 The trial ended with feedback (900 ms), to maximize their feeling of engagement in the task. The feedback informed participants whether they gained or lost points (depending on the condition), which would be converted to money at the end of the task. This was mentioned explicitly to the participants, to introduce monetary incentivization and arouse motivated participation. During the gain or loss-trials, participants gained or did not lose the points in case of a correct trial, and would either lose or not gain the points in case of an incorrect trial. However, participants were instructed that they could earn the most points (and thereby the most money) when they reported their confidence as accurately as possible.
Each session of the task, consisted of 24 trials of each condition (i.e. -100, 0 and 100 points), with a condition manipulation on a trial-by-trial basis. A random inter-trial interval (ITI) of minimal 4500 ms and maximal 6000 ms, ensured that participants could not predict the occurrence of the stimuli. The duration of one session was approximately 20 minutes. 2.3.2 Generated stimuli
Prior to the perceptual confidence task, participants performed a 144 trials calibration session, later used to adjust the main task stimuli per subject. This calibration session included the discrimination of the Gabor patch with the highest contrast, without the incentive and subjective confidence rating (Fig. 1). During the calibration, contrast
differences between the Gabor patches (i.e. difficulty) were adapted every 12 trials, such that an overall performance of approximately 75% correct was reached (Lebreton et al., 2017). Setting a fixed performance accuracy for every participant would allow participants to deviate from their performance, regarding their subjective confidence, in a similar way as other participants; max. 25% overconfidence or max. 75% underconfidence could be reached by each participant (see Supplementary materials, for more details about the calibration).
9
Figure 1. Trial within the perceptual confidence task and the calibration session
Participants had to judge which Gabor patch had the highest contrast (moment of choice). Thereafter participants were presented with an incentive cue, which varied among the conditions (100 points in the gain condition, -100 points in the loss condition and 0 points in the neutral condition). In addition, a confidence scale from 50 to 100% appeared on the screen, and participants were asked to report their confidence about their preceding choice, by using the scale (rating moment). The trial ended with feedback, which informed participants whether they gained or lost points. The calibration session before the actual task included only the Gabor patch stimuli and the moment of choice.
2.3.3 Task measurement
Confidence accuracy was calculated by taking the difference between the averaged
subjective confidence and the corresponding average performance, resulting in a deviation score, expressed in percentage:
Where n is the total number of trials, Ck is the rated subjective confidence at trial k, and Pk is
10 A confidence accuracy of zero indicates a perfect match between the overall subjective
confidence rating and the overall objective performance, whereas a negative or positive percentage reflects overconfidence or underconfidence, respectively (i.e. mismatch). The higher one’s ability to match subjective confidence onto one’s objective performance, the more accurate their subjective confidence, indexed by zero. Confidence accuracy was assessed for each condition separately, by taken the mean subjective confidence rating and mean performance of the condition specific trials only.
2.4 Choice at Risk task
The Choice at Risk task was designed in MATLAB® (MathWorks). In addition, the VBA toolbox (http://mbb-team.github.io/VBA-toolbox/) was used to implement a variational Bayesian approach (VBA) to interpret the data through computational modelling. The following functions were used within the VBA toolbox: performing efficient and robust parameter estimation on nonlinear models and optimizing the experimental design in the aim of maximizing the statistical power of model-based data analysis.
2.4.1 Task design
In this computer task, participants were presented with two lotteries on each of 100 trials: a reference lottery and a lottery of varying winning and losing probabilities (introducing risk), next to varying amounts of potential gains and losses. The reference lottery always displayed 50% probability of winning 5 points and a 50% probability of losing 2 points (Fig. 2).
Participants had to choose a lottery. Between consecutive trials, a fixation cross was
presented. Participants were instructed that they could win extra money, depending on the amount of points they earned during the game, to promote active participation and guide participants to make decisions that maximize their earnings.
11
Figure 2. Different decisions during the Choice at Risk Task
Two trials of the Choice under Risk Task are presented. In each trial a reference and variable lottery were presented as a circle painted partly red and green, indicating the winning and losing probability respectively. Participants were asked to choose one of the lotteries. No explicit instructions were given, only that they should be guided by their intuition and that the earned points would be converted to money.
2.4.2 Task measurements
During the task a utility function v(x) was modelled based on the behavioural choices, to obtain a subject-specific utility function. This utility function represents the perceived satisfaction of a person when gaining or losing a certain number of goods (i.e. points). The following equation, proposed by Tversky and Kahneman (1992) was used to model the utility function:
v(x)= x α if x ≥ 0 (F.1)
- λ (-x)β if x ≤ 0
where v measures the subjective value of the consequence x and parameters α, β and λ are constant. The curvature of the function is determined by parameter α(β). Concavity in the function illustrates diminishing marginal utility for greater goods (e.g. when a decision maker would receive $50, the utility he gains with this monetary stake is more than half the utility gained by receiving $100). In contrast, a convex utility function implies increased obtained marginal utility for greater goods. Importantly, the utility curvature captures a subject-specific risk attitude (Chumbley et al., 2014; Bombardini et al., 2012): concavity in the utility function (α(β) < 1) reflects a risk-averse attitude, whereas a convex utility function
12 (α(β) > 1) reflects risk-seeking.
Next to the curvature of the utility function, subjects’ loss aversion was estimated. This was also derived from the utility function including the parameter λ, which is the coefficient of loss aversion.Loss aversion is present when losing a number of goods has a greater psychological weight than winning the same number of goods (losses loom larger than gains; Kahneman & Tversky, 1992). In the model, the parameter λ captures the valence symmetry in subjective value. Here, λ = 1 would indicate no valence difference between losses and gains, λ > 1 would indicate loss aversion and λ < 1 would indicate loss-seeking. Loss aversion is of interest, because it could contribute to risk attitude.
2.4.3 Generated stimuli
During the task the combination of lottery stimuli for each trial was selected on the basis of participant’s previous decisions. The design of the task was updated on this trial-by-trial basis to get the most efficient design for estimating subject-specific parameters. This was implemented using the VBA_designEfficiency() function of the VBA toolbox.
2.5 fMRI data acquisition
Participants lay in a 3T MRI-scanner at the Amsterdam Medical Centre (AMC) with their head in a standard 8-channel whole-head coil. A T1 structural scan was acquired to use as anatomical reference with the following parameters: TR of 7 s, TE of 3.2 ms, 150 slices, slice thickness of 1 mm, field of view (FOV) of 256x 237x 180 mm, with a transverse slice
orientation. A multi-echo gradient-echo planar imaging (EPI) sequence was used to measure the BOLD signals during the perceptual confidence task. The specifics of the sequence were set as follows: TR 2.375 s, TEs for 3 echoes: 9ms, 26.4ms and 43.8 ms, voxel size of 3 mm, 37 slices, slice thickness of 3 mm, FOV of 224x 224x 122 mm, slice gap of 0.3 mm, with an interleaved slice acquisition. The multi-echo EPI sequence was used to reduce image distortion and increase BOLD sensitivity in regions that are more affected by scan artefacts, such as the VMPFC (Fernandez et al., 2017; Bhavsar et al., 2014).
13
3. Data analysis
3.1 Quality check of the behavioural data
The behavioural data had been checked for extreme high and low numbers of left presses, average accuracy and whether participants actively indicated their confidence, by checking whether they changed the initial position of the cursor of the confidence bar. Participants were excluded when their accuracy turned out to be lower than 60% or higher than 85%, to avoid large differences in the max. possible overconfidence and underconfidence each participant could reach. Based on this criterion, two HC and one PG were excluded.
Furthermore, two HC were excluded based on inactive participation during the confidence rating, indexed by the mean change of the initial position of the confidence bar cursor.
3.2 Statistical analysis of the behavioural data
Confidence accuracy and the effect of monetary incentive
Repeated measures ANOVA was performed, with group as between factor (PG vs. control), monetary incentive as within-factor (gain, neutral and loss) and confidence accuracy as dependent variable.
Correlation between confidence accuracy and risk attitude
The Pearson product-moment correlation was computed to investigate whether there was a significant linear relationship between confidence accuracy (irrespective of incentive
condition) and risk attitude (i.e. parameter α). Firstly, the group difference in risk attitude was explored to determine whether the correlation analysis could be ran on the whole sample, in case of no group difference. Separate independent-samples t-tests were ran to compare the risk attitude (α) and loss aversion (λ) between groups.One PG and one HC were excluded from the analysis because of their missing data for the Choice at Risk Task.
3.3 fMRI pre-processing
fMRI pre-processing and its statistical analysis were completed in MATLAB R2016b, using the software: Statistical Parametric Mapping 12 (SPM12). Weights for a BOLD contrast-to-noise ratio map (CNR map) for each echo was estimated, based on the first thirty volumes acquired before the start of the task. Weighted summation was then used to combine all
14 three echoes into a single data set. Non-brain tissue was removed during segmentation. The fMRI data was realigned to the 31st volume with a linear rigid body transformation of 6
degrees of freedom (DOF; 3 translations and 3 rotations). Next, fMRI data was spatial smoothed with a full-width at half-maximum Gaussian kernel of 6 mm. Furthermore, participant’s fMRI data was co-registered to their T1-weighted structural image. As a next step, fMRI data was spatially normalized to 3 mm MNI-space (Montreal Neurological
Institute), using the 3 mm MNI152 standard-space T1-weighted brain as reference. Successful normalization was subjectively inspected.
3.4. Quality check of the fMRI data
Functional data was checked for signal intensity spikes to assess the quality of the acquired functional volumes. Furthermore, the motion of the participants was manually checked by visualizing the motion parameters in the x, y and z-direction. Data of three PG were
excluded from fMRI analysis because their movement exceeded far above 3.5 mm. Still, some participants showed artefacts in certain functional volumes, due to sudden motion in
combination with interleaved scanning. In order to reduce these systematic motion-related artefacts, known to be especially prevalent in patient-control studies (Power et al., 2012), the Art-Repair toolbox (Mazaika et al., 2009) was used. This SPM toolbox can detect and repair bad volumes within the preprocessed data, which were identified by excessive scan-to-scan motion or large variations in average global intensity.
All volume outliers, after setting a threshold of 1.5% variation from the mean BOLD signal intensity, were repaired with the interpolation technique. The number of participants whose certain volumes needed to be repaired were 8 in the PG group (average number of repaired volumes = 21) and 6 in the HC group (average number of repaired volumes = 6). In a later step, after the first-level analysis, the global quality of the intervention was checked. This was done by calculating the reduction in standard deviation (SD) of the estimation error, for three important contrasts of the first-level analysis that was based on the repaired data, compared to the use of the original data. Average improvement in SD was 7.28% for the PG group and 1.32% for the HC group.
15 (also known as the residual of your fitted model).
3.5 Statistical analysis of the fMRI data
GLM design
fMRI data were statistically analysed with a two-level procedure using the general linear model (GLM) as implemented in SPM12. At the first-level, for each participant incentive regressors were constructed, using the onsets and durations from individual task-logfiles. These incentive regressors were modeled for two events separately, namely during the moment of choice and during confidence rating. Thus, the full model contained 6 ( 3 incentive conditions x 2 moments of interest) event-related regressors. In addition, to investigate the parametric effect of individual subjective confidence levels on the VMPFC activation, each regressor was linearly modulated with a parametric modulator (pmod), namely the subjective confidence rating of the specific trial. Furthermore, accuracy (correct vs. incorrect answer) was included as a second parametric modulator. However, this pmod is beyond the scope of in this paper in the second-level analysis. The estimated realignment parameters were also included as regressors of non-interest to account for head motion. Next, a set of hidden regressors were included in the model which functioned as the high pass temporal filter with a cut-off of 128s, to filter low-frequencies from the signal. An overview of the GLM design is displayed in figure 7 (see Supplementary materials).
Contrasts
All contrasts of regression coefficients were estimated at subject level. Next, individual contrast estimates of all subjects were entered in second-level analysis.
Firstly, to check whether our incentive manipulation worked, three contrasts were set at the moment of confidence rating. Gain-trials (100) vs. neutral-trials (0) and gain vs. loss-trials (-100) were compared. Furthermore, loss-trials were compared to neutral-trials. As this was done as a proof of principle we focused only on the HC group. These contrasts were
analysed at whole-brain level, with special interests in the striatal response (Delgado et al., 2005).
Furthermore, to investigate the BOLD activation related to subjective confidence, a contrast was set to compare the moment of choice modulated by the pmod subjective confidence rating (irrespective of incentive condition) to an implicit baseline1. This was
16 tested with a one-sample t-test for the HC and PG group separately. Activation maps were inspected at whole-brain level.
Next, this contrast was used to investigate between-group differences in VMPFC activation related to subjective confidence. This was tested with an independent sample t-test. To create the ROI for this analysis, an 8 mm sphere around the MNI coordinates of the VMPFC cluster (x,y,z: -2, 52, -2 mm), as indicated in the study by Lebreton et al. (2015), was used. This mask was created with the MarsBaR toolbox in SPM12.
To determine significant voxel activation, an uncorrected p < 0.001 intensity threshold was used for all voxels, and followed by a threshold at cluster level using family-wise error
(FWE) correction. Only activation below pFWE_clu < 0.05 was reported as significant. For the
ROI analysis, a small volume correction was used based on the applied mask. Significant activation at a FWE-corrected threshold of p < 0.05 at voxel level was reported. The
anatomical labelling (aal) function of the WFU Pickatlas 2.5 toolbox in SPM12 was used for classifying the brain regions that showed significant BOLD activation.
4. Results
4.1 Group differences in confidence accuracy and the effect of monetary incentive The Mauchly’s test indicated that the assumption of sphericity was met ( X2 (2) = 4.639,
p < 0.05) for the within-subject variable. The Shapiro-Wilk test indicated no violation of normality for all conditions within each group. Equal variance across groups was not met for the data in the gain condition only. Because there is no non-parametric version of the
repeated measures ANOVA and data-transforming would not preserve negative values for confidence accuracy, the planned ANOVA maintained the best option for analysis.
The repeated measures ANOVA showed no significant main effect of group on confidence accuracy, F(1,30) = 0.284, p = 0.598 (Fig. 3). There was also no significant main effect of monetary incentive on confidence accuracy, F(2,60) = 1.809, p = 0.173 (Fig. 3).
17 Furthermore no interaction was found between group and incentive on confidence accuracy, F(2,60) = 0.793, p = 0.457 (Fig. 3).
Figure 3. Confidence accuracy of PG and HC in three different incentivized conditions.
Displayed is the computed confidence accuracy of PG and HC over three incentivized conditions (loss, neutral and gain). A confidence accuracy of zero indicates a perfect match between overall subjective confidence level and overall performance. A positive or negative percentage indicates a mismatch; negative indicates underconfidence, i.e. one’s subjective confidence level is lower than one’s performance, whereas a positive percentage indicates overconfidence, i.e. one’s subjective confidence level is higher than one’s performance. Mean scores with SD error bars are displayed.
4.2 fMRI results
Incentive effect
The manipulation check of the effect of monetary incentive in the HC group, revealed a significantly higher BOLD activation of the right caudate nucleus during rating, when comparing gain-trials to neutral-trials. Furthermore, the right inferior temporal lobe and the right inferior part of the parietal lobe showed also more BOLD activation during gain-trials compared to neutral-trials. In addition, significantly higher activation was also found in the superior, medial and inferior frontal gyrus and cingulate gyrus (see Fig. 4 and Table 2 for specifics). When comparing the loss-trials to neutral-trials, the inferior parietal lobe and
-15 -10 -5 0 5 10 15 20 25 30 Confi d en ce a cc u ra cy ( %)
lose (-100 ) neutral (0) gain (100)
Incentive (points)
PG HC
18 inferior frontal gyrus showed again significantly higher BOLD activation. This was
accompanied by higher BOLD activation in the right supplementary motor area and angular gyrus (see Fig. 4 and Table 2 for specifics). No significantly higher BOLD activation was found in the brain during the moment of rating in the gain-trials compared to the rating in the loss-trials and vice versa.
Table 2. fMRI results of the incentive effect in HC at whole-brain level.
Listed below are the regions that showed a significant BOLD activation during the moment of rating when comparing the incentive conditions.
Contrast Brain region Cluster size
(k)
T x y z a pFWE_clu
Gain > neutral Caudate nucleus Right inferior temporal lobe
Right inferior parietal lobe Superior frontal gyrus
Medial frontal gyrus Inferior frontal gyrus
Cingulate gyrus 142 1093 1199 43 560 43 41 5.64 9.97 7.91 7.45 7.39 6.18 5.23 9 11 -4 54 -52 -10 -39 -46 44 39 14 56 6 23 44 51 11 14 6 -31 29 < 0.0001 < 0.0001 < 0.0001 0.029 < 0.0001 0.029 0.035 Loss > neutral
Inferior parietal lobe Inferior frontal gyrus
Supplementary motor cortex Angular gyrus 214 157 111 45 398 8.12 7.26 5.77 5.24 6.17 -36 -46 44 48 11 26 39 26 14 3 20 47 36 -52 44 < 0.0001 < 0.0001 < 0.0001 0.034 < 0.0001 Gain > loss Loss > Gain
a Coordinates in Montreal Neurological Institute (MNI) space (x, left to right; y, posterior to anterior; z, inferior to superior). MNI-coordinates and T values are shown of local maxima for each significant cluster. FWE corrected p-values at cluster level are reported.
19
Figure 4. Effect of different incentives in the brain.
For the contrast gain vs. neutral, HC showed significant BOLD activation in the caudate nucleus and varies frontal and parietal areas. Varies frontal and parietal areas showed also more BOLD activation in the loss-trials, compared to neutral. Results are depicted with a FWE-corrected threshold of p < 0.05 at cluster level. The colour bar represents corresponding T values. IPL, Inferior partietal lobe; IFG, Inferior frontal gyrus; SFG, Superior frontal gyrus; MFG, medial frontal gyrus; SMA, Supplementary motor area.
Brain activation related to subjective confidence level in HC and PG
In HC, at whole-brain level, BOLD activation in the left putamen (peak MNI coordinates [-30,-13,-1] mm , k = 230, T = 6.15, pFWE_clu < 0.0001) was linearly related to subjective
confidence level (Fig. 5). Furthermore, the right amygdala ( peak MNI coordinates [30,-4,-19] mm , k = 137, T = 6.43, pFWE_clu < 0.0001 ) and precentral gyrus (peak MNI coordinates
[-30,-31,62] mm , k = 135, T = 6.76, pFWE_clu < 0.0001) also vary linearly with subjective confidence
level (Fig. 5).
In PG, at whole-brain level, no significant BOLD activation was found to have a linear
relationship with subjective confidence level. Hence, in both groups BOLD activation in the VMPFC was not modulated by subjective confidence.
20
Figure 5. Brain activation related to subjective confidence in HC.
HC showed significant BOLD activation in the left putamen, right amygdale and precentral gyrus, linearly related to subjective confidence levels. Results are depicted with a FWE-corrected threshold of p < 0.05, at cluster level . The colour bar represents corresponding T-values.
Group difference in VMPFC activation related to subjective confidence
The ROI analysis indicated no group difference in VMPFC activation, linearly related to subjective confidence level, when using a FWE-corrected threshold of p < 0.05 at voxel level.
4.4 Group differences in risk attitude and loss aversion
To test for group differences in risk attitude, as indexed by parameter α, the non-parametric Mann-Whitney U test was conducted, because the assumption of normality was violated. There was no significant difference in risk attitude between the PG group and HC group, ( U = 105, Z = -0.127, p = 0.917). Both PG ( 0.78 ± 0.6) and HC (0.73 ± 0.5) showed to be risk-averse, indicated by α < 1.
The Shapiro-Wilk test indicated that the λ parameter, a measure for loss aversion, was normally distributed for both groups. However, the Levene's test turned out significant, indicating unequal variances between groups. Degrees of freedom were adjusted, using the Welch-Satterthwaite method, to compensate for this violation.PG (0.91 ± 0.57) showed to be loss-seeking ( λ < 1), while HC (2.3 ± 1.31) showed loss aversion (λ > 1). Test results indicated that PG were significant less loss averse than HC, t(28) = 3.52, p = 0.0001.
21
4.5 Correlation between confidence accuracy and risk attitude
The Shapiro-Wilk test indicated that the α parameter was not normally distributed. Therefore, the Spearman's rank-order correlation (non-parametric equivalent), with one-tailed testing (p < 0.01) was ran to determine the relationship between risk aversion and confidence accuracy. Since the Spearman’s correlation is not very sensitive to outliers, one detected medium outlier was kept in the data. The analysis was performed on the whole sample, because there were no significant group effect on risk attitude.
There was a moderate, negative correlation between confidence accuracy and risk attitude, which was statistically significant (rs = -0.427, p = .009). Hence, the more risk-seeking, the less overconfidence (Fig. 6).
Figure 6. Relationship between confidence accuracy and risk attitude
The scatterplot shows the correlation between subject’s risk attitude, as indexed by the α parameter, and confidence accuracy. α < 1 reflects a risk-averse attitude, whereas α > 1 reflects risk-seeking. A confidence accuracy of zero indicates a perfect match between overall subjective confidence level and overall performance. A positive or negative percentage indicates less confidence accuracy; a negative score indicates underconfidence, i.e. one’s subjective confidence level is lower than one’s
performance, whereas a positive score indicates overconfidence, i.e. one’s subjective confidence level is higher than one’s performance. The blue dots represent the PG individuals and the red dots indicate the HC individuals.
22
5. Discussion
The aim of the current study was to investigate whether confidence accuracy is impaired in PG and whether this is reflected by aberrant VMPFC activation, compared to HC. In
addition, we investigated whether confidence accuracy is modulated by monetary incentive in PG and HC. Our last interest was to investigate the association between overconfidence and risk attitude.
Confidence accuracy in PG and HC
In contrast to our hypothesis, PG did not show significant more overconfidence than HC. Numerically, PG seemed slightly more overconfident in the gain condition compared to the neutral and loss condition, however this did not turn out to be significant. This is not surprising, given the great variance (indexed by the SD) in confidence accuracy in both groups. This resulted in a lower power to detect group differences. In our study the age between our participants varied a lot (although the groups were matched on age), the youngest participant was 18 and the oldest 64 years. Wong et al. (2012) assessed the confidence accuracy of old adults and young adults on a episodic memory task and found age-declined confidence accuracy. Diminishing confidence accuracy with age has also been found for our perceptual confidence task (Palmer et al., 2014). The role of age in confidence accuracy might be an explanation for the great variance in confidence accuracy. A solution for increasing the power of our study and decreasing the within-group variance would be to include more participants and to make age subgroups within the PG and HC group.
Incentive effect on confidence accuracy
No modulatory effect of monetary incentive on confidence accuracy was found in PG. Surprisingly, this effect was also absent in the HC group. This finding contradicts with the study of Lebreton et al., (2017) who replicated in three independent samples a biasing effect of monetary incentive on subjective confidence in HC; loss prospects improved confidence accuracy, but gain prospects elicit more overconfidence. This effect was found in a similar sample size (n = 24), with the same task design. Still, there are some differences that could explain why we did not find this monetary incentive effect in our study. Our sample was on
23 average 10 years older, compared to the sample of Lebreton et al. (2017) who included only young adults. It might be possible that the sensitivity of confidence for monetary incentive varies among ages, with elderly being less sensitive to monetary incentive. However, there are no studies that could (dis)confirm this. Including age subgroups would be interesting for future analysis. Another reason for the discrepancy could be the use of different incentive cues. Our study used the visualization of winning and losing points as incentive cue, whereas Lebreton et al. (2017) used 1 euro images as cue. The direct representation of winning or losing money might have had more valence and reminded participants more about the real monetary payment. However, we found an incentive effect in the brain that suggest that our task elicited valence differences, rendering this account less plausible.
Incentive effect in the brain
During the moment of rating, differences in BOLD activation were found between the incentive conditions. The gain-trials, compared to the neutral-trials, elicit expected activation in the caudate nucleus. This region of the reward-system has shown to be involved in
preceding and anticipating reward in learned visuomotor paradigms (Hikosaka et al., 1989). The role of the caudate nucleus in reward-related learning has also been confirmed by human studies (Delgado et al., 2005; Haruno et al., 2004). Furthermore, the caudate nucleus is essential for goal-directed behaviour (Grahn et al., 2009; Grahn et al., 2008), which might indicate that participants felt more engaged in rating their subjective confidence during the gain-trials, compared to the neutral-trials. Higher BOLD activation of the caudate nucleus was not found when comparing the gain condition with the loss condition, this suggests some activation of the caudate nucleus when one could potentially lose points. This is not surprising, since the caudate nucleus responds to changes in motivational context, rather than reward only (Delgado et al., 2004). Lastly, both gain and loss-trials showed to induce more activation in regions of the frontoparietal network, such as the inferior frontal gyrus (IFG), superior frontal gyrus (SFG) and inferior parietal lobe (IPL). This frontoparietal network is responsible for visuospatial attention (Corbetta et al., 2011; Simpson et al., 2011). This might indicate that participants were more attention-oriented during the rating moment of the gain and loss-trials compared to neutral-trials. These results suggest that the incentive cues (i.e. 100 and -100 points) had a higher valence compared to the neutral cue. Thus, the
24 observed effects of incentive on the BOLD activation, suggest that the incentive manipulation in our design succeeded. Performance was not affected by our incentive manipulation: the average accuracy scores were 75%, 75.3% and 73.8% for the gain-, neutral- and loss-trials respectively.
Activation modulated by subjective confidence level
This study showed that the BOLD activation in the left putamen, the right amygdale and the precentral gyrus varied with subjective confidence levels; BOLD activation in those regions linearly increased with increased subjective confidence. To our best knowledge, the function of these regions have not yet been related to subjective confidence. The putamen has shown to be involved in reward prediction (Schultz et al., 1998) and habitual behaviour (Tricomi et al., 2009; Seger et al., 2011), whereas the amygdale is involved in the process of emotions, such as happiness (Gur et al., 2002; Phelps et al., 2005), next to emotional learning, emotion regulation and memory (Phelps et al., 2005). Regarding the nature of the task and the known functionality of the putamen and amygdale, our finding might indicate that participants felt more happy and expected reward during high confidence trials, because one knew (or thought) one made the right choice and would win or not lose points; the more confident, the more rewarded and happier a person felt. While this seems likely, our hypothesis about the linear relationship between subjective confidence and the putamen and amygdale relies on abductive reasoning and needs further exploration.
The precental gyrus, also referred to as the primary motor area, commands voluntary movements of skeletal muscles within the body (Scott et al., 1995; Kakei et al., 1999; Meier et al., 2008). Subjective confidence modulated the activation of the precentral gyrus, which indicates a greater motor response when participants were more confident about their choice. This suggests that participants pressed the button box with more conviction and enthusiasm when they were more confident about their choice.
In PG we did not find any BOLD activation modulated by subjective confidence levels. However, it seems to be unlikely the case that no brain region is involved in the subjective confidence of PG. A Methodological limitation is probably underlying our finding. Our fMRI results are based on 12 PG, which is far below the advised number of subjects (n > 20)
25 critical to ensure sufficient statistical power. The difference in sample size between the PG and HC, and therefore the difference in statistical power, might explain why we did not found activation of the putamen, amygdale and precentral gyrus in both groups. Absence of VMPFC activation (scaling issue)
The hypothesized activation of the VMPFC was absent in PG, and even in the larger sample of HC. Interestingly, our colleague Monja Hoven (2017) was able to show significant BOLD activation of the VMPFC that was linearly modulated by subjective confidence. This was found for a similar HC group (including only a higher female ratio), whose data was derived from the same study.The only difference in her analysis was the scaling of the behavioural data that was used for the parametric modulator; confidence rating were z-scored to normalize the behavioural data, whereas in our analysis original confidence ratings were entered as parametric modulator. This emphasizes the importance of one’s decision in whether one normalizes the behavioural data, before using it as parametric modulator. This methodological issue has been discussed exhaustively in the paper of Lebreton and
Palminteri (2016), who explain that one’s choice depends on one’s hypothesis about how the BOLD signal scales with behaviour.
When one normalizes one’s behavioural data, the BOLD signal will encode the variable x on an identical scale across subjects (the relationship between BOLD signal and variable x will be the same for each subject). In this way, between-group or individual differences will reflect differences in behaviour instead of differences in scaling law between groups or individuals. So, the behavioural differences are not reflected by differences in BOLD signals but by the relationship between BOLD signals and behaviour (different coding).
In contrast, without normalization it is assumed that behavioural differences are due differences in BOLD signal. One could hypothesize that possible behavioural differences in subjective confidence level between HC and PG are due aberrant VMPFC activity in PG. While our choice (not normalizing the pmod) was mostly exploratory, we wanted to be able to detect such group difference in VMPFC activity, related to subjective confidence.
With our current approach, we assumed that every value of the rating scale (50-100 %) would be represented in the brain. However, people within the same group might differ in how they report their minimal and maximal confidence feeling. One individual could feel very confident and classify this as 80% confidence, whereas someone else would classify it as
26 100% confidence. However, for both persons this represents their maximal confidence level and therefore express a similar brain signal A. This will lead to a discrepancy about how brain activation A is labelled with subjective confidence level (i.e. one with 80% and one with 100%), while in both persons the activation reflects their maximal subjective confidence level. As a consequence of this within-group variance, it will be harder to find a linear relationship between subjective confidence and the BOLD signal of the brain region coding confidence at the group level.With normalization you take into account each subject-specific confidence range, by converting the confidence scores to standard deviations (SDs) from their average reported confidence. In this way, the highest possible confidence level will be coded for each subject differently. In the example above, the 80% and 100% of both persons will now be coded as a similar behavioural response, since it is converted to a similar SD value. Hence, this explains why the study of Monja Hoven (2017), but not our study, found BOLD activation in the VMPFC related to variations in subjective confidence levels. This stresses out the importance of taken into account individual differences in subjective rating.
Group difference in VMPFC activation
Again in contrast with our hypothesis, we did not find group differences in VMPFC activation that was related to subjective confidence. This is in line with our behavioural result which indicated no group difference in confidence accuracy. However, this fMRI result is difficult to interpret, since the results are based on unequal sample sizes with a too small PG group and more importantly we did not find activation of the VMPFC modulated by the subjective confidence, in both groups separately. As discussed earlier, these results could imply a lack of power to detect group differences, next to the use of a normalized pmod for fMRI analysis.
Risk attitude and confidence accuracy
Contradictory to our hypothesis, a moderate significant negative correlation between overconfidence and risk attitude was found. We expected to find overconfidence in risk-seeking individuals but our study showed that overconfidence was actually present in the risk-averse individuals. Furthermore, overconfidence decreased to zero when individuals were less risk-averse. Interestingly, the risk-seeking individuals in our sample showed to have a better confidence accuracy (close to zero), than risk-averse individuals (when
27 ignoring the outliers). This finding is not in line with Murad et al. (2016) who found a
positive relation between risk-seeking and subjective confidence level. However, they did not control for subject’s actual performance by investigating confidence accuracy.
Our hypothesis was built on the rationale that highly confident individuals are more willing to take a risk, because they are confident that it will turn out well. However, our data now could be interpreted as following: individuals who know that they can judge their decisions well (i.e. accurate confidence), trust themselves more in taking a risk. Even though this seems not unlikely, our finding should be interpreted with caution; in our sample only a few participants were risk-seeking and our measure for risk attitude might have been not that optimal. In our study we estimated the subject-specific parameter α, which determines the curvature of the utility function (to asses risk attitude) and has not been used that often. People with a convex function, experiencing increased marginal utility for greater goods, are classified as risk-seeking individuals. People who experience decreased marginal utility for greater goods (concave function), are classified as the risk-averse individuals. However, this underlying assumption, that one will be more risk-seeking when greater goods satisfy one far more than smaller goods, might not be that correct. An alternative reason for being more risky could be that one is less sensitive to losses and more sensitive to rewards. Our study showed that HC were more averse than PG, with PG being seeking and HC loss-averse. This indicates that losses loomed larger than gains for HC, but not for PG. This is in line with studies that showed hyposensitivity for losses in PG (Lole et al., 2015; Goudriaan et al., 2006; De Ruiter et al., 2009).
Surprisingly, PG did not differ in their risk attitude from HC; both groups showed to be risk-averse. This finding is in contrast with varies behavioural studies that used different risk tasks, and all indicated that PG take more risk than HC (e.g. Brand et al., 2005; Goudriaan et al., 2006; Power et al., 2012; Kraplin et al., 2014). Our inability to reveal this characteristic of PG (risky behaviour) in the present study, suggests that our measure computes another construct. It would be good to examine risky decision-making in this group of PG with a more common task, such as the Iowa Gambling Task. As a next step, comparing the results could confirm whether the tasks assess a different construct.
28
Unequal sample
Another limitation of the study could be the differences in characteristics between our groups. Matching succeeded successfully only on age and gender. Due to the loss of subjects (based on the quality of their data) and the collaboration with other studies, we were not able to match all demographic variables. Hence, we focused on age and gender, in our opinion, the most potential variables that could act as an confounding factor. Since, studies have shown that confidence accuracy declines with aging, our non-significant difference in age might have been too robust. However, it is unlikely that the age difference (i.e. higher average age in HC) contributed to the absence of group-differences in confidence accuracy. On the other hand, the difference in education level needs some attention. Academic achievement has shown to have a modest positive relationship with self-esteem (Valentine et al., 2004; Fathi-Ashtiani et al., 2007; Pullman et al., 2008). Hence, the higher educational level of the HC group might have resulted in a higher esteem in this group. Whether self-esteem affects confidence accuracy is unknown. Intuitively, less self-self-esteem is more likely to contribute to underconfidence in one’s performance, whereas high self-esteem could
promote overconfidence in one’s performance. Based on this, it is unlikely that our results were confounded by educational level through self-esteem, since we found overconfidence in both groups and no group differences. Still, a lower educated HC group is desirable to exclude other possible, however unknown, effects of education level.
The average FTND-score indicated that the smokers were low nicotine-dependent in the HC group and low to moderate in the PG group. This non-significant difference in FTND-score, seems not worrisome. Especially, since there is no knowledge that would hypothesize a link between nicotine dependence and confidence accuracy. Limited findings have shown more risky behaviour in smokers compared to controls (Lejuez et al., 2003; Ert et al., 2013) and suggest that smokers’ risk taking is driven by their tendency to be more easily tempted by immediate high rewards (Ert et al., 2013). The PG group, which included a larger
proportion of smokers, seemed to be more sensitive for rewards than for losses, whereas HC showed to be loss-averse. We cannot exclude the possibility that the unequal proportion of smokers contributed to this finding. Therefore, in further analysis, an equal proportion of smokers is desirable.
29
Recommendations for future research
Evidence as Parametric modulator
This study used confidence ratings as parametric modulator (pmod) to detect the brain regions involved in subjective confidence. Future analyses for this study and other studies regarding this topic, could use a more objective measure for confidence as an additional pmod. Confidence has already been studied as an objective mathematical quantity. Sanders et al. (2016) has shown that subjective confidence judgements can be predicted from a statistical model for confidence formation. This model proposes that confidence builds on perceptual evidence. In this study, higher confidence levels for correct decisions were found, when the amount of evidence increased and, in contrast, confidence decreased with
increased levels of evidence for incorrect decisions. More recently, this positive and negative correlation between subjective confidence and the amount of evidence in HC was replicated (Lebreton et al., 2017). These studies indicate that in easy decisions (i.e. more evidence is present) one is more sure one was correct or incorrect, resulting in high or low subjective confidence, respectively.
In our study the trials included a range of individual difficulties, by varying the contrast difference between the Gabor patches. The difference in contrast could be used to quantify the evidence available per trial (Lebreton et al., 2017). The trial-specific evidence could be used as pmod, to investigate which region vary in BOLD activation with the amount of evidence for correct and incorrect choices, separately. Interestingly, it has already been found that the VMPFC was more activated during trials including high levels of evidence,
compared to trials including less evidence (Rolls et al., 2010). With this finding, the authors proposed that the VMPFC is engaged in computing the confidence in our choices. The combination of evidence and subjective confidence rating as pmod, might capture the
complete confidence formation in the brain; from the first computation, based on the amount of evidence, till the rise of its subjective confidence feeling. In addition, it would be
interesting to compare the use of evidence as pmod with the use of subjective confidence rating as pmod, in HC and PG. Since, the amount of evidence and subjective confidence are correlated in HC, overlap between the results would be expected in HC. However, it is not known whether the specific pattern between subjective confidence and amount of evidence
30 (Sanders et al., 2016; Lebreton et al., 2017) is also present in PG. Investigating the role of evidence in the confidence formation of PG could provide some interesting insights. Another approach for determining confidence accuracy
This study determined subject’s confidence accuracy by calculating the mismatch between average subjective confidence and average performance. However calculating confidence accuracy for each trial separately is more desirable. This way inter-subject variability could be taken into account and makes it able to calculate the confidence accuracy on a trail-by-trial basis. Furthermore, our approach does not investigate confidence in correct and
incorrect trials separately. One could be inappropriate highly confident in incorrect trials, but low confident in correct trials. This information gets lost, when taking the average subjective confidence and average performance. Future analyses could use the amount of evidence given in the specific trial (which determines the difficulty of the choice in that trial), to assess whether the subjective confidence rating was appropriate given the amount of evidence and the correctness of the choice. The difference between subject’s reported confidence level and the expected confidence level (based on the evidence and trial performance) can be used as a measure of confidence accuracy on a trial-by-trial basis. The trial-specific confidence
accuracies can be entered in a mixed model.
Conclusion
The current study showed no difference in confidence accuracy between PG and HC; both groups showed to be equivalent overconfident, to a small extent. Furthermore, confidence accuracy of both groups seemed to be insensitive to monetary incentive. However, a larger sample size with age subgroups should be analysed again to confirm these findings. Our methodology did not find VMPFC activation that was related to subjective confidence in both groups, whereas another approach did find this in HC. This emphasizes the importance of normalizing the parametric modulator in fMRI analysis, when individual differences in the behavioural measure (regarding the scaling) are plausible. Moreover, this study was not able to show group differences in VMPFC activation modulated by subjective confidence. The negative relationship between overconfidence and risk attitude, indicates that
overconfidence does not contribute to risky behaviour. However, this negative relationship needs to be interpreted with caution; a more general approach for assessing risk is needed.
31 Most importantly, a greater (clinical) sample size is needed to validate our findings. Future studies should build on this study, using level of evidence as an additional parametric modulator.
REFERENCES
Bartra, O., McGuire, J. T., & Kable, J. W. (2013). The valuation system: a coordinate-based meta-analysis of BOLD fMRI experiments examining neural correlates of subjective value. Neuroimage, 76, 412-427.
Bechara, A., Tranel, D., & Damasio, H. (2000). Characterization of the decision-making deficit of patients with ventromedial prefrontal cortex lesions. Brain, 123(11), 2189-2202.
Bhavsar, S., Zvyagintsev, M., & Mathiak, K. (2014). BOLD sensitivity and SNR characteristics of parallel imaging-accelerated single-shot multi-echo EPI for fMRI. Neuroimage, 84, 65-75. Bombardini, M., & Trebbi, F. (2012). Risk aversion and expected utility theory: an experiment with large and small stakes. Journal of the European Economic Association, 10(6), 1348-1399. Brand, M., Kalbe, E., Labudda, K., Fujiwara, E., Kessler, J., & Markowitsch, H. J. (2005). Decision-making impairments in patients with pathological gambling. Psychiatry research, 133(1), 91-99.
Brevers, D., Cleeremans, A., Bechara, A., Greisen, M., Kornreich, C., Verbanck, P., & Noël, X. (2013). Impaired self-awareness in pathological gamblers. Journal of Gambling Studies, 29(1), 119-129.
Brevers, D., Bechara, A., Cleeremans, A., & Noël, X. (2013). Iowa Gambling Task (IGT): twenty years after–gambling disorder and IGT. Frontiers in psychology, 4, 665.
Brevers, D., Cleeremans, A., Bechara, A., Greisen, M., Kornreich, C., Verbanck, P., & Noël, X. (2014). Impaired metacognitive capacities in individuals with problem gambling. Journal of gambling studies, 30(1), 141-152.
Chumbley, J. R., Krajbich, I., Engelmann, J. B., Russell, E., Van Uum, S., Koren, G., & Fehr, E. (2014). Endogenous cortisol predicts decreased loss aversion in young men. Psychological science, 25(11), 2102-2105.
Clark, L. (2014). Disordered gambling: the evolving concept of behavioral addiction. Annals of the New York Academy of Sciences, 1327(1), 46-61.
32 Clark, L., Bechara, A., Damasio, H., Aitken, M. R. F., Sahakian, B. J., & Robbins, T. W. (2008). Differential effects of insular and ventromedial prefrontal cortex lesions on risky decision-making. Brain, 131(5), 1311-1322.
Corbetta, M., Shulman, G. L.( 2011). Spatial neglect and attention networks. Annual Review Neuroscience, 34, 569–99.
Delgado, M. R., Stenger, V. A., & Fiez, J. A. (2004). Motivation-dependent responses in the human caudate nucleus. Cerebral Cortex, 14(9), 1022-1030.
Delgado, M. R., Miller, M. M., Inati, S., & Phelps, E. A. (2005). An fMRI study of reward-related probability learning. Neuroimage, 24(3), 862-873.
De Martino, B., Fleming, S. M., Garrett, N., Dolan, R. J. (2013). Confidence in value-based choice. Nature Neuroscience, 16(1):105–110.
De Ruiter, M. B., Veltman, D. J., Goudriaan, A. E., Oosterlaan, J., Sjoerds, Z., & Van Den Brink, W. (2009). Response perseveration and ventral prefrontal sensitivity to reward and punishment in male problem gamblers and smokers. Neuropsychopharmacology, 34(4), 1027-1038.
Ert, E., Yechiam, E., Arshavsky, O. (2013). Smokers’ Decision Making: More than Mere Risk Taking. PLoS ONE, 8(7): e68064.
Fathi-Ashtiani, A., Ejei, J., Khodapanahi, M. K., & Tarkhorani, H. (2007). Relationship between self-concept, self-esteem, anxiety, depression and academic achievement in adolescents. Journal Applied Science., 7, 995-1000.
Fernandez, B., Leuchs, L., Sämann, P. G., Czisch, M., & Spoormaker, V. I. (2017). Multi-echo EPI of human fear conditioning reveals improved BOLD detection in ventromedial
prefrontal cortex. NeuroImage, 156, 65-77.
Fleming, M., Weil, R.S., Nagy, Z., Dolan, R.J. & Reed, G. (2010). Relating Introspective Accuracy to Individual Differences in Brain Structure. Science, 329, 1541-1543.
Gaher, R. M., Hahn, A. M., Shishido, H., Simons, J. S., & Gaster, S. (2015). Associations between sensitivity to punishment, sensitivity to reward, and gambling. Addictive behaviors, 42, 180-184.
Giardini, F., Coricelli, G., Joffily, M., & Sirigu, A. (2008). Overconfidence in predictions as an effect of desirability bias. Advances in decision making under risk and uncertainty, 163-180. Goodie, A.S. (2005). The Role of Perceived Control and Overconfidence in Pathological Gambling. Journal of Gambling Studies ,21(4),481-502.
33 Goudriaan, A. E., Oosterlaan, J., de Beurs, E., & van den Brink, W. (2006).
Psychophysiological determinants and concomitants of deficient decision making in pathological gamblers. Drug and alcohol dependence, 84(3), 231-239.
Grahn, J. A., Parkinson, J. A., & Owen, A. M. (2008). The cognitive functions of the caudate nucleus. Progress in neurobiology, 86(3), 141-155.
Grahn, J., Parkinson, J.A., & Owen, A.M. (2009). The role of the basal ganglia in learning and memory: Neuropsychological studies. Behavioural Brain Research 199 (1),144–145.
Gur, R. C., Schroeder, L., Turner, T., McGrath, C., Chan, R. M., Turetsky, B. I., ... & Gur, R. E. (2002). Brain activation during facial emotion processing. Neuroimage, 16(3), 651-662.
Haruno, M., Kuroda, T., Doya, K., Toyama, K., Kimura, M., Samejima, K., ... & Kawato, M. (2004). A neural correlate of reward-based behavioral learning in caudate nucleus: a functional magnetic resonance imaging study of a stochastic decision task. Journal of Neuroscience, 24(7), 1660-1665.
Heatherton, T. F., Kozlowski, L. T., Frecker, R. C., & FAGERSTROM, K. O. (1991). The Fagerström test for nicotine dependence: a revision of the Fagerstrom Tolerance Questionnaire. Addiction, 86(9), 1119-1127.
Hikosaka, O., Sakamoto, M., & Usui, S. (1989). Functional properties of monkey caudate neurons. III. Activities related to expectation of target and reward. Journal of
Neurophysiology, 61(4), 814-832.
Hoven, M. (2017). Confidence calibration in pathological gambling and obsessive-compulsive disorder. Unpublished manuscript.
Jiménez-Murcia, S., Fernández-Aranda, F., Mestre-Bach, G., Granero, R., Tárrega, S., Torrubia, R., ... & Moragas, L. (2016). Exploring the relationship between reward and punishment sensitivity and gambling disorder in a clinical sample: A path modeling analysis. Journal of Gambling Studies, 1-19.
Jönsson F. U., Olsson, H., Olsson, M. J, (2005). Odor emotionality affects the confidence in odor naming. Chemical Senses, 30(1):29–35.
Kakei, S., Hoffman, D. & Strick, P. (1999). "Muscle and movement representations in the primary motor cortex". Science. 285 (5436), 2136–2139.
Kepecs, A., Uchida, N., Zariwala, H. A., & Mainen, Z. F. (2008). Neural correlates, computation and behavioural impact of decision confidence. Nature, 455(7210), 227.
Kiani, R., Corthell, L., & Shadlen, M. N. (2014). Choice certainty is informed by both evidence and decision time. Neuron, 84(6), 1329-1342.
