MetacognitionEmotionWorry

pated before the decision. Under worry individuals tend to have better metacognition in

nd cobetwre int

processes such as emotions. The only exception that we are aware is Garnkel et al. (2013) who show that the level of meta-rtbeat. This resultHere we wion.judgments

average to real success (Harvey, 1997) and metacognitive accuracy refers to the discrimination (or resolution)variations of condence match the variation of performances (Fleming & Dolan, 2012). A potential relationship b

http://dx.doi.org/10.1016/j.concog.2014.08.0061053-8100/ 2014 Elsevier Inc. All rights reserved.

Address: School of Economics and Finance, Queensland University of Technology, QUT Gardens Point Z Block Level 8 Ofce Z833, 2 George St.,Brisbane City, QLD 4000, Australia.

E-mail address: sebastien.massoni@gmail.com

Consciousness and Cognition 29 (2014) 189198

Contents lists available at ScienceDirect

Consciousness and Cognition

journal homepage: www.elsevier .com/locate /concogcognition in a memory task is modulated by the timing of the stimulus with respect to the phase of the heaconrms the idea of the present study: the internal processes may have an effect on the metacognition.characterizes this effect and shows that metacognition is improved by the emotional valence of the decis

Wemeasure metacognition via two different abilities: calibration (or bias) reveals how close condenceant to

are onof howetweension and are useful to make accurate judgments (Damiaso, 1994; LeDoux, 1996). In this study we will focus on a specicaspect of cognition: metacognition, i.e. the knowledge an individual has about his own cognition. How emotion interactswith this specic kind of cognition is for the moment an open question.

Recent studies have focused on inter-subject and inter-task variation in metacognition (Fleming, Weil, Nagy, Dolan, &Rees, 2010; Song et al., 2011; McCurdy et al., 2013) but none have examined howmetacognition could be affected by internalMotivationAttentionReaction time

1. Introduction

How humans combine emotion aThe initial philosophical dissociationseded by a view in which emotions aterms of the two measures. Furthermore understanding the formation of condence is bet-ter explained with taking into account the level of worry in the model. This study showsthe importance of an emotional component in the formation and the quality of the subjec-tive probabilities.

2014 Elsevier Inc. All rights reserved.

gnition during their decision process is a central question in the behavior sciences.een rational and emotional decision (Plato, Descartes, and Kant) has been super-egrated in the decision process. Emotions are an important component of the deci-Emotion as a boost to metacognition: How worry enhancesthe quality of condence

Sbastien Massoni QuBE School of Economics and Finance, Queensland University of Technology, Australia

a r t i c l e i n f o

Article history:Received 26 February 2014Available online 3 October 2014

Keywords:Condence

a b s t r a c t

Emotion and cognition are known to interact during human decision processes. In thisstudy we focus on a specic kind of cognition, namely metacognition. Our experimentinduces a negative emotion, worry, during a perceptual task. In a numerosity task subjectshave to make a two alternative forced choice and then reveal their condence in this deci-sion. We measure metacognition in terms of discrimination and calibration abilities. Ourresults show that metacognition, but not choice, is affected by the level of worry antici-

190 S. Massoni / Consciousness and Cognition 29 (2014) 189198metacognition and emotion is hypothesized at the light of advances on the neural basis of condence judgments. Recentstudies show that metacognition is associated with activity in the anterolateral prefrontal cortex (PFC) (Yokoyama &et al., 2010 using a short-memory task) and the lateral PFC (Fleming, Huijgen, & Dolan, 2012; Fleming et al., 2010;Rounis, Maniscalco, Rothwell, Passingham, & Lau, 2010 using a perceptual task). Emotions have been more extensively stud-ied and robust evidence has been found in favor of a central role of the orbitofrontal cortex (Bechara, Damasio, & Damiaso,2000) and the amygdala (Seymour & Dolan, 2008) in the emotional part of decision-making. Nevertheless, some studies doc-ument a network of brain area activated during the cognitive emotional integration rather than specic area for the twoaspects (Pessoa, 2008) and this integration of emotions in cognition may occur in lateral PFC (Gray, Braver, & Raichle, 2002).This view gives support to the existence of an emotional effect on the metacognitive abilities. A more intuitive way of think-ing about the link between emotion and metacognition is provide by the attentional effect of emotion. The affective signif-icance of a stimulus is known to induce changes in sensory processing and attention (Vuilleumier, 2005; Yiend, 2010).Negative emotions, such as anxiety and worry, are generally associated with a decrease in attentional control (Eysenck,Derakshan, Santos, & Calvo, 2007) and dysregulation of attentional focus (Bishop, 2008). Nevertheless a differentiationbetween the effects of state and trait anxiety (Pacheco-Unguetti, Acosta, Callejas, & Lupianez, 2010) gives support to a poten-tial positive impact of stated worry on metacognition by increased willingness to (over)control information. This tendency toover-react under worry could have a positive impact on the quality of metacognition with the use of more precise ratingstrategies and thus better discrimination ability. We might also expect better calibration with a diminishing of the overcon-dence by a depressive effect.

Our experimental design induces emotions by framing effects. We use loss vs. gain and high vs. low stake frames to gen-erate variations on the self-reported level of worry of subjects. Loss aversion is the tendency to weight loses greater thanequivalent gains in decision-making (Kahneman & Tversky, 1979). Recently this well documented bias has been studied withthe help of neuronal data and there is evidence that loss aversion could be linked to emotional interactions on the cognitiveprocess (see Takahashi, 2013). Indeed De Martino, Kumaran, Seymour, and Dolan (2006), De Martino, Camerer, and Adolphs(2010) found an activation of the amygdala in the case of loss frame. Even if another fMRI study (Tom, Fox, Trepel, &Poldrack, 2007) was not able to replicate this nding (they found activation of the lateral PFC and the striatum), there existsevidence for an emotional aspect for loss choices compared to gains (Sokol-Hessner, Camerer, & Phelps, 2013; Sokol-Hessneret al., 2009, conrm this hypothesis using physiological and fMRI measures). Thus using a loss vs. gain frame in an experi-ment should induce variations in the worry felt by subjects facing their choices. We also use a high stake vs. low stake frameto increase the variations of the stated level of worry. High stake decisions are known to be more emotionally demandingthan low stake cases (see Kunreuther & et al., 2002, for a review of high stakes decision making). Overall we can expectto succeed in inducing some variations of a negative emotion that will be measured in terms of a worry scale. We deneworry as a cognitive phenomenon [. . .] concerned with future events where there is uncertainty about the outcome, thefuture being thought about is a negative one, and this is accompanied by feeling of anxiety (MacLeod, Williams, &Bekerian, 1991 p. 478). This approach focuses on the central role of uncertainty (Dugas, Gosselin, & Ladouceur, 2001)and makes sense in our design where the outcomes are uncertain and may be negative with important losses. The inuenceof anxiety on decision-making is well-studied (Hartley & Phelphs, 2012) but it remains unclear that worry leads always toworse decisions. As our design is not based on cognitive tasks but perceptual ones with emotionally neutral stimuli (but wor-rying frames), we expect to nd a positive effect of worry on metacognition due to an increase of energy and attentionaleffort. We assume that metacognition will be improved when subjects reveal their condence after making a decision undera worried mood. This assumption that worse mood leads to better metacognition is also supported by results from neuro-psychiatric disorders studies. Metacognition, dened as insight or awareness of illness, is improved by negative moods(David, Bedford, Wiffen, & Gilleen, 2012). We can hypothesize that this link is also valid for healthy individuals and thus thatmetacognition is improved by worried mood.

2. Methods

2.1. Participants

The experiment was conducted in May and July 2012 at the Laboratory of Experimental Economics in Paris (LEEP) of theUniversity of Paris 1. Subjects were recruited by standard procedure in the LEEP database and gave written informed consentto take part in the experiment. 103 healthy subjects (54 men; age 1838 years, mean age, 22.9 years, most enrolled as under-graduate students at the University of Paris) participated in this experiment for pay. The sessions lasted around 120 min andsubjects were paid on average 27.1. We excluded 6 subjects from analysis due to insufcient variation (s.d. < 0.03) of con-dence or worry. The nal sample included 97 subjects for analysis.

2.2. Stimuli

The experiment was conducted in MATLAB using Psychophysics Toolbox version 3 (Brainard, 1997). We use a 2AFC num-erosity task, which is known to be convenient to t SDT models (Nieder & Dehaene, 2009) and may be positively affected byemotions (see Phelps, Ling, & Carrasco, 2006, for the effect of emotions on perception). The stimuli consisted of two circleswith

S. Massoni / Consciousness and Cognition 29 (2014) 189198 191Fig. 1. Experimental design. (A) describes the timeline of the experiment. First subjects observe the characteristics of the bet: the amount of the stake is 20or 200 and the goal is 2, 3 or 4 successes over 5. In the loss frame they play for not losing and they lose if they fail more than 5 minus the goal times over 5;in the gain frame they play for wining and they win if they succeed at least the goal times over 5. Then they give their level of worry against this bet on ascale between 0 and 9 and their level of condence in succeeding to the bet on a scale between and 100. After that they do 5 trials of perceptual task withthe following sequence: after observing a xation cross, subjects initiate the stimuli which consist of two circles with a certain amount of dots inside (B is anexample of our stimuli). While one circle contains always 50 dots the other circle contains 50 + xc dots. The value of xc is determined by a psychophysicalstaircase in order to obtain a success rate of 71%. After observing the stimulus for 700 ms subjects have to make their decision and indicate whether it wasthe right or the left circle that contained the most of dots. Then they have to give the level of condence on the accuracy of this decision on a scale goingfrom 0% to 100% with steps of 5. (C) explains the mechanisms of the condences elicitation by probabilities matching. It consists in asking individualswhether they prefer to be paid according to the correctness of their answer or according to a specied lottery. A number l1 is drawn between 0 and 100. Iftheir condence is higher than this number they will be paid on the accuracy of their decision: they win one point if the answer is correct and lose 1 pointa certain number of dots in each circle (see Fig. 1B). All dotswere of the same size andwe control for the distance between eachdot. One of the two circles always contained 50 dots while the other contained 50 + xc dots. Before the experiment we esti-mated the value of xc needed to obtain a success rate of 71% using a psychophysical staircase (Levitt, 1971; see below). Theposition of the circle containing the greater number of dots was randomly assigned to be on the left or right on each trial.

2.3. Task and procedure

2.3.1. Practice and thresholdingSubjects initially performed practice trials of the dots task without condence ratings, in which full feedback was given.

We used these trials to calibrate difculty of the dots task. The calibration phase was done by one-up two-down staircase(Levitt, 1971): after two consecutive correct answers one dot was removed, and after one failure one dot was added. Westopped the calibration after 30 reversals in the staircase and the value of xc was calculated as the mean dot number acrossthe two last reversals of the staircase. Subjects then performed 10 trials of the tasks with condence elicitation and feedbackboth on their accuracy and on the results of the elicitation mechanism. Finally they performed 30 trials without feedback tocheck whether the calibration procedure had succeeded.

2.3.2. Experiment phaseThe experimental design comprised two aspects, a bet phase in which we induced emotions and a perceptual phase in

which subjects performed the numerosity task with choice and condence. The experiment consisted in 64 blocks of 5 trials.Each block is dened by a risky bet that subjects have to face. These bets have three main components: their goal (the objec-tive of the bet is to obtain at least 2, 3 or 4 success over the 5 next trials), the amount of money in play (20 or 200) and theframe (bet for win or to avoid loss i.e. a gain or loss framing). The sequence of a block was the following. First subjectsobserved the three characteristics of the bet. Then they made a judgment about their level of worry about this bet on a10-point scale from 0 (any worry) to 9 (very worried). We also asked them to reveal their premium coverage for the bet1

and their condence in the bets success. Then they performed 5 trials of the perceptual task. No feedback was provided eitheron the results of the bet or on the success of each trial. Each trial of the task was based on the following sequence (see Fig. 1A).

otherwise. If the condence is lower than l1 they will be paid according to the following lottery: they will have a probability l1 of winning 1 point and aprobability 100 - l1 of losing 1 point. A second number l2 is then drawn between 0 and 100; if it is higher than l1 they win otherwise they lose.

1 We elicited this certainty equivalent by a BDM mechanism (Becker, DeGroot & Marschak, 1964). We do not provide a full explanation of this aspect of theexperiment as we do not use it in further analysis.

First two outline circles were displayed with xation crosses at their center. The subject initiated the trial by pressing thespace key on a standard computer keyboard. The dot stimuli then appeared for 700 ms, and subjects were asked to respondleft or right by pressing the f or j keys, respectively. There was no time limit for responding. After responding subjects were

L( x with probability p, and probability (1 p). A ran ber q is then drawn in th l[0 ller than p, the subject is p ding to the lottery L(E). Ot he subject is paid accordin -te elds x with probability q a th probability (1 q).

r scoring rule i.e. it provi tives for subjects to reve ubjective probability truth eGajdos, Massoni, & Vergnaud, 2014). Suppose that the subject has a probability of success is p but reports a probability

) yields a higher p (q). Thus, the sub payoff is higher w pthan when reporting r < p.

192 S. Massoni / Consciousness and Cognition 29 (2014) 189198Similarly, if the subject reports r > p, the payments (according to her subjective probability p) are described in the follow-ing table.

q < p < r p < q < r p < r < q

Reports r > p L(p) L(p) L(q)Reports p L(p) L(q) L(q)

Whenever p < q < r, L(q) yields a higher payment than L(p). Thus, the expected payoff is higher when reporting p thanwhen reporting r > p.

A major advantage of this mechanism is that it provides the subject incentives to reveal her subjective probabilities truth-fully, while not being contaminated by his attitude towards risk.2 Furthermore Gajdos, Massoni, & Vergnaud (2014) show thatthe elicited condences t the SDT predictions and thus it supports the use of SDT-measures such as the meta-d0.

In practice the Matching Probabilities mechanism is implemented using a scale of 0100, with steps of 5 (see Fig. 1C).After having made the decision, subjects are told that they are entitled to a ticket for a lottery based on their answers accu-racy. This lottery gives them 1 point if their answer is correct, and 1 point otherwise. Subjects have then to report on a scaleranging from 0 to 100 the minimal percentage of chance p they require to accept an exchange between their lottery ticketand a lottery ticket that gives p chances of winning 1 point and 100 p chance of losing 1 point. A number l1 is drawnaccording to a uniform distribution between 0 and 100. If l1 is smaller than p, subjects keep their initial lottery ticket. Ifl1 is higher than p, they are paid according to a lottery that gives them l1 chances of winning. In this case, a random drawdetermines the payment: a number l2 is determined using a uniform distribution between 0 and 100, the lottery leads toa win if l1 is higher than l2.

2 On the contrary to the post-wagering method (Persaud, McLeod, & Cowey, 2007) which is the other main method uses to elicit condence with incentivesand which suffers from a risk preferences dependence (see Dienes & Seth, 2010; Schurger & Sher, 2008; Fleming & Dolan, 2009).Whenever r < q < p, L(p ayment than L jects expected hen reportingReports p L(p) L(p) L(q)r p. If r < p, the lotteries according to which the subject (given the subjective probability p) is paid are represented inthe following table, as a function of the random value q.

q < r < p r < q < p r < p < q

Reports r < p L(p) L(q) L(q)This a prope des incen al their s fully (se,1]. If q is smary L(q) that yiaid accornd x wiherwise, t g to a lot

p) that yields x with dom num e intervaasked to indicate their level of condence in their choice on a gauge from 0% to 100% with steps of 5%, using the up and downkeys, again with no time limit on the response.

Overall the experimental phase was divided into 64 blocks (and thus 320 trials of the perceptual task) with the followingcharacteristics: 32 with a goal of 2/5, 16 with a goal of 3/5, 16 with a goal of 4/5; 32 with a stake of 200, 32 with a 20; 32with the loss frame, 32 with the gain frame. Overall, subjects gave 320 ratings of condence and 64 levels of worry with thethree characteristics balanced between trials.

2.3.3. PaymentSubjects payment comprised 5 for participation and two variable parts: one bet was randomly chosen and subjects

either won the stake, received nothing if they were not covered by the BDM insurance mechanism, or received the certaintyequivalent in case of coverage by the insurance; they also accumulated points according to the accuracy of their stated con-dence. The incentive mechanism used was the probability matching rule (Fig. 1C) by which they won or lost points on eachtrial. This mechanism can be seen as a generalization of the no-loss gambling studied by Dienes and Seth (2010). It consists inasking individuals whether they prefer to be paid according to the correctness of their answer or according to a speciedlottery. To elicit a subjects subjective probability about an event E, the subject is asked to provide the probability p thatmakes him indifferent between a lottery L(E) that gives a positive reward x if E happens, and x otherwise and a lottery

As the mechanism might seem complicated we explained in details to subjects how their stated condence will deter-

We dene two components of metacognition as calibration and discrimination. Calibration is obtained by computing the0

S. Massoni / Consciousness and Cognition 29 (2014) 189198 193mean of the difference between condence and accuracy. Discrimination is measured by the difference between meta-d andd0. The measure of d0 is: d0FC = Z(prob(Response = Right|Stimulus = Right)) Z(prob(Response = Right|Stimulus = Left)). Meta-d0 is a new measures introduced by Maniscalco and Lau (2012) that controls for a performance confound in the discrimina-tion measure. It is known that type II decisions are affected by the treatment of the type I signal (Galvin, Podd, Drga, &Whitmore, 2003), but meta-d0 denes the level of d0 that an SDT-ideal observer would need to generate this set of type IIanswers. Thus a SDT-optimal observer will have a null difference between meta-d0 and d0 while for a real subject the valueof this difference is a robust measure of metacognitive ability which is independent of the type I and type II criterion (seeBarrett, Dienes & Seth, 2014, for a detailed analysis). Meta-d0 is computed using the Matlab code of Maniscalco and Lau(2012) available on their website (http://www.columbia.edu/~bsm2105/type2sdt/). Barrett, Dienes & Seth (2014) show thatthe stability of this measure may be questionable in the case of a very small number of trials for the different types ofresponses (levels of condence for correct and incorrect answer and for right and left stimuli). But the high number of trialsand their distribution excludes this instability in the present study.

A dynamic measure of type I performances is derived from the parameters of the diffusion model (Ratcliff & McKoon,2008, for a review) and is computed on Matlab using the D-MAT toolbox (Vandekerckhove & Tuerlinckx, 2007, 2008). Thistoolbox performs an estimation of the different parameters of a diffusion model: three parameters for the decision process(the boundary separation, a, which captures the speed-accuracy trade-off; the drift rate, z, showing the amount of informa-tion; and the starting point, z, which gives information about a bias in the decision); one parameter for the non-decision (thenon-decision time, Ter, that measures all the other process involved, e.g. motor reaction time); and three parameters of inter-trial variability (the inter-trial standard deviation of the drift rate, eta, for the variation of attention; the inter-trial ranges ofbias, sz, and non-decision time, st, that captures the subjects variability for both measures). We then t different models onthe data at an individual level: rst an estimation on the whole dataset, then some models allowing one or several param-eters to vary under high and low worry while the other parameters stay identical under the two conditions. The differentmodels estimated are the following: variation of the boundary only, of the drift only, of the boundary and the drift, of theboundary, the drift and the non-decision time, of the previous ones and the bias, of the previous ones and the parametersof inter-trial variability. We can then compare the parameters under high and low worry using paired t-tests.

The relationships between different measures were analyzed with Pearsons product-moment correlations. Comparisonsof their means were conducted using paired t-tests. Comparisons of condence distributions are based on two-sample Kol-mogorovSmirnov tests and their representation is done by a kernel estimation of the density on the trials under high andlow worry.

We examine whether condence is better explained by adding an emotional component. We compare two models: one inwhich condence is explained by the level of accuracy and the reaction time and one in which we add the normalized level ofworry in addition. Regressions on the full dataset were done by ordinary least squares (OLS) method with a clustering of thestandard errors by individuals. Models comparisons are based on different measures of goodness-of-t provided by the t-stat command under Stata. The support for one model against the others was obtained using the grades of evidence ofRaftery (1995) which compare the Bayesian information criterion (BIC; Schwarz, 1978) of each model by the following equa-tions: BIC = 2lnL + k lnn where L is the log-likelihood of the model, k the numbers of free parameters and n the number ofobservations. The difference in terms of BIC provides some supports for one model against the other with the followinggrades: none for a negative difference; weak for a value between 0 and 2, positive between 2 and 6; strong between 6and 10; and very strong for a difference higher than 10. We additionally compared each model to a null intercept-only modelin terms of reduction of the variance.

3. Results

We rst checked whether the experimental design has succeeded in implementing different levels of worry at an indi-vidual level. The mean level of worry was 4.75 (s.d. 2.64) and varied by individuals from 0.27 to 8.45 (s.d. 1.57). In orderto control for a potential bias in the use of the worry scale and to have comparable values across subjects we transformthe reported level of worry into a normalized level of worry by computing the individual z-score of worry.

We expected a higher level of worry in the loss frame, for high stakes and for high difculties. We examined this by per-forming an OLS regression of these characteristics on the worry. Table 1 presents these effects.mine their payment and showed how different rating strategies will lead to different earning schemes. The accumulatedpoints paid at the exchange rate of 1 point = 0.05.

2.4. Data analysis

In order to do mean comparison we split the data into two groups. We dened a block of 5 trials as high worry if thenormalized worry is strictly positive and as low worry if it is negative or null. This split leads to 48% of trials with highworry and 52% with low worry.

Table 1Mean effects of the characteristics on the level of worry.

Goal 3/5 vs. 2/5 Goal 4/5 vs. 2/5 Frame loss vs. frame gain High stake vs. low stake

194 S. Massoni / Consciousness and Cognition 29 (2014) 189198These results show that our experimental design was efcient at inducing changes of worry by varying the characteristicsof the bet. Thus we can study the effects of variation of worry on type I and type II decisions.

We rst examine whether the worry affects the type I decision. Task performance on high worry trials (mean 68.5%, s.d.0.08) does not differ from those with low worry (mean 68.3, s.d. 0.09; t(96) = 0.27, P = .790) as well as reaction time (mean1.203, s.d. 0.32 vs. mean 1.201, s.d. 0.35; t(96) = 0.12, P = .904). This lack of effect of the level of worry in the treatment of thesignal is conrmed in terms of SDT modeling by comparing the d0FC (no signicant difference: mean 1.064, s.d. 0.50 vs. mean1.072, s.d. 0.53; t(96) = 0.14, P = .882) and by tting a diffusion model on the dataset. Diffusion models have been used toemphasis emotional impacts on the decision process even if the raw comparison of reaction times does not reect an effect(White, Ratcliff, Vasey, & McKoon, 2010). However all parameters of the dynamic model (drift, threshold, non-decision time,bias and internal variations) were not statistically different with high or low worry. Table 2 shows the mean values of theparameters for the whole model and for models allowing one or several parameters to vary under high and lowworry. All thedifferences between the parameters estimated under high and low worry are statistically non-signicant.

Furthermore allowing only one parameter to vary does not improve the likelihood of the model against a model withoutvariation induced by worry. Thus we can assert that the level of worry has no impact on the type I decision.

We next examined the effects of worry on type II decisions and metacognitive abilities. First we note that condence andworry are weakly but signicatively correlated (r = .02, P < .0001) on the whole dataset. This link is due to condence onhigh worry trials being systematically lower (mean 69.5%, s.d. 0.12) than on low worry ones (mean 70.6%, s.d. 0.12;t(96) = 2.40, P = .009). As performance is not affected this effect implies signicantly lower overcondence for trials withhigh worry (mean +0.98%, s.d. 0.14) than with low one (mean +2.26%, s.d. 0.14; t(96) = 1.72, P = .044 see Fig. 2A). Notethat the bias for worrying trials is not statistically different from 0 (t(96) = 0.695, P = .489) and reects perfect calibration(consistent with the review of Baranski & Petrusic, 1994, on perceptual tasks). This effect shows a positive impact of worryon the calibration index and thus on metacognitive ability. However it is for the moment not possible to exclude that it ismore an automatic effect rather than an increase of the metacognitive ability. Indeed the better t of the calibration could beattributed to an effect of worry on condence, together with a constant performance. A more robust and signicant effect isobserved on discrimination ability (see Fig. 2B). The difference between meta-d0 and d0 is systematically lower for the highworry trials (mean 0.3308, s.d. 0.45) than for to the low worry ones (mean 0.4282, s.d. 0.52; t(96) = 1.88, P = .032). As themeta-d0 difference is not affected by the mean level of condence we can assert that a high level of worry leads to an increasein metacognitive ability. On our dataset we cannot attribute this effect to an increase of the attentional effort as the reactiontimes for condence are not statistically different.3 Thus declaring a high level of worry changes the formation of condenceand increases its quality. To conrm this hypothesis we perform a KolmogorovSmirnov test of difference in the distribution ofcondence for high and low worry trials. The two distributions are statistically different according to this test (D = 0.0317,P < .0001) and this conrms that subjects use a different (and a better) rating strategy under high worry (see Fig. 3).

We have found a signicant and positive impact of worry on calibration and discrimination ability. It is interesting tocheck whether this improvement is linked between these two components. The correlation between the variations withworry of the calibration index and the meta-d0-d0 difference is statistically signicant (r = .41, P < .0001). Thus we can assertthat worry has a positive effect on metacognition and subjects improve their abilities on both aspects: calibration and dis-crimination. It is interesting to see whether the level of worry felt before the trials is a good predictor of condence in a per-ceptual task. We compare a model in which condence is explained by the accuracy and reaction time of the trial against amodel in which condence is explained by the two previous variables and the level of worry anticipated for the trials. Thesetwo nested models were estimated by an OLS regression on the whole dataset with individual clusters of the standard errors.The rst model provides the following coefcients: 0.0576 (t(31040) = 8.65, P < .0001) for the accuracy, 0.0535(t(31040) = 6.01, P < .0001) for the reaction time and 0.7228 (t(31040) = 41.69, P < .0001) for the intercept while the model

Worry +25.25%*** +56.1%*** +13.3%*** +14.0%***

% of subjects 51% 70% 48% 38%

*** Means statistically signicant at 1% and comes from an OLS regression. Percentages of subjects affected by the characteristics are those for which wehave a signicant effect (at 10% or less) in individual OLS regressions.with normalized worry gives the following ones: 0.0576 (t(31040) = 8.65, P < .0001) for the accuracy, 0.0535(t(31360) = 6.02, P < .0001) for the reaction time, 0.0053 (t(31040) = 2.79, P = .006) for the normalized worry and0.7228 (t(31040) = 41.68, P < .0001) for the intercept. Both models outperform a null intercept-only model in terms of reduc-tion of the variance (for the model without worry: F(2,96) = 49.88, P < .0001 and with worry: F(2,96) = 36.98, P < .0001. Thecomparison of the goodness-of-t of these two models is done in terms of BIC difference. The difference of 11.018 in BIC pro-vides a very strong support for the model with worry.4 If we perform model comparison at an individual level, we obtain

3 This measure of reaction time has to be used with caution as the times to answer are affected by scale usage.4 Note that if we use the level of worry reported and not its normalized value we obtain a difference of 273 in terms of BIC between the two models.

Table 2Mean values of the diffusion model parameters for the different models estimated. (1) is the model estimated on the whole data set; (2) allows the boundary separation, a, to vary under high and low worry; (3) allowsthe drift rate, v, to vary; (4) allows the boundary separation and the drift rate to vary; (5) allows the previous parameters and the non-decision time, Ter, to vary; (5) allows to previous parameters and the bias, z, tovary; (6) allows the previous parameters and the inter-trial variability parameters (eta for the drift rate, sz for the bias and st for the non-decision time) to vary. When the parameters are estimated under high and lowworry a t-test of difference is displayed with the value of the difference, the t statistic and the P-value.

Parameters (1) (2) (3) (4) (5) (6) (7)

a 0.1735 High 0.2591 Low 0.2414 0.2008 High 0.2591 Low 0.2424 High 0.2715 Low 0.2392 High 0.2731 Low 0.2409 High 0.2435 Low 0.2336diff = 0.0099 diff = 0.0167 diff = 0.0323 diff = 0.0321 diff = 0.0099t = 0.2845, P = .777 t = 0.3862, P = .700 t = 0.7270, P = 469 t = 0.7192, P = .474 t = 0.3051, P = .761

v 0.1512 0.1638 High 0.1621 Low 0.1849 High 0.1353 Low 0.1402 High 0.1585 Low 0.1639 High 0.1572 Low 0.1639 High 0.1408 Low 0.1609diff = 0.0228 diff = 0.0049 diff = 0.0054 diff = 0.0334 diff = 0.0201t = 0.8575, P = .394 t = 0.1094, P = .913 t = 0.0706, P = .944 t = 0.4286, P = .669 t = 0.7038, P = .484

z 0.0902 0.1073 0.1035 0.1036 0.1052 High 0.1066 Low 0.1059 High 0.1103 Low 0.1083diff = 0.0007 diff = 0.0020t = 0.3032, P = .770 t = 1.2021, P = .233

Ter 0.8141 0.8908 0.8114 0.8606 High 0.8737 Low 0.8871 High 0.8758 Low 0.8924 High 0.9142 Low 0.9286diff = 0.0134 diff = 0.0166 diff = 0.0144t = 0.6334, P = .528 t = 0.7461, P = .458 t = 0.5795, P = .564

eta 0.2190 0.2755 0.2705 0.2675 0.2741 0.2713 High 0.2835 Low 0.2918diff = 0.0084t = 0.5770, P = .566

sz 0.0850 0.0912 0.0891 0.0871 0.0951 0.0952 High 0.1036 Low 0.1043diff = 0.0007t = 0.1480, P = .883

st 0.5276 0.6075 0.4990 0.5985 0.5786 0.5829 High 0.6090 Low 0.6396diff = 0.0306t = 0.6364, P = .566

S.Massoni/Consciousness

andCognition

29(2014)

189198

195

196 S. Massoni / Consciousness and Cognition 29 (2014) 189198Fig. 2. Metacognitive abilities. Comparison of the mean values of the metacognitive abilities measured on the trials with worry (in white) and on the trialswithout worry (in gray). The error bars represent the standard errors of each mean. (A) Comparison of the calibration index. The decrease of overcondenceunder worry is statistically signicant at 5% (P = .044). (B) Comparison of the difference between meta-d0 and d0 . The increase of discrimination ability underworry is statistically signicant at 5% (P = .032).strong support for the model with worry, as the mean difference in BIC is 6.497 (s.d. 11.21) with a range of values from 0.053 to73.001. We thus conclude that taking into account the level of worry in a model of condence formation improves the t of thismodel when explaining subjective probabilities in a perceptual task.

4. Discussion

We studied how emotions impact the metacognitive abilities. We induced emotions by an experimental paradigm withframing effects based on loss aversion and the magnitude of the stakes. Even if this design is less common than mood induc-tions by priming or pharmacological manipulations (see Phelps, 2009, for a review of the available mechanisms), it has theadvantage of being more ecologically valid and can match with the experimental constraints of multiple elds (psychology,economics, and neurosciences). We measured worry by a self-report scale for this same reason of ecological validity. Thismeasurement might be less accurate than physiological measures but it has been shown to be efcient for relating currentlyexperienced emotions (Mauss & Robinson, 2009). Furthermore in our design worry could vary over the experiment and isendogenous to the choice. This is preferable to inducing exogenous worry (priming) or between subjects variations (phar-macological). We found that a negative emotion, worry, was associated with an increase of the quality of calibration (dim-inution of the bias) and discrimination (decrease of the difference between meta-d0 and d0). The distribution of statedcondence under worry was different from those without worry. This change of ratings strategy could be associated witha higher willingness for control due to a negative emotional mood. Assuming that emotion provides additional energy,the lack of effect on the type I decision gives support to the idea that this supplement of attention and control is investedin the type II decision. Note that in our study the rewards are based on the accuracy of the elicited condence and not onthe accuracy of the perceptual decision. Even if we cannot exclude that it may bring the attention of the subjects moreon the condence than on the task we assume that this effect is not enough strong to invalidate our results. First the accuracy

Fig. 3. Distribution of condence. Representation of the condence distributions for high worry (in blue) and low worry (in magenta) trials with a kernelestimation. The condences under high worry are more bunched in the low level while the condences under low worry have more high values. Thisconrms the difference of use of the scale under worry as shown by the KolmogorovSmirnov test of difference. (For interpretation of the references to colorin this gure legend, the reader is referred to the web version of this article.)

of the condence is linked to the accuracy of the answer. Furthermore the results of the literature show that the performancein a discrimination task is increased for emotional stimuli or emotional cues and not for a neutral stimulus as ours

S. Massoni / Consciousness and Cognition 29 (2014) 189198 197(Zeelenberg, Wagenmakers, & Rotteveel, 2006; Yiend, 2010). Nevertheless one drawback of our experiment is that therecording of reaction times for condence was imperfect and forbids a dynamic analysis of condence. Indeed it could beinteresting to see whether the effect of emotions will change the dynamic process of subjective probability formation. Thisquestion could be addressed by tting a dynamic type II diffusion model (Pleskac & Busemyer, 2010; Ratcliff & Starns, 2009).

If worry has such positive impacts on the metacognition it would be useful to understand what factors could explain theindividual tendency to worry (see Keogh, French, & Reidy, 1998, for an extensive study). We studied the relationshipsbetween mean level of worry by subjects and different psychological measures asked in a pre-experiment questionnaire.The tendency to worry, measured by the PSWQ (Meyer, Miller, Metzger, & Borkovec, 1990) is highly correlated with themean level of worry (r = .2169, P = .034) showing that an aggregate measure could be useful to anticipate the individual levelof worry felt facing choices. We also measure two factors of the Big Five Inventory (John, Donahue, & Kentle, 1991) and nd asignicant positive relationship of worry with neuroticism (r = .2047, P = .045) and a non-signicant negative one with con-scientiousness (r = .1201, P = .244). A scale of mood introspection (BMIS, Mayer & Gaschke, 1988) providing different oppo-sitions of moods5 did not give any signicant relationship with worry. Other factors such as preferences (risk and loss aversion,impatience, risky behaviors, etc.) lead also to non-signicant results. These correlations between personality and worry help usto dene a typology of worried individuals. Thus even if the emotional mood is not the primary focus of a study it will be helpfulto measure it, at least indirectly by questionnaires, to take into account the potential bias induced by worry in type II decision.

In the present study we take worry as a proxy of a negative emotion (anxiety) but the emotions have a multiplicity ofcomponents and implications for decision-making (Pster & Bohm, 2008, for a review). Our result that links emotions toan increase in metacognitive abilities should be replicated with different types of emotions, with the same and oppositevalence, in order to be generalized.

Acknowledgments

This work was supported by the ANR France under Grant Riskemotion ANR-08-RISKNAT. I am grateful to Steve Flem-ing, Thibault Gajdos, Geraint Rees, Jean-Christophe Vergnaud and two anonymous reviewers for insightful comments.

References

Baranski, J., & Petrusic, W. (1994). The calibration and resolution of condence in perceptual judgments. Perception and Psychophysics, 55(4), 412428.Barrett, A., Dienes, Z., & Seth, A. K. (2014). Measures of metacognition on signal-detection theoretic models. Psychological Methods, 18(4), 535552.Bechara, A., Damasio, H., & Damiaso, A. R. (2000). Emotion, decision making and the orbitofrontal cortex. Cerebral Cortex, 10(3), 295307.Becker, G. M., DeGroot, M. H., & Marschak, J. (1964). Measuring utility by a single-response sequential method. Behavioral Science, 9(3), 226232.Bishop, S. J. (2008). Trait anxiety and impoverished prefrontal control of attention. Nature Neuroscience, 12, 9298.Brainard, D. H. (1997). The psychophysics toolbox. Spatial Vision, 10, 433436.Damasio, A. R. (1994). Descartes Error: Emotion, Reason, and the Human Brain. New York: Grosset/Putnam.David, A. S., Bedford, N., Wiffen, B., & Gilleen, J. (2012). Failures of metacognition and lack of insight in neuropsychiatric disorders. Philosophical Transactions

of the Royal Society B, 367(1594), 13791390.De Martino, B., Camerer, C. F., & Adolphs, R. (2010). Amygdala damage eliminates monetary loss aversion. PNAS, 107(8), 37883792.De Martino, B., Kumaran, D., Seymour, B., & Dolan, R. J. (2006). Frames, biases, and rational decision-making in the human brain. Science, 313(5787),

684687.Dienes, Z., & Seth, A. (2010). Gambling on the unconscious: A comparison of wagering and condence ratings as measures of awareness in an articial

grammar task. Consciousness & Cognition, 19, 674681.Dugas, M. J., Gosselin, P., & Ladouceur, R. (2001). Intolerance of uncertainty and worry: Investigating specicity in nonclinical sample. Cognitive Therapy and

Research, 25(5), 551558.Eysenck, M. W., Derakshan, N., Santos, R., & Calvo, M. G. (2007). Anxiety and cognitive performance: Attentional control theory. Emotion, 7(2), 336353.Fleming, S. M., & Dolan, R. J. (2009). Effects of loss aversion on post-decision wagering: Implications for measures of awareness. Consciousness and Cognition,

19(1), 352363.Fleming, S. M., & Dolan, R. J. (2012). The neural basis of accurate metacognition. Philosophical Transactions of the Royal Society B, 367(1594), 13381349.Fleming, S. M., Huijgen, J., & Dolan, R. J. (2012). Prefrontal contributions to metacognition in perceptual decision-making. Journal of Neuroscience, 32(18),

61176125.Fleming, S. M., Weil, R. S., Nagy, Z., Dolan, R. J., & Rees, G. (2010). Relating introspective accuracy to individual differences in brain structure. Science,

329(5998), 15411543.Gajdos, T., Massoni, S., Vergnaud, J.-C. (2014). Condence measurement in the light of signal detection theory. Unpublished manuscript.Galvin, S., Podd, J., Drga, V., & Whitmore, J. (2003). Type 2 tasks in the theory of signal detectability: Discrimination between correct and incorrect decisions.

Psychonomic Bulletin & Review, 10(4), 843876.Garnkel, S. N., Barrett, A. B., Minati, L., Dolan, R. J., Seth, A. K., & Critchley, H. D. (2013). What the heart forgets: Cardiac timing inuences memory for words

and is modulated by metacognition and interoceptive sensitivity. Psychophysiology, 50(6), 505512.Gray, J. R., Braver, T. S., & Raichle, M. E. (2002). Integration of emotion and cognition in the lateral prefrontal cortex. PNAS, 99(6), 41154120.Hartley, C. A., & Phelphs, E. A. (2012). Anxiety and decision-making. Biological Psychiatry, 72(2), 113118.Harvey, N. (1997). Condence in judgment. Trends in Cognitive Sciences, 1(2), 7882.John, O. P., Donahue, E. M., & Kentle, R. L. (1991). The big ve inventory versions 4a and 54. Berkeley, CA: University of California, Berkeley, Institute of

Personality and Social Research.Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263291.Keogh, E., French, C. C., & Reidy, J. (1998). Predictors of worry. Anxiety, Stress & Coping: An International Journal, 11(1), 6780.Kunreuther, H. et al (2002). High stakes decision making: Normative, descriptive and prescriptive considerations. Marketing Letters, 13(3), 259268.

5 We use the following BMIS index: pleasant vs. unpleasant, arousal vs. calm, positive vs. tired, and negative vs. relaxed.

LeDoux, J. (1996). The Emotional Brain: the Mysterious Underpinnings of Emotional Life. New York: Simon and Schuster.Levitt, H. (1971). Transformed up-down methods in psychoacoustics. Journal of the Acoustical Society of America, 49(2), 467477.MacLeod, A. K., Williams, M. G., & Bekerian, D. A. (1991). Worry is reasonable: The role of explanations in pessimism about future personal events. Journal of

Abnormal Psychology, 100, 478486.Maniscalco, B., & Lau, H. (2012). A signal detection theoretic approach for estimating metacognitive sensitivity from condence ratings. Consciousness and

Cognition, 21(1), 422430.Mauss, I. B., & Robinson, M. D. (2009). Measures of emotion: A review. Cognition & Emotion, 23(2), 209237.Mayer, J. D., & Gaschke, Y. N. (1988). The experience and meta-experience of mood. Journal of Personality and Social Psychology, 55(1), 102111.McCurdy, L. Y., Maniscalco, B., Metcalfe, J., Liu, K. Y., de Lange, F. P., & Lau, H. (2013). Anatomical coupling between distinct metacognitive systems for

memory and visual perception. The Journal of Neuroscience, 33(5), 18971906.Meyer, T. J., Miller, M. L., Metzger, R. L., & Borkovec, T. D. (1990). Development and validation of the Penn state worry questionnaire. Behaviour Research and

Therapy, 28(6), 487495.Nieder, A., & Dehaene, S. (2009). Representation of number in the brain. Annual Review of Neuroscience, 32(1), 185208.Pacheco-Unguetti, A. P., Acosta, A., Callejas, A., & Lupianez, J. (2010). Attention and anxiety: Different attentional functioning under state and trait anxiety.

Psychological Science, 21(2), 298304.Persaud, N., McLeod, P., & Cowey, A. (2007). Post-decision wagering objectively measures awareness. Nature Neuroscience, 10, 257261.Pessoa, L. (2008). On the relationship between emotion and cognition. Nature Reviews Neuroscience, 9(2), 148158.Pster, H.-R., & Bohm, G. (2008). The multiplicity of emotions: A framework of emotional functions in decision making. Judgment and Decision Making, 3(1),

517.Phelps, E. A., Ling, S., & Carrasco, M. (2006). Emotion facilitates perception and potentiates the perceptual benets of attention. Psychological Science, 17(4),

292299.Phelps, E. A. (2009). The study of emotions in neuroeconomics. In P. W. Glimcher, C. F. Camerer, E. Fehr, & R. A. Poldrack (Eds.), Neuroeconomics: Decision

making and the brain (pp. 233248). London, England: Academic Press.Pleskac, T. J., & Busemyer, J. R. (2010). Two-stage dynamic signal detection: A theory of choice, decision time, and condence. Psychological Review, 117(3),

864901.Raftery, A. E. (1995). Bayesian model selection in social research. Sociological Methodology, 25, 111163.Ratcliff, R., & McKoon, G. (2008). The diffusion decision model: Theory and data for two-choice decision tasks. Neural Computation, 20(4), 873922.

198 S. Massoni / Consciousness and Cognition 29 (2014) 189198Ratcliff, R., & Starns, J. J. (2009). Modeling condence and response time in recognition memory. Psychological Review, 116(1), 5983.Rounis, E., Maniscalco, B., Rothwell, J., Passingham, R., & Lau, H. (2010). Theta-burst transcranial magnetic stimulation to the prefrontal cortex impairs

metacognitive visual awareness. Cognitive Neuroscience, 1(3), 165175.Schurger, A., & Sher, S. (2008). Awareness, loss aversion, and post-decision wagering. Trends in Cognitive Sciences, 12, 209210.Schwarz, G. (1978). Estimating the dimension of a model. Annals of Statistics, 6, 461464.Seymour, B., & Dolan, R. J. (2008). Emotion, decision making and the amygdala. Neuron, 58(5), 662671.Sokol-Hessner, P., Camerer, C. F., & Phelps, E. A. (2013). Emotion regulation reduces loss aversion and decrease amygdala responses to losses. Social Cognitive

& Affective Neuroscience, 8(3), 341350.Sokol-Hessner, P., Hsu, M., Curley, N. G., Delgado, M. R., Camerer, C. F., & Phelps, E. A. (2009). Thinking like a trader reduces individuals loss aversion. PNAS,

106(13), 50355040.Song, C., Kanai, R., Fleming, S. M., Weil, R. S., Schwarzkopf, D. S., & Rees, G. (2011). Relating inter-individual differences in metacognitive performance on

different perceptual tasks. Consciousness and Cognition, 20(4), 17871792.Takahashi, H. (2013). Molecular neuroimaging of emotional decision-making. Neuroscience Research, 75(4), 269274.Tom, S. M., Fox, C. R., Trepel, C., & Poldrack, R. A. (2007). The neural basis of loss aversion in decision-making under risk. Science, 315(5811), 515518.Vandekerckhove, J., & Tuerlinckx, F. (2007). Fitting the Ratcliff diffusion model to experimental data. Psychonomic Bulletin & Review, 14(6), 10111026.Vandekerckhove, J., & Tuerlinckx, F. (2008). Diffusion model analysis with MATLAB: A DMAT primer. Behavior Research Methods, 40(1), 6172.Vuilleumier, P. (2005). How brains beware: Neural mechanisms of emotional attention. Trends in Cognitive Sciences, 9(12), 584594.White, C., Ratcliff, R., Vasey, M., & McKoon, G. (2010). Anxiety enhances threat processing without competition among multiple inputs: A diffusion model

analysis. Emotion, 10(5), 662677.Yiend, J. (2010). The effects of emotion on attention: A review of attentional processing of emotional information. Cognition & Emotion, 24(1), 347.Yokoyama, O. et al (2010). Right frontopolar cortex activity correlates with reliability of retrospective rating of condence in short-term recognition

memory performance. Neuroscience Research, 68(3), 199206.Zeelenberg, R., Wagenmakers, E-. J., & Rotteveel, M. (2006). The impact of emotion on perception: Bias or enhanced processing. Psychological Science, 17(4),

287291.

Emotion as a boost to metacognition: How worry enhances the quality of confidence1 Introduction2 Methods2.1 Participants2.2 Stimuli2.3 Task and procedure2.3.1 Practice and thresholding2.3.2 Experiment phase2.3.3 Payment

2.4 Data analysis

3 Results4 DiscussionAcknowledgmentsReferences