Authors: Roeland Heerema, Mathias Pessiglione
Categories: Article, Computational neuroscience, Human behaviour
Source: Communications Psychology
Authors: Roeland Heerema, Mathias Pessiglione
When making decisions, humans are susceptible to all sorts of biases, relative to rational norms. An important factor is incidental changes in affective states, such as variations in mood between happiness and sadness. We previously developed a computational model, in which mood affects choice by forming a predisposition to face costs and seek more rewards. Here, we generalized this theory to account for how specific inductions of happiness and sadness affect different types of economic decisions involving a tradeoff between costs (risk, delay, effort) and benefits (financial rewards). Across exploratory and confirmatory studies (N = 94), we observed a consistent bias exerted by transitory mood states, whether they were assessed through self-reports (rated happiness minus sadness) or inferred from physiological measures (valence of facial expression times intensity of autonomous arousal). This choice bias was best explained by our computational model, with a mood-scaled bonus added to the value of the more rewarded but more costly option, irrespective of the cost type (risk, delay or effort). Additionally, gaze tracking during decision making confirmed that the choice bias was driven by an early preference for the mood-congruent option. Together, these results demonstrate the feasibility of predicting irrational choices from objective measures of affective states.
According to rationality axioms of economic decision theory, our preferences regarding attributes of choice options should be stable and consistent^1^. However, people vary in their attitude towards key attributes, such as risk, showing systematic deviations from rational choice behaviour^2^. Among the sources of variability are changes in internal states, such as mood fluctuations. Field observations have reported spectacular cases in which incidental mood changes spill over onto unrelated decisions. For example, people take more risk when the weather is nice or after a victory of the local sports team, both in personal gambling and on the stock market^3–5^. These real-world mood effects have been reproduced in experimental mood fluctuations induced by series of positive or negative feedbacks incline participants into accepting or declining risky challenges^6,7^.
Although mood-related changes in preferences are clearly irrational when mood triggers are independent from choice outcomes, they may arise from what would be considered as adaptations from a behavioural ecology perspective, which is notably taken in foraging theory^8^. In recent theoretical work, we have shown, using simulations of foraging behaviour, that mood-related changes in preferences are indeed adaptive when sources of costs and rewards in the environment are correlated, as is the case across seasonal fluctuations^9^. Mood is defined here as an affective state that can be positioned on a continuum from sadness to happiness and that varies with the experience of positive and negative outcomes. Our model assumes that mood shifts the tradeoff between costs and benefits in the decision to forage for food reward. This assumption is applied not only to risk attitude but also to time preference and effort when in a good mood, theoretical agents are more willing to take more risks, wait for longer delays and exert more strenuous efforts for the prospect of collecting bigger rewards.
We have validated our computational model of choice behaviour, using an experimental design where changes in mood were induced by series of positive or negative feedback to the answers that participants provided to general knowledge quiz questions^10^. Mood was assessed with a subjective rating on a scale graduated from very bad to very good, and its effects on choices were captured by a bonus parameter added to the value of bigger but more costly rewards. However, subjective ratings have been criticized on the grounds that they are bounded by the quality of insight and susceptible to a potential demand effect (when participants can guess what experimenters expect them to report). The first aim of the present study was to circumvent these limitations by collecting objective physiological markers (pupil diameter and skin conductance) of autonomic arousal, which by definition cannot be voluntarily controlled and therefore remain immune to demand effects^11–14^. As autonomic arousal only reflects the intensity of mood changes^15,16^, we also recorded the electromyographic activity of the smiling and frowning muscles (zygomaticus and corrugator), which respectively reflect the positive and negative valence of mood states, i.e. the relative expression of happiness and sadness^17–20^. Thus, the idea was to build and validate an analytic pipeline that would enable inferring mood from objective markers that do not depend on subjective reports.
The second aim of the present study was to generalize our theoretical account of mood effects on economic decisions to more transient states of happiness and sadness that have equally been explored as emotional states. There is no consensual taxonomy of affective states, which form a superordinate category that includes both emotions and moods^21^. Emotions are generally considered as multi-faceted phenomena, triggered by well-defined events that give rise to specific subjective experience, physiological responses, cognitive appraisals, and action plans^22–25^. They have been classically divided into discrete categories (such as anger and fear) that are shared across cultures and species^26,27^ and associated with specific behaviours (such as aggression and escape). Moods are usually defined as more diffuse or global affective states that can last long enough to impact incidental decisions across different contexts. In this sense, they would represent temporal extensions of emotions described as transitory sadness/happiness that may follow negative/positive experience. Our working hypothesis is that happiness and sadness should induce the same shift in preference (bonus for bigger but more costly rewards), irrespective of whether their manifestations are best described as short-lived emotions or long-lasting moods. The prediction is therefore that procedures meant to induce emotions of happiness and sadness should result in the same effects on risk, delay and effort discounting that we observed in our previous study when inducing episodes of positive and negative mood^10^.
To induce happiness and sadness, we used guided imagery vignettes engaging foreground attention combined with congruent music in the background, a procedure that was previously validated based on subjective experience^28,29^. We opted for this procedure because it was reported to be the second most effective one in a meta-analysis^30^, the first one being movies, which we discarded because variations in luminance would have blurred measures of pupil size. Vignettes and music can be used as well to induce other basic emotions, such as anger and fear, which we took as controls for the specificity of mood effects on choices. The idea was not to establish an exhaustive mapping of how emotions affect decisions, but to test prototypic examples that can be compared with sadness and opposed to happiness (because of their negative valence). Hereafter, we keep the term mood to designate the position of the affective state on the sadness-happiness dimension, with the aim of maintaining a link with our theoretical model and empirical observations exposed in previous studies, even if moods in the present study were more transient. We note that such transitory affective states have been called mood by other groups^31,32^, including when using the same induction procedure with text and vignette^28^. Although this is clearly not the only possible meaning of the word, this unidimensional definition is pertinent because mood is typically said to vary from ‘good’ to ‘bad’^33,34^, and because it provides continuity with mood disorders in psychiatry (extreme states of sadness/happiness being seen in depressive/manic episodes).
We are certainly not the first to explore the impact of affective states on cost/benefit tradeoffs. The very idea that incidental emotions can influence judgement and decision-making has been discussed by philosophers for centuries, and empirically tested in different fields of research such as economics^35^, psychology^36^ and neuroscience^37^. However, meta-analyses have emphasized that the evidence is not consistent across studies^38^, as can be illustrated by the literature on risk attitude. Happiness following receipt of a candy gift was reported to increase risk taking^39^, but other studies found that the effect was restricted to low-risk gambles, the opposite effect being observed with high-risk gambles^40^. Anger induced with biographic recall or video clips was once shown to increase risk taking^41^, but another study only found the effect in men but not women^42^. Fear induced by threat of electric shocks was said to increase risk aversion in financial decisions^43^ but fear induced by a writing task was found to increase risk taking in human interactions^44^. In studies inducing negative emotions with written scenarios, some observed that anxious and sad states were respectively driving risk aversion and risk seeking^45^, but others^46^ observed that all negative emotions (including anger, fear and sadness) increased risk seeking. A similar degree of inconsistency appears in the literature on how emotions impact decisions that involve delay an effort. As a systematic review would be beyond the scope of this experimental paper, we refer readers to our online table (https://github.com/MBB-team/MoodChoicePhysiology2025) that lists all papers (up to this date) reporting effects of incidental emotions on economic choices involving cost-benefit tradeoff.
Several factors may have contributed to the non-replicability of the links between emotions and decisions. A first factor is the diversity of methods used to induce emotions and to probe decisions, a reason why we systematically employed the same procedure to induce all emotions, and the same format for all types of cost in our economic choice task. Another issue is the use of comparisons between groups, which may occasion false positives related to individual preferences, especially when combined with a small sample size. Indeed, studies using within-subject designs to compare emotional states induced by music, as we have done here, found more robust results, with more risk seeking in happy states and more risk aversion in sad states^47,48^. An even more critical issue is that the emotional state targeted by the induction is rarely assessed, such that decisions are compared between conditions (e.g., between groups that do or do not receive a happiness-enhancing gift). We argue here that decisions should be related to individual experiences, which may differ between participants for the same induction. For this reason, we collected both subjective ratings of experienced emotion and objective markers of intensity and valence. The aim of the analyses was then to test the link between mood indicated by either self-reports or physiological markers, and the preference for costly options expressed in economic choices.
These studies were not preregistered. However, our predictions regarding the effects of mood changes on economic choices were directly determined by simulations of our computational model presented in a previous publication^9^, and our analytic pipeline for assessing those effects was identical to that used in another previous publication^10^.
Participants were recruited through subject pools of the Paris Brain Institute and of the RISC (‘Relais d’Information sur les Sciences de la Cognition’), a collective of labs that conduct experiments in psychology and neuroscience in the Paris region. The announcement of the study stated that we were interested in “cognition, emotion, and physiology”, without mention of the scientific purpose (testing the effects of emotions on decisions). Candidates were eligible if they were native French speakers between 18 and 50 years old, with no history of psychiatric or neurological disorders. They were not allowed to consume alcohol or any psychotropic drugs during the day of the experiment. Further inclusion criteria applied specifically to physiological recordings. For eye-tracking, candidates had to be able to do the entire experiment without glasses or contact lenses. For electromyography, the requirements were not to have any facial tattoos or piercings, not to have make-up on, and to be clean-shaved around the mouth and cheeks.
In total, 4 samples of participants were recruited and tested on slightly different versions of the experimental design. The first two studies (N = 35) were considered as exploratory, because they included two emotions (anger and fear) for which the effect was uncertain, in addition to the two emotions (happiness and sadness) that were meant to generalize the effect observed with mood manipulations in our previous study^10^. The last two studies (N = 59) were considered as confirmatory, because the objective was to replicate the impact of mood (scored as rated happiness minus rated sadness) on economic choices that we observed in exploratory studies. Other important differences between studies were the display of choice options (written numbers versus graphical illustrations) and the recording of physiological measures (see Supplementary Note 1 for details). In the end, 97 individuals registered to take part in one of the four studies. One invited person could not be included because his beard prevented facial electrodes to be properly placed. One person’s participation was prematurely aborted due to technical errors, and one participant’s dataset was lost due to overwriting. Thus, the total number of included datasets was 94 (35 men and 59 women; mean age 26.2 ± 0.55). In accordance with French regulations, information about race/ethnicity was not recorded.
Candidates who enroled online filled in trait questionnaires that measured, depending on the version of the study, a subset of the impulsivity, depression, anxiety, apathy, hypomania, anger rumination, vividness of mental imagery, and openness to experience. The scores on some of these scales have served as exclusion criteria if they were in the range of de facto clinical values. For a demographic overview summarizing descriptive statistics of each study, see Supplementary Table 1.
Invited participants gave written informed consent to partake in the studies, which were approved by the ethics committee of the Sorbonne University in Paris. In the oral briefing prior to the experiment, participants were explained that the first goal was to do their best to experience the emotions that the presented text and music would convey, and that the second goal was to be truthful in expressing their preferences in the choice task, as there were no right or wrong answers. They were also told that the payment for their participation would depend on the choices that they would make during the experiment, three of which would be randomly selected and implemented (see below).
The main experiment was composed of 60 or 75 inductions that were divided into 2 or 3 sessions with short breaks in between (see Supplementary Fig. 1 for an illustration of the differences between studies). The 10-s emotion induction relied on a specific imagery scenario (only used once) presented as a text vignette in a red coloured font that was luminance-adjusted to the grey background to minimize the effect on pupil dilation. The vignette was accompanied by music (except for neutral inductions), which kept on playing during choices in order to maximize the expected emotional spill-over. Participants made a short series of risk, delay, and effort choices, interleaved with a 1–2 s jittered fixation cross. The minimum decision time was fixed to 0.5 s. The number of decisions presented after an induction differed between studies, ranging from 6 to 12. At the end of the choice battery, music stopped playing and participants rated their emotions. To ensure proper emotional washout before the following induction, the screen was then turned black for 8 s with a message that “Please relax and empty your mind as you await the next trial”. Then, a 2-s fixation cross preceded the next emotion induction.
During instructions, participants viewed a few vignettes for familiarisation with the emotion induction and rating procedures prior to the main experiment. The vignettes consisted of short, personally relevant sentences, such as (for happiness): “You leave for a holiday. Upon your arrival, your destination looks like paradise.”, or (for sadness): “Your pet that you were very attached to just passed away.” For the neutral induction, the vignettes were impersonal, encyclopaedic sentences adopted from Wikipedia, for “The novel is a literary genre that is essentially characterized by a fictional narrative.” We initially collected 24 vignettes per category and had them rated by an independent pilot group of participants (see Supplementary Table 3 for a full overview). Some of the vignettes were adapted from Mayer and colleagues^28^, but most of them were originally written because we needed many more for all our inductions.
The vignettes were paired with emotionally laden pieces of music in all but the neutral induction. The music extracts consisted of diverse fragments of classical music from Baroque to contemporary composers (notably including compositions for movies and video games). Classical music is a genre that can be associated with each of our intended emotion categories, which made it suitable for our purposes. We included only instrumental music because we did not want lyrics to interfere with the text contents of the vignettes. We avoided overly rhythmical pieces that could incite the participant to tap along, which could possibly interfere with the physiological recordings. Per emotion category, we initially collected 16 pieces (see Supplementary Table 2 for a full overview), some of which were adapted from Mayer and colleagues^28^ and from Schulreich and colleagues^48^; others were created for the occasion. Each music extract had the sound peak amplitude normalized to -1dB and was introduced/concluded by a 1 s fade in/out. We used Audacity® for sound editing.
During the experiment, rating screens displayed the question “To which extent did you feel each of the following emotions?”. Emotions were happiness and sadness in all cases, fear and anger only in exploratory studies, and curiosity only in confirmatory studies. Participants answered by placing a cursor on continuous rating scales that ranged from 0, labelled “not at all”, to 10, labelled “maximally”. The ratings were collected after the choices to prevent explicit self-report from interfering with the impact of affective states on decision making.
Prior to the main experiment, participants also viewed instructions and examples of the economic decisions that they would be making later on following emotion inductions.
To probe risk discounting, participants were offered either a safe option with a variable small reward (<30€), or a lottery where 30€ could be won, but 10€ could be lost. The probability of winning and the risk of losing (1—winning probability) were either written as percentages on screen (exploratory studies) or illustrated as pie charts (confirmatory studies). These choices implement an intuitive notion of risk (probability of losing money instead of winning money), by combining probability discounting and loss aversion.
To probe delay discounting, participants were offered the option to receive a variable small reward that would be paid out in cash after the experiment, or a larger sum of 30€ that would be received by wire transfer after a delay of up to 1 year indicated on screen. In exploratory studies, the delay was explicitly written, whereas in confirmatory studies, it was indicated within a calendar pictogram.
To probe effort discounting, participants were offered either to pedal on a fitness bike for 15 min at an indicated effort level and obtain 30€, or to spend the 15 min on the bike without doing any physical exercise (0-effort level) and receive a lower sum of money. The 0-effort level was implemented to match the opportunity cost of time between the two options. Effort levels were either explicitly written (exploratory studies) or indicated as the slope of a schematized route (confirmatory studies). For participants to understand the meaning of effort levels, they were trained beforehand on a fitness bike during a calibration session. The maximal effort level was adjusted individually by asking participants to maintain a steady speed of 50 rotations per minute (RPM), while the power level was increased step by step every 20 s. Participants then experienced various effort levels from a flat slope (no pedalling) to a 45° slope (50 RPM at 80% of the maximal power they could sustain for consecutive 20 s).
Before the main experiment, participants viewed 10 example choices per cost type, covering the full range of rewards and costs. Participants then proceeded to a choice calibration that was meant to identify individual indifference curves for each cost type, i.e. the pairs of options with equal subjective value. This was achieved using a bisection method adapted from Wiehler and colleagues^49^ in exploratory studies, and using online trial generation previously described in Heerema et al. in confirmatory studies^10^. With the Variational Bayesian Analysis (VBA) toolbox^50^, discount functions were fitted to the calibration choices and then used to generate trials such that half the choices were at indifference, one third near indifference (±10% of equivalent small reward for the selected cost level) and one out of six further from indifference (±20%). In the main experiment, small rewards and costs associated with the big reward (30€) were drawn such that choice trials were matched between the different emotions induced.
After completing all trials, participants viewed the three randomly selected choices that would be implemented (one per cost type). This involved watching the outcome of a lottery (in case the risky option was chosen) and spending 15 minutes on the fitness bikes (in all cases). All rewards were immediately given in cash, except if the delayed option was chosen, in which case it was delivered through bank transfer. The selection of choices was actually pseudo-randomized to minimize payoff variance around the usual compensation for behavioural experiments. The average payoff across studies was 48.60 ± 1.50€ (for a duration between two and three hours).
At the very end of the experiment in confirmatory studies, participants filled in a short paper debriefing questionnaire. They were asked about the purpose of the study, whether they succeeded in experiencing the emotions, and whether they thought their emotions had influenced their decisions (see Supplementary Note 6 for the results).
The behavioural dependent variables were subjective ratings of experienced emotion, choices of costly versus uncostly option, and choice response times (RT). RT data were trimmed (by removing RT shorter than 0.75 s or mean minus 3 SD, and RT longer than 10 s or mean plus 3 SD) and detrended (by removing the linear effect of trial number). Rating and RT data were z-scored before being compared between inductions. Choices were first analysed in a model-free manner, as described in the Results. Following a random-effect approach, summary statistics (e.g., difference in choice rate between emotional conditions, weight of subjective rating in linear regressions, computational parameter fitted to choice data, etc.) were first computed at the individual level and then tested for significance at the group level, using two-tailed Student’s t-tests as implemented in MATLAB® R2021a.
Choices were further analysed using computational modelling, to better specify the effect of mood fluctuations on decision-making. Choice probability was modelled with a softmax decision rule, which applies a sigmoidal mapping from decision value to selection probability. Decision value is the difference between option values, computed with a discount function specific to the considered cost type (see below). The parameterisation of the choice and discount functions gives rise to a space of possible model variants, which we navigated as follows.
In the softmax decision function, stochasticity (i.e., the sigmoid slope) was adjusted by the weight (inverse choice temperature β1) on decision value. A priori preference for the costly option (i.e., the sigmoid intercept) was captured by a bonus (choice bias β0) added to the weighted decision value. We tested variants where choice bias and inverse temperature were fixed (to 0 and 1, respectively) or variable, either across cost types or per cost type.
In the discount functions, three parameters were considered. In their simplest expression, discount functions have one a discount factor, i.e. a weight on the cost (kR, kD, kE). Extensions include a weight on reward kRew common across cost types, and a power parameter on the cost to control the curvature of the discount function (γR, γD, γE). As to the form of the discount functions themselves, we limited the possibilities to the models that were previously validated for the same types of economic choices^10^. For delay discounting, we opted for an exponential decay model, a convex function that asymptotically declines to zero as time to payment extends further^49,51^. For effort discounting, subjective values can be negative for high levels of effort, hence we selected an additive cost term that was quadratic when the power on cost was fixed^52–54^. For risk discounting, we retained a linear function where expected gain is discounted by subtracting expected loss.
Below, the three decision functions and the three discount functions are written with their maximal number of free parameters. Note that discount functions can be equally used to compute the subjective values of costly and uncostly rewards Rew (the cost being set to zero in the latter case).
For risk (of losing L with probability 1-P):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {V}{R}={k}{{Rew}}\cdot {Rew}\cdot P-{k}{R}\cdot L\cdot {(1-P)}^{{\gamma }{R}}
\usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ P\left({risky}\right)={\left(1+ exp (-{\beta }_{1,R}\left({V}_{R,{costly}}-{V}_{R,{uncostly}}\right)-{\beta }_{0,R})\right)}^{-1} $$\end{document}Prisky=1+exp(−β1,RVR,costly−VR,uncostly−β0,R)−1 For 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {V}_{D}={k}_{{Rew}}\cdot {Rew}\cdot exp \left(-{k}_{D}\cdot {D}^{{\gamma }_{D}}\right) $$\end{document}VD=kRew⋅Rew⋅exp−kD⋅DγD4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ P\left({delayed}\right)={\left(1+ exp (-{\beta }_{1,D}\left({V}_{D,{costly}}-{V}_{D,{uncostly}}\right)-{\beta }_{0,D})\right)}^{-1} $$\end{document}Pdelayed=1+exp(−β1,DVD,costly−VD,uncostly−β0,D)−1 For 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {V}_{E}={k}_{{Rew}}\cdot {Rew}-{k}_{E}\cdot {E}^{{\gamma }_{E}} $$\end{document}VE=kRew⋅Rew−kE⋅EγE6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ P\left({effortful}\right)={\left(1+ exp (-{\beta }_{1,E}\left({V}_{E,{costly}}-{V}_{E,{uncostly}}\right)-{\beta }_{0,E})\right)}^{-1} $$\end{document}Peffortful=1+exp(−β1,EVE,costly−VE,uncostly−β0,E)−1 In total, the combinations of described variants gave rise to a space of 36 models that are listed in Supplementary Table 4. Models were inverted and compared using the Variational Bayesian Analysis (VBA) toolbox, accessible at [https://mbb-team.github.io/VBA-toolbox](https://mbb-team.github.io/VBA-toolbox). VBA applies approximate Bayesian inference to nonlinear state-space model fitting under the mean-field and Laplace approximations^50^. VBA takes in the trial-wise option features and fits the designated parameters while taking account of their specified prior means and variances. Parameters *k*, *β*~*1*~, and *γ* were constrained to be positive and had distributions centred on 1. Parameters *β*~*0*~ were free to vary and had a prior distribution centred on 0. VBA outputs the posterior parameter estimates and the model evidence, which provides a tradeoff between accuracy (goodness of fit) and complexity (number of free parameters). For Bayesian model selection^55,56^, group-level random-effects analyses were applied to individual log-model evidence, with the underlying assumption that different participants’ behaviour can be best explained by different models. Outcomes of the analyses were the expected frequency with which each model prevails in the population, and the exceedance probability that a given model is more frequent than all the others in the model space. The winning model, described in “Results”, reached a model frequency of 26.9% and exceedance probability of 92.6% (Supplementary Table 4). The winning model was subjected to a recovery analysis, in which choices were simulated with parameters drawn from their empirical distribution and with costs and rewards drawn from uniform distributions within the range used in the experiments. Simulated choices were then fitted by inverting the winning model, and the obtained parameter values were each regressed against all simulated parameter values. We found that the correlation between the simulated and the fitted parameters was consistently high (all R > 0.64) and that only very little variance was explained by other parameters. We thus concluded that the model parameters were well recoverable. ### Physiology #### Data collection In a subset of participants, sensors from a BIOPAC® MP 150 kit were placed for physiological recordings. On the left hand, one Ag-AgCl (signal) electrode was placed on the fingertip of the middle finger and another (baseline) electrode on the back of the ring finger to measure galvanic skin responses. On the left thumb, a pulse oximeter was placed to obtain a photoplethysmogram for heart period measurement (not reported). Participants placed their left hand on the lab desk and were instructed to avoid moving it during the study. We further placed electrode pairs on the zygomaticus, corrugator, frontalis and depressor muscles on the left part of the face for electromyography. These facial muscles are respectively used for smiling, frowning, pouting, and raising one’s eyebrows. Electrodes placed on the depressor had the tendency to loosen and signals from the frontalis showed considerable crosstalk from the corrugator’s activity. Since the zygomaticus and corrugator are commonly dichotomized as signalling positive and negative emotions^19,20^, we decided to restrict recordings of facial musculature to those muscles after the first sample of participants. Signals were collected at a 1000 Hz sampling rate using BIOPAC’s AcqKnowledge® software. Once all sensors were placed, participants were seated comfortably at approximately 60 cm from a large computer screen (HP S230tm, 50.9 × 28.6 cm active area). A headrest was placed against the back of the head to reduce movement for studies in which pupil size was recorded. An infrared eye-tracking device from The Eye Tribe® (sampling 60 Hz) was positioned under the computer screen and centred on the participant’s eyes. Pupil dilation responses were measured in all cases, but tracking gaze position requires a short calibration where participants follow a moving dot with their eyes, which was only performed in confirmatory studies. For a full overview of the different physiological recordings per study, see Supplementary Fig. 1e. The physiological recordings common to exploratory and confirmatory studies consisted of pupillometry, electrodermal activity, pulse oximetry, and electromyographic (EMG) activity of the zygomaticus, and corrugator muscles. Two full datasets were lost (one saving error and one early interruption), and three EMG datasets were discarded after quality check (too many artefacts), leaving 69 participants for pupil size and skin conductance, and 66 participants for zygomaticus and corrugator EMG activity. #### Data preprocessing Below we describe the preprocessing steps for each of the physiological signals. Data filtering and smoothing was done using functions from Fieldtrip^57^; see [http://fieldtriptoolbox.org](http://fieldtriptoolbox.org). Skin conductance time series were consecutively bandpass-filtered between 1/40 Hz and 1 Hz (4^th^-order Butterworth filter), smoothed with a 1 s boxcar kernel, downsampled to 10 Hz, and epoched to the full duration of the induction (10 seconds). A few participants moved their hand, which caused artefacts in the electrodermal signals. All signals measured during artefactual inductions (0.8 ± 0.4%) were visually detected and removed. The remaining time series were standardized across inductions and baseline-corrected by subtracting the first data point after induction started. EMG time series of the zygomaticus (smiling) and corrugator (frowning) were high-pass filtered at 25 Hz. Line noise was filtered out with a 48-52 Hz stopband filter, after which the signal was rectified by taking the analytic envelope. The signal was then epoched to the induction duration and visually inspected. Inductions with large, prolonged artefacts (1.7 ± 0.4%) were rejected. The signal was then downsampled to 50 Hz and transformed by taking its natural logarithm, so that a normal distribution of the data was approached. Remaining outliers were automatically detected if their value was more than 3 standard deviations above the mean; these samples were removed and the signal was interpolated. Then, the signal was smoothed with a 2 s boxcar kernel and standardized across all inductions, separately for each muscle. Finally, the signals were baseline-corrected. Eye-tracking time series were preprocessed following the pipeline described in Wiehler et al.^49^. Pupil diameter was estimated from eye-tracking data and corrected for the estimated distance between eyes and screen (although headrest minimized changes in this distance). Samples that were not within the median ±5 standard deviations of the signal were discarded as outliers. Eye blinks were detected and labelled by the eye-tracker, along with the 6 samples (100 ms) preceding and following the blink. Signal samples containing outliers, blinks, or absence of a detected pupil were removed and interpolated but flagged for the analysis (see below). Next, a bandpass filter between 1/128 Hz and 1 Hz was applied and the signal was smoothed with a 0.5 s boxcar kernel. The pupil diameter was averaged between the two eyes and the resulting signal was standardized. Next, the signal was epoched to the induction duration and baseline-corrected. Inductions with more than 2/3 of samples being originally flagged as artefacts (4.1 ± 1.3%) were excluded from analyses. #### Statistical analysis Significant clusters in the time window corresponding to induction duration were detected using the VBA toolbox to correct for multiple comparisons (i.e., multiple time points). The correction applied here controls for family-wise error (FWE) based on random field theory, given the estimated temporal dependency of successive samples in the autocorrelated signal. This procedure was used to detect time clusters for which F-statistics were significantly different between emotional and neutral inductions, or between happiness and sadness inductions. For assessing the links with ratings and choices, we needed to reduce the dimensionality of physiological responses to one estimate per induction. For the pupil dilation and EMG responses, this was done by taking the median of the signal over the entire 0-10 s induction window, following common practice^11,58^. The amplitude of the skin conductance response was estimated using PsPM^59,60^, which first extracts a canonical response function and then uses it to deconvolve the skin conductance signal by fitting a general linear model. For each physiological response, the estimates were corrected for the session number (categorical variable), to account for changes in recording quality due to readjustments during breaks. From the resulting corrected response estimates, we extracted valence and arousal, the two dimensions onto which emotions are classically mapped^61–64^. The proxy for valence was construed as a net facial expression signal (the standardized difference between the zygomaticus and corrugator responses). The proxy for arousal was construed as a mean autonomic signal (the average of pupil dilation and skin conductance response). ### Eye-tracking analysis In confirmatory studies, we obtained raw coordinates of gaze position on screen at a sampling rate of 60 Hz. Blinks were automatically detected by the eye-tracker and removed from the signal along with the preceding and following 10 samples. Coordinates were then normalized with respect to the dimensions of the screen and samples that were out of screen bounds were rejected. Data were stored from the onset of any choice trial to when the decision was recorded. Next, each time sample was categorized based on its coordinates. There were four elements of interest on the reward and cost associated with each of the two options. In confirmatory studies, cost levels were symbolically represented in pictograms and monetary rewards were written below. The left vs. right position of the costly and uncostly options was counterbalanced across choice trials and occupied symmetrical frames on screen. For categorization of gaze coordinates, a margin was allowed around the four frames, as illustrated in Fig. 7a, while avoiding any overlap between frames. For the main analysis, we simply counted the number of time samples when gaze was located in each of the four frames. In a separate analysis, we made heatmaps of gaze distribution on screen by convolving each sample with a 2D Gaussian kernel (mean of 1 and standard deviation corresponding to a 3° visual angle). Trial-wise maps were obtained by summation of time-resolved maps across the first second of decision-making. These heatmaps were then averaged across choice trials and compared between experimental conditions. ### Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. ## Results Separate exploratory and confirmatory samples (total *N* = 94; see Supplementary Table 1 for a demographic overview) were recruited to assess the effects of happiness and sadness inductions on economic choices. For all behavioural data (ratings and choices), we therefore report the results of exploratory (*N* = 35) and confirmatory (*N* = 59) studies separately. However, a single pool of participants was considered for the analysis of physiological responses, which were only collected in a subsample (*N* = 66), and for the analysis of gaze position, because eye-tracking data were only collected in confirmatory studies (*N* = 59). For all results we report the effect size, 95% confidence interval, t-statistic (with degrees of freedom), and p-value. Data were visually inspected to be approximately normally distributed using quantile-quantile plots, so as to assure that assumptions for statistical testing were met. ### Emotion manipulation check We opted for a within-subject design, where each participant is presented with stimuli from different emotional categories (Fig. 1a). In exploratory studies (*N* = 35), these categories were happiness, sadness, anger, fear, and an emotionally neutral control condition. In confirmatory studies (*N* = 59), we retained only the happy, sad, and neutral stimuli. Our approach to inducing emotions has been the same in each we presented a text vignette describing an emotional scenario for 10 s, combined with a piece of congruent music that started playing as the vignette appeared on screen (Fig. 1b). In the neutral condition, the vignette was an encyclopaedic sentence without music. Subsequently, participants performed a short battery of economic choices while the music kept on playing (Fig. 1c). Next, the music stopped and participants rated the extent to which they had experienced each of the emotions featured in the experiment. This sequence (rating after choice) was adopted to avoid participants adjusting their decisions to their declared emotions. After an emotional washout period of 8 s, during which participants were told to clear their mind, the next induction followed. The duration of the sequence of events around an induction was about one minute. The order of inductions was pseudo-randomized such that the emotional inductions were distributed uniformly over sessions, which were separated by short breaks. Details about design differences between exploratory and confirmatory studies are compiled in Supplementary Fig. 1.Fig. 1Experimental design (confirmatory studies).**a** Example of a 60-induction sequence, in which a series of 3 inductions presented each of the 3 emotions, following a pseudo-random order that avoided repeating the same emotion on two consecutive inductions. In this confirmatory study, the experiment was divided into 3 sessions separated by 2 breaks. **b** Example screens presented to induce an emotion (using a text vignette accompanied by congruent music) and assess economic preferences (using a battery of 12 choice trials interleaved with fixation crosses). With button presses, participants indicated their preference for the large-but-costly or the small-but-uncostly reward. The position of choice options on screen was counterbalanced across trials. At the end of the trial, music stopped playing and participants positioned a cursor on a visual analogue scale to rate the emotions that they had just experienced. **c** Example choice trial screens for the different types of cost. Participants decided between receiving a large reward in exchange for a cost, or a smaller but costless reward. Costs entailed either a risk of losing money (with probability indicated by the pie chart), a delay until payment (with duration indicated on the calendar), or a physical effort to be exerted on a fitness bike (with power level indicated by the slope). In the exploratory studies, 4 emotions (happiness, sadness, anger, fear) were induced and rated by participants, and choices were presented with written numbers instead of symbolic representations (see Supplementary Fig. 1 for illustration). #### Subjective rating Our main objective here is to validate happiness, sadness and neutral inductions with subjective ratings, which are illustrated in Fig. 2a (exploratory studies) and Fig. 2b (confirmatory studies). We defined as ‘target ratings’ the ratings given to the emotional dimension targeted by the induction. For all inductions in both studies, target ratings were significantly higher than the mean of other ratings, showing that the four emotions were successfully induced. The difference was more clear-cut with happiness induction, as there was some degree of covariance between the ratings of the negative emotions, which is a well-documented phenomenon^28,30,65^. Nevertheless, the difference between the target rating and the mean of all other ratings was highly significant in all happiness (exploratory: Δ = 43.4% [38.3:48.6], *t*(34) = 17.2, *p* = 2.24e–18; Δ = 32.0% [29.0:35.1], *t*(58) = 21.0, *p* = 8.43e–29), sadness (exploratory: Δ = 32.0% [27.9:36.1], *t*(34) = 15.8, *p* = 3.16e–17; Δ = 26.8% [24:29.5], *t*(58) = 19.5, *p* = 3.76e–27), anger (Δ = 31.3% [26.4:36.2], *t*(34) = 12.9, *p* = 1.06e–14), and fear (Δ = 36.0% [31.2:40.8], *t*(34) = 15.2, *p* = 1.01e–16). When regressing the target rating against the rating of the same emotion on the subsequent trial, we found no significant relation (exploratory: b~next rating~ = 0.032 [–0.026:0.091], *t*(34) = 1.1, *p* = 0.269; b~next rating~ = 0.012 [–0.013:0.038], *t*(58) = 0.9, *p* = 0.348). Thus, there was no statistically significant evidence that the induced affective state survived the 8-second washout period.Fig. 2Validation of emotion induction by subjective ratings.**a**, **b** Subjective ratings in the exploratory studies (top, *N* = 35) and confirmatory studies (bottom, *N* = 59). Each graph includes all the ratings made after the induction of a specific emotion. Each box shows the rating of a specific emotional dimension (H = happiness, S = sadness, A = anger, F = fear). Statistical results are only shown for the main comparison (happiness vs. sadness ratings). **c** Mood scores (happiness minus sadness ratings). Boxes show mood scores obtained in the exploratory (exp.) and confirmatory (conf.) studies, for each of the main emotional inductions (happiness, sadness and neutral inductions). Statistical tests compare mood scores to zero. In all graphs, box plots show median, first and third quartiles, minimum and maximum, and outliers. ♣ denotes *p* < 1E–20. Critically, happiness and sadness ratings were significantly different after both happiness and sadness inductions, but not after neutral inductions (Fig. 2a, b). Across inductions, happiness and sadness ratings were anti-correlated (exploratory: R = –0.258 [–0.295:–0.220], *t*(34) = –13.8, *p* = 1.64e–15; R = –0.445 [–0.483:–0.407], *t*(58) = –23.2, *p* = 5.02e–31), suggesting that they could be reduced to a single dimension. We therefore defined a mood score as the standardized difference between the rated happiness and sadness (Fig. 2c), which was calculated for every induction (whether the target was happiness, sadness or neutral). The mood score was significantly positive following happiness induction (exploratory: M = 1.12 [1.09:1.15], *t*(34) = 68.3, *p* = 5.67e–38; M = 1.11 [1.07:1.14], *t*(58) = 68.9, *p* = 2.40e–57), close to zero in the neutral condition (exploratory: M = -0.04 [–0.07:–0.001], *t*(34) = –2.1, *p* = 0.047; M = –0.01 [–0.04:0.02], *t*(58) = –0.5, *p* = 0.640), and significantly negative following sadness induction (exploratory: M = –1.09 [–1.13:–1.06], *t*(34) = –66.9, *p* = 1.10e-37; M = -1.10 [-1.13:-1.07], *t*(58) = –70.7, *p* = 5.53e–58). The impact of emotion inductions was preserved throughout the experiment, although we observed a slight decline of the effect size (by 7.2% after 60 trials, on average), meaning that happiness and sadness ratings both decreased with inductions (exploratory: b = –0.0003 [–0.001:0.001], *t*(34) = –0.5, *p* = 0.603; b = -0.0014 [–0.002:–0.001], *t*(58) = –4.2, *p* = 8.63e–05). This decrease with time on task was orthogonal to our main effect of interest, because inductions of happiness and sadness were regularly alternated along the experiment. In an open-answer debriefing questionnaire presented after completion of the experiment, we asked participants to reflect on the emotions they had experienced. A large majority of participants reported that they had genuinely experienced happiness and sadness, and the intensity of their reported emotions during debriefing was correlated with their average respective ratings during the experiment (see Supplementary Note 6 for a full analysis of these debriefing questionnaires). #### Physiological responses To assess whether induced emotions were truly experienced (and not just reported), we measured (in a subset of participants, see Supplementary Fig. 1e for details) physiological indicators of autonomic arousal (pupil diameter and skin conductance) and electromyographic (EMG) activity of facial muscles that indicate the valence of emotional experience, namely the zygomaticus (active when smiling), and the corrugator (active when frowning). Significant effects of inductions were observed for all emotions (including anger and fear in exploratory studies), at least in the pupil dilation response. For our purposes, we focus here on the differences between happiness, sadness and neutral inductions. Regarding arousal signals (Fig. 3a), we compared both happiness and sadness to the neutral induction. Significant clusters, after correction for multiple comparisons based on random-field theory (see “Methods” for details), were observed in the pupil dilation response during both inductions of happiness (0.8s-10s, 554 samples, *p* = 3.97e–9) and sadness (1.25–10 s, 527 samples, *p* = 1.88e–8) and in the skin conductance response during both inductions of happiness (3.86s-10s, 63 samples, *p* = 7.23e–7) and sadness (8.71s–10s, 14 samples, *p* = 0.034). As a time-locked analysis may not be optimal for measuring phasic skin conductance responses to affective changes, we estimated the amplitudes of time-variable activity peaks in subsequent analyses. Regarding valence signals (Fig. 3b), we compared between happiness and sadness, and found significant clusters in the activity of both the zygomaticus (1.11s-10s, 446 samples, *p* = 1.73e–12) and corrugator (0.82s-10s, 461 samples, *p* = 1.04e–13).Fig. 3Validation of emotion induction by physiological responses.Physiological signals of arousal (**a** pupil diameter and skin conductance) and physiological signals of valence (**b** zygomaticus and corrugator EMG activity) were measured in a subset of participants after happiness and sadness inductions (*N* = 66, left column) and after anger and fear inductions (*N* = 35, right column). Time-resolved baseline-corrected responses are shown over the 10-s duration of the induction, locked to text/music onset. Solid lines and shaded areas represent mean responses and their standard errors. Time clusters with significant differences between inductions, after correction for multiple comparisons using Random Field Theory, are denoted by horizontal lines for both arousal measures (yellow: happiness versus neutral, sadness versus neutral) and valence measures (green: happiness versus sadness). We further verified the validity of the emotion induction procedure by testing the link between physiological responses and subjective ratings. To this end, we collapsed the time-resolved activity into a single number by taking the median activity over the induction window, except with skin conductance data, for which we took response peak estimates using the PsPM toolbox^59^. Across inductions, the target rating (happiness or sadness) was significantly correlated with pupil dilation (R = 0.047 [0.003:0.092], *t*(68) = 2.1, *p* = 0.037) and skin conductance response (R = 0.078 [0.038:0.117], *t*(68) = 3.9, *p* = 1.95e–04). These correlations confirm the relationship between rated intensity and physiological arousal following emotional inductions, discarding potential confounds inherent to the comparison with the neutral induction, such as the mere presence of music. In addition, the mood score was significantly correlated across inductions with facial EMG activity; positively for the zygomaticus (R = 0.231 [0.171:0.290], *t*(64) = 7.7, *p* = 1.05e-10) and negatively for the corrugator (R = –0.248 [–0.299:–0.197], *t*(64) = –9.7, *p* = 3.50e–14). ### Effects on choice behaviour Following each induction, participants were presented with a short series (6–12 trials, depending on the study) of economic choices. Each choice featured two one offered a large, fixed reward (e.g., 30€ in confirmatory studies) in exchange for a cost, while the other option was costless but came with a smaller, variable reward (Fig. 1c). The costs in question a risk of losing money in a lottery, a delay until payment, or a physical effort to be exerted on a fitness bike after the experiment. In exploratory studies, the costs were explicitly written on the screen, while in confirmatory studies, they were indicated in a pictogram representing a wheel-of-fortune, a calendar, or a slope, with a red shaded area proportional to the cost level. Importantly, participants were told that their final endowment would depend on 3 selected trials—one per cost type—to be drawn after the experiment from all decisions that were made during the study. Prior to the experiment, a choice calibration was implemented to get an estimate of each participant’s indifference curves, which outline pairs of options for which choice probability is 50/50% (see “Methods”). These curves were used in exploratory studies to generate trials while controlling for the distance to indifference, with the idea that affective state may have a higher impact on choices for which the participant has no strong preference for one of the two options. #### Model-free results We compared choice rates between emotional and neutral inductions, pooling data over all cost types as done in our previous study (Fig. 4a). Compared to neutral induction, participants chose costly options more frequently after happiness induction (Δ = 1.7% [0.5:2.8], *t*(93) = 2.75, *p* = 0.007) and less frequently after sadness induction (Δ = –1.4% [–2.9:–0.02], *t*(93) = –2.03, *p* = 0.046). As a consequence, the difference between happiness and sadness was significant, in both the exploratory and confirmatory studies (Δ~exp~ = 5.6% [1.8:9.4], *t*(34) = 3.0, *p* = 0.005; Δ~conf~ = 1.6% [0.3:2.9], *t*(58) = 2.51, *p* = 0.015). There was no significant change in costly choice rate across trials, hence no statistically significant evidence that the task was long enough to induce cognitive fatigue, which we previously found to affect economic decisions^49^.Fig. 4Model-free effects of mood induction on choice behaviour.**a** Effects of emotional inductions (relative to neutral induction) on choice rates (% of costly options selected). Bars show means across the different types of cost (risk, delay, effort), in exploratory and confirmatory studies, after the different emotional inductions (H = happiness, S = sadness, A = anger, F = fear). Statistical tests show the difference between happiness and sadness inductions, which are compared across all participants in an adapted ‘Raincloud plot’^86^. Data points represent participants, and grey lines connecting data points have shades proportional to their slope. **b** Effects of mood score (standardized difference between happiness and sadness ratings) on costly choice rate in the exploratory and confirmatory studies. Logistic regressions were done separately per cost type. Box plots show median, first and third quartiles, minimum and maximum, and outliers, for individual participants’ regression estimates (beta weights). **c** Effects of happiness and sadness inductions (relative to neutral induction) on standardized choice response time (RT), plotted separately for choices in which the uncostly / uncostly option was selected. In all **p* < 0.05, ***p* < 0.01, ****p* < 0.001, N~exploratory~ = 35, N~confirmatory~ = 59, N~total~ = 94. To assess whether the effect on choice rate was not just linked to the type of induction but also to the subjective experience, we ran a logistic regression of choice rate against mood score (Fig. 4b), excluding subjects with extreme preferences (<5% or >95%). The weights of mood score on costly choice rate (regression estimates) were positive in all cases, but only significant for risk and effort (b~R~ = 0.07 [0.03:0.12], b~D~ = 0.06 [–0.01:0.14], b~E~ = 0.09 [0.04:0.13]; *p*~*R*~ = 0.003, *p*~*D*~ = 0.109, *p*~*E*~ = 4E–4). However, there was no statistically significant evidence for a true difference in regression estimates between cost types, as a one-way ANOVA showed no significant effect (*F*(2,227) = 0.14, *p* = 0.866). There was no significant difference either between regression estimates obtained in the exploratory and confirmatory studies (b~exp~ = 0.10 [0.03:0.16], b~conf~ = 0.04 [0.01:0.08]; Δ = 0.05 [–0.01:0.12], *t*(91) = 1.65, *p* = 0.102). When pooling choices across studies and cost types, the global weight on mood score was highly significant (*b*~*all* costs~ = 0.06 [0.03:0.10], *t*(92) = 4.10, *p* = 9E–5). Thus, the more they experienced happiness, and the less they experienced sadness, the more inclined participants were to choose largely rewarded options that were risky, delayed, or effortful. Along with this effect, we observed (Fig. 4c) that choice reaction time (RT) varied with both mood score (rated happiness minus sadness) and the eventually selected option (costly versus uncostly). In a multivariate regression against RT, we found a main effect of mood (b = –0.04 [–0.06:-0.01], *t*(92) = –3.18, *p* = 0.002), no effect of chosen option (b = 0.05 [–0.03:0.12], *t*(92) = 1.26, *p* = 0.211), but a significant interaction (b = 0.03 [3E–4:0.06], *t*(92) = 2.01, *p* = 0.048). Indeed, the mood effect was greater and only significant when the chosen option was the costly one (Δ = –0.09 [–0.14:–0.04], *t*(93) = –3.45, *p* = 8E–4). Thus, participants were globally faster to make decisions as they felt happier, and this effect was particularly salient for decisions in which the costly option was ultimately selected. Contrary to happiness and sadness inductions, anger and fear inductions did not significantly affect economic decisions in the exploratory study (Fig. 4a). Indeed, relative to the neutral induction, choice rates following anger and fear inductions were not significantly different for any of the cost types (all –2.1% < Δ~anger~, Δ~fear~ < 5.8%). Across cost types, the average difference with respect to neutral was also close to zero (Δ~anger~ = 0.6% [–1.6:3.0], Δ~fear~ = 0.7% [–2.3:3.9]; *t*~*anger*~(34) = 0.59, *t*~*fear*~(34) = 0.48; *p*~*anger*~ = 0.560, *p*~*fear*~ = 0.633). In a logistic regression of standardized anger and fear ratings against choices, there was no significant effect for any of the cost types (all |*b*~*anger*~ | < 0.05, |*b*~*fear*~ | < 0.08). Across cost types, the regression estimates were again close to zero (*b*~*anger*~ = –0.02 [–0.09:0.05], *b*~*fear*~ = –0.01 [–0.08:0.06]; *t*~*anger*~(34) = –0.7, *t*~*fear*~(34) = –0.2; *p*~*anger*~ = 0.493, *p*~*fear*~ = 0.831). #### Model-based results We used computational modelling to better capture the role of mood in decision-making and that of the other factors (costs and rewards). We first identified the model that best explained choices independently of mood inductions, and then complemented the winning model by integrating trial-wise mood scores. The equations listed below correspond to the winning model of a Bayesian comparison conducted in another study employing the same choice task^10^ and replicated with the present dataset (see “Methods”). Choices were modelled with a standard softmax decision function (Eq. (7)) that expresses the selection probability of each option as a sigmoidal transformation of decision value *DV* (difference between option values), weighted by an inverse choice temperature *β*~*1*~ and added to a choice bias *β*~*0*~. In the equation below, we use a single generic notation for simplicity, even if the choice bias and choice temperature parameters (*β*~*0*~ and *β*~*1*~) are actually specific to each cost type (see “Methods” for cost-specific notations).7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{P}}\left({{costly}}\right)=\frac{1}{1+ exp (-{{{\beta }}}_{0}-{{{\beta }}}_{1}\cdot {{DV}})} $$\end{document}Pcostly=11+exp(−β0−β1⋅DV) In this decision function, the decision value *DV* is defined as the value difference between the costly and uncostly options, i.e. *V(costly) – V(uncostly)*. The weight and bias parameters (*β*~*1*~ and *β*~*0*~) respectively adjust the slope and the intercept of the sigmoid curve. The slope *β*~*1*~ captures the consistency of choices (how strictly they follow the decision value), by opposition to their stochasticity (sometimes called temperature). At the extremes, a step function (*β*~*1*~ = ∞) would be totally deterministic and a flat function (*β*~*1*~ = 0) totally random. The intercept *β*~*0*~ captures the choice bias, i.e. the inclination to choose one option or the other when the decision value is null (*DV* = *0*). Here, higher *β*~*0*~ means greater propensity to choose the costly option. The subjective option values, *V(costly)* and *V(uncostly)*, are modelled using three different discount functions, depending on the type of cost (risk, delay or effort). Discount functions integrate objective rewards with the costs weighted by a discount factor *k*~*R*~, *k*~*D*~, or *k*~*E*~ (Eqs. (8–10)). The curvatures of discount functions are controlled by power parameters *γ*~*R*~, *γ*~*D*~, and *γ*~*E*~.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {V}_{R}={Rew}\cdot P-{k}_{R}\cdot L\cdot {\left(1-P\right)}^{{\gamma }_{R}} $$\end{document}VR=Rew⋅P−kR⋅L⋅1−PγRwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ P $$\end{document}P is the probability of winning the reward *Rew* and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ L $$\end{document}L the amount that can be lost by taking the risky option.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {V}_{D}=Rew\cdot exp (-{k}_{D}\cdot {D}^{{\gamma }_{D}}) $$\end{document}VD=Rew⋅exp(−kD⋅DγD)where *D* is the delay of reward delivery.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {V}_{E}={Rew}-{k}_{PE}\cdot {E}^{{\gamma }_{E}} $$\end{document}VE=Rew−kPE⋅EγEwhere *E* is the effort level. Note that the same discount function is used to compute the value of costly options (for which the reward *Rew* is 30€ and the cost higher than 0), and the value of uncostly options (for which *Rew* varies and cost is 0, such that *V = Rew*). To check that the set of discount and decision functions described above were best capturing the new choice dataset, we reconducted a Bayesian model comparison. The full space contained 36 different models with systematically varying parametrizations, described in detail in “Methods” and in Supplementary Table 2. We used the Variational Bayesian Analysis toolbox^50^ to invert the models and compare their log evidence, which offers a tradeoff between accuracy (goodness of fit) and complexity (degrees of freedom). The winning model was submitted to a recovery analysis, which showed good identifiability of all parameters (see “Methods”). The fitted discount functions were quasilinear for risk, convex for delay and concave for effort (see Fig. 5a). Thus, subjective option values were always positive for delayed rewards, reflecting the fact that they are always better than no reward at all. In contrast, subjective option values of risky and effortful rewards could be negative, accounting for the possibility that participants would rather give up money than take too great risks or make too demanding efforts.Fig. 5Model-based effects of mood induction on choice behaviour.**a** Discount functions for each of the cost types, obtained from fitting the winning model to the choice data. Each curve describes how the subjective value of the objective big reward (30€) diminishes with increasing costs (risk, delay or effort). Thin grey lines represent individual curves, the thick black line is the mean curve across all participants. **b** Observed psychometric curves. Dots represent costly choice rates following happiness (yellow) and sadness (blue) inductions, as a function of binned decision value (difference between costly and uncostly options), calculated using the discount functions of the winning model. The neutral condition has been removed because it would not be visible between the two curves. **c** Correlation between mood scores and choice residuals. The green line and shaded area display the mean regression fit and its standard error across participants. Choice residuals were obtained by subtracting the probability predicted by the model from the observed choice rate and then binned according to sorted mood score. Colours of the dots illustrate mood score (same as x-axis) along a dimension going from sadness to happiness. Both in panels (**b**) and (**c**), error bars represent standard errors of the mean. **d** Modelled psychometric curves. Modelled choice rates were obtained here with a softmax decision function that included a bias scaled to mood score. This mood-related choice bias captures the change in the intercept of the psychometric functions between happiness (yellow) and sadness (blue) inductions. Solid lines represent the mean and shaded areas the standard error of the mean across participants. **e** Distribution of mood-related bias parameters obtained in the exploratory (Exp.) and confirmatory (Conf.) studies. Dots are individual participants. Horizontal lines show the mean across all participants (from both studies). Stars denote distribution density of fitted parameters that differ from the null (zero mood bias). ***p* < 0.01, ****p* < 0.001, N~exploratory~ = 35, N~confirmatory~ = 59, N~total~ = 94. Psychometric curves (Fig. 5b) showed that costly choice rates were shifted upwards following happiness induction and downwards after sadness induction, relative to neutral induction. This confirms our theory that mood exerts an additive bias favouring the costly options, on top of the attributes (rewards and costs) that are integrated in the computation of option values. Indeed, mood score significantly explained choice residuals (Fig. 5c), which represent the variance across trials that was unaccounted for by the model (R = 0.041 [0.024:0.058], *t*(93) = 4.8, *p* = 5.06e–06). We therefore updated our model to include mood as a modulator of the choice bias in the softmax function, with parameter *β*~*M*~ shared across cost types.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{P}}\left({{costly}}\right)=\frac{1}{1+ exp (-({{{\beta }}}_{0}+{{{\beta }}}_{{{M}}} \cdot {{Mood}})-{{{\beta }}}_{1}\cdot {{DV}})} $$\end{document}Pcostly=11+exp(−(β0+βM⋅Mood)−β1⋅DV) In this augmented softmax function, higher *β*~*M*~ means that positive and negative mood have a greater impact on the choice bias (i.e., irrespective of option values, happiness inclines choices towards costly options and sadness towards uncostly options). We compared this model to an alternative model that accounts for mood effects by modulating the weight on reward in discount functions. In a Bayesian model comparison, the additive variant in Eq. (11) was selected as the most plausible, with a 90.4% model frequency and 100% exceedance probability. The updated choice model with the additive mood-related bias could indeed account for the offset of psychometric curves between happiness and sadness inductions (Fig. 5d). The fitted mood-related bias parameter (Fig. 5e) was significantly positive in both exploratory studies (*β*~*M, exp*~ = 0.203 [0.080:0.327], *t*(34) = 3.3, *p* = 0.002) and confirmatory studies (*β*~*M, conf*~ = 0.111 [0.039:0.183], *t*(58) = 3.1, *p* = 0.003), and highly significant across studies (*β*~*M*~ = 0.145 [0.082:0.209], *t*(93) = 4.5, *p* = 1.76e–05). Thus, the mood-related choice bias is a key computational parameter, accounting for how much cost-benefit tradeoffs in a given individual are affected by experienced happiness and sadness. To assess the efficiency of our design, we conducted sensitivity and recovery analyses for this mood-related bias parameter β~M~ (see Supplementary Fig. 2 and Supplementary Note 5 for details). Given the mean and variance of β~M~, power calculation showed that a sample of *N* = 38 participants was sufficient to detect a significant mood-related bias at *p* < 0.05 with 80% power. The sensitivity analysis showed that, for the current sample size of *N* = 94, the minimal detectable effect with *p* < 0.05 would be β~M~ = 0.091 for 80% power, or β~M~ = 0.105 for a 90% power. This is well below the effect that was actually observed (β~M~ = 0.145), which corresponds to a medium effect size (Cohen’s d = 0.48). Recovery analysis showed that the β~M~ fitted on simulated data (with parameters sampled from observed distributions) purely captured the variance in the true mood-related bias (with a regression weight of b = 0.87, the weights of all other computational parameters being below 0.01). Critically, when repeating this analysis with increasing sample size, the proportion of variance in the true β~M~ explained by the fitted β~M~ reached a plateau between 30 and 45 participants, suggesting that including more participants would not improve the assessment of the mood-related choice bias in the population. To examine whether participants were aware of this bias exerted by mood on their choices, we examined responses made to the debriefing questionnaire. Interestingly, the mood-related bias parameter was no different between participants who guessed versus not guessed the purpose of the study, who reported or not genuine experience of happiness and sadness, or who thought or not that their choices had been influenced by their emotions (see Supplementary Note 6 for a full analysis of this questionnaire). #### Link between choices and physiology In the preceding analyses, choices are shown to be influenced by mood score, which could be susceptible to demand effects as is the case for any self-report assessment. We thus investigated whether choices could be directly related to physiological measures of affective states taken during the just preceding induction. Because happiness and sadness ratings depend on both the intensity and valence of experienced emotion, we used both markers of autonomic arousal (pupil diameter, skin conductance) and of facial expression (zygomaticus and corrugator EMG activity). These measures were correlated with each other across inductions (Fig. 6a), the correlation being positive for pupil dilation and skin conductance response (R = 0.110 [0.070:0.149], *t*(68) = 5.5, *p* = 5.12e–07), and negative for zygomaticus and corrugator activity (R = –0.119 [–0.196, –0.042], *t*(64) = –3.1, *p* = 0.003). However, none of these signals taken in isolation was significantly correlated with the costly choice rate following the induction. We thus aimed to construct a physiological proxy for mood that would combine the different measures (Fig. 6b). In principle, arousal reflects the intensity of an emotional response, whereas valence indicates the sign (positive or negative). We therefore multiplied the physiological measure of arousal (*Aφ* = mean of pupil diameter and skin conductance) and that of valence (*Vφ* = sigmoid function of the difference between zygomaticus and corrugator EMG activity, orthogonalised with respect to arousal). Because the effect of mood was found to decline across the series of choices, we also integrated post-induction trial number *N* to account for temporal dissipation of the affective state. Thus, our physiological proxy for mood was *Mφ* = *Vφ * Aφ / N*. It was significantly correlated with mood score (R = 0.233 [0.184:0.281], *t*(64) = 9.5, *p* = 6.79e–14). Note that this correlation between physiological and reported mood was observed despite the fact that we did not fit any behavioural measure (there is no free parameter in the mood proxy) and did not inform the analysis about which emotion was induced.Fig. 6Prediction of choice behaviour from physiological mood proxy.**a** Correlation matrix of physiological and behavioural measures. Pupil dilation and skin conductance response are taken during emotional induction, relative to baseline. Mood score is the standardized difference between happiness and sadness ratings. Choice rate is the % of costly options selected. For all correlations, cells mention mean Pearson coefficients and p-values obtained from two-tailed t-tests. Significant cells (with *p* values surviving corrections for multiple comparisons) are highlighted in colours, others are in grey. **b** Two mood-related dimensions extracted from physiological recordings. Top: an arousal dimension *Aφ*, obtained by averaging pupil dilation and skin conductance response. Bottom: a valence dimension *Vφ*, calculated as a sigmoid function of the difference between the zygomaticus and corrugator EMG responses. Measures on both dimensions are plotted against mood score (dots are means in each bin, error bars are standard errors of the mean across participants). **c** Correlation between mood proxy and choice residuals. The physiological mood proxy *Mφ* is the product of physiological arousal and valence, divided by the post-induction choice number. Choice residuals are costly choice rates after subtraction of choice probability predicted by the model that does not incorporate the mood-related bias. The green solid line shows the mean regression across participants, the shaded area is the 95% confidence interval. Dots are means in each bin, error bars are standard errors of the mean across participants. The colour of the dots represents the mean mood score per bin. **d** Correlation between the effects of mood proxy and mood score (*Mφ* and *M*) on choices. Dots represent individual estimates of weights *b*~*Mφ*~ and *b*~*M*~ in the regression explaining choice residuals. The straight line shows the regression of *b*~*Mφ*~ against *b*~*M*~, surrounded by the 95% confidence interval. ***p* < 0.01, *N* = 65. Next, we investigated whether the physiological proxy for mood could be related to the choices that participants made (Fig. 6c). We found, in a linear regression of residuals from the winning choice model (without the mood-related bias parameter), that the mood proxy *Mφ* was a significant predictor (*b*~*M*~*φ* = 0.057[0.020:0.094], *t*(63) = 3.1, *p* = 0.003). This may not be surprising, since the mood proxy regressor Mφ was significantly correlated with the mood score regressor M, which was itself a predictor of choice residuals (cf. Fig. 5c). Thus, the self-reported and physiological mood proxies *M* and *Mφ* explained some common variance in decisions (Fig. 6d), and indeed, their influence on choice residuals (when tested in separate regressions) was correlated across participants (R = 0.363, *p* = 0.003). #### Link between choices and attention Our theory is that mood serves as a prior preference, specifying a default option to be taken in the absence of specific information about option attributes^9^. Thus, mood should affect the visual exploration of choice options in the early phase of decision making. More precisely, a predisposition to selecting either choice option should result in attention being preferentially directed towards that option immediately following its presentation. To test this prediction, we used eyetracking to record the gaze position on screen in confirmatory studies, where cost levels are indicated by a shaded red area within pictograms. As a consequence, the (counterbalanced) location of the costly option was immediately apparent in the visual field. The distribution of gaze between the two options was quantified in a time-resolved manner (see “Methods”). During the first second following option display, more attention was paid to the costly option overall, and even more so after happiness induction relative to sadness induction (Fig. 7a, b). This pattern was suggestive of a modulation of attention by mood, which was indeed significant in a direct regression of the time spent looking at the costly option (within the first second) against mood score (*b*~*Mood*~ = 0.102 [0.021:0.184], *t*(58) = 2.5, *p* = 0.015). In turn, the proportion of time spent during the first second was associated with a higher choice rate (*b*~*gaze*~ = 0.503 [0.216:0.790], *t*(57) = 3.5, *p* = 9e-4) and a shorter RT (*b*~*gaze*~ = –0.323 [-0.480:-0.166], *t*(58) = –4.1, *p* = 1e–4) when the costly option was eventually selected. Thus, an early attentional bias in favour of the costly option made the selection of this option both more likely and more rapid (Figs. 7c, d).Fig. 7Effects of mood induction on visual attention.**a** Difference in gaze distribution between happiness and sadness inductions during the first second following option display (shaded window in the time course below). In the heatmap, screen pixels receiving more gaze after sadness and happiness inductions appear in blue and yellow, respectively. Note that half the trials have been flipped such that the costly option was always on the right in this analysis, although it was not the case in reality (because the side of costly option display was counterbalanced across trials). **b** Time course of visual attention paid to the two options, within the 0–3 s window following option display. Gaze rate is obtained at each time point by dividing the number of samples inside each option frame by the number of trials. It declines with time as decisions are already made in a growing number of trials (median RT is shown by the dotted vertical line). The two curves do not add up to 100% because gaze can be located outside the two option frames. The solid line and shaded area represent the mean and its standard error across participants. **c** Costly option choice rate plotted as a function of costly option gaze proportion (during the first second following option display). **d** Choice RT plotted as a function of costly option gaze proportion (during the first second following option display), specifically for trials in which the costly option was selected. Both in panels (**c**) and (**d**), the solid lines represent linear regression fits and shaded areas the 95% confidence interval across trials and participants. In all panels, *N* = 59. ## Discussion In this experiment, we have induced two emotions (happiness and sadness), using written imagery scenarios accompanied by congruent music, and tested their effects on a battery of economic choices involving tradeoffs between monetary rewards and different types of costs (risk, delay, effort). We found that induced happiness and sadness influenced decision-making in a symmetrical manner that was consistent across cost happiness biased decisions towards large-but-costly rewards and sadness towards uncostly-but-small rewards. These effects on choice bias are in line with the predictions of our theory on the functional role of mood fluctuations between happiness and sadness^9^, which predispose to foraging action when big rewards are available at low costs in the environment. Importantly, the choice bias could be predicted not only from a mood score based on self-reports (rated happiness minus rated sadness), but also from a physiological proxy built on measures of arousal (skin conductance and pupil size) and valence (zygomaticus versus corrugator contraction), both mood constructs being completely uninformed by experimental conditions (emotions targeted by inductions). These findings contribute to a growing body of literature that intends to specify how incidental emotions may bias economic decisions^35–37^. However, this literature has not yet converged on firm conclusions, due to mixed evidence for most of the effects tested^38^. We intended to provide more solid methodology by recruiting a large sample (N = 94) across variants of the design, by inducing the two emotions in a same participant using the same procedure, by validating their effects at both the experiential and physiological levels, and by systematically testing their impact on different kinds of economic choices within a same study. As a result, the effects of happiness and sadness states were highly significant in direct comparisons between inductions as well as in regressions against subjective ratings. In contrast, we observed no significant effect of control emotions (anger and fear) on choices, although these emotions were successfully induced, as demonstrated by both self-reports and physiological recordings. This may seem surprising, given the amount of anecdotal evidence that whatever the trigger, we tend to be aggressive towards many people when we feel anger, and afraid of many things when we feel fear. One interpretation could be that anger and fear are emotions that operate in specific (often social) contexts^66^ and not when facing a computer in a secure environment as is a research laboratory. A related interpretation would be that anger and fear do not diffuse into decisions in general, but just prime some specific actions (like fight or flight) that could not be expressed in our economic choice paradigm. Yet, null results should be interpreted with caution, and further studies may help specify in what context and on which decisions these and other negative emotions could have an impact. In any case, our results suggest that happiness and sadness may not operate on the same processes as other affect listed as basic emotions^26,27^, since they have a more pervasive effect on different types of decisions. Indeed, they can be considered as the two poles of a single dimension akin to mood, which is known to bias choices in both real-life behaviours^3–5^ and lab experiments^6,7,10^. It is striking that the difference between episodes of positive and negative mood in these studies yielded the same choice bias as observed in the present study with the difference between happiness and sadness inductions. Besides the difference in induction methods (imagination of personally-relevant situations here versus feedback on personal performance in previous studies), there is a difference in time scale. In our previous studies, mood fluctuated at the time scale of ~15 minutes, through the integration of recurrent feedback. Here, experience of happiness and sadness lasted about ~1 minute in total, since ratings were related to the last induction only, and tended to fade away before the next induction. We note that mood fluctuations as they spontaneously occur in real life are usually observed at longer time scales, from hours to weeks^67,68^. A way to bring these results together is to consider that positive/negative mood represents some temporal integration of momentary happiness/sadness^9,69^, with the same effect on decision-making (a bias toward seeking larger reward, even if costly). Because it is incidental, the effect of mood may be seen from an economic perspective as a case of irrational choice behaviour^70–72^. However, we have shown using model simulations^9^ that in certain environments, where sources of reward and cost are correlated between them and across time, the effect of mood fluctuations may actually be adaptive. This is because they help with engaging foraging actions during periods when rewards are abundant and costs relatively moderate (like during summer for moderate climates). Critically, the decision function that we used in our simulations was the same as in the winning model here, with mood being additive to decision value. It suggests that mood serves as prescribing a global default in the absence of further information (about values specific to the considered options), an agent is more likely to forage when in a good mood (facing costs to get more rewards) and to withdraw when in bad mood (i.e., being content with little reward and avoiding costs). The direct effect of mood on decision-making distinguishes our model from others that suggested a similar adaptive role for mood fluctuations (generalizing across reward sources) but through a different the modulation of learning rates^73^. The additive effect of mood was also reflected in the interaction (between emotional states and final decisions) observed on choice response time. Relative to uncostly actions, the decision to take the costly actions was faster when in line with the default option (in states of happiness) than when against the default option (in states of sadness). Moreover, the notion of mood as a prior preference for large-but-costly rewards was supported by the effects of emotional states on looking patterns. There is ample evidence that people look first at the option they choose in the end^74,75^, although the direction of causality is not clear. We found evidence indeed that, during the early phase of decision making, participants spent more time looking at the costly (versus uncostly) option in states of happiness (versus sadness). In turn, the attention paid to the costly option was associated with a higher chance that this option would be chosen and a shorter RT when this option was chosen, as would be predicted by the attentional drift-diffusion model (aDDM^74^). Yet, the pattern of results observed here was not entirely compatible with the aDDM, because the uncostly option was not chosen faster after sadness inductions, even if it was preferentially looked at (relative to happiness inductions). Nevertheless, the early effect on gaze direction denotes a prior intention to take the action favoured by the affective state, before even seeing the specifics of choice options. This interpretation would be in line with the functional role ascribed to emotions in driving attention so that a priori valuable stimuli are preferentially processed^76–78^. It is also compatible with the idea that decision-making is cognitively demanding and hence that it would benefit from shortcuts provided by emotions which may pre-specify what the best option is likely to be, given the context^79,80^. ### Limitations One potential limitation of our study is that the within-subject design makes the induction participants easily understand that they are supposed to alternate between happiness, neutral and sadness states, depending on the vignette. This could make the manipulation vulnerable to demand effects, i.e., participants could simply comply with the experimenter’s expectations and report being happy/sad after obvious happiness/sadness inductions. Yet this would not explain variations in the intensity of the emotional state within happiness and sadness categories. Also, although it is easy to guess what type of rating is expected, it is not so easy to guess what effect on choices might be expected (even experts disagree, depending on which papers they have read). One could argue that taking less risk after a fear induction is an obvious expected behaviour, but precisely, participants did not conform to that expectation. Moreover, participants who retrospectively reported in a debriefing questionnaire that they felt very emotional, or understood the purpose of the study, or noticed the effects of inductions on their choices, did not exhibit a greater choice bias than those who declared not having a clue. It must be noted, however, that retrospective insight is notoriously limited, so debriefing self-reports alone may not be taken as conclusive evidence for discarding demand effects. A more definitive argument comes from physiological while the facial muscles may be contracted on demand, one cannot directly control pupil size or skin conductance. So, the fact that the choice bias could be predicted from physiological responses speaks against a demand effect and for a true emotional involvement. Also, they eliminate the issue of a possible reverse because ratings were made after choices, one could imagine that ratings were impacted by choices, and not the opposite. This reverse interpretation is not possible for the effects of inductions, and the related physiological responses, because the induced mood states were reconstructed from physiological measures recorded before choices were displayed to participants. Interestingly, facial muscles mostly predicted the sign (i.e., valence) and not the intensity of perceived emotion. To estimate the intensity, we needed autonomic arousal measures (pupil size and skin conductance), which by definition are beyond voluntary control. As suggested before^15,16^, physiological measures may be used as low-level readouts of emotional states in affect-induction paradigms where explicit ratings are omitted, which would avoid potential interference with emotion perception due to the overshadowing effect^81–83^. The next step would be to directly record signals from emotionally-relevant brain areas such as the amygdala, insula, and hippocampus. Although impressive progress has been made in the decoding of naturalistic mood fluctuations from neural recordings^84,85^, the utility of such measures for decoding induced emotions remains to be proven. Ultimately, neural investigations are needed to provide a mechanistic account of how affective states of the brain can bias unrelated choice behaviours. ## Supplementary information Transparent Peer Review file Supplementary Material Reporting summary