Many people believe they would be happier if only they had more money. And events such as winning the lottery or receiving a large pay rise do make people happy, at least temporarily. But recent studies suggest that the main factor driving happiness on such occasions is not the size of the reward received. Instead, it is how well that reward matches up with expectations. Receiving a 10% pay rise when you were expecting 1% will make you feel happier than receiving 10% when you had been expecting 20%.

This difference between an expected and an actual reward is referred to as a reward prediction error. Reward prediction errors have a key role in learning. They motivate people to repeat behaviours that led to unexpectedly large rewards. But they also enable people to update their beliefs about the world, which is rewarding in itself. Could it be that reward prediction errors are associated with happiness mainly because they help us understand the world a little better than before?

To test this idea, Blain and Rutledge designed a task in which the likelihood of receiving a reward was unrelated to the size of the reward. This study design makes it possible to separate out the contributions of learning versus reward to moment-by-moment happiness.

In the task, volunteers had to decide which of two cars would win a race. In the ‘stable’ condition, one of the cars always had an 80% chance of winning. In the ‘volatile’ condition, one car had an 80% chance of winning for the first 20 trials. The other car then had an 80% chance of winning for the next 20 trials. The volunteers were not told these probabilities in advance, but had to work them out by playing the game. However, on every trial, the volunteers were shown the reward they would receive if they chose either of the cars and that car went on to win. The size of the rewards varied at random and was unrelated to the likelihood of a car winning.

Every few trials, the volunteers were asked to indicate their current level of happiness on a scale. The results showed that volunteers were happier after winning than after losing. On average they were also happier in the stable condition than in the volatile condition. This was especially true for volunteers with pre-existing symptoms of depression. Moreover, happiness after wins did not depend on how large the reward they got was, but instead simply on how surprised they were to win.

These results suggest that how we learn about the world around us can be more important for how we feel than rewards we receive directly. Measuring happiness in various types of environment could help us understand factors affecting mental health. The current results suggest, for example, that uncertain environments may be especially unpleasant for people with depression. Further research is needed to understand why this might be the case. In the real world, rewards are often uncertain and infrequent, but learning may nevertheless have the potential to boost happiness.

Decisions are guided by beliefs about states of the world. Some states are directly observable, like the potential prize for a bet. Other states, like the probability of winning, may not be directly observable but can be inferred from past events. Thus, learning from experience is essential for adaptive behaviour. In the standard theoretical framework, learning is driven by how unexpected the outcome is (i.e. by the prediction error): the difference between outcome and prediction (Barto, 1995). Sensitivity to the prediction error (i.e. the learning rate) flexibly adapts to environmental statistics. The decisions of both humans and non-human primates are consistent with a higher learning rate in more volatile environments, and subjects are more likely to stay on the same option after positive compared to negative feedback when reward probabilities change more frequently (Behrens et al., 2007; Browning et al., 2015; Donahue and Lee, 2015; Massi et al., 2018; Mathys et al., 2011).

Emotions are widely believed to play a role in adaptive behaviour (Fredrickson, 2004), but no computational framework exists to link them. Unexpected outcomes influence affective states, so that bad outcomes feel worse when unexpected than when expected, and good outcomes feel better when unexpected than when expected (Mellers et al., 1997; Shepperd and Mcnulty, 2002). It has recently been shown that reward expectations and reward prediction errors (RPEs), the difference between experienced and predicted rewards, can explain changes in affective state in the context of decision-making under uncertainty when learning is not required (Rutledge et al., 2014; Rutledge et al., 2015). A number of studies have found results consistent with the idea that happiness is modulated by past prediction errors (Otto et al., 2016) including a recent report showing in students that prediction errors due to exam performance influence real-world emotions (Villano et al., 2020). Mood has been proposed to represent environmental momentum, whether an environment is getting better or worse, which could be a useful variable for adaptive behaviour (Eldar et al., 2016; Eldar and Niv, 2015).

Impairments in reward and emotion processing are associated with depression (Shepperd and Mcnulty, 2002). Learning in depression has been extensively studied, but there is not consistent evidence for a specific deficit (Blanco et al., 2013; Cella et al., 2010; Chase et al., 2010; Gillan et al., 2016; Herzallah et al., 2013; Kunisato et al., 2012; Mueller et al., 2015; Pechtel et al., 2013; Robinson et al., 2012; Taylor Tavares et al., 2008; Thoma et al., 2015; Vrieze et al., 2013), as reviewed in Huys et al., 2013; Scholl and Klein-Flügge, 2018. The ability of individuals to appropriately adjust learning rates to environmental volatility is associated with anxiety symptoms (Browning et al., 2015). Individuals with high trait anxiety show reduced ability to adjust updating of outcome expectancies for aversive outcomes to the volatility of the environment compared to individuals with low trait anxiety. Failure to appropriately adjust learning to environmental volatility has not been established in depression. Affective states reflect subjective estimates of uncertainty, which predict the dynamics of subjective and physiological stress responses (de Berker et al., 2016). Overall mood during risky decision-making tasks is reduced with increasing depression severity, both in the laboratory and using remote smartphone-based data collection (Rutledge et al., 2017).

Here, our goal was to quantify the relationship between mood and adaptive behaviour in two common reinforcement learning tasks (Figure 1): one in which reward probabilities do not change (stable) and one in which reward probabilities periodically change (volatile). We addressed the following questions: (1) Is mood more sensitive to learning-relevant or learning-irrelevant variables in established reinforcement learning models? (2) Do mood dynamics adjust to environmental volatility? (3) Are mood dynamics affected by depression in the context of learning?

Figure 1

Download asset Open asset

To investigate the relationship between mood dynamics and adaptive behaviour, we adapted a task design employed in previous studies (Behrens et al., 2007; Massi et al., 2018). Participants repeatedly made decisions between two cars racing against each other (corresponding to a one-armed bandit because only one car can win each race). The reward magnitude associated with each car was explicit and assigned randomly on each trial, whereas the outcome probability was implicit and had to be inferred from the outcomes of previous trials. Participants (n = 75) completed two learning tasks with a break in between tasks (counterbalanced across participants). In the stable environment, the ‘best’ car won 80% of the races on average. In the volatile environment, the ‘best’ car had an 80% probability of winning and the ‘best’ car switched every 20 trials (see Figure 1). Participants were asked to report their current happiness after every three to four trials. Unlike in previous studies in humans using a similar task design (Behrens et al., 2007; Browning et al., 2015), participants were informed that they would complete stable and volatile tasks. Which environment they were in was indicated at the start of each task. Subjects were given no guidance as to how they should use this information, although we expected this manipulation to increase the difference in behavioural sensitivity between environments relative to previous studies.

Learning rates change with environmental volatility

Participants chose the option with the higher expected value (i.e. choice accuracy) more often than chance (see Figure 2A) in the stable environment (82.1 ± 1.1% (mean ± SEM), z = 7.5, p < 10⁻¹³) and in the volatile environment (73.3 ± 1.0%, z = 7.5, p < 10⁻¹³). Participants chose the higher probability option more often than chance in the stable environment (80.8 ± 1.0%, mean ± SEM, z = 7.5, p < 10⁻¹³) and in the volatile environment (62.6 ± 1.1%, z = 7.2, p < 10⁻¹²). To ensure that participants incorporated information about the magnitudes of potential rewards into their decisions, we considered only trials where the car with the lower outcome probability had the higher expected value. Subjects chose the low probability car in these trials more often than chance (stable: 61.2 ± 3.0%, z = 3.5, p < 0.001; volatile: 75.6 ± 2.7%, z = 6.4, p < 10⁻⁹). These results are consistent with participants integrating both probability and reward to make their decisions.

Figure 2

Download asset Open asset

Multiplicative and additive models that integrate probability and reward in different ways have been widely used to explain behaviour across stable and volatile environments (Behrens et al., 2007; Browning et al., 2015; Donahue and Lee, 2015; Massi et al., 2018). Both types of models include the same learning component for updating the probability estimate for each car to win (Equations 1 and 2), based on the probability prediction error (PPE), the difference between the outcome (0 or 1) and the estimated probability of winning:

(1) ${\displaystyle P_{car1wins}\left(t+1\right)=P_{car1wins}\left(t\right)+\alpha PPE\left(t\right),}$

(2) ${\displaystyle PPE\left(t\right)=Outcome\left(t\right)-P_{car1wins}\left(\mathrm{t}\right),}$

where P_{car1 wins}(t) is the estimated probability for car 1 winning on trial t, and α is the learning rate. For choices to car 2, a similar equation applies. Multiplicative and additive models differ regarding implementation of choice predictions with probability and reward magnitude either integrated multiplicatively (Behrens et al., 2007; Browning et al., 2015) or additively (Donahue and Lee, 2015; Farashahi et al., 2017; Massi et al., 2018).

The multiplicative selector resembles the maximisation of expected utility common to economic decision models (Kahneman and Tversky, 1979):

(3) ${\displaystyle E{U}_{car1}(t)=\mathrm{m}\mathrm{a}\mathrm{x}\left[\mathrm{m}\mathrm{i}\mathrm{n}\left[\eta ((P_{car1wins}(t)-0.5)+0.5),1\right],0\right]\times R_{car1wins}(t)}$

(4) ${\displaystyle P_{car1chosen}\left(t\right)=\frac{1}{1+\exp \left(-\beta \left(E{U}_{car1}\left(t\right)-E{U}_{car2}\left(t\right)\right)\right)},}$

where η is a free parameter related to the level of risk aversion and β is the inverse temperature (i.e. choice stochasticity or precision), and $R_{car1wins}\left(t\right)$ is the reward magnitude if car 1 is chosen and wins. The multiplicative model captured choice data in both stable (pseudo-r² = 0.54 ± 0.02, mean ± SEM) and volatile environments (pseudo-r² = 0.41 ± 0.02).

In contrast, the additive selector is implemented as follows:

(5) ${\displaystyle P_{car1chosen}\left(t\right)=\frac{1}{1+\exp \left(-\beta \left(\varphi \mathrm{\Delta }Probability\left(t\right)+\left(1-\varphi \right)\mathrm{\Delta }Reward\left(t\right)\right)\right)}}$

where ∆Probability corresponds to the difference in estimated probability between the options and ∆Reward corresponds to the difference in normalised reward magnitude between the options, and ϕ represents the relative weight of probability and magnitude on choice. The additive model captured choice data in the stable (pseudo-r² = 0.62 ± 0.02) and volatile environments (pseudo-r² = 0.45 ± 0.02, see Figure 2A).

Model comparison demonstrated that the additive model better explained choice data with the same number of parameters as the multiplicative model in both stable (∆BIC = 690) and volatile environment (∆BIC = 321, see Table 1). Our results are consistent with similar findings obtained in highly trained non-human primates using a task design in which changes between stable and volatile environments were signalled (Massi et al., 2018) as in the present study.

Table 1

Model	Number of parameters	Stable pseudo-r2	Volatile pseudo-r2	Stable BIC	Volatile BIC	Stable ΔBIC	Volatile ΔBIC
Additive	3	0.62	0.45	4134	5605	0	0
Multiplicative	3	0.54	0.41	4824	5926	690	321
Probability only	2	0.35	0.16	6104	7635	1970	2030
Magnitude only	1	0.14	0.23	7473	6739	3338	1134

We next examined whether model fits were consistent with subjects integrating both potential reward magnitudes and probabilities to make decisions. The relative weight of probability and reward magnitude in the additive model (ϕ) was balanced on average (stable: ϕ = 0.57 ± 0.017 [mean ± SEM]; volatile: ϕ = 0.44 ± 0.027), suggesting that subjects integrated both probabilities and reward magnitudes to make decisions. Omitting ϕ and evaluating simpler models that considered only probabilities (α and β) or reward magnitudes (β only) resulted in worse fits (see Table 1). Lower BIC for additive and multiplicate models compared to the simpler models confirmed that subjects considered both probability and reward magnitude when making decisions.

We then asked whether greater environmental volatility was associated with higher learning rates as observed in previous studies (Behrens et al., 2007; Browning et al., 2015; Massi et al., 2018). A simple prediction for standard reinforcement learning models is that subjects should stay more on the same option after winning than losing (Figure 2B). Subjects did not stay more on the same option after winning than losing for the stable environment (difference in choice proportion, high probability car: 4.1 ± 1.9%, mean ± SEM, z = 1.7, p = 0.098; low probability car: 1.3 ± 3.2%, z = -0.43, p = 0.67). Subjects stayed on the same option more after winning than losing in the volatile environment (difference in choice proportion, high probability car: 22.1 ± 2.3%, z = 6.6, p < 10⁻¹⁰, low probability car: 21.4 ± 3.7%, z = 4.8, p < 10⁻⁵). Subjects stayed on the same option after winning compared to losing more in volatile than stable environments (difference volatile – stable, high probability car: 18.1 ± 2.7%, z = 5.7, p < 10⁻⁷, low probability car: 19.8 ± 4.1%, z = 4.0, p < 10⁻⁴; see Figure 2B).

We then checked that the predictions generated by a reinforcement learning model fit separately to each environment correspond to observed behavioural patterns described above in model-independent analyses (predicted difference in choice proportion after winning and losing in stable [high probability car: 8.8 ± 1.6%, z = 5.2, p < 10⁻⁶, low probability car: 6.1 ± 2.6%, z = 1.5, p = 0.12] and volatile environments [high probability car: 21.5 ± 2.2%, z = 6.9, p < 10⁻¹¹, low probability car: 21.1 ± 2.7%, z = 5.9, p < 10⁻⁸; see Figure 2B]). The model predictions were able to capture observed differences in behaviour following wins and losses and also the difference in in choice proportion after winning and losing between volatile and stable environments (high probability car: 12.8 ± 2.4%, z = 4.6, p < 10⁻⁵; low probability car: 14.3 ± 3.6, z = 3.4, p < 0.001). We found that learning rates (Figure 2C) were substantially higher in volatile than stable environments (stable α = 0.16 ± 0.02, mean ± SEM; volatile α = 0.47 ± 0.03; difference volatile – stable: 0.31 ± 0.03, z = 6.9, p < 10⁻¹¹). Overall, these results demonstrate that the learning rate increases substantially in the volatile compared to the stable environment, in line with previous studies (Behrens et al., 2007; Browning et al., 2015; Massi et al., 2018).

Happiness is more sensitive to learning-relevant than learning-irrelevant variables

We next examined how happiness changes over time during the tasks. Subjects varied their happiness ratings in both the stable (SD = 24.2 ± 1.1, mean ± SEM) and volatile (SD = 25.0 ± 1.1) environments. They were happier on average after winning than after losing (stable: 63.8 ± 1.9 vs 34.5 ± 2.0, z = 7.5, p < 10⁻¹³; volatile: 61.9 ± 1.8 vs 33.6 ± 2.0, z = 7.5, p < 10⁻¹³, Figure 3A). Participants were happier on average in the stable environment than in the volatile environment (stable: 55.0 ± 1.7, volatile: 49.5 ± 1.6, z = 3.7, p < 0.001). This effect may be at least partly due to lower choice accuracy in volatile environments, and the difference in average happiness between environments was correlated between participants with the difference in choice accuracy in terms of EV maximisation (Spearman’s ρ(73) = 0.24, p < 0.05).

Figure 3 with 1 supplement see all

Download asset Open asset

Previous studies have reported that momentary happiness in response to outcomes in a probabilistic reward task were explained by recent RPEs when maximising cumulative reward does not require learning (Rutledge et al., 2014; Rutledge et al., 2015). In our task, maximising cumulative reward requires learning the outcome probability. In this context, PPEs (depending on whether the outcome was a win or a loss and the subjective probability of that outcome) are relevant to learning but RPEs (depending on the magnitude of the reward received and the expected value of the chosen option) are not relevant to learning or future behaviour. Reward information was choice relevant and choices were driven by the (additive) integration of the estimated outcome probability and the magnitude of potential rewards. Therefore, we tested whether happiness was more strongly associated with the PPEs used for learning or alternatively by RPEs that incorporate learning-irrelevant reward magnitudes. We compared two models which both use the subjective probability as estimated in the additive choice model to compute prediction errors:

(6) ${\displaystyle Happiness\left(t\right)=w_0+w_{\widehat{PPE}}\sum _{j=1}^t{\gamma }^{t-j}\widehat{PP{E}_j},}$

where ${\displaystyle \widehat{PPE}}$ refers to the probability prediction error (PPE), defined as the difference between the outcome (one for win, 0 for loss) and the subjective probability estimated from the additive choice model, w₀ is a constant term, w_PPE is a weight capturing the influence of past PPEs, and 0 ≤ γ ≤1 is a forgetting factor that makes events in more recent trials more influential than those in earlier trials;

(7) ${\displaystyle Happiness\left(t\right)=w_0+w_{\widehat{RPE}}\sum _{j=1}^t{\gamma }^{t-j}\widehat{RP{E}_j},}$

where ${\displaystyle \widehat{RPE}}$ is the difference between reward magnitude and the expected value of the chosen option computed based on the subjective probability estimated from the additive choice model. Reward magnitudes were rescaled from 0 to 1.

Mood fluctuations were better explained by a model including past PPEs than by a model including past RPEs, both in the stable (BIC_PP̂E = −698, BIC_RP̂E = −299, ∆BIC = 399) and volatile (BIC_PP̂E = −559, BIC_RP̂E = −319, ∆BIC = 240) environments (see Table 2 and Figure 4A). This result holds for a broader model space including other definitions of the prediction error terms (see Table 2 and Figure 4):

(8) ${\displaystyle Happiness\left(t\right)=w_0+w_{PPE}\sum _{j=1}^t{\gamma }^{t-j}PP{E}_j,}$

Figure 4 with 1 supplement see all

Download asset Open asset

Table 2

Model	Number of parameters	Stable mean r2	Volatile mean r2	Stable BIC	Volatile BIC	Stable ΔBIC	Volatile ΔBIC
${\displaystyle \widehat{PPE}}$	3	0.50	0.44	−698	−559	184	587
${\displaystyle \widehat{RPE}}$	3	0.38	0.39	−299	−319	583	826
PPE	3	0.48	0.42	−640	−436	242	710
RPE	3	0.36	0.36	−223	−212	658	934
${\displaystyle \hat{P}+\widehat{PPE}}$	4	0.58	0.62	−882	−1146	0	0
${\displaystyle \widehat{EV}+\widehat{PPE}}$	4	0.56	0.53	−752	−691	130	454
${\displaystyle \widehat{EV}+\widehat{RPE}}$	4	0.43	0.48	−224	−370	657	775
${\displaystyle \hat{P}+\widehat{RPE}}$	4	0.47	0.51	−370	−504	511	641
$R-\stackrel{-}{R}$	3	0.36	0.44	−242	−471	640	675
$R-RP$	4	0.35	0.42	47	−193	929	952
Win −loss	4	0.57	0.63	−848	−1181	34	−35

where PPE refers to the objective PPE defined as the difference between the outcome sign and the objective probability of the chosen option (0.2 or 0.8), and also:

(9) ${\displaystyle Happiness\left(t\right)=w_0+w_{RPE}\sum _{j=1}^t{\gamma }^{t-j}RP{E}_j,}$

where RPE is computed by taking the difference between the reward magnitude and the objective expected value (potential reward multiplied by the objective probability of the chosen option) as above.

We then tested whether mood fluctuations were additionally sensitive to current expectations, as shown in risky choice tasks that do not require learning (Rutledge et al., 2014; Rutledge et al., 2015). Again, two types of expectations may explain mood fluctuations: a subjective probability relevant to learning, or expected values that incorporate learning-irrelevant reward magnitudes. We compared the following models:

(10) ${\displaystyle Happiness\left(t\right)=w_0+w_{\hat{P}}\sum _{j=1}^t{\gamma }^{t-j}\left(\hat{P}-0.5\right)+w_{\widehat{PPE}}\sum _{j=1}^t{\gamma }^{t-j}\widehat{PP{E}_j},}$

where $\widehat{P}$ is the probability estimated with the additive choice model, and:

(11) ${\displaystyle Happiness\left(t\right)=w_0+w_{\widehat{EV}}\sum _{j=1}^t{\gamma }^{t-j}\left(\widehat{EV}-\stackrel{-}{EV}\right)+w_{\widehat{PPE}}\sum _{j=1}^t{\gamma }^{t-j}\widehat{PP{E}_j},}$

where ${\displaystyle \widehat{EV}}$ is the product between ${\displaystyle \hat{P}}$ and the reward magnitude and this term is mean-centred.

The model including choice probability better explained happiness ratings (stable: mean r² = 0.58; volatile: mean r² = 0.62) than the model including the expected value in the stable (BIC_P̂+PP̂E = −882, BIC_EV+PP̂E = −752, ∆BIC = 130) and in the volatile (BIC_P̂+PP̂E = −1147, BIC_EV+PP̂E = −691, ∆BIC = 454) environments (see Table 2 and Figure 4B). The probability and PPE weights were significantly different from 0 at the group level in both the stable (w_P̂ = 0.74 ± 0.09, z = 6.2, p < 10⁻⁹; w_PP̂E = 1.32 ± 0.06, z = 7.5, p < 10⁻¹⁴) and the volatile environments (w_P̂ = 0.94 ± 0.09, z = 6.7, p < 10⁻¹⁰; w_PP̂E = 1.14 ± 0.05, p < 10⁻¹⁴, Figure 3C).

We next extended the model space with plausible alternative models. We included two models incorporating the history of reward magnitude. In the first model, we centred the reward magnitude regressor for each participant. This model thus predicts that reward magnitudes larger than the averaged reward magnitude will increase happiness, and that the larger the reward magnitude, the greater the happiness.

(12) ${\displaystyle Happiness\left(t\right)=w_0+w_R\sum _{j=1}^t{\gamma }^{t-j}\left(R_j-\stackrel{-}{R}\right),}$

where R_j is the reward magnitude at trial j and $\stackrel{-}{R}$ is the average reward magnitude. Instead of assuming a reference point of the average reward, we also used a free parameter in a subsequent model above and below which reward magnitudes increase or decrease happiness, respectively:

(13) ${\displaystyle Happiness\left(t\right)=w_0+w_R\sum _{j=1}^t{\gamma }^{t-j}\left(R_j-RP\right),}$

where RP is a free parameter corresponding to the reference point in an individual subject. If this value is 0, receipt of rewards always increases happiness in proportion to the reward magnitude, so this model also provides a test of whether failing to obtain reward decreases happiness during reinforcement learning. The average reward magnitude was 25.5 ± 2.5 (mean ± SD) points in the stable environment and 22.6 ± 3.0 points in the volatile environment. The reference point RP estimated in model 14 was on average 14.2 ± 9.5 points and greater than 0 in the stable environment (z = 7.0, p < 10⁻¹¹) and 14.8 ± 7.3 points and greater than 0 in the volatile environment (z = 7.4, p < 10⁻¹²). This result supports the idea that obtaining 0 points on a trial is aversive: failing to obtain reward decreases happiness in our tasks. We also included two additional models incorporating ${\displaystyle \widehat{RPE}}$.

(14) ${\displaystyle Happiness\left(t\right)=w_0+w_{\widehat{EV}}\sum _{j=1}^t{\gamma }^{t-j}\left(\widehat{E{V}_j}-\stackrel{-}{EV}\right)+w_{\widehat{RPE}}\sum _{j=1}^t{\gamma }^{t-j}\widehat{RP{E}_j},}$

where ${\displaystyle \widehat{EV}}$ corresponds to the expected value of the chosen option, corresponding to the weighted sum of probability and reward estimated based on each individual participant’s choices with the additive choice model, ${\displaystyle \overline{EV}}$ corresponds to the averaged expected reward, and ${\displaystyle \widehat{RPE}}$ corresponds to the difference between the outcome reward magnitude and ${\displaystyle \widehat{EV}}$. We also included a model combining the estimated probability with the reward prediction error.

(15) ${\displaystyle Happiness\left(t\right)=w_0+w_{\hat{P}}\sum _{j=1}^t{\gamma }^{t-j}\left(\widehat{P_j}-0.5\right)+w_{\widehat{RPE}}\sum _{j=1}^t{\gamma }^{t-j}\widehat{RP{E}_j},}$

Besides the constant and the forgetting factor, the ${\displaystyle \hat{P}+\widehat{PPE}}$ model includes two parameters, one for predictions (w_P̂) and one for probability prediction error (w_PP̂E).

We also asked whether the history of wins (excluding any information about reward magnitude) and losses could account for happiness by fitting the following model:

(16) $Happiness\left(t\right)=w_0+w_{win}{\sum }_{j=1}^t{\gamma }^{t-j}wi{n}_j-w_{loss}{\sum }_{j=1}^t{\gamma }^{t-j}los{s}_j$

As reported in Table 2, the model evidence for this new model was similar to the ${\displaystyle \hat{P}+\widehat{PPE}}$ model overall. We next used estimated model frequency to compare both models. The ${\displaystyle \hat{P}+\widehat{PPE}}$ is preferred to the win-loss model in the stable environment (EF_P̂+PP̂E = 0.65 ± 0.05, EF_win-loss = 0.35 ± 0.05, exceedance probability = 0.99). However, both models performed similarly in the volatile environment (EF_P̂+PP̂E = 0.50 ± 0.06, EF_win-loss = 0.50 ± 0.06, exceedance probability = 0.48; see Figure 4—figure supplement 1).

We next asked whether an alternative analysis could test whether happiness was influenced by trial-by-trial probability estimates, the key difference between the models. If the weights for the two terms of the ${\displaystyle \hat{P}+\widehat{PPE}}$ model are identical, the equation mathematically reduces to a constant plus an exponentially weighted average of previous wins. However, the w_PP̂E parameter was larger than w_P̂ in both stable (z = 6.0, p < 10⁻⁸) and volatile tasks (z = 2.35, p < 0.05). The difference between w_PP̂E and w_P̂ was significantly larger in the stable compared to the volatile task (z = 3.7, p < 0.001). Comparison of weights across tasks therefore suggests a reduced impact of expectations on happiness as environmental volatility increases. We next computed the residuals of the win-loss model (which does not include probability estimates) and tested for a correlation with trial-by-trial probability estimates. Because prediction errors are equal to outcomes minus expectations and numerically w_P̂ is lower than w_PP̂E in both environments, the overall influence of probability on happiness should be negative after accounting for the impact of wins and losses. In the stable environment, we found the expected negative correlation between the win-loss model residuals and trial-by-trial probability estimates (average Spearman’s ρ(73) = −0.06 ± 0.03, z = 2.2, p = 0.03). This relationship was not present in the volatile environment (average Spearman’s ρ(73)=−0.02 ± 0.03, z = 0.65, p = 0.51). A potential explanation for this pattern of results is that expectations cannot affect happiness when participants do not have strong predictions, as it is the case immediately after reversals in the volatile condition. This would be consistent with findings from the animal literature showing that dopamine early in training does not represent prediction errors (Coddington and Dudman, 2018).

Finally, we focused on the win-loss model. We asked whether weights from the win-loss model were positively correlated, consistent with similar but opposite impacts. We found instead a negative correlation across participants between win and loss weights (stable: Spearman’s ρ(73) = −0.56, p < 10⁻⁶; volatile: Spearman’s ρ(73) = −0.68, p < 10⁻²⁰), suggesting that individuals that respond to wins tend to respond less to losses and vice versa. Indeed, comparing the weight of wins and losses shows that participants reacted more strongly on average to losses than to wins (difference in stable: 0.69 ± 0.12, z = 5.2, p < 10⁻⁶; difference in volatile: 0.31 ± 0.13, z = 2.7, p < 0.01; see Figure 3—figure supplement 1). Given that participants received positive feedback on average in 81% of trials in the stable environment and 63% of trials in the volatile environment, asymmetric responses to wins and losses are consistent with happiness reflecting knowledge of the underlying structure of both environments. Interestingly, the difference between win and loss weights was not correlated across participants with overall performance including the percentage of trials with positive feedback (stable: Spearman’s ρ(73) = −0.05, p = 0.65; volatile: Spearman’s ρ(73) = −0.18, p = 0.12) or the percentage of trials where the higher expected value option was chosen (stable: Spearman’s ρ(73) = 0.04, p = 0.74; volatile: Spearman’s ρ(73) = −0.17, p = 0.14).

Our results suggest that although reward information influences choice, contrary to what would be predicted from the literature, RPEs and reward magnitudes do not explain happiness when this information is not necessary for participants to learn the structure of the environment. Happiness reflects knowledge of the underlying structure of the environment in a way that cannot be explained by simple performance metrics. RPEs are relevant to learning in many paradigms, and happiness should relate to RPEs in such tasks because of their value for learning the structure of environment. Learning and reward are dissociable in our paradigm, and we find in this context that RPEs and reward magnitudes do not explain happiness.

Sensitivity of mood dynamics to learning variables depends on volatility but not on learning rate

We found that the learning rate (i.e., behavioural sensitivity to PPE) was approximately three times higher in the volatile compared to the stable environment. Furthermore, mood dynamics were highly sensitive to PPE. However, PPE weights were actually higher in stable than volatile environments (∆w_PP̂E = 0.18 ± 0.05, z = 3.5, p < 0.001; see Figure 3C). Futhermore, the difference between w_PP̂E and w_P̂ was greater in stable than volatile environments (stable – volatile: 0.39 ± 0.11, z = 3.7, p < 0.001), consistent with a greater influence of trial-by-trial probability estimates on happiness in stable environments (see previous section). PPE weights were not correlated between participants with the learning rate in the stable (Spearman’s ρ(73) = −0.10, p = 0.40) and volatile (Spearman’s ρ(73) = −0.1, p = 0.37) environments nor was the difference of PPE weights across environments related to the difference in learning rate (Spearman’s ρ(73) = −0.02, p = 0.84). Instead, PPE weights were highly consistent across environments (Spearman’s ρ(73) = 0.44, p < 0.001, see Figure 5A).

Figure 5 with 1 supplement see all

Download asset Open asset

The happiness model forgetting factor γ determines how many previous trials influence current affective state. When γ is equal to 1, mood is equally influenced by all previous trials, when γ is equal to 0, mood is influenced by only the most recent trial. If the change in forgetting factor mirrors behaviour, forgetting factors should be lower in volatile than stable environments, reflecting integration over fewer trials and consistent with the higher learning rates observed in volatile compared to stable environments. Instead, the forgetting factor was slightly higher on average in the volatile environment (stable: γ = 0.59 ± 0.04, volatile γ = 0.63 ± 0.03, corresponding to current happiness being influenced by 6–7 previous trials on average in both environments, stable – volatile: ∆γ = 0.05 ± 0.03, z = 1.9, p = 0.064; see Figure 5B). Higher values for γ in volatile environments are not consistent with happiness integrating over fewer trials as behaviour would predict. Furthermore, the change in happiness forgetting factor was not correlated across participants with the learning rate difference between environments (Spearman’s ρ(73) = −0.08, p = 0.50). A linear regression with ten previous probability prediction errors as independent variables confirmed this model-based result (see Figure 5—figure supplement 1). To further test for a relationship between the forgetting factor and the learning rate, we switched the learning rates estimated from stable and volatile conditions in each individual before re-fitting happiness data (i.e., we used the ‘wrong’ learning rate to estimate probabilities and PPEs before fitting the happiness model and estimating a forgetting factor). This did not substantially affect estimates of the happiness forgetting factor (stable γ = 0.58 ± 0.03, volatile γ = 0.62 ± 0.03, stable – volatile: ∆γ = 0.032 ± 0.026, z = 1.4, p = 0.16). The resulting forgetting factor estimates were highly correlated with forgetting factors estimated using the actual learning rates (stable: Spearman’s ρ(73) = 0.63, p < 10⁻⁸; volatile: Spearman’s ρ(73) = 0.56, p < 10⁻⁶). The happiness forgetting factor was highly consistent across environments (Spearman’s ρ(73) = 0.41, p < 0.001, see Figure 5C), suggesting that the number of previous trials that affective state depends on may be a trait-like feature of individuals unrelated to environmental volatility.

Depressive symptoms are associated with reduced happiness in volatile environments

Previous studies have linked learning rates to anxiety (Browning et al., 2015; Pulcu and Browning, 2019) and individuals with high trait anxiety showed less ability to appropriately adjust updating of outcome expectancies between stable and volatile environments. We found that depressive symptoms (PHQ) were uncorrelated across participants with choice accuracy (stable: Spearman’s ρ(73) = −0.06, p = 0.63; volatile: Spearman’s ρ(73) = −0.09, p = 0.46; volatile – stable: Spearman’s ρ(73) = 0.04, p = 0.71) and all parameters estimated in the additive choice model (α, stable: Spearman’s ρ(73) = 0.08, p = 0.51; volatile: Spearman’s ρ(73) = 0.10, p = 0.38; volatile – stable: Spearman’s ρ(73) = 0.19, p = 0.1; ϕ, stable: Spearman’s ρ(73) = 0.20, p = 0.09; volatile: Spearman’s ρ(73) = 0.01, p = 0.94; volatile – stable: Spearman’s ρ(73) = −0.20, p = 0.09; β, stable: Spearman’s ρ(73) = −0.04, p = 0.76; volatile: Spearman’s ρ(73) = −0.07, p = 0.53; volatile – stable: Spearman’s ρ(73) = 0.04, p = 0.75). In the stable environment, where uncertainty and volatility are low, average happiness did not correlate across participants with depressive symptoms (Spearman’s ρ(73) = 0.07, p = 0.58; Figure 6A, left panel). In the volatile environment, where uncertainty is high and volatility is high, average happiness was correlated with depressive symptoms, with lower happiness associated with higher depressive symptoms (Spearman’s ρ(73) = −0.23, p = 0.043; Figure 6A, central panel). Finally, the difference between average happiness between volatile and stable environments was also correlated with depressive symptoms even after standardising the variables (Wilcox and Tian, 2008) (volatile – stable: Spearman’s ρ(73) = −0.28, p = 0.014; Figure 6A, right panel). Baseline mood parameters estimated using our happiness model fit to non-z-scored happiness ratings showed the same relationship to depressive symptoms (stable: Spearman’s ρ(73) = −0.07, p = 0.54; volatile: Spearman’s ρ(73) = −0.28, p = 0.017; volatile – stable, standardised: Spearman’s ρ(73) = −0.32, p = 0.0049; see Figure 6B).

Figure 6

Download asset Open asset

No other happiness model parameters were correlated with depressive symptoms (w_P̂, stable: Spearman’s ρ(73) = 0.20, p = 0.09; volatile: Spearman’s ρ(73) = 0.15, p = 0.19; volatile – stable: Spearman’s ρ(73) = −0.02, p = 0.84; w_PP̂E, stable: Spearman’s ρ(73) = 0.09, p = 0.45; volatile: Spearman’s ρ(73) = 0.07, p = 0.54; volatile – stable: Spearman’s ρ(73) = −0.03, p = 0.81; γ, stable: Spearman’s ρ(73) = 0.06, p = 0.63; volatile: Spearman’s ρ(73) = −0.09, p = 0.44; volatile – stable: Spearman’s ρ(73) = −0.12, p = 0.28).

We found that subjects tracked outcome probabilities and made decisions by integrating both learned probability and explicit reward magnitudes. Learning rates adapted to environmental volatility, with a higher learning rate in the more volatile environment consistent with previous studies (Behrens et al., 2007; Browning et al., 2015; Massi et al., 2018). That behaviour was consistent with an additive choice model (Donahue and Lee, 2015; Farashahi et al., 2017; Massi et al., 2018; Rouault et al., 2019) which is consistent with recent empirical evidence (Farashahi et al., 2019; Koechlin, 2020) that humans and non-human primates adopt a multiplicative strategy under risk when probabilities are explicit, but both species adopt an additive strategy under uncertainty when probabilities must be learned.

Our tasks required learning the probability of getting a reward and the reward magnitude was explicitly given. In such an environment, mood dynamics were more closely related to learning-relevant variables than learning-irrelevant variables and we found convergent evidence that this was the case across both stable and volatile learning tasks. Mood was sensitive to the combined influence of past chosen subjective probabilities and past PPEs. Parameters for PPE and forgetting factors estimated from happiness ratings were correlated across stable and volatile environments. Finally, we found that although choice accuracy and choice model parameters were not affected by depressive symptoms when changes between safe and volatile environments are signalled, the decrease in happiness observed in the volatile relative to the stable environment was correlated with symptom severity. The same pattern was present for the baseline mood parameter in the happiness model. Experiencing a stable environment with low uncertainty and volatility could attenuate the expression of depressive symptoms on mood. Risky decision tasks used in previous studies (Rutledge et al., 2017; Rutledge et al., 2015; Rutledge et al., 2014) maximise irreducible uncertainty (i.e., risky options had a 50% probability of each option), which is more comparable to the volatile environment in the current study and may explain the previous finding of a link between baseline mood parameters and depressive symptoms. Computational models that capture ecologically relevant learning and decision processes may provide a critical advantage for understanding the mechanisms that underlie psychiatric symptoms (Scholl and Klein-Flügge, 2018). Our findings suggest that subjective feelings measured during tasks in depression-relevant domains may provide additional information not captured by computational models of learning and decision-making. One reason depression might reduce mood more in volatile than stable environments could be an increase in the number of negative prediction errors experienced. Misestimation of the level of uncertainty may also lead to a tendency for negative events to disproportionally affect depressed individuals, and this uncertainty misestimation is believed to contribute to depression and anxiety (Pulcu and Browning, 2019). That the learning rate difference between the volatile and the stable environment did not correlate with anxiety symptom severity is consistent with previous findings (Browning et al., 2015) that anxiety is linked to learning deficits for aversive but not appetitive outcomes. In the aversive domain, anxiety severity might be associated with differences in behavioural adaptation to volatility changes as well as mood.

It is not yet established what the neural signal associated with PPEs is. On the one hand, reward prediction errors have been associated with neuromodulator dopamine and are thought to be linked to ventral tegmental area (Bayer and Glimcher, 2005; Cohen et al., 2012; Hart et al., 2014; Montague et al., 1996; Pessiglione et al., 2006; Schultz et al., 1997). In most studies, RPEs and PPEs are equivalent and therefore the specific link between PPEs and dopamine is less documented. The probability of obtaining reward and probability prediction error have been associated with VTA activity correcting for expected value (Behrens et al., 2007), suggesting that PPEs may be represented by dopamine. Boosting dopamine levels pharmacologically during risky decision making increases the happiness resulting from smaller rewards to a level similar to that resulting from larger rewards (Rutledge et al., 2015). Although happiness in that study was influenced by the history of RPEs, dopamine drug impacts were limited to rewards. Dopamine has also been associated with other signals than prediction errors, for example incentive salience which might relate to ‘wanting’ and might influence choice and action (Berridge, 2012; Smith et al., 2011; Zhang et al., 2009).

Some studies suggest that dopaminergic activity in the midbrain is linked to information-seeking to reduce uncertainty about an upcoming reward, even though such information is not instrumental (Bromberg-Martin and Hikosaka, 2009; Brydevall et al., 2018; Charpentier et al., 2018; Gruber and Ranganath, 2019). The intrinsic reward resulting from reducing uncertainty could influence mood, and mood ratings could then be used to quantify the relative subjective weight of extrinsic and intrinsic reward.

Previous studies using risky decision tasks where reward and probability were explicitly represented (Rutledge et al., 2014; Rutledge et al., 2015) showed that mood dynamics were explained by past expected values and RPEs. The present design allows us to dissociate the impact of learning-relevant and learning-irrelevant information for mood in two different standard learning environments. Our results suggest that when goal attainment requires adaptive behaviour, mood dynamics reflect learning-relevant information. Consistent with the results obtained from risky decision tasks used in previous studies (Rutledge et al., 2014; Rutledge et al., 2015; Rutledge et al., 2017), potential rewards were a key determinant of behaviour in our task. However, rewards were not a determinant of mood in the current study, in contrast to previous results in risky decision tasks (Rutledge et al., 2014; Rutledge et al., 2015; Rutledge et al., 2017). Unlike most reinforcement learning experiments, our task design allows dissociating the impacts of PPEs and RPEs on behaviour and mood. Here, our results suggest that happiness does not always depend on reward and preferentially reflects learning about the structure of the environment. However, if learning the structure of the environment requires tracking changing reward magnitudes, we would expect that happiness would track learning-relevant variables (e.g., reward magnitudes and RPEs in such an environment).

This result might imply a role for mood in learning in line with influential proposals (Eldar et al., 2016; Eldar and Niv, 2015). However, mood did not reflect all learning-relevant information that influenced behaviour. Happiness forgetting factors corresponding to the number of past trials that influence affective state were highly correlated across environments and acted more as a stable trait that differed between individuals and did not adjust to environmental volatility. Overall, our findings show that mood dynamics are sensitive to depressive symptoms and reflect variables relevant to adaptive behaviour irrespective of environmental volatility.

Data and code are available online (https://github.com/BastienBlain/MSWB_LearningNotReward; copy archived at https://archive.softwareheritage.org/swh:1:rev:b7c4a0cd761dcf249c72caf809dd81af24c4a49b/).

1. Book

   1. Barto AG

   (1995)

   Adaptive critics and the basal ganglia

   In: Houk J. C, Davis J. L, editors. Models of Information Processing in the Basal Ganglia, Computational Neuroscience. The MIT Press. pp. 215–232.

   • Google Scholar
2. 1. Bayer HM
   2. Glimcher PW

   (2005) Midbrain dopamine neurons encode a quantitative reward prediction error signal

   Neuron 47:129–141.

   https://doi.org/10.1016/j.neuron.2005.05.020

   • PubMed
   • Google Scholar
3. 1. Beck AT
   2. Steer RA
   3. Ball R
   4. Ranieri W

   (1996) Comparison of beck depression inventories -IA and -II in psychiatric outpatients

   Journal of Personality Assessment 67:588–597.

   https://doi.org/10.1207/s15327752jpa6703_13

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

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

   Nature Neuroscience 10:1214–1221.

   https://doi.org/10.1038/nn1954

   • PubMed
   • Google Scholar
5. 1. Berridge KC

   (2012) From prediction error to incentive salience: mesolimbic computation of reward motivation

   European Journal of Neuroscience 35:1124–1143.

   https://doi.org/10.1111/j.1460-9568.2012.07990.x

   • PubMed
   • Google Scholar
6. Software

   1. Blain B

   (2020) MSWB_LearningNotReward, version swh:1:rev:b7c4a0cd761dcf249c72caf809dd81af24c4a49b

   Software Heritage.

   https://archive.softwareheritage.org/swh:1:dir:e9174f278be2f11c918d94546c15347580285ac9;origin=https://github.com/BastienBlain/MSWB_LearningNotReward;visit=swh:1:snp:8db3baca6ab7c4542691213db4576c2def9c6016;anchor=swh:1:rev:b7c4a0cd761dcf249c72caf809dd81af24c4a49b/
7. 1. Blanco NJ
   2. Otto AR
   3. Maddox WT
   4. Beevers CG
   5. Love BC

   (2013) The influence of depression symptoms on exploratory decision-making

   Cognition 129:563–568.

   https://doi.org/10.1016/j.cognition.2013.08.018

   • PubMed
   • Google Scholar
8. 1. Bromberg-Martin ES
   2. Hikosaka O

   (2009) Midbrain dopamine neurons signal preference for advance information about upcoming rewards

   Neuron 63:119–126.

   https://doi.org/10.1016/j.neuron.2009.06.009

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

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

   Nature Neuroscience 18:590–596.

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

   • PubMed
   • Google Scholar
10. 1. Brydevall M
    2. Bennett D
    3. Murawski C
    4. Bode S

    (2018) The neural encoding of information prediction errors during non-instrumental information seeking

    Scientific Reports 8:1–11.

    https://doi.org/10.1038/s41598-018-24566-x

    • PubMed
    • Google Scholar
11. Book

    1. Burnham KP
    2. Anderson DR

    (2004) Model Selection and Multi-Model Inference: A Practical Information-Theoretic Approach

    Springer.

    https://doi.org/10.1007/b97636

    • Google Scholar
12. 1. Cella M
    2. Dymond S
    3. Cooper A

    (2010) Impaired flexible decision-making in major depressive disorder

    Journal of Affective Disorders 124:207–210.

    https://doi.org/10.1016/j.jad.2009.11.013

    • PubMed
    • Google Scholar
13. 1. Charpentier CJ
    2. Bromberg-Martin ES
    3. Sharot T

    (2018) Valuation of knowledge and ignorance in mesolimbic reward circuitry

    PNAS 115:E7255–E7264.

    https://doi.org/10.1073/pnas.1800547115

    • PubMed
    • Google Scholar
14. 1. Chase HW
    2. Frank MJ
    3. Michael A
    4. Bullmore ET
    5. Sahakian BJ
    6. Robbins TW

    (2010) Approach and avoidance learning in patients with major depression and healthy controls: relation to anhedonia

    Psychological Medicine 40:433–440.

    https://doi.org/10.1017/S0033291709990468

    • PubMed
    • Google Scholar
15. 1. Coddington LT
    2. Dudman JT

    (2018) The timing of action determines reward prediction signals in identified midbrain dopamine neurons

    Nature Neuroscience 21:1563–1573.

    https://doi.org/10.1038/s41593-018-0245-7

    • PubMed
    • Google Scholar
16. 1. Cohen JY
    2. Haesler S
    3. Vong L
    4. Lowell BB
    5. Uchida N

    (2012) Neuron-type-specific signals for reward and punishment in the ventral tegmental area

    Nature 482:85–88.

    https://doi.org/10.1038/nature10754

    • PubMed
    • Google Scholar
17. 1. Daunizeau J
    2. Adam V
    3. Rigoux L

    (2014) VBA: a probabilistic treatment of nonlinear models for neurobiological and behavioural data

    PLOS Computational Biology 10:e1003441.

    https://doi.org/10.1371/journal.pcbi.1003441

    • PubMed
    • Google Scholar
18. 1. de Berker AO
    2. Rutledge RB
    3. Mathys C
    4. Marshall L
    5. Cross GF
    6. Dolan RJ
    7. Bestmann S

    (2016) Computations of uncertainty mediate acute stress responses in humans

    Nature Communications 7:10996.

    https://doi.org/10.1038/ncomms10996

    • PubMed
    • Google Scholar
19. 1. Donahue CH
    2. Lee D

    (2015) Dynamic routing of task-relevant signals for decision making in dorsolateral prefrontal cortex

    Nature Neuroscience 18:295–301.

    https://doi.org/10.1038/nn.3918

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

    (2016) Mood as representation of momentum

    Trends in Cognitive Sciences 20:15–24.

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

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

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

    Nature Communications 6:6149.

    https://doi.org/10.1038/ncomms7149

    • PubMed
    • Google Scholar
22. 1. Farashahi S
    2. Donahue CH
    3. Khorsand P
    4. Seo H
    5. Lee D
    6. Soltani A

    (2017) Metaplasticity as a neural substrate for adaptive learning and choice under uncertainty

    Neuron 94:401–414.

    https://doi.org/10.1016/j.neuron.2017.03.044

    • PubMed
    • Google Scholar
23. 1. Farashahi S
    2. Donahue CH
    3. Hayden BY
    4. Lee D
    5. Soltani A

    (2019) Flexible combination of reward information across primates

    Nature Human Behaviour 3:1215–1224.

    https://doi.org/10.1038/s41562-019-0714-3

    • PubMed
    • Google Scholar
24. 1. Fredrickson BL

    (2004) The broaden-and-build theory of positive emotions

    Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences 359:1367–1377.

    https://doi.org/10.1098/rstb.2004.1512

    • PubMed
    • Google Scholar
25. 1. Gillan CM
    2. Kosinski M
    3. Whelan R
    4. Phelps EA
    5. Daw ND

    (2016) Characterizing a psychiatric symptom dimension related to deficits in goal-directed control

    eLife 5:e11305.

    https://doi.org/10.7554/eLife.11305

    • PubMed
    • Google Scholar
26. 1. Gruber MJ
    2. Ranganath C

    (2019) How curiosity enhances Hippocampus-Dependent memory: the prediction, appraisal, curiosity, and exploration (PACE) Framework

    Trends in Cognitive Sciences 23:1014–1025.

    https://doi.org/10.1016/j.tics.2019.10.003

    • PubMed
    • Google Scholar
27. 1. Hart AS
    2. Rutledge RB
    3. Glimcher PW
    4. Phillips PE

    (2014) Phasic dopamine release in the rat nucleus accumbens symmetrically encodes a reward prediction error term

    The Journal of Neuroscience 34:698–704.

    https://doi.org/10.1523/JNEUROSCI.2489-13.2014

    • PubMed
    • Google Scholar
28. 1. Herzallah MM
    2. Moustafa AA
    3. Natsheh JY
    4. Abdellatif SM
    5. Taha MB
    6. Tayem YI
    7. Sehwail MA
    8. Amleh I
    9. Petrides G
    10. Myers CE
    11. Gluck MA

    (2013) Learning from negative feedback in patients with major depressive disorder is attenuated by SSRI antidepressants

    Frontiers in Integrative Neuroscience 7:67.

    https://doi.org/10.3389/fnint.2013.00067

    • PubMed
    • Google Scholar
29. 1. Huys QJ
    2. Pizzagalli DA
    3. Bogdan R
    4. Dayan P

    (2013) Mapping anhedonia onto reinforcement learning: a behavioural meta-analysis

    Biology of Mood & Anxiety Disorders 3:12.

    https://doi.org/10.1186/2045-5380-3-12

    • PubMed
    • Google Scholar
30. 1. Kahneman D
    2. Tversky A

    (1979) Prospect theory: an analysis of decision under risk

    Econometrica 47:263–391.

    https://doi.org/10.2307/1914185

    • Google Scholar
31. 1. Koechlin E

    (2020) Human Decision-Making beyond the rational decision theory

    Trends in Cognitive Sciences 24:4–6.

    https://doi.org/10.1016/j.tics.2019.11.001

    • PubMed
    • Google Scholar
32. Book

    1. Kreft IG
    2. De Leeuw J

    (1998) Introducing Multilevel Modeling

    Sage.

    https://doi.org/10.1007/978-0-387-73186-5_1

    • Google Scholar
33. 1. Kroenke K
    2. Spitzer RL
    3. Williams JB

    (2001) The PHQ-9: validity of a brief depression severity measure

    Journal of General Internal Medicine 16:606–613.

    https://doi.org/10.1046/j.1525-1497.2001.016009606.x

    • PubMed
    • Google Scholar
34. 1. Kunisato Y
    2. Okamoto Y
    3. Ueda K
    4. Onoda K
    5. Okada G
    6. Yoshimura S
    7. Suzuki S
    8. Samejima K
    9. Yamawaki S

    (2012) Effects of depression on reward-based decision making and variability of action in probabilistic learning

    Journal of Behavior Therapy and Experimental Psychiatry 43:1088–1094.

    https://doi.org/10.1016/j.jbtep.2012.05.007

    • PubMed
    • Google Scholar
35. 1. Massi B
    2. Donahue CH
    3. Lee D

    (2018) Volatility facilitates value updating in the prefrontal cortex

    Neuron 99:598–608.

    https://doi.org/10.1016/j.neuron.2018.06.033

    • PubMed
    • Google Scholar
36. 1. Mathys C
    2. Daunizeau J
    3. Friston KJ
    4. Stephan KE

    (2011) A Bayesian foundation for individual learning under uncertainty

    Frontiers in Human Neuroscience 5:39.

    https://doi.org/10.3389/fnhum.2011.00039

    • PubMed
    • Google Scholar
37. 1. Mellers BA
    2. Schwartz A
    3. Ho K
    4. Ritov I

    (1997) Decision affect theory: emotional reactions to the outcomes of risky options

    Psychological Science 8:423–429.

    https://doi.org/10.1111/j.1467-9280.1997.tb00455.x

    • Google Scholar
38. 1. Montague PR
    2. Dayan P
    3. Sejnowski TJ

    (1996) A framework for mesencephalic dopamine systems based on predictive hebbian learning

    The Journal of Neuroscience 16:1936–1947.

    https://doi.org/10.1523/JNEUROSCI.16-05-01936.1996

    • PubMed
    • Google Scholar
39. 1. Mueller EM
    2. Pechtel P
    3. Cohen AL
    4. Douglas SR
    5. Pizzagalli DA

    (2015) Potentiated processing of negative feedback in depression is attenuated by anhedonia

    Depression and Anxiety 32:296–305.

    https://doi.org/10.1002/da.22338

    • PubMed
    • Google Scholar
40. 1. Otto AR
    2. Fleming SM
    3. Glimcher PW

    (2016) Unexpected but incidental positive outcomes predict Real-World gambling

    Psychological Science 27:299–311.

    https://doi.org/10.1177/0956797615618366

    • PubMed
    • Google Scholar
41. 1. Pechtel P
    2. Dutra SJ
    3. Goetz EL
    4. Pizzagalli DA

    (2013) Blunted reward responsiveness in remitted depression

    Journal of Psychiatric Research 47:1864–1869.

    https://doi.org/10.1016/j.jpsychires.2013.08.011

    • PubMed
    • Google Scholar
42. 1. Pessiglione M
    2. Seymour B
    3. Flandin G
    4. Dolan RJ
    5. Frith CD

    (2006) Dopamine-dependent prediction errors underpin reward-seeking behaviour in humans

    Nature 442:1042–1045.

    https://doi.org/10.1038/nature05051

    • PubMed
    • Google Scholar
43. 1. Pulcu E
    2. Browning M

    (2019) The misestimation of uncertainty in affective disorders

    Trends in Cognitive Sciences 23:865–875.

    https://doi.org/10.1016/j.tics.2019.07.007

    • PubMed
    • Google Scholar
44. 1. Robinson OJ
    2. Cools R
    3. Carlisi CO
    4. Sahakian BJ
    5. Drevets WC

    (2012) Ventral striatum response during reward and punishment reversal learning in unmedicated major depressive disorder

    American Journal of Psychiatry 169:152–159.

    https://doi.org/10.1176/appi.ajp.2011.11010137

    • PubMed
    • Google Scholar
45. 1. Rouault M
    2. Drugowitsch J
    3. Koechlin E

    (2019) Prefrontal mechanisms combining rewards and beliefs in human decision-making

    Nature Communications 10:1–16.

    https://doi.org/10.1038/s41467-018-08121-w

    • PubMed
    • Google Scholar
46. 1. Rutledge RB
    2. Skandali N
    3. Dayan P
    4. Dolan RJ

    (2014) A computational and neural model of momentary subjective well-being

    PNAS 111:12252–12257.

    https://doi.org/10.1073/pnas.1407535111

    • PubMed
    • Google Scholar
47. 1. Rutledge RB
    2. Skandali N
    3. Dayan P
    4. Dolan RJ

    (2015) Dopaminergic modulation of decision making and subjective Well-Being

    Journal of Neuroscience 35:9811–9822.

    https://doi.org/10.1523/JNEUROSCI.0702-15.2015

    • PubMed
    • Google Scholar
48. 1. Rutledge RB
    2. Moutoussis M
    3. Smittenaar P
    4. Zeidman P
    5. Taylor T
    6. Hrynkiewicz L
    7. Lam J
    8. Skandali N
    9. Siegel JZ
    10. Ousdal OT
    11. Prabhu G
    12. Dayan P
    13. Fonagy P
    14. Dolan RJ

    (2017) Association of neural and emotional impacts of reward prediction errors with major depression

    JAMA Psychiatry 74:790–797.

    https://doi.org/10.1001/jamapsychiatry.2017.1713

    • PubMed
    • Google Scholar
49. 1. Scholl J
    2. Klein-Flügge M

    (2018) Understanding psychiatric disorder by capturing ecologically relevant features of learning and decision-making

    Behavioural Brain Research 355:56–75.

    https://doi.org/10.1016/j.bbr.2017.09.050

    • PubMed
    • Google Scholar
50. 1. Schultz W
    2. Dayan P
    3. Montague PR

    (1997) A neural substrate of prediction and reward

    Science 275:1593–1599.

    https://doi.org/10.1126/science.275.5306.1593

    • PubMed
    • Google Scholar
51. 1. Schwarz G

    (1978) Estimating the dimension of a model

    The Annals of Statistics 6:461–464.

    https://doi.org/10.1214/aos/1176344136

    • Google Scholar
52. 1. Shepperd JA
    2. Mcnulty JK

    (2002) The affective consequences of expected and unexpected outcomes

    Psychological Science 13:85–88.

    https://doi.org/10.1111/1467-9280.00416

    • PubMed
    • Google Scholar
53. 1. Smith KS
    2. Berridge KC
    3. Aldridge JW

    (2011) Disentangling pleasure from incentive salience and learning signals in brain reward circuitry

    PNAS 108:E255–E264.

    https://doi.org/10.1073/pnas.1101920108

    • PubMed
    • Google Scholar
54. Book

    1. Spielberger CD

    (1983) State Trait Anxiety Inventory for Adults: Sampler Set: Manual, Test, Scoring Key, the Corsini Encyclopedia of Psychology

    Mind Garden.

    https://doi.org/10.1002/9780470479216

    • Google Scholar
55. 1. Stephan KE
    2. Penny WD
    3. Daunizeau J
    4. Moran RJ
    5. Friston KJ

    (2009) Bayesian model selection for group studies

    NeuroImage 46:1004–1017.

    https://doi.org/10.1016/j.neuroimage.2009.03.025

    • PubMed
    • Google Scholar
56. 1. Taylor Tavares JV
    2. Clark L
    3. Furey ML
    4. Williams GB
    5. Sahakian BJ
    6. Drevets WC

    (2008) Neural basis of abnormal response to negative feedback in unmedicated mood disorders

    NeuroImage 42:1118–1126.

    https://doi.org/10.1016/j.neuroimage.2008.05.049

    • PubMed
    • Google Scholar
57. 1. Thoma P
    2. Norra C
    3. Juckel G
    4. Suchan B
    5. Bellebaum C

    (2015) Performance monitoring and empathy during active and observational learning in patients with major depression

    Biological Psychology 109:222–231.

    https://doi.org/10.1016/j.biopsycho.2015.06.002

    • PubMed
    • Google Scholar
58. 1. Villano WJ
    2. Otto AR
    3. Ezie CEC
    4. Gillis R
    5. Heller AS

    (2020) Temporal dynamics of real-world emotion are more strongly linked to prediction error than outcome

    Journal of Experimental Psychology: General 149:1755–1766.

    https://doi.org/10.1037/xge0000740

    • PubMed
    • Google Scholar
59. 1. Vrieze E
    2. Pizzagalli DA
    3. Demyttenaere K
    4. Hompes T
    5. Sienaert P
    6. de Boer P
    7. Schmidt M
    8. Claes S

    (2013) Reduced reward learning predicts outcome in major depressive disorder

    Biological Psychiatry 73:639–645.

    https://doi.org/10.1016/j.biopsych.2012.10.014

    • PubMed
    • Google Scholar
60. 1. Wilcox RR
    2. Tian T

    (2008) Comparing dependent correlations

    The Journal of General Psychology 135:105–112.

    https://doi.org/10.3200/GENP.135.1.105-112

    • PubMed
    • Google Scholar
61. 1. Zhang J
    2. Berridge KC
    3. Tindell AJ
    4. Smith KS
    5. Aldridge JW

    (2009) A neural computational model of incentive salience

    PLOS Computational Biology 5:e1000437.

    https://doi.org/10.1371/journal.pcbi.1000437

    • PubMed
    • Google Scholar

Author details

1. Bastien Blain

   Max Planck UCL Centre for Computational Psychiatry and Ageing Research, University College London, London, United Kingdom

   Contribution

   Conceptualization, Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing

   For correspondence

   b.blain@ucl.ac.uk

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-7735-6043

2. Robb B Rutledge

   1. Max Planck UCL Centre for Computational Psychiatry and Ageing Research, University College London, London, United Kingdom
   2. Wellcome Centre for Human Neuroimaging, University College London, London, United Kingdom
   3. Department of Psychology, Yale University, New Haven, United States

   Contribution

   Conceptualization, Resources, Formal analysis, Supervision, Validation, Investigation, Project administration, Writing - review and editing

   For correspondence

   robb.rutledge@yale.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-7337-5039

• 6,001

  views

• 733

  downloads

• 35

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.