Do propranolol and amisulpride modulate confidence in risk- taking? [version 1; peer review: 1 approved with reservations]

Background: Making rational choices and being able to consciously reflect on the goodness of these choices is important for successfully navigating the world. Value-based decisions have been extensively studied, but we know little about the factors that influence our confidence in value-based choice. Particularly, we know very little about the neurotransmitters that may mediate these processes. Methods: In this double-blind, placebo-controlled study design involving 61 healthy human subjects (30 female), we assessed the contributions of dopamine (400 mg amisulpride) and noradrenaline (40 mg propranolol) to value-based decision making and the subjective confidence therein in a monetary risky gambling task. Results: We did not find any significant effect of either of the two pharmacological manipulations, neither on value-based decision making, nor on subjective confidence. Conclusion: We discuss these (null) findings, and release all relevant data and code. This will allow researchers to further interrogate the data, to counteract publication biases in favour of significant findings, and to use our study as a source for balanced meta-analyses.


Introduction
Considering the probability and magnitude of a potential outcome is critical for adequate decision making. Over the last decades, value-based decision making has been intensively studied (Brandstätter et al., 2006;Kahneman & Tversky, 1979;Rangel et al., 2008). Despite this wealth of research, relatively little is known about the factors that influence our conscious perception about whether such a choice was good. A feeling of confidence is important to evaluate our decisions and guide our future choices (Fleming & Dolan, 2012). If confidence does not adequately reflect one's choice, then this can be detrimental to long-term outcomes ( Indeed, in a previous study, we showed that noradrenaline plays a role in perceptual metacognition (Hauser et al., 2017a), and another study found that dopaminergic modulation impairs memory-related metacognition (Clos et al., 2019). In this study, we explore whether those findings transfer to the domain of value-based decision making.
Both dopamine and noradrenaline are also thought to play key roles in value-based decision making. Dopamine has been shown to encode expected value (Schultz, 1998;Tobler et al., 2005), modulate processing of reward magnitude and probability (Burke et al., 2018) . Thus, in this study we additionally investigated whether we could replicate previous findings from studies using pharmacological manipulations of dopamine and noradrenaline in value-based decision making. In addition, we explored whether noradrenaline or dopamine have distinct influences on the confidence in value-based choice.
To assess the noradrenergic and dopaminergic influence on value-based decision making and confidence, we used two pharmacological manipulations in a task that was similar to a previously used task (De Martino et al., 2013). We selected drugs that are known to have a selective and high affinity for either dopamine or noradrenaline, and which have been used in studies using value-based decision making tasks (Burke et al., 2018;Jocham et al., 2011;Sokol-Hessner et al., 2015). Consequently, we used the β-adrenoceptor antagonist propranolol (40mg) for noradrenaline manipulation and the D2/D3 receptor antagonist amisulpride (400mg) for impacting dopamine functioning. In this double-blind, placebo-controlled study design (61 healthy subjects in total), we failed to find any reliable influence of amisulpride or propranolol on value-based decision making. Similarly, we did not find any association between any of the catecholamines and confidence in our task. We make all data openly available for further (meta-)analyses.

Methods
The study was conducted at University College London from January 2019 to February 2019.

Participants
We recruited 61 healthy participants aged 18-35 years (mean = 23.28, standard deviation [s.d.] = 3.624) for a doubleblind, placebo-controlled, between-subjects study. The study was advertised on local subject pools, where participants could sign up on a voluntary basis. Each participant was randomly allocated to one of three groups (using randperm function in MATLAB), assuring equal gender balance in each group. The randomisation was managed by a team member who did not undertake the testing of participants. The placebo group had 21 participants, whereas the propranolol and amisulpride groups consisted of 20 participants each. The extra male participant in the placebo group was recruited due to a technical error in a different (main) task of the experiment. The sample size was selected based on previous studies that used the same pharmacological manipulation ( Volunteers with a history of neurological or psychiatric disorders, current health issues, regular medications (except contraceptives), or prior allergic reactions to drugs were excluded from the study. Participants were reimbursed for their participation on an hourly basis and to ensure motivation in this task, one trial was randomly selected at the end of the task and was given as a bonus payment. On average, subjects received £10 per hour.

Ethical considerations
The study was approved by the Research Ethics Committee of University College London (study number: 14261/001) and all participants provided written informed consent.

Procedure
Sixty-one healthy volunteers were randomly assigned to one of three groups: propranolol, amisulpride and placebo. The propranolol group received 40mg of propranolol, a non-selective b-adrenoceptor antagonist, while the amisulpride group was administered 400mg amisulpride, a selective D2/D3 antagonist. The placebo group did not receive any active drug. Propranolol and amisulpride have different pharmacokinetic properties and thus were administered at different times to ensure the peak effect at similar times ( Figure 1A

Task
The risky gambling task that participants played was set up on a computer. Participants were asked to repeatedly choose between two risky options with varying probabilities and reward magnitudes without any time limit ( Figure 1B). Participants were asked to press the left arrow key to choose the option on the left and the right arrow key to choose the option on the right. The ranges of stimuli were adapted from a previously used task (Hunt et al., 2013), and they were drawn such that reward magnitude and probability were never identical across the two options. The reward magnitude ranged from £1 to £9 and the probability ranged from 10% to 90%. Therefore, participants had to integrate across reward magnitude and probability to make optimal choices. Participants were presented with each combination of options (n = 54) twice, counterbalanced across left-right spatial configurations. After every choice, participants were asked to indicate their confidence level (i.e. "How confident are you that the choice you made was the right one for you?") on a continuous scale between 1 (low confidence) and 10 (high confidence). Participants were given 6,000 ms to choose their level of confidence in their decision. Every participant completed 108 trials divided in 4 blocks. They also completed 10 practice trials before starting the task to ensure their understanding of the task. It took participants up to 10 minutes to complete the task.

Statistical analyses
We compared behavioural measures using one-way ANOVAs with drug group as between-subject factor (placebo, propranolol, amisulpride). We further compared the measures using ANCOVAs with drug group as between-subject factor and controlling for intellectual abilities (WASI score) and positive and negative affect (PANAS scores). Significant effects were further explored using (independent samples) t-tests. We report effect sizes using partial eta squared (η²) for ANOVAs and Cohen's d (d) for t-tests. The analyses were conducted in MATLAB (R2020a) (open alternative: GNU Octave) and SPSS software (version 26, IBM).
To support the null-hypotheses further, we conducted Bayesian ANOVAs and Bayesian ANCOVAs using JASP software (Version 0.14.1). We use Bayes Factor 01 (BF01) to report evidence in favour of the null hypothesis.

Procedure
To examine the role of catecholamines in value-based choice and confidence, we ran a double-blind, placebo-controlled, between-subjects study. We excluded four participants from the main analysis. Two participants were excluded due to poor Figure 1. Study design and task procedure. (A) Before completing the task, participants were administered drugs: either amisulpride which blocks dopaminergic D2/D3 receptors, propranolol that blocks noradrenergic β-receptors or inert substances for placebo group. Different administration times were chosen to ensure the peak effect at similar times as the drugs have different pharmacokinetic properties. (B) Participants were presented with two choice options with varying reward and probability, and had to choose one. After each choice, they were asked their level of confidence for choosing the best option. Participants did not see the outcome of the gamble, but they were instructed that one of the gambles will be chosen randomly to win a bonus in the end of the task. performance as seen in the low proportions of choosing the option with higher reward magnitude and probability in 'no-brainer' trials (67.5% and 55%, while the mean of the other subjects was at 98.1%, s.d. = 3.169; the excluded participants were from amisulpride and placebo group respectively, both female). Additional two participants were excluded (1 male from placebo group, 1 female from propranolol group) due to perfect choice patterns that prevented reliable estimation of a logistic fit, leaving nineteen participants to each group (c.f. Table 1). Because our final drug groups differed marginally in IQ and positive and negative mood (cf. methods for details), we decided to report all findings with and without controlling for these covariates.
All volunteers played a version of a monetary risky gambling task ( Figure 1B). They were asked to make repeated choices between two choice options with varying probabilities and reward magnitudes (e.g., 60% of chance of receiving £3 vs 25% chance of receiving £9). The ranges of stimuli were adapted from a previously used task (Hunt et al., 2013). From here on, we refer to the monetary amount as 'reward magnitude', the probability of receiving the monetary amount as 'probability', and to the multiple of the amount of money and probability as 'expected value' (EV). After every choice, participants reported their confidence from a scale of 1 to 10, which reflects the degree of subjective certainty that they had in making the best choice for themselves (De Martino et al., 2013). Subjects never saw the outcome of their choice.

Similar performance between groups
Before investigating the role of noradrenaline and dopamine in value-based decisions and confidence, we were interested in assessing whether the groups performed similarly in the simplest performance metrics of the task. Firstly, we examined 'no-brainer' trials, in which both reward magnitude and probability of one choice option were higher than that of the other, thus suggesting that one should pick the option with higher reward magnitude and probability. There were no differences between groups in these 'no-brainer' trials (F(2,54) = 1.31, p = 0.279, η 2 = 0.046; controlling for IQ and positive and negative mood: F(2,51) = 2.00, p = 0.146, η 2 = 0.073). Reassuringly, there was a significant effect of IQ on no-brainer trials (F(1,51) = 7.30, p = 0.009, η 2 = 0.125).
Next, we looked at the proportion of trials where individuals chose the optimal option, i.e. the option with higher EV (all participants: mean = 57.6%, s.d. = 2.854). We did not find any group differences (F(2,54) = 0.36, p = 0.700, η 2 = 0.013; controlling for covariates: F(2,51) = 0.30, p = 0.744, η 2 = 0.012). Using Bayesian statistics, we found moderate evidence that the null model (i.e. no difference) is 5 times more likely (BF01 = 5.51) than a model that distinguishes between groups, suggesting that groups did not behave differently in choosing optimal options. When further controlling for covariates, we found strong evidence for no difference between groups (BF01 = 99.5).
How do reward magnitude, probability and EV predict choice?
In gambles like these, the optimal strategy would be to calculate the expected value (EV) of each option and chose the one with the higher EV. However, humans most often act sub-optimally, with a substantial proportion of subjects being more strongly influenced by either probability or magnitude, rather than their product.
To evaluate the influence that reward magnitude, probability and EV had on participants' choices, we thus followed a similar analysis to Hunt and colleagues (Hunt et al., 2013). We performed a logistic regression to predict choosing the option on the right side at every trial using reward magnitude, probability Table 1. The drug groups did not differ in gender, nor in age. The final groups differed marginally in intellectual abilities (adapted WASI matrix test), and positive and negative affect score from PANAS questionnaire, which is why we chose to control for these covariates and report effects with and without controlling for them.
No drug effects on how reward magnitude, probability and EV affect choice To examine whether catecholamines affect how the reward magnitude, probability and EV influence one's choice, we compared the parameter estimates from the logistic regression between groups ( Figure 2B) We found no evidence that the null model is more likely than a model that distinguishes between groups for both magnitude and probability (reward magnitude: BF01 = 1.01; probability: BF01 = 1.06; EV: BF01 = 2.21), however when further controlling for the covariates we found a moderate evidence for the support of null model (reward magnitude: BF01 = 13.65; probability: BF01 = 5.13; EV: BF01 = 8.32). These results suggest that in our sample, the catecholamines had little to no role in how participants used reward magnitude, probability, and EV to guide one's choice.

Proportion of chosen uncertain options is similar between groups
Noradrenaline has been suggested to be linked to uncertainty processing (Yu & Dayan, 2005), thus we investigated whether there were any differences in how groups chose options with higher uncertainty. To measure uncertainty preference, we investigated the proportion of choices in which participants chose the option that had a probability that was closer to 50% (excluding trials with the same distance from 50%). This metric captures how willing you are to make a choice that is uncertain, as 50% is the most uncertain of all options.

Expected value and confidence
Next, we were interested in how the confidence reports relate to the relevant task variables. Optimally, confidence should reflect the (absolute) EV difference between the options, because this signifies the difficulty of a choice. However, absolute EV difference accounted on average only for 10.4% of the confidence variance (mean correlation between confidence and EV difference: r = 0.32, s.d = 0.170), suggesting that confidence (similarly to value-based choice above) does not solely rely on the objectively optimal criterion, EV.
To further evaluate other factors that affect confidence we fit a linear regression predicting the level of confidence using the EV difference and the chosen EV as predictors. The chosen EV is important because it reflects the overall goodness of the choice as it takes into account the total monetary amount that could be won and the probability of the chosen option. We found that both predicted the confidence level (EV difference: mean = 0.32, s.d. Next, we compared both predictors between groups to explore the drug effects on the confidence level predictors. We did not find any differences between groups in how the chosen EV influenced confidence (F(2,53) = 0.16, p = 0.856, η 2 = 0.006; controlling for covariates: F(2,50) = 0.32, p = 0.727, η 2 = 0.013).
There was moderate evidence that the null model is more likely compared to model that distinguishes between groups (BF01 = 3.18), and strong evidence for null model when further controlling for covariates (BF01 = 16.92).
Choice and confidence consistency In this task, each item pairing was repeated twice (once in each spatial configuration), and therefore it was possible to examine choice consistency by investigating the chosen option at both times when the same item pairing was presented. On average only 13.2% (s.d. = 8.26%) of choices were reversed on the second encounter, showing that participants were consistent in their choices. To explore whether there were any drug effects on choice consistency in the context of this task, we compared the consistency between groups. We did not find any differences between groups ( These findings thus suggest that confidence reports were meaningfully linked to task relevant variables and subsequent behaviour, but we did not observe any drug-induced changes in the utilisation of confidence.

Discussion
Making rational choices and being able to consciously reflect on the goodness of these choices is important for successfully navigating the world. Previously, it has been proposed that dopamine and noradrenaline play a role in how reward magnitude and probability are perceived and integrated (Burke et al., 2018;Jocham et al., 2011;Rogers et al., 2004). In our study, the two pharmacological manipulations we used to target noradrenaline (β-adrenoceptor antagonist propranolol) and dopamine (D2/D3 receptor antagonist amisulpride) did not have a reliably significant effect on value-based decision making. Furthermore, we did not find any influence of noradrenaline or dopamine manipulation in the subjective confidence in one's choice.
We did not find any differences between groups in how participants chose more optimal options (i.e. options with higher expected value), thus not replicating findings by Burke and colleagues (Burke et al., 2018). Their findings suggested that participants who received the same dose of amisulpride (400mg) chose more expected value options (Burke et al., 2018) than those who received a placebo. This may be due to their study's substantially bigger sample size (placebo group (n = 48) and amisulpride group (n = 45)) that allowed to detect smaller effect sizes. A further difference is that the task used by Burke and colleagues only had 20 trials per person, using an adaptive task structure that adjusted each trial according to participants' risk attitude. The task additionally included potential losses as choice options. Thus, the differences in the task structures or our smaller sample size may have accounted for the lack of replication of their findings.
Our findings suggest that participants use information about the reward magnitude and probability to guide their choice, replicating findings by Hunt and colleagues (Hunt et al., 2013) in a similar task. We did not find any drug effects on how reward magnitude and probability are integrated to the choice. Previously, pharmacological dopamine manipulations have found that dopamine plays a role in value-based decision making, including modulation of how reward magnitude and probability are processed after amisulpride administration (Burke et al., 2018). The lack of findings here may be due to the difference in the task structures or lower sample size compared to the study by Burke and colleagues (Burke et al., 2018) as mentioned above.
We observed that confidence is influenced by the overall goodness of the choice (i.e. potential overall outcome), as well as the difference in expected value between options (i.e. choice difficulty). Furthermore, we found that choices that had lower confidence were associated with changes of mind suggesting that confidence is useful in trying a different option when the previous choice did not have the expected consequence. We did not find any influence of amisulpride nor propranolol in the construction of confidence ratings, suggesting that dopamine and noradrenaline did not play a role in confidence in this task context. In conclusion, we did not find sufficient evidence that amisulpride and propranolol affected value-based decision making or the confidence in one's choice in the context of this study. We report these results here together with the release of all relevant data and code. This will allow researchers to further interrogate the data, to counteract publication biases in favour of significant findings, and to use our study as a source for balanced meta-analyses. Did the authors somehow account for individual catecholamines baseline levels? If not, this should be discussed.

8.
Because the task was distributed in varying delay (140 -160 mins) and had an unrestricted response time, the drug effect might be also varying because of the duration from the drug administration to the time performing the task (even maybe in each trial). The reaction times are not reported so it is not clear if this might affect and if the model should include that factor.

9.
We need to stress again, that an introduction and/or discussion of how these drugs could have impacted or have impacted in previous work the neurobiology of decision making is completely missing from the manuscript and would be an important piece for future researchers to find the relevant publication to derive useful hypotheses to test with appropriate paradigms. 10.

Discussion:
Indicating that the task design was different between the current study and previous ones is important, however, it is not entirely clear how the authors link it to absence of an effect. It would be desirable, if the authors could elaborate on why and how having a dynamic task with included losses can affect the expression of drug effects.

1.
The fourth paragraph of the discussion (about confidence) is simply a reframing of the Results section. The authors didn't provide any reason to support their results and didn't link their findings to previous research.

2.
As it is written, it does not bring much novelty except detailed description of null results. To enrich the manuscript, the authors should carefully rewrite the discussion section and add much more neurophysiological detail. The results here need to be discussed in light of previous findings, potential hypothesis about null findings should be elaborated. The literature on dopamine and norepinephrine is extremely rich and full of null, confirmatory or contradicting findings, which makes it actually easy to relate the current findings to previous ones.

3.
"Suggesting that dopamine and noradrenaline did not play a role in confidence in this task context". The authors obviously tried to make this statement as attenuated and cautious as possible, but it remains bold. To rule out the role of dopamine, one would need not only act on both striatum and prefrontal cortex but also have another group, where DA function would be enhanced. This is not the case in the present study, suggestion about dopamine (at least) not playing a role seems to be bold and needs rewriting.

4.
Discussing the role of norepinephrine in processing uncertainty is not entirely appropriate in the context of the present task. Uncertainty is rather an umbrella term, and it would be desirable, if the authors could incorporate a discussion about the differences of noradrenergic effects on its components, namely risk and ambiguity, from previous studies. This point is important because the current task only contains a risk component because 5.
the odds of all options are well defined.
The authors suggest that the drugs effect could be explained (in part) by the difference in negative mood. Is it possible that the drugs induce that very same effect even before the tasks, when participants take the survey?