A selective effect of dopamine on information-seeking.

Humans are motivated to seek information from their environment. How the brain motivates this behavior is unknown. One speculation is that the brain employs neuromodulatory systems implicated in primary reward-seeking, in particular dopamine, to instruct information-seeking. However, there has been no causal test for the role of dopamine in information-seeking. Here, we show that administration of a drug that enhances dopamine function (dihydroxy-L-phenylalanine; L-DOPA) reduces the impact of valence on information-seeking. Specifically, while participants under Placebo sought more information about potential gains than losses, under L-DOPA this difference was not observed. The results provide new insight into the neurobiology of information-seeking and generates the prediction that abnormal dopaminergic function (such as in Parkinson's disease) will result in valence-dependent changes to information-seeking.

RCT Entities:   Population Interventions Outcomes

Introduction

Curiosity, commonly defined as the desire for knowledge, is a fundamental part of human nature (Kidd and Hayden, 2015; Loewenstein, 1994). In humans, it manifests as information-seeking behaviors such as asking questions, reading, conducting experiments, and online searches. Such behavior is integral to learning, social engagement, and decision-making (Kidd and Hayden, 2015; Loewenstein, 1994; Sakaki et al., 2018). Despite information-seeking being central to behavior, we know remarkably little about the biological mechanisms that control it.

It has been suggested that information-seeking relies on the same neural system as reward-seeking (Bromberg-Martin and Hikosaka, 2009; Bromberg-Martin and Hikosaka, 2011; Blanchard et al., 2015; Charpentier et al., 2018; Ligneul et al., 2018; Kobayashi and Hsu, 2019; Kang et al., 2009; Smith et al., 2016; Tricomi and Fiez, 2012; Jessup and O'Doherty, 2014; Gruber et al., 2014; van Lieshout et al., 2018), implying that the opportunity to gain knowledge has intrinsic value (Grant et al., 1998). This assumption is supported by correlational studies showing that the opportunity to gain information is encoded in regions rich in dopaminergic neurons (e.g. Ventral Tegmental Area, Substantia Nigra) and their targets (e.g. Nucleus Accumbens, Orbital Frontal Cortex) (Bromberg-Martin and Hikosaka, 2009; Bromberg-Martin and Hikosaka, 2011; Blanchard et al., 2015; Charpentier et al., 2018; Ligneul et al., 2018; Kobayashi and Hsu, 2019; Kang et al., 2009; Smith et al., 2016; Tricomi and Fiez, 2012; Jessup and O'Doherty, 2014; Gruber et al., 2014; van Lieshout et al., 2018). For example, information prediction error signals have been identified in dopamine-rich brain regions (Bromberg-Martin and Hikosaka, 2009), which analogous to reward prediction errors (Schultz et al., 1997) are theorized to provide reinforcement for seeking-information. These signals have been observed even when information is non-instrumental (Bromberg-Martin and Hikosaka, 2009) (i.e. cannot be used to gain future rewards or avoid future harm), consistent with the idea that the brain treats the opportunity to gain knowledge as a higher order reward (Bromberg-Martin and Hikosaka, 2009; Bromberg-Martin and Hikosaka, 2011; Blanchard et al., 2015; Grant et al., 1998). Such coding may be adaptive because information could turn out to be useful in the future even if it appears useless at present (Eliaz and Schotter, 2007).

Thus, one hypothesis is that dopamine boosts information-seeking. However, another possibility is that dopamine selectively affects the impact of valence on information-seeking. In particular, it has been shown that individuals seek information more when information is about future gains than losses (Charpentier et al., 2018; Thornton, 2008; Persoskie et al., 2014; Dwyer et al., 2015; Caplin and Leahy, 2001; Kőszegi, 2010; Golman et al., 2017). For example, investors monitor their portfolio more frequently when they expect their worth has gone up rather than down (Karlsson et al., 2009); some people refuse to receive results of medical tests for fear of bad news (Hertwig and Engel, 2016); and monkeys prefer to know in advance the size of rewards they are about to receive particularly when they expect large rewards (Bromberg-Martin and Hikosaka, 2009; Bromberg-Martin and Hikosaka, 2011; Blanchard et al., 2015). In humans, dopaminergic midbrain regions have been shown to code for the opportunity to receive information in a valence-dependent manner (Charpentier et al., 2018), suggesting that the intrinsic utility of knowledge is modulated by valence.

To test the above competing hypotheses, we enhanced dopamine function in humans by administrating L-DOPA and asked them to perform an information-seeking task (Charpentier et al., 2018). We compared their performance to participants who received Placebo to examine whether and how dopamine alters non-instrumental information-seeking.

Results

Two hundred and forty-eight participants performed an information-seeking task adapted from our previous publication (Charpentier et al., 2018), in which 16 participants did not complete the task in full; therefore, data of 232 subjects was analyzed. The study was a double-blind pharmacological intervention where one group of participants received Placebo (n = 116, females = 72, mean age = 24.36, Table 1) and the other received L-DOPA (150 mg) (n = 116, females = 71, mean age = 25.44, Table 1).

Table 1.

Demographics.

Demographics	Placebo mean (SD)	L-DOPA mean (SD)	p-Value
Age (years)	24.36 (7.91)	25.44 (7.92)	0.301
Gender	Females N= 72	Females N= 71	0.893
Income (1-9)	4.85 (2.38)	4.61 (2.54)	0.462
Education Level (1-10)	7.09 (1.72)	7.39 (1.50)	0.157

There were no differences between groups in terms of demographics. p-Value is of independent sample t-test , or in the case of gender of X2. Education was measured on a scale ranging from 1 (no formal education) to 10 (Doctoral degree ). Annual household income was measured on a scale from 1 (less than 10K) to 10 (more than 100K).

Participants began the task 40 min after receiving L-DOPA or Placebo (as in Guitart-Masip et al., 2012; Sharot et al., 2009; Sharot et al., 2012), as the half-life of L-DOPA is 90 min. They were endowed with £5 at the beginning of each of the four blocks to invest in two of five stocks in a simulated stock market. There were 50 trials per block. On each trial, participants observed the evolution of the market (i.e. whether the market was going up or down) and the exact value of the market (Figure 1). They then bid for a chance to know (or remain ignorant about) the value of their portfolio. Specifically, they indicated how much they were willing to pay to receive or avoid information about the value of their portfolio on a scale ranging from 99 p to gain knowledge through 0 p (no preference) to 99 p to remain ignorant. The more they were willing to pay, the more likely their choice was to be honored. Information was non-instrumental; it could not be used to increase rewards, avoid losses, or make changes to portfolio.

Figure 1.

Stock market task.

(A) Participants observed the evolution of a financial market after investing in two of its five companies. They then indicated whether they believed their portfolio value likely went up or down relative to the previous trial and indicated their confidence in their answer. They then indicated how much they were willing to pay to receive or avoid information about their portfolio value. Next, their portfolio value in points was presented on screen or hidden (‘XX points’ was shown).

L-DOPA did not alter general information-seeking

L-DOPA administration did not alter general aspects of information-seeking (Figure 2). In particular, there were no difference between the Placebo and L-DOPA groups in the average number of trials in which participants selected to pay for information (Placebo = 71.16 trials, L-DOPA = 72.89 trials, t(230) = 0.226, p=0.821, independent samples t-test ), pay to avoid information (Placebo = 27.79 trials, L-DOPA = 27.97 trials, t(230) = 0.036, p=0.971), or not to pay at all (i.e. entered 0 p: Placebo = 93.86 trials, L-DOPA = 92.65 trials, t(230) = 0.145 p=0.885). There was also no difference in the average amount each group paid to receive information (Placebo = 18.18 p, L-DOPA = 15.68 p, t(228) = 0.928, p=0.355) or avoid it (Placebo = 11.34 p, L-DOPA = 9.95 p, t(223) = 0.587, p=0.558). These results suggest that dopamine does not generally alter information-seeking. Finally, there was no difference across groups in the number of trials participants missed (that is trials in which they were too slow in responding: Placebo = 7.03 trials, L-DOPA = 6.50 trials, t(230) = 0.283, p=0.777), suggesting no difference in engagement with the task.

Figure 2.

L-DOPA does not alter general aspects of information-seeking.

There were no differences in general information-seeking between those who received Placebo and those who were administered L-DOPA. In particular, there were no differences in the average number of trials on which the participants decided to receive or to avoid information or were indifferent (i.e. paid 0). Furthermore, there was no difference across groups in the number of trials participants missed (that is trials in which they were too slow in responding). Error bars SEM.

L-DOPA diminished the effect of valence on information-seeking

In this task, we had previously shown that despite participants wanting information both when the market was going down and when it was going up (Charpentier et al., 2018), information-seeking was modulated by the expected valence of the outcome (Charpentier et al., 2018). In particular, we had reported that participants were more likely to pay for information when the market was going up rather than down and more likely to pay to avoid information when the market was going down rather than up (Charpentier et al., 2018). This is because people expected to learn about gains when the market was going up and expected to learn about losses when the market was going down (Charpentier et al., 2018). The second factor we had reported to influence information-seeking was the absolute amount of change in the market. Participants were willing to pay more for information when there were big changes in the market. Here, we examine whether dopamine modulates these effects on information-seeking.

On each trial, we calculated the Willingness To Pay (WTP) for information. WTP is coded positively if participants indicated they wanted to receive information and negatively if they wanted to avoid information (Charpentier et al., 2018). We then ran a Linear Mixed Model to predict WTP on each trial from the two factors we had previously shown to impact information-seeking in this task (Charpentier et al., 2018): (i) valence (quantified as signed market change, which is the amount by which the market went up or down); (ii) absolute market change; as well as from (iii) group (L-DOPA or Placebo). All three factors were included as fixed and random effects, as were the interactions of group with each of the other two factors. Random and fixed intercepts were also included in the model.

The results revealed an interaction between group and valence on the WTP for information (β = −0.15, CI = −0.29 /- 0.01, t(230.52) = 2.15, p=0.032) as well as a main effect of valence (β = 0.20, CI = 0.11/0.30, t(229.60) = 4.11, p=0.0001) and a main effect of absolute market change (β = 0.41, CI = 0.25/0.58, t(231.35) = 4.87, p=0.0001). There was no interaction between group and absolute market change (β = 0.07, CI = −0.17/0.30, t(232.42) = 0.576, p=0.565) nor a main effect of group (β = −2.61, CI = −7.50/2.28, t(232.53) = 1.045, p=0.297).

The interaction indicates that expected valence differentially effected the desire for information in the Placebo and L-DOPA groups. To tease apart the interaction, we next ran two mixed linear models separately for the Placebo and L-DOPA groups. WTP was entered as the dependent factor and valence and absolute market change as fixed and random factors. The model included fixed and random intercepts. This revealed a significant effect of valence in the Placebo group (main effect of signed market change: β = 0.20, CI = 0.08/0.33, t(115.25) = 3.18, p=0.001, Figure 3a), but lack thereof in the L-DOPA group (main effect of signed market change: β = 0.05, CI = −0.005/0.11, t(115.28) = 1.78, p=0.076, Figure 3a). Both groups showed a main effect of absolute market change (Placebo: β = 0.41, CI = 0.23/0.59, t(117.58) = 4.52, p=0.0001; L-DOPA: β = 0.48, CI = 0.33/0.63, t(113.54) = 6.266, p=0.0001, Figure 3a). These results suggest that L-DOPA selectively reduced the impact of the expected valence of information on the desire for knowledge.

Figure 3.

L-DOPA reduces the effect of valence on information-seeking.

(a) A mixed linear regression predicting Willingness To Pay (WTP) for information revealed an interaction between group (Placebo/L-DOPA) and valence (the amount by which the market went up or down), with no interaction between group and absolute market change. To tease apart the interaction, we ran linear mixed models separately for the L-DOPA and Placebo groups. Plotted are the fixed effects of those models. As observed, this revealed a significant effect of valence on information-seeking in the Placebo group but lack thereof in the L-DOPA group. Absolute change was a significant predictor in both groups. This indicates a reduction in the influence of valence on information-seeking under L-DOPA. (b) To further characterize the effect of valence and drug on information-seeking, we run separate mixed linear models for each group and polarity predicting WTP from market change, trial number and the interaction of the two. Plotted are the fixed effects of market change for each. As can be observed under L-DOPA market change was a significant predictor of information-seeking about potential losses and gains - the greater the expected gain/loss the more participants were willing to pay for information. In contrast, under Placebo market change was a significant predictor of information-seeking about potential gains, but not losses. These results show that L-DOPA selectively alters information-seeking about losses. (c) Plotted is the effect of market change on WTP for information controlling for any effects of trial number. As can be observed the slopes are significantly positive for all groups/conditions except for the Placebo group in the loss domain. Clouds are based on Standard Errors of the fixed effect. Error bars SEM, * p <0.05, ** p < 0.01, *** p < 0.001.

The same results are observed also when using a simpler model with WTP as a dependent measure and only one independent factor - valence - coded in a binary fashion (1 for market up and 0 for market down) as fixed and random variable with fixed and random intercepts. We find a significant effect of valence in the Placebo group (β = 1.85, CI = 0.64/3.05, t (116.88)=3.01, p=0.003) with WTP for information being greater for trials in which the market went up (indicating potential gains) than down (indicating potential losses), and lack thereof in the L-DOPA group (β = 0.35, CI = −0.21/0.91, t(117.58),=1.22 p=0.224). This shows that under Placebo participants desired information more when the market was up vs down, whereas under L-DOPA the desire for information was not altered by valence.

L-DOPA selectively alters information-seeking about potential losses

Our results indicate that L-DOPA selectively reduces the impact of valence on information-seeking. Next, we ask if this effect is due to L-DOPA altering information-seeking about potential losses, about potential gains, or both. Moreover, we ask whether the effect of L-DOPA emerged over the course of the experiment or whether it was apparent from the very beginning.

To that end, we ran two separate mixed effect linear model predicting the WTP for information – one for trials in which the market went up (potential gain trials) and one for which the market went down (potential loss trials). The independent factors included (i) market change, (ii) trial number, and (iii) group (L-DOPA/Placebo). As each model now includes only one polarity - either market going up or down - signed market change and absolute market change are perfectly correlated. Thus, only one factor ‘market change’ is added. In the loss domain, the greater the ‘market change’ the greater the expected losses. In the gain domain, the greater the ‘market change’ the greater the expected gains. All three factors and their interactions were included as fixed and random effects. Random and fixed intercepts were also included in the model. The results revealed an interaction between market change and group in the loss domain (β = 0.40, CI = 0.07/0.74, t (751.900)=2.35, p=0.018), but not in the gain domain (β = 0.10, CI = −0.29/0.50,t (490.600)=0.505, p=0.614) where instead there was a main effect of market change (β = 0.70, CI = 0.42/.98, t (495.800)=4.896, p=0.0001). No other effects were significant.

To characterize the interaction of interest (between market change and group) in the loss domain and lack thereof in the gain domain, we ran four linear mixed models - one for each group and valence polarity. WTP was the dependent factor, and the independent factors were (i) market change and (ii) trial number. Both factors and their interactions were included as fixed and random effects. Random and fixed intercepts were also included in the model. This revealed that under L-DOPA participants were willing to pay more for information the greater the gains (effect of market change: β = 0.79, CI = 0.53/1.05, t (313.300)=5.98, p = 0.0001 Figure 3b) and the greater the losses (effect of market change: β = 0.53, CI = 0.33/0.74, t (484.700)=5.033, p=0.0001, Figure 3b). In contrast, under Placebo participants were willing to pay more for information the greater the gains (effect of market change: β = 0.69, CI = 0.40/0.99, t (219.92)=4.64, p=0.0001, Figure 3b) but did not show this effect for losses (effect of market change: β = 0.12, CI = −0.13/0.39, t (335.800)=0.956, p=0.33, Figure 3b). For L-DOPA in the gain domain, there was an additional interaction between trial and market change (β = −0.002, CI = −0.004 /- 0.006), t (382.20) = 2.74, p=0.006). No other effects were significant.

The results show that under L-DOPA, participants’ desire for information increased as the expected magnitude of the outcome increased - participants were willing to pay more for information as potential gains and losses increased (Figure 3C). In contrast, under Placebo, participants’ desire for information increased as potentials gains increased but remained constant and relatively low for potential losses (Figure 3C).

The effect of L-DOPA on information-seeking for losses is not explained by changes in expectations

We next ask whether the selective effect of L-DOPA on information-seeking about losses can be explained by a selective effect of L-DOPA on expectations about losses. To test participants’ expectations regarding their outcomes, we asked participants whether they believed their stocks went up or down after observing the global market change. This was done by having participants rate their expectations on a scale ranging from 1 (decreased a lot) to 9 (increased a lot). We then entered these ratings into Linear Mixed Model predicting expectation ratings. The independent factors were: (i) valence (signed market change), (ii) absolute market change, and (iii) group (L-DOPA or Placebo). All three factors were included as fixed and random effects as were the interactions of group with each of the other two factors. Random and fixed intercepts were also included in the model. There was no main effect of group (β = 0.06, CI = −0.05/0.17, t (257.400)=1.01, p=0.310), nor an interaction between group and valence (β = −0.01, CI = −0.03/0.02, t (232.900)=0.63, p=0.529) nor an interaction between group and absolute market change (β = 0.00, CI = −0.01/0.01, t (242.900)=0.266, p=0.790). There was a main effect of absolute market change (β = −0.01, CI = −0.02 /- 0.0001, t (241.000)=2.28, p=0.023) and of valence (β = 0.21, CI = 0.19/0.22, t (233.000)=23.264, p=0.0001). The latter confirms that participants’ expectations about their outcomes were linked to the observed trends in the market.

These results suggest that L-DOPA did not affect participants’ expectations. To further examine whether there may be an effect of L-DOPA on expectations that altered over time, we added to the model above trial number as a fixed and random factor as well as all the two- and three-way interactions of trial number with the other factors. Once again, this neither revealed an effect of group on expectations (β = 0.05, CI = −0.49/0.60, t(312.800) = 0.186 p=0.852) nor were any of the interactions between group and any of the other factors significant (all Ps > 0.329). These results suggest that L-DOPA selectively altered the effect of valence on information-seeking without altering outcome expectations.

Discussion

Humans and non-human animals seek information even when information cannot be used to alter outcomes (Bromberg-Martin and Hikosaka, 2009; Bromberg-Martin and Hikosaka, 2011; Blanchard et al., 2015; Charpentier et al., 2018). This observation led to the notion that knowledge may have evolved to carry intrinsic value (Bromberg-Martin and Hikosaka, 2009; Bromberg-Martin and Hikosaka, 2011; Blanchard et al., 2015; Grant et al., 1998). Indeed, it has been shown that the opportunity to receive non-instrumental information is encoded by the same neural system as for primary rewards (Bromberg-Martin and Hikosaka, 2009; Bromberg-Martin and Hikosaka, 2011; Blanchard et al., 2015; Charpentier et al., 2018; Ligneul et al., 2018; Kang et al., 2009; Gruber et al., 2014; van Lieshout et al., 2018). As this system includes regions rich in dopamine (the findings triggered the hypothesis that dopamine plays a critical role in non-instrumental information-seeking (Bromberg-Martin and Hikosaka, 2009; Bromberg-Martin and Hikosaka, 2011; Blanchard et al., 2015; Charpentier et al., 2018). By manipulating the dopamine levels in humans, we were able to directly test this hypothesis.

Our results show that L-DOPA has a selective effect on non-instrumental information-seeking. Administration of L-DOPA dampened the effect of valence on non-instrumental information-seeking, altering non-instrumental information-seeking about potential losses without impacting non-instrumental information-seeking about potential gains. Specifically, while participants under Placebo sought information more about potential gains than losses (an effect observed in the past [Charpentier et al., 2018]), under L-DOPA this difference was not observed. Moreover, under L-DOPA, participants’ WTP for information increased as potential gains and losses increased. In stark contrast, under Placebo, participants’ WTP for information increased as potential gains increased but remained constant and relatively low as potential losses increased.

An intriguing question concerns the mechanism by which L-DOPA alters information-seeking about potential losses. The effect could not be explained by changes to participants’ mood, as there were no differences in participants’ self-reported subjective state under Placebo and L-DOPA (see Table 2). Neither could it be explained by reduced attention and/or engagement, as participants under L-DOPA did not miss more trials than those under Placebo. L-DOPA also did not alter expectations of outcomes. Thus, modulation of outcome expectations (that is how much is expected to be lost/gained) cannot explain the results. Moreover, as the task did not involve learning (past outcomes had no impact on future outcomes, see supplementary results), L-DOPA did not affect learning about potentials outcome gains and losses.

Table 2.

Subjective State Questionnaire.

Subjective State Questionnaire (Joint Formulary Committee, 2009) revealed no differences in subjective state between groups. p-Value relates to independent sample t-test.

Subjective State Questionnaire	Before the task			After the task
	Placebo mean (SD)	L-DOPA mean (SD)	p-Value	Placebo mean (SD)	L-DOPA mean (SD)	p-Value
Alert to drowsy	2.68 (1.19)	2.62 (1.08)	0.687	3.60 (1.41)	3.85 (1.57)	0.208
Calm to excited	2.33 (1.11)	2.29 (1.03)	0.808	2.34 (1.09)	2.27 (1.22)	0.632
Strong to feeble	2.68 (1.01)	2.63 (1.01)	0.699	2.97 (1.13)	3.15 (1.35)	0.264
Muzzy to clear headed	4.47 (1.26)	4.48 (1.11)	0.956	3.70 (1.23)	3.41 (1.39)	0.099
Coordinated to clumsy	2.28 (1.14)	2.22 (1.06)	0.722	2.80 (1.16)	3.02 (1.31)	0.187
Lethargic to energetic	3.89 (1.14)	3.94 (1.18)	0.736	3.20 (1.24)	3.00 (1.43)	0.263
Contented to discontented	2.18 (1.01)	2.12 (0.83)	0.620	2.58 (1.16)	2.65 (1.19)	0.644
Troubled to tranquil	4.83 (1.02)	4.66 (1.02)	0.201	4.52 (1.13)	4.55 (1.13)	0.858
Slow to quick witted	4.32 (1.11)	4.28 (1.04)	0.761	3.63 (1.30)	3.31 (1.30)	0.069
Tense to relaxed	4.64 (1.11)	4.67 (0.98)	0.803	4.47 (1.14)	4.42 (1.23)	0.733
Attentive to dreamy	2.78 (1.26)	2.73 (1.10)	0.740	3.47 (1.34)	3.42 (1.38)	0.804
Incompetent to proficient	4.56 (0.98)	4.70 (0.86)	0.260	4.14 (1.16)	4.02 (1.26)	0.450
Happy to sad	2.43 (1.06)	2.34 (0.84)	0.453	2.62 (1.12)	2.57 (0.95)	0.712
Antagonistic to friendly	5.08 (0.97)	5.07 (0.81)	0.942	4.65 (0.95)	4.60 (1.04)	0.703
Interested to bored	2.35 (1.21)	2.28 (1.02)	0.639	3.52 (1.45)	3.63 (1.50)	0.587
Withdrawn to sociable	4.34 (1.17)	4.36 (1.18)	0.868	3.90 (1.17)	3.83 (1.39)	0.672

One possibility is that L-DOPA altered expectations not about outcomes per-se, but about the affective impact of negative information. A negative cue (e.g. watching the financial market fall) triggers expectations not only about the material outcome (the amount one has likely lost) but also about how bad it would be to receive information about that loss (Bromberg-Martin and Sharot, 2020). L-DOPA may have triggered less pessimistic expectations regarding the latter, altering the value of information about losses, which could have changed information-seeking in the loss domain. To illustrate this point, imagine two participants who accurately expect to lose £100 when they observe the market falling. One participant predicts that learning about the loss will have little negative impact, whereas the other predicts a large negative impact. Dopamine dips could signal both elements separately when observing the cue. As L-DOPA is thought to interfere with such dips (Ungless et al., 2004; Satoh et al., 2003), it could result in less pessimistic expectations about the value of bad news and thus more information-seeking. This possibility can be investigated in the future by recording participant’s actual and predicted expectations regarding the affective impact of information.

It is important to keep in mind that our task exclusively examined non-instrumental information about gains and losses. As dopamine is known to play an important role in reward-guided learning and decision-making, it is possible that dopamine plays a more general role in information-seeking when information has instrumental value and/or for non-valenced information. Future studies are needed to investigate the role of dopamine in those situations.

Because information-seeking is integral to decision-making (Kidd and Hayden, 2015; Loewenstein, 1994), understanding its biological basis is important for understanding impairments in these domains. Our results suggest that patients with deficiency to the dopamine system may exhibit abnormal patterns of information-seeking, which may provide a marker of their condition. For example, patients with low levels of dopamine function, such as patients with Parkinson’s disease, may be less likely to seek information regarding negative events. The findings also generate predictions of how prescription drugs targeting dopamine function may alter patients’ information-seeking behavior. For example, patients taking L-DOPA may increase self-exposure to negative information, which may induce negative affect.

Materials and methods

Participants

Two hundred and forty-eight subjects were recruited via the University College London psychology online system and assigned randomly to receive Placebo (123) or L-DOPA (125). Sample size was calculated based on our previous studies (Sharot et al., 2009; Sharot et al., 2012) looking at dopamine effects on decision-making. All participants filled in the informed consent and a screening form for significant medical conditions, medications, and illicit drugs. All subjects were paid for their participation. The study was double-blind and approved by the UCL ethics committee (Project ID Number: 8127/001).

Data from five subjects was lost due to technical error, and 11 subjects did not complete the task due to either feeling nausea (five subjects), power outage (one subject) or lack of interest/motivation (five subjects). Thus, we obtained full data sets from 232 participants (Placebo group: n = 116, females = 72, mean age = 24.36, SD = 7.918; L-DOPA group: n = 116, females = 71, mean age = 25.44, SD = 7.926). Education level was measured on a scale from 1 (no formal educatio) to 10 (Doctoral Degree). Income was measured on a scale from 1 (annual household income £10,000 or less), to 9 (annual household income over £100000). There were no significant differences between the groups in terms of age (t(230) = 1.036, p=0.301), income (t(228) = 0.737, p=0.462), gender (X2(1) = 0.018, p=0.893), and education level (t(230) = 1.420, p=0.157).

Procedure and task

Participants were administered either Placebo or L-DOPA (150 mg of levodopa, 37.5 mg of carbidopa, and 200 mg of entacapone) upon arrival to the lab in a double-blind fashion. They then completed a brief questionnaire - the Subjective State Questionnaire (SSQ) (Joint Formulary Committee, 2009). They began the task 40 min after the administration of L-DOPA/Placebo (L-DOPA half-life is 90 min and peaks at 60 min). The task took about 60 min to complete after which they completed the SSQ (Joint Formulary Committee, 2009) again. There was no differences between the Placebo and L-DOPA groups across SSQ (Joint Formulary Committee, 2009) items either before or after the task (see Table 2).

The task, known as the Stock Market Task, was adapted from our previous study (Charpentier et al., 2018). This task is composed of four blocks of 50 trials each. At the beginning of each block, each participant received 50 points, worth £5, which they had to invest in 2 of 5 five fictitious companies which compose a ‘global market’. On each trial, participants first observed changes in market value (a dynamic increase or decrease in the curve lasting 2.3 s). The market value fluctuations reflected changes in the overall market; therefore, it partially indicated changes in the participant’s own portfolio value. Unbeknown to the participants, on each trial, there was a 65% probability that their actual portfolio value would change consistent with the market trend. After observing the global market change, participants were asked to predict how their portfolio value likely changed relative to the previous trial from 1 (decreased a lot) to 9 (increased a lot) and their confidence in their answer from 1 (not confident at all) to 9 (extremely confident). They had up to 8 s to perform each rating. Sixty-four subjects (34 subjects received Placebo and 30 L-DOPA) were asked to state their expectation and confidence on their answer only on blocks 3 and 4, while all other subjects were asked to respond on every trial.

Participants were then given the chance to discover their portfolio value on that trial. Subjects had up to 8 s to state how much they were willing to pay to either receive or avoid information about their portfolio value. They could state their decision using a scale ranging from 99 p to avoid information (‘NO’), through 0, to 99 p to receive information (‘YES’) (p indicated pence). Position of ‘YES’ and ‘NO’ (left/right) were counterbalanced across participants. They were informed that the more they paid the greater the probability that their wish would be honored. When 0 p was selected, information was delivered 50% of the time. If they selected an amount between 1 p and 20 p, their request was honored on 55% of the trials, between 21 p and 40 p - 65%, and so on up to 95%. Participants were not aware of these exact mathematical relationships. After that, the current value of their portfolio was shown on screen or hidden (that is ‘XX points’ was shown) for 3 s. In this study, information was not instrumental, in the sense that it could not be used to change the portfolio.

At the end of the task, one trial was randomly selected and participants received the value of their portfolio on that trial (e.g, portfolio value of 60 points=£6). If on that trial they decided to pay a certain amount to receive or avoid information and their wish was honored (e.g. they paid 40 p to receive information and they received it), then that amount was deducted from the portfolio value (e.g. £6-£0.40 = £5.60).

Data analysis

First, we investigated the effect of dopamine manipulation on general aspects of information-seeking by comparing the number of trials in which subjects decided to pay to receive information, avoid information, or pay nothing, the average amount they paid to receive information, the average amount they paid to avoid information and number of missed trials between the L-DOPA and the Placebo groups with an independent samples t-test.

Then, we computed willingness to pay (WTP) on every trial with amount paid to avoid information scored negatively, and amount paid to receive information positively (zero is simply coded as zero). For each trial, a Linear Mixed Model was run to predict WTP from the two factors we had previously shown to impact information-seeking in this task (Charpentier et al., 2018) (i) valence (quantified as signed market change, which is the amount by which the market went up or down); (ii) absolute market change; as well as from (iii) group (L-DOPA or Placebo). All three factors were included as fixed and random effects, as were the interactions of group with each of the other two factors. Random and fixed intercepts were also included in the model. All linear mixed models were run in R using the lmer function (lme4 package) using maximum likelihood estimation method, the BOBYQA (Bound Optimization BY Quadratic Approximation) optimizer and a maximum number of iterations of 100,000.

As the model revealed a group by valence interaction, we next ran two mixed linear models separately for the Placebo and L-DOPA groups to tease apart that interaction. WTP was entered as the dependent factor and valence and absolute market change as fixed and random factors. The model included fixed and random intercepts. We also ran simpler models for each group separately, with WTP as a dependent measure and valence, coded in a binary fashion (market up/down), as fixed and random variable with fixed and random intercepts.

As the above analysis revealed a significant effect of valence in the Placebo group but not the L-DOPA group, we asked if the effect is due to L-DOPA altering information-seeking about potential losses, about potential gains, or both. Moreover, we ask whether the effect of L-DOPA emerged over the course of the experiment or whether it was apparent from the very beginning. Thus, we ran two separate mixed effect linear model predicting the WTP for information – one for trials in which the market went up (potential gain trials) and one for which the market went down (potential loss trials). The independent factors included (i) market change, (ii) trial number (iii), and group (L-DOPA/Placebo). As each model now includes only one polarity - either market going up or down - signed market change and absolute market change are perfectly correlated. Thus, only one factor ‘market change’ is added. In the loss domain, the greater the ‘market change’ the greater the expected losses. In the gain domain, the greater the ‘market change’ the greater the expected gains. All three factors and their interactions were included as fixed and random effects. Random and fixed intercepts were also included in the model. We followed up with four linear mixed models - one for each group and valence polarity. WTP was the dependent factor and the independent factors were (i) market change (ii) and trial number. Both factors and their interactions were included as fixed and random effects. Random and fixed intercepts were also included in the model.

Finally, we examined whether participants’ expectations are affected by L-DOPA. To this aim, we run a Linear Mixed Model predicting expectations with the following independent factors: (i) valence (signed market change), (ii) absolute market change, and (iii) group (L-DOPA or Placebo). All three factors were included as fixed and random effects as were the interactions of group with each of the other two factors. Random and fixed intercepts were also included in the model. To further examine whether there may be an effect of L-DOPA on expectations that alters over time, we added to the model above trial number as a fixed and random factor as well as all the two- and three-way interactions of trial number with the other factors.

Data availability

Anonymized data are available on GitHub (https://github.com/affective-brain-lab/A-Selective-Effect-of-Dopamine-on-Information-Seeking-Valentina-Vellani-; Vellani, 2020 copy archived at swh:1:rev:71ef6f1b6a438236450207810b0630c8738336b8).

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

Your investigation of the role of dopamine in human information seeking using a well-powered pharmacological intervention has provided novel observations for existing theories. The finding that L-DOPA administration in healthy volunteers reduces the impact of valence on information seeking in a stock market task makes testable predictions regarding alterations of information seeking in patients suffering from abnormal dopaminergic function. Congratulations for this insightful study.

Decision letter after peer review:

Thank you for submitting your article "A selective effect of dopamine on information-seeking" for consideration by eLife. Your article has been reviewed by three peer reviewers, including Valentin Wyart as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Christian Büchel as the Senior Editor.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

This manuscript describes a human pharmacology study which aims at characterizing the role of dopamine in information seeking. The authors contrast two hypotheses: 1) a general increase in information seeking, and 2) a selective increase in information seeking about potential losses. Using a stock “market task” tested on two groups of subjects (N = 116 each) under placebo and L-DOPA, they provide support for the second hypothesis. They report that subjects under placebo seek information about potential gains more than potential losses (as previously reported), and that L-DOPA reduces this asymmetry by increasing information seeking about potential losses.

The research question is of interest to a large research community interested in the substrates of information seeking under uncertainty. The use of a pharmacological protocol tested on a large group of subjects appears adequate to study the research question, and the reported results could provide new insights on information seeking. However, despite these merits, the reviewers have raised concerns regarding the analyses conducted to obtain some of the key results – concerns that the authors should address in a revised version of the manuscript. The outcome of the additional analyses used to address the reviewers' concerns will be critical for us to decide whether the manuscript can be published at eLife. The paragraphs below describe the main concerns that have been discussed among reviewers, and that should be addressed explicitly in a revised version of the manuscript. The separate reviews from the three reviewers are attached at the bottom of this message for your reference, but they do not require point-by-point responses.

Main concerns that require revisions:

1) Status of information seeking in the task. As emphasized by the authors in the description of their task, information seeking in their “stock market” task is purely non-instrumental. It cannot be leveraged by subjects to gain information that would help them maximizing rewards in the task – something which subjects should be aware of from the task instructions. We agree that information seeking behaviors can be studied in situations where it has no instrumental value. However, it is unclear how one can assume that findings obtained in a condition where information seeking is divorced from reward maximization would generalize to conditions where information seeking has an instrumental value. This is important in this study because the authors study the effect of dopamine on information seeking, and dopamine is also widely known to play an important role in reward-guided learning and decision-making. It is thus important to state, and discuss, the possibility that the results obtained in this study may not apply to tasks where information seeking can be leveraged to maximize rewards.

2) Handling of “indifferent” trials and exclusion of subjects. It is unclear to me how the authors dealt with the large number of “indifferent” information seeking ratings provided by the subjects (this is the most populated class of trials in Figure 2). The authors should clarify whether and how this large number of trials was handled in the analyses. We recommend reporting whether the groups differed in the number of trials they were indifferent under the same conditions (e.g., market-up vs. market-down) – ideally the proportion of each of these 3 choice types would be presented at least in a supplementary figure. Also, the exclusion of participants based on the empirical variability of their choices appears suboptimal. The authors have adopted a two-step approach when analyzing willingness to pay and choice – each participant's betas from individual regressions were entered into t-tests against zero, as well as independent-sample t-tests for comparisons between groups. We recommend the authors instead run a linear mixed effects model, including interactions with treatment group alongside the effects of the market variables. The benefit of this approach would be that all participants can be included and the number of tests to run would be minimized – thus reducing the type-1 error rate. Further, due to the hierarchical fitting and shrinkage, the obtained parameters would be more robust. As an example, for the analysis of the choice data, a number of subjects needed to be excluded because they did not have sufficient observations in either level of the DV. For these data, we recommend the authors instead use a generalized linear mixed effects model with a binomial link function to replace the logistic regression.

3) Fixed-effects analyses in linear mixed models. The reliance on fixed-effects estimation of interaction terms rather than random-effects estimation seems problematic, given that all the key results in the manuscript depend on interaction terms. It appears that the authors did not fit the data for every participant separately but instead just fit the ratio of participant choices across trials separately for each group. If this is what they did, it is quite unusual and there is no rationale for it mentioned in the text. More commonly the model would have been fit on the individual participant level – preferably using a hierarchical approach. It is also completely unclear why they suddenly introduce this type of modeling at this stage. If these analysis choices have a clear rationale, they should be clarified explicitly in the main text since the key results depend on it. However, by aggregating behavioral measures within groups and trials and market condition, and then analyzing the data with group as intercept and trial and the market condition as predictors, the authors are effectively using a fixed-effects approach and thus do not account for between-participant variability. In practice, the authors are generating two artificial participants (placebo and L-dopa) and then assessing the difference between those two, without accounting for between-participant variability. We recommend the authors instead run a linear mixed effects model across all subjects, with subject as random intercept and group as a between-subject regressor, and test for interactions between group, trial and the predictors of interest.

4) Reward history effects to link the results more directly to DA and reward-guided learning. The authors interpret their effects in part through the lens of DA influences on encoding of positive vs. negative RPEs, but this interpretation does not rest on specific analyses of the behavioral data. Because the “stock market” task is a prediction task, it would be very useful to know better how subjects learnt to make expectations in this task under placebo and L-DOPA. For this purpose, the authors could fit a canonical RL model (e.g., with a Rescorla-Wagner rule) to estimate the learning parameters of the subjects in the two groups. They could also (or instead) look at trial history effects (prior outcomes vs. expectations) on subsequent expectations and choices in the two groups. Because the authors are proposing in the Discussion that dopamine affects predictions, it would be very valuable to model the formation of predictions (through basic RL, for example) to know whether and how dopamine (L-DOPA) influences this learning process.

5) Implications of significant time-dependent interactions. The fact that it is the three-way interaction between group, valence and time that is significant, not the two-way interaction between group and valence has implications. This result suggests that there was no maintained valence effect but instead an effect that diminishes over time. It is difficult to know whether that is because the drug effects got weaker or other reasons, but in any case it deserves more attention – and explicit discussion – as the paper currently mostly reads as if there was a straightforward valence effect without changes over time. This is of course unless the authors can provide additional evidence for a time-independent valence by group interaction.

Reviewer #1:

This manuscript describes a human pharmacology study which aims at characterizing the role of dopamine in information seeking. The authors contrast two hypotheses: 1) a general increase in information seeking, and 2) a selective increase in information seeking about potential losses. Using a stock “market task” tested on two groups of subjects (N = 116 each) under placebo and L-DOPA, they provide support for the second hypothesis. They report that subjects under placebo seek information about potential gains more than potential losses (as previously reported), and that L-DOPA reduces this asymmetry by increasing information seeking about potential losses.

The studied question is of interest to a large research community interested in the substrates of information seeking under uncertainty. The pharmacological protocol on large groups of subjects appears adequate to test the studied question, and the reported results provide new insights on information seeking. The manuscript is also clearly written and introduces the existing research appropriately. However, I have concerns regarding the status of information seeking in the tested “stock market” task, and regarding some of the analyses conducted to obtain some of the key results – concerns which the authors should address in a revised version of the manuscript.

Main concerns:

1) Status of information seeking in the task. As emphasized by the authors in the description of their task, information seeking in their “stock market” task is purely non-instrumental. It cannot be leveraged by subjects to gain information that would help them maximizing rewards in the task – something which subjects should be aware of from the task instructions. I completely understand the authors' line of reasoning that information seeking behaviors are observed even in situations where the information has no instrumental value, and thus that information seeking can be studied in their protocol. However, I don't see how one can assume that the authors' findings, obtained in a condition where information seeking is divorced from reward maximization, would extend to conditions where information seeking has an instrumental value. This is important in this study because the authors study the effect of dopamine on information seeking, and dopamine is also widely known to play an important role in reward-guided learning and decision-making. The pattern of effects observed here in a study where information seeking is non-instrumental may thus change substantially when information seeking interacts with reward maximization. I am not asking here for additional data collection, of course, but I think it would be important to discuss not only the strengths (i.e., the fact that information seeking can be studied in isolation when divorced from reward maximization), but also the limitations of the current study (e.g., that the results obtained in this study may change in a task where information seeking can be leveraged to maximize rewards).

2) Inclusion of “indifferent” trials in reported analyses. After reading the manuscript, it was unclear to me how the authors dealt with the large number of “indifferent” information seeking ratings provided by the subjects (this is the most populated class of trials in Figure 2). The authors should clarify how this large number of trials was included in the analyses. Also, the exclusion of participants based on their choice variability seems suboptimal. Could the authors use a hierarchical model to retain all participants in the analyses? It seems like the current approach (performing single-subject analyses and then averaging parameter estimates at the group level) is suboptimal compared to a hierarchical model which could include all subjects in the analysis, even those with limited choice variability.

3) Fixed-effects analyses in Linear Mixed Models. The sole reliance on fixed-effects estimation of interaction terms (all the key results depend on interaction terms) rather than random-effects estimation seems problematic. After reading the manuscript, I was unclear regarding the exact models that were used, and thus uncertain of their validity. In any case, the authors could in theory (unless they rely solely on between-subject variability to assess the interaction terms) design hierarchical mixed-effects models to describe and estimate interaction terms, which I think would be more appropriate given their design.

4) All-or-none expectation ratings, and modeling of gradual learning of expectations. Figure 4B seems to suggest that expectation ratings provided by subjects were “all-or-none” (either strongly toward “expect loss”, or strongly toward “expect gain”). Is it the case? The authors could plot the overall distribution of expectation ratings on Figure 4B next to the current plot. Because the “stock market” task is a prediction task, it would be very useful to know a bit better how subjects learnt to make expectations in this task under placebo and L-DOPA. Could the authors use a canonical RL model (e.g., with a Rescorla-Wagner rule) to estimate the learning parameters of the subjects in the two groups over the course of each block of the task? The authors are proposing in the Discussion that dopamine affects predictions, and it would be very valuable to model the formation of predictions (through basic RL, for example) to know whether and how dopamine (L-DOPA) influences this learning process.

Reviewer #2:

In the present manuscript the authors investigate the role of dopamine in information seeking. Using L-DOPA, which increases dopamine levels, compared to placebo, they test two competing hypotheses about dopamine's involvement in information seeking. The first hypothesis states that increased dopamine leads to increased information seeking, whereas the second hypothesis states that dopamine selectively modulates the impact of valence processing on information seeking. Participants performed a stock market task, in which they could bid to receive or avoid non-instrumental information about their stocks after observing a couple of marked changes (up or down) and rating the expected change to their portfolio and the confidence therein. Echoing previous observations, participants (in the placebo group) were overall more likely to seek information when the market was going up and following larger absolute changes. In line with the second hypothesis, that dopamine selectively modulates valence effects, the authors find no differences in overall information seeking or in the effect of the absolute size of changes on information seeking between L-DOPA and Placebo. Instead, they find that L-DOPA selectively reduces the effect of valence on information seeking.

This work applies a novel approach (pharmacological intervention) to the study of the mechanisms underlying information seeking. This is an area of broad interest and, given the novelty of the experimental manipulation, readers would be interested in it whatever the outcome. I therefore think this could be a valuable contribution to the literature, but I have several concerns about the existing analyses and information provided that need to be addressed.

Major comments:

1) The authors chose a two-step approach when analyzing willingness to pay and choice – each participant's betas from individual regressions were entered into t-tests against zero, as well as independent sample t-tests for comparisons between groups. I would recommend that they instead run a linear mixed effects model, including interactions with treatment group alongside the effects of the market variables. The benefit of this approach would be that all participants can be included and the number of tests to run (with false positive rates increasing accordingly) would be minimized. Further, due to the hierarchical fitting and shrinkage, the obtained parameters would be more robust.

2) For the analysis of the choice data, a number of subjects needed to be excluded because they did not have sufficient observations in either level of the DV. For these data, I would recommend the authors instead just use a generalized linear mixed effects model with a binomial link function to replace the logistic regression (or justify why the current approach is preferable). It also wasn't clear how indifferent trials were handled in this analysis. Were these trials coded as 0 or just excluded? In any case, it would be informative to see how information seeking and avoidance each separately compared to indifference.

3) The authors should clarify the approach to trial-level analysis. As far as I can tell, they aggregated relative frequency of bid to receive vs bid to avoid, (or, in a subsequent analysis, average expectation) within groups and trials and market condition and then analyzed the data with group as intercept and trial and the market condition as predictors. If that's the case, they are effectively generating two artificial participants (placebo and L-dopa) and then assessing the difference between those two, without accounting for between-participant variability. I would recommend instead running a linear mixed effects model across all subjects (collapsing across groups), with subject as random intercept and group as a between subject regressor, and test for interactions between group, trial and the predictors of interest. Given the large number of trials that participants were indifferent, for analyses that focus on the difference between receive vs. avoid, I would recommend reporting whether the groups differed in the number of trials they were indifferent under the same conditions (e.g., market-up vs. market-down) – ideally the proportion of each of these 3 choice types would be visualized at least in the supplement (or reported in a table).

4) The authors interpret their effects in part through the lens of DA influences on encoding of positive vs. negative RPEs. Could they test this more directly by looking at trial history effects (prior outcomes vs. expectations) on subsequent expectations and choices in the two groups?

Reviewer #3:

This study is about the causal effect of changing dopamine in humans indirectly with L-DOPA on information seeking behaviours. To measure information seeking the authors use a decision task in which participants observe a stock market and can decide to pay for information about the consequences on their own personal portfolio. Alternatively, they can be indifferent or even pay to not have to see information. Decisions are made on a sliding scale with increasing money amounts being associated with higher probabilities of the preference being realized. To get additional information about participants internal expectations, they are also asked about their own predictions of portfolio changes and confidence in their prediction.

In the placebo group the authors found a clear preference for paying to see positive information or not paying to avoid positive information (at least early in the task), despite there not being any relevance of this information in increasing or decreasing future rewards. However, participants who were given L-DOPA do not show such a strong valence effect on information seeking behaviour and were instead more driven by the absolute market change (which is related to how much their portfolio might change up or down).

Additionally, L-DOPA increased expectations of portfolio value, but again, this might be more pronounced early in the task.

Overall, the task and data have potential to be interesting to the field. However, certain aspects of the task appear a bit under-analyzed and it is sometimes unclear what the exact effects really are as responses are sometimes highly aggregated and other unusual analytical choices were made. Therefore, it is quite difficult for me to assess what my recommendation would be once those additional analyses and illustrations are made.

Major concerns:

1) One of the most important problems I had with the paper was trying to find out what was driving the effects of L-DOPA. If I didn't misunderstand the behavioural responses, participants could make one out of three choices. A) They could pay to be more likely to see information, B) be indifferent and get info 50% of cases or C) pay to reduce the probability to receive information. Additionally, to this there is an amount they pay to avoid or get information but not when they are indifferent. This means behaviour under L-DOPA can change by either changing the proportions of seek, indifferent or avoid or change the amount participants are willing to pay for each or a mixture of both. However, the authors only present the data in two ways. Either they binarize the data into proportions of seeking vs not (essentially treating indifference like avoidance) or by making a more quantitive measure of willingness to pay (WTP) in which they combine amounts to seek and amounts to avoid by sign-flipping the avoid amounts (it is unclear from the Materials and methods what they do with the indifference trials, but I assume they just ignore them?). From both analyses it is unclear what exactly drives the change they observed. For example, is it a decrease in avoidance or an increase in seeking when the market goes down, that leads to smaller signed effects?

To answer these questions the authors should A) run three logistic regressions, the seeking or not, indifferent or not and avoiding or not (or find another way to look at each choice type) B) they should tease apart the quantitive effects included in the WTP metric further into amounts payed to avoid and amounts payed to seek as these could be very different things. This is important because statements like : "Specifically, administration of LDOPA increased information-seeking about potential losses without impacting information-seeking about potential gains." Rely on the idea that their metric is specifically measuring information seeking, while it is a combination of information seeking, information avoidance and indifference. If I am not mistaken their results could equally well be due to a decreased avoidance behaviours or less indifference (except changed indifference wouldn't obviously affect WTP unless indifference trials were more likely to go into one of the other two categories).

2) It would be more intuitive to show the effect of market change in positive and negative change separately rather than having absolute and signed changes in one regression. That way readers could assess whether both conditions have a modulation by market changes in the control group (because their signed vs unsigned is equally consistent with no effect in negative and a twice as large effect as the absolute in the positive case). I think it is important to know how the regressors add up in this regard because there are quite different interpretations if there is a complete lack of an effect in negative vs increased effect in positive change condition. Including the signed and unsigned regressors this way and testing them against zero is somewhat misleading (if I understood their analysis correctly) because the absolute change regressor can be significant event if their effect only exists in up market trials because of the way a regression adds and subtracts different regressor together and the two regressors are uncorrelated but not independent! This also means this statement is not necessarily true: "Note, that the absence of difference in the impact of absolute market change on information-seeking across groups indicates that both groups encoded the magnitude of the market-change equally well." As the controls might not have encoded the magnitude at all in the negative condition.

3) The way the authors ran the linear mixed effects models seemed a bit strange to me. They write: "We then ran a Linear Mixed Model predicting information seeking exactly as described above (that is the proportion of subjects who selected to seek information minus the proportion who selected to avoid information on each trial)". This suggests the authors did not fit the data for every participant separately but instead just fit the ratio of participant choices across trials separately for each group. If this is what they did, it is quite unusual and there is no rationale for it mentioned in the text. More commonly the model would have been fit on the individual participant level (potentially using a hierarchy). It is also completely unclear why they suddenly introduce a linear mixed effect style model, as they could have used one from the start or analyzed the data without it as they did in the first results they report (logistic and linear regressions). If these analysis choices have a clear rationale, it wasn't obvious from the text. In the Materials and methods again no explanation of why these unusual analysis choices were made but it again sounds like thy used aggregate values across all participants of a group for the fit instead of fitting each participant: "To explore whether the effect of LDOPA emerged over the course of the experiment or appeared from the very beginning, on each of the 200 trials we quantified information seeking as the proportion of participant that paid to receive information minus the proportion who paid to avoid information. We did this separately for each group for trials in which the market was going up and for which the market was going down. Then, we performed a Linear Mixed Model predicting information-seeking with group (LDOPA, Placebo), valence (market up, market down) and time (trial number: 1-200) and their interactions as fixed effect and group as random effect with fixed and random intercept. To further investigate the effect of valence over time we run the same Linear Mixed Model (without valence of course) predicting information seeking separately when market was down and when it was up." This also means that they have to calculate aggregate values for other potentially interesting regressors like expectation "To investigate the effect of LDOPA on expectations we calculated the mean expectation rating of participants on each trial separately for each group when the market was going up and when it was going down. Then we run a Linear Mixed Model predicting expectation rating with group (LDOPA, Placebo), valence (market up, market down) time (trial number: 1-200) and their interactions as fixed effect and group as random effect with fixed and random intercepts." Because expectation isn't used separately for every participant, the aggregate regressor completely neglects to explain any variation between participants that could feasibly explain participants choices.

Furthermore, the linear mixed effects models use the proportion measure, ignoring any magnitude without any reasons for why this analytical choice was made. This feels like throwing out valuable information without giving a reason. If using the magnitude doesn't work the reader should know this, as it is very valuable information.

4) I was quite surprised by one finding later in the paper. "The interaction between valence and group did not reach significance (β = 0.11, F(1,792) = 3.306, p = 0.069), instead there was an interaction between group, valence and time β = 0.001, F(1,792) = 9.450, p = 0.002. As can be observed in Figure 4B the three-way interaction effect was due to LDOPA increasing positive expectations when the market went down more towards the end of the task, but increasing positive expectations when the market was going up more towards the beginning of the task." As this result suggests that there was no maintained valence effect but instead an effect that diminishes over time. While, it is difficult to know whether that is because the drug effects got weaker or other reasons, I think it deserves more attention as the paper currently mostly reads as if there was a straightforward valence effect without changes over time (unless they have other evidence for a non-time dependent valence effect which I missed).

Main concerns that require revisions:

1) Status of information seeking in the task. As emphasized by the authors in the description of their task, information seeking in their “stock market” task is purely non-instrumental. It cannot be leveraged by subjects to gain information that would help them maximizing rewards in the task – something which subjects should be aware of from the task instructions. We agree that information seeking behaviors can be studied in situations where it has no instrumental value. However, it is unclear how one can assume that findings obtained in a condition where information seeking is divorced from reward maximization would generalize to conditions where information seeking has an instrumental value. This is important in this study because the authors study the effect of dopamine on information seeking, and dopamine is also widely known to play an important role in reward-guided learning and decision-making. It is thus important to state, and discuss, the possibility that the results obtained in this study may not apply to tasks where information seeking can be leveraged to maximize rewards.

We completely agree with the reviewers. As we stated in the original Discussion “It is important to keep in mind, however, that our task exclusively examined non-instrumental information about gains and losses. It is possible that dopamine plays a more general role in information-seeking for instrumental information and/or non-valenced information. Future studies are needed to investigate these possibilities.”

One cannot assume that the findings will (or will not) generalize to instrumental information-seeking, which is an intriguing empirical question to be answered with future studies. We now expand on this point in the revised manuscript.

2) Handling of “indifferent” trials and exclusion of subjects. It is unclear to me how the authors dealt with the large number of “indifferent” information seeking ratings provided by the subjects (this is the most populated class of trials in Figure 2). The authors should clarify whether and how this large number of trials was handled in the analyses. We recommend reporting whether the groups differed in the number of trials they were indifferent under the same conditions (e.g., market-up vs. market-down) – ideally the proportion of each of these 3 choice types would be presented at least in a supplementary figure.

Thank you for the opportunity to clarify how trials in which the subject selects to pay 0p (i.e. “indifferent trials”) are handled. In the Willingness to Pay (WTP) analysis these trials are handled as any other trials. We simply insert the number “0” as the dependent measure. This is not a problem, as the dependent measure varies from -99p all the way to +99p. We now make this clearer in the manuscript.

For analysis in which binary “choice” was the dependent measure (e.g., seek information/ not seek information) indifferent trials were categorized as ones in which the subject chooses not to seek information. Following the reviewer comment, however, it became apparent that there are several ways in which the WTP scale can be divided into two categories. We originally binned WTP < = 0 responses as one category and WTP >0 as the other (as explained above). However, one could also divide the data into trials in which the subject chooses to actively avoid information (WTP <0) or not (WTP > = 0). One could also not include zeros at all and bin trials into actively seek information (WTP >0) and actively avoid information (WTP <0). Given that the participants reported their decisions on a continuous scale which provides greater sensitivity, we decided to simply use the continuous WTP scale in all our analysis (as done previously – Charpentier et al., 2019) rather than bin the responses by dividing the scale into two. In this analysis indifference trials are simply coded as zero and thus there is no ambiguity.

Following the reviewers’ recommendation we now report that number of trials in which subjects selected to pay 0p (“indifferent trials”) did not differ between groups under the same condition, and we present these numbers in the supplementary material.

Also, the exclusion of participants based on the empirical variability of their choices appears suboptimal. The authors have adopted a two-step approach when analyzing willingness to pay and choice – each participant's betas from individual regressions were entered into t-tests against zero, as well as independent-sample t-tests for comparisons between groups. We recommend the authors instead run a linear mixed effects model, including interactions with treatment group alongside the effects of the market variables. The benefit of this approach would be that all participants can be included and the number of tests to run would be minimized – thus reducing the type-1 error rate. Further, due to the hierarchical fitting and shrinkage, the obtained parameters would be more robust. As an example, for the analysis of the choice data, a number of subjects needed to be excluded because they did not have sufficient observations in either level of the DV. For these data, we recommend the authors instead use a generalized linear mixed effects model with a binomial link function to replace the logistic regression.

Thank you for this excellent suggestion. We now replaced all our previous analysis with a linear mixed effects model (when examining WTP) exactly as recommended. The analysis reveals the same results as originally reported. For WTP there is a significant interaction between group and valence on the Willingness To Pay for information (β = -0.15, CI = -0.29/-0.01, t(230.52) = 2.15, p = 0.032) as well as a main effect of valence (β = 0.20, CI = 0.11/0.30, t(229.60) = 4.11, p = 0.0001) and absolute market change (β = 0.41, CI = 0.25/0.58, t(231.35) = 4.87, p = 0.0001). There was no interaction between group and absolute market change (β = 0.07, CI = -0.17/0.30, t(232.42) = 0.576, p = 0.565), nor a main effect of group (β = -2.61, CI = -7.50/2.28, t(232.53) = 1.045, p = 0.297).

While we now focus on the continuous scale and no longer bin responses into two group, we nonetheless conducted the suggested analysis for the purpose of this response and found the same results. When dividing the data into two bins – actively select information (WTP> 0) or not (WTP=< 0) and analyzing the data with logistic Linear Mixed Model we found a significant interaction between group and signed market change (β = -0.016, SE = 0.007, z = -2.214, p = 0.027), as well as a main effect of valence (β = 0.026, SE = 0.005, z = 4.939, p = 0.0001) and absolute market change (β = 0.057, SE = 0.009, z = 6.083, p = 0.0001). There was no interaction between group and absolute market change (β = -0.007, SE = 0.013, z = -0.529, p = 0.596), nor a main effect of group (β = 0.136, SE = 0.267, z = 0.511, p = 0.609).

3) Fixed-effects analyses in linear mixed models. The reliance on fixed-effects estimation of interaction terms rather than random-effects estimation seems problematic, given that all the key results in the manuscript depend on interaction terms. It appears that the authors did not fit the data for every participant separately but instead just fit the ratio of participant choices across trials separately for each group. If this is what they did, it is quite unusual and there is no rationale for it mentioned in the text. More commonly the model would have been fit on the individual participant level – preferably using a hierarchical approach. It is also completely unclear why they suddenly introduce this type of modeling at this stage. If these analysis choices have a clear rationale, they should be clarified explicitly in the main text since the key results depend on it. However, by aggregating behavioral measures within groups and trials and market condition, and then analyzing the data with group as intercept and trial and the market condition as predictors, the authors are effectively using a fixed-effects approach and thus do not account for between-participant variability. In practice, the authors are generating two artificial participants (placebo and L-dopa) and then assessing the difference between those two, without accounting for between-participant variability. We recommend the authors instead run a linear mixed effects model across all subjects, with subject as random intercept and group as a between-subject regressor, and test for interactions between group, trial and the predictors of interest.

Thank you for this helpful suggestion. We now replaced all our previous analysis with a linear mixed effects model across all subjects, with subject as random intercept and group as a between-subject regressor, and test for interactions between group, trial and the predictors of interest, exactly as recommended. The analysis reveals the same key results as originally reported. Specifically, we find an interaction between market change and group in the loss domain (β = 0.40, CI = 0.07/0.74, t (751.900) = 2.35, p = 0.018), but not in the gain domain (β =0.10, CI = -0.29/0.50, t (490.600) = 0.505, p = 0.614) on the willingness to pay for information. To tease apart the interaction of interest we run linear mixed models for each group and each valence polarity separately. This revealed that under L-DOPA participants were willing to pay more for information the greater the gains (effect of market change: β = 0.79, CI = 0.53/1.05, t (313.300) = 5.98, p = 0.0001) and the greater the losses (effect of market change: β = 0.53, CI = 0.33/0.74, t (484.700) = 5.033, p = 0.0001). In contrast, under Placebo participants were willing to pay more for information the greater the gains (effect of market change: β = 0.69, CI = 0.40/0.99, t (219.900) = 4.64, p = 0.0001) but did not show this effect for losses (effect of market change: β = 0.12, CI = -0.13/0.39, t (335.800) = 0.956, p = 0.339).

While we follow the reviewers’ request and no longer include the analysis looking at proportion of participants selecting information over time, we would like to clarify that we conducted this analysis because it has been previously conducted by us on a similar task (Charpentier et al., 2019 – PNAS, supplementary material, top of page 2 and Figure S4). It was also conducted in the past when looking at how LDOPA alters choice in a reinforcement learning task (e.g., Pessiglione et al., 2006 – Nature, Figure 1B, left panel).

4) Reward history effects to link the results more directly to DA and reward-guided learning. The authors interpret their effects in part through the lens of DA influences on encoding of positive vs. negative RPEs, but this interpretation does not rest on specific analyses of the behavioral data. Because the “stock market” task is a prediction task, it would be very useful to know better how subjects learnt to make expectations in this task under placebo and L-DOPA. For this purpose, the authors could fit a canonical RL model (e.g., with a Rescorla-Wagner rule) to estimate the learning parameters of the subjects in the two groups. They could also (or instead) look at trial history effects (prior outcomes vs. expectations) on subsequent expectations and choices in the two groups. Because the authors are proposing in the Discussion that dopamine affects predictions, it would be very valuable to model the formation of predictions (through basic RL, for example) to know whether and how dopamine (L-DOPA) influences this learning process.

In the Discussion we speculate that DA may influence learning about the affective impact of information. For example, a subject may believe they will not have much of a negative response to knowing they lost but when the information is provided they do have a strong negative reaction. This triggers a negative PE which may influence future information-seeking choices. This negative PE may be dampened by DA. This speculation could be tested in the future by measuring subjects’ predicted affective responses vs actual affective responses.

While affective responses are not record, the reviewer is suggesting that we examine more traditional learning about outcome (i.e. portfolio value). In particular, how the difference between prior portfolio value and prior expectations about portfolio value impact subsequent expectations. This is a reasonable request and indeed DA has been shown to affect this type of traditional learning (Frank et al., 2004, Pessiglione et al. 2006). However, our task was not designed as a learning task of this sort. First, there is no incentive for expectations to be accurate, nor do choices effect outcomes. Second, portfolio values (i.e. “outcomes”) are only observed on half the trials. Third, past outcomes have no bearing on present outcomes. The best way to make a prediction is to simply report the market change on that trial regardless of history. The likelihood that the portfolio will follow the same exact trend as the market is 65% and in 35% it would change in the opposite dircetion with a randomly generated magnitude. Thus, if the market goes up, the most rational response is to expect your portfolio to go up.

Nevertheless, we followed the request and examined if the difference between prior outcomes observed by the subject and prior expectations were relate to subsequent expectations. As detailed below, this did not reveal an effect, which is rationale, as the change in portfolio on the last trial is unrelated to change in portfolio on the next trial. In particular, we run a mixed linear model entering participants’ expectation rating on trial t (note – the rating is always regarding the expected portfolio change relative to the previous trial) as the dependent measure, and as the independent measure the change in outcome on last trial (portfoliot-1 minus portfoliot-2) minus expectation rating on t-1. We of course only included trials in which portfolio was in fact observed on the last trial. There was no effect (L-DOPA: β = -0.001, CI = -0.01/0.008, t (120.787) = 0.315, p = 0.753; Placebo: β = 0.001, CI = -0.009/0.12, t (108.300) = 0.304, p = 0.762).As participants often did not observe the portfolio on trial t-2 we ran the analysis again this time instead of inserting portfoliot-2 in the equation above we inserted the portfolio value last observed before t-1. Again, this did not reveal an effect (L-DOPA: β = -0.001, CI = -0.008/0.006, t (112.877) = 0.330, p = 0.742; Placebo: β = -0.0007, CI = -0.006/0.005, t (102.700) = 0.253, p = 0.800). We also examined if past outcome (portfoliot-1) was related to current prediction rating, it was not (L-DOPA group: β = 0.0001, CI = -0.003/0.002, t (519.000) = 0.464, p = 0.643; Placebo group: β = 0.0009, CI = -0.001/0.003, t (213.500) = 0.621, p = 0.535).We then run all these models again, this time predicting WTP for information. As expected none showed a significant effect (all P > 0.240). We now report these results in supplementary material.

5) Implications of significant time-dependent interactions. The fact that it is the three-way interaction between group, valence and time that is significant, not the two-way interaction between group and valence has implications. This result suggests that there was no maintained valence effect but instead an effect that diminishes over time. It is difficult to know whether that is because the drug effects got weaker or other reasons, but in any case it deserves more attention – and explicit discussion – as the paper currently mostly reads as if there was a straightforward valence effect without changes over time. This is of course unless the authors can provide additional evidence for a time-independent valence by group interaction.

Please note that the lack of two-way interaction and existence of three-way interaction is not for the information seeking data, but rather describe the expectations data. Indeed, for participants’ expectations there is no two-way interaction. In fact, when we use the new analysis recommended by the reviewers (mixed linear model) the three-way interaction does not hold. This further supports our conclusion that our key findings regarding information-seeking (that is the significant two-way interaction between group and valence on information-seeking) is not simply a consequence of the effects of L-DOPA on expectations. In the revised manuscript this is now made clear throughout.

Key resources table

Reagent type (species) or resource	Designation	Source or reference	Identifiers	Additional information
Software, algorithm	SPSS	SPSS	RRID:SCR 002865	Version 25
Software, algorithm	MATLAB	MATLAB	RRID:SCR_001622	Version R2020a
Software, algorithm	R	R	RRID:SCR_001905	R-4.0.1
Chemical compound, drug	levodopa	Orion Pharma (UK) Limited	PubChem CID:6047	150 mg
Chemical compound, drug	carbidopa	Orion Pharma (UK) Limited	PubChem CID: 34359	37.5 mg
Chemical compound, drug	entacapone	Orion Pharma (UK) Limited	PubChem CID: 5281081	200 mg