nature human behaviour
Article
https://doi.org/10.1038/s41562-024-01894-9
Comparing experience- and
description-based economic
preferences across 11 countries
Received: 23 February 2023
Accepted: 19 April 2024
Published online: xx xx xxxx
Check for updates
Hernán Anlló 1,2,3 , Sophie Bavard 1,3,4, FatimaEzzahra Benmarrakchi 3,5,
Darla Bonagura3,6, Fabien Cerrotti1,3, Mirona Cicue7, Maelle Gueguen 3,6,
Eugenio José Guzmán 8, Dzerassa Kadieva 9, Maiko Kobayashi2,
Gafari Lukumon5, Marco Sartorio10, Jiong Yang 11, Oksana Zinchenko 3,12,
Bahador Bahrami 3,13, Jaime Silva Concha 3,8, Uri Hertz 3,7,
Anna B. Konova3,6, Jian Li 3,11,14, Cathal O’Madagain 3,5, Joaquin Navajas3,10,15,16,
Gabriel Reyes 3,8, Atiye Sarabi-Jamab3,17, Anna Shestakova 3,12,
Bhasi Sukumaran 3,18, Katsumi Watanabe 2,3 & Stefano Palminteri 1,3,19
Recent evidence indicates that reward value encoding in humans is
highly context dependent, leading to suboptimal decisions in some
cases, but whether this computational constraint on valuation is a
shared feature of human cognition remains unknown. Here we studied
the behaviour of n = 561 individuals from 11 countries of markedly
different socioeconomic and cultural makeup. Our findings show that
context sensitivity was present in all 11 countries. Suboptimal decisions
generated by context manipulation were not explained by risk aversion,
as estimated through a separate description-based choice task (that is,
lotteries) consisting of matched decision offers. Conversely, risk aversion
significantly differed across countries. Overall, our findings suggest that
context-dependent reward value encoding is a feature of human cognition
that remains consistently present across different countries, as opposed
to description-based decision-making, which is more permeable to
cultural factors.
Cross-cultural differences in economic decision-making processes
have been investigated in several domains, such as risk preference
and behavioural game theory. Although several qualitative features
seem to be preserved (such as prospect theory-like preferences and
delay discounting1,2), evidence has repeatedly shown culturally driven
differences in many decision-making traits3–5.
To date, efforts to assess the cross-cultural stability of decisionmaking processes have mainly (if not only) focused on what can be
defined as description-based paradigms (that is, using tasks where all
of the decision-relevant information, such as prospective outcomes
and their costs, can be inferred from explicit cues or instructions6–8).
A full list of affiliations appears at the end of the paper.
Nature Human Behaviour
However, little is known concerning the cross-cultural stability
(or lack thereof) of experience-based decisions, which encompass all
situations where the decision-making variables have to be inferred
from past experience9,10. One prominent conceptual framework with
which to investigate experience-based decision processes is reinforcement learning (RL). The RL computational framework encompasses
the ensemble of cognitive mechanisms and behaviours involved in
the acquisition of knowledge through trial and error. More specifically, models of RL propose computational solutions to a broad range
of value maximization problems (such as foraging, navigation and
economic decision-making) by decomposing these problems in their
e-mail: hernan.anllo@cri-paris.org; stefano.palminteri@ens.fr
Article
elementary building blocks (action, state and rewards). The empirical and experimental foundations of this formal understanding of
the learning process span multiple disciplines, from neuroscience to
artificial intelligence11.
The lack of cross-cultural investigation of human RL processes is
particularly problematic, given that RL is a pervasive cognitive process,
with many important implications for mental health, education and
economics12–15. Despite its general adaptive value (seek rewards and
avoid punishments), laboratory-based research has illustrated that RL
processes in many circumstances deviate from statistical and normative standpoints16,17. Determining whether such RL biases are cultural
artefacts, or rather stable components of human decision processes,
can provide additional fundamental hints to enable understanding of
the computational constraints of bounded rationality18,19.
Among several features characterizing human RL, the notion
of outcome (or reward) context dependence has recently risen to
prominence16. More specifically, a series of studies conducted mostly
with Western, educated, industrialized, rich and democratic (WEIRD)
populations20 have shown that in many RL tasks participants encode
outcomes (that is, rewards and punishments) in a context-dependent
manner21–24. While there may not be a consensus yet concerning the
exact functional form of such context dependency, the available findings seem to favour the idea that subjective outcomes are calculated
relatively, following some form of range normalization25–27. Such context dependence-induced rescaling of subjective outcomes is often
interpreted as a consequence of efficient information coding in the
human brain28,29. According to this hypothesis, this feature can be
understood as the result of fundamental neurocomputational constraints akin to those observable in perceptual decision-making30–32.
In accordance with this proposal of outcome context dependence in
RL as a form of efficient coding, multiple studies using similar tasks
in different species have consistently found evidence of range value
adaptation, which suggests we may be looking at a general principle
of brain functioning33,34.
One well-known consequence of context dependence in RL is that,
in some cases, it can induce suboptimal decisions25–27. In particular
learning contexts, individuals mistakenly attribute higher subjective
values to objectively worse options because of how these options
are appraised in relation to the local reward distribution, resulting
in choices that fail to maximize reward. If indeed there exists such a
fundamental computational constraint in the human brain, the behavioural signatures of context dependence should be a stable feature of
decision-making, and thus persist across different populations and
cultures. In the present work, we set out to test this hypothesis by leveraging a task capable of eliciting context-dependent RL behaviours and
deploying it across 11 countries of remarkably different socioeconomic
and cultural makeup (Argentina, Iran, Russia, Japan, China, India, Israel,
Chile, Morocco, France and the United States). This allowed us to test
the cross-cultural stability of context-dependent value encoding in
human RL, and thus assess its putative role as a core computational
process of experience-based decision-making.
In addition, we also administered to our participants a descriptionbased decision-making task that included decision contexts overlapping with those presented in the RL task. The rationale behind this
second task was twofold. First, it allowed us to determine the extent
to which choice behaviour measured in the RL task can be explained
by risk aversion, estimated using standard procedures in behavioural
economics (that is, using lottery tasks). Second, it gave us the opportunity to compare the variability of experience- and description-based
decision-making processes across countries.
Our results indicate a remarkable similarity in how context dependence affects decisions from experience and generates suboptimal
choice across countries, consistent with the idea that it may derive from
deep and conserved constraints on cognition. Our results also showed
that risk aversion inferred from the description-based lottery task
Nature Human Behaviour
https://doi.org/10.1038/s41562-024-01894-9
could not account for these effects. Interestingly, description-based
decisions were also found to be highly variable across countries, further
confirming the functional dissociation between the behaviour elicited
by the two modalities6,7,35. Exploratory analyses using independent
socioeconomic, cultural and cognitive measures taken from our samples further showed that the origin of cross-country differences in
description-based decisions is multifactorial, as was previously found
for risk and other cognitive domains5,36,37. Overall, our results suggest
that reinforcement (experience-based) decision processes are much
more culturally stable than description-based ones and have important
implications for theories of bounded rationality18,19. We conclude this
work by discussing the possible implications of these results for the current implementation of policies and interventions aimed at contrasting
the economic and social burden of biased decision-making worldwide.
Results
Behavioural protocol
Our behavioural protocol consisted of a RL (that is, experience-based)
task, in the form of a previously validated two-armed bandit task26,
followed by a description-based decision-making task consisting of
choices between lotteries (Fig. 1a). Both decision-making tasks were
preceded by dedicated instructions and a short training session and
succeeded by a series of questionnaires directed at obtaining information on participants’ socioeconomic, cultural and cognitive features, as
well as general demographics (Supplementary Fig. 1). In a two-armed
bandit task, participants make trial-by-trial choices between two possible options (which would be conceptualized as lotteries, following the
traditional nomenclature in economics). Each option has a given probability of providing a certain reward, and participants’ choices can affect
reward maximization. Crucially, they initially ignore the value of each
option, but, as trials advance, participants progressively accumulate
feedback information and can learn an experiential notion of the value
of each option. In the present work, our bandit task design and implementation reproduced that of Bavard et al.26. Thus, the RL task consisted
of two phases: a learning phase and a transfer phase. During the learning
phase, participants were presented with eight abstract icon cues, each
representing a lottery of non-disclosed expected value, paired in four
stable decision contexts. In the learning phase, each decision context
featured only two possible outcomes: either 10/0 points or 1/0 points.
The outcomes were probabilistic (75 or 25%). For convenience, contexts
were labelled by taking into account the difference in expected value
between the most and least rewarding options (that is, the expected
value-maximizing (correct) and value-minimizing (incorrect) options)
(Fig. 1b). In the ensuing transfer phase, these same eight lotteries were
rearranged into new decision contexts (as was previously done in similar designs for humans and birds22,26,33,34,38). In addition to the change in
decision contexts, the key difference between the learning and transfer
phases was that during the learning phase participants were presented
with complete feedback whereas in the transfer phase no feedback was
provided, so that choices could only be based on values learned during
the learning phase (Fig. 1b). Finally, we conducted an additional task,
which we identified as the lottery task (Fig. 1c). There, the values of the
options were explicitly disclosed, as the abstract cues were replaced by
cue cards informing reward magnitudes and probabilities in an explicit
numerical manner. The lottery task featured the same decision contexts as those used in the transfer phase plus four additional contexts
designed to better assess risk preferences. These last contexts consisted
of choices comparing varying probabilities of winning 10 points (100,
75, 50 and 25%) against the certainty of winning 1 point. The present
work consists of a direct cross-cultural extension of the hypotheses
and analytical pipeline already exposed in Bavard et al.26 (tightly linked
to previous studies from our laboratory and collaborating teams22). In
this previous instance, the authors used the same outcome measures
and a computational approach largely overlapping with the present
one. Thus, while we understand the rationale behind preregistration,
Article
a
https://doi.org/10.1038/s41562-024-01894-9
RL task
Design
Training
(~5 min)
b
∆EV = 5
∆EV = 6.75
Explicit choices
(32 choices)
Transfer
(120 choices)
c
RL task
P: 0.75 0.25
M: 10
10
d
Learning
(120 choices)
Lottery task
0.75 0.25
1
1
0.75 0.25
10
10
0.75 0.25
1
1
∆EV = 0.5
∆EV = 5
∆EV = 0.5
∆EV = 2.25
∆EV = 7.25
∆EV = 1.75
Learning phase
choice sets
(with feedback)
Transfer phase
choice sets
(without feedback)
Included sites
e
Questionnaires
(~10 min)
Lottery task
∆EV = 9
10
1
100% 100%
∆EV = 6.5
10
1
75%
100%
∆EV = 4
10
1
50% 100%
∆EV = 1.5
10
1
25%
100%
10
1
75%
75%
∆EV = 6.75
10
1
25%
25%
∆EV = 2.25
10
1
75%
25%
∆EV = 7.25
10
1
25%
75%
∆EV = 1.75
Risk aversion
assessment
(without feedback)
Transfer phase
choice sets
(without feedback)
Country characteristics
HDI
Cultural distance
France
Distance from India
0.75
0.50
0.25
0.2
China
Israel
Japan
0.1
Morocco
United States
Iran
Argentina
Russia
Chile
India
0
United States
Israel
Japan
France
Chile
Argentina
Russia
Iran
China
Morocco
India
0
0
0.05
0.10
0.15
Distance from United States
Fig. 1 | Behavioural protocol and sample. a, Outline of the experimental design,
including training, the RL task, the lottery task and questionnaires.
b, Probabilities (P) and magnitudes (M) of each of the lotteries for the learning
and transfer phases of the RL task, together with the differences in expected
values (∆EV) between options for each local decision context. Complete
feedback was provided during the learning phase (factual and counterfactual
feedback), whereas no feedback was provided during the transfer phase.
c, Probabilities and magnitudes of each of the lotteries for the lottery task,
together with the differences in expected values between options for each local
decision context. No feedback was provided. d, Geographical locations of the
participating countries. Dots represent the cities where data collection was
conducted (that is, New Jersey, Haifa, Tokyo, Paris, Santiago de Chile, Buenos
Aires, Moscow, Tehran, Beijing, Rabat and Chennai), colour coded as a function
of each country’s HDI score (see left panel of e). e, Country macrometric
characteristics, including HDI scores (left) and cultural distance between each
country, India and the United States (right).
we posit that in this particular case its absence is counterbalanced by
the coherence between the existing published analytical pipelines
and the present one. Of note, analysis of preferences and choices in
the lottery task (a novelty of this study) followed the same logic as that
of the RL task. Finally, analyses on the possible correlations between
socioeconomic/cultural factors and outcome measures were explicitly defined as exploratory (as no specific hypothesis was proposed).
(Fig. 1e, left). To assess the cultural spread of the selected countries, we
used the 1981–2014 dataset of Muthukrishna and colleagues’ cultural
distance metric40, to estimate the cultural difference between each
of the selected countries with respect to the United States and India,
which represented the highest and lowest HDI values in our sample
(Fig. 1e, right).
To ensure that our samples would adequately represent the culture
of the country to which they belonged, inclusion criteria required that
participants: (1) had the target country nationality; (2) resided in the
target country; (3) had completed at least the full basic education cycle
in the target country; and (4) spoke the country’s official language as
their native language. These criteria were assessed for each participant
during a video meeting before launching the experiment. The meeting,
task instructions and questionnaires were delivered in each country’s
official language by local researchers.
Additionally, to confirm the diversity of the sample beyond country macrometrics, participants completed individual questionnaires
on socioeconomic status41, individualistic/collectivistic tendencies42
and centrality of religiosity in their social environment43, as well as a
Population demographics
Our main goal was to test the replicability of context dependence in
RL across countries (while disentangling it from risk aversion, as it is
standardly assessed in behavioural economics using lottery-based
tasks). Thus, our final sample included 11 countries (United States,
Israel, Japan, France, Chile, Argentina, Russia, Iran, China, Morocco and
India), covering a total of five continents and ten languages (Fig. 1d).
Country selection was aimed at portraying a gradual spread across the
United Nations’ Human Development Index (HDI)39. This coefficient is
built with many metrics, such as gross domestic product, industrialization, mean education level, income inequality and liberty indexes
Nature Human Behaviour
Article
https://doi.org/10.1038/s41562-024-01894-9
Table 1 | Demographic and sociocultural metrics and sample sizes
United
States
Israel
Japan
France
Chile
Argentina
Russia
Iran
China
Morocco
India
All
P
51
58
55
58
59
51
58
60
53
56
64
623
–
Completion
issues
0
7
3
3
5
1
7
6
1
2
8
43
–
Rollout issues
1
1
2
1
0
0
1
5
3
3
2
19
–
n (initial)
Exclusions
n (final)
50
50
50
54
54
50
50
49
49
51
54
561
–
Mean (s.d.)
age (years)
26.5 (4.2)
26 (2.9)
20.6 (1.7)
28.9 (5.7)
22.5 (2.2)
22.5 (3.6)
26.3
(4.1)
27 (5.4)
23.4 (2.8)
21.8 (2.9)
23.1 (4.9)
24.4
(4.6)
<0.0001
Gender (%
female)
74
70
58
67
65
72
50
65
49
47
53
60.9
0.99
University
education (%)
95a
100
100
100
100
100
100
100
100
100
100
–
HDI (2019)
0.926
0.919
0.919
0.901
0.851
0.845
0.824
0.783
0.761
0.686
0.645
–
Cultural distance
From United
States
–
0.1060
0.1222
0.1195
0.0627
0.0638
0.1369
0.0959
0.1618
0.1573
0.0845
–
–
From India
0.0845
0.1454
0.1200
0.2811
0.0491
0.0525
0.0814
0.0669
0.1474
0.0975
–
–
–
Socioeconomic status (mean (s.d.))
Childhood
3.9 (0.3)
4.8 (0.3)
6.1 (0.2)
4.8 (0.2)
5.9 (0.3)
6.1 (0.2)
4.3
(0.3)
5.1 (0.3)
4.2 (0.3)
4.6 (0.3)
5.2 (0.3)
–
<0.0001
Adulthood
3.9 (0.3)
3.5 (0.2)
5.7 (0.3)
3.9 (0.3)
4.0 (0.2)
4.9 (0.2)
4.2
(0.2)
5.2 (0.3)
4.8 (0.3)
3.8 (0.3)
5.1 (0.3)
–
<0.0001
Social
hierarchy
5.4 (0.3)
6.1 (0.2)
7.0 (0.2)
5.9 (0.2)
6.7 (0.2)
6.6 (0.2)
5.5
(0.2)
6.8 (0.2)
5.2 (0.3)
6.1 (0.3)
6.0 (0.3)
–
<0.0001
Individualistic and collectivistic tendencies (mean (s.d.))
Vertical
individualistic
18 (0.9)
22 (0.8)
23 (0.8)
18 (1.0)
17 (1.0)
18 (1.0)
21 (0.7)
23 (0.9)
26 (0.8)
25 (0.9)
24 (0.7)
–
<0.0001
Horizontal
individualistic
29 (0.6)
28 (0.7)
25 (0.8)
28 (0.6)
29 (0.6)
27 (0.7)
26
(0.7)
31 (0.6)
28 (0.8)
31 (0.5)
28 (0.8)
–
<0.0001
Vertical
collectivistic
24 (1.0)
26 (0.7)
21 (0.9)
24 (0.7)
25 (0.9)
19 (0.7)
19
(0.7)
21 (1.0)
27 (0.7)
30 (0.8)
30 (0.9)
–
<0.0001
Horizontal
collectivistic
28 (0.8)
28 (0.8)
26 (0.9)
27 (0.6)
31 (0.6)
31 (0.5)
25
(0.7)
25 (0.7)
26 (0.7)
30 (0.7)
28 (0.8)
–
<0.0001
Centrality of religiosity in social environment (mean (s.d.))
Experiences
8.0 (0.6)
6.8 (0.5)
5.8 (0.4)
6.8 (0.5)
7.5 (0.5)
5.7 (0.4)
6.4
(0.4)
9.1 (0.5)
4.0 (0.3)
13.0 (0.4)
11.0 (0.5)
–
<0.0001
Role in
ideology
9.9 (0.6)
9.0 (0.6)
8.0 (0.4)
8.9 (0.6)
10.5 (0.4)
7.1 (0.5)
8.3
(0.6)
11.0 (0.6)
5.3 (0.4)
14.0 (0.3)
11.0 (0.5)
–
<0.0001
Religious
thought
7.6 (0.4)
6.4 (0.4)
7.7 (0.3)
8.2 (0.5)
6.6 (0.4)
7.5 (0.4)
7.3
(0.4)
7.8 (0.4)
5.8 (0.4)
11.0 (0.4)
9.1 (0.5)
–
<0.0001
Private life
7.8 (0.4)
6.0 (0.5)
7.3 (0.4)
6.9 (0.5)
7.6 (0.5)
5.9 (0.4)
6.1
(0.4)
7.7 (0.6)
5.4 (0.4)
12.0 (0.5)
10.0 (0.5)
–
<0.0001
Public life
5.6 (0.5)
6.2 (0.5)
5.7 (0.3)
5.9 (0.4)
5.0 (0.4)
4.7 (0.4)
4.4
(0.3)
5.4 (0.4)
4.1 (0.3)
9.2 (0.5)
8.6 (0.5)
–
<0.0001
a
Of the 78% of US participants who chose to disclose their education level. P values were Bonferroni corrected for the comparisons presented in this table. P values were calculated by
conducting separate linear and mixed-effects linear regressions, where the country variable was used as a predictor.
cognitive reflection test44 (see Methods for a detailed description of
each metric).
Sample sizes for each country were set based on a power analysis
conducted based on the online results of Bavard et al.26 (n = 46 per country; see Methods). After exclusions (failure to complete the task, n = 43;
troubleshooting/translation issues during task rollout, n = 19), the final
sample comprised the remaining n = 561 participants (342 female; mean
(s.d.) age = 24.4 (4.6) years; n = 51 on average per country). Separate
linear regressions, using each of the demographic and sociocultural
indexes as predictors of nationality, confirmed that country samples
Nature Human Behaviour
were significantly different in many respects. A summary of these differences, demographic information, sample sizes and exclusions can
be found in Table 1. Detailed results of the regressions can be found in
Supplementary Table 1.
Experience-based RL task
First, we looked at performance in the RL task. We focused on correct responses (that is, the probability of picking the expected
value-maximizing choice) as the behavioural dependent variable.
The correct response rate was analysed separately in each RL phase
Article
(that is, learning and transfer), as a function of decision context
(a within-participants variable) and country (a between-participants
variable). We also compared the correct response rate against chance
level (or indifference; P = 0.5) to assess learning and preferences. As in
previous studies using the same or similar designs22,26, of particular
relevance for the demonstration of outcome context dependence
were: (1) the comparison of accuracies between the ∆EV = 5.0 and the
∆EV = 0.5 decision contexts in the learning phase (where an absence
of difference—the magnitude effect—is taken as a sign of relative value
learning); and (2) the preference expressed in the ∆EV = 1.75 decision
context of the transfer phase (where below-chance accuracy is taken
as an indicator of context-dependent outcome encoding).
The results showed that the average correct response rate for
the learning phase was significantly different from the chance level of
0.5 for all countries and decision contexts (Fig. 2a), which confirmed
that learning had occurred (pooled sample at ∆EV = 5.0: 0.8 ± 0.2;
t(560) = 42; P < 0.001; d (95% confidence interval (CI)) = 1.8 (1.66, 1.92);
for ∆EV = 0.5: 0.8 ± 0.2; t(560) = 38; P < 0.001; d = 1.6 (1.49, 1.74); see
Supplementary Table 3 for model selection and Supplementary Table 4
for full regression results). Although we found significant differences
in aggregate performance between countries (country main effect:
χ2(10) = 58; P < 0.001), learning and above-chance performance levels
were observable in all samples and contexts (Supplementary Fig. 2).
Importantly, we did not find statistical evidence for magnitude
effects in any of our country samples, and learning performance
remained consistently above chance for the ∆EV = 5.0 and ∆EV = 0.5
conditions in all samples (decision context main effect: χ2(1) = 2;
P = 0.142; decision context × country interaction: χ2(10) = 12; P = 0.289).
Furthermore, a corrected Akaike information criterion (AICc) weight
ratio analysis of regression models fitted to our data also pointed to
a lack of magnitude effect in our sample (that is, a model including
decision context as a regressor was 0.01 times more likely to predict
correct choices than the same model without it). As an additional index
of relative evidence of one model over the other, Bayes factor computation strongly favoured the null model (BF < 0.001).
We then turned to analysis of the transfer phase (Fig. 2b). In this
case, correct choice rates were strongly modulated across decision
contexts (decision context main effect: χ2(3) = 326; P = < 0.001). When
assessing the statistical evidence in favour of a country effect, we only
found marginal results (country main effect: χ2(10) = 18; P = 0.049; decision context × country interaction: χ2(30) = 41; P = 0.093). Upon further
inspection, an AICc weight ratio analysis of regression models fitted
to our data pointed to a lack of country effect within our sample (that
is, a model including country as a regressor was 0.0003 times more
likely to predict correct choices than the same model without it). As
an additional index of relative evidence of one model over the other, a
Bayes factor computation strongly favoured the null model (BF < 0.001;
see also the Supplementary Information). Thus, we concluded that
this marginal result was not indicative of a significant country effect.
Replicating previous findings, indicating that participants could
successfully retrieve and generalize the values learned during the learning phase, correct choice rates in the ∆EV = 7.25 and ∆EV = 6.75 decision
contexts were well above the chance level (for ∆EV = 7.25: 0.7 ± 0.3;
t(560) = 15; P < 0.001; d = 0.6 (0.55, 0.73); for ∆EV = 6.75: 0.56 ± 0.4;
t(560) = 3.5; P < 0.001; d = 0.15 (0.07, 0.23)). Crucially, however, accuracy in the ∆EV = 1.75 context was consistently below the chance level
for all countries, indicative of suboptimal preferences induced by
context dependence (pooled sample: 0.33 ± 0.3; t(560) = −12; P < 0.001;
d = −0.5 (−0.6, −0.4); see individual per-country t-tests in Supplementary Table 5). Crucially, the presence of suboptimal behaviour in the
∆EV = 1.75 context was observable in every country (Supplementary
Table 5), with no significant differences between countries (Fig. 2e, left;
see Supplementary Table 6 for post-hoc pairwise contrasts).
These results replicated the same suboptimal response patterns for the ∆EV = 1.75 decision context already seen in a previous
Nature Human Behaviour
https://doi.org/10.1038/s41562-024-01894-9
publication26, and were consistent with other previous findings showing evidence of context dependence13,22–24. Chiefly, the observed behavioural signatures of outcome context dependence were cross-culturally
stable in the RL task.
Contrary to what a model encoding values on an absolute scale
would have predicted, performance was not affected by the outcome
magnitude during the learning phase: this constitutes a positive manifestation of context-dependent adaptive coding28. Additionally, preferences were globally below chance in the ∆EV = 1.75 condition. Namely, a
previously optimal option (EV = 0.75) was preferred to a previously suboptimal option (EV = 2.5) despite its expected value being higher in the
new decision context. This illustrated the already known negative side of
outcome context dependence in the context of RL: suboptimal decisions
may arise when options are extrapolated from their original context.
Description-based lottery task
We then analysed participants’ preferences in the description-based
lottery task (Fig. 2c,d). We first considered choices in the decision contexts aimed at benchmarking risk preferences, where a sure small payoff
(1 point) was presented against risky options with varying probabilities
of delivering a bigger payoff (10 points). These four decision contexts
allowed us to estimate risk preference, quantified as the decrease in
expected value-maximizing choice rates as the probability for obtaining the larger payoff decreased (that is, the propensity to choose the
objectively higher value option as the levels of risk for that option
increased). The results showed a coherent modulation of decision
context on choice behaviour: as the involved risk increased, choice
ratios for the objectively higher value offers decreased for all countries
(pooled sample; for ∆EV = 9: 0.94 ± 0.1; t(560) = 60; P < 0.001; d = 2.6; for
∆EV = 6.5: 0.79 ± 0.2; t(560) = 23; P < 0.001; d = 1; for ∆EV = 4: 0.72 ± 0.3;
t(560) = 16; P < 0.001; d = 1; for ∆EV = 1.5: 0.53 ± 0.4; t(560) = 2; P = 0.088;
d = 0; decision context main effect: χ2(3) = 326; P = <.001; see Supplementary Table 3 for model selection and Supplementary Table 4 for
full regression results). Interestingly, although risk affected expected
value maximization in all country samples, it did so differently across
countries (country main effect: χ2(10) = 57; P < 0.001; country × decision context interaction: χ2(30) = 100; P < 0.001; see Supplementary
Table 5 for per-country t-test analyses). This indicated that preferences
expressed in the description-based task were not cross-culturally stable,
unlike behaviour observed in the RL task.
After verifying the presence of risk aversion in the benchmark
decision contexts of the lottery task, we analysed preferences in the
decision contexts homologous to those of the transfer phase in RL
(Fig. 2d). This allowed us to directly compare between experienceand description-based preferences. We focused mainly on the behaviour expressed for the ∆EV = 1.75 decision context, where a tendency
to significantly choose suboptimal choices can be interpreted as a
sign of context dependence in the RL task. Crucially, and contrary
to RL behaviour, the results showed that in all countries the correct
choice rate was significantly above chance for this decision context in
the description-based task (pooled sample; for ∆EV = 7.25: 0.9 ± 0.1;
t(560) = 58; P < 0.001; d = 2.4; for ∆EV = 6.75: 0.9 ± 0.1; t(560) = 51;
P < 0.001; d = 2; for ∆EV = 2.25: 0.9 ± 0.1; t(560) = 47; P < 0.001; d = 2;
for ∆EV = 1.75: 0.6 ± 0.4; t(560) = 9; P < .001; d = 0.4). Additionally,
the ∆EV = 1.75 lottery context presented evidence of significant
between-country differences that were absent in RL (Fig. 2e, right;
country × decision context interaction: χ2(30) = 68; P = < 0.001; see Supplementary Table 6 for post-hoc pairwise contrasts). To directly compare between descriptive and experiential choices for the ∆EV = 1.75
context, we modelled preferences in this decision context by including
an additional regressor (decision type; levels: RL and lottery). The
results indicated a significant decision modality effect (χ2(1) = 216;
P = < 0.001) that confirmed the difference between the two tasks.
Overall, the results from the lottery task illustrated two important points. First, we were able to detect significant across-country
Article
https://doi.org/10.1038/s41562-024-01894-9
b
1.0
0.5
0.8
0.3
0.6
0.4
0.2
P (EV maximizing)
RL task (learning phase)
Difference
0.1
–0.1
–0.3
0
–0.5
1 P (75%)
vs 1 P (25%)
∆EV = 0.5
10 P (75%)
vs 10 P (25%)
∆EV = 5.0
RL task (transfer phase)
1.0
0.8
0.6
0.4
Country
0.2
0.8
United States
Israel
Japan
France
Chile
Argentina
Russia
Iran
China
Morocco
India
0.6
All
0
∆EV 0.5
minus
∆EV = 5
Decision context
Decision context
d
Lottery task (risk benchmark)
1.0
P (EV maximizing)
0.8
0.6
0.4
0.2
0
10 P (100%)
vs 1 P (100%)
∆EV = 9.00
10 P (50%)
vs 1 P (100%)
∆EV = 4.00
10 P (75%)
vs 1 P (100%)
∆EV = 6.5
Lottery task (transfer contexts)
1.0
0.4
0.2
0
10 P (25%)
vs 1 P (100%)
∆EV = 1.5
Decision context
Decision context
Country pairwise contrast (∆EV 1.75)
RL task
Lottery task
Chile
Chile
France
France
Japan
Japan
India
Morocco
Iran
China
Russia
Chile
Argentina
France
Israel
Japan
Israel
United States
United States
Israel
United States
0.2
0
P < 0.05
India
Russia
Argentina
Iran
Russia
Argentina
Distance
0.4
Russia
Iran
Chile
China
Iran
Argentina
Morocco
China
Israel
Morocco
Japan
India
United States
India
France
e
10 P (25%)
vs 1 P (75%)
∆EV = 1.75
10 P (25%)
vs 1 P (25%)
∆EV = 2.25
10 P (75%)
vs 1 P (75%)
∆EV = 6.75
10 P (75%)
vs 1 P (25%)
∆EV = 7.25
Morocco
P (EV maximizing)
c
10 P (25%)
vs 1 P (75%)
∆EV = 1.75
10 P (25%)
vs 1 P (25%)
∆EV = 2.25
10 P (75%)
vs 1 P (75%)
∆EV = 6.75
10 P (75%)
vs 1 P (25%)
∆EV = 7.25
China
P (EV maximizing)
a
Fig. 2 | Behavioural results. a–d, Proportion of correct answers (that is, choices
that maximize expected value) in the RL task (learning phase (a) and transfer
phase (b)) and lottery task (benchmark of risk preferences (c) and transfer
decision contexts (d)) for each individual country (dots) and the average of all
countries (boxes) for each of the two (a) and four (b–d) task decision contexts. In
a, the difference between the big (∆EV = 5.0) and the small (∆EV = 0.5) magnitude
context is shown to the right. In c, the decision contexts were presented to
estimate risk aversion. In d, the decision contexts were homologous to those
of the transfer phase. e, Country pairwise contrasts for the ∆EV = 1.75 decision
context. Shown are the Euclidean distances between the mean proportion of
correct answers of each country during the RL task (left) and lottery task (right).
The bars represent s.e.m. The midline of each box represents the mean of all
countries. Bounds of boxes represent 95% confidence intervals of the mean. Red
boxes represent a significant pairwise contrast. In a–d, correct choice rates were
analysed independently for samples of the United States (n = 50), Israel (n = 50),
Japan (n = 50), France (n = 54), Chile (n = 54), Argentina (n = 50), Russia (n = 50),
Iran (n = 49), China (n = 49), Morocco (n = 51) and India (n = 54).
behavioural differences in our sample. This excludes that absence of
an effect in the RL task may be due to a general inability of our protocol
to detect behavioural differences. Second, these findings showed that
risk aversion, as inferred from preferences expressed in the lottery
task, could not account for preferences in the RL task. This was specifically true for the key ∆EV = 1.75 decision context, where we observed
a clear case of preference reversal when comparing the two decision
modalities45.
and parsimoniously capture the behavioural consequences of both
context-dependent outcome encoding (in the RL task) and decreasing
marginal utility (in the lottery task). In both tasks, the subjective value
of a given outcome or payoff was adjusted through the implementation
of a free parameter (0 ≤ ν ≤ 1) as follows:
10p × ν, if Robj,t = 10p
Rscaled,t = {
Robj,t
otherwise
Computational results
To quantify the observed decision-making strategies in a systematic
manner that encompassed all decision contexts across all tasks, we formalized choice behaviour using simple models built around the notion
of subjective outcome scaling. This choice was motivated by the fact
that this outcome-scaling process, described below, could satisfactorily
Nature Human Behaviour
where Rscaled,t represents the scaled subjective outcome and Robj,t the
objective unscaled outcome at trial t. For RL trials, we embedded the
scaling process within a fully parameterized version of the standard
Q-learning algorithm, where option-dependent Q values were learnt
from the range-adapted reward term Rscaled. The algorithm also included
Article
Nature Human Behaviour
a
Scaling parameter (ν)
1.00
0.75
νRL
0.50
νLOT
0.25
Un
ite
d
St
a
te
s
Is
ra
el
Ja
pa
Fr n
an
ce
C
h
Ar
i
ge le
nt
in
a
Ru
ss
ia
Ira
n
C
h
M ina
or
oc
co
In
di
a
0
b
Country pairwise contrast
νRL
νLOT
Distance
0.2
0.1
0
P < 0.05
ite
d
St
a
Is tes
r
Ja ael
Fr pa
a n
Ar Cnce
ge hi
n le
Ru tina
ss
i
Ira a
C
M h n
or in
oc a
Un
ite In co
d di
St a
a
Is tes
r
Ja ael
Fr pa
a n
Ar Cnce
ge hi
n le
Ru tina
ss
i
Ira a
C
M h n
or in
oc a
In co
di
a
India
Morocco
China
Iran
Russia
Argentina
Chile
France
Japan
Israel
United States
Un
free inverse temperature (β), forgetfulness (φ) and learning rate (α)
parameters, inasmuch as the RL process consists of acquiring value
from experience and subsequently storing that value in memory for
value actualization and learning11. For the lottery task trials, we formalized choice behaviour based on the subjective expected value that
participants attributed to each choice as a function of its inherent risk,
by multiplying Rscaled,t by reward probability (as is customarily done
in standard linear utility models46). While we retained choice inverse
temperature (β) for this instance of the model, no memory actualization or learning processes were expected to take place during lottery,
which rendered φ and α unnecessary. We differentiated between scaling
and inverse temperature in RL and lottery decision contexts by fitting
specific parameters as νRL and βRL and νLOT and βLOT, respectively. We
made sure that our fitting procedure allowed us to correctly recover
the parameters in simulated datasets, as well as produce simulations
that would closely replicate the observed behavioural data (see Supplementary Information for the procedure and results of the simulations
and parameter recovery).
Utilizing the same scaling parameter (ν) in both models was a
crucial step in the formalization, as it allowed us to compare experiential and descriptive adaptation mechanisms in the same terms while
integrating all of the possible decision contexts. We expected νRL to
reflect context-dependent range value adaptation in the RL task and νLOT
to capture marginally decreasing utility (and therefore risk aversion)
in the lottery task. It follows that νRL was expected to remain invariant across country samples, confirming that relative value encoding
occurred universally and independent of risk preferences. Conversely,
we expected νLOT to differ significantly between countries, in line with
the observed risk aversion behaviours for each country sample, and
to be decorrelated from νRL.
As shown in Fig. 3a, scaling patterns conformed to these hypotheses. First, we found minimal to no evidence for differences between
countries in νRL (νRL ~ country; sum of squares (SS) = 0.98; degrees of
freedom (d.f.) = 10; P = 0.066). We confirmed this lack of effect through
AICc weight ratio analysis: we considered a full model including country
as a predictor, and as null an identical model not including it. The results
strongly disfavoured country as a relevant predictor of νRL in terms of
information loss (that is, the full model having 0.23 times the strength
of the null model). As an additional index of relative evidence of one
model over the other, Bayes factor computation strongly favoured the
null model (BF < 0.001). Second, evidence showed that νLOT differed
significantly across country samples (νLOT ~ country; SS = 3; d.f. = 10;
P = 0.004). Here, the AICc weight ratio strongly favoured the country
effect model (the full model being 16.65 times stronger than the null
model). Finally, as seen in Fig. 3b, between-country pairwise contrasts
revealed significant differences in νLOT (see Supplementary Table 9 for
post-hoc pairwise contrasts). Indeed, νLOT differed substantially across
countries, from quite substantial risk aversion (median νLOT = 0.28 in the
Chilean sample) to moderate to high risk aversion (median νLOT = 0.62
in the Israeli sample).
Crucially, νLOT values were highly correlated with the risk aversion
behavioural patterns previously observed in the ∆EV = 1.5 (R = 0.84
(95% CI = 0.81, 0.86) and P < 0.001) and ∆EV = 1.75 lottery trials (R = 0.64
(95% CI = 0.59, 0.69) and P < 0.001) and decorrelated from νRL (R = 0.08
(95% CI = 0, 0.16) and P = 0.235) (Supplementary Fig. 4 and Supplementary Table 7).
In summary, our computational approach confirmed strong evidence for stable cross-country outcome context dependence in the RL
task using a compact computational measure. A similar analysis performed in the lottery task confirmed that the preferences in the RL task
could not be accounted for by risk aversion inferred from the lottery
task. Crucially, these results also confirmed a difference in the stability of experience- and description-based processes across countries.
To discard that the differences found in scaling between phases
could be confounded by differences in task performance (that is, lack
https://doi.org/10.1038/s41562-024-01894-9
Fig. 3 | Computational results. a, Values of the scaling free parameter estimated
during the RL task (νRL) and lottery task (νLOT). b, Country pairwise contrasts for
the scaling parameters. Shown are the Euclidean distances between the means
of the scaling parameters of each country during the RL task (left) and lottery
task (right). Translucent dots represent individual participants’ values. Dots with
a bold outline represent the mean. Bars represent s.e.m. Red boxes represent a
significant pairwise contrast.
of learning or inattention), we reanalysed and refitted the data after
excluding all participants who had less than 100% accuracy in choices
involving fully dominated options in the lottery task (as seen in previous studies on economic preferences47,48). In such contexts (that is,
∆EV = 7.25 and ∆EV = 9), suboptimal choices can be ascribed to general
inattention, or the use of task-irrelevant heuristics (for example, basing
choices on a cue’s visual features and so on). These analyses, available in
the Supplementary Information, confirmed that this strict elimination
criterion improved overall performance (and resulted in less stochastic
choices, as proxied by the increase of both βRL and βLOT). However, even
after exclusion of these participants (n = 124; total remaining, n = 437),
we were still able to replicate all of the behavioural and computational
patterns of the results presented thus far (Supplementary Figs. 5–8).
Drivers of risk aversion differences
Our main goal was to test whether the behavioural and computational signatures of context-dependent outcome encoding in RL
would replicate across samples from different countries and cultural
backgrounds, and whether or not said preferences would differ from
those of a description-based task. Indeed, we found positive evidence
showing that context dependence as captured in experience-based
decision-making tasks is stable across the included countries and distinct from risk aversion in tasks from description. Importantly, we did
not have any specific directional prediction on which cultural or socioeconomic factors would influence preferences in general (and more
Article
specifically, risk aversion in the lottery task). However, in an exploratory
manner, we evaluated whether the cultural and socioeconomic metrics
we had obtained characterized the differences in risk aversion between
samples. We did so by producing separate linear regressions of the
scaling (νRL and νLOT) and inverse temperature (βRL and βLOT) parameters
against our country- and participant-level cultural, economic and cognitive metrics. The results of these exploratory analyses (Supplementary
Table 12) showed that single-dimension subjective metrics did not
significantly predict the values of the outcome-scaling parameters
for either task. In contrast, country-level macrometrics composed of
multiple dimensions (that is, HDI and cultural distance) did improve
the models. This fell in line with previous findings on intercultural risk
preferences, which show that individual differences rarely inform risk
preferences, but country-level macrometric indexes are marginally
better5,36,37. It should be noted, however, that even when significant
the correlation magnitudes were considerably small. Nonetheless, it
should be noted that cultural metrics generally predicted changes in
νLOT, but not νRL, which was consistent with the robustness of RL biases
to cultural factors, as well as the gap between experiential and descriptive choices found in our main results.
Discussion
As a phenomenon, culture has been defined as the ensemble of transmissible social and cognitive features that determine the common
identity and way of life of a group of people49. Cross-cultural research
usually focuses on identifying how said features can be organized in
larger coherent constructs that act as cultural vectors, shaping preferences and behaviour50. Perhaps the most researched among these
constructs is the collectivism versus individualism spectrum50, which
scores tendencies to act at the behest of oneself versus the interests of
the collective42. Other well-researched cultural constructs include the
analytic versus holistic thought spectrum51 (object-focused reasoning
versus context-focused reasoning) and tight versus loose normativity
spectrum52 (strong versus lax enforcement of social norms). When it
comes to studying decisions across different cultures, these broad
indexes can be difficult to unify under a common theoretical and methodological framework, which leads to results not always being consistent51. However, despite some notable exceptions, evidence from the
past two decades has shown that WEIRD countries broadly lean towards
individualistic behaviour and analytic thought, while Eastern countries
exhibit behaviours consistent with collectivism and holistic thought50.
These cultural determinants have in turn been shown to shape several
aspects of decision and choice behaviour, including risk preferences
(for example, individualism positively correlates with loss aversion53),
heuristics (for example, analytic populations are more thorough54) and
reference point adaptation (for example, holistic populations adjust
reference points more often55).
In the present work, we sought to assess the cross-cultural stability
of another recently discovered but well-documented feature of human
behaviour: context-dependent RL. It is important to underscore that
however robust, the vast majority of the results concerning context
effects in human RL to date come from WEIRD samples16,21–26,56. This
severely limited the interpretation of context-dependent outcome
encoding as a fundamental building block of human RL. In this Article,
we aimed to address this issue by showing evidence of the generalizability of outcome context dependence in samples from 11 countries
of different sociocultural makeup. Outcome context dependence was
evident both from behavioural signatures (that is, magnitude invariant
performance in the learning phase and persistent suboptimal preferences in the transfer phase) and from analysis of the key parameter of
our computational model (that is, νRL). In addition to our RL task, we
also administered a description-based task featuring an overlapping set
of decision contexts. This allowed us to demonstrate that risk aversion
(as is standardly inferred in behavioural economics from lottery tasks)
could not account for behavioural signatures of context dependence in
Nature Human Behaviour
https://doi.org/10.1038/s41562-024-01894-9
the RL task (especially suboptimal preferences in the transfer phase).
Furthermore, we have also shown that while experience-based processes and preferences were remarkably stable across the included
countries, description-based processes were not.
By replicating the finding of context-dependent RL outside the
WEIRD space, our work shows that this cognitive feature is not likely
to be a simple cultural artefact57,58. Of course, we acknowledge that our
current sample is not diverse enough to argue for a definitive universality of contextual value encoding in RL. We also acknowledge that our
samples may be neglecting within-country variations (some of the
included countries contain within themselves very different ethnic
and linguistic communities that we did not cover). However, the fact
that our results would show this bias consistently throughout samples
constitutes strong evidence in that direction, particularly since our
samples were distinct enough to elicit between-country differences in
description-based choices. Future research efforts seeking to extend
the present findings should consider testing in rural versus urban population settings59 and across different social layers within the same societies2. Additionally, further replications should consider larger sample
sizes, both to study more complex interactions and to disambiguate
the status of near-significant effects present in the current study.
The presence of context-dependent RL across such a diverse sample falls in line with numerous previous findings pointing to the reliability of the phenomenon. Multiple studies have shown the flexibility
of context dependence38, its validity for non-binary outcomes24 and
non-binary decision spaces60 and different temporal learning dynamics61. Furthermore, instances of context-dependent value learning have
also been observed reliably in a wide range of non-human animals, as
diverse as mammals, birds and insects34,62. The coincidence between
our present cross-cultural results and the ample array of cross-species
previous findings, reinforces the notion that RL processes may be
largely hard coded and evolutionarily stable63. Indeed, despite the
incidental generation of suboptimal preferences (for example, in the
transfer phase), context-dependent value learning probably presents
an overall adaptive value. Theoretical propositions suggest that the
normativity of context-dependent value learning can be traced to at
least two, not mutually exclusive sources. First, it is possible that outcome context dependence in RL may constitute just another manifestation of the adaptive coding phenomenon28,29. In adaptive coding theory,
the neural representations of objective variables are transformed
as a function of their underlying distribution as a means to adjust to
neural constraints in information processing30,64,65. Second, it is also
possible that context-dependent value learning serves the purpose of
maximizing performance (that is, fitness) in many ecological foraging
situations66. Namely, encoding the convenience of a choice with respect
to its alternatives in context (that is, storing the result of a computation rather than all of its components) would be much less resource
intensive and ecological than committing to memory large repertoires
of absolute values dissociated from their contexts67.
A crucial contribution of the present work is the analysis of behavioural performance in a description-based decision-making task featuring the same decision problems as in the transfer phase (in addition to
other benchmark decision problems). This allowed us, first and foremost, to rule out the possibility that an absence of cross-cultural variation in context-dependent value learning could be merely due to our
inability to detect any cross-cultural differences in choice behaviour
in our sample. This was not the case, as we observed that behavioural
risk preferences elicited during the lottery task were significantly
different across countries. As with previous cross-cultural studies on
decision-making, differences in lottery-elicited risk preferences were
found to be multicausal5,36,37. Possible causes for this lack of clarity in the
aetiology of risk preferences can be traced to the diversity of methods
used to quantify risk aversion across studies and to the fact that most
of the tested predictors evaluated so far have been shown to account
for only small fractions of the total variance37. As stated, pinpointing
Article
the cultural drivers of differences in risk preferences across countries
was beyond the scope of the present work. Given their effect size and
exploratory nature, these results can not be interpreted at the moment
as anything more than venues for future research. Still, our findings
highlight the necessity of developing a unified strategy for quantifying risk preferences that may take into account the socioeconomic,
demographic and cognitive characteristics of intercultural samples68.
Importantly, the addition of a lottery task featuring decision contexts homologous to those of the RL task allowed us to directly compare
experience- and description-based choice behaviour. This led us to
show that in otherwise comparable decision contexts risk aversion as
inferred from a standard lottery task does not explain preferences in
the transfer phase of a RL task. This was particularly noteworthy for
the ∆EV = 1.75 decision context, in which suboptimal choice preferences are customarily considered a hallmark of context dependence
in value learning23,26,34. Indeed, in the present work, preference reversal
in this context was observable for all countries during RL, and shown
to be different from risk-driven choice behaviour, thus calling for an
alternative explanation.
These differences between the RL and lottery tasks, concerning
both subjective outcome encoding and cross-cultural stability, were
well recapitulated by our modelling approach. We devised a simple
parsimonious outcome-scaling process that, fitted to both experiential
and described versions of our decision problems, led to the emergence
of two clearly distinguishable sets of values for the scaling parameter. It
is important to underscore that, while for parsimony and commensurability purposes we modelled preferences in RL and lottery tasks with
the same outcome-scaling model, this does not imply the assumption
that both tasks share similar computational processes. Indeed, based
on the present and other behavioural findings13,21,26 it is likely that
these different value-scaling schemes arise from different underlying
computations altogether: respectively, outcome range adaptation in
RL and diminishing marginal utility in lottery (see the Supplementary
Information for further consideration). It is nonetheless important to
note that here we are not claiming that context-dependent valuation is
exclusive to choices based on experience (or reinforcement). In fact,
many contextual effects have been documented in descriptive choices
(such as the decoy effect). Further studies should determine whether
such effects of description-based choices are cross-culturally stable.
The present results broadly fit within the larger framework of
the experience–description gap by showing that preferences for the
same decision contexts are strongly affected by the modality in which
the problems are presented6,7,69. This begs the question of whether or
not differences in probability weighting, which are robustly reported
between experience- and description-based decisions, could explain
the observed discrepancy, and more specifically the preference reversal
in the ∆EV = 1.75 decision context8. Prima facie, the fact that the 1 point
with 75% chance option would be preferred to the 10 points at 25%
chance option is compatible with the traditional experienced-based
pattern of underweighting rare events7,70. However, it should be noted
that for the preference reversal to derive solely from different probability weightings it would require a probability distortion much larger than
what has commonly been observed in experiments and meta-analyses
to date8,71. Furthermore, the learning phase of our experience-based
task featured complete feedback—a manipulation that makes feedback
information independent from choice and thus decreases or even
eliminates insufficient probability sampling (which is the traditional
explanation for the classical probability weighting of experience-based
choices). Finally, the underweighting of rare events would not explain
the absence of a magnitude effect during the RL learning phase. Conversely, outcome context dependence does provide a satisfactory and
parsimonious explanation for the observed choice patterns in both the
learning and transfer phases.
Finally, we offer some reflection on the implications of our findings for behavioural science-inspired interventions in policy-making.
Nature Human Behaviour
https://doi.org/10.1038/s41562-024-01894-9
In recent years, the idea that descriptive models of behavioural
decision-making should be used to inform better policies (top down)
or for designing better decision architectures (bottom up) has gained
traction72–74. In the long term, this approach may help to improve both
individual and collective decision-making in domains where biases and
suboptimal decision-making represent key bottlenecks (for example,
issues such as choice of vaccination or behaviours favouring environmental protection). Historically, decision models in (behavioural)
economics, nudging and behaviourally inspired policies have been
based on description-based choice behaviour. Our results show that,
compared with description-based processes, experience-based decision models are much more stable on a cross-cultural level, possibly
capturing deep and preserved features of human cognition. We therefore believe that, especially if this pattern is confirmed and generalized
to other tasks and processes, the present work calls for a better consideration of experience-based decision models in designing behavioural
science-informed public policies in general.
Methods
Participants
Recruitment was conducted locally, through the standard channels
of each participating institution (for example, dedicated mailing
lists, flyers and online advertisements). Sample size was determined
through a power analysis based on the behavioural results of an online
experiment by Bavard et al.26. For the ∆EV = 1.75 context of said experiment (blocked trials and complete feedback version), online participants reached a difference between choice rate and chance (0.5) of
0.27 ± 0.30 (mean ± s.d.). To obtain the same difference with a power
of 0.95, the MATLAB function samsizepwr.m indicated that 46 participants per country were needed. Samples were allowed to exceed this
limit by up to 20%, to ensure that the desired power would be achieved
regardless of potential participant exclusions. Exclusion criteria consisted of failure to complete the task (n = 43) and troubleshooting/
translation issues during the online task rollout (n = 19). A remainder
of n = 561 participants (342 female; mean (s.d.) age = 24.4 (4.6) years)
comprised the final sample.
Ethics
The research was carried out following the principles and guidelines
for human experimentation provided in the Declaration of Helsinki
(1964; revised in 2013). This study belongs to a series of experiments
approved by the INSERM Ethics Evaluation Committee (IRB00003888).
Wherever needed, this ethical authorization was seconded with further
authorizations at the local level at the behest of each participating
institution (for Japan, the Waseda University Ethics Committee (2019357(1)); for the United States, Rutgers Institutional Review Board (IRB)
(Pro2019000049); for Israel, the University of Haifa IRB (psychology
ethics committee 038/20); for Russia, the Higher School of Economics Committee on Interuniversity Surveys and Ethical Assessment of
Empirical Research; for India, the Memorandum of Understanding, with
SRM University granting validity to French approval IRB00003888; and
for China, the School of Psychological and Social Sciences at Peking
University (approval number 2018-03-01)). All of the remaining collaborators did not need unique ethics approval because of compatibility between local requirements and the existing standards in France
and other countries. All participants provided standardized written
informed consent before inclusion.
Payment
To sustain motivation throughout the experiment, participants were
given a bonus depending on the number of points won in each task. To
ensure motivation would be even across countries, each participating
institution calculated the average cost of a local university lunch (an
inter-country average cost of €5.8 ± 2.82) and divided it by the total
number of points to be potentially won throughout the experiment
Article
(that is, 1,275 points for a perfect run, with an average value of points
of €0.0045 ± 0.002 and an average bonus reward of €5.4 ± 1.53). In
addition to the bonus accrued through point accumulation, all participants received a flat participation rate equivalent to an additional
student lunch (see Supplementary Table 2 for average bonuses in local
currencies).
Behavioural task
There were two behavioural tasks: the RL task and the lottery task
(Fig. 1a). The RL task was a direct reproduction of the probabilistic
instrumental learning task performed in experiment 7 of Bavard et al.26.
Participants were asked to choose on a trial basis between the undisclosed lotteries of different two-armed bandit problems, with the goal
of maximizing overall reward. The lottery task consisted of a standard
economic decision-making task, where participants had to choose on
a trial basis between two lotteries of known expected value, again with
the intention of maximizing overall reward.
In the RL task, the lotteries for each decision context were represented by abstract stimuli (cues) taken from randomly generated
identicons. Identicons were generated so that hue and saturation had
similar values within the HSLUV colour scheme (www.hsluv.org). In the
lottery task, cue cards displaying the reward and probability values for
each option were used instead. For all tasks, each decision context was
formed by two cues, one at each side of the screen, equidistant from the
screen centre. Each trial consisted of a single decision context. Stimulus
location was pseudo-randomized, so that every cue would appear an
equal number of times on each side of the screen.
In the RL task, participants had to complete a learning phase and
then a transfer phase16,21–26,56. In the learning phase (Fig. 1b, top) cues
appeared in four different fixed pairs (that is, decision contexts). Within
pairs, each cue would lead to possible zero and non-zero outcomes
with reciprocal probabilities (0.75/0.25 and 0.25/0.75). Each decision
context featured only two possible outcomes: either 10/0 points or 1/0
points. Contexts were labelled by taking into account the difference
in expected value between options (that is, two ∆EV = 5.0 and two
∆EV = 0.5 decision contexts). Once a choice was made by clicking on a
cue, a fixed 500 ms delay ensued, after which factual and counterfactual choice feedback was displayed for 1,000 ms in the form of 10, 1 or
0 points cue cards. After learning phase completion, the subtotal of
points earned was displayed, together with its monetary equivalent in
the local currency. In the transfer phase, cues were rearranged into four
new pairs (∆EV = 7.25, ∆EV = 6.75, ∆EV = 2.25 and ∆EV = 1.75). Crucially,
the probability of obtaining a specific outcome from each cue remained
the same as in the learning phase (Fig. 1b, bottom). In the lottery task
(Fig. 1c), participants had to choose between explicit cue cards, which
were paired reproducing the four decision contexts of the transfer
phase, and another four decision contexts comparing varying probabilities of winning 10 points (100, 75, 50 and 25%) versus the certainty
of winning 1 point (∆EV = 9, ∆EV = 6.5, ∆EV = 4 and ∆EV = 1.5). Neither the
transfer phase nor the lottery task presented any post-choice feedback:
choices were followed by a fixed 500 ms delay interval, after which
‘???’ cue cards were displayed for 1,000 ms. Each decision context of
the RL task (four in the learning phase and four in the transfer phase)
was presented 30 times, for a total of 240 trials. Decision contexts of
the lottery task (four reproducing transfer and four benchmarking
risk aversion) were presented four times each, for a total of 32 trials.
The presentation order of decision contexts was pseudo-randomized
within each phase so that all trials of a given decision context would be
clustered (that is, blocked stimuli presentation).
Questionnaires
After completing the behavioural experiment, participants were
required to complete several psychometric and socioeconomic questionnaires. Socioeconomic questionnaires included the individualistic
and collectivistic tendencies inventory42, the perceived socioeconomic
Nature Human Behaviour
https://doi.org/10.1038/s41562-024-01894-9
status in childhood, adulthood and social hierarchy questionnaires41
and the centrality of religiosity questionnaire43. The sole goal of these
questionnaires was to confirm that samples were socioculturally different from each other, as simply belonging to different countries
may not have ensured a difference. Psychometric questionnaires were
incorporated for purely exploratory purposes, including the Ten-Item
Personality Inventory75 and the extended version of the Cognitive
Reflection Test44. The order of the questionnaires, as well as the questions within each questionnaire, was randomized (see Supplementary
Information for a technical description of each questionnaire and
exploratory analyses).
Country metrics
Questionnaires gave us the opportunity to assess different dimensions
of the socioeconomic and cultural makeup of each country sample
from participants’ own subjective answers. To quantify the socioeconomic and cultural profile of each country sample in a macrometric
way, we also incorporated into the analysis each country’s HDI score39,
as well as the cultural distance between countries40. Both of these coefficients were computed by combining large numbers of economic, educational, political and psychosocial markers. Under the same rationale
as the questionnaires, inclusion of these metrics was not hypothesis
driven, but rather served to establish the differences between country
samples and enable exploratory analyses (see Supplementary Information for details on metrics).
Procedure
Testing was conducted in a hybrid face-to-face/online format, where
participants met a local experimenter for an online live debrief held in
their local language to verify identity and cultural affiliation. After the
interview, participants received a personalized link to a Gorilla server
(www.gorilla.sc) where the experiment was hosted. After clicking on
the link, participants were sent to a consent form, which they had to
complete in order to access the experiment. The experiment started
by providing written instructions on how to perform the task. It was
explained to participants that they would have to choose between
two different options over several trials, with the goal of maximizing
overall point reward. They were told that they would have to make this
decision without necessarily knowing the probability and magnitude of
rewards for each option at first. Finally, it was explained at length that
their final payoff would be affected by their choices, as rewards were
convertible to actual currency. The possible outcomes in points (0, 1
and 10 points) were explicitly shown, as well as the points-to-currency
conversion rate for their country (for example, 1 point = €0.005 in
France; see Supplementary Table 2). Instructions were followed by a
short training session of 12 trials, designed to familiarize participants
with response modality. Participants could decide to repeat the training session up to twice before starting the experiment. After finishing
the training session, participants had to complete the RL task (learning
and transfer), lottery task and sociocultural questionnaires, in that
order. Presenting the tasks in this particular order, rather than balancing task presentation order, was deemed preferable to prevent participants from entering the RL task with previous reward distribution
knowledge that could affect performance76–78. Namely, the lottery task:
(1) provided participants with the exact value of all choices and their
range; and (2) revealed the configuration of all decision contexts in the
transfer phase. Such information could push participants to implement rogue policies that would turn the RL task into a matching task
(for example, actively searching for which implicit cue corresponds
to which explicit cue). As an additional measure to prevent the use
of alternative strategies, the existence of the transfer phase was not
disclosed until the end of the learning phase. Crucially, before starting
the transfer phase, participants were made explicitly aware of the fact
that they would be presented with the same cues they had seen during
the learning phase, but combined in different pairs. Before starting
Article
https://doi.org/10.1038/s41562-024-01894-9
the lottery task, participants were shown an example of a cue card
with its explicit reward probability and magnitude written on it and
were again instructed to choose the option that they thought would
maximize overall point reward. Following completion of the lottery
task, participants had to answer all sociocultural and psychometric
questionnaires. The order of the questionnaires, as well as the order
of each item within the questionnaires, was randomized. Completing the full experiment, including consent and questionnaires, took
approximately 25 min (average response time per trial: 1.46 ± 6.7 s;
median: 0.96 s). Once finished with the experiment, participants were
given a personalized completion code and were tasked with sending
this code to the experimenter by email to signal completion and trigger
payment. The online debrief, task instructions and questionnaires were
all delivered in each country’s official language by local researchers.
Statistical analyses
All of the statistical analyses were performed and visualized using
R79–81. The main dependent variable was the correct choice rate (that is,
choices directed towards the option with the highest expected value).
Statistical effects were assessed by phase, using generalized linear
mixed-effect models with a random intercept per participant79, with
decision context and country of sample as categorical predictors (that
is, P(correct) ~ decision context × country + ε; see Supplementary Information for model selection). P values were computed through analysis
of deviance (type II Wald χ2 test) and we report χ2, degrees of freedom
and P values. The proportion of variance explained per predictor was
not reported because of how variance is partitioned in mixed models82.
In cases where only one data point per participant was available (for
example, differences in parameter values across countries), statistical significance was evaluated through standard linear models using
country as a categorical predictor (for example, νRL ~ country). For those
analyses, we report the F statistic, sum of squares, P value and Cohen’s
F. Post-hoc contrasts were calculated with their respective confidence
intervals, through estimated marginal means analysis, and P values
were Benjamini–Hochberg corrected. In particular, whenever we had
to assess whether choice rate performances were significantly different
from chance, we performed additional t-tests against the chance level
(0.5). In those cases, we report the t-statistic, P value and Cohen’s d to
estimate effect size. The significant association between continuous
quantities (for example, between parameter value and performance
for a given decision context) was tested through correlation analysis,
for which we report the t-statistic, degrees of freedom, P values and the
R coefficient as the effect size. To prove lack of effect, we conducted
AICc weight ratio analyses83,84 using a model containing the tested
predictor (full) and its equivalent minus said predictor (null).
Computational analyses
The SCALING model was built around the notion of value scaling.
Value scaling for both the RL and lottery tasks was arbitrated by the
free parameter (ν) designed to capture value adaptation as follows:
then modelled participants’ choice behaviour using a softmax decision
rule that yielded the probability that for a state s a participant would
choose, say, option a over option b according to:
1
P(a)t =
1 + e β×(Q(b)t −Q(a)t )
where β is the inverse temperature parameter. Low inverse temperatures (β → 0) cause the action to be stochastically equiprobable. High
inverse temperatures (β → +∞) result in choices deterministically determined by the difference between the Q values11. Our algorithm also
included a forgetfulness parameter ϕ (0 ≤ ϕ ≤ 1) that allowed us to
account for the possibility of forgetting the option values when moving
from the learning to the transfer phases of the RL task. The Q values
used to fit (and simulate) the transfer phase choices ( Q (∶)TRA ) were
calculated from the Q values of the learning phase Q (∶)LEA as follows:
Q (∶)TRA = Q (∶)LEA × ϕ
For the lottery task, expected utilities (EU) of individual lotteries
were calculated based on the described probability (P) of its non-zero
outcome and the subjective rescaled rewards ( Rscaled,t, calculated as for
the learning task). For example, the expected value of lottery a was
calculated as follows:
EU(a) = R(a)scaled,t × P(a)
Choice probabilities were also instantiated through a softmax rule,
as follows (probability of choosing lottery a over lottery b):
1
P(a)t =
1 + e β×(EU(b)−EU(a))
Since the lottery task does not involve learning or memory processes, its model lacked any notion of learning rate and forgetting
parameter. The RL task and lottery model shared the scaling parameter
and inverse temperature that were fitted specifically for each task (νRL
and νLOT, and βRL and βLOT).
Model parameters were fitted using maximum likelihood estimation with gradient descent, as implemented in MATLAB. Finally,
in the ‘Alternative models’ section in Supplementary Information,
we compare SCALING with three alternative computational models
to discard other possible interpretations of our data. These include
the ABSOLUTE model, which encodes outcomes on an absolute scale
independent of the decision context in which they were presented;
the ABSOLUTE-RISK model, which rescales rewards for the RL task
trials using the νLOT parameter fitted on lottery task trials, to evaluate
whether risk aversion predicts preference reversal; and the NEGLECT
model, which assumes that participants only learned the probabilities
behind each choice, but ignored reward magnitude.
Reporting summary
Further information on research design is available in the Nature
Portfolio Reporting Summary linked to this article.
10p × ν, if Robj,t = 10p
Rscaled,t = {
Robj,t
otherwise
where Rscaled,t represents the scaled objective reward Robj,t at trial t and
0 ≤ ν ≤ 1. For RL task trials, we used a simple Q-learning model11 to estimate for each choice context (or state) the expected reward (Q) of each
option and pick the one that maximized this expected reward Q. At trial
t, option values (for example, of the chosen option c) were updated
according to the delta rule:
Data availability
Data for the present study are available for free (for non-commercial
use only) from our OSF.io repository (https://osf.io/yebm9/?view_
only=). Source data are provided with this paper.
Code availability
Main analysis scripts are available (for non-commercial use only)
from the Human Reinforcement Learning Team GitHub repository
(https://github.com/hrl-team/WEIRDbandit).
Q(c)t+1 = Q(c)t + αc × (R(c)scaled,t − Q (c)t )
References
where αc is the learning rate for the chosen option, which, multiplied by
the difference between Rscaled,t and Qt, is the prediction error term. We
Nature Human Behaviour
1.
Ruggeri, K. et al. Replicating patterns of prospect theory for
decision under risk. Nat. Hum. Behav. 4, 622–633 (2020).
Article
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
Ruggeri, K. et al. The globalizability of temporal discounting.
Nat. Hum. Behav. 6, 1386–1397 (2022).
Hallsson, B. G., Siebner, H. R. & Hulme, O. J. Fairness, fast and
slow: a review of dual process models of fairness. Neurosci.
Biobehav. Rev. 89, 49–60 (2018).
Kim, B., Sung, Y. S. & McClure, S. M. The neural basis of cultural
differences in delay discounting. Phil. Trans. R. Soc. B 367,
650–656 (2012).
Rieger, M. O., Wang, M. & Hens, T. Risk preferences around the
world. Manag. Sci. 61, 637–648 (2013).
Garcia, B., Cerrotti, F. & Palminteri, S. The description–experience
gap: a challenge for the neuroeconomics of decision-making
under uncertainty. Phil. Trans. R. Soc. B 376, 20190665
(2021).
Hertwig, R. & Erev, I. The description–experience gap in risky
choice. Trends Cogn. Sci. 13, 517–523 (2009).
Wulff, D. U., Mergenthaler-Canseco, M. & Hertwig, R. A
meta-analytic review of two modes of learning and the
description–experience gap. Psychol. Bull. 144, 140–176 (2018).
Niv, Y. Reinforcement learning in the brain. J. Math. Psychol. 53,
139–154 (2009).
Wimmer, G. E., Daw, N. D. & Shohamy, D. Generalization of value in
reinforcement learning by humans. Eur. J. Neurosci. 35, 1092–1104
(2012).
Sutton, R. S. & Barto, A. G. Reinforcement Learning: An
Introduction 2nd edn (MIT Press, 2018).
Frank, M. J., Seeberger, L. C. & O’reilly, R. C. By carrot or by stick:
cognitive reinforcement learning in parkinsonism. Science 306,
1940–1943 (2004).
Vandendriessche, H. et al. Contextual influence of reinforcement
learning performance of depression: evidence for a negativity
bias? Psychol. Med. 53, 4696–4706 (2022).
Plonsky, O., Roth, Y. & Erev, I. Underweighting of rare events in
social interactions and its implications to the design of
voluntary health applications. Judgm. Decis. Mak. 16, 267–289
(2021).
Ho, T. H., Camerer, C. F. & Chong, J.-K. Self-tuning experience
weighted attraction learning in games. J. Econ. Theory 133,
177–198 (2007).
Palminteri, S. & Lebreton, M. Context-dependent outcome
encoding in human reinforcement learning. Curr. Opin. Behav.
Sci. 41, 144–151 (2021).
Palminteri, S. & Lebreton, M. The computational roots of positivity
and confirmation biases in reinforcement learning. Trends Cogn.
Sci. 26, 607–621 (2022).
Kahneman, D. Maps of bounded rationality: psychology for
behavioural economics. Am. Econ. Rev. 93, 1449–1475 (2003).
Todd, P. M. & Gigerenzer, G. Bounding rationality to the world.
J. Econ. Psychol. 24, 143–165 (2003).
Henrich, J., Heine, S. & Norenzayan, A. The weirdest people in the
world? Behav. Brain Sci. 33, 61–83 (2010).
Palminteri, S. et al. Contextual modulation of value signals in
reward and punishment learning. Nat. Commun. 6, 8096
(2015).
Bavard, S. et al. Reference-point centering and range-adaptation
enhance human reinforcement learning at the cost of irrational
preferences. Nat. Commun. 9, 4503 (2018).
Klein, T., Ullsperger, M. & Jocham, G. Learning relative values in
the striatum induces violations of normative decision making.
Nat. Commun. 8, 16033 (2017).
Hayes, W. M. & Wedell, D. H.Reinforcement learning in and out
of context: the effects of attentional focus. J. Exp. Psychol. Learn.
Mem. Cogn. 49, 1193–1217 (2023).
Juechems, K. & Summerfield, C. Where does value come from?
Trends Cogn. Sci. 23, 836–850 (2019).
Nature Human Behaviour
https://doi.org/10.1038/s41562-024-01894-9
26. Bavard, S., Rustichini, A. & Palminteri, S. Two sides of the same coin:
beneficial and detrimental consequences of range adaptation in
human reinforcement learning. Sci. Adv. 7, eabe0340 (2021).
27. Hayes, W. M. & Wedell, D. H. Testing models of context-dependent
outcome encoding in reinforcement learning. Cognition 230,
105280 (2023).
28. Rustichini, A., Soukupova, M. & Palminteri, S. Adaptive coding is
optimal in reinforcement learning. SSRN https://doi.org/10.2139/
ssrn.4320894 (2023).
29. Padoa-Schioppa, C. & Rustichini, A. Rational attention and
adaptive coding: a puzzle and a solution. Am. Econ. Rev. 104,
507–513 (2014).
30. Fairhall, A. et al. Efficiency and ambiguity in an adaptive neural
code. Nature 412, 787–792 (2001).
31. Sato, T. et al. An excitatory basis for divisive normalization in visual
cortex. Nat. Neurosci. 19, 568–570 (2016).
32. Carandini, M. & Heeger, D. J. Summation and division by neurons
in primate visual cortex. Science 264, 1333–1336 (1994).
33. Freidin, E. & Kacelnik, A. Rational choice, context dependence,
and the value of information in European starlings (Sturnus
vulgaris). Science 334, 1000–1002 (2011).
34. Pompilio, L. & Kacelnik, A. Context-dependent utility overrides
absolute memory as a determinant of choice. Proc. Natl Acad. Sci.
USA 107, 508–512 (2010).
35. Garcia, B. Experiential values are underweighted in decisions
involving symbolic options. Nat. Hum. Behav. 7, 611–626 (2023).
36. Gandelman, N. & Hernández-Murillo, R.Risk aversion at the
country level. Fed. Res. Bank St. Louis Rev. 97, 53–66 (2015).
37. Haridon, O. & Vieider, F. All over the map: a worldwide
comparison of risk preferences. Quant. Econ. 10, 185–215 (2019).
38. Juechems, K., Altun, T., Hira, R. & Jarvstad, A. Human value
learning and representation reflect rational adaptation to task
demands. Nat. Hum. Behav. 6, 1268–1279 (2022).
39. Human Development Report 2020: The Next Frontier: Human
Development and the Anthropocene (United Nations Development
Programme, 2020).
40. Muthukrishna, M. et al. Beyond Western, educated, industrial,
rich, and democratic (WEIRD) psychology: measuring and
mapping scales of cultural and psychological distance. Psychol.
Sci. 31, 678–701 (2020).
41. Griskevicius, V. et al. When the economy falters, do people spend
or save? Responses to resource scarcity depend on childhood
environments. Psychol. Sci. 24, 197–205 (2013).
42. Triandis, H. C. & Gelfland, M. J. Converging measurement of
horizontal and vertical individualism and collectivism. J. Pers.
Soc. Psychol. 74, 118–128 (1998).
43. Huber, S. & Huber, O. The centrality of religiosity scale (CRS).
Religions 3, 710–724 (2012).
44. Toplak, M. E., West, R. F. & Stanovich, K. E. Assessing miserly
information processing: an expansion of the cognitive reflection
test. Think. Reason. 20, 147–168 (2014).
45. Lichtenstein, S. & Slovic, P. The Construction of Preference
(Cambridge Univ. Press, 2006).
46. Cartwrigth, E. Behavioural Economics (Routledge, 2018).
47. Alós-Ferrer, C. et al. Preference reversals: time and again. J. Risk
Uncertain. 52, 65–97 (2016).
48. Alós-Ferrer, C. & Granic, G. D. Does choice change preferences?
An incentivized test of the mere choice effect. Exp. Econ. 26,
499–521 (2023).
49. Smith, S. Cultural Anthropology (Allyn and Bacon, 1997).
50. Yates, F. & de Oliveira, S. Culture and decision making. Organ.
Behav. Hum. Decis. Process. 136, 106–118 (2016).
51. Choi, I., Choi, J. A. & Norenzayan, A. in Blackwell Handbook of
Judgment and Decision Making (eds Koehler, D. J. & Harvey, N.)
504–524 (Blackwell Publishing, 2004).
Article
52. Gelfand, M. J. et al. Differences between tight and loose cultures:
a 33-nation study. Science 332, 1100–1104 (2011).
53. Kitayama, S. & Cohen, D. Handbook of Cultural Psychology
2nd edn (Guilford Press, 2018).
54. Yates, J. F. et al. Indecisiveness and culture: Incidence, values,
and thoroughness. J. Cross Cult. Psychol. 41, 428–444 (2010).
55. Arkes, H. R., Hirshleifer, D., Jiang, D. & Lim, S. S. A cross-cultural
study of reference point adaptation: evidence from China,
Korea, and the US. Organ. Behav. Hum. Decis. Process. 112,
99–111 (2010).
56. Spektor, M. & Seidler, H. Violations of economic rationality
due to irrelevant information during learning in decision from
experience. Judgm. Decis. Mak. 17, 425–448 (2022).
57. Barret, H. C.Towards a cognitive science of the human:
cross-cultural approaches and their urgency. Trends Cogn. Sci.
24, 620–638 (2020).
58. Nielsen, M., Haun, D., Kartner, J. & Legare, C. H. The persistent
sampling bias in developmental psychology: a call to action.
J. Exp. Child Psychol. 162, 31–38 (2017).
59. Linnell, K. J. & Caparos, S. Urbanisation, the arousal system, and
covert and overt attentional selection. Curr. Opin. Psychol. 32,
100–104 (2020).
60. Bavard, S. & Palminteri, S. The functional form of value
normalization in human reinforcement learning. eLife 12,
e83891 (2023).
61. Hayes, W. M. & Wedell, D. Effects of blocked versus interleaved
training on relative value learning. Psychon. Bull. Rev. 30,
1895–1907 (2023).
62. Solvi, C. et al. Bumblebees retrieve only the ordinal ranking of
foraging options when comparing memories obtained in distinct
settings. eLife 11, e78525 (2022).
63. Kacelnik, A., Vasconcelos, M. & Monteiro, T. Testing cognitive
models of decision-making: selected studies with starlings.
Anim. Cogn. 26, 117–127 (2023).
64. Rangel, A. & Clithero, J. A. Value normalization in decision
making: theory and evidence. Curr. Opin. Neurobiol. 22,
970–981 (2012).
65. Louie, K. & Glimcher, P. W. Efficient coding and the neural
representation of value. Ann. NY Acad. Sci. 1251, 13–32 (2012).
66. McNamara, J. M., Trimmer, P. C. & Houston, A. I. The ecological
rationality of state-dependent valuation. Psychol. Rev. 119,
114–119 (2012).
67. Hunter, L. E. & Daw, N. D. Context-sensitive valuation and learning.
Curr. Opin. Behav. Sci. 41, 122–127 (2021).
68. Frey, R., Pedroni, A., Mata, R., Rieskamp, J. & Hertwig, R.
Risk preference shares the psychometric structure of major
psychological traits. Sci. Adv. 3, e1701381 (2017).
69. Madan, C. R., Ludvig, E. A. & Spetch, M. L. Comparative
inspiration: from puzzles with pigeons to novel discoveries with
humans in risky choice. Bahav. Process. 160, 10–19 (2019).
70. Zilker, V. & Pachur, T. Nonlinear probability weighting can reflect
attentional biases in sequential sampling. Psychol. Rev. 129,
949–975 (2022).
71. Erev, I. et al. Choice prediction competition: choices from
experience and from description. J. Behav. Decis. Mak. 23,
15–47 (2010).
72. Thaler, R. H. & Sunstein, C. R. Libertarian Paternalism Is Not an
Oxymoron Public Law and Legal Theory Working Paper No. 43
(Univ. Chicago, 2003).
73. Grüne-Yanoff, T., Marchionni, C. & Feufel, M. Toward a framework
for selecting behavioural policies: how to choose between boosts
and nudges. Econ. Philos. 34, 243–266 (2018).
74. Brown, P., Cameron, L., Wilkinson, M. & Taylor, D. in The Handbook
of Behaviour Change (eds Hagger, M. et al.) 617–631 (Cambridge
Univ. Press, 2020).
Nature Human Behaviour
https://doi.org/10.1038/s41562-024-01894-9
75. Gosling, S. D., Rentfrow, P. J. & Swann, W. B. Jr. A very brief
measure of the big five personality domains. J. Res. Pers. 37,
504–528 (2003).
76. Doll, B. B., Jacobs, W. J., Sanfey, A. G. & Frank, M. J. Instructional
control of reinforcement learning: a behavioral and
neurocomputational investigation. Brain Res. 1299, 74–94 (2009).
77. Li, J., Delgado, M. & Phelps, E. How instructed knowledge
modulates the neural systems of reward learning. Proc. Natl Acad.
Sci. USA 108, 55–60 (2010).
78. Wang, Z. & Taylor, M. E. Interactive reinforcement learning
with dynamic reuse of prior knowledge from human and agent
demonstrations. In Proc. 28th International Joint Conference on
Artificial Intelligence (IJCAI'19) 3820–3827 (AAAI Press, 2019).
79. Bates, D., Maechler, M., Bolker, B. & Walker, S. Fitting linear
mixed-effects models using lme4. J. Stat. Softw. 67, 1–48 (2015).
80. R Core Developmemt Team R: A Language and Environment for
Statistical Computing (R Foundation for Statistical Computing,
2014).
81. Wickham, H. ggplot2: Elegant Graphics for Data Analysis
(Springer, 2009).
82. Rights, J. D. & Sterba, S. K. Quantifying explained variance
in multilevel models: an integrative framework for defining
R-squared measures. Psychol. Methods 24, 309–338 (2019).
83. Burnham, K. P., Anderson, D. R. & Huyvaert, K. P. AIC model
selection and multimodel inference in behavioral ecology:
some background, observations, and comparisons. Behav. Ecol.
Sociobiol. 65, 23–35 (2011).
84. Wagenmakers, E. J. & Farrell, S. AIC model selection using Akaike
weights. Psychonom. Bull. Rev. 11, 192–196 (2004).
Acknowledgements
We thank a number of colleagues and peers, including the members
of the Human Reinforcement Learning laboratory and all of the
senior researchers who provided feedback during the multiple
conference presentations in which this work was featured. We
also thank Waseda University and the École Normale Supérieure
Department of Cognitive Studies for aiding us with the many
logistical obstacles that we had to overcome in order to kickstart
this study during the thick of the COVID-19 pandemic. We especially
thank all of the participants who kindly contributed their time to
make this study a reality. S.P. is supported by the European Research
Council under the European Union’s Horizon 2020 research and
innovation programme (RaReMem: 101043804), the Agence
Nationale de la Recherche (CogFinAgent: ANR-21-CE23-0002-02;
RELATIVE: ANR-21-CE37-0008-01; RANGE: ANR-21-CE28-0024-01)
and the Alexander von Humboldt-Stiftung. O.Z., D.K. and A.S. were
supported by the Basic Research Program at the National Research
University Higher School of Economics (HSE University). U.H. and
M.C. were supported by the Israel Science Foundation (1532/20).
K.W. was supported by JSPS KAKENHI (22H00090) and JST
Moonshot Research and Development (JPMJMS2012). A.B.K., M.G.
and D.B. were supported by the National Institute on Drug Abuse
(R01DA053282 and R01DA054201 to A.B.K.). J.N. was supported by
the James S. McDonnell Foundation 21st Century Science Initiative in
Understanding Human Cognition—Scholar Award (#220020334) and
by a Sponsored Research Agreement between Meta and Fundación
Universidad Torcuato Di Tella (#INB2376941). The funders had no role
in study design, data collection and analysis, decision to publish or
preparation of the manuscript.
Author contributions
H.A. is the lead author and researcher responsible for study design,
coordination and management between teams, data management
and collection and analysis, visualization and writing of the paper.
S.P. was the main senior supervisor, who worked hand in hand
Article
with H.A. on every aspect of this work, including collaboration
management, design, hypothesis development, supervision of
the analysis, interpretation of the results, visualization and writing.
S.B. was the main author behind the original design that this study
replicated and contributed greatly to ensuring that our design
indeed reproduced theirs. F.B., D.B., F.C., M.C., M.G., E.J.G., D.K.,
M.K., G.L., M.S., J.Y and O.Z. reviewed and supported the design
of the experiment and its hypotheses. They also took charge of
translation and deployment of the experiment in each of their
countries, collected data locally and revised the paper. B.B., J.S.C.,
U.H., A.B.K., J.L., C.O., J.N., G.R., A.S.-J., A.S., B.S. and K.W. are senior
supervisors who monitored the study locally, providing insight on
the experimental design and commentary on the final version of the
paper. In addition, K.W. provided essential scientific and logistical
support in deploying the experiment worldwide.
Competing interests
The authors declare no competing interests.
https://doi.org/10.1038/s41562-024-01894-9
Correspondence and requests for materials should be addressed to
Hernán Anlló or Stefano Palminteri.
Peer review information Nature Human Behaviour thanks
Thomas J. Faulkenberry and the other, anonymous, reviewer(s) for
their contribution to the peer review of this work. Peer reviewer reports
are available.
Reprints and permissions information is available at
www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds
exclusive rights to this article under a publishing agreement with
the author(s) or other rightsholder(s); author self-archiving of the
accepted manuscript version of this article is solely governed by the
terms of such publishing agreement and applicable law.
Additional information
Supplementary information The online version contains supplementary
material available at https://doi.org/10.1038/s41562-024-01894-9.
1
© The Author(s), under exclusive licence to Springer Nature Limited
2024, corrected publication 2024
Human Reinforcement Learning Team, Laboratory of Cognitive and Computational Neuroscience, Paris, France. 2Faculty of Science and Engineering,
Waseda University, Tokyo, Japan. 3Intercultural Cognitive Network, Paris, France. 4General Psychology Lab, Hamburg University, Hamburg, Germany.
5
School of Collective Intelligence, Université Mohammed VI Polytechnique, Rabat, Morocco. 6Department of Psychiatry, University Behavioral Health
Care and Brain Health Institute, Rutgers University—New Brunswick, Piscataway, NJ, USA. 7Department of Cognitive Sciences, University of Haifa,
Haifa, Israel. 8Facultad de Psicología, Universidad del Desarrollo, Santiago de Chile, Chile. 9International Laboratory for Social Neurobiology, Institute
for Cognitive Neuroscience, HSE University, Moscow, Russia. 10Laboratorio de Neurociencia, Universidad Torcuato Di Tella, Buenos Aires, Argentina.
11
School of Psychological and Cognitive Sciences and Beijing Key Laboratory of Behavior and Mental Health, Peking University, Beijing, China. 12Centre for
Cognition and Decision Making, Institute for Cognitive Neuroscience, HSE University, Moscow, Russia. 13Department of Psychology, Ludwig Maximilian
University, Munich, Germany. 14IDG/McGovern Institute for Brain Research, Peking University, Beijing, China. 15Escuela de Negocios, Universidad Torcuato
Di Tella, Buenos Aires, Argentina. 16Consejo Nacional de Investigaciones Científicas y Técnicas, Buenos Aires, Argentina. 17School of Cognitive Sciences,
Institute for Research in Fundamental Sciences, Tehran, Iran. 18Department of Clinical Psychology, SRM Medical College Hospital and Research Centre,
Chennai, India. 19Departement d’études cognitives, Ecole normale supérieure, PSL Research University, Paris, France.
e-mail: hernan.anllo@cri-paris.org;
stefano.palminteri@ens.fr
Nature Human Behaviour
Last updated by author(s): Hernán Anlló (last edit 19/02/2024)
Reporting Summary
Nature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.
nature portfolio | reporting summary
Corresponding author(s): Stefano Palminteri, Hernán Anlló
Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.
A description of all covariates tested
A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)
For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.
For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.
Software and code
Policy information about availability of computer code
Data collection
Data was collected through the Gorilla.sc platform for online experiments (BUILD 20201002 )
Data analysis
Statistical analyses and figure production was handled with R (ver 4.2). Additional instances of model fitting and sample power estimations
were handled with Matlab (ver 2022b)
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.
Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
Data for the present study is available for free (for non-commercial use only), in our OSF.io repository (https://osf.io/yebm9/?view_only=).
April 2023
- Accession codes, unique identifiers, or web links for publicly available datasets
- A description of any restrictions on data availability
- For clinical datasets or third party data, please ensure that the statement adheres to our policy
1
Policy information about studies with human participants or human data. See also policy information about sex, gender (identity/presentation),
and sexual orientation and race, ethnicity and racism.
Reporting on sex and gender
Sex data was collected as part of a standard demographic survey, with the sole purpose of describing the tested population
and ensuring our sample was balanced. No hypotheses concerning predictions based on sex or gender were made.
Reporting on race, ethnicity, or Samples came from nationals of 11 different countries: USA, Iran, Israel, Chile, Argentina, Russia, Morocco, Japan, China,
India, France.
other socially relevant
groupings
Population characteristics
Age (mean(SD), in years) = 24.4(4.6). Gender (% of females) = 60.9.
Recruitment
Recruitment was open, conducted through the standard university channels of each of the participating institutions (flyers,
mailing lists). We are aware that this mode of recruitment implies an overrepresentation of university students in our sample
to which we admit. This implies that generalizing the results obtained in this study to a broader population (e.g. rural
populations, non-educated populations, elderly groups) cannot be done directly and requires further research targeting
these groups exactly,
Ethics oversight
This protocol was approved by the INSERM Ethical Review Committee/IRB00003888 on 13 November 2018
nature portfolio | reporting summary
Research involving human participants, their data, or biological material
Note that full information on the approval of the study protocol must also be provided in the manuscript.
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences
Behavioural & social sciences
Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf
Life sciences study design
All studies must disclose on these points even when the disclosure is negative.
Sample size
Describe how sample size was determined, detailing any statistical methods used to predetermine sample size OR if no sample-size calculation
was performed, describe how sample sizes were chosen and provide a rationale for why these sample sizes are sufficient.
Data exclusions
Describe any data exclusions. If no data were excluded from the analyses, state so OR if data were excluded, describe the exclusions and the
rationale behind them, indicating whether exclusion criteria were pre-established.
Replication
Describe the measures taken to verify the reproducibility of the experimental findings. If all attempts at replication were successful, confirm this
OR if there are any findings that were not replicated or cannot be reproduced, note this and describe why.
Randomization
Describe how samples/organisms/participants were allocated into experimental groups. If allocation was not random, describe how covariates
were controlled OR if this is not relevant to your study, explain why.
Blinding
Describe whether the investigators were blinded to group allocation during data collection and/or analysis. If blinding was not possible,
describe why OR explain why blinding was not relevant to your study.
Behavioural & social sciences study design
All studies must disclose on these points even when the disclosure is negative.
This is an online behavioral experiment that consists of completing a reinforcement learning task, an economics decision-making
task, and fill questionnaires concerning participants' cultural traits (e.g. collectivistic vs. individualistic tendencies). Thus, this
experiment gathers both quantitative and qualitative data.
Research sample
Full description of the research sample is available on Table 1 of the manuscript. Samples are representative of their country's
cultural traits (but see limitations outlined in Recruitment, above). In order to ensure that our samples would adequately represent
the culture of the country to which they belonged, inclusion criteria required that participants: (1) had the target country nationality,
(2) resided in the target country, (3) had completed at least the full basic education cycle in the target country, and (4) spoke the
country’s official language as their native language. These criteria were assessed for each participant during a video meeting prior to
launching the
experiment. The meeting, task instructions, and questionnaires were delivered in each country’s official language, by local
researchers. Rationale for this sample was to seek individuals both within and outside the WEIRD sphere. Number of countries
included was ultimately limited by our resources.
April 2023
Study description
2
Sample size was determined through a power analysis based on the behavioral results of Bavard et al, 2021. In the critical decision
context of said experiment (expected value difference between options = 1.75), online participants reached a difference between
choice rate and chance (0.5) of 0.27 +/- 0.30 (mean +/- SD). To obtain the same difference with a power of 0.95, the MATLAB
function "samsizepwr.m" indicated that 46 participants per country were needed. Samples were allowed to exceed this limit by up to
20%, to ensure the desired power would be achieved regardless of potential exclusions. These numbers were also verified through
simulation-based power analyses Sampling strategy was quota sampling (where each stratum was one of the selected countries).
Quotas were determined by sample size.
Data collection
Testing was conducted in a hybrid face-to-face/online format, where participants met a local experimenter for an online live debrief
held in their local language to verify identity and cultural affiliation. After the interview, participants received a personalized link to a
Gorilla server (www.gorilla.sc) where the experiment was hosted. After clicking on the link, participants were sent to a consent form,
which they had to complete in order to access the experiment. Once finished with the experiment, participants were given a
personalized completion code, and were tasked with sending this code to the experimenter by email to signal completion and trigger
payment. The online debrief, task instructions and questionnaires were all delivered in each country's official language, by local
researchers. Experimenters in charge of providing the links and conducting the testing were not blinded to all conditions, since the
main condition was belonging to a certain country, and that made blinding impossible in the context of our task. Said experimenters
were also aware that the main hypothesis consisted of comparing decision strategies across countries (without further detail). They
did not mention this to participants, who were unaware that the same experiment was being conducted in several countries
simultaneously.
Timing
Data collection started on 06/2020 and finished on 04/2022
Data exclusions
The only exclusion criterion established was task completion. Across all countries, 43 participants failed to complete the task and
were excluded. An additional 19 participants were excluded because of troubleshooting/translation issues during the online task
rollout.
Non-participation
Reasons for not finishing the online task were not provided by the participants (who just left their sessions open and never finished
the task, or closed it mid-way). This is a common occurrence in online experiments. A total of 43 participants out of 623 across all
countries dropped. An additional
Randomization
Participants were allocated to groups as afunction of the country they belonged to. This was done naturally, as separate recruitments
took place locally at each of the 11 countries involved.
nature portfolio | reporting summary
Sampling strategy
Ecological, evolutionary & environmental sciences study design
All studies must disclose on these points even when the disclosure is negative.
Study description
Briefly describe the study. For quantitative data include treatment factors and interactions, design structure (e.g. factorial, nested,
hierarchical), nature and number of experimental units and replicates.
Research sample
Describe the research sample (e.g. a group of tagged Passer domesticus, all Stenocereus thurberi within Organ Pipe Cactus National
Monument), and provide a rationale for the sample choice. When relevant, describe the organism taxa, source, sex, age range and
any manipulations. State what population the sample is meant to represent when applicable. For studies involving existing datasets,
describe the data and its source.
Sampling strategy
Note the sampling procedure. Describe the statistical methods that were used to predetermine sample size OR if no sample-size
calculation was performed, describe how sample sizes were chosen and provide a rationale for why these sample sizes are sufficient.
Data collection
Describe the data collection procedure, including who recorded the data and how.
Timing and spatial scale
Indicate the start and stop dates of data collection, noting the frequency and periodicity of sampling and providing a rationale for
these choices. If there is a gap between collection periods, state the dates for each sample cohort. Specify the spatial scale from which
the data are taken
Data exclusions
If no data were excluded from the analyses, state so OR if data were excluded, describe the exclusions and the rationale behind them,
indicating whether exclusion criteria were pre-established.
Reproducibility
Describe the measures taken to verify the reproducibility of experimental findings. For each experiment, note whether any attempts to
repeat the experiment failed OR state that all attempts to repeat the experiment were successful.
Randomization
Describe how samples/organisms/participants were allocated into groups. If allocation was not random, describe how covariates were
controlled. If this is not relevant to your study, explain why.
Blinding
Describe the extent of blinding used during data acquisition and analysis. If blinding was not possible, describe why OR explain why
blinding was not relevant to your study.
Yes
No
April 2023
Did the study involve field work?
3
Field conditions
Describe the study conditions for field work, providing relevant parameters (e.g. temperature, rainfall).
Location
State the location of the sampling or experiment, providing relevant parameters (e.g. latitude and longitude, elevation, water depth).
Access & import/export Describe the efforts you have made to access habitats and to collect and import/export your samples in a responsible manner and in
compliance with local, national and international laws, noting any permits that were obtained (give the name of the issuing authority,
the date of issue, and any identifying information).
Disturbance
Describe any disturbance caused by the study and how it was minimized.
Reporting for specific materials, systems and methods
nature portfolio | reporting summary
Field work, collection and transport
We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.
Materials & experimental systems
Methods
n/a Involved in the study
n/a Involved in the study
Antibodies
ChIP-seq
Eukaryotic cell lines
Flow cytometry
Palaeontology and archaeology
MRI-based neuroimaging
Animals and other organisms
Clinical data
Dual use research of concern
Plants
Antibodies
Antibodies used
Describe all antibodies used in the study; as applicable, provide supplier name, catalog number, clone name, and lot number.
Validation
Describe the validation of each primary antibody for the species and application, noting any validation statements on the
manufacturer’s website, relevant citations, antibody profiles in online databases, or data provided in the manuscript.
Eukaryotic cell lines
Policy information about cell lines and Sex and Gender in Research
Cell line source(s)
State the source of each cell line used and the sex of all primary cell lines and cells derived from human participants or
vertebrate models.
Authentication
Describe the authentication procedures for each cell line used OR declare that none of the cell lines used were authenticated.
Mycoplasma contamination
Confirm that all cell lines tested negative for mycoplasma contamination OR describe the results of the testing for
mycoplasma contamination OR declare that the cell lines were not tested for mycoplasma contamination.
Commonly misidentified lines
Name any commonly misidentified cell lines used in the study and provide a rationale for their use.
(See ICLAC register)
Palaeontology and Archaeology
Provide provenance information for specimens and describe permits that were obtained for the work (including the name of the
issuing authority, the date of issue, and any identifying information). Permits should encompass collection and, where applicable,
export.
Specimen deposition
Indicate where the specimens have been deposited to permit free access by other researchers.
April 2023
Specimen provenance
4
Dating methods
Tick this box to confirm that the raw and calibrated dates are available in the paper or in Supplementary Information.
Ethics oversight
Identify the organization(s) that approved or provided guidance on the study protocol, OR state that no ethical approval or guidance
was required and explain why not.
Note that full information on the approval of the study protocol must also be provided in the manuscript.
Animals and other research organisms
Policy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research, and Sex and Gender in
Research
Laboratory animals
For laboratory animals, report species, strain and age OR state that the study did not involve laboratory animals.
Wild animals
Provide details on animals observed in or captured in the field; report species and age where possible. Describe how animals were
caught and transported and what happened to captive animals after the study (if killed, explain why and describe method; if released,
say where and when) OR state that the study did not involve wild animals.
Reporting on sex
Indicate if findings apply to only one sex; describe whether sex was considered in study design, methods used for assigning sex.
Provide data disaggregated for sex where this information has been collected in the source data as appropriate; provide overall
numbers in this Reporting Summary. Please state if this information has not been collected. Report sex-based analyses where
performed, justify reasons for lack of sex-based analysis.
Field-collected samples
For laboratory work with field-collected samples, describe all relevant parameters such as housing, maintenance, temperature,
photoperiod and end-of-experiment protocol OR state that the study did not involve samples collected from the field.
Ethics oversight
Identify the organization(s) that approved or provided guidance on the study protocol, OR state that no ethical approval or guidance
was required and explain why not.
nature portfolio | reporting summary
If new dates are provided, describe how they were obtained (e.g. collection, storage, sample pretreatment and measurement), where
they were obtained (i.e. lab name), the calibration program and the protocol for quality assurance OR state that no new dates are
provided.
Note that full information on the approval of the study protocol must also be provided in the manuscript.
Clinical data
Policy information about clinical studies
All manuscripts should comply with the ICMJE guidelines for publication of clinical research and a completed CONSORT checklist must be included with all submissions.
Clinical trial registration Provide the trial registration number from ClinicalTrials.gov or an equivalent agency.
Study protocol
Note where the full trial protocol can be accessed OR if not available, explain why.
Data collection
Describe the settings and locales of data collection, noting the time periods of recruitment and data collection.
Outcomes
Describe how you pre-defined primary and secondary outcome measures and how you assessed these measures.
Dual use research of concern
Policy information about dual use research of concern
Hazards
Could the accidental, deliberate or reckless misuse of agents or technologies generated in the work, or the application of information presented
in the manuscript, pose a threat to:
No Yes
Public health
National security
Crops and/or livestock
Ecosystems
April 2023
Any other significant area
5
Does the work involve any of these experiments of concern:
No Yes
Demonstrate how to render a vaccine ineffective
Confer resistance to therapeutically useful antibiotics or antiviral agents
Enhance the virulence of a pathogen or render a nonpathogen virulent
Increase transmissibility of a pathogen
Alter the host range of a pathogen
Enable evasion of diagnostic/detection modalities
Enable the weaponization of a biological agent or toxin
Any other potentially harmful combination of experiments and agents
nature portfolio | reporting summary
Experiments of concern
Plants
Seed stocks
Report on the source of all seed stocks or other plant material used. If applicable, state the seed stock centre and catalogue number. If
plant specimens were collected from the field, describe the collection location, date and sampling procedures.
Novel plant genotypes
Describe the methods by which all novel plant genotypes were produced. This includes those generated by transgenic approaches,
gene editing, chemical/radiation-based mutagenesis and hybridization. For transgenic lines, describe the transformation method, the
number of independent lines analyzed and the generation upon which experiments were performed. For gene-edited lines, describe
the editor used, the endogenous sequence targeted for editing, the targeting guide RNA sequence (if applicable) and how the editor
was applied.
Describe any authentication procedures for each seed stock used or novel genotype generated. Describe any experiments used to
assess the effect of a mutation and, where applicable, how potential secondary effects (e.g. second site T-DNA insertions, mosiacism,
off-target gene editing) were examined.
Authentication
ChIP-seq
Data deposition
Confirm that both raw and final processed data have been deposited in a public database such as GEO.
Confirm that you have deposited or provided access to graph files (e.g. BED files) for the called peaks.
May remain private before publication.
For "Initial submission" or "Revised version" documents, provide reviewer access links. For your "Final submission" document,
provide a link to the deposited data.
Files in database submission
Provide a list of all files available in the database submission.
Genome browser session
Provide a link to an anonymized genome browser session for "Initial submission" and "Revised version" documents only, to
enable peer review. Write "no longer applicable" for "Final submission" documents.
Data access links
(e.g. UCSC)
Methodology
Replicates
Describe the experimental replicates, specifying number, type and replicate agreement.
Sequencing depth
Describe the sequencing depth for each experiment, providing the total number of reads, uniquely mapped reads, length of reads and
whether they were paired- or single-end.
Antibodies
Describe the antibodies used for the ChIP-seq experiments; as applicable, provide supplier name, catalog number, clone name, and
lot number.
Peak calling parameters Specify the command line program and parameters used for read mapping and peak calling, including the ChIP, control and index files
used.
Data quality
Describe the methods used to ensure data quality in full detail, including how many peaks are at FDR 5% and above 5-fold enrichment.
Software
Describe the software used to collect and analyze the ChIP-seq data. For custom code that has been deposited into a community
repository, provide accession details.
April 2023
6
Plots
Confirm that:
The axis labels state the marker and fluorochrome used (e.g. CD4-FITC).
The axis scales are clearly visible. Include numbers along axes only for bottom left plot of group (a 'group' is an analysis of identical markers).
All plots are contour plots with outliers or pseudocolor plots.
A numerical value for number of cells or percentage (with statistics) is provided.
Methodology
Sample preparation
Describe the sample preparation, detailing the biological source of the cells and any tissue processing steps used.
Instrument
Identify the instrument used for data collection, specifying make and model number.
Software
Describe the software used to collect and analyze the flow cytometry data. For custom code that has been deposited into a
community repository, provide accession details.
Cell population abundance
Describe the abundance of the relevant cell populations within post-sort fractions, providing details on the purity of the
samples and how it was determined.
Gating strategy
Describe the gating strategy used for all relevant experiments, specifying the preliminary FSC/SSC gates of the starting cell
population, indicating where boundaries between "positive" and "negative" staining cell populations are defined.
nature portfolio | reporting summary
Flow Cytometry
Tick this box to confirm that a figure exemplifying the gating strategy is provided in the Supplementary Information.
Magnetic resonance imaging
Experimental design
Design type
Indicate task or resting state; event-related or block design.
Design specifications
Specify the number of blocks, trials or experimental units per session and/or subject, and specify the length of each trial
or block (if trials are blocked) and interval between trials.
Behavioral performance measures
State number and/or type of variables recorded (e.g. correct button press, response time) and what statistics were used
to establish that the subjects were performing the task as expected (e.g. mean, range, and/or standard deviation across
subjects).
Acquisition
Imaging type(s)
Specify: functional, structural, diffusion, perfusion.
Field strength
Specify in Tesla
Sequence & imaging parameters
Specify the pulse sequence type (gradient echo, spin echo, etc.), imaging type (EPI, spiral, etc.), field of view, matrix size,
slice thickness, orientation and TE/TR/flip angle.
Area of acquisition
State whether a whole brain scan was used OR define the area of acquisition, describing how the region was determined.
Diffusion MRI
Used
Not used
Preprocessing
Provide detail on software version and revision number and on specific parameters (model/functions, brain extraction,
segmentation, smoothing kernel size, etc.).
Normalization
If data were normalized/standardized, describe the approach(es): specify linear or non-linear and define image types used for
transformation OR indicate that data were not normalized and explain rationale for lack of normalization.
Normalization template
Describe the template used for normalization/transformation, specifying subject space or group standardized space (e.g.
original Talairach, MNI305, ICBM152) OR indicate that the data were not normalized.
Noise and artifact removal
Describe your procedure(s) for artifact and structured noise removal, specifying motion parameters, tissue signals and
physiological signals (heart rate, respiration).
April 2023
Preprocessing software
7
Volume censoring
Define your software and/or method and criteria for volume censoring, and state the extent of such censoring.
Model type and settings
Specify type (mass univariate, multivariate, RSA, predictive, etc.) and describe essential details of the model at the first and
second levels (e.g. fixed, random or mixed effects; drift or auto-correlation).
Effect(s) tested
Define precise effect in terms of the task or stimulus conditions instead of psychological concepts and indicate whether
ANOVA or factorial designs were used.
Specify type of analysis:
Statistic type for inference
Whole brain
ROI-based
Both
Specify voxel-wise or cluster-wise and report all relevant parameters for cluster-wise methods.
(See Eklund et al. 2016)
Correction
Describe the type of correction and how it is obtained for multiple comparisons (e.g. FWE, FDR, permutation or Monte Carlo).
nature portfolio | reporting summary
Statistical modeling & inference
Models & analysis
n/a Involved in the study
Functional and/or effective connectivity
Graph analysis
Multivariate modeling or predictive analysis
Functional and/or effective connectivity
Report the measures of dependence used and the model details (e.g. Pearson correlation, partial correlation,
mutual information).
Graph analysis
Report the dependent variable and connectivity measure, specifying weighted graph or binarized graph,
subject- or group-level, and the global and/or node summaries used (e.g. clustering coefficient, efficiency,
etc.).
Multivariate modeling and predictive analysis Specify independent variables, features extraction and dimension reduction, model, training and evaluation
metrics.
April 2023
8