# Cognitive processes across the globe

## Study Information

### 1. Title

Cognitive processes across the globe: Webcam-based eye-tracking analysis of ingroup bias

### 2. Authorship

Rima-Maria Rahal^1^ & Frederik Schulze Spüntrup^2^

^1^ Max Planck Institute for Research on Collective Goods<br>
^2^ Institute for Globally Distributed Open Research and Education (IGDORE)

### 3. Description
We study the cognitive mechanisms underlying ingroup favoritism, comparing decision processes captured via eye-tracking in several diverse societies. Participants decide whether to benefit in- or outgroup members at a cost to themselves, or to maximize their own payoffs. Eye gaze captured during these decisions is analyzed.

Whether and to what degree culture modifies cognition has been an area of research often limited by possibilities to gather relevant data across societies. In this project, we leverage the possibilities of webcam-based eye-tracking to study cultural variations of cognitive processes underlying ingroup favoritism. We aim to sample 100 participants each from 17 societies using Prolific and MTRUK. Participants are assigned to an ingroup based on a color perception task, complete a group reinforcement stage and then make decisions to allocate points between themselves and random matched players in a repeated decomposed dictator game, where they face either an in- or an outgroup member. During this task, we record eye-gaze with participants’ webcams. We expect that ingroup favoritism in choices depends on culture, such that in more collectivist societies, decision makers make more prosocial decisions and are biased towards more prosociality when facing an in- rather than an outgroup member. This behavioral ingroup favoritism effect is expected to be mirrored in gaze data: in more collectivist societies, decision makers facing an in- rather than an outgroup member show higher fixation counts and longer dwell times, attend to a higher number of available information and fixate less on their own outcomes, while these differences diminish in more individualistic societies. We additionally predict that in more collectivist societies, decision makers are more likely to seek out information about others’ group membership, both in terms of visual attendance during the decomposed dictator game, and explicitly in a final dictator game task where participants are asked to explicitly state if they prefer to learn about the matched players’ group membership. Further, we explore the relation of other individual- and country-level predictors previously related to ingroup favoritism with eye-gaze.

### 4. Hypotheses

Visualization: https://docs.google.com/presentation/d/1Mzk_eVEEYEhdQyvgBL9KAOs1udKuuuBV7nsyXpKcYRY/edit?usp=sharing

#### Individual-Level Predictors: Choice Behavior

1. Hypothesis I1 **Main effect of SVO:** Higher likelihood of making prosocial rather than selfish choices for people with higher deviations from pure individualism towards altruistic Social Value Orientation (SVO).
2. Hypothesis I2 **Main effect of decision setting:** Higher likelihood of making prosocial rather than selfish choices in decisions vis á vis an ingroup rather than an outgroup member.
3. Hypothesis I3 **Interaction of SVO x decision setting:** Increasing likelihood of making prosocial rather than selfish choices between in- vs. outgroup decision setting, with increasing deviation from pure individualism towards altruistic SVO.

#### Individual-Level Predictors: Gaze Behavior
4. Hypotheses I4 **Interaction of decision setting x SVO** (Replication of Rahal et al., 2020):
1. I4a **on decision times:** While individualists take longer to make in- vs. outgroup decisions, this difference in decision times is decreased with increasing deviation from pure individualism towards altruistic SVO.
2. I4b **on number of fixations:** While individualists show a higher number of fixations in in- vs. outgroup decision settings, this difference in the number of fixations is decreased with increasing deviation from pure individualism towards altruistic SVO.
3. I4c **on amount of inspected information:** While individualists show a higher amount of inspected information in in- vs. outgroup decision settings, this difference in the amount of information inspected is decreased with increasing deviation from pure individualism towards altruistic SVO.
4. I4d **on proportion of attention to own outcomes:** Individualists focus proportionately more on own outcomes than more prosocial decision makers. In in- vs. outgroup decision settings, the proportion of attention to own outcomes decreases overall. The difference in the proportion of attention allocated to own outcomes between in- and outgroup decisions increases with increasing deviation from pure individualism towards altruistic SVO.
5. Hypothesis I5: Participants' **explicit preference** for uncovering information about others' group membership predicts a higher likelihood of **freely gazing** at information about others' group membership.
6. Hypothesis I6: Participants who **identify more strongly** with the ingroup and **like the ingroup more** are more likely to **freely gaze** at others' group membership.

#### Country-Level Predictors: Choice Behavior

1. Hypothesis C1 **Religiosity x Decision Setting**: The difference in the probability of making a prosocial rather than selfish choice when facing the in- vs. outgroup decreases as the rated importance of religiosity in life increases.
2. Hypothesis C2 **Governmental Effectiveness x Decision Setting**: The difference in the probability of making a prosocial rather than selfish choice when facing the in- vs. outgroup decreases as governmental effectiveness increases.
3. Hypothesis C3 **Disease Burden x Decision Setting**: The difference in the probability of making a prosocial rather than selfish choice when facing the in- vs. outgroup increases as the disease burden increases.
4. Hypothesis C4 **COVID Burden x Decision Setting**: The difference in the probability of making a prosocial rather than selfish choice when facing the in- vs. outgroup increases as the COVID burden increases.
5. Hypothesis C5 **Individualism x Decision Setting**: The difference in the probability of making a prosocial rather than selfish choice when facing the in- vs. outgroup decreases as individualism increases.

#### Country-Level Predictors: Gaze Behavior
6. Hypotheses C6 **Interaction of decision setting x country-level predictors** (Ingroup Favoritism in Choices is mirrored in Gaze Behavior):
1. C6a (1 to 5) **on decision times:** The difference in decision times between decisions facing an in- vs. outgroup (where ingroup decisions take longer) increases as (1) religiosity decreases, (2) governmental effectiveness decreases, (3) disease burden and (4) COVID burden increase, and (5) individualism decreases.
2. C6b (6 to 10) **on number of fixations:** The difference in fixation counts between decisions facing an in- vs. outgroup (where fixation counts are higher in ingroup decisions) increases as (1) religiosity decreases, (2) governmental effectiveness decreases, (3) disease burden and (4) COVID burden increase, and (5) individualism decreases.
3. C6c (11 to 15) **on amount of inspected information:** The difference in the amount of information inspected between decisions facing an in- vs. outgroup (where more information is inspected in ingroup decisions) increases as (1) religiosity decreases, (2) governmental effectiveness decreases, (3) disease burden and (4) COVID burden increase, and (5) individualism decreases.
4. C6d (16 to 20) **on proportion of attention to own outcomes:** The difference in the proportion of attention allocated to own outcomes between in- and outgroup decisions (where more attention is allocated to own outcomes when facing an out- rather than an ingroup member) increases as (1) religiosity decreases, (2) governmental effectiveness decreases, (3) disease burden and (4) COVID burden increase, and (5) individualism decreases.
7. Hypothesis C7 (1 to 5): Participants are more like to **freely gaze** at others’ group membership if they are from countries with lower religiosity, lower governmental effectiveness, higher disease burden, higher COVID burden, and lower individualism.

#### Meta-Analytic Effects
1. Hypothesis M1: We predict between-country heterogeneity in the size of the effect of the in- vs. outgroup decision setting on the probability of making a prosocial decision, albeit predicting an overall significant effect of ingroup favoritism.
2. Hypothesis M2a-d: We predict between-country heterogeneity in the size of the effect of the in- vs. outgroup decision setting on (a) decision times, (b) fixation counts, (c) the amount of information inspected an (d) the proportion of attention allocated to own outcomes (negative), albeit predicting an overall significant effect of ingroup favoritism.

## Design Plan

### 5. Study Type

Experiment

### 6. Blinding

Participants, will not know the treatment group to which they have been assigned.

### 7. Is there any additional blinding in this study?

No.

### 8. Study design
Participants take part in a 17 (countries; endogenous variable, between subjects) x 2 (group membership: in vs. outgroup; manipulated variable, within subjects) x SVO (continuous; endogenous variable, between subjects) mixed design.

### 9. Randomization
In each trial of the main task, participants face a decision task were the other player is either an in- or an outgroup member (within subjects).

In the main task, we fully counterbalance the display order of items on the screen between subjects. Own payoff is randomly either displayed in the top or bottom row, with other payoff in the corresponding row. Group membership is ramdonly either displayed in the top or bottom row, with random number in the corresponding row. The column in which payoffs for option 1, option 2, or the other player information is displayed is also determined at random.

In the final dictator game, randomly, participants face a decision task were the other player is either an in- or an outgroup member, or the group membership of the other player is not disclosed.

## Sampling Plan

Participants are recruited in spring 2023. Participants are invited to sign up if they are between 18 and 35 years old, are not wearing glasses during the experiment, speak English well, and can complete the study on a desktop computer and have access to their webcam. Where made possible through the participant pool provider, we choose a balanced distribution of the study to participants in terms of gender.

For each country from which we sample, we run a separate sub-experiment to ensure achieving 100 participants per country.

Participants on receive a fixed payment of £2.50, and a variable payment for the incentivized parts of the experiment (SVO slider: 100 points worth £0.30, group reinforcement task: £0.5 bonus for winning team, main decision task: 100 points worth £0.30, final dictator game: 100 points worth £0.30). The study is planned to take 20 min, with an expected payment of £3.48, which corresponds to an hourly expected wage at app. £10.44. (minimum wage in the UK). For each country, we adapt the payment information based on the native currency, which may lead to fluctuations due to currency exchange rates.

We invite participants to join from the following 17 countries:

- Afghanistan
- Australia
- Bangladesh
- Chile
- Germany
- Greece
- Hungary
- India
- Italy
- Mexico
- Philippines
- Poland
- Portugal
- South Africa
- Spain
- United Kingdom
- United States

Afghanistan, Bangladesh, India and Philippines are sampled via MTRUK. All other countries are sampled via Prolific.

### 10. Existing data

### 11. Explanation of existing data

### 12. Data collection procedures

First, participants give indicate informed consent and give permission to use their webcam. They then indicate their age and gender, and complete the 6-item version of the SVO slider task (Murphy, Ackermann, & Handgraaf, 2011).

In the group allocation task, participants see 8 colors bars (2 clearly blue, 2 clearly green, and 4 mixed blue-green tones) in random order, for each of which they indicate whether they perceive them as green or blue. Based on which color they indicated more often, they are allocated to Team Green or Team Blue, and informed of their group membership.

In the group reinforcement task, participants complete a 12-trial reaction time competition, aiming to press the correct key in response to a yellow, purple or blue-green star appearing in random order as quickly as possible. Correct responses earn 10 points, while false responses lose 10 points. Participants are informed that they compete against a player from the other team and that they can win 100 bonus points for completing the task faster than the other player. Whichever team has more points in the end will earn a bonus payment. At the end of the task, participants are shown how fast they were and how many correct responses they gave, but whether they won will only be calculated after the study has taken place.

In the subsequent main task, participants face 60 trials (2 filler) of deciding between two options (one selfish and one prosocial option). They are shown how much they would earn and how much the other player would ear for each option. In addition, they receive two pieces of information about the other player: whether they are an in- or outgroup member (1 or 0) and a random number assigned to this person (between 10 and 89). They complete three checks for understanding, in which they indicate for an example trial how many points the other player would get if they chose option f, whether the other player in this example was an ingroup member, and how many points they would get if they chose option j. Following each question, the correct answer is displayed for all participants.

Next, participants are instructed for eye-tracking, and a 9-point calibration procedure, followed by a 9-point validation procedure takes place.

Thereafter, participants are reminded about the main task, and complete four practice trials. Before moving on to the main task, a calibration (9 points) and validation (4 points) procedure is run, which is repeated before trials 12, 24, 36, and 48 . Participants complete all 60 trials of the main task, during which eye-data is recorded.

Participants then face another dictator game, where, before deciding how to allocate 100 points between themselves and another player, they indicate whether they would like to learn the other players' group membership (yes or no). Regardless of the answer, they are randomly presented either with an in- or outgroup member, or with a participant with unknown group membership. They then decide how many points to allocate to this person, while keeping the remainder to themselves.

Further, we elicit expectations of preferential treatment from ingroup members in the final dictator game (yes vs. no). Next, we elicit expectations about others' behavior in this final dictator game. Participants indicate how many of 100 participants in this study they think:
- want to know the other players' group membership
- give more points to participants from the same team
- expect to get more points from participants from the same team.

We then run a manipulation check and ask participants to indicate which team they belonged to.

To elicit identification with both in- and outgroup, participants complete an adapted version of the Inclusion of Other in the Self Scale (Aaron et al., 1991). Then, they indicate their agreement with the following questions regarding the in-and outgroup, respectively, on a 7-point Likert scale:
- Attitude towards the groups
- I see myself as a member of TEAM GREEN [TEAM BLUE],
- Liking of the groups
- I like TEAM GREEN [TEAM BLUE].

Participants then complete an 8-item shortened version of the horizontal and vertical individualism and collectivism scale (Triandis & Gelfand, 1998), where they answer on a 9-point scale (1 = never to 9 = always).

Finally, we assess whether participants wore glasses during the study (yes or no), whether English is their native language (yes or no), in which country they spent the most time before turning 18 and in which country they currently live.

In an open response field, participants can provide feedback to the experiment in the end.

### 13. Sample size
We aim to collect data from 1700 participants, 100 from each country.

### 14. Sample size rationale
Sample size considerations are based on a power analysis using G*Power, where we used the model for repeated measures ANOVA (within-between interatcion) because logistic models with repeated measurements are not available. A meta-analysis by Balliet et al. (2014) shoed a small effect for ingroup favoritism through categorization in dictator games (d = 0.19). In the power analysis, we used this effect size (equivalent to f = 0.095), for 2 groups (in- vs. outgroup) and 29 measurements (per group), setting alpha = 0.05, correlation among the repeated measures = 0.5 and eta = 1. To achieve a power of 99%, data from 88 participants would be necessary.

We anticipate that the effect size of differences in eye-gaze may be smaller than the behavioral effect, and therefore aim to collect data from 100 participants per country. A sensitivity analysis suggests that this would afford detecting an effect of size f = 0.064 (d=0.128) with 80% power.

Further, sample size is restrained by feasibility concerns for an international data collection.

### 15. Stopping rule
Data collection continues for 2 weeks (per country), or until the desired sample size is reached, which ever is reached first. If after 1 week, the sample is not full, we remove the age and vision aid restrictions on the sample. We continue to invite participants until 100 completed responses (i.e., not counting partial responses) are obtained or the temporal cut-off kicks in.

## Variables

### 16. Manipulated variables
- Group membership of the matched player (in- or outgroup)

### 17. Measured variables

Behavior in main task per trial:
- Prosocial decision (0 or 1)

Gaze Behavior in main task per trial:
- Decision time (time between trial onset and logged-in decision)
- Number of fixations (fixation count to specific AOI)
- Fixation durations
- Number of inspected information (AOIs fixated at least once)
- Proportion of fixations to others' outcomes relative to own outcomes
- Proportion of fixation to others' group membership relative to random number

We define six equally sized, non-overlapping Areas of Interest (AOIs) for analyzing the gaze data directed to the four payoff values and two other player values. These AOIs are defined as 25.7% x 28% screen width x height set in a rectangle around the respective AOI midpoint. Between AOIs, a margin of 5.65% horizontally and 10% vertically is in place.

Further, gazes are categorized as directed towards labels (Option F, Option J, etc.). Labels are defined as 25.7% x 9.16% screen width x height set in a rectangle around the respective label midpoint.

All other gazes are defined as outside of relevant areas.

We define fixations as gazes with a duration of >50ms and a dispersion of >3% of the screen width and height.

Futher Variables:
- Country
- SVO
- Explicit preference to uncover group information
- Points given in final dictator game
- Expectations of preferential treatment
- Expectations of others' interest to uncover group information
- Expectations of others' preferential treatment
- Expectations of others' expectation of preferential treatment
- Identification with ingroup
- Liking of both groups
- Attitutde towards both groups
- Horizontal and vertical individualism and collectivism
- Glasses worn
- English is native language
- Childhood country
- Current country
- Proportion of trials in which group membership was freely looked up relative to all trials completed

### 18. Indices
#### Individual-level variables:
- SVO: calculated as in Murphy et al., 2011
- Preferential liking for ingroup: difference between the average scores of liking items towards in- and outgroup
- Preferential attitude towards ingroup: difference between the average scores of attitude items towards in- and outgroup
- Horizontal and vertical individualism and collectivism: mean scores for items of the HI, HC, VI and VC scales.

#### Country-level variables:
We use country-level data of:
- Religiosity: World Value Survey wave 2017-2022; item "religion is important in life"; https://www.worldvaluessurvey.org/WVSOnline.jsp?WAVE=4&COUNTRY=495&WAVE=4&COUNTRY=495
- Government effectiveness: World Bank Governance Indicator wave 2021; item government effectiveness; https://info.worldbank.org/governance/wgi/
- Burden of Disease: Global Burden of Disease Collaborative Network. Global Burden of Disease Study 2019; Disability-Adjusted Life Years per 100,000; https://ourworldindata.org/grapher/dalys-rate-from-all-causes?tab=table
- Burden of COVID: WHO COVID-19 Dashboard (May 3, 2023); Total confirmed deaths due to COVID-19 per million people; https://ourworldindata.org/grapher/total-covid-cases-deaths-per-million?tab=table
- Individualism: IDV scores reported in Hofstede (2001); https://clearlycultural.com/geert-hofstede-cultural-dimensions/individualism/

## Analysis Plan

### 19. Statistical models

#### Individual-Level Predictors: Choice Behavior
Hypotheses I1 to I3:
Logistic multilevel model predicting prosocial decisions with fixed effects of ingroup*svo, and with random effects of participant and item, and controlling for a fixed effect of the trial number, the platform on which participants were recruited, and whether they speak English as a second language.

For this analysis we run an additional santiy check, comparing trials where group information was fixated (where ingroup favoritism should be present) with trials where it was not (and consequently, no ingroup favoritism can emerge).

#### Individual-Level Predictors: Gaze Behavior
Hypotheses I4a to I4d:
Linear multilevel models predicting (a) decision times, (b) fixation counts, (c) number of inspected information and (d) proportion of attention to own outcomes with fixed effects of ingroup*svo, and with random effects of participant and item, and controlling for a fixed effect of the trial number, the platform on which participants were recruited, and whether they speak English as a second language.

For this analysis we exlcude all trials where no gaze (regardless of the temporal cut-off) was directed at group information.

Hypothesis I5 and I6:
Logistic multilevel model predicting whether group information was fixated with fixed effects of svo, diff_like, diff_attitude, and with random effects of participant and item, and controlling for a fixed effect of the trial number, the platform on which participants were recruited, and whether they speak English as a second language.

For this analysis we drop the temporal criterion for fixations and also include gazes with durations below 50ms.

#### Country-Level Predictors: Choice Behavior
Hypotheses C1 to C5:
Logistic multilevel model predicting whether a prosocial decision was made with fixed effects of ingroup*government_effectiveness, ingroup * individualism, ingroup * religiosity, ingroup *pathogen_stress, ingroup * covid_stress, and with random effects of country, participant, and item, and controlling for a fixed effect of the trial number, the platform on which participants were recruited, and whether they speak English as a second language.

#### Country-Level Predictors: Gaze Behavior
Hypotheses C6a to C6d (1 to 20):
Linear multilevel models predicting (a) decision times, (b) fixation counts, (c) proportion of information attended to, and (d) proportion of attention to own outcomes with fixed effects of ingroup*government_effectiveness, ingroup * individualism, ingroup * religiosity, ingroup *pathogen_stress, ingroup * covid_stress, and with random effects of country, participant, and item, and controlling for a fixed effect of the trial number, the platform on which participants were recruited, and whether they speak English as a second language.

For this analysis, we exlcude all trials where no gaze (regardless of the temporal cut-off) was directed at group information.

Hypothesis C7 (1 to 5):
Logistic multilevel model predicting whether group information was fixated with fixed effects of svo, diff_like, diff_attitude and expl_grouplook, government effectiveness, individualism, religiosity, pathogen_stress and covid_stress, and with random effects of country, participant and item, and controlling for a fixed effect of the trial number, the platform on which participants were recruited, and whether they speak English as a second language.

For this analysis we drop the temporal criterion for fixations and also include gazes with durations below 50ms.

#### Meta-Analytic Effects: Ingroup Favoritism in Choices
We use metafor to fit a random effects meta-analytic models, where each country is used as a study. For each country, we extract the effects size estimates and sampling variance for the fixed effect of ingroup from a logistic multilevel model predicting whether a prosocial decision was made with fixed effects of ingroup * svo, with random effects of participant and item, controlling for a fixed effect of the trial number, the platform on which participants were recruited, and whether they speak English as a second language.

#### Meta-Analytic Effects: Ingroup Favoritism in Gazes

We use metafor to fit a random effects meta-analytic models, where each country is used as a study. For each country, we extract the effects size estimates and sampling variance for the fixed effect of ingroup from linear multilevel models predicting (a) decision times, (b) fixation counts, (c) number of inspected information and (d) proportion of attention to own outcomes with fixed effects of ingroup*svo, and with random effects of participant and item, and controlling for a fixed effect of the trial number, the platform on which participants were recruited, and whether they speak English as a second language.

### 20. Transformations

We log-transform decision time data if the data deviates from a normal distribution (as indicated by a Shapiro–Francia test of normality).

Continuous predictors in regressions with interaction terms are centered.

### 21. Inference criteria
p < 0.05

### 22. Data exclusion
Data from participants who fail to indicate correctly which team they were allocated to are excluded from the analyses.

Data from filler trials, where there is no difference between the options for the other players' outcomes, are excluded from the analyses.

Trials will be excluded in case more than 75% of the fixations are not allocated to one of the pre-defined areas of interest (AOI) or directed at the labels. Moreover, we exclude trials shorter than 200 ms and longer than 3 standard deviations above individual mean response time.

### 23. Missing data
Data from participants where no gaze recordings were logged is excluded from the analyses.

### 24. Exploratory analysis
Because participants' display sizes and seating positions from the screen, as well as other hardware conditions will vary, it is difficult to anticipate necessary data handling and exclusion steps. We will explore if the results remain qualitatively unaffected if we do the following:
- We explore if the results remain robust to a different definition of AOIs as 14% x 14% screeen width x height rectangles, where labels are defined as 14% x 11.775% (top row) and 14% x 14% (bottom row). This AOI definition would minimize false positives but would excluse many gazes falling outside of the small areas, such that the exclusion criterion to remove trials with more than 75% of fixations outside of AOIs and labels would not be used.
- We explore if the results remain robust to an absolute definition of dispersion (as 30px), AOIs (as 300px x 200px) and labels (as 300px x 75px). These definitions would be unresponsive to the varying screen sizes participants may use, but it would ensure that the same pixel sizes are applied to all participants.
- In addition to the exclusions registered above, we will assess if excluding trials with problematic fixation data sensu Yang & Krajbich (2021) leads to different conclusions. Problematic trials are trials where either more than 80% of fixations are directed at the center of the screen or the sampling interval is higher than 750ms (10 times larger than expected given pilot data in a similar experiment).
- We explore if data quality differs between participants who indicate having worn glasses during the study and those without glasses.
- We explore if results remain qualitatively unaffected when using raw gaze data instead of fixation because of concerns over system latency. If there is delay in processing the webcam data, it could affect the accuracy of fixation detection and introduce further latency to the fixation durations.

## Other

### 25. Other

[](https://hackmd.io/Ll0Iedx2S76VDL27uYD4rQ)