'FIFA Experiments 2019'
(AsPredicted #31,014)
Download PDF

---

---

1) Have any data been collected for this study already?
No, no data have been collected for this study yet.

2) What's the main question being asked or hypothesis being tested in this study?
We expect that levels of discrimination vary by incentive structure (e.g. whether participants rank players or give ratings as well as whether participants get paid according to the strength of their team). Furthermore, we hypothesize that participants discriminate more in the games where they have to make decisions in a low information environment (“pick one”, “lottery”).

3) Describe the key dependent variable(s) specifying how they will be measured.
We use several outcome variables. In our case, discrimination consists of the systematic preference for lighter skin colour. Depending on the game, the outcome variable is a rating of a player’s skills or a preference for one player over one or a group of other players.

4) How many and which conditions will participants be assigned to?
Participants will be randomly allocated to different conditions.
At the meta-level, we indicate that the initial team is either “already quite good” or “has much room for improvement”. The actual strength of the three players is fixed at an average of 75.
At the meta-level, we either show all 4 skill indicators or only a random subset of 3 of them. With only three skill indicators shown, decision-making occurs under uncertainty.
At the meta-level, we either offer a certain pay-off (€0.50 plus average team strength at the end), or an uncertain pay-off that depends on beating a rival team (€1.50 in winning, €0.50 if losing). In both instances, the possible range of pay-offs is €0.50 to €1.50. In the second variant, the probability of winning depends on the average team strength at the end. For the randomly selected control group, we don’t offer any financial incentive on top of the 0.50€ participation reimbursement.
At the meta-level, the initial three players vary in skin colour.
There are five different rounds (“rating”, “paired conjoint”, “auction”, “pick one”, “lottery”).
In each round, participants are asked to rate or select players drawn from a random sample of about 500 portraits of real football players in Germany. The probability of players being sampled depends on their skin colour, which is automatically computed using the pixel composition of the portraits.
Rating: participants rate the average skill of 3 players, seeing one player at a time (that the player rated closest to his actual strength will be added to the team will be revealed ex-post, to avoid strategic miss-rating of weaker players).
Paired conjoint: participants twice pick one of two players offered, seeing both players at the same time.
Auction: participants pick one out of 5 players. The players are shown one at a time, and the participants choose whether to pick the player or not. If a player is picked, the remaining players are not shown. If the first 4 players are not picked, the 5th player is automatically picked.
Pick one: participants twice pick one out of 5 players. The 5 players are shown all at the same time.
Lottery: participants twice pick either a player with (partially) known skills, or a player with unknown skills, where the probability of the player having higher and lower skills is shown.

5) Specify exactly which analyses you will conduct to examine the main question/hypothesis.
We will analyse each manipulation independently, using regression analysis. The manipulations at the meta-level are additional predictors for each of the five rounds. We will use ordinary least squares (OLS) regression models, with clustered standard errors to account for answers given by the same participant.
The outcome variables will be the rating, or whether a player was picked. The predictor variables will be the manipulation. We will run separate models for each of the rounds. In the basic models, we only include the outcome (rating/picked) and the manipulation of the round, plus relevant control variables (the other players shown to the participants). In each instance, we also include a model with the meta-level manipulations as additional predictor variables, and a model where we control for participant-level variables:
• age as continuous variable
• gender with two or three categories [three categories if more than 50 participants choose ‘other’]
• level of education as categorical [6 categories, unless fewer than 50 participants fall into any of the categories, in which case we combine adjacent categories]
• German nationality [binary variable]
• closest political party [10 categories, unless fewer than 50 participants fall into any of the categories, in which case we code these parties as ‘other’]
• interest in football as continuous variables [5 categories treated as continuous]
• days playing computer-sports games as continuous variable
• frequency of watching football games [4 categories, unless fewer than 50 participants fall into any of the categories, in which case we combine adjacent categories]
• contact with foreigners as continuous variables [2 variables]
• preference for German workers [4 categories, unless fewer than 50 participants fall into any of the categories, in which case we combine adjacent categories]
• German Bundesland based on IP location

6) Describe exactly how outliers will be defined and handled, and your precise rule(s) for excluding observations.
We time all participants. Participants using less than 1/3 of the median time overall and for each screen will be considered “speeders”, and we run separate models to check whether their inclusion in the models has a substantial impact on the results.
Secondly, we time the authors playing as fast as they reasonably can as an alternative benchmark. Participants using less than 80% of this benchmark overall and for each screen will be considered “speeders”, and we run separate models to check whether their inclusion in the models has a substantial impact on the results.
We will remove participants with invalid IP addresses, and participants who – despite existing security checks – manage to take the survey more than once.

7) How many observations will be collected or what will determine sample size?
No need to justify decision, but be precise about exactly how the number will be determined.
We have determined N=9,630 as the minimum to get enough statistical power for all the conditions, but will use a larger sample if we receive additional funding. We assumed an effect size on a continuous scale 0-100: 3%, based on results from an existing experiment where we used a similar setup, and also Bryson and Chevalier 2015 Labour Economics. We targeted alpha = 0.05; power = 0.8; SD = 0.175.

8) Anything else you would like to pre-register?
(e.g., secondary analyses, variables collected for exploratory purposes, unusual analyses planned?)
We will conduct the survey experiment via Clickworker (self-administered). As a fixed rate we pay a standard fee of €0.50 to each respondent who completes the survey (see fee-recommendations < https://www.clickworker.com/survey-participants-for-online-surveys>), which should pay above the minimum wage in Germany. Together with the incentive payments (expected payoff = ((0.75 * 1) + 0.75 + 0)/3 = €0.50) this results in an expected payoff per unit of €1.