Interactive versus non-interactive implementations of "Dictator Games"

The main hypotheses concern fundamental differences in giving decisions between the two protocols: The non-interactive protocol predicts intermediate payments if a subject has some distributional and altruistic concern for the recipient's material well-being. By contrast, Nash equilibria of the games induced by the interactive protocol are characterized by extremal payments (either full or null payments), and in particular by increased levels of zero-giving for low prices of redistribution.

Additional hypotheses can be tested by varying the 'price of redistribution' p, that is, the factor by which each unit of material payoff that is given up is multiplied and sent to the recipient: in the interactive protocol, the Nash equilibrium prediction is that the peak at null (full) decreases (increases) in p.

We do a standard economic experiment conducted online. Recruitment will be done through a pre-registered subject pool from Amazon's Mechanical Turk, as used by the Decision Science Laboratory of ETH Zurich. Earnings will include a fixed show-up fee, and a variable bonus that depends on decisions of the subjects. We will not exclude any specific set of subjects except for ones that took part in dictator game experiments in the last three months, and we will pay subjects based on the decisions they take during the experiment according to online experiments standards. We aim to run the study before January 2018.

We plan to run 5 sessions of two-times-both treatments involving 20 active decision-making participants per session per treatment = 200 active subjects per treatment, which leads to a total of 600 subjects (of which 200 are inactive "recipients" in one of the treatments).

2017-11-21

2017-11-22

The above mentioned 20 dictator games will be played under the two treatments/protocols described above. In each treatment, one randomly selected game will be chosen for payment. In each treatment, each dictator plays 20 dictator games. The games vary with respect to p, the price of redistribution of the amount given: given any p, the recipient receives p-times what was given. We chose all prices of redistribution from 0.1 to 2.0, and dictators will face them in random order.

We will compare the resulting two giving distributions using various standard statistics including the arithmetic mean, median, variance, as well as higher moments of the distribution and other distributional measures such as bimodality, etc. In addition, we will define dummy variables for extremal (zero and full) giving decisions.

To test distributional differences between the two treatments we will use the following non-parametric tests:

• distributional differences using Kolmogorov-Smirnov

• differences in mode using Mann-Whitney-Wilcoxon if KS test shows no significant distributional differences

• Equality of variance tests: Levene’s Test (using mean), Brown–Forsythe test (using median) if KS test shows no significant distributional differences

In addition, we identify whether extremal payments are more likely to be made under interactive versus non-interactive and whether zero-giving occurs more frequently for low p's, and full giving more frequently for high p's. These tests will rely on

• Logit regression for dummy “giving was zero” and for dummy "giving was full". Interaction effects between treatment and level of p are included, controlling for subject fixed effects as each dictator decided for all values of p. Regressions are repeated adding variables related to beliefs (as collected via the exit survey); in particular the following variables are added: we ask subjects to specify how many tokens they expected a random other subject to give when in the role of the dictator. We would add this value as a control. Moreover, we ask whether they would have given more (less) if they knew more (less) was given, and we will add two dummies respectively specifying whether the answer to this question is yes or no.

We run a standard two-group, between-subject experiment, with random allocation of roles (dictator vs recipient) in the non-interactive protocol. We will balance sessions by day and time.

In each session, half of the subjects per session are randomly matched into one of two treatments and into one of the two games played per treatment, each subject playing exactly one protocol. In addition, there will be random allocation of roles (dictator vs recipient) in the non-interactive protocol. Active players will face decision situations in random order. Payment is based on one randomly selected decision per dictator.

Since we are comparing the protocols between-subject, we implement the experiment in the interest of budget and reasonable statistical power to have at least n=200 active subjects per treatment.

A total of 600 participants is about as much as we can afford given our budget situation.

Recruiting fewer subjects is not desirable as previous experimental from dictator games (under either protocol) consistently shows that a proportion of subjects (up to 50% in total) always and independent of treatment give either nothing or half, thus reducing the variation of people that could be affected by treatment variations.

Yes

Since we are comparing the protocols between-subject, we implement the experiment in the interest of budget and reasonable statistical power to have at least n=200 active subjects per treatment.

A total of 600 participants is about as much as we can afford given our budget situation.

Recruiting fewer subjects is not desirable as previous experimental from dictator games (under either protocol) consistently shows that a proportion of subjects (up to 50% in total) always and independent of treatment give either nothing or half, thus reducing the variation of people that could be affected by treatment variations.

Since we are comparing the protocols between-subject, we implement the experiment in the interest of budget and reasonable statistical power to have at least n=200 active subjects per treatment. A total of 600 participants is about as much as we can afford given our budget situation. Recruiting fewer subjects is not desirable as previous experimental from dictator games (under either protocol) consistently shows that a proportion of subjects (up to 50% in total) always and independent of treatment give either nothing or half, thus reducing the variation of people that could be affected by treatment variations.

Since we are comparing the protocols between-subject, we implement the experiment in the interest of budget and reasonable statistical power to have at least n=200 active subjects per treatment.

A total of 600 participants is about as much as we can afford given our budget situation.

Recruiting fewer subjects is not desirable as previous experimental from dictator games (under either protocol) consistently shows that a proportion of subjects (up to 50% in total) always and independent of treatment give either nothing or half, thus reducing the variation of people that could be affected by treatment variations.