Do moral transgressions lead to pro-social effort? A real-effort experiment

Participants are randomly assigned to either the control group (manual decision whether or not to take the money from the charity) and the treatment group (computer decides whether the money is transferred to the charity or to the participant).
Additionally, participants are randomly assigned to three different orders in the letter counting task (part 2 of the study) to control for potential order effects. Participants hence either first work for their own account and then for the charity’s account or vice versa.

2022-02-15

2023-12-31

The main, countervailing, hypotheses we test are:
H1: Those who actively take the money exert relatively higher effort for the charity compared to those who did not take the money – this would mean that moral licensing dominates the outcome.
H2: Those who actively take the money exert relatively lower effort for the charity compared to those who did not take the money – this would mean that the stability of social preferences dominates the outcome.
Our former experiment suggests that H2 holds.

Our primary outcomes are (i) is whether those who actively take the money exert relatively higher or lower effort for the charity compared to those who did not take the money, and (ii) how the impact of getting the money is affected by whether it’s the own active decision or a computer draw.
Our main regression is

PerformanceforCharity=β_0+β_1 MoneyTransferred+β_2 TreatGroup+β_3 (MoneyTransferred×TreatGroup)+β_4 PerformanceforOwnAccount+ Controls+ϵ

The main coefficient of interest is β_3 – if it is positive, this means that if the money was actively taken then subjects in the treatment group have higher outcomes when working for the charity, controlled for the outcome when working for their own account. A positive coefficient would hence point to the dominance of moral licensing (i.e. those who actively take the money subsequently exert higher effort for the charity), and a negative coefficient to the dominance of stable preferences.
What about the rest? As we include this interaction, the coefficient β_1 refers only to the case where the random draw has transferred the money. A positive coefficient β_1 would hence mean that unintentional unfairness increases pro-social effort. Respectively, the coefficient β_2 refers only to the case where the random draw did not transfer the money. We do not make any specific hypothesis about this case. Finally, β_4 is expected to be positive, as it can be seen as a proxy for capabilities and motivation.
In addition to the parametric analysis, we will perform Wilcoxon rank sum tests to compare the difference in the performance when working for the own and the charity’s account of
Those who took the money deliberately to those who deliberately did not take it
Those who got the money from the computer to those who did not get the money from the computer.
Those who took the money deliberately to those who got the money from the computer
Those who did deliberately not take the money to those who did not get the money from the computer.
We will also use Wilcoxon sign-rank test to compare performance in the two real effort tasks within-subject in each group.

Our first secondary outcome is whether the outcome when working for the own and the charity’s account differs at all. This will be done with a regression with the outcome as dependent variable, the dummy that takes the value 0 (1) when working for the own (the charity’s) account, a dummy that takes the value 0 (1) when working first for the own (the charity’s) account, and the personal control variables.
Our second kind of secondary outcome is the impact of the control variables (demographics and personality, in particular, with regards to social attitudes).

The experiment will be conducted online and participants will be recruited using Prolific. In Part (1) of the study, subjects either decide actively whether they transfer money designated for donation to their own account (treatment group) or whether this is decided by a random computer draw. In Part (2) all subjects work on a real-effort task (counting letters in senseless paragraphs), both for their own and for the charity’s account. A 50-50 random computer draw decides whether subjects work first for their own or the charity’s account.

Not available

Subjects will be randomly assigned to the treatment and the control group (computer decides vs. participants decide themselves), and randomly assigned to the order in which they work on the task (own account vs. charity’s account). The computer will also randomly determine which of the two real effort tasks (own vs. charity) is paid-out.

Individual. No clustering.

No

0. No clustering.

400 individuals, thereof 200 get to decide for themselves whether or not to take the money from the charity, and for the other 200 the computer decides. Of the 200 in each group, 100 will work first for their own account and then for the charity’s account and the other 100 work in opposite order.

400

We plan to run a preliminary pilot study with 40 subjects in November/December 2022. If this pilot reveals a need for readjustment, we will update our analysis plan accordingly.

Regression: The regression analysis allows detecting a Small-Medium effect size of 0.0302 (assuming 400 subjects, 4 tested predictors, 10 predictors overall (approximated), and β=0.8,α=0.05). Between-subject Wilcoxon rank-sum tests (assuming a two-tailed normal parent distribution and β=0.8,α=0.05) comparing two sub-groups of 100 subjects each allows detecting a medium effect size of 0.047. If order effects are found to be irrelevant, we can instead restrict attention to the treatment group vs. the control group, with 200 subjects each. This allows detecting a smaller effect size of 0.0287. Within-subject Wilcoxon sign-rank (assuming a two-tailed normal parent distribution and β=0.8,α=0.05). A Wilcoxon sign-rank test (within-subject) looking at each sub-group (of 100 subjects) allows detecting a Small-Medium effect size of 0.028. If order effects are found to be irrelevant, we can instead restrict attention to the treatment group vs. the control group, with 200 subjects each. This allows detecting a smaller effect size of 0.0203. All calculations were made using G*Power 3.1’s sensitivity analysis (using “R2 increase” for regressions and the respective tests as mentioned above).