The Role of Peer Effects and Simple Gamification Mechanisms for Green Crowdfunding

Observational Study:
All visitors of EcoCrowd have the possibility to take part in the survey.

Natural field experiment as a randomized trial without non-compliance
Using a full factorial-design, participants of the survey are randomly assigned to one out of 5 treatment groups or the control group in equal shares in order to study the effects of emotional primes on decision to financially contribute to an environmental organization as in comparison to a donation or reward based crowdfunding project or keeping money.

Natural field experiment as a randomized trial with non- compliance
Second, participants and non-participants of the survey are further assigned to one out of two treatments and one control group in equal shares. Treatment group 1 has the possibility to play the game.
Within the game the participants are free to answer four funny designed self-revelation questions regarding their environmental relevant behavior and how they would behave in specific situations. The second group gets specially designed sharing application that contrary to the standard Facebook-Plugin contained on the website allows sharing content with selected friends (“placebo” group). The application will be presented in the same fashion (pop-up modal dialog of same size, comparable message) to the visitors of the crowdfunding platform. The control group gets the website in its original state (no treatment group).

2018-08-29

2019-06-05

Propensity of posting content of the crowdfunding website on Facebook
Financial involvement (propensity to contribute: yes /no; how much, which project type)
Number of likes as a measure of visibility effectiveness, visits EcoCrowd /projects of alter / contribution of alter

The experiment consits of two parts.
1) Observational (online) survey with randomized trial without non-compliance with treatment randomized at indivdual level
2) Field experimental randomized trial with non-compliance (Facebook Game) with treatment randomized at individual level

Given that previous research found that the different randomization methods perform similarly in samples greater than 300, we use pure randomization within a between-subjects design.

Individual

No

The survey will consist of approximatly 2500 individuals.
The natural field experiment as a randomized trial with non-compliance consists of 15.000 individuals.

The (online) observational survey will consist of approximatly 2500 individuals. The natural field experiment as a randomized trial with non-compliance consists of 15.000 individuals.

The survey will consist of approximately 416 individuals in the 5 treatments and 416 in the control group.
The natural field experiment with non-compliance will consist of approximately 5000 individuals in the 2 treatments and 5000 in the control group.

Statistical power calculations were undertaken to determine the appropriate sample size (for details see attached pre-analysis plan). 1) Survey and randomized trial without non-compliance For the questionnaire we use a probit model for data generation. We simulated 1000 runs of this model for different values of the effect and sample size and estimated the statistical power by estimating the percentage of positive findings given the significance level of 5%. For this we used R's base functions and the glm function from the MASS package. The results show that achieving a statistical power of 80% is only feasible for effect sizes that are greater than or equal to 0.2 (which approximately corresponds to an increase of 6.8% in the outcome from the baseline given the treatment). To achieve this a sample size of approx. N=2500 is necessary. 2) Field experimental randomized trial with non-compliance (Facebook Game) For the Facebook game we assumed the following dependency graph on the random variates involved. This corresponds to the classical setup of a randomized trial with non-compliance: The outcome Y is stochastically dependent on a confounding variable X as well as on the action/treatment A which itself depends on an instrumental variable Z. Greek letters placed on the arrows in the dependency graph indicate the parameter which stochastically influences the variable at the end of the arrow given that the value of the variable at the beginning of the arrow is 1. For the data generation process we modeled Z and X as independent Bernoulli random variables with success probability frac=1/3 (“probability of offer to treatment”) and frac2=1/10 respectively. A and Y are then mixtures of Bernoulli random variables. Again we simulated from this setup 1000 runs for each scenario and calculated the percentage of positive findings under a significance level of 5% in a linear IV regression setting. For the latter we used the R package ivreg. In the following figure we show how different effect sizes, i.e. β_2∈{0.03,0.06,0.1,0.2,0.3}, cf. the legend of the graph, result in different power estimates. Each of the nine plots in the figure corresponds to a parameter combination of α_2∈{0.01,0.05,0.15} and β_1∈{0.01,0.05,0.1} with α_2 (β_1) being shown in the columns (rows) of the three-by-three grid. We believe that a realistic setting for sharing is α_2=0.05,β_1=0.01, see the plot in the upper middle cell of Figure 3, and for donating is α_2=0.05,β_1= 0.1, see the plot in the lower middle cell. Therefore, we conclude that we need a sample size of N=15000 in order to guarantee a power of about 80 percent for the effect sizes which seem realistic (i.e. effect sizes greater or equal to 0.1 or a 10% increase in the probability to share content or donate).