Experimental Design
This experiment is a 3x2 design based on an emission game in which we manipulate the audit mechanism (random vs. competitive) and the level of information firms have about each other’s emissions (No Information vs. Imperfect Information vs. Perfect Information).
In this game, n firms produce a level of output e that they have to self-report to a third-party. Firms receive a benefit from their production and pay a tax t on their reported output. Each round, one participant is audited by the third-party. Under the random audit mechanism, the audit probability is fixed and the same for each participant (1/n). Under the competitive audit mechanism, the third-party compares firms’ reports against a reference value and assign a larger audit probability to firms that report less and a lower audit probability to firms that report more. Participants that are audited face a penalty c for each unit of output underreported.
In our experiment, participants are randomly matched in groups of 3 and play 7 incentivized rounds of the emission game under a unique audit mechanism and a unique level of information. Each round is composed of 4 stages: an output stage, an information stage, a report stage and a feedback stage. Participants are randomly rematched at the beginning of each round.
In the output stage, participants choose a level of output. In the information stage, they receive information about the output of their fellow group members. In the Perfect Information treatment, participants perfectly observe the actual level of output of their fellow group members. In the Imperfect Information treatment, participants observe the level of output of their fellow group members with some noise. The noise is either +10 with probability 0.25, -10 with probability 0.25 and 0 with probability 0.5. In the No Information treatment, participants do not receive any information about their fellow group members. In the report stage, participants choose how many units of output they wish to report. In the feedback stage, participants are informed about whether they have been audited and their earnings for this round. At the end of the experiment, 4 rounds are randomly selected for payment. This design allows us to test the following hypotheses:
Hypothesis 1. Given that firms have no information about each others’ emissions, then CAMs induce: (a) the same level of emission than random audits and (b) higher emission reports than random audits.
Hypothesis 2. Given that firms have perfect information about each others’ emissions, then CAMs induce: (a) lower emissions than random audits and (b) higher emission reports than random audits.
Hypothesis 3. Given that firms have perfect information about each others’ emissions, CAMs induce socially efficientemissions.
Hypothesis 4. CAMs induce: (a) lower emissions and (b) lower emission reports when firms have perfect information about each others’ emissions as compared to when firms have no information.
Conjecture 1. Given that firms have limited information about each others’ emissions, CAMs induce: (a) lower emissions than random audits and (b) higher emission reports than random audits.
The data will be collected online in an environment that mimics the laboratory ("quasi-lab experiment"). Using a combination of oTree and HeiCONF (a videoconferencing software), we will conduct 8 experimental sessions varying between 15 and 21 participants. In each session, an experimenter will explain the unfolding of the session, read the instructions and assist the participants throughout the experiment. Our subject pool will be composed mainly of students from local engineering, business, and medical schools in Lyon (France).
In the case of subjects dropping out, their actions in the group will be replaced by the actions of a backstop player. The observations for participants that are matched with the backstop player will be excluded from the analyses for the period in which the encounter happened. For example, if participant 1 and participant 2 are matched with the backstop player in period 5, we will exclude the data for both participants in that particular period from the analyse. The data from the backstop player will also be excluded from the analyses.