Disentangling Device and Observability Effects in Dishonesty: Evidence from Probability-Equivalent Randomization Tasks

The study consists of a series of incentivized laboratory experiments in which participants observe realizations from randomization devices and report the outcome for monetary compensation. The objective is to study how the structure and implementation of the randomization device affect dishonest reporting, holding constant the probability distribution of outcomes and the associated payoffs.

Participants are assigned to one of several experimental conditions that vary along three main dimensions. First, we compare binary and multi-outcome environments by implementing coin-based and die-based tasks with equivalent winning probabilities. Second, we vary the cognitive complexity of the randomization process by contrasting simple devices with composite mechanisms that generate identical distributions through sequential steps. Third, we manipulate the format of the randomization device. The latter distinguishes between conditions in which participants generate outcomes themselves (e.g., by flipping a coin or rolling a die) and conditions in which outcomes are generated by the computer within the experimental software.

When outcomes are generated by the participant, only reported values are observed, and dishonesty can be inferred only from aggregate deviations from the theoretical distribution. When outcomes are generated by the computer, the realization is recorded by the software, allowing the researcher to compare realized and reported outcomes at the individual level. Although all decisions remain anonymous and unsupervised, the computer-generated implementation may affect participants’ perception of privacy.

Each participant completes multiple independent rounds of the task within a single session, and monetary payoffs depend on reported outcomes. The design does not involve deception and maintains identical incentives across comparable treatments.

The experiment consists of eight treatments in a between-subjects design, combining four task structures with two modes of implementation of the randomization device.
Each participant is assigned to a single treatment and completes five independent repetitions of the given task. No feedback is provided between rounds, and each round involves a new independent randomization.
The four task structures are as follows.
1. a binary coin task in which a coin is flipped and participants report the result. If the reported outcome is heads, the participant receives a fixed positive payoff (14€) and zero otherwise.
2. a probability-matched die task in which an eight-sided die is rolled and participants report the realized number. If the reported outcome is 1, 2, 3 or 4, the payoff is 0€, while if the reported number is 5, 6, 7 or 8, the payoff is again 14€. This payoff structure ensures that the probability of success is identical to that of the coin task.
3. a multi-outcome die task in which an eight-sided die is rolled and participants report the realized number. Payoffs increase monotonically with the reported value for outcomes 1 (2€) through 7 (14€), while outcome 8 yields 0€ payoff. This generates a uniform distribution over eight outcomes with a non-linear payoff structure.
4. a composite randomization task in which three sequential coin tosses are performed and participants report the resulting sequence of outcomes: Heads,Heads, Heads; Heads, Heads, Tails; etc. Each possible sequence of outcomes is mapped to a number between 1 and 8, and the payoff structure coincides with that of the task 3. This task thereby replicates exactly the distribution of outcomes and payoffs of the eight-sided die.

Each of these four task structures is implemented under two modes of implementation:
(i) In the participant-generated condition, participants physically execute the randomization device (coin or die) and report the outcome. The realization is not observed or recorded by the experimenter or the software, and dishonesty can only be inferred at the aggregate level by comparing the distribution of reports to the theoretical benchmark.
(ii) In the computer-generated condition, the outcome is generated using a standard computer-based randomization procedure implemented in z-Tree. The realization is recorded by the program, and participants are asked to report the outcome without any supervision. This allows for the identification of individual-level misreporting as the discrepancy between realized and reported values. Although all decisions remain anonymous and there is no direct supervision, this implementation may affect perceived privacy relative to the participant-generated condition.

Participants in each session complete five rounds of the assigned task and at the end of the session one round is randomly selected for payment. Total earnings consist of the sum of task earnings and a show-up fee of 4€.

After completing the task, participants answer a questionnaire including demographics, field of study, risk attitudes, and measures related to moral attitudes.

The design allows for both aggregate inference of dishonesty and direct measurement of individual misreporting in half of treatments, while holding constant payoff distributions across comparable conditions.

2026-03-30

2026-04-30

The primary outcomes are the reported outcome in each round. At the aggregate level, we analyze mean reported values and distributional deviations. In treatments with observable outcomes, we additionally construct individual-level measures of misreporting, including indicators for any misreporting, upward misreporting, and maximal misreporting.

The reported outcome constitutes the fundamental behavioral variable and is recorded either as a binary indicator (in coin-based tasks) or as a numerical value (in multi-outcome tasks). Aggregate dishonesty is constructed by comparing the empirical distribution of reported outcomes with the theoretical distribution implied by truthful reporting. This includes differences in mean reported outcomes relative to the expected value under honesty and formal distributional comparisons.

In treatments where outcomes are generated and recorded by the computer, individual-level dishonesty is directly observed. A binary indicator of dishonesty is defined as whether the reported outcome differs from the realized outcome. In addition, the magnitude of dishonesty is measured as the difference between reported and true outcomes, allowing for an analysis of the frequency, intensity and motives of misreporting.

Secondary outcomes include measures related to the cognitive and individual correlates of reporting behavior. These comprise response times during the reporting decision, individual characteristics such as gender, measures of cognitive ability, and survey-based measures capturing moral attitudes, risk preferences, and background characteristics (including field of study and demographics).

Response times are recorded for each reporting decision and will be used as a proxy for the cognitive effort associated with reporting. In particular, longer response times may reflect the additional cognitive cost of manipulating a reported outcome relative to truthfully reporting the realized value.

Individual characteristics are used to study heterogeneity in dishonest behavior. Gender will be included as a pre-specified dimension of heterogeneity, given its relevance in the literature on ethical decision-making. Cognitive ability measures collected in the post-experimental questionnaire will be used to examine whether the propensity to misreport varies with participants’ ability to process probabilistic or sequential tasks.

Survey-based measures of moral attitudes will be constructed from questionnaire responses and used to analyze whether individual differences in moral views are associated with reporting behavior. Risk preferences and field of study will be included as additional covariates in heterogeneity analyses.

These secondary outcomes are not the primary focus of the study but are intended to provide complementary evidence on the mechanisms underlying dishonest reporting and to explore potential sources of heterogeneity across participants.

The study employs a between-subjects experimental design in which participants are randomly assigned to one of eight treatments. Each participant takes part in a single experimental session and is exposed to only one treatment condition. Within each session, participants complete five independent repetitions of a reporting task in which they observe the realization of a randomization device and report the outcome for monetary compensation.

The treatments vary along two key dimensions. The first dimension concerns the structure of the randomization device. Participants are assigned either to binary-outcome tasks or to multi-outcome tasks, and within each category, the randomization process is implemented either through a simple device or through a sequential composite mechanism that generates an equivalent probability distribution. The second dimension concerns the mode of implementation of the randomization device. In some treatments, the outcome is generated by the participant using a physical device (such as a coin or die), while in others the outcome is generated by the experimental software (z-Tree).

In all treatments, the probability distribution of outcomes and the associated payoff structure are held constant across comparable conditions. This ensures that any differences in reporting behavior can be attributed to the structure or implementation of the randomization device rather than to differences in expected monetary incentives.

Participants receive monetary payoffs based on their reported outcomes in one randomly selected round. The experimental design does not involve deception, and all decisions are made anonymously and without direct supervision by the experimenter.

The experiment consists of eight treatments constructed by combining four task structures with two modes of implementation of the randomization device. Participants are randomly assigned to one of the eight treatments and remain in that condition for the duration of the session.

Each participant completes five independent rounds of the reporting task. In each round, a new random outcome is generated, and participants report the realized outcome. No feedback about past outcomes or earnings is provided between rounds, and each round is independent in both realization and incentives. At the end of the session one round is chosen randomly for payment.

The four task structures are defined as follows. In the first task, participants face a binary randomization device consisting of a coin toss. A positive payoff (14€) is associated with one outcome of heads, while the other outcome (tails) yields zero payoff. In the second task, participants face a probability-matched die task in which an eight-sided die is rolled and the same positive payoff (14€) is assigned to a subset of outcomes (e.g., values 5 through 8), ensuring that the probability of receiving the positive payoff is identical to that in the coin task.

In the third task, participants face a multi-outcome randomization device consisting of an eight-sided die. Each outcome corresponds to a different payoff, with payoffs increasing monotonically in the reported value for outcomes 1 (2€) through 7 (14€), while one outcome (e.g., 8) yields zero payoff. This generates a uniform distribution over eight possible outcomes combined with a non-linear payoff schedule.

In the fourth task, a composite randomization procedure consisting of three sequential coin tosses is performed. Each possible sequence of coin outcomes is mapped uniquely to one of eight numerical outcomes, thereby replicating exactly the probability distribution and payoff structure of the previous, eight-sided die task. This treatment isolates the effect of procedural complexity while holding the outcome distribution constant.

Each of these four task structures is implemented under two modes of implementation. In the participant-generated condition, individuals physically execute the randomization device (coin or die) and privately observe the outcome. They then report the outcome through the computer interface. The realization is not observed by the experimenter or recorded by the software, so dishonesty can only be inferred at the aggregate level.

In the computer-generated condition, the outcome is generated within the experimental software (z-Tree) using a built-in randomization procedure and is recorded by the program. Participants observe the outcome on the screen and are asked to report it. This allows for a direct comparison between the realized and reported outcomes at the individual level. Although all decisions remain anonymous and there is no direct monitoring by the experimenter, this implementation may affect participants’ perception of privacy relative to the participant-generated condition.

At the end of the task, participants complete a questionnaire collecting information on demographics (including gender), field of study, risk and loss preferences, cognitive ability, and moral attitudes. These measures are used for heterogeneity analysis and to explore potential mechanisms underlying reporting behavior.

Participants are randomly assigned to treatments through a computerized randomization process that ensures that each participant has an equal probability of being allocated to any of the treatment conditions. Randomization is performed independently for each participant and does not depend on participant characteristics or on the behavior of other participants.

The unit of randomization is the individual participant. Each participant is assigned to exactly one treatment and remains in that treatment throughout the session.

No

480 participants

Each of the 480 participants completes five independent rounds of the reporting task, resulting in a total of 2,400 observations at the decision level.

The sample is evenly distributed across the eight treatments, with 60 participants assigned to each treatment condition. Since each participant completes five rounds, each treatment yields 300 observations at the decision level.