Freedom to choose or opportunity to exploit? Autonomy in evaluation timing, effort provision, and signaling concerns in a principal-agent setting

We conduct a between-subjects online experiment within a principal-agent framework. Participants in the role of agents perform a real-effort task over two periods, with their output quantified in points earned. Principals benefit financially from each point achieved by their assigned agent and evaluate the agent’s performance by setting a threshold (point category) that determines whether the agent receives a discretionary bonus in addition to their flat fee.

The key intervention systematically varies whether the timing of this subjective performance evaluation (SPE) is determined exogenously (EXO) or chosen by the agent (ENDO). In both treatments, there are two possible evaluation points: either in the middle of the two working periods (middle) or at the end (end).

In addition, to isolate the role of signaling concerns, we introduce a NO-SIGNAL treatment. As in the other treatments, agents work for a flat fee and a potential discretionary bonus. Mirroring the choice structure in the ENDO condition, agents can choose the timing of their evaluation (middle or end). However, unlike in the other treatments, the bonus is not based on the principal’s evaluation. Instead, it depends on whether the agent meets a predetermined performance threshold. This threshold is not disclosed to participants (consistent with the other treatments) and is derived from the decisions made by principals in the EXO treatment. To ensure comparability across conditions, the agent is still matched with a principal, who benefits from the agent's performance.

2025-08-12

2025-08-22

For the main analysis, we collect data on employees’ output, measured in points achieved in the real-effort task in the main part (Part 2). We compare the output (i.e., points achieved) between the ENDO and EXO treatments overall and per period to examine the influence of autonomy in choosing the evaluation timing (see separate analysis plan).

To examine two key influencing factors, we first analyze the share of “shirkers” between the ENDO and EXO treatments at the individual level.
Second, we analyze the distribution of evaluation timing choices (middle vs. end) in the ENDO and NO-SIGNAL treatments, as well as output overall and per period between the NO-SIGNAL versus ENDO and EXO treatments.

Employees’ output by evaluation time (middle vs end), Managers’ evaluation behavior, employees’ output based on the manager’s bonus decision, managers’ beliefs about employees’ output, employees’ beliefs about managers’ behavior

To answer our research questions, we conduct a between-subjects online experiment in a principal-agent setting. Participants are assigned to one of two roles: agent (employee) or principal (manager). Participants will remain in the same role throughout the experiment and will be paired with a player from the other role.

To ensure that the real-effort task is completed by human subjects themselves, we conducted extensive pre-testing and made targeted adjustments to the task design. Specifically, we introduced features intended to reduce the likelihood that large language models (LLMs) or image recognition algorithms could solve the task efficiently.
To assess the robustness of these adjustments, we submitted the task to several publicly available and mostly used AI systems (ChatGPT 4o, Gemini 2.5, and Claude.ai Sonnet 4). Solutions by these models were either unsuccessful or more time-demanding than completing the task manually. Given that participants are incentivized based on their performance, we can reasonably assume that they are motivated to complete the task themselves.

Not available

Computer

Experimental session

No

We plan to gather data on 136 groups per treatment.

In total about 816 participants. In the analysis, we will exclude participants that do not work at all.

For each of the experimental treatments (ENDO / EXO / NO-SIGNAL treatment) 136 groups, each consisting of an employee and a manager (between-subject design).

The distribution of evaluation timings (middle vs end) observed in the ENDO treatment is replicated in the EXO treatment by randomly assigning evaluation timings based on the ENDO frequencies.

For the calculation of the required sample size, we set the error probability to 0.05, the power to 0.80, and the minimum detectable effect size to 0.35. Based on a two-sided Wilcoxon-Mann-Whitney test, these parameters yield a required sample size of 136 groups (272 participants) per treatment.