The Feedback Fallacy: Praise, Criticism, and Regression to the Mean

Participants in the role of principal observe an agent's interim performance on a task and may send a verifiable feedback message. Agents subsequently choose effort for a final stage.

The experiment is a computerized laboratory experiment implemented in z-Tree (Fischbacher, 2007). Each session involves a variable number of participants (typically between 16 and 24, depending on attendance), who are randomly assigned to the role of "employer" (principal) or "worker" (agent). Roles are fixed for the duration of the experiment. The experiment consists of 20 rounds, with random rematching of principal-agent pairs in each round (stranger matching).

Each round consists of two stages. In each stage, the agent chooses an effort level (integer from 0 to 10), which costs c(e) = 0.25e². The outcome of each stage (low, medium, or high performance) is determined stochastically by the agent's effort and task difficulty. Task difficulty is either easy (θ = 2/5, probability 1/2) or difficult (θ = 1/10, probability 1/2), drawn at the start of each round and constant across both stages. Neither party is told the task difficulty. The exact outcome probabilities are implemented in the z-Tree program.
Payoffs from outcomes: low performance yields 0 for the worker and 0 for the employer; medium performance yields 40 for the worker and 20 for the employer; high performance yields 100 for the worker and 120 for the employer.

After stage 1, the principal observes the outcome and may send a message. If the outcome is high, the principal can send "Well Done" or no message. If the outcome is low, the principal can send "Not Good Enough" or no message. If the outcome is medium, no message can be sent. After receiving any message or no message, the agent chooses effort for stage 2. At the end of each round, both parties learn the outcomes of both stages.

2026-06-08

2026-10-01

Principal's binary message choice conditional on a high first-stage outcome (Well Done vs. no message), modeled as a function of cumulative prior disappointing experience with praise.

"Disappointing experience with praise" is constructed as the cumulative count of prior rounds in which the principal sent "Well Done" and the subsequent second-stage outcome was not high.

Agent's second-stage effort, modeled as a function of received feedback message and first-stage effort.
Frequency of performance reversion: proportion of cases in which a high first-stage outcome is followed by a non-high second-stage outcome, and proportion of cases in which a low first-stage outcome is followed by a non-low second-stage outcome.
Principal's binary message choice conditional on a low first-stage outcome (Not Good Enough vs. no message), as a function of cumulative prior encouraging experience with criticism.

Encouraging experience with criticism" is the cumulative count of prior rounds in which the principal sent "Not Good Enough" and the subsequent second-stage outcome was not low. All cumulative experience variables are lagged and constructed only from prior rounds.

Computerized laboratory experiment using a principal-agent design with 20 rounds and stranger matching. Participants are randomly assigned to fixed roles (principal or agent) at the session start. In each round, principals observe an agent's interim performance on a two-stage task and may send a verifiable feedback message before the agent chooses effort for the final stage. Sessions are conducted at the BELA laboratory at NES.

The experiment is a computerized laboratory experiment implemented in z-Tree (Fischbacher, 2007). Each session involves a variable number of participants (typically between 16 and 24, depending on attendance), who are randomly assigned to the role of "employer" (principal) or "worker" (agent). Roles are fixed for the duration of the experiment. The experiment consists of 20 rounds, with random rematching of principal-agent pairs in each round (stranger matching).

Task structure. Each round consists of two stages. In each stage, the agent chooses an effort level (integer from 0 to 10), which costs c(e) = 0.25e². The outcome of each stage (low, medium, or high performance) is determined stochastically by the agent's effort and task difficulty. Task difficulty is either easy (θ = 2/5, probability 1/2) or difficult (θ = 1/10, probability 1/2), drawn at the start of each round and constant across both stages. Neither party is told the task difficulty. The exact outcome probabilities are implemented in the z-Tree program.
Payoffs from outcomes are as follows: low performance yields 0 for the worker and 0 for the employer; medium performance yields 40 for the worker and 20 for the employer; high performance yields 100 for the worker and 120 for the employer.

After stage 1, the principal observes the outcome and may send a message. If the outcome is high, the principal can send "Well Done" or no message. If the outcome is low, the principal can send "Not Good Enough" or no message. If the outcome is medium, no message can be sent. After receiving any message or no message, the agent chooses effort for stage 2. At the end of each round, both parties learn the outcomes of both stages.

Session procedures. We aim to recruit approximately 35–50 principals (70–100 participants in total), distributed across four to six sessions of variable size depending on attendance. Each session lasts approximately 40 minutes. At the beginning of each session, participants receive written instructions, which are also read aloud by the experimenter. Participants have an opportunity to ask questions privately. A practice round is conducted before the main experiment begins.

All subjects receive a guaranteed show-up fee of 550 ECU, plus performance-based earnings. At the end of each session, earnings are determined by a random selection of 4 out of 20 rounds. The exchange rate is 1 ECU = 1 RUB (Russian rubles). Expected earnings are approximately 900 RUB per participant for a 40-minute session.

After the experiment, subjects complete a short questionnaire collecting age, gender, field of study/major, optional GPA or GPA range, and questions about feedback strategies and attitudes toward positive and negative feedback. Principals are additionally asked about their perception of the effectiveness of their feedback.

Random assignment by the z-Tree software at the session start (for role assignment) and at the start of each round (for pair rematching).

Individual within session for role assignment (principal vs. agent); pair-level for stranger matching across rounds. All participants within a session are exposed to the same procedure.

No

4-6 experimental sessions

70-100 participants (35-50 principals, 35-50 agents); 20 rounds per participant, generating 1400-2000 round-level observations per role

Not applicable: the study does not use between-subjects treatment arms. All participants experience the same protocol; variation in feedback received by agents is generated endogenously by principals' decisions conditional on observed outcomes.

Power analysis is conducted via simulation calibrated to pilot estimates of the relationship between cumulative disappointing experience with praise and the probability that a principal sends a "Well Done" message conditional on a high first-stage outcome. With the planned recruitment of 35–50 principals (70–100 participants in total) across approximately four to six sessions, the design achieves at least 80% power at α=0.05 to detect effects as small as 50% of the pilot point estimate. At N=40 principals, power exceeds 91% under a 50% effect-size attenuation and remains above 70% under a 75% attenuation (i.e., when the true effect is only one-quarter of the pilot estimate). With the planned recruitment of 35–50 principals, the design achieves at least 80% power at α=0.05 to detect effects as small as 50% of the pilot point estimate. At N=40 principals, power exceeds 91% under a 50% effect-size attenuation and remains above 70% under a 75% attenuation (i.e., when the true effect is only one-quarter of the pilot estimate).