Mechanism Design for All

We study how to explain economic mechanisms to lay participants.

We study how to effectively explain a complex economic mechanism — specifically, a VCG mechanism with two goods. We have nine between-subjects treatments. In each session, participants are read aloud one of nine explanations of the mechanism: an uninformative (“null”) explanation, an explanation produced by a large language model (AI), or one of seven explanations produced by individual economist experts. After the explanation, all participants bid in four incentivized VCG auctions against two computerized opponents (values for all bidders are independently drawn from the same distribution), then answer a battery of incentivized multiple-choice comprehension questions about the mechanism and complete an unincentivized exit questionnaire eliciting attitudes. Cross-cutting the explanation manipulation, sessions are also randomly assigned to one of two stakes levels (normal or double). Based on the results of this experiment, we may run follow-up experiments that study whether input from science communication experts improves explanations.

2026-09-21

2026-10-11

1. Bidding behavior. (a) Mean absolute deviation of each participant’s bid from the dominant-strategy bid, averaged across the four auctions, with per-auction deviations winsorized at 10; (b) share of dominant-strategy bids. We will appropriately rescale (a) in the double-stakes treatments when pooling across regular stakes and double stakes.
2. Comprehension. Total number of correct answers on the incentivized multiple-choice comprehension questions, both overall and broken down by prespecified topic / difficulty sub-group.
3. Attitudes. Responses to the exit-questionnaire attitude items, broken down by topic.

· Subgroup analyses of all of the above by participant characteristics (major, gender, math proficiency, attention-check performance, socioeconomic status, and other demographics).
· Explanation-feature analyses: each explanation will be scored on a prespecified feature index (e.g., presence of menu description, worked examples, explicit bidding advice), and we will explore the relationship between feature scores and outcomes.

Subjects will read an explanation, participate in a mechanism, and answer comprehension questions.

All sessions follow the same internal structure; what varies between sessions is the explanation read aloud and the stakes level.
A session proceeds as follows:
1. Explanation. The session’s randomly assigned explanation is read aloud to participants. There are nine explanation conditions: one null (uninformative) explanation; one explanation produced by a large language model; and seven explanations produced by individual economist experts.
2. Auctions. Participants then complete four incentivized VCG auctions against two computerized opponents. Each auction sells two goods; participants’ values and computerized opponents’ values are drawn independently from the same distribution.
3. Comprehension questions. Participants answer a battery of incentivized multiple-choice questions about the mechanism. Questions are pre-organized into topic / difficulty sub-groups. The order of questions within each sub-group is randomized at the participant level; the order of sub-groups across the questionnaire is fixed.
4. Exit questionnaire. Participants complete an unincentivized exit questionnaire with attitude items (organized by topic) and demographic / self-report measures.
All sessions use the auctions-first ordering described above.
Cross-cutting the nine explanation conditions, sessions are randomly assigned to one of two stakes levels:
· Normal stakes: $0.30 per correct comprehension answer (up to a $10.50 maximum); expected auction earnings of approximately $2 per round.
· Double stakes: Both comprehension and auction payments doubled, so perfect comprehension plus expected auction earnings yield approximately $37 in expected bonus payments (in addition to a show-up fee).
The double-stakes condition targets approximately ⅙ of participants, but actual proportions will vary with session-day attendance and are not committed to exactly ⅙.
Sessions will be run at Cornell University and Ohio State University, with additional sites under discussion (potentially the University of Pittsburgh and the University of California, San Diego).

By computer.

Individual.

No

We will assign treatments at the session level because instructions will be read aloud, but subjects do not interact with each other, so we will treat each subject as an independent observation.

660 subjects (with variation depending on attendance).

660.