The Comparative Blind Spot: Misconceptions About Comparative Advantage

In the experiment, subjects complete production tasks that require one finished product consisting of two complementary parts. Producing each part incurs a potentially different cost. Subjects are given the opportunity to work with another producer (a computer player), who also needs one complete product and may face different production costs. Subjects first decide whether to cooperate and then propose how to divide the gains from cooperation through a transfer.

The main treatment variation is the subject and the other producers’ production costs. Cost structures fall into the following treatment conditions:
- AA: Each producer has an absolute advantage in a different good, so efficient cooperation can be identified using absolute advantage.
- CA (Adv.): Cooperation can only be based on comparative advantage, and the subject is universally advantaged.
- CA (Disadv.): Cooperation can only be based on comparative advantage, and the subject is universally disadvantaged.
- NA: There are no gains from cooperation

We'll compare subjects' decisions across these treatments. See the attached document for rationale and details.

2026-08-11

2026-10-01

Our primary cooperation outcome, cooperation, is a binary variable equal to one if the subject chooses either production plan that involves working with another producer, and zero if the subject chooses to produce independently.

We use two measures of the proposed division of gains: first, the raw proposed transfer amount and second, the proposed share of the cooperation surplus.

The raw proposed transfer amount is the transfer requested from or offered to the other producer. Positive values request a payment from the other producer, whereas negative values offer a payment to it. We analyze this measure for all cooperation proposals. We prioritize this measure because it is observed for all subjects who cooperate.

The proposed share of the cooperation surplus is defined conditional on efficient cooperation because otherwise the cooperation surplus is zero. It is the fraction of the total cooperation surplus that the subject would receive if the proposal were accepted. It is based on the subject’s reduction in final cost relative to producing independently, including the proposed transfer, and therefore captures the division of gains implied by the proposal regardless of whether it is accepted.

The secondary outcomes are:
• Beliefs: For a subset of main tasks, perceived total gain from cooperation without a transfer; perceived gain to the subject without a transfer; and perceived gain to another producer without a transfer.
• Incomplete-information choices: Responses to questions when transfers are unavailable and when they are available.
• Choices under calculated costs: For a subset of main tasks, cooperation choices and, conditional on cooperation, transfer amounts.
• Vignette choices: Cooperation choices in vignette scenarios that present less abstract decision contexts.
• Other measures: stated support for free trade; the aggregate zero-sum-belief index; the trade-domain zero-sum belief; and the Cognitive Reflection Test score.

These secondary outcomes are used in exploratory analyses of the reasons underlying the observed choice patterns and the extent to which similar patterns arise in less abstract decision contexts.

The experiment proceeds in five main parts. Subjects first complete the main task, in which they choose production plans and transfers across a series of cost structures. Second, subjects make choices under calculated costs: instead of the per-unit production costs shown in the main task, we display, for each available production plan, the total production cost the subject would incur without a transfer. Third, we elicit beliefs about the total gains and the gains to each party from cooperation without transfers. Fourth, subjects assess whether mutually beneficial cooperation is possible given incomplete information about another producer’s costs. Finally, subjects respond to cooperation vignettes and additional survey questions.

See the attached document for rationale and details.

Not available

We have a within-subject design, where each subject complete 24 tasks that fall into different treatments. The selection of the tasks and the order of the tasks are randomly decided by the survey script in Qualtrics.

Individual (Within-subject)

Yes

N=250 subjects on Prolific

The experiment has a within-subject design, and the observation is at subject * task level (each subject will complete 24 tasks) At recruitment, we screen out anyone who fails any of the following checks: • a reCaptcha score below 0.7; • missing the attention-check question; • failing a technical check in which a sequence of numbers is displayed in video format. In the main analysis, we further exclude subjects who: • are not rated “high” on Prolific’s authenticity checks for both LLM and Bot; or • correctly answer the question “How many elements does the largest sporadic simple group have? If you do not know it, answer ‘don’t know’.” Almost no human knows this figure, whereas an AI does, so a correct answer flags a likely bot or AI-assisted response. We aim for a final sample of N = 250 subjects satisfying all of these criteria.

Each subject completes a given number tasks fall into the following treatment conditions:
- AA treatment: 10 tasks
- CA (Adv.) treatment: 6 tasks
- CA (Disadv.) treatment: 6 tasks
- NA treatment: 2 tasks

The tasks completed by each subject are drawn from a set of 29 possible tasks, with the specific subset varying across subjects so that all 29 tasks are represented in the sample.