Ratings, Externalities and Competition

We vary how information about two product attributes is displayed to consumers in a laboratory Bertrand duopoly. In each matching group, two sellers simultaneously choose quality in attribute A, quality in attribute B, and a price. A single consumer observes both offers and buys at most one unit, or abstains.

The consumer's payoff depends only on attribute A and the price, so attribute B has no consumption value. Attribute B stands for an investment, such as reducing an externality, that is costly to the seller and not directly valued by the buyer.

The two treatments differ only in what the consumer sees:

Separate ratings: the consumer observes attribute A and attribute B for both sellers, each displayed as its own rating, together with both prices.

Combined rating: the consumer observes a single rating for each seller, equal to 0.75 times attribute A plus 0.25 times attribute B, together with both prices.

The question is whether bundling attribute B into the displayed rating induces sellers to invest in it even though the consumer places no value on it.

2026-09-22

2027-12-28

Investment in attribute B by sellers.

Attribute B investment is a seller's choice on a 0 to 100 scale in each round. The outcome is constructed at the level of the matching group, which is the unit of randomisation and the unit of analysis: for each matching group we take the mean of attribute B investment across both sellers and across all rounds in the analysis window, giving one value per group.

The analysis window is round 4 to the last round the group completed. Rounds 1 to 3 are excluded to allow for learning. This window is pre-specified on the basis of a pilot session, in which the treatments were indistinguishable over the first three rounds and separated cleanly from round 4 onward. A group must complete at least six rounds to enter the analysis.

Investment in attribute A by sellers, and prices posted by sellers.

Both are constructed exactly as the primary outcome: the mean across both sellers and all rounds in the analysis window, giving one value per matching group.

Attribute A is predicted to be 50 under separate ratings and between 45 and 50 under the combined rating, so the two treatments are predicted to be similar or close on this dimension. Prices are predicted to be 25 under separate ratings and between 22.50 and 27.78 under the combined rating, again a small difference relative to the difference predicted for attribute B.

Subjects are randomly assigned to fixed matching groups of three: two sellers and one buyer. Roles and group membership are fixed for the whole session, so each matching group is a single independent observation.

Each round has three stages. The two sellers simultaneously choose quality in attribute A, quality in attribute B, and a price. The buyer then sees both offers, displayed in randomised order, and chooses one of them or neither. Finally all group members receive feedback.

Groups play up to 15 rounds with fixed matching and no rematching. A session ends after round 15 or after one hour of play, whichever comes first. The stopping rule is announced in the instructions and the number of rounds each group completes is recorded.

Sessions begin with six instruction screens, a practice screen on which subjects can enter hypothetical combinations of the two quality levels and a price and see the resulting seller profit, consumer payoff and rating, and five incentivised comprehension questions with answers displayed afterwards. Subjects receive a starting capital of 50 points and cumulative earnings are floored at zero, so no subject can finish with negative earnings. Points are converted to cash at a rate announced in advance.

The two treatments differ only in whether the buyer sees two separate attribute ratings or one combined rating. All other features are identical.

Not available

Subjects are assigned at random to computer terminals on arrival at the laboratory, and the software forms matching groups of three from consecutive terminal numbers. Groups with an odd index are assigned to the separate-ratings treatment and groups with an even index to the combined-rating treatment. Both treatments are therefore present in every session.

The matching group of three subjects (two sellers and one buyer). Treatment is assigned at the group level, within session. There is no second level of randomisation.

Yes

36 matching groups (each consisting of 3 subjects: 2 sellers and 1 buyer).

108 subjects (72 sellers and 36 buyers). At the seller-round level this gives up to approximately 1,080 seller decisions, but the unit of analysis is the matching group, so the analysis is based on 36 independent observations.

18 matching groups separate ratings (54 subjects), 18 matching groups combined rating (54 subjects).

The outcome is mean attribute B investment per matching group, measured in points on a 0 to 100 scale. The matching group is both the unit of randomisation and the unit of analysis, so clustering is fully accounted for by aggregating to the group level before testing. A pilot session gives a pooled between-group standard deviation of attribute B of 4.0 points, but estimated on only 3 degrees of freedom, so the 95 percent interval for the standard deviation runs from 2.3 to 15.1. We therefore report minimum detectable effects across a range of standard deviations rather than relying on the pilot point estimate. Because the primary test is nonparametric and the outcome is bounded below at zero, power was simulated rather than taken from a normal-theory formula. The simulation draws the group-level outcome as a normal censored at zero, to reproduce the mass of groups that settle on zero investment, and applies the Mann-Whitney rank-sum test at the 5 percent two-sided level. With 18 groups per arm, the minimum detectable difference at 80 percent power is 4.08 points if the standard deviation is 4, 5.04 points if it is 5, 6.09 points if it is 6, 8.07 points if it is 8, and 10.12 points if it is 10. The predicted treatment effect is 15 to 16.67 points, and the pilot difference over the analysis window was 14.06 points. Expressed relative to the predicted effect, the design detects effects of roughly 25 to 40 percent of the predicted size for standard deviations between 4 and 6. At the pilot effect size, simulated power exceeds 0.99 for any standard deviation up to 8, and is 0.97 even at a standard deviation of 10. A t-test applied to the same simulated data gives minimum detectable effects within about 1 percent of the rank test, so the choice between the two does not drive the sample size. A normal-theory t formula would give figures about 5 percent smaller, but that formula assumes an uncensored normal outcome and therefore understates the minimum detectable effect for this design. Simulated size of the rank test is approximately 0.05 under equal variances.