Online Dispute Resolution: value of bid revision, algorithmic closure, and mechanism transparency

Participants are randomly assigned to one of four dispute-resolution procedures that differ in how settlements are reached after the two sides state how much of a fixed 100-point pie they want. Each participant has an outside option between 0 and 100, uniformlly drawn. Participant knows her outside option, but not the one of the opponent. Barg is direct bargaining without platform mediation. ODR is a one-shot structured procedure that converts the two requests into a nonbinding settlement recommendation when the requests are jointly feasible using second-best mechanism (Chaterjee and Samuelson,1983; Compte and Jehiel, 2009). SeqODR adds one opportunity to revise requests after an initial impasse. AlgODR allows the platform to use an internal rule to generate a compromise recommendation in some cases when the initial requests are incompatible.

The intervention is motivated by both theory and practice. The theory benchmark is the canonical sealed-bid bargaining mechanism of Chatterjee and Samuelson (1983), implemented in nonbinding form with veto rights as in Compte and Jehiel (2009). The practical motivation is that real ODR systems vary substantially in whether they allow repeated submissions, whether they use structured recommendations, whether they attempt algorithmic gap-closing when bids are incompatible, and how much they disclose about those rules. The slide deck specifically discusses variation across systems such as PayPal Resolution Center, Smartsettle, Cybersettle, Matterhorn, and Tyler ODR. This experiment is designed to isolate those platform-design features in a controlled setting.
In every round, two participants bargain over 100 points. Each participant privately observes an outside option drawn independently from a uniform distribution between 0 and 100. If no agreement is reached, participant i receives their own outside option v_i. If an agreement is reached, participant i receives the accepted. One point corresponds to £0.02. Participants know their own outside option and the distribution from which outside options are drawn, but not the counterpart’s realization.
The treatment details are as follows.
In Barg, participants bargain directly. Each participant can make up to two simultaneous offers, can revise or cancel offers, and can accept the opponent’s offer. Bargaining ends when one offer is accepted, when a participant exits, or through stochastic breakdown after the first minute.
In ODR, each participant submits a request r_i. If r_1+r_2>100, the platform recommends breakdown. If r_1+r_2≤100, the platform recommends x_i=r_i+(100-r_1-r_2)/2. Both participants then simultaneously accept or reject the recommendation.
In SeqODR, the first stage is identical to ODR, but when the initial requests are incompatible, both participants receive one opportunity to revise their requests. The same recommendation rule is then applied to the revised requests.
In AlgODR, the first stage is identical to ODR when requests are jointly feasible. When requests are incompatible, the platform uses a fixed ex ante prediction rule to generate predicted outside options p_i. If p_1+p_2>100, it recommends breakdown. If p_1+p_2≤100, it recommends x_i=p_i+(100-p_1-p_2)/2. The prediction rule will be estimated based on baseline data of ODR treatment by running regression of the outside option on the bid of the participants. Note we will also for riobustness run a second version of treatment, where the algorithm will be trained on intial submissions of AlgODR treatment, to avoid potential effects of selected labels problem.

2026-06-09

2026-06-30

The primary outcome variables are: breakdown at the pair-round level; individual realized payoff in points; the initial submitted request in ODR, SeqODR, and AlgODR, the revised request in SeqODR; and acceptance or rejection of the platform recommendation in the mechanism-based treatments. The main outcome is payoff.
The outcomes only analyzed for markets where agreement is feasible, namely sum of outside option is below 100.

This is an online experiment conducted on Prolific. Participants are randomly assigned to one of four treatment arms and remain in that arm for all four rounds. Within each treatment arm, participants are organized into groups of four and are randomly rematched into bargaining pairs in each round. In every round, a pair decides how to divide 100 points while each participant privately knows their own outside option, which is independently drawn from a uniform distribution between 0 and 100. One point equals £0.02.
We exclude participants who finish the round faster than 5 seconds.

Randomization is done by the experimental software using OTree random assignment. At study entry, participants are assigned with equal probability to Barg, ODR, SeqODR,or AlgODR. Within each treatment arm, the software then forms groups of four and randomly rematches participants into pairs in each of the four rounds.

Primary unit of randomization: groups of 4 individual participants.

Yes

200 groups per treatment

3,200 participants in total, corresponding to 12,800 participant-round observations. In expectation, 6,400 could be used for analyses as the sum of outside options will be below 100.

200 groups of four per treatment arm, that is 800 participants in each of Barg, ODR, SeqODR, and AlgODR.
Additionally 150 for sophisticated Algorithm treatment will be collected later.

For continuous individual-level outcomes, including realized payoff and submitted requests, the study has approximately 80 percent power to detect effects of about 0.12 to 0.15 standard deviations for pairwise comparisons between two arms when the intra-cluster correlation is between 0.05 and 0.20. If the standard deviation of realized payoffs is approximately 25 points, this corresponds to a minimum detectable effect of approximately 3 to 4 payoff points. For binary pair-level outcomes such as breakdown, using the conservative benchmark of a 50 percent baseline breakdown rate, the study has approximately 80 percent power to detect differences of about 7.5 to 9 percentage points for pairwise comparisons between two arms when the intra-cluster correlation is between 0.05 and 0.20. For binary individual-level outcomes such as acceptance of a recommendation, the corresponding minimum detectable effect is approximately 6 to 8 percentage points, although the effective sample size for acceptance outcomes may be smaller because acceptance is only observed when a platform recommendation is made. These calculations are intended to guide the main pairwise comparisons and will be complemented by regression specifications that use all feasible observations and include clustered standard errors at the group level. The primary inference will focus on effect sizes and confidence intervals, not only on statistical significance.