Who Should Price Genomic Data? Reselling versus Agency Contracts for Moving Statistical Power to Underrepresented Ancestries

Most genetic studies of disease have been run on people of European ancestry. As a result, risk scores built from those studies work far worse for everyone else: accuracy falls to about half of the European level for South and East Asians and to 39.4% for Africans (Polygenic risk score portability for common diseases across genetically diverse populations). Statistical methods can partly close this gap by borrowing strength from large European datasets when analyzing small non-European ones. One such method raised the effective sample size of Biobank Japan studies by 49% (Leveraging the local genetic structure for trans-ancestry association mapping). These methods quietly assume that the large dataset is available for free. That is often false for the custodians whose data would help most, such as direct-to-consumer genetics platforms and proprietary or ancestry-specific reference collections. Whether their data reaches a small cohort depends on a commercial arrangement that nobody has modeled.

Operations research has a mature theory for this kind of arrangement. It compares reselling, where an intermediary buys goods at a wholesale price and sets its own retail price, with agency selling, where the producer sets the price on a platform and pays the platform a commission on each sale. This proposal applies that theory to genomic data. The value of data to a buyer is not a linear demand curve here. It is a measured statistical-power curve that saturates with sample size and depends in a non-monotone way on how genetically similar the buyer's population is to the seller's. The central question is which contract format delivers the most prediction accuracy to underrepresented cohorts, and whether the answer differs from what the textbook linear-demand models predict.

Core Claim

We claim that the choice between reselling and agency contracts for genomic summary data acts as a hidden lever on cross-ancestry statistical power. Under reselling, the custodian adds a wholesale markup and the broker adds a retail markup on top. Economists call this double marginalization. Two stacked markups price out budget-limited minority cohorts, so less statistical power flows to exactly the groups that need it. Agency selling removes one markup and, more interestingly, gives the platform a reason to reveal what it privately knows about genetic similarity between cohorts. This matters because the value of borrowed data depends heavily on that similarity.

The specific new content has two parts. First, we build buyer demand from empirically measured power-transfer curves instead of assumed linear demand. Second, we show that the platform's information advantage is naturally a genomic quantity: knowledge of where local genetic correlation is shared across ancestries. It is not the scalar forecast-accuracy parameter used in commerce models. If the power curves turn out to be close to linear, the genomics adds nothing and the standard models apply directly. That outcome is an explicit falsification condition, not something we assume away.

Where the Two Literatures Stand

The contract literature supplies the method and a set of predictions to test. Distribution Contract Analysis on e-Platform by Considering Channel Role and Good Complementarity solves leader–follower pricing games, known as Stackelberg games, by working backward from the last move. It finds that goods sold wholesale carry higher retail prices than goods sold under agency contracts, and that suppliers prefer agency contracts whether they lead or follow in setting prices. We reuse its solution method. We also test whether its supplier-preference result survives when demand is a saturating power curve.

Reselling or agency selling? Sales mode selection for a manufacturer in private label competition under information asymmetry models a retailer with private demand forecasts. It shows two things. Agency selling makes truthful information sharing the retailer's best strategy. And better retailer forecasts push manufacturers toward agency selling. That paper motivates our information subgame, but its setting differs from ours in two ways. Its information is a single forecast-accuracy number, and its informed party sits downstream of a data-poor producer. In genomics the data-rich party is the upstream seller, and the "forecast" is a map of which genome regions share architecture across ancestries. The file flags this reversed setting as open.

The genomics literature supplies the value function. We read these methods as producers of a price-relevant quantity, not as competitors to be beaten. Leveraging the local genetic structure for trans-ancestry association mapping (LOG-TRAM) borrows from a European auxiliary cohort only inside 1-megabase windows where local heritability and genetic correlation are shared. Multiethnic polygenic risk prediction in diverse populations through transfer learning (TL-Multi) models target effects as European effects plus a sparse correction. It reports that transfer helps most at moderate genetic correlation, between 0.4 and 0.6: near-identical populations need only ordinary meta-analysis, and very different ones break the sparsity assumption.

That non-monotonicity is the crux of our argument. It means the willingness to pay for a given dataset varies sharply across buyers, and it varies in a way only someone who knows the genetic correlation map can predict. Both methods treat access as costless, and that is the assumption we remove.

Adjacent work constrains the design. Data-marketplace papers such as Model-based Pricing for Machine Learning in a Data Marketplace and Dealer: End-to-End Data Marketplace with Model-based Pricing price data by the accuracy improvement it produces. They do not compare intermediary formats, and they do not ask who holds markup power. Personalizing or Not: Dynamically Personalized Federated Learning with Incentives handles participation incentives through a learning-rate adjustment, not through prices or contracts. We borrow the performance-indexed pricing idea from the first group as a pricing rule inside our game.

Research Program

The work has three linked stages.

Stage 1: measure how much prediction accuracy a buyer gains from purchased data. Using UK Biobank as the auxiliary (seller) cohort, we will subsample it to a grid of purchased sample sizes. For each size, we run LOG-TRAM and TL-Multi against Biobank Japan summary statistics as the target cohort, across the 29 traits LOG-TRAM studied. We then record the out-of-sample prediction gain, measured as the increase in variance explained ($R^2$) over a target-only model.

To vary genetic similarity beyond one population pair, we will stratify traits by their estimated trans-ancestry genetic correlation. We will also restrict borrowing to subsets of genomic windows, which simulates sellers whose data overlaps the buyer's architecture to different degrees. The output is an empirical value surface: $R^2$ gain as a function of purchased sample size, trait genetic correlation, and fraction of shared local structure. We then fit a parametric form that the game can use. The candidate is saturating in purchased sample size, because each additional sample adds less once effects are well estimated.

Stage 2: embed the value surface in a three-player pricing game. The players are a data custodian, an intermediary, and a population of buyer cohorts. Each buyer has a budget and a value per unit of $R^2$ gained.

Under reselling, the custodian sets a wholesale price, and the broker then sets an access price. Under agency selling, the platform first posts a commission rate, and the custodian then sets the access price. We also include the compensation-for-information mechanism from Reselling or agency selling? Sales mode selection for a manufacturer in private label competition under information asymmetry as a reselling variant.

The equilibria will be solved numerically by backward induction, because the fitted value curves will generally not have closed forms. The outcomes we report are custodian, intermediary, and total profit, plus one outcome no commerce model tracks: total $R^2$ gain delivered to cohorts with non-European ancestry.

Stage 3: the information subgame. The intermediary observes the local genetic-correlation map from all the cohorts it hosts. The custodian and buyers see only a noisy version. We ask under which format the intermediary reveals that map. Revealing it lets well-matched buyers pay more and lets poorly matched buyers avoid wasted purchases. The commerce result predicts revelation under agency selling. We will check whether this still holds when value is non-monotone in genetic correlation. In that case a broker might profit from pooling mismatched buyers.

Baselines

The naive reference point is target-only analysis. The minority cohort buys nothing and uses only its own data. For Biobank Japan, this is the single-ancestry analysis that LOG-TRAM improves on. For the lupus case, the file reports single-ancestry Lassosum at an area under the curve (AUC) of 0.6943 (Multiethnic polygenic risk prediction in diverse populations through transfer learning). In the game this corresponds to no trade. Any contract format must deliver positive $R^2$ gain relative to this point, or it has failed.

The upper reference point is free release. The custodian publishes its data at zero price, and buyers receive the full statistical-transfer benefit. The strongest published statistical benefit in the file is LOG-TRAM's. It reports a 49% average effective-sample-size increase for Biobank Japan across 29 traits and a 17% to 24% $R^2$ improvement in an independent WeGene Chinese cohort (Leveraging the local genetic structure for trans-ancestry association mapping). TL-Multi reports an AUC of 0.7131 against 0.7098 for meta-analysis on lupus. We will rerun both methods ourselves on the same summary statistics and splits. Every contract outcome is then reported as the fraction of this free-release gain that it recovers. The benchmark is the right one because it shows how much statistical power the market leaves on the table.

The strongest economic competitor is performance-indexed posted pricing without format choice, in the style of Model-based Pricing for Machine Learning in a Data Marketplace. Here a single seller prices access by the buyer's predicted accuracy gain. We found no reported result on Lacuna for model-based pricing applied to genomic data, nor for reselling-versus-agency equilibria in any data market. We will implement this rule on the same value surface and the same buyer population. We will also give it the same pricing flexibility as our custodian. Agency selling beats it if it recovers a larger share of free-release $R^2$ gain at equal or higher custodian revenue.

The key ablation replaces our measured value surface with the linear demand used in the commerce papers, calibrated to match the same average willingness to pay. If equilibrium format choice and delivered $R^2$ are unchanged, the genomic calibration adds nothing.

Why not just…?

Why not just release all summary statistics for free? For academic consortia, largely free release is already the norm, and we treat it as the upper bound. The problem we address concerns custodians who do not release and have commercial reasons not to: direct-to-consumer platforms, proprietary reference panels, and ancestry-specific collections. For these custodians, free release is not an available action. The fair question is how much of the free-release benefit each feasible contract recovers. If agency selling recovers nearly all of it, that is a useful policy result. If every format recovers little, that argues for public subsidy, and our model quantifies how large the subsidy would need to be.

Why not just use the existing linear-demand contract models? They may suffice. The linear-demand ablation is designed to settle this. Our reason to doubt it is TL-Multi's finding that transfer value peaks at moderate genetic correlation and falls on both sides. A linear curve cannot represent a buyer who gains little from a very similar seller because ordinary meta-analysis already works, and also little from a very dissimilar one because the sparsity assumption breaks. If the fitted surface is close to linear in the parameter ranges the data actually support, we will say so, and the contribution shrinks to an application note.

Why not just price data by accuracy improvement and ignore the intermediary format? Model-based pricing sets the right price signal, but it assumes a single seller. In practice, genomic data reaches small cohorts through consortia, brokers, and hosting platforms, and each can add a margin. Whether the second margin exists is exactly the difference between reselling and agency selling. Performance-indexed pricing inside a reselling chain can still suffer double markups. The head-to-head comparison in the baselines tests whether the format matters once pricing is already performance-indexed.

Decisive Checks

Three results would settle the proposal.

The first is the shape of the Stage 1 value surface. The claim needs clear saturation in purchased sample size and a peak at intermediate genetic correlation that survives bootstrap resampling of traits. If the surface is flat or linear, the premise fails.

The second is the delivered-power comparison. Under realistic buyer budgets, agency selling should deliver measurably more $R^2$ gain to non-European cohorts than reselling, expressed as a fraction of the free-release benchmark. The gap should also widen as buyer budgets tighten, which is the signature of double marginalization.

The third is the information subgame. If the platform reveals its genetic-correlation map under agency selling but withholds it under reselling, the commerce result transfers. If it pools mismatched buyers under both formats, we will have found a case where the commerce result fails. That would be a negative result that matters for anyone designing a genomic data exchange.