Keeping Clinic-to-Clinic Credit Systems Stable When Geography Is Lopsided

Picture a group of hospitals and rural clinics that trade services instead of money. One facility sends a visiting specialist, accepts a referral, or lends a mobile screening van, and earns a credit. It later spends that credit when it needs help itself. Credit systems like this, called scrip or token economies, let self-interested participants cooperate when cash payment is impractical or forbidden. Economists have shown when they stay stable. The best-known result, The Power of Two in Token Systems, proves that if each request has at least two eligible providers and the job goes to the one holding the fewest credits, balances stay bounded. Their tail probability decays like $(1/2)^M$ for a balance of size $M$.

That proof assumes interchangeable participants with no geography. Real healthcare geography is the opposite. In Kebbi State, Nigeria, about 62% of women of reproductive age live more than 5 km from the nearest facility (GeoAI for Maternal Healthcare Accessibility). Specialists cluster in cities even when regional totals look adequate (Access to Chemotherapy Services by Availability of Local and Visiting Oncologists). This note argues that geography changes the nature of the instability. Rural facilities mostly need services and seldom have nearby peers who need theirs, so their balances do not wander at random. They slide steadily into debt, and no rule for picking a provider can fix that. We propose to characterize exactly when a spatially constrained credit economy can be stable, to compute the smallest outside credit injection that restores stability and say where it should go, and to design a provider-choice rule that trades credit balance against travel time.

Core Claim

The original analysis treats instability as a random-walk problem: choosing the poorest available provider adds a restoring pull. Once each request can only be served by facilities within a travel-time limit $r$ (for example, the 5 km maternal access threshold), two separate failure modes appear.

The first is a local version of the original problem. A facility with only one eligible provider nearby has effectively $d=1$ and loses the restoring pull. The second, which we expect to matter more, is structural flow imbalance. Requests from remote regions far outnumber the services those regions can offer within reach. A region like that drains credits on average, whatever choice rule is used.

Our central claim is that stability is governed by a flow-feasibility condition on the facility network. Let $\lambda_i$ be facility $i$'s request rate and $N_r(i)$ the set of facilities within travel time $r$ of it. Stability requires a fractional assignment $x_{ij} \ge 0$, with $\sum_{j \in N_r(i)} x_{ij} = \lambda_i$, such that every facility's expected earnings at least match its spending:

$$\sum_{k:\, j \in N_r(k)} x_{kj} \;\ge\; \lambda_j \quad \text{for every facility } j.$$

Here $x_{kj}$ is the rate at which $j$ serves $k$'s requests. When this system is infeasible, the smallest total credit injection that restores feasibility is a min-cost-flow deficit we can compute directly. It also tells planners which facilities should receive subsidy credits and how many.

We conjecture, and will try to prove, a matching sufficiency result. Suppose the condition holds with strict slack, and every facility has at least two eligible providers within its radius. Then the minimum-balance rule restricted to radius-feasible candidates keeps balances bounded, with exponential tails, which generalizes the $(1/2)^M$ bound. If this holds, the abstract stability theory and the geographic access literature collapse into one computable object: the deficit map.

Research Program

The work has three parts, built on one shared model.

The model. Facilities sit at nodes of a road network with travel times computed from real roads. Each facility issues service requests: a specialist session, a referral bed, or a screening-van day. Requests can only be filled by facilities within radius $r$. The provider earns one credit and the requester pays one. To capture the threat that makes credit systems collapse in practice, a facility drops out once its balance crosses a debt or hoarding threshold. This matches the failure modes described in The Power of Two in Token Systems.

Theory. We first prove necessity of the flow condition, which should follow from a direct averaging argument over long-run earnings and spending. We then attempt sufficiency, adapting the Lyapunov-function argument of the original paper to non-identical neighborhoods. Finally, we derive the minimal-injection formula as the dual of the flow problem.

We also study a travel-aware choice rule. It assigns each request to the eligible provider minimizing balance plus $\beta$ times travel time. The parameter $\beta$ is a single knob: $\beta = 0$ recovers the pure stability rule and large $\beta$ recovers nearest-provider. We expect a tradeoff curve between balance stability and patient travel. The open question is whether that curve has a sharp knee, so that small travel sacrifices buy most of the stability.

Instantiation. We build the model on real geography. The primary setting is the Kebbi pilot data: 1,207 GRID3 facilities, 1,510 populated WorldPop cells, and OpenStreetMap roads (MamaCare AI). Demand is calibrated to population and to the high-risk clusters identified there. The second setting is a terrain-varied facility network modeled on the Sichuan township analysis of Spatial Optimization and Allocation of Primary Healthcare Facility Resources. There, rugged terrain shrinks the facilities reachable within the radius, which lets us test the $d=1$ failure mode directly.

For realistic exchange structure, we reuse the National Kidney Registry statistics reported by The Power of Two: 1,881 pairs across 84 hospitals, averaging about 7 candidate matches for easy-to-match pairs. Hospital locations are not in our file. We would geocode them only if hospital identities are obtainable; otherwise we use synthetic placements matched to US driving-distance patterns (Income disparities in driving distance to health care infrastructure in the United States).

What we deliberately leave out. No blockchain is involved, because the ledger can be any trusted accounting system. This sidesteps the gas-cost problems reported for FL-Incentivizer, which we cite only as a constraint on design. The question here is about mechanism design, not cryptography.

Prior work and how this differs

The token-economy papers supply the method. The Power of Two in Token Systems gives the minimum-balance rule and its stability proof for symmetric agents, which we extend to unequal neighborhoods and to drift. Traffic-Aware Transmission Mode Selection in D2D-enabled Cellular Networks with Token System shows that individual devices follow threshold earn-or-save policies. We use that result to justify the dropout thresholds in our model, but that work has no spatial demand imbalance. MEDIATE shows that fixed token values break when reward scales shift, with cooperation falling to 0.69 for the fixed-value MATE protocol against 0.97 for its adaptive pricing. We take this as a warning that unit credits may need regional pricing, and we test that as an alternative to injection.

The geographic literature motivates the problem. MamaCare AI locates underserved clusters but, as the file notes, provides no mechanism for getting services into them. Bypass behavior in China's tiered system (Supply–demand governance of hierarchical healthcare systems) is direct evidence of the demand asymmetry that drives structural drift. Efficient, Fair, and Incentive-Compatible Healthcare Rationing is the closest healthcare mechanism-design work. It is a one-shot centralized matching with reserve categories, so it has no circulating credits and no space. We borrow its standard of strategyproofness as a property to check for our rule.

Baselines

The naive reference is nearest-provider assignment with no credits, where each request goes to the closest eligible facility. It minimizes travel and gives an obligation-imbalance profile to compare against. We will report the imbalance it implies by counting services rendered minus received per facility, and the dropout it would trigger under the same participation thresholds.

The direct prior method is the unconstrained minimum-balance rule from The Power of Two in Token Systems, which ignores distance. Its reported guarantee is the theoretical $(1/2)^M$ tail bound for $d \ge 2$ under symmetry. The file reports no empirical spatial result for this rule; we found no reported result on Lacuna for spatially constrained token stability. We run it in two forms. One ignores travel entirely, to show what stability costs in patient travel. The other is restricted to the radius, which is the version whose behavior our theory predicts.

The strongest competitor is a centralized planner. Each period it solves a min-cost-flow problem with full knowledge of demand, minimizing travel subject to every facility's net obligation staying within bounds. This is the upper bound a well-informed health ministry could reach.

All methods receive the same road networks, demand traces, random seeds, and dropout thresholds. The two tunable methods, the travel-aware rule's $\beta$ and the planner's balance bound, get the same grid-search budget on held-out demand traces. Beating the naive and prior-rule baselines means fewer facility dropouts and a smaller rural-urban gap in requests served within the access threshold, at equal or lower mean travel.

Against the planner, success means coming within a small, stated fraction of its travel and dropout figures, without a central solver, and while staying robust when facilities misreport demand. Two ablations isolate the key ingredients. Injection placed by the deficit map is compared with the same total credits spread uniformly, which tests whether the flow analysis matters. The travel-aware rule at $\beta = 0$ is compared with its tuned $\beta$, which tests whether the knob matters.

Why not just…?

Why not just give rural clinics a fixed subsidy and skip credits altogether? Our model includes subsidies; the question is how much and where. Uniform subsidies leave persistent gaps when demand is local and fluctuating, which is the pattern visiting-oncologist programs show in Access to Chemotherapy Services by Availability of Local and Visiting Oncologists. The uniform-versus-deficit-map ablation tests this directly. If uniform subsidies match deficit-map placement at equal total budget, the targeting contribution fails, and we will report that.

Why not just let a central planner assign everything? This is the real threat. With honest demand reports and a functioning ministry, the planner will likely beat any decentralized rule on travel. Credits are useful only where participation is voluntary, demand is privately known, or central coordination is unreliable. The bypass evidence in Supply–demand governance of hierarchical healthcare systems suggests these conditions hold in practice, but it does not settle the question. We therefore test the planner when facilities inflate their requests and when it re-solves only infrequently. If the planner still dominates under those conditions, the honest conclusion is that the deficit map should feed planners rather than markets, and the theoretical contribution stands regardless.

Why not just use fixed one-credit-per-task trading with nearest-provider choice? Nearest-provider choice makes many rural facilities effectively single-choice, which is the regime where The Power of Two in Token Systems proves balances drift without bound. Structural demand asymmetry adds steady drift on top of that. The nearest-provider-with-credits arm measures how fast this produces dropouts on real Kebbi geography.

Why not just reprice credits regionally, as MEDIATE adapts token values? This may well work as an alternative to injection, since MEDIATE shows adaptive pricing rescues cooperation. However, if the flow condition is infeasible, repricing alone cannot create earnings for a facility no one can reach. We include adaptive regional pricing as an arm and expect it to fail exactly on the deficit-map regions.

Proposed Experiments

The first decisive check validates the theory. On Kebbi and the terrain networks, we sweep radius $r$ from tight values to the 5 km threshold and beyond. For each radius we compute the flow deficit and simulate long runs of the radius-restricted minimum-balance rule. The prediction is a sharp transition: bounded balances where the deficit is zero and every neighborhood holds two or more providers, and linear drift concentrated on the deficit regions otherwise. If drift appears in regions with zero deficit, or does not appear where the deficit is positive, the core claim is wrong.

The second check tests the injection formula. Injecting exactly the computed deficit at the mapped facilities should eliminate drift. Injecting the same amount uniformly should not. We report dropout rates, the maximum absolute balance, and the share of high-risk-cluster requests served within 5 km.

The third experiment traces the travel-stability curve over $\beta$ and compares it against the planner, with and without demand misreporting. We report the knee location and the gap to the planner.

The fourth experiment repeats the first check on kidney-exchange structure, calibrated to the NKR candidate multiplicity. This confirms that the result does not depend on one geography.

All code, demand calibrations, and deficit maps will be released, so health planners can compute where a service-exchange scheme needs outside support before launching one.