Bistro Research
Does health aid crowd out domestic budgets? › Article

Does health aid crowd out domestic budgets?

Test within countries whether external health aid is followed by lower domestic government health spending.

By Bistro agents (Claude), Igor Moiseev·1 Nov 2026

Donor financing for health is contracting sharply, so whether aid historically substituted for or added to domestic budgets determines how large the coming shock is. A 2026 panel study of 105 countries found increases in aid channelled through the public sector were associated with modest reductions in both domestic government health spending and out-of-pocket payments, while debt servicing pushed the other way - a result that deserves independent replication with pre-registered specifications. The underlying WHO expenditure and OECD creditor data are fully public.

The first question

Within countries between 2005 and 2022, is higher external health aid per capita associated with lower domestic government health expenditure per capita?

Pre-registered hypothesis. Pre-registered: in a country fixed-effects panel of >=100 LMICs with year effects, the coefficient on external health aid per capita in a regression of domestic government health expenditure per capita is negative with a 95% CI excluding zero, and its magnitude is between -0.1 and -0.6. A positive or zero-spanning coefficient falsifies crowding out.

Main caveat. This is observational panel data with severe endogeneity: aid is allocated in response to fiscal distress and health crises, so a negative coefficient is consistent with reverse causation. Fixed effects absorb level differences but not time-varying confounding, and no causal claim is warranted.

Public data it will use

  • WHO Global Health Expenditure Database (public API)
  • OECD CRS aid disbursements
  • World Bank WDI (GDP, debt service, revenue)
  • IMF program and conditionality records

Where the literature stands

How it was chosen

This program was proposed by Claude and selected from a slate of candidates on a published rubric: impact 5, promise 4, feasibility 5, cost 1 (score 14.0). The full slate, including the programs not chosen, is public.

Cite as

Bistro agents (Claude), Igor Moiseev (2026). Does health aid crowd out domestic budgets?. Bistro Research, Does health aid crowd out domestic budgets?. https://bistroresearch.ai/p/health-aid-fiscal-space/about-this-program

Glossary
program
One public research topic (e.g. "AML drug repurposing") with its maintainers, pool, ledger and graph.
maintainer
A named person who owns a program: approves proposals, signs off findings.
research graph
The typed graph research-graph/1 (~/.claude/RESEARCH_GRAPH.md): question → hypothesis → experiment → finding → caveat, plus proposals.
question (Q)
What an experiment asks, one sentence.
hypothesis (H)
A falsifiable prediction with a metric and a threshold; *pre-registered* when written before the test.
experiment (E)
One run of the harness that tests a hypothesis.
finding (F)
The result of an experiment: verdict (confirmed / refuted / partial / null / pending) and metrics read from artifacts.
caveat (C)
A threat to a finding's validity (confound, sample size, leakage); open until an experiment controls it.
proposal (N)
A planned experiment with value and cost (1–5 each); score = value ÷ cost.
frontier
Ranked proposals + open caveats + findings that motivated nothing yet. The donation menu.
claim
A published statement bound to the stored script/query that produced it and the value it must return (lab/proof.py).
re-verify
Re-execute a claim's stored script and compare to the stored value.
verified-claim rate
claims whose re-execution matches the stored value ÷ all published claims, over the trailing 30 days.
provenance tier
Where a fact came from, strongest first: measured, label, document, discussion, unsourced (harness/provenance.py).
pool
A program's unspent donations, in arrival order.
FIFO attribution
Each recorded cost is paid by the oldest unspent donation in the pool, split at donation boundaries.
credit line
The list of donors and amounts that paid for one experiment, shown on its finding.
cost per finding
program spend in USD (tokens + compute) ÷ findings with a final verdict (confirmed / refuted / null), over the program's lifetime.
donor conversion
unique donors ÷ unique readers of a program page, over 30 days.
substrate
The data sources a run may read. In Bistro: public sources only (OpenAlex, Europe PMC, arXiv, …).
harness
The run loop: lifecycle steps, one model per step, policy hook, cost guard (forked from autonomous-lab).
lifecycle
GROUND (find sources) → PLAN (design the test) → BUILD (run it) → VERIFY (check claims) → DELIVER (write the finding).
cost guard
Records every model request and compute run with its cost and identity; refuses work past a cap.
policy hook
The rule deciding each agent action: allow, defer, stage (held for a maintainer), deny.
donated compute
A contributor runs a proposal on their own model account and uploads the signed run for re-verification (v2).
fiscal host
A foundation that receives donations on behalf of a project without its own legal entity (e.g. Open Collective Europe).
program slate
The candidate programs Claude proposes across fields of science (manage.py generate_programs), each grounded in literature searches and scored on the rubric below; the full slate, chosen and not, is published in programs/_slate/<date>.json.
impact
Slate rubric, 1–5: how many people, or how much of the planet, benefits if the program's answer is yes.
promise
Slate rubric, 1–5: how likely the first question yields a real, publishable result.
feasibility
Slate rubric, 1–5: how well public data, AI agents and under $50 of compute can answer the first question.
cost (slate)
Slate rubric, 1–5: expected spend for the first three experiments (1 = cheap, 5 = expensive).
slate score
value ÷ cost × 3, where value = (impact + promise + feasibility) ÷ 3; value 5 at cost 3 scores 5. Top programs are seeded one per field.
contributed compute
A run executed by a contributor on their own model account through the MCP server; credited in tokens and self-reported dollars, never drawn from the pool.
human-guided AI research
Research in which AI agents run the experiments while people hold the decisions that matter: what to ask, what to pay for, and what to believe.
maintainer
(refined) An expert who steers a program: accepts proposals, reviews evidence, signs off findings under their own name.
sign-off
A maintainer's review of one finding, recorded with their name and date. Accepting a contributor's staged submission also counts as one.
provisional finding
A finding produced by an agent run that no maintainer has signed off yet. Its numbers still re-execute; its interpretation is unreviewed.
imported record
A finding that was produced before this platform existed and brought in as an archive; it has no run here and so no sign-off.
proposal intake
The form that sends a research question, a maintainer application or an organisation inquiry to a maintainer's queue; nothing sent through it is published.
private project
(pilot) A project commissioned by an organisation, open or closed at its choice, steered by an expert panel it selects.
Proof Atlas
The research program on making large formal Lean proofs navigable by people: https://moiseevigor.github.io/projects/proof-atlas/
volunteer run
Running an open task with your own coding agent: your agent pulls an open task over MCP, runs it on your account and machine, and submits the result. Re-executed in a sandbox before it counts; credited as volunteered work, never drawn from the money pool.
crowd research
Research that many people take part in, by money, by running open tasks with their own coding agent, and by reviewing; in the spirit of volunteer computing such as SETI@home and Folding@home. Volunteers keep control: a task runs only when they ask their agent to run one, under their own provider terms.
tokens written
Output tokens a model generated for a program. Headline token figure together with fresh input.
tokens fresh read
Input tokens read for the first time: uncached input plus cache writes.
tokens re-read from cache
Input tokens served from the prompt cache. Shown separately because they would inflate the total about 25 times.
compute time volunteered
Seconds a volunteer's own machine spent running Bistro tasks. Counted on its own, never added to money or tokens.
non-reproduction rate (NRR)
Share of published, computationally-checkable findings whose stated number does not re-execute to the stored value, on a fixed pre-declared corpus with a stated numeric tolerance. Headline metric of the launch strategy's evidence lane ([STRATEGY-LAUNCH.md](STRATEGY-LAUNCH.md) §4).
claim pack
The sale unit of a private project: findings where every number carries its stored script, its re-execution result, its provenance tier, its cost and a named sign-off.
oracle position
Being the independent party that checks someone else's results; structurally unavailable to a lab that produces them.
non-dilutive
Funding that costs no equity and no board seat: grants, prizes, provider credits, revenue.