MacKenzie Scott's giving, in QALYs | Max Ghenis
Home Writing Projects Setup Protocol CV X LinkedIn GitHub Bluesky
Assumptions<br>Drag to see the estimate move. The model reruns in your browser.<br>Evidence stanceskeptical<br>Skeptical weights each effect by how well its study identifies causation. Credulous trusts every cited effect at face value.
Realization (mode)mode 0.80<br>Central value (mode) of a 0.55–1.10 triangular draw for the share of the studied effect a marginal unrestricted grant delivers — the whole distribution is sampled, not this number alone.
Total giving (2026 $)$30.3B<br>Real 2026 dollars. Default inflates each year's gifts ($26.39B nominal, 2020–2025) to ~$30.3B with CPI-U.
Discount rate3.0%<br>Annual discount on future life-years.
Allocation across causesHealth - insurance & access5%
Health - community health centers5%
Health - mental & behavioral2%
Economic security - cash & financial4%
Economic security - food security2%
Economic security - housing & homelessness4%
Economic security - workforce & mobility13%
Education (HBCUs, comm. college, scholarships)19%
Equity & justice17%
Civic / democracy / philanthropy infra11%
Arts & culture3%
Environment / climate8%
Other / general community6%
FastStandardHigh<br>Reset
Running Monte Carlo…
The most important control is evidence stance : from<br>skeptical (~70,000 QALYs — each effect weighted by how well its study<br>identifies causation) to credulous (~200,000 — every cited effect at<br>face value). That gap, not the dollar figure, is the real uncertainty.
About this model
MacKenzie Scott's Yield Giving network has made over $26 billion<br>in 2,700+ gifts since 2019 — $26.3 billion through 2025 by<br>CNBC's year-end accounting. This page asks what that buys in<br>quality-adjusted life-years, the unit health economists use to<br>compare a death averted against years lived in better health.
It is a GiveWell-style model: 13 intervention<br>archetypes, each cost-per-QALY drawn where possible from a published<br>causal estimate<br>(Medicaid mortality,<br>community health centers,<br>supportive housing,<br>collaborative-care depression),<br>each effect shrunk toward zero in proportion to how well its study<br>identifies causation, and the whole thing rerun through thousands of<br>Monte Carlo draws each time you move a slider.
I built the model with Claude; every estimate here is a model output,<br>not a measured fact. The Python package, tests, and sources are on<br>GitHub; this page runs a checked TypeScript<br>implementation in the browser, reading the exported parameter file.
How it works
Each Monte Carlo draw takes the giving — each year's gifts<br>inflated to 2026 dollars<br>($26.39 billion nominal ≈ $30.3 billion, so the dollars and the<br>cost-effectiveness evidence share one price level) — allocates it<br>across the archetypes (a Dirichlet whose centers come from<br>Scott's own gift database —<br>dollar amounts are disclosed for about two-thirds of the money, and<br>each organization's dollars are split across its reported focus<br>areas and mapped to the 13 archetypes; the undisclosed remainder is<br>imputed from her announced year totals, scaled by each recipient's<br>pre-gift IRS 990<br>revenue), assigns each a<br>cost-per-QALY, and multiplies by two independent discounts:
The gift-size data yields one measured regularity along the way:<br>across the 1,313 disclosed gift–revenue pairs, gift size scales<br>with the recipient's pre-gift revenue to the power 0.41<br>(R² 0.37) — a 10× larger organization receives about 2.5× more<br>money, not 10× more. That fitted elasticity, not proportionality,<br>weights the imputation of the undisclosed gifts.
Causal credibility — how well the effect is<br>identified, drawn from the study's design tier: a<br>lottery RCT (income → mortality)<br>is trusted; an associational SNAP correlation<br>is shrunk hard; an assumption-only bucket (arts, civic) goes to near<br>zero. This is the evidence stance slider.
Realization — the fraction of the studied effect a<br>marginal unrestricted grant actually delivers.
The model also prices the same dollars at the global-health<br>frontier. GiveWell's current impact<br>estimates put the 2022–2024 program averages at ~$4,000 per life<br>saved (Malaria Consortium) to ~$5,500 (AMF nets). A child death<br>averted at ~age 1 is ~25 discounted QALYs under this model's<br>own conventions (~65 remaining years, 3% discount, utility ~0.87), so<br>those endpoints — inflated to 2026 dollars — become roughly $175–$241<br>per QALY-equivalent; I model the benchmark as loguniform $150–$260,<br>handicapped with the same realization and credibility as Scott's<br>portfolio (and rescaled at other discount rates) so the comparison<br>is like-for-like. At the skeptical defaults, the frontier delivers<br>roughly 1,500× more health per marginal dollar.
That multiple is a marginal comparison — the next dollar,<br>not the whole portfolio. Frontier-priced opportunities are scarce:<br>GiveWell directed $397 million in all<br>of 2024 and moves its<br>cost-effectiveness bar with the money it expects to raise —<br>funding down to ~6× cash when flush, back up...