Introduction

MacKenzie Scott gave away $7.17 billion in 2025 (Huddleston 2025); Giving USA counts $6.65 billion of it against $19.2 billion in U.S. megagift dollars — about a third of that year’s megagift value (Indiana University Lilly Family School of Philanthropy 2026; Childress 2026). (The bases differ: Yield Giving’s own annual total is the larger figure.) Her disclosed lifetime total is $26.39 billion across 2,711 gifts to 2,545 organizations since 2019 (Yield Giving 2026). Almost none of it is denominated in health, and, to my knowledge, no published estimate prices the portfolio in outcomes of any kind.

This paper does so in the unit health economists use to compare a death averted against years lived in better health: the quality-adjusted life-year. The exercise is deliberately in the style of a charity evaluator’s cost-effectiveness analysis (GiveWell 2025), scaled from a program to a portfolio, with one addition the setting forces: because most of the portfolio funds interventions whose health effects have never been causally measured, the model makes its epistemics explicit. Every effect is shrunk toward zero in proportion to how well its supporting study identifies causation, the shrinkage schedule is a slider rather than a buried constant, and the headline is reported across the full range of evidence philosophies — from counting only randomized evidence to taking every published estimate at face value.

The paper makes four contributions. It provides what I believe is the first outside, portfolio-level cost-effectiveness estimate of Scott’s giving, and I have found no comparable outside estimate for any other individual mega-donor — funder-run portfolio metrics exist, including GiveWell’s aggregate impact estimates (GiveWell 2025), Robin Hood’s benefit–cost framework (Robin Hood Foundation 2023), the Constellation Fund’s portfolio reporting (Constellation Fund 2026), and social-rate-of-return analysis of USAID’s Development Innovation Ventures portfolio (Center for Effective Global Action 2022); the difference here is independence from the donor, a subject who publishes no outcome data, and explicit evidence weighting: a best-guess median of 202,000 QALYs from $26.39 billion nominal ($30.3B in 2026 dollars), a blended $150,067 per QALY, and a median benefit–cost ratio of 4.9 under the HHS regulatory value-per-QALY distribution (Kearsley 2026). It estimates extreme geographic concentration of impact under its pricing assumptions: the 4.8% of dollars funding global health and development delivered outside the United States accounts for 69.1% of estimated health impact. It reports a new empirical regularity in the gift data itself: across 1,313 disclosed gift–filing pairs, gift size scales with recipient pre-gift revenue to the power 0.41 (R² 0.37) — Scott’s giving is far flatter across organization size than proportional. And it demonstrates a working method for auditable, AI-assisted cost-effectiveness modeling: the model was built with a frontier language model, reviewed and corrected in public through the revision history summarized in the robustness section, audited by two independent model families against primary filings, and ships with versioned parameters, machine-checked citations, and a browser implementation that reruns the full Monte Carlo under any reader’s assumptions.

The unit of account deserves a caveat before any number. QALYs measure health. Most of Scott’s portfolio aims at other things — education, economic mobility, equity, civic infrastructure — whose primary benefits are not health benefits. A finding that a dollar of education philanthropy produces few QALYs is a statement about the lens, not a verdict on the gift. The estimates are a partial, health-only accounting under positive-effect priors — not a lower bound on total wellbeing: the lens omits the primary intended benefits of most of the portfolio, and the model places no probability on zero or harmful effects, a structural choice the limitations section returns to.

Data

Model

Each Monte Carlo draw allocates the giving across archetypes (a Dirichlet centered on the derived shares), assigns each archetype a cost per QALY drawn from a distribution anchored in published estimates, and multiplies the implied QALYs by two independent discounts. Causal credibility asks whether the measured effect is real. Each archetype carries the identification tier of its best supporting evidence, and the tier sets a Beta-distributed shrinkage factor toward the null: a lottery-randomized wealth shock (Cesarini et al. 2016) or the Medicaid mortality quasi-experiments — the ACA expansions (Miller et al. 2021), with cost–benefit arithmetic from the pre-ACA state expansions (Sommers 2017) whose spending component draws on the Oregon Health Insurance Experiment (Finkelstein et al. 2012) — keeps most of its effect; a child-targeted cash quasi-experiment (Aizer et al. 2016), with related designs catalogued in the parameter file, anchors the cash bucket; the community-health-center rollout design (Bailey and Goodman-Bacon 2015) and similar contested quasi-experiments keep less; associational and assumption-only categories are shrunk near zero. Realization asks whether a marginal philanthropic dollar delivers the studied effect at all — transport to different populations, overhead, and funging, net of possible capacity effects: surveyed Scott recipients report stronger capacity and reach (Buteau et al. 2023), self-reports on whose basis the model permits realization above one (an author judgment, not a CEP effect estimate). It is one global triangular draw with mode 0.80 on support 0.55–1.10.

The two multipliers are designed to partition the doubt — transportability concerns in realization, identification concerns in credibility, cost-per-QALY priors stated conditional on the effect being real and delivered — though the separation is imperfect at the prior boundaries: several anchors already average over mixed null-and-positive evidence before the credibility discount applies (a limitation noted below). Allocation uncertainty is a Dirichlet with concentration 60 around the derived centers — a judgment prior, implying for example a roughly 1.3–10% ninety-percent range on the global-health share — and represents subjective allocation uncertainty rather than sampling uncertainty propagated from the gift data pipeline. The tier ordering is the defensible part; the levels are judgment, proposed by a language model and reviewed by the author, and the results section reports the full consequence of disagreeing with them.

Health effects arrive in different natural units — lives saved, life-years, depression-free days — and are converted to discounted QALYs at a 3% annual rate under source-specific conventions: adult mortality anchors sample roughly 26 remaining life-years at a utility draw centered near 0.78, the frontier’s child-mortality benchmark uses a fixed ~65 remaining years at utility 0.87, and direct cost-per-QALY anchors keep their sources’ conventions. Every parameter’s unit and dollar vintage is machine-validated at load. The global-health frontier benchmark is handicapped within the same framework: GiveWell’s reported recent program averages (GiveWell 2025), converted to roughly $175–$241 per discounted QALY-equivalent under this model’s own conventions (a model conversion, not a GiveWell QALY figure), modeled as loguniform $150–$260, sharing each draw’s realization multiplier, and receiving RCT-grade credibility — the strongest tier — where portfolio archetypes receive their own.

Results

At the best-guess defaults (100,000 draws), the portfolio produces a median of 201,923 QALYs, with a 90% interval of 94,000 to 419,000 and a mean of 222,060. The interval is a probabilistic-sensitivity range under the stated priors, not a confidence interval; Monte Carlo error on the median is a few hundred QALYs. The blended cost-effectiveness — dollars given over QALYs produced, at the median — is $150,067 per QALY. Monetized under the value-per-QALY distribution HHS recommends for regulatory analysis (central $726,000 at 3%; Kearsley (2026), Table 3), the median draw-level benefit–cost ratio is 4.9 (90% interval 2.1–11.0, reflecting both health-output and valuation uncertainty). Figure 1 shows the full distribution.

Robustness

Evidence philosophy dominates everything else

The dominant sensitivity is not a parameter but a philosophy. At the near-RCT-only floor — non-randomized tiers shrunk to a credibility mean of 0.01, the randomized tier kept at its drafted mean — the median collapses to 9,114 QALYs, dominated by the randomized-tier mental-health bucket; at near face value (all tier means 0.999) it rises to 414,264 (Figure 2). That 45-fold range exceeds the spread from any single distributional parameter — though structural choices can move the headline more: with no geographic repricing at all, every health dollar at U.S. anchors, the median is roughly 70,000 (the pre-correction model in the §5.3 table), so the routing assumption alone spans ~3× — and the gap between the face-value and best-guess estimates is a direct reading of how much of the estimated impact rests on evidence below randomized or strong quasi-experimental identification standards. The Spearman tornado (Figure 3) tells the same story within the default stance: the top three drivers are the global-health allocation share, the realization factor, and global-health credibility.

Construction and auditability

The model was built with a frontier language model (Anthropic’s Claude), with the author reviewing every parameter and assumption; adversarial review by a second model family (OpenAI’s), plus reader corrections, produced the revision history above. Three mechanisms make the result checkable rather than merely stated. Parameters live in one typed file in which every value carries its unit, dollar vintage, and inline citation, machine-validated at load; an annotated bibliography maps each quantitative input to its primary source and price-level conversion. Sync tests force the published web parameters, allocation shares, and geography exports to equal fresh derivations from the committed data snapshots, and the body of this manuscript interpolates its model-derived numbers from those same artifacts at render time (the abstract’s figures are typed copies of the interpolated values; externally sourced constants carry citations). And the audit trail is data: the 990 match audit ships as its applied per-row overlay, and the 50-organization geography audit ships with per-organization sources and confidence grades — both applied by the pipeline rather than folded silently into parameters.

The browser implementation is a checked TypeScript port that reruns the full Monte Carlo client-side; every figure in this paper can be reproduced, and every assumption overridden, at maxghenis.com/mackenzie-scott-qaly. The repository, including this manuscript’s source, is at github.com/MaxGhenis/mackenzie-scott-qaly.

Limitations

The credibility tiers and realization distribution are judgment priors with no literature calibration; the ordering is defensible, the levels are not measured, and the stance sweep is the honest bound on their influence. The QALY lens counts only health, so the estimate is a partial accounting that omits both the non-health benefits where Scott concentrates — education, mobility, equity — and any possible harms; it is not a floor on total wellbeing. Credibility shrinks toward zero but never below it, so the model places no mass on harmful giving; the closest available zero scenario is the near-RCT-only stance floor, which drives unsupported archetypes to ~1% credit. Allocation splits each organization’s dollars equally across its focus areas and service locations; both are coarse, and the geography audit covers only the 50 largest “global”-listed organizations. Realization is one global draw, so it cannot vary by cause. Archetype anchors treat heterogeneous organizations as draws from one distribution. Within each cause, delivery regions are priced identically, so the geographic decomposition exposes assumptions rather than measuring regional differences. The undisclosed third of dollars is imputed from a fitted elasticity, not observed. The routed global-health bucket prices its unattributable and Europe-&-Central-Asia-plus-Canada dollars — roughly a fifth of the bucket, whose true income mix is unidentified — at low-and-middle-income anchors, an upward-bias channel. Several cost priors average over mixed null-and-positive evidence before the credibility discount applies, so the two-multiplier separation of doubts is imperfect at those boundaries. The Dirichlet allocation spread is a subjective prior; uncertainty from the elasticity fit, unmatched filings, the focus-area crosswalk, and the geography audit is not propagated. Allocation shares are derived in nominal dollars and applied to the real total; re-deriving in real dollars moves shares by hundredths of a percentage point. The geography audit’s scalar U.S. shares and regional decompositions were reconciled separately and can differ by fractions of a point per organization. Finally, several anchors rest on a small number of quasi-experimental papers whose external validity to Scott’s grantees is exactly what the realization factor is guessing at.

Conclusion

Measured in health, MacKenzie Scott’s $26 billion buys an estimated median of about 202,000 quality-adjusted life-years — several times its cost under the government’s own regulatory valuation — and most of that estimated health comes from the twentieth of her dollars funding health and development delivered abroad. The number is a model output under stated priors, not a fact, and the model’s most important output is not the point estimate but the machinery around it: explicit evidence weights, a 45-fold stance range that any reader can traverse, versioned corrections, and audits committed as data. Portfolio-scale philanthropy is measurable in outcomes without waiting for donors to publish outcome data — their own gift lists, recipients’ tax filings, and the published causal literature suffice — and the same machinery applies to any large giver whose public disclosures carry comparable gift-level allocation and geography fields. The invitation of the interactive version is the paper’s real conclusion: disagree with a parameter, move it, and watch what survives.