How Much Tax Do U.S. Billionaires Pay? Evidence from Administrative Data

Why it matters, in detail. Population rate = T/Y with T including social contributions and Y excluding transfers; top-400 payroll taxes are about zero, so only the benchmark moves. Gap (top 400 minus population): -6.4 pp (paper); -3.2 to -3.5 pp (social-insurance benefits in Y, CBO-type); -1.4 to -1.9 pp (all benefits in Y); -0.5 to -1.3 pp (plus refundable credits netted); -1.1 pp (OASDI contributions excluded); +1.6 pp (all contributions excluded). With F3's consistent exclusion of pure gains (top 400 26.7%), the CBO-type gap is -0.3 to -0.6 pp and the broader variants turn positive (+1.0 to +2.4 pp).

Evidence. Sec. 2.1 pp. 4-5; Sec. 2.2 p. 8 ('Social Security and Medicare taxes'); Sec. 3.2 p. 9 and fn. 12; Table 2 and notes. Adjudicator recomputation from BEA NIPA, Sept-2026 vintage (scripts/population_rates.py): the paper's definition gives 31.2% literally (30.6% excluding voluntary SMI premiums; the authors' own DINA gives 30.1%, per Claude's sz2026_macro_check), so alternatives are scaled to the paper's 30.2%. For 2018-20 (2018-19 in parentheses): social-insurance benefits (Social Security $1,027bn, Medicare $784bn, UI $194bn; total $2,045bn) added to the denominator 27.0% (27.3%); all government social benefits ($3,401bn) added 25.2% (25.7%); and refundable credits netted 24.3% (25.1%); OASDI contributions excluded 24.9% (24.8%); all social-insurance contributions excluded 22.2% (top 400 about 23.7%). Claude's script gives 24.8/22.1/25.2%; Codex's expanded-income variant (market-price NNP plus benefits, credits netted) 23.8%; Splinter (2025, Table 1, adjustment 3) 25%.

Why this severity. The paper's stated estimand (a macro tax-to-income ratio) reproduces and the 24% level is unaffected, but the 6.4-point gap behind the 'less than the average American' message halves to 3.2-3.5 points under the CBO-type treatment, shrinks to 1-2 points if all benefits are added or Social Security is treated as saving, and reverses only if all payroll taxes are excluded. This is a material, undisclosed sensitivity of one headline comparison to a stated convention, not an error; combined with F3 it removes most of the gap.

The authors' response. Not addressed in the paper, whose framework (Sec. 2.1, following Saez & Zucman 2023) counts all compulsory taxes, including payroll taxes, and keeps transfers out of income by design. No reply to Splinter (2025) was found (web searches, October 2026).

Why it matters, in detail. Consistent exclusion: top 400 26.7% vs 30.2% (-3.5 pp). Consistent inclusion: 23.8% vs 29.4% (-5.6 pp). Symmetric fn-10 treatment: about 23.2% vs about 29% (about -5.7 pp). The asymmetry thus accounts for 0.7-2.9 points of the 6.4-point gap and compounds F1.

Evidence. Sec. 2.2 pp. 7-8; fn. 10 p. 8; App. C; Table A6. Adjudicator reproduction (scripts/reproduce_headline.py): removing pure gains with all taxes kept 26.75% (Claude 26.7%, Splinter 27%); removing all included gains 27.3%; fn-10 variant (also removing an approx. 28.8% tax on pure gains) 23.2%. Including 50% of realized gains (SOI Table 1.4, $973bn a year in 2018-20) in the population denominator gives 29.4% (Claude). Removing capital-gains taxes (approx. 1.2-1.5% of basic-price national income, adjudicator approximation) from the population numerator gives about 29%.

Why this severity. A confirmed internal asymmetry that lowers the top-400 rate relative to the benchmark. Depending on which consistent rule is used, the 6.4-point gap becomes 3.5 points (exclude pure gains for everyone), 5.6 points (include half of realized gains for everyone) or about 5.7 points (remove gains and their taxes on both sides), and the top-400 level rises by 0-2.9 points; noticeable but not transformative on its own.

The authors' response. Fn. 10: excluding pure gains and the taxes paid on them 'leaves effective tax rates virtually unchanged' (adjudicator: 23.2% vs 23.8%). This addresses the top-400 level but not the asymmetry with the population benchmark.

Why it matters, in detail. Benchmark private income = r_e x (1 - s_e) x W_Forbes. With Table A3 incomes the top-400 ETR is 23.4% (-0.4 pp) and the top-100/next-300 gradient is 22.6/24.6% instead of 22.0/26.6%. The validation's own uncertainty is about +/-1-1.5 pp on the 2018-20 level; for 2010-17, where nothing is observed, the paper's +/-20% private-yield band (21.7-26.4% with taxes fixed, 22.3-25.7% with taxes scaled, in 2018-20 terms) is the only guide.

Evidence. Table 1 implied yields: top 100 public 4.568% vs private 4.561%; next 300 5.005% vs 5.008% (adjudicator; Claude 4.564/5.007%). Table 1 private income times units 43.3/38.6/81.7 $bn vs Table A3 39.7/45.9/85.6 $bn. Substituting Table A3 incomes gives 23.4% (top 100 22.6%, next 300 24.6%, vs 22.0%/26.6%) (adjudicator, Claude, Codex). Forbes private wealth 30% (50%) too high under the benchmark method: 25.5-26.0% (26.4-28.0%), depending on whether the property tax imputed on private wealth falls with it. A +/-10% error in the listed dividend yield moves private C-corp income to about $20-50bn (Claude). Restatement check: scripts/restatement_check.py (S-corps $10.35bn vs $10.55bn; partnerships $23.69bn vs $33.32bn). Possible double counting of operating-passthrough dividends: about $1.6bn if passthroughs carry the C-corp stakes at cost, up to about $21bn if their books include the C-corps' full profits. App. A step 2 (fixed end-2019 stakes); Table A1 notes (single August-2019 public share).

Why this severity. The administrative check lands within about half a point of the benchmark for 2018-20, and its plausible range (roughly $64-100bn of private income) maps into about 22.8-24.7%, so the 2018-20 level is loosely validated. But the published benchmark does depend on Forbes valuations, the within-top-400 gradient halves on administrative incomes, and the 2010-17 levels, hence the size of the decline, rest on an untested equal-yield assumption.

The authors' response. The paper calls equal yields 'a natural benchmark to use in light of the available evidence' (p. 7) and reports a +/-20% private-yield sensitivity (Fig. 3(d)); it does not discuss the dependence of the published benchmark on Forbes valuations.

Why it matters, in detail. About 5.6 of the 7.7-point fall in the top-400 ETR between 2014-17 and 2018-20 comes from the corporate component, driven by tau falling from 26.3% to 16.1%. On identical fixed weights the subset's tau falls 9.1 pp with txt, 4.4 pp with txpd, 2.2-3.9 pp with TCJA one-time items removed or re-timed, and 10.9 pp with txc. In ETR terms the 2014-17 to 2018-20 decline is 5.2-6.0 pp with cash taxes and 4.0-4.9 pp with re-timed items (benchmark 7.8 pp); the decline from the 2010-17 average is 3.0-4.1 pp (benchmark 5.7 pp). Whether Berkshire, with its 2017 remeasurement benefit, is in the paper's tau sample is unknown; if it is, the one-time items roughly offset (decline 7.4 pp). On an accrual view that assigns the transition tax to the pre-2018 years in which the foreign profits accrued, the paper's measure is defensible; the problem is the mislabelling and the unreported sensitivity.

Evidence. BSYZ p. 8; App. A p. 23 (txt includes 'all income taxes owed in a given year, current plus deferred'); Table A2 panel C (top 400: 27.8/26.3/16.1%; 2019 18.2%, foreign -0.5%); Table 2 notes; fn. 21; Fig. 2(a), pp. 10-11. Boll, Saez & Zucman (2026, w35218, revised Aug 2026), appendix item on global book taxes (transition-tax re-timing; 'Using txpd generates an average income tax rate of 17.19% which is almost identical to the 16.96% tax rate we obtain in our computation using txt', 2019-25 sample). Adjudicator approximate replication (scripts/xbrl_tax_panel.py, xbrl_weighted_tau.py, xbrl_tcja_adjustment.py; SEC XBRL company facts for 14 large top-400-owned listed firms covering about 60% of top-400 public profits; approximate 2019 Forbes-400 stakes held fixed as in App. A; transition-tax amounts from XBRL, 10-K text and BSZ 2026): ownership-weighted rate excluding Berkshire 26.8/27.3/18.3% with provisions (tracking the paper's 27.8/26.3/16.1%), 23.8/24.7/20.3% with cash taxes paid, 26.3/31.3/20.5% with current tax expense, and 21.0-22.9% in 2014-17 (18.8-19.0% in 2018-20) when one-time TCJA items are removed or re-timed. Mapped into the paper's framework, the top-400 ETR path becomes 25.7/30.0/24.9% with cash taxes and 27.5/28.0-29.1/24.0-24.2% with TCJA items removed or re-timed, against 27.5/31.5/23.75%. Codex's firm examples (Berkshire, Microsoft) show provisions and cash payments differing in both directions.

Why this severity. The 2018-20 level moves by only +0.3 to +1.5 points and the decline survives every concept tested, but its size, a headline result, is roughly a quarter to a half smaller with cash taxes paid (the concept the paper's own table names) or with the authors' later transition-tax treatment; with current tax expense it is as large or larger. The replication is approximate (subset of firms, approximate stakes, unknown treatment of Berkshire), so this is a likely, material concept-dependence of C2's magnitude rather than an established error.

The authors' response. Not in this paper. In Boll, Saez & Zucman (2026, NBER w35218) the authors adjust Compustat book taxes for the one-time transition tax and report that cash taxes paid give almost the same rate as book taxes for 2019-25 (17.19% vs 16.96%).

Why it matters, in detail. HS rate = T/(dW + T). Selecting on 2019 wealth raises dW (winners) and lowers the rate; selecting on 2010 wealth lowers dW (mean reversion, deaths) and raises it. The paper reports only the lower one; a symmetric presentation would bracket the top-400 rate at roughly 20-25% against 20-21.5% for the population.

Evidence. Sec. 3.2 p. 10; Table 2 panel B and notes; IRS-SOI Table 3. Adjudicator (scripts/reproduce_headline.py): panel B wealth change $4,680.7m, required total taxes $1,259.3m, implied corporate plus property taxes $812.7m; panel A 24.6% vs panel B 21.2% with identical corporate and property taxes, and 23.4% vs 20.1% with the contemporaneous top-400 path (about $733m). Claude: 21.9-23.4% vs 18.4-20.2%. Codex: 21.2% not reproducible. Population Haig-Simons rate recomputed at 20.3-21.1% (Claude, NIPA and Z.1).

Why this severity. Affects a secondary robustness result: the cohort choice moves the top-400 Haig-Simons rate by about 3.4 points and turns 'similar to the population' into 'above it', and the level cannot be verified. Neither cohort is unbiased (panel A has mean reversion and decedents), so this qualifies, but does not reverse, a core finding.

The authors' response. The IRS-SOI companion (p. 8) describes the two panels as 'similar ... consistent with a relatively modest turnover at the very top'.

Why it matters, in detail. Private property taxes = 0.0019 x $4,676m = $8.9m per unit. Each 0.1 pp on the private rate adds about $4.7m of taxes (and, by construction, of income), roughly +0.8 pp on the ETR. The parameter is not covered by Fig. 3(c)-(d).

Evidence. Sec. 2.2 p. 8; App. A p. 22; Table 2 ('Other taxes' 3.4 points); Saez & Zucman (2026) appendix (property taxes 'roughly equal to 1% of Amazon's capital stock, the US-wide average business property tax rate'); Boll, Saez & Zucman (2026) appendix (land and buildings only for Alphabet, Meta, Nvidia, Oracle). Table A4: private wealth of the top 120 in 2016 is 45.5% private C-corps, 17% finance, 5.5% real estate, 31.9% other passthroughs. DINA business property taxes are about 0.75% of business wealth (Claude). Scenarios (adjudicator, Claude): 0.40% of private value gives 25.4% (26.0% without adding the tax to income); 0.73% gives 27.8% (29.4%).

Why this severity. An untested imputation worth 3.4 points; plausible private rates of 0.3-0.5% of value would add about 1-2.5 points to the 2018-20 rate, narrowing the population gap, but the direction for listed firms may be the opposite. A robustness concern with limited demonstrated impact.

Why it matters, in detail. Over-attribution x scales wealth-based income (public and private profits) and corporate/property taxes by (1 - x) while individual taxes stay fixed; with individual taxes about 46% of the total, the ETR rises by roughly 0.08 pp per 1% of over-attribution.

Evidence. IRS-SOI companion p. 3 (matching on the listed individual). Forbes March lists 2018-20: '& family' entries hold 6.9-7.4% of top-367 US wealth (Claude). Raub, Johnson & Newcomb (2010, p. 131): Forbes treats trust assets as the income beneficiary's and 'joint wealth, including wealth owned by extended family members in some cases, as if it belongs to the listed individual'. Smith, Zidar & Zwick (2023, p. 552): 2,370 adults 'possibly represented'. Scenarios (adjudicator): 10% over-attribution gives 24.6%, 25% gives 26.2%. Internal checks against a large effect: matched 1040 dividends ($15.3bn) exceed predicted dividends on Forbes listed stakes ($11.0bn), and 2019 administrative private income is close to the Forbes-based figure.

Why this severity. An undisclosed, one-directional mismatch; a likely 5-15% over-attribution would raise the 2018-20 rate by about 0.4-1.2 points (2.4 points at 25%), a modest but real narrowing of the population gap. Splinter's 39% haircut (+8 points) is not supported.

Why it matters, in detail. Next-300 public yield 5.0% vs observed 4.4% raises next-300 economic income by about 8%; ETR +0.6 pp (top 400), +1.9 pp (next 300); the decline is overstated by about 0.8 pp.

Evidence. App. A pp. 22-23; Table A2 panel A; Table A1 line 8. Adjudicator (scripts/reproduce_headline.py): with each group's own observed yield the top-400 ETR is 24.3% in 2018-20 (next 300 28.6%) and 31.2% in 2014-17, so the 2014-17 to 2018-20 decline is 6.9 instead of 7.8 points (Claude: 24.3%, 31.2%).

Why this severity. A disclosed choice with a stated rationale; +0.6 point on the 2018-20 level and about 0.8 point on the decline.

The authors' response. App. A: the rule is meant to neutralise the influence of the five largest firms on subgroup yields observed in 2010-13.

Why it matters, in detail. Consistent +/-20% band 22.3-25.7% vs stated 21.7-26.4%; macro-yield scenario 16.7-16.8%.

Evidence. Sec. 3.4 p. 12; Intro p. 2; Fig. 3(c)-(d). Adjudicator: +/-20% private income gives 21.67%/26.40% with taxes fixed (matching the figure) and 22.28%/25.66% with corporate taxes scaled; macro yield 16.7% (Claude 16.8%; Codex the same +/-20% figures).

Why this severity. Presentation: the stated band is wider than the consistent one, and the omitted scenario lies on the side favourable to the paper's thesis.

Why it matters, in detail. Average and marginal rates converge at very high wages, so 45% approximates a pure top earner's all-in rate; the comparison of 23.8% with about 40-45% is robust.

Evidence. Abstract; Intro p. 2; fn. 11 p. 9. 37 + 3.8 + 5 = 45.8%, about 45.1% with employer Medicare in the denominator (Codex); the federal average rate on $10m of wages is about 36.4% under 2019 brackets (adjudicator).

Why this severity. A reasonable illustration for very high wage earners; any realistic measured benchmark (roughly 38-48% depending on the state) leaves the comparison with billionaires intact. A labelling issue.

The authors' response. Fn. 11 discloses the construction.

Why it matters, in detail. The individual-tax component of the 2014-17 to 2018-20 decline (about 2 of 7.7 points) cannot be attributed to TCJA expensing rather than to 2020 or the CARES Act from three-year averages; the corporate component is tied to the statutory cut (but see F2).

Evidence. Sec. 3.4 p. 11; Sec. 3.3 pp. 10-11; P.L. 116-136 sec. 2304; IRS Publication 536 (2019) p. 1 (Claude). The IRS tables give only three-year averages.

Why this severity. Concerns the explanation and generalisability of C2, not the measured decline; the effects cannot be separated with the published averages.

The authors' response. Intro p. 2 describes 2018-20 as 'a period that includes various economic and policy changes including the Tax Cuts and Jobs Act'.

Why it matters, in detail. The estate component is constant at $2.64m per unit by construction; its ratio to wealth mechanically falls 0.057% -> 0.040% -> 0.033%.

Evidence. Intro p. 3; Sec. 3.2 p. 10; IRS-SOI p. 4 ('transfer taxes are averaged across all years 2010 to 2020') and Table 2 (identical 2,354/3,368/3,477 rows in all periods). $2.642m per unit divided by wealth of 4,620, 6,573 and 8,015 $m gives 0.057%, 0.040% and 0.033%.

Why this severity. Wording; no effect on the levels.

Why it matters, in detail. The low gross-estate/Forbes ratio (27%; 39% for singles) is consistent with Forbes overstatement and family/trust attribution as well as with legal undervaluation; the 0.8%/7% estate-tax rates are relative to Forbes wealth and inherit that ambiguity.

Evidence. BSYZ Sec. 3.2 p. 10; Raub et al. (2010) pp. 130-134 (local copy /mnt/zdata/downloads/raub-johnson-newcomb2010/raub2010.txt): trusts and family wealth (p. 131); discounts raise the Forbes ratio from .47 to .54; 'contrary to the Forbes' assertion that its methodology is conservative, these results suggest that it is actually generous in some cases' (p. 134).

Why this severity. Interpretive; it bears on how much weight Forbes-based denominators deserve (F4, F7) but changes no number.

Why it matters, in detail. Only the US/foreign split of corporate taxes, the reading of two figures and the reader's ability to reproduce are affected.

Evidence. Locations as listed. Adjudicator check of the 42% rule ($4.47bn vs $2.08bn of foreign taxes) and of the labels in the extracted text and the IRS table images; Codex foreign_tax_note (2.698% vs 1.4%) and label_checks; Claude I13.

Why this severity. Typos and loose wording; totals are unaffected.

Why it matters, in detail. 6.91% vs the stated 6.0%; the remaining gap to the top-400 passthrough yield is about 1.6 pp, not about 0.7 pp.

Evidence. App. B.3 p. 29; Table A5 (8.3% S-corporation return). Codex scorp_illiquidity; adjudicator recomputation.

Why this severity. Affects one supporting comparison only.

Why it matters, in detail. A like-year comparison (US 2014-17 about 31% vs France 2016 26%) supports the statement, but the figure's US bar is post-TCJA; converting the Dutch average as well would widen, not narrow, the Dutch gap.

Evidence. Intro p. 3; Conclusion pp. 12-13; Fig. 3(f) and notes; Bach et al. (2023) 26%; Bruil et al. (2025) 'somewhat below 20%' and 45% (sources verified by Claude and Codex).

Why this severity. Presentation; the within-country patterns are faithfully reported.

The authors' response. Intro p. 3 compares France in 2016 with the US 'pre-Tax Cuts and Jobs Acts' (30%).

Why it matters, in detail. Proprietary Forbes stakes and Compustat items cannot be independently audited; the adjudicator's public-data approximation reproduces the provision-based tau path within about 1-2 pp.

Evidence. Paper App. A-C; IRS-SOI Table 3 (blank corporate-tax rows); Codex reproduction log (Table A3 rebuilt to $86.1bn vs $85.6bn; Haig-Simons not reproducible); author and NBER pages searched by both auditors (no package found).

Why this severity. Confidential microdata cannot be released and the published tables are unusually detailed; the gap mainly prevents verifying C7 and the corporate-tax inputs, which F2 checks only approximately.

Why it matters, in detail. Wealth-ranked top 400 (BSYZ) 23.8% vs income-ranked top 400 about 29% in the authors' current DINA series, against a population rate of about 28.8% on that basis; the similarity with the 2019 vintage is coincidental across groups and vintages.

Evidence. BSYZ p. 3; Saez & Zucman (2019, BPEA) pp. 440 and 466-467; PSZ2020 and PSZ2022 AppendixTablesII(Distrib).xlsx and SaezZucman2026Data.xlsx, sheet 'taxrates', taxtop400 (scripts/dina_top400_vintages.py: 2018 27.23/29.31/29.44%; taxall 28.34/28.74/28.84%). Also recorded in this project's adjudication of Saez & Zucman (2019, Triumph of Injustice), F5.

Why this severity. Concerns a corroboration remark, not the paper's estimates; it does show that the 'below average' result depends on ranking by wealth rather than by income.

Why it matters, in detail. Each 10 pp of the listed-sales gains reclassified as past retained earnings removes about $2.1bn ($5.7m per unit) from income, about +0.3 pp on the ETR.

Evidence. Sec. 2.2 p. 8; App. C pp. 29-31 and fn. 36; Table A6 (49/64/85% reproduce as 48.8/64.3/85.0%). Codex: logical counterexample (all sales from a firm with a 73% retained-earnings share would give 19.9%); Nike's 2019 10-K equity statement (a $4.1bn repurchase debit to retained earnings). Adjudicator: the largest sellers' firms (Amazon, Meta, Microsoft) have low accounting retained-earnings shares; a plausible 20-30% share would lower the includible share to about 42-46% and raise the ETR by about 0.2-0.5 point.

Why this severity. An overstated label; the plausible effect on the level is below one point (about +2 points only in an implausible extreme).

The authors' response. App. C: the low-end scenario 'is by construction highly conservative'.

Why it matters, in detail. Numerator omissions (sales and residential property taxes) of about 0.4-1.5 pp against a denominator omission (net imputed rent) of about 0.2-0.5 pp; net +0.2 to +1 pp.

Evidence. Sec. 2.2 p. 8; Table 2 'Other taxes'; Saez & Zucman (2026) data assign sales taxes of 2.2% and residential property taxes of 0.17% of income to the income-ranked top 400 (Claude). Adjudicator: residential property tax on $100-500m of homes (about 1% of value) and sales tax on consumption are of the order of 0.4-1.5 points of economic income, offset by about 0.2-0.5 point of omitted net imputed rent.

Why this severity. Small and partly offsetting; plausible net effect of about +0.2 to +1 point on the level.

The authors' response. Sec. 2.2: sales taxes are disregarded because they are 'negligible at the top, where saving rates tend towards 100%'.

Why it matters, in detail. Up to one quarterly state payment shifted out of 2018 (about 8% of the 2018-20 three-year total) implies about +0.2 pp on the ETR and a slightly smaller decline.

Evidence. IRS-SOI p. 4 (Schedule A amounts 'reported without cap'); IRS Table 2: state and local taxes 13.889 -> 10.807 $m per unit; AGI 185.374 -> 172.669 (-6.9% nominal); AGI less gains and dividends 34.38 -> 16.04 (-53%) (adjudicator, Claude, Codex). Splinter (2025) p. 5, which also labels per-unit thousands as billions (Codex).

Why this severity. A timing effect worth about +0.2 point at most on the 2018-20 level; the larger correction proposed by Splinter is not supported.

The authors' response. IRS-SOI companion p. 4: state taxes are taken from Schedule A and are 'reported without cap even after TCJA capped such deductions'.