Emmanuel Saez, Gabriel Zucman · Letter, January 18, 2019; updated letter on the Ultra-Millionaire Tax Act of 2021, February 24, 2021

Scoring of Senator Warren's Wealth Tax Proposal (letters to Senator Elizabeth Warren)

16 problems found: 2 major · 4 moderate · 10 minor

What the paper claims

About 75,000 households (less than 0.1%) would be liable for the 2% (> $50m) + 1% (> $1bn) wealth tax in 2019.
The tax would raise about $212bn in 2019 (about 1.0% of GDP) and about $2.75tn over 2019-2028, of which about $0.3tn from the billionaire surtax.
The assumptions are conservative: a 15% avoidance/evasion haircut is 'a large response in light of existing estimates', and growth of the taxable base at the pace of the economy is 'conservative'.
Ultra-Millionaire Tax Act of 2021 (2% + 1%): about 100,000 families liable; $252bn in 2023 (about 1.0% of GDP); $3.0tn over 2023-2032, of which $0.4tn from the surtax; DINA gives 'fairly close (and even slightly higher)' revenue.
With the billionaire surtax raised to 4% (6% top marginal rate), revenue would be $351bn in 2023 and $3.9tn over 2023-2032.
Context: wealth concentration has risen sharply (top 0.1% ~7% to ~20%, bottom 90% ~35% to ~25%; top 0.1% owns almost as much as the bottom 90%; the bottom-90% fall is primarily due to debt); billionaire wealth is $4.2tn, 40% higher than before COVID; top 0.00025% wealth rose from 2.6% to 21.6% of national income.
Context: relative to wealth, the top 0.1% pay 3.2% in taxes (4.3% with the wealth tax) while the bottom 99% pay 7.2%.

Every problem found

1majorconfirmed · previously known · internal inconsistency

The authors' own method says people hide more wealth when the tax rate is higher, but they applied the same 15% to billionaires facing 3% or 6% rates. Applied consistently, most of the extra $0.9 trillion promised from the higher billionaire rate disappears; the effect on the main estimates is only a few percent.

Technical detail and evidence

Why it matters, in detail. h=e*tau with e=8 gives 24% at tau=3% and 48% at tau=6% (log-linear 21%/38%). Holding h=15% overstates revenue by 3-6% for the 2%/3% schedules and by 17-30% for the 2%/6% schedule: C5's $3.9tn becomes ~$2.7-3.2tn and the surtax increment over panel A falls from $0.88tn to ~$0.07-0.31tn. PWBM made the same point for the campaign's 2%/6% plan ($2.7tn vs $3.75tn).

Evidence. 2019 letter pp.2-3 pt 4; 2021 letter p.3 b)-e) and table p.4 panel B; PWBM (Dec 12 2019) App. D citing Max Ghenis. scripts/haircut_rules.py (authors' Sept-2019 file): rate-consistent rule -5.9% (linear) / -3.2% (log-linear) vs uniform 15%. scripts/rebuild_2021.py (SCF 2019 + Forbes real-time 24 Jan 2021, aged as described): panel A -6.3% (linear) / -3.4% (log-linear) vs a 15%-of-wealth baseline; panel B 2023 -27%/-18.5%, 10-yr -25.7%/-17.4% (vs liability baseline -30.0%/-22.2%); billionaires supply 39% (A) and 56% (B) of 2023 revenue. Implied C5: $3,906bn -> about $2.7-3.2tn; 4%-surtax increment $0.88tn -> about $0.07-0.31tn.

Why this severity. For C5, applying the authors' own rule consistently lowers the ten-year figure by roughly 17-30% and eliminates most of the advertised gain from the higher billionaire rate, a major change to a headline claim of the 2021 letter; for C2 and C4 the effect is small (3-6%).

The authors' response. No direct response located. Panel B does lower billionaire base growth by 3 points a year (2021 p.3 d), but that is a mechanical (tax-paid) adjustment, not an avoidance response.

2majorlikely · previously known · speculative assumption

The estimates assume the very rich would hide or shelter only 15% of their wealth. The authors call this cautious, but it is near the low end of what studies (including one co-written by Zucman) suggest for the very wealthiest; with other reasonable assumptions the revenue would be roughly a fifth to a third lower.

Technical detail and evidence

Why it matters, in detail. Revenue scales with the post-haircut base above the exemption, ~1.3-1.4x as elastic as wealth. h=15% (semi-elasticity ~7.5-8 at a 2% rate) sits at the bottom of the available evidence (PWBM -13; Advani-Tarrant 7-17; JJKZ ceiling DiD ~20+ per pp over 8 years; Swiss >=43). On the authors' own micro-file, e=13 rate-scaled lowers 2019 revenue 24% ($203.5bn -> $154.8bn), e=17 38%, and the letters' own averaging with the updated Swiss estimate 13-19%; near-zero responses would raise it up to 26%. The uncertainty is two-sided but skewed toward larger avoidance, and no range is reported.

Evidence. 2019 letter p.2 pt 2 and fn 2; 2021 letter p.2 pt 2, fn 3 and table note. JJKZ NBER w24371 (Mar 2018) p.2 ('about 30% after 8 years for the very wealthy'), p.3 ('particularly effective avoidance and evasion vehicles'), p.14 ('applies to the sample of moderately wealthy people around the kink, not necessarily to the super wealthy'). Bruelhart-Gruber-Krapf-Schmidheiny, CESifo WP 7908 (Oct 28 2019), abstract: 'at least 43% after 6 years'. Advani & Tarrant (Fiscal Studies 2021, s.3): SZ's 7.5 'somewhat arbitrary', well-designed tax 7-17. PWBM (Dec 2019): weighted -13. Kopczuk (BPEA 2019 comment, pp.4-5): bunching elasticities 'a severe understatement'; 26% of the base between threshold and 2x threshold. scripts/scf_portfolio.py: business equity 51.6% (SCF 2016) and 53.9% (SCF 2019) of assets of NW>=$50m households; market-priced financial assets 27-30%. scripts/haircut_rules.py on the authors' Sept-2019 file (15% = $203.5bn): e=13 uniform -17.6%, e=13 rate-scaled -23.9%, e=17 rate-scaled -37.6%; letters' own averaging with the Oct-2019 Swiss estimate (e=11.6; 23% haircut) -13.3% uniform / -19.0% rate-scaled; with Danish DiD (e=16.7) -29%/-37%; no avoidance +25.8%. scripts/rebuild_2021.py: 2021 panel A with e=13: -24% (log-linear) to -30% (linear). Balkir, Saez, Yagan & Zucman (NBER w34170, 2025): Forbes decedents' gross estates 39% of Forbes wealth, attributed to estate-tax undervaluation.

Why this severity. The headline depends directly on this parameter; moving it within the range supported by evidence available when the letters were written (semi-elasticity 11.6-17 instead of ~8) lowers 2019 and 2021-A revenue by roughly 13-38%, up to a third of the headline. The 'conservative' label is not supported, although lower responses (which would raise revenue) cannot be excluded.

The authors' response. Avoidance is a policy choice; the Warren design has no exemptions, third-party reporting, a 30% audit floor, exit tax and IRS funding (SZ, 'How would a progressive wealth tax work?', Feb 2019; BPEA 2019). Response to Summers-Sarin (June 25 2019): 15% is 'a middle-ground'; 'it is possible that avoidance would be below 15%'; ~80% of top-0.1% wealth has market prices. Bunching captures avoidance, DiD effects partly real saving. PWBM's 2021 enforcement-adjusted semi-elasticity (-8.4 to -9.4) is close to the letters' level, though applied at marginal rates.

3moderateconfirmed · previously known · version instability

In 2019 the authors averaged two data sources; one of them was later revised by the authors themselves and now shows about a quarter less wealth above $50 million. With the revised figures the 2019 estimate would be about 13-16% lower, roughly $2.3 trillion instead of $2.75 trillion.

Technical detail and evidence

Why it matters, in detail. Replacing the DINA 2019 projection (top-0.1% share 22.2%) with the authors' Feb-2022 vintage (18.5%) lowers the averaged base above $50m from $9.34tn to $7.81tn and 2019 revenue from $212bn to $178bn (-16% including F7/F8; -13% from the revision alone); taxpayers fall from ~75,000 to ~61,000. For 2021, DINA is within ~1-5% of SCF+Forbes revenue, not above it.

Evidence. 2019 letter pp.1-2, p.2 pt 3; 2021 letter pp.1-2 and fn 1. scripts/dina_vintages.py (15% haircut, tax units): spreadsheet 78,434/$10,502bn; OPA anonymised 77,151/$10,413bn; Jul-2019 vintage 2019 projection 75,901/$10,207bn (top-0.1% 22.2%); Feb-2022 vintage 2019 55,881/$7,718bn (18.5%); Sept-2020 vintage 2018 $7,743bn vs earlier 2018 file $9,463bn (-18%). scripts/reconcile_2019.py: revision alone -$28bn (-13%); cumulative with F7/F8 $178.3bn (-16.1%), x13 = $2.32tn. scripts/dina_vs_scf_2019.py: sub-billionaire base (billionaires replaced by Forbes in both) SCF 2019 $5,230bn vs Sept-2020 vintage scaled to the SCF aggregate $5,052bn (-3%), Feb-2022 vintage $4,431-4,756bn (-9% to -15%), superseded Jul-2019 projection $6,364bn (+22%).

Why this severity. On the authors' own current data the 2019 headline falls by about 13-16%, a noticeable but not transformative change; the 2021 headline rests on SCF+Forbes, and the DINA statement there is secondary.

The authors' response. The authors themselves adopted the revisions (heterogeneous returns on fixed income; SZ 2020, 2022) while disputing Smith-Zidar-Zwick's lower top tail. No response on the 2021 DINA statement.

4moderateconfirmed · newly identified here · overclaiming

The estimates come as precise numbers, but the survey behind them includes only a handful of super-rich families; one family alone accounts for about an eighth of the survey's taxable wealth. Statistical uncertainty alone is roughly plus or minus 10% of the total, and the letters say nothing about it.

Technical detail and evidence

Why it matters, in detail. Var = replicate variance + 1.2 x between-implicate variance gives CV 16.6% on the SCF base; the extreme tail is carried by records with weights of 500-1,600, which also drives the irreproducible 2021 billionaire count. Translated to total revenue, ~+/-10% (1 s.e.) from sampling alone.

Evidence. scripts/scf_se.py (999 replicate weights x multiplicity factors, 5 implicates): SCF 2019 base above $50m (unaged, no haircut, no Forbes) $7,208bn, SE $1,199bn (CV 16.6%; 95% CI $4.9-9.6tn). scripts/scf_concentration.py: household yy1=2365 (NW $669m, weight ~1,585 across implicates) = 13.6%; top 10 households 35.9%. 2021 table p.4 (10,971; 100,449; 3,284; 1,005). 2019 letter p.1: 'None of these sources provides perfect estimates' is the only caveat.

Why this severity. An important undisclosed caveat: with the SCF supplying ~60% of 2021 revenue, sampling error alone is about +/-10% (1 s.e.) on the total, before source (SCF vs DINA 28%) and vintage (-24%) uncertainty.

The authors' response. 2019 letter: 'None of these sources provides perfect estimates'; June 2019 response: the no-avoidance figure 'involves some uncertainty'.

5moderateplausible · previously known · external inconsistency

Records from the existing estate tax show the very rich reporting far less wealth at death than surveys and Forbes lists suggest, and the letters do not explain this gap. If even part of it reflects how the rich value assets for tax purposes, a wealth tax would collect perhaps a tenth to a fifth less than estimated, though far more than the tiny amounts some critics claimed.

Technical detail and evidence

Why it matters, in detail. Estate returns imply reported values of ~26-57% of the wealth that SCF/DINA attribute to the $50m+ group (at 1-2% mortality); Forbes-to-estate matches give 39-50%. Either channel (overstated measured wealth or valuation discounting) implies a base below 85% of measured wealth; valuing Forbes-400 records at 70-50% lowers revenue 11-18%. The Summers-Sarin $25-75bn/yr is not a credible score (it imports estate-specific deductions, trusts and weak enforcement).

Evidence. IRS SOI estate tax returns filed in FY2017, Table 1 (17es01fy.xlsx): 423 returns >= $50m, gross estate $68.66bn, net estate tax $9.32bn. Units >= $50m held ~$12-13tn at 2016 levels (SCF 2016 $10.8tn + Forbes 400; DINA 2016 $12.1tn), implying ~$120-260bn of gross estates a year at 1-2% wealth-weighted mortality. Raub, Johnson & Newcomb (2010) as cited by SZ (June 2019 fn 4): Forbes decedents' estate wealth about half of Forbes; Balkir, Saez, Yagan & Zucman (NBER w34170, 2025, p.3 and Sec. on effective estate tax rates): gross estate 39% of Forbes wealth. scripts/haircut_rules.py: Forbes records at 70%/50% of Forbes before the 15% haircut lower 2019 revenue 10.9%/18.1%. Summers & Sarin (Washington Post, Apr 4 2019): $25bn (statutory) or ~$75bn (effective rate) per year.

Why this severity. A plausible but undemonstrated downward risk to the base of roughly 10-20% (partial-Forbes scenarios); the estate-tax analogy itself is not like-for-like, so it qualifies rather than overturns the headline.

The authors' response. SZ, 'Response to Summers and Sarin' (June 25 2019): the $25bn implies 80-90% avoidance; estate data put the top 0.1% below 10% of wealth; 'about 50% of estate taxes owed are avoided'; estate audits fell from 65% (1975) to 8.6% (2018); the Warren design (no exemptions, third-party reporting, exit tax) differs from the estate tax. Balkir et al. (2025, with Saez and Zucman) attribute the Forbes-estate gap to estate-tax undervaluation techniques.

6moderateplausible · previously known · speculative assumption

The letters assume the taxed wealth keeps growing as fast as the economy and call this cautious because the rich got richer faster in the past. But paying 2-3% a year itself slows the growth of big fortunes; accounting for that alone lowers the ten-year total by up to about a tenth.

Technical detail and evidence

Why it matters, in detail. With fixed nominal thresholds the base grows ~a*g; decumulation of ~1.6% of wealth a year subtracts ~2 points. The 2019 multiplier of 13 is matched by CBO-type 4% growth without the drag (12.84) but falls to 11.66 with it; the 2021 letter's 12.0 is roughly neutral (11.7 with the drag). Offsets to income and estate taxes lower net revenue further (PWBM; Tax Foundation -$142bn).

Evidence. 2019 letter p.3 pt 6; 2021 letter p.3 d). scripts/growth_paths.py (authors' Sept-2019 file): tax = 1.61% of reported wealth; 4% wealth growth, fixed thresholds: revenue growth 5.47%/yr, multiplier 12.84; plus mechanical decumulation: 3.44%/yr, 11.66 (-9% vs 12.875); 5.5% growth plus decumulation: 12.93. PWBM (Mar 15 2021): $2.71tn for 2023-2032 including decumulation, income/estate offsets and enforcement revenue (5% wealth growth) vs $3.03tn; Tax Foundation (2020): $2.64tn conventional, $2.3tn dynamic, $2.2tn net of income/payroll offsets. Bill text: no inflation adjustment of the $50m/$1bn thresholds.

Why this severity. A robustness concern of limited demonstrated size: about 0 to -10% on the 2019 ten-year total and about -3% on 2021 panel A; the 'conservative' label is asserted rather than shown.

The authors' response. The letters argue that top wealth historically grew 2.8-4.5 points a year faster than average, more than the tax rate, so top wealth would still grow at least as fast as the average; the 2021 6% variant reduces billionaire growth by 3 points.

7minorconfirmed · newly identified here · arithmetic or transcription

A formula used to estimate the wealth of billionaires just below the Forbes 400 was applied to the wrong quantity, overstating the billionaire part of the tax by about 14%. The effect on the overall total is small (under 2%).

Technical detail and evidence

Why it matters, in detail. Surtax base overstated by ~$0.35tn (+16%), surtax revenue by $3.0-3.6bn/yr, headline by 1.4-1.7%; the Pareto count (911) also exceeds Forbes's own count (607).

Evidence. 2019 letter p.3 pt 4 and fn 3; spreadsheet rawdata!C17 = 0.01*(400*((1-0.15)*7.2-1)*1.25) = 25.6. scripts/letter_arithmetic.py (analytic and numerical integration): missing slab $153-157bn; surtax base $2,201-2,205bn vs $2,550-2,560bn; surtax revenue $22.0bn vs $25bn (25.6); total -1.4% to -1.7%; 10-yr surtax ~$286bn. scripts/forbes_checks.py: Forbes Mar-2019 list 607 US billionaires ($3,111bn); 167 ranked 401+ with 85% wealth >= $1bn add $53bn (2.6%).

Why this severity. A genuine error, but it changes the 2019 total by under 2% and the rounded surtax claim (~$0.3tn) survives.

The authors' response. Implicitly corrected in the authors' Sept-2019 data-based re-score (surtax base ~$2.20tn).

8minorconfirmed · newly identified here · internal inconsistency

When updating the survey to 2019, the authors in effect counted the wealth of the 400 richest Americans twice in the national total. Fixing it lowers the estimate by about 1%.

Technical detail and evidence

Why it matters, in detail. SCF records are scaled by 94/(94-2.89) ~ 1.032 too much, raising the SCF base 3.6%, the SCF count 8% and averaged 2019 revenue 1.3%; the corrected projection ($7,907bn) matches the actual 2019 SCF + Forbes 400 base ($7,902bn).

Evidence. 2019 letter p.2 pts 1 and 3; spreadsheet rawdata!A4, B12 (=400+71743), B16 (=271+400). scripts/scf2016_leg.py: SCF->$94tn + F400 reproduces 72,143 (71,743 SCF) and 671 billionaires exactly, base $8,189bn (sheet $8,179bn); SCF->$94tn-F400 gives 66,565 and $7,907bn (-3.6%); effect on averaged revenue -$2.83bn (-1.33%). Authors' Feb-2019 wealthtax.do: 'Rescaling scf to match total total wealth reported in DINA excluding the f400'; 'benchmarking on DINA aggregate household wealth of 2019 of $94.0Tr'. Ex post: actual SCF 2019 + Forbes 400 base $7,902bn.

Why this severity. About 1.3% on the 2019 headline and 8% on the SCF taxpayer count.

The authors' response. Corrected in the authors' own Feb-2019 Stata program (SCF scaled to $94tn minus the Forbes 400); the January letter was not revised.

9minorconfirmed · newly identified here · misrepresentation of source

The 2019 letter says its 5.5% yearly growth figure comes from official budget projections, which actually projected about 4%. Because the tax thresholds are not adjusted for inflation, the mistake happens to make little difference to the total, but the source is misstated.

Technical detail and evidence

Why it matters, in detail. 5.5% is not a CBO/JCT projection; with CBO-type 4% wealth growth and fixed thresholds the multiplier is 12.84 vs 13, so the misattribution matters little absent decumulation (F5).

Evidence. 2019 letter p.3 pt 6 and fn 4. scripts/cbo_check.py: CBO Aug-2018 2018-28 nominal GDP 3.99%/yr, GDP price index 2.13%, real 1.82%, adult population 0.78%; Jan-2019 3.93%, 2.05%, 1.84%, 0.77%; Feb-2021 2021-31 nominal 4.09% (consistent with the 2021 letter's 4%). scripts/letter_arithmetic.py: multiplier 12.006 at 4% vs 13 (-7.6%); scripts/growth_paths.py: 12.84 with unindexed thresholds.

Why this severity. The source is misstated, but the net effect on the ten-year total is small once unindexed thresholds are considered.

10minorconfirmed · newly identified here · documentation gap

Unlike the 2019 estimate, the 2021 update cannot be rebuilt from what was published, and it quietly changed how the 15% assumption is applied in a way that could raise revenue by about 7%. The authors' own later recalculation of the 2019 plan (about 4-6% lower) was never added to the 2019 letter.

Technical detail and evidence

Why it matters, in detail. The four 2021 statistics cannot be jointly reproduced; the liability-vs-wealth haircut is worth +6.9% on identical data; cross-version changes in data (SCF/DINA/Forbes 400/Pareto -> SCF/Forbes real-time), growth (5.5% -> 4%) and haircut definition are not reconciled.

Evidence. 2021 letter p.1 (two reasons for the increase), p.2 pts 1-2, p.3 a)-b), table p.4; 2019 letter p.2 pt 2. scripts/haircut_rules.py: 15% of liability vs of wealth: OPA file $216.5bn vs $202.5bn (+6.9%); Sept-2019 file $217.5bn vs $203.5bn. scripts/rebuild_2021.py: base matched with end-2020 factor ~1.05 (liability reading) or ~1.17 (wealth reading); best joint max error 14% / 37%; 663 US billionaires (Forbes real-time, 24 Jan 2021). Sanders letter (Sept 22 2019) addendum and fn 4: Warren plan $200.1bn at 16%, 70,187 taxpayers, 982 billionaires; 'differs very slightly from the original scoring ... $210 billion'. 212 x 1.055^4 = $263bn.

Why this severity. The 2021 base is approximately reproducible and within ~10% of an independent conventional estimate (PWBM); the undocumented haircut convention could matter by up to ~7%.

The authors' response. The Sanders letter (Sept 2019, fn 4) disclosed that the re-score 'differs very slightly' from the January scoring.

11minorconfirmed · newly identified here · internal inconsistency

The 2021 letter has several slips (a wrong multiplier in the text, 'billion' for 'trillion', a share of 'families' computed with the wrong population count, ten-year totals labelled as 2022 dollars though they include inflation, and a footnote that does not match the text). None changes the headline numbers.

Technical detail and evidence

Why it matters, in detail. None changes the nominal headline; the 'share of families' is understated by ~30% (0.054% vs 0.077% of households) and the real-dollar label misstates value by ~11%.

Evidence. 2021 letter p.3 a), d), e); table p.4; p.7 and fn 6; Warren bill summary (Mar 2021). scripts/letter_arithmetic.py: 10.4622; 131.36 x 10.462 = 1,374; growth implied by 10.17 = 0.374%; 100,449/185m = 0.0543%; real value $2,700bn; 1.27/10 = 0.127 = 0.60% of $21.3m. SCF 2019 household weights sum to 128.6m (scripts/rebuild_2021.py).

Why this severity. Presentational; the table totals ($3,027bn; $3,906bn) are internally consistent.

12minorconfirmed · newly identified here · version instability

The letter assumes the tax starts one year earlier than the bill introduced five days later says. Depending on how the ten-year window is counted, the total could be slightly higher or about 8% lower; the yearly amounts are unaffected.

Technical detail and evidence

Why it matters, in detail. A one-year valuation-date mismatch; the ten-year total moves by -8% (fixed 2023-2032 window) to +4% (first ten payments).

Evidence. 2021 letter p.1 and table notes p.4; S.510 (117th Congress, introduced Mar 1 2021) new IRC sec. 2901(a) and sec. 2(f) (govinfo XML). scripts/letter_arithmetic.py: nine payments 2024-2032 $2,774bn (-8%); ten payments 2024-2033 $3,147bn (+4%). PWBM (Mar 15 2021) followed the letter's timing; under CBO/JCT conventions (FY2022-31) $2.05tn.

Why this severity. Annual amounts are unaffected and the direction of the ten-year effect depends on convention (-8% to +4%); the letter predates the introduced text.

13minorconfirmed · newly identified here · selective presentation

The 2021 letter says billionaire wealth rose 40% 'since before the COVID crisis', but the comparison starts in the middle of the March 2020 market crash (and is misdated as 2019). Measured from just before the crash, the rise was about 16%.

Technical detail and evidence

Why it matters, in detail. The pandemic-era gain is overstated by a factor of ~2.5 relative to the pre-crisis level (+16% vs '+40%'); the 'top 400' ratio is for ~450 people.

Evidence. 2021 letter pp.6-7 and figure p.7; Warren bill summary repeats '40 percent higher than before the COVID crisis began'. scripts/forbes_checks.py: Forbes real-time 24 Jan 2021 663 US billionaires $4,160bn; 18 Mar 2020 $2,953bn (Forbes 2020 list $2,949bn), ratio 1.41; 19 Feb 2020 $3,591bn, ratio 1.16; Forbes Mar-2019 list $3,111bn, ratio 1.34; top 450 $3,852bn = 21.6% of 2020 national income ($17,843bn, current BEA vintage), top 400 21.0%.

Why this severity. Secondary rhetoric outside the revenue score.

14minorconfirmed · previously known · selective presentation

The comparison holds in the authors' own data, which count tax units, but in the Federal Reserve's household-based estimates the top 0.1% own well under half as much as the bottom 90%. Wealth is highly concentrated either way.

Technical detail and evidence

Why it matters, in detail. Top-0.1%/bottom-90% wealth ratio 0.65-0.86 in the authors' tax-unit series vs 0.41 in the DFA household series; the claim is not robust to unit and wealth concept.

Evidence. 2019 letter p.4; 2021 letter p.6; Warren bill summary. SaezZucman2019BPEAData.xlsx sheet DataFig2 (2016): top 0.1% B=19.61%, D (SCF+Forbes)=19.35%, E (revised)=17.83%; bottom 90% P=25.75%, Q (SCF+Forbes)=22.47%, R (revised)=27.53%. Fed DFA dfa-networth-shares.csv (current vintage) 2019:Q4 TopPt1 12.4, Next40+Bottom50 30.1. scripts/dina_vintages.py: Feb-2022 vintage 2019 top-0.1% share 18.5%.

Why this severity. A contextual claim that does not enter the revenue score.

The authors' response. SZ argue the DFA under-covers the top and that unit differences explain much of the gap (see us-wealth-taxes-08).

15minorconfirmed · newly identified here · arithmetic or transcription

A few small slips in the 2019 letter, such as a fiscal-year GDP figure, rounding up a multiplier, and slightly overstated Forbes-400 growth figures, each matter by about 1% or less.

Technical detail and evidence

Why it matters, in detail. +1% on C2 from rounding; the top-wealth growth gap used to justify 'conservative' growth is overstated by ~0.5-1.5 points a year.

Evidence. 2019 letter p.1, p.3 pt 6, fn 4; 2021 letter p.3 d); Sanders letter (Sept 2019) p.1. CBO Aug-2018 econproj: FY2019 GDP $21,141bn, CY2019 $21,383bn (scripts/cbo_check.py and inline check); scripts/letter_arithmetic.py: 12.8754.

Why this severity. Each slip matters by about 1% or less.

16minorlikely · newly identified here · internal inconsistency

The letters say the middle class lost ground mainly because it took on more debt, including consumer and student debt. The authors' own published breakdown shows debt explains only about a third of the decline and consumer and student debt almost none of it. This is a side remark, not part of the revenue estimate.

Technical detail and evidence

Why it matters, in detail. Accounting share of debt in the 1978/1980/1986-2012 decline ~0.30-0.44; non-mortgage debt ~0 (vintage to 2012; accounting, not causal).

Evidence. 2019 letter p.4; 2021 letter p.6. SZ (2016) AppendixTables(Distributions).xlsx, sheet bot90: bottom-90% share 36.4% (1986) -> 25.0% (2012): home mortgage -3.6 pp, non-mortgage debt +0.1 pp, gross homes -5.8 pp, bonds -4.9 pp, pensions +4.4 pp; debt share of the fall 32% (1978-2012), 44% (1980-2012), 30% (1986-2012). SZ 2016 QJE Sec. V.B: 'mostly the consequence of an upsurge in debt'. Codex evidence_checks (debt 2.913 of 9.101 pp, 1978-2012).

Why this severity. A contextual claim outside the revenue score.

The authors' response. SZ (2016, QJE) interpret the decline as mostly due to an upsurge in debt that offset the rise of pensions.

Claims examined and rejected

Problems one auditor raised that did not survive checking.

2021 panel B applies slower growth only to wealth above $1bn, ignoring billionaires' $50m-$1bn slab (~$29bn over 10 years) (I13 item 6). — Not an error: the $0.95bn slab of each billionaire's 2% base is fixed per person and does not scale with that person's wealth, so slowing only the above-$1bn component is a defensible approximation; the effect would be under 1% anyway.
The 2021 tax-burden paragraph is stale (I13 item 7). — The paragraph is explicitly labelled as 2019 projections, and the added burden relative to wealth (~1.1%) is about the same under the 2021 plan; not an inconsistency.
The then-current DINA vintage implies ~14% less 2021 revenue than SCF+Forbes ($217bn vs $252bn) (I3). — Not confirmed: a like-for-like comparison (same aggregate, billionaires replaced by Forbes in both) puts the Sept-2020 vintage only ~1-5% below SCF in total revenue. Only the 'slightly higher' part of the letter's claim fails (retained in F3).
The 1,005 billionaires contradict the letter's statement that SCF billionaires were replaced by Forbes (I12). — The stated sequence (replace at end-2020, then age to end-2022) mechanically admits heavily weighted SCF records crossing $1bn; retained only as a documentation and sampling-fragility point (F10, F6).
Applying the letters' own linear rule lowers 2019 revenue by ~10% (I1/I2). — Refined: on the same dataset the rate-consistency effect alone is -5.9% (linear) or -3.2% (log-linear); the remaining ~4% is the difference between the authors' Sept-2019 file and the January spreadsheet (F10).
The 5.5% CBO misattribution lowers the ten-year total by ~8% (I10). — Refined: with unindexed thresholds, CBO-type 4% wealth growth gives a 12.84 multiplier, close to the letter's 13; retained as a misattribution with little net effect (F9), the downside being captured in F5.
The estate-tax benchmark implies the effective base could be ~40-60% below the letters' assumption (I6 magnitude). — Rejected as a magnitude: the estate tax's marital/charitable deductions, trusts, lifetime gifts and weak enforcement make the analogy not like-for-like, and the Summers-Sarin $25-75bn/yr is not a credible score; retained only as qualitative evidence with an 11-18% sensitivity (F4).
The start-date mismatch with S.510 is a moderate issue reducing the 2023-2032 total by 8.3% (I5). — Severity lowered to minor: the -8.3% holds only if the calendar window is held fixed; counting the first ten payments raises the total ~4%, and annual amounts are unaffected (F12).
The Pareto error is a moderate issue (I2). — Severity lowered to minor: it changes the headline by 1.4-1.7% and the rounded surtax claim survives (F7).
The SCF calibration discrepancy is only 'likely' and is not double counting (I3). — Upgraded to confirmed: both the taxpayer count (72,143 = 71,743 + 400) and the billionaire count (671 = 271 + 400) are reproduced exactly only under the double-count scaling, and the authors' own Feb-2019 code corrects it; no individual is counted twice, but Forbes wealth enters the aggregate twice (F8).
Corpus critiques of tax-rate/incidence conventions (us-wealth-taxes-12 to -16) and of Social Security wealth (-09) apply to the letters' tax-burden and inequality paragraphs. — Dismissed (in agreement with Codex): the letters explicitly compare taxes to wealth and make no income-based progressivity claim, and Social Security wealth is not part of a wealth-tax base.
The 5.5% CBO/JCT attribution could not be verified (limitation). — Resolved: verified against CBO's Aug-2018 and Jan-2019 projection workbooks (3.9-4.0% nominal GDP growth); a confirmed misattribution (F9).

Audit notes

What reproduced. Reproduced: (i) the 2019 letter's arithmetic step by step, and the authors' spreadsheet exactly (75,288.5 taxpayers; base $9,340.5bn; $212.41bn; x13 = $2,761bn); (ii) the SCF leg exactly from the public SCF 2016 extract plus the Forbes 400 (72,143 taxpayers, 671 billionaires, base $8,189bn vs $8,179bn), but only under the aggregate double count (F8); (iii) the DINA leg within 1-3% (OPA anonymised file 77,151/$10,413bn; Jul-2019 projection 75,901/$10,207bn vs 78,434/$10,502bn); (iv) the Forbes 400 inputs exactly; (v) the 2021 table arithmetic (3,029 and 3,907 vs 3,027 and 3,906), the 4% multiplier (12.006) and 1,374 = 131.36 x 10.462; (vi) the Forbes real-time $4.16tn (663 billionaires) and the 21.6% fractile ratio; (vii) the cited elasticity figures and the IRS estate anchor (423 returns, $68.66bn, $9.32bn); (viii) the SCF 2019 base's sampling SE ($1.20tn, CV 16.6%) and concentration (one household 13.6%). Discrepant: the Pareto add-on (surtax base $2.20tn, not $2.5tn); the 5.5% CBO attribution (CBO 3.9-4.0%); the text's 10.17 multiplier; the March-2019 date and the '+40% since before COVID' (+16% from the pre-crash peak). Not reproducible: the 2021 taxpayer (100,449) and billionaire (1,005) counts jointly with the base (inputs undocumented); the 2021 claim that DINA is 'slightly higher' with the then-current vintage; the 1982 2.6% ratio (list not obtained). 2019 reconciliation: $212.4bn -> $208.8bn (Pareto) -> $206.1bn (SCF scaling) -> $178.3bn (authors' revised DINA; -16%, ~$2.3tn x13) -> $167.7bn with the letters' own rate-scaled rule (-21%) or $135.7bn with PWBM's semi-elasticity 13 (-36%). 2021: on a reconstruction, the own rule lowers panel A by 3-6% (vs a wealth baseline) and panel B by 17-26% (ten-year), so $3.9tn becomes about $2.7-3.2tn.

Limits of this audit. No 2021 code or data exist publicly, so 2021 sensitivities use an adjudicator reconstruction (SCF 2019 + Forbes real-time list) and report relative effects only. The Sept-2020 DINA file for 2019 was not obtained (that vintage's 2018 file and the Feb-2022 2019 file were used); the authors' confidential January-2019 DINA file is unavailable (public versions within 1-3%). The true avoidance response of a US wealth tax cannot be known from existing evidence: F1, F2 and F4 report sensitivities to published alternatives, not corrections. The estate-tax comparison depends on assumed mortality (1-2%). The SCF variance ignores aging and the Forbes splice. The debt decomposition uses the 2016 vintage through 2012. The bill-timing reading is textual (PWBM followed the letter). The 1982 Forbes list was not obtained. Several data files (CBO workbooks, Forbes real-time snapshots via Wayback, SCF original releases and replicate weights) came from auditor A's downloads; all key computations were re-run with the adjudicator's own scripts. National income uses the current BEA vintage. The BPEA 2019 paper, the Sanders letter and the campaign's 2%/6% estimate were used only as cross-checks.