Emmanuel Saez, Gabriel Zucman · Brookings Papers on Economic Activity, Fall 2019: 437-511

Progressive Wealth Taxation

16 problems found: 1 major · 5 moderate · 10 minor

What the paper claims

Aggregate U.S. household wealth rose from about 3 to about 5 times national income between 1980 and 2018, driven by asset prices rather than capital accumulation.
Wealth concentration has risen sharply and the data sources can be reconciled: the top 0.1% share doubled from under 10% (1980) to about 20%; the Forbes 400 share rose from 0.9% (1982) to 3.3% (2018); correcting estate data for the rising longevity of the rich closes about half of the gap between estate-based and other estimates.
The wealth tax base is large: about $12 trillion above the 99.9th percentile in 2019, so a perfectly enforced 1% tax raises about $120 billion; bases above $50m are similar across three of four sources; hiding 15% of wealth cuts the base by 13-18%.
Summers and Sarin's estate-tax-based estimate implies about 85% avoidance; correcting it for spousal/charitable deductions (x3) and differential mortality (x1.9) approximately reconciles it with the estate-based base.
The richest report only about half their economic income; the U.S. tax system is a 'giant flat tax' in which the top 400 pay the lowest total rate, less than the middle class; a Warren wealth tax would double the top-400 rate.
Avoidance and evasion are a policy choice: with a comprehensive base, third-party reporting, FATCA and an exit tax, a U.S. wealth tax can be enforced, with 15% hiding as the 'strong enforcement' benchmark.
In a tractable model of billionaire taxation the revenue-maximizing wealth tax rate is 1/(1+e_T), where e_T is wealth-weighted years of exposure; with e_T = 15 it is about 6.25%, and Forbes-400 simulations confirm that the Sanders schedule (up to 8%) is slightly below and a 10% 'radical' tax approximately at the revenue maximum.
Had the Warren, Sanders or radical tax been in place since 1982, the 2018 Forbes-400 wealth share would have been 2.0%, 1.3% or 1.0% instead of 3.3%; Forbes-400 wealth grew about 4.5 points a year faster than average.

Every problem found

1majorlikely · previously known · speculative assumption

The paper's model says billionaires' fortunes could be taxed at about 6% a year before the tax starts bringing in less money, but the model assumes billionaires do nothing to shrink or hide their taxable wealth. Using the authors' own estimate of how much the rich hide, the best rate for revenue falls to about 4%, and lower if they also save less. So the claim that an 8-10% billionaire tax is about the best for revenue does not hold up, although taxes of a few percent would still raise substantial money.

Technical detail and evidence

Why it matters, in detail. Mechanism: tau_R = 1/(1+e) with e counting only mechanical erosion; any non-mechanical response (reporting avoidance, accelerated giving or consumption, lower saving) raises the base elasticity. Using the paper's own semi-elasticity of reported wealth (7.5-8) gives tau_R of about 4.2-4.4%; PWBM's 13 gives 3.5%; JJKZ total/mechanical ratios of 1.53-2.38 give 4.2-2.7%. The revenue curve is flat near its peak, so the Sanders schedule (5.2% average) still raises 76-98% of the maximum, but the radical 10% schedule (7.2% average) falls to 49-84% and below Warren, reversing the paper's $66bn > $62bn > $49bn ranking. Direction: tau_R overstated by roughly 30-55%; transferring JJKZ's Danish ratios to U.S. billionaires is uncertain (JJKZ themselves note mechanical shares are larger for proportional taxes and high returns), hence 'likely'.

Evidence. BPEA pp. 495-498, 501; workbook DataFig6!A2 ('Assumes zero evasion rate') and A5. Kopczuk (2019, BPEA comment p. 516 fn. 3): behavioural responses are 'assumed away in the simulations'. JJKZ (2020, QJE 135(1), pp. 331, 355, 360, 381-383 and Table III): 30-year total/mechanical effects for the very wealthy 0.651/0.273 = 2.38 (R = 5%) and 0.639/0.419 = 1.53 (R = 7%); JJKZ p. 355 note mechanical effects 'would have been much larger' under a proportional tax. Adjudicator script scripts/revenue_max_rate.py: paper 6.25%; with the paper's own semi-elasticity 7.5/8 (R = tau(1-tau)^15 e^(-k tau)) 4.31%/4.23% (4.51%/4.42% with consistent payment timing); PWBM's 13: 3.50%; JJKZ ratios 1.53/2.38: 4.19%/2.72%; JJKZ R=7% plus k=8: 3.16%. Revenue at the simulated Warren/Sanders/radical average rates (2.8/5.2/7.2%) relative to the maximum: 0.77/0.98/0.99 (paper's model), 0.93/0.98/0.84 (k=8), 0.98/0.91/0.71 (k=13), 1.00/0.76/0.49 (JJKZ R=5%).

Why this severity. Under the paper's own avoidance elasticity the revenue-maximizing rate falls by about a third, and under Danish-type real responses by up to about half, enough to reverse the statements that Sanders' rates are below, and a 10% tax at, the revenue maximum (the radical schedule would raise less than Warren's). The qualitative point that billionaire taxes of a few percent sit near the revenue maximum survives, and the descriptive findings are untouched.

The authors' response. The paper itself notes (p. 498) that tau_R would be lower if billionaires accelerate giving or raise consumption, that the model assumes perfect enforcement, and that the section I.C evasion model could be added; it does not do so. No direct response to Kopczuk's footnote located.

2moderateconfirmed · previously known · speculative assumption

The paper assumes the rich would hide only 15% of their wealth from the tax, whatever the rate. That number is an average of four foreign studies and is driven almost entirely by one study the authors themselves distrust; it should also rise when rates are higher. Applied consistently, the Sanders plan's revenue and its effect on billionaires' tax rates are overstated by roughly a quarter to two-fifths, while the Warren figures barely change.

Technical detail and evidence

Why it matters, in detail. With semi-elasticity k, reported wealth is roughly exp(-k x mtr) of true wealth. At k = 8 hiding is about 15% at 2%, 21% at 3% and 33-47% at 5-8%, so a flat 15% understates base erosion only where marginal rates exceed about 2%: little for Warren (-2% to -4% revenue), a lot for Sanders (-21% to -23%, or -44% at k = 13). Without the Swiss study the four-study mean implies about 2% hiding; with JJKZ's difference-in-differences in place of bunching it implies about 16%. The 15% is therefore at the low end of defensible values for a 2% tax and too low for higher rates.

Evidence. Precursor (Feb. 5, 2019) section 2.1 and fn. 10; BPEA pp. 459-461, 468, 471-472, 482. Adjudicator re-scoring of the authors' September 2019 scoring file (wealthtaxcalc/wealth.dta; scripts/score_wealth_tax.py), reproducing Auditor B: Warren $203.5bn with flat 15% vs $199.7bn / $196.4bn with rate-scaled hiding exp(-k x marginal rate), k = 7.5 / 8, and $166.6bn at k = 13; Sanders $340.3bn vs $269.6bn / $261.4bn and $192.0bn (-21% to -44%). Forbes-400 Sanders liability falls from 5.16% to 3.16% of wealth at k = 8. Auditor A's Figure 5 replication: Sanders top-400 rate 72.9% (flat) vs 57-58% (rate-scaled; paper 74.8%). Later evidence: Brulhart et al. (2022) semi-elasticity of at least 43; Advani and Tarrant (2021) 7-17; PWBM (2019) weighted 13.

Why this severity. The Warren figures quoted in the text change by only 2-4%, and Table 2 also shows a 50% scenario. But the Sanders effects in Figure 5 and the 'strong enforcement' benchmark for high-rate schedules are overstated by roughly 20-45%, and the enforcement claim (C6) rests on an assumption that is internally inconsistent with its own semi-elasticity logic.

The authors' response. The authors discount the Swiss estimate as weakly identified and argue that avoidance depends on design and enforcement (comprehensive base, third-party reporting, exit tax); BPEA pp. 471-472 acknowledge that bunching may miss the full response. No response to PWBM's rate-scaling point located.

3moderateconfirmed · previously known · methodology

The paper's striking claim that the 400 richest Americans pay a lower total tax rate than middle-class families rests on gaps of only half a point to two and a half points. Changing one accounting choice the paper does not mention, whether part of billionaires' stock-market gains counts as income, is enough to put their rate level with or above the middle class. What does hold is that the very richest pay roughly what ordinary families pay, and much less than the merely rich.

Technical detail and evidence

Why it matters, in detail. The top-400 rate is about 23-24% with the 45% fiscal-income cap and the pure-gains denominator. It is about 26% in the public file (roughly 25% if the same shift applied to the published figure) under the denominator the BPEA text describes, against 22.7-25.6% for P0-P60. Within the Table 3 range of the cap (0.41-0.51) it moves only about +/-0.9 points, so the cap itself is not the fragile part; the capital-gains convention and the transfer/credit conventions are. Including gains is defensible (taxes on realized gains are in the numerator; Balkir et al. include half of realized gains), but it is undisclosed and decisive for 'lowest of all groups'. Robust conclusion: no progressivity at the very top, with the top 400 roughly 9-11 points below P99.9-99.99.

Evidence. Workbook DataFig5: top 400 23.04% vs P30-40 23.52%, P40-50 24.20%, P10-20 24.25%, P20-30 24.50%, P50-60 25.39%, P0-10 25.62%; P99.9-99.99 33.15%. Adjudicator public-DINA replication of the TOI code (scripts/fig5_sensitivity.py, adapting Auditor A's port; usdina2018.dta): baseline top 400 23.7% (paper 23.0%), with P30-40 22.7%, P40-50 23.3%, P50-60 25.0%. A fiscal/pre-tax cap of 0.42 gives 23.2% and 0.51 gives 24.6%. Removing the pure-gains addition (realized gains above 3% of national income, about 1.7% of NI in 2018) raises the top 400 to 26.0% (+2.3 points), above every group from P0 to P60, and P99.9-99.99 to 36.8%. Balkir, Saez, Yagan and Zucman (2025, NBER w34170, Tables 1-2): top 0.0002% 23.8% vs 30.2% for all, fiscal/economic income 42%. Splinter (2020, NTJ appendix Fig. B7; as summarized in critique corpus us-wealth-taxes-12): counting refundable credits lowers SZ bottom-quintile rates by at least 17 points.

Why this severity. It qualifies a headline claim of the conclusion: a 2-3 point change in defensible choices reverses the ordering against the middle deciles. The top-400 level (about 23-26%) and the sharp drop from P99.9-99.99 to the top 400 are robust and were later corroborated.

The authors' response. The TOI online appendix (Oct. 2019) discloses and motivates the pure-gains adjustment and says the 45% ratio 'could be refined'. The authors defend national-accounts conventions (transfers are not negative taxes; corporate tax on shareholders). Their later administrative study (Balkir et al. 2025) finds 23.8% for 2018-2020.

4moderateconfirmed · previously known · speculative assumption

Estate-tax records suggest wealth is much less concentrated than other sources show. The paper closes much of that gap by assuming the very wealthy now live much longer than average, using life-expectancy data for high earners rather than for the wealthy themselves. That is plausible but untested, and modest changes to the assumption undo a large part of the reconciliation; the paper's other wealth measures are unaffected.

Technical detail and evidence

Why it matters, in detail. The correction is parametric: share x (1+x)^(1/a). Its level depends on the income-to-wealth mortality mapping (factor 1.0-1.9 moves the estate base above the top-0.1% threshold between about $5.9T and $8.9T and 'implied evasion' between 32% and 55%). Its trend depends on the 1980 anchor (2.2 of the 8.3-point rise). The 'closes about half of the gap' claim (pp. 456-457; fn. 3) is therefore conditional, and the 33% residual cannot be read as an evasion rate. Direction of bias unknown.

Evidence. BPEA pp. 454-457, fn. 24; workbook DataFig2 H71 = 1 to H102 = 1.9 (linear), I = G x H^(1/1.5); corrected share 7.77% (1981) to 16.10% (2012), a rise of 8.33 points. Initial factor 1.45 gives a rise of 6.15 points, 26% smaller (Auditor B; Kopczuk 2019, p. 523: -2.1 points). Factor 1.5 gives a 2009-12 share of 13.7% and implied evasion of 41%; no correction gives 10.5% and 55% (Auditor A). Inputs verified by the auditors: fn. 24 arithmetic (1.905), SCF age shares 4.1% / 11.5%, Chetty top-1% relative mortality 0.406 to 0.307 (A exact; B 0.411 to 0.307).

Why this severity. It affects the estate-data reconciliation, a secondary core claim, and moves the estate-based base by up to a third. It does not affect the capitalization or SCF+Forbes series on which the headline base rests.

The authors' response. The paper flags (p. 457) that mortality of the super-wealthy may differ from that of high earners and calls for internal-data estimates; it argues that a raw estate share near 10% would make U.S. wealth as equal as Denmark's (p. 453).

5moderateconfirmed · previously known · methodology

How much the richest 0.1% own depends heavily on an assumption about the interest rate they earn on bonds and deposits. The paper's corrected estimate uses a smaller adjustment than its own survey evidence suggests; using the survey evidence trims the top 0.1% share from about 18% to about 16% and the taxable base by about a tenth. Concentration still roughly doubled since 1980 in every series, including critics' series.

Technical detail and evidence

Why it matters, in detail. Capitalization factors near 1/r make top fixed-income wealth hostage to one yield parameter (aggregate taxable rate 0.79% in 2016). Moving the ratio from 1.8 to the paper's own 2016 SCF value or to SZZ's lowers the revised share by 1.6-1.8 points. The plausible 2016 range is about 15.6-21%, and the base above the top-0.1% threshold about $10-14T, against the reported 'about $12 trillion', with no band disclosed. The doubling since the late 1970s holds in every series, including SZZ's (7.1% to 15.7%). SZ (2020) counter that SCF yields are biased up by pass-through interest, so the right ratio is disputed.

Evidence. Workbook DataFig2-extra!R53:R57 = 1.8 (note R61: 'Rise to 180% in 2012 and after'); DataFig3: SCF top 0.1% 2.50% vs aggregate 0.789% (2016), 2.07% vs 0.959% (2013); BPEA p. 452. Adjudicator script scripts/revised_cap_2016.py (matching Auditors A and B): 2016 revised share 17.83% at 1.8; 17.21% at 2.16; 16.23% at 3.17; 16.03% at 3.5; 15.56% at the AAA ratio 4.65. The 2019 base above the top-0.1% threshold falls from $11.8T to $10.3-10.7T. Header/formula mismatch: multiplier label 2.5 vs 3 used (18.09% at 2.5). SZZ (2023, QJE pp. 550-552): 15.7% in 2016 (adults). Auditor A: July-2019 public DINA files give 21.1-21.4% for 2016 (not re-verified). Kopczuk (2019, pp. 521-522): headline rise revised by 3.4 points between SZ (2016) and this paper.

Why this severity. It lowers the revised 2016 share by about 10% and the headline base (the mean of three sources) by a few percent. The doubling of concentration and the large base survive.

The authors' response. SZ (2020, NBER w27921; 2022, w30396) argue that SCF top yields are biased upward by interest received through pass-through businesses. They accept capitalizing bond funds from non-qualified dividends and a modest premium (about 1.4 times for the top 1%), and dispute SZZ's boutique-fund extrapolation.

6moderateconfirmed · newly identified here · documentation gap

The 15-year figure behind the 6.25% result and the long-run simulations cannot be checked, because the data and code have not been released, and the figure reported (time spent on the Forbes list) is not exactly what the formula needs. Since the data only start in 1982, the true long-run figure is probably larger, which would lower the revenue-maximizing rate even before any behavioural response.

Technical detail and evidence

Why it matters, in detail. tau_R(e_36) is an upper bound on the long-run mechanical rate if e_T rises with T, as the paper states. Whether 15 is wealth-weighted is not stated, and weighting matters: the top 15 alone average 22-27 years. Neither the e_T statistic, the $49/62/66bn revenues nor the 'approximately revenue-maximizing' radical schedule (DataFig6!A5) can be checked independently.

Evidence. BPEA pp. 496-498; workbook DataFig6!A1. The public scoring code (wealthtaxcalc) and the Forbes 2011-2018 panel cannot recreate the history (Auditor B). Table 4 vs the second results block in DataFig6: Page under Warren 35.3 vs 34.3, Sanders 19.5 vs 18.3, radical 13.3 vs 12.1. Table 4 is consistent with W x 0.97^n (Bezos n = 20, Gates 32, Buffett 36). The top 15's implied exposure is 26.6 years wealth-weighted (24.2 at Warren-taxed weights, about 22 at weights taxed at 6.25%), but they hold only about a third of Forbes-400 wealth. With mechanical exposure e = 20-28, tau_R would be 3.5-4.8% (scripts/revenue_max_rate.py).

Why this severity. The Section V quantities (e_T, Table 4 counterfactuals, Figure 6 paths and revenues) cannot be verified, and the 'long-run' label overstates the mechanical revenue-maximizing rate by an unknown amount. There is no evidence that the outputs are wrong.

7minorconfirmed · previously known · internal inconsistency

Figure 5 is not quite what its labels say. The 'poorest tenth' leaves out about one adult in ten with very low incomes, the income measure includes some stock-market gains, and the 'top 400' are the 400 highest earners rather than the 400 richest. These choices are explained only in a companion book appendix. They do not change the main message, and the last one actually makes the paper understate how much a wealth tax would raise billionaires' tax rates.

Technical detail and evidence

Why it matters, in detail. The plotted bottom decile is roughly the second decile of adults. The denominator is national income plus about 1.7%. The top-400 group has pre-tax income/wealth about 1.7 times the Forbes 400's, so wealth-tax increments for actual billionaires would be about 1.5-1.8 times those shown (e.g., Warren +34-41 points rather than +23).

Evidence. TOI tax-by-gperc.do (read as text) and appendix. Public-file reconstruction: 23.3m of 243.0m adults (9.6%) dropped, so 'P0-10' is about P9.6-P18.6 (Auditors A and B; Kopczuk 2019 p. 524: about 26m). Keeping them puts the bottom-decile rate near 50% (adjudicator replication 49.5%). The pure-gains addition is about 1.7% of national income in 2018. TOI full2018.xlsx top-400 row: pre-tax income $416m per adult on wealth of $4.27bn (9.7%; 10.7% with gains), vs about 5.6% economic income/wealth for the Forbes 400 (Balkir et al. 2025, Table 1).

Why this severity. These are documentation and labelling problems, disclosed only in a companion appendix. The top-400 mislabel makes the paper's wealth-tax increments conservative, and the substantive consequence for the ordering is covered in F3.

The authors' response. The TOI appendix explains the minimum-wage cut (to avoid infinite rates) and the pure-gains adjustment (to avoid artificial rate spikes).

8minorconfirmed · newly identified here · arithmetic or transcription

One step in the paper's rebuttal of Summers and Sarin counts estate debts twice, so the correction factor should be about 2.7 rather than 3. The rebuttal still roughly works.

Technical detail and evidence

Why it matters, in detail. The combined correction is 2.66 x 1.9 = 5.05 rather than 5.7. Summers-Sarin's $25bn becomes $126bn, a base of about $6.3T at 2%, instead of $142.5bn ($7.1T); both sit close to the estates-based $6.8T.

Evidence. IRS SOI 17es01fy.xlsx, all returns with gross estate of $50m or more: gross $68.66bn; debts $4.33bn (6.31%); total deductions $43.72bn = 67.96% of net estate; spouse 40.15%, charity 20.18%; taxable estate $24.21bn = 37.64% of net estate, a factor of 2.657 (adjudicator script scripts/irs_estate_2017_check.py, matching Auditors A and B).

Why this severity. The reconciliation still approximately holds.

9minorconfirmed · newly identified here · internal inconsistency

A small timing slip in the model's formula makes the computed best rate slightly too low (6.25% instead of about 6.7%). It does not change any conclusion.

Technical detail and evidence

Why it matters, in detail. The inverse-elasticity rule holds once W_A is defined before the current payment and e counts only prior payments; with 15 included payments the mechanical optimum is 6.67% rather than 6.25%.

Evidence. BPEA pp. 495-497. Consistent timing gives R = tau W_A/(1-tau), so tau_R = 1/e when e counts the current payment: 6.67% for e = 15 (Auditor B; scripts/revenue_max_rate.py).

Why this severity. This is a one-period indexing slip. It moves tau_R by 0.4 points in the authors' favour (opposite to F1) and does not affect the simulations.

10minorconfirmed · newly identified here · methodology

One unusual estate, Steve Jobs's, counts very heavily in the estate-based estimate because he died young, and it inflates the paper's estate-based numbers by up to about a tenth. Removing it entirely would understate wealth, so the right figure lies somewhere in between; the effect is small.

Technical detail and evidence

Why it matters, in detail. The estate-multiplier estimator is unbiased but noisy at the top. Rare high-weight young decedents stand in, on average, for real wealth, so dropping Jobs biases the estimate downward. A three-year window containing one weight-200 death probably over-represents such events, so the defensible level lies between 14.4% and 15.95%. The paper does not disclose that its Table 2 estate base carries part of a spike it attributes to a single estate.

Evidence. Workbook DataFig2 G101 = 3/5 F101 + 2/5 F102; G102 = 3/5 F102 + 2/5 F101 (raw and smoothed 2011-12 mean both 10.34%); Table2!M21 = AVERAGE(DataFig2!J99:J102) = 15.95%. Replacing raw 2011 with the 2009/2012 mean (adjudicator recomputation, matching Auditor A) gives a corrected share of 14.40%, a base above the top-0.1% threshold of about $8.0T instead of $8.9T, implied evasion of about 38% instead of 32%, and a 1981-2012 rise of about 6.5 points instead of 8.3.

Why this severity. The effect is at most about 10% of a secondary series, and the direction of the right correction is debatable.

The authors' response. Footnote 26 identifies the spike; the smoothing is described as noise reduction (p. 456).

11minorconfirmed · previously known · misrepresentation of source

The paper cites an IRS study as evidence that the rich increasingly dodge the estate tax. The study itself says the gap mostly reflects legal valuation rules, family ownership and Forbes being generous, and that it did not grow over time. It still shows that legal valuations of large fortunes can be far below Forbes' figures.

Technical detail and evidence

Why it matters, in detail. The source supports some legal valuation avoidance (discounts up to 50%, trusts) but not a rising trend. Its view that Forbes is generous bears on the Forbes-based top end: the Forbes 400 are about 30% of the SCF+Forbes base above $50m. The estate/Forbes gap is not an evasion rate.

Evidence. Raub, Johnson and Newcomb (2010), NTA Proceedings pp. 130-134 (read directly). Ratio about 0.50 overall, with a slight decline of 0.6% a year over 1982-2009. Adding jointly held assets for married decedents raises it from 0.46 to 0.53; adding discounts, 0.47 to 0.54; both, 0.58. Quote: 'Previous researchers have suggested that differences between tax and Forbes estimates are due to evasion; however ... they will fall within legally defensible parameters.' Balkir et al. (2025, p. 3): taxable estates are about 30% of Forbes wealth on average.

Why this severity. No estimate in the paper depends on the citation. It misuses the source for an interpretive claim (rising avoidance) that is otherwise supported mainly by falling audit rates.

The authors' response. SZ's June 25, 2019 response to Summers and Sarin reads the same gap as 'about 50% of estate taxes owed are avoided'.

12minorconfirmed · newly identified here · selective presentation

The paper's comparison of audit rates mixes the largest estates in 1974 with all estate returns in 2017, many of which owe no tax. Big estates are audited far less than in 1975, but the drop is roughly from two-thirds to between one-fifth and one-third, not to under a tenth.

Technical detail and evidence

Why it matters, in detail. For estates with tax at stake, coverage fell from about 65% to about 19-31%, not to 8.6%. The weakening of enforcement is real but about half as steep as the juxtaposition suggests.

Evidence. IRS 1975 Annual Report Table 2: 29,069 returns with gross estate of $300k or more, 18,834 examined (64.8%); $300k in 1974 is about $1.5m in 2017 dollars. IRS Data Book FY2018 Table 9a: 33,690 returns, 2,898 examined (8.6%); under $5m 21,223 (2.7%); $5-10m 8,324 (12.6%); $10m or more 4,143 (31.0%); $5m or more 18.7% (adjudicator check of both files, matching Auditors A and B).

Why this severity. Both figures are accurate and the decline is real; only its size is exaggerated.

13minorconfirmed · newly identified here · selective presentation

The paper quotes the part of an economists' poll that supports it (a wealth tax would reduce inequality if it were enforced) but not the part that does not (most of the same experts thought it would be much harder to enforce than existing taxes).

Technical detail and evidence

Why it matters, in detail. Expert opinion on enforceability, the premise of C6, runs 82:9 against the paper's framing; the cited question concerns effects conditional on successful enforcement.

Evidence. BPEA pp. 487-488 and fn. 76. Kent Clark Center 'Wealth Taxes' page (local copy): Q.B 19% strongly agree + 54% agree, 12% disagree (confidence-weighted). Q.A: 34% strongly agree + 48% agree, 9% disagree (Saez among the disagreeing) that the tax 'would be much more difficult to enforce than existing federal taxes because of difficulties of valuation and the ways by which the wealthy can under-report their true wealth'.

Why this severity. The omission is rhetorical and has no effect on any estimate.

14minorconfirmed · newly identified here · documentation gap

A table reports survey 'sample sizes' that count each household five times, so the evidence rests on about 20 and 100 families rather than 86 and 465; the numbers themselves check out.

Technical detail and evidence

Why it matters, in detail. The effective sample is about 4-5 times smaller than labelled, which widens the uncertainty around the 3.2% ratio and the factor of 2 used to map the IRS top 400 to the Forbes 400.

Evidence. Adjudicator script scripts/scf2016_table3.py on the public SCF 2016 summary extract (matching Auditor B): the top 0.001% by wealth is 85 records from 21 distinct families (YY1), with income/wealth 3.27%; the top 0.01% is 464 records from 100 families (3.21%); the top 0.001% by income is 63 records from 15 families.

Why this severity. The point estimates reproduce; only the stated precision is overstated.

The authors' response. Page 466 calls them 'records' and acknowledges small samples.

15minorconfirmed · newly identified here · arithmetic or transcription

There are several small numerical and labelling slips in the paper's main revenue table and text; none changes a conclusion.

Technical detail and evidence

Why it matters, in detail. Collectively these indicate a lightly documented workbook (hard-coded cells; header/formula mismatches; multiple simulation vintages), but none moves a headline by more than about 5%.

Evidence. (a) Table 2 note and p. 459 say hiding '10 percent or 50 percent'; the rows use 15% and 50% (Table2!K28 = 0.15). (b) Tax units are 173m (p. 459), 174.2m (Table2!K8) and 175m (fn. 19). (c) '2019' Financial Accounts wealth is 2018 x 1.055 and national income 2018 x 1.05 (Table2!K6, J6). (d) The revised fixed-threshold bases use an ad hoc 0.95 factor (Table2!G13, H13). (e) The estate Pareto coefficient is fixed at 1.5 while the paper's own estate shares imply 1.458 (+2.9% base). (f) The SCF base above the top-0.1% threshold is hard-coded at $11.45T, against $11.06T implied by its own share (19.38%) and threshold ($40.59m). (g) Page 462 moves from a 15% vs 20% share gap (25%) to '33 percent' evasion, a figure that also embeds the 10.8% pension discount. (h) Page 447: 0.48% x $17T = $81.6bn, not $90bn. (i) The DataFig2-extra!AL1 header says multiplier 2.5; the formulas use 3. (j) 'The 75,000 families with net worth above $50 million' (p. 483) vs Table 2's 97k (capitalized) and 91k (SCF) and 92.6k in the authors' scoring file. (k) The 1982 Forbes-400 share is 0.934% (DataFig6!B8; '0.9 percent', pp. 439, 500) vs 0.83% in the cited source (Zucman 2019, Table C3); the higher start understates the rise. (l) Two Section V output blocks give different Table 4 values (Page 35.3 vs 34.3). (m) The 1976 vs 2017 estate-tax/wealth comparison mixes federal+state (0.72%) with federal-only (0.135%) taxes and different denominators; like-for-like ratios still show a 4.6-5.6x decline.

Why this severity. Each slip is small, a few percent at most, and several run in the conservative direction.

16minorconfirmed · newly identified here · misrepresentation of source

Four supporting statements get their sources or history wrong (for example, Britain did impose top inheritance tax rates as high as America's). None affects the main results.

Technical detail and evidence

Why it matters, in detail. Item (c) weakens one argument the paper offers against SZZ's AAA-rate capitalization. Items (a), (b) and (d) are historical or citation inaccuracies with no effect on the estimates.

Evidence. (a) Fn. 2 (p. 438): 'No European country ever imposed similarly high top inheritance tax rates'. Piketty (2014) Table S14.2 shows UK top rates of 75% (1946-48), 80% (1949-68), 85% (1969-71) and 75% (1972-83), and Germany 60% (1946-48), vs the U.S. 77%; the U.S. was, however, first to reach 70% (1936). (b) Page 487 says Guvenen et al. (2019) find a wealth tax 'increases aggregate productivity and output (7.5 percent in consumption-equivalent terms)'; the 7.5% is a newborn welfare gain (NBER w26284 abstract). (c) Fn. 21 (p. 452) puts domestic and foreign corporate bonds at about 30-40% of mutual funds' bonds, citing table L.122. L.122 shows 'corporate and foreign bonds' at 45-47% of bonds (54-56% excluding munis) in 2016-18 (2021Q2 vintage, adjudicator; similar in the current vintage, Auditor A), with no split for foreign sovereigns. (d) Page 457 says 1976 had 'no marital deduction'; a deduction of up to half the adjusted gross estate existed from 1948, was increased in 1976 and became unlimited in 1981 (Jacobson, Raub and Johnson 2007, pp. 122-123).

Why this severity. None of these statements carries any of the core findings.

Claims examined and rejected

Problems one auditor raised that did not survive checking.

The conclusion's 'Forbes 400 wealth grew about 4.5 percentage points faster annually than average' is inconsistent with the paper's own Figure 6 normalization, which implies about 3.5 points (A-I11, first part). — The statement is accurate under its natural reading. Forbes-400 total (average) wealth grew from $92bn in 1982 to $2,892bn in 2018 (10.05% a year), while aggregate household wealth grew from $8.90T to $88.66T (6.59% a year), about 5.3-5.6% a year per household or family, a gap of about 4.5-4.7 points. Figure 6's constant-population-fraction share (3.5 points) measures a different object. Auditor B also declined to call it an error. The separate 1982-share discrepancy is retained in F15(k).
The top-400 rate is mechanically produced by the 45% cap: without the cap it would be 30.1%, and at the internal-file ratio of 0.65 it would be 26.6% (A-I4). — Rejected as evidence against the level. The uncapped ratio (0.94 in the public file) is the artifact the cap corrects, because DINA imputes wealth from realized income. The authors flag the 0.65 internal-file ratio as biased upward (fn. 39). Matched administrative data later give 42% and a 23.8% rate (Balkir et al. 2025). Within the paper's own Table 3 range (0.41-0.51) the rate moves only about +/-0.9 points (adjudicator replication: 23.2-24.6%). The fragility of the ordering is retained in F3.
Feeding JJKZ's behavioural responses into the paper's formula roughly halves the revenue-maximizing rate to about 2.3-3.2% (1.45% with eight-year ratios) (A-I1, magnitude). — Partly rejected as overstated; the issue itself is retained as major (F1). The ratios come from a progressive Danish tax near its threshold, and JJKZ (p. 355) say mechanical effects 'would have been much larger' for a proportional tax. JJKZ Table III gives a total/mechanical ratio of 1.53 at a 7% return, implying about 4.2%. The paper's own avoidance elasticity gives about 4.2-4.4%. A defensible range is about 3-4.5%, not 1.5-3.2%.
A rough wealth-weighted tenure of about 28 years for the top 15 shows e_T = 15 understates exposure (A-I2, inference). — Not established. Recomputing from Table 4 (W x 0.97^n) gives 24-27 years for the top 15 (about 22 at weights taxed at 6.25%), but the top 15 hold only about a third of Forbes-400 wealth, and the remaining fortunes are younger, so the full wealth-weighted e_T could be near 15. The non-reproducibility and truncation points are retained in F6.
The estate smoothing 'keeps the Steve Jobs outlier', inflating the estate share, base and trend by about 10-20% (moderate) (A-I5). — Downgraded to minor (F10). The estate-multiplier estimator is unbiased but noisy, so excluding a valid high-weight young decedent biases the estimate downward; the correct figure lies between the auditor's no-spike value and the paper's.
The Raub et al. misreading is a moderate issue affecting the reconciliation and the Forbes-based base (A-I8, severity). — Severity reduced to minor (F11). The paper's sentence is literally accurate, no estimate depends on it, the source does document legal valuation discounts (a form of avoidance), and the Forbes-generosity reading mainly concerns hard-to-value private fortunes rather than the listed-equity fortunes at the top.
Consistent measures make the 1976-2017 decline in estate taxes as a share of top wealth larger than the paper states (A-I15(i)). — Not established. Like-for-like comparisons from the Estates1976 sheet give 0.62% to 0.135% (federal, Kopczuk-Saez denominator), 0.73% to 0.135% (federal, SZ denominator) and 0.85% to 0.151% (federal+state, SZ denominator): a decline of 4.6-5.6 times, against the paper's 0.7% to 0.13% (5.4 times). Immaterial either way; the mixed definitions are noted in F15(m).
The 15% derivation uses Seim at 0.5 whereas the BPEA text cites 0.2 (A-I3, sub-point). — Immaterial: the four-study mean moves from 8.0 to 7.9 and the haircut from 16% to 15.7%. The substantive problems with the haircut are retained in F2.
The 2018 Figure 5 rates were projected before 2018 tax data existed and are therefore unreliable (part of A-I7; also critique us-wealth-taxes-14). — Rebutted for the top-400 level by later matched administrative data: Balkir et al. (2025) find 23.8% for 2018-2020, against 23.0% in the paper. Concerns about the bottom and middle groups are conventions, retained in F3 and F7.
The July-2019 public DINA micro-files give 2016 top-0.1% shares of 21.1-21.4%, so the 19.6% capitalized share used in Table 2 reflects an undisclosed vintage choice (part of A-I9). — Not treated as an issue against the findings. It is not re-verified here, and if correct it makes the paper's base conservative. It is reported only as the upper end of the uncertainty range in F5.
Table 2 applies revised shares to the unadjusted common 2019 aggregate instead of carrying the revised (larger) balance sheet, understating the revised top-0.1% base by $1.11T (Auditor B's 'important candidate', not classified as an error by B). — Agreed with Auditor B: the paper discloses that all sources are extrapolated with a common Federal Reserve aggregate (p. 459), a defensible normalization rather than an arithmetic error. The alternative would raise, not lower, the base.

Audit notes

What reproduced. Reproduced exactly or closely: Table 1 and Figure 1 arithmetic (net worth $88.66T, 5.03 times national income; 645 inherited source cells match, per Auditor B); all Table 2 formulas, including base = share/a, thresholds and the hiding scale-downs (13.0% / 17.7%) (A, B, adjudicator); the revised capitalization share of 17.83% and its composition (A, B, adjudicator); the corrected estate series (15.95% for 2009-12) and the mortality arithmetic (1.905; 1.9^(2/3) = 1.53) (A, B, adjudicator); the Chetty top-1% mortality trend 0.406 to 0.307 (A exact, B close); Table 3's SCF ratios of 3.2% (A, B, adjudicator: 3.27% / 3.21%) and column 4 inputs (A); the SCF+Forbes 2016 top-0.1% share (19.0-19.3% vs 19.35%) and SCF bases (A, B); Figure 5 baseline rates, matched exactly to the TOI appendix by B and reproduced on public DINA within about 1-3 points by A and the adjudicator (top 400 23.7% vs 23.0%); IRS estate and audit figures (A, B, adjudicator); the Forbes 2018 total of $2,891.7bn (B); the internal consistency of Table 4 and Figure 6 (A, B); and the authors' scoring file (Warren $203.5bn, Sanders $340.3bn at 15%) (B, adjudicator). Not reproducible: the Section V historical Forbes-400 simulations (e_T = 15, the Table 4 counterfactual columns, Figure 6 paths, $49/62/66bn), because the panel, inheritance links and code are not public. Discrepant: the x3 deduction factor (2.66); the 1982 Forbes share vs the cited source (0.934% vs 0.83%); 75,000 vs 91-97k families above $50m; the hard-coded SCF base (11.45 vs 11.06); fn. 21's bond composition; the fee arithmetic ($81.6bn vs $90bn); the two simulation output blocks. New adjudicator computations: the revenue-maximizing rate under the paper's own avoidance elasticity (about 4.2%) and the revenue ranking of the three schedules; the removal of the pure-gains denominator in Figure 5 (top 400 23.7% to 26.0%); the Swiss share of the 15% haircut (89%).

Limits of this audit. The Forbes 1982-2018 panel and the Section V simulation code are not public, so e_T, Table 4 and Figure 6 could be checked only for internal consistency. Figure 5 sensitivities use the public DINA 2018 micro-file and a Python port of the TOI code (adapted from Auditor A's), which reproduce the published rates only to within about 1-3 points; the published internal-file effect of removing the pure-gains adjustment is inferred, not computed. Behavioural quantifications (own-elasticity, PWBM and JJKZ ratios) are scenarios: transferring Danish, Swiss or bunching estimates to U.S. billionaires is uncertain. Not re-verified by the adjudicator: Auditor A's July-2019 DINA vintage shares (21.1-21.4%), A's rate-scaled Figure 5 Sanders replication (57-58%; cross-checked only with the scoring file), the Chetty mortality tabulation, and Splinter's credit-adjusted rates (taken from the critique corpus). Summers and Sarin's op-eds (paywalled) and Scheve and Stasavage (2016) were not read; their content is taken from the paper and the corpus. Legal feasibility, administrative and compliance costs, and general-equilibrium effects were not adjudicated. Later evidence used as corroboration (Balkir et al. 2025) is co-authored by Saez and Zucman.