← All papers
Emmanuel Saez, Gabriel Zucman · NBER Working Paper 27921, October 2020
Trends in US Income and Wealth Inequality: Revising After the Revisionists
16 problems found: 1 major · 5 moderate · 10 minor
What the paper claims
- Smith-Yagan-Zidar-Zwick's 25% capital share of pass-through profits is inconsistent with these firms' capital stock; about 50% is capital, and a majority of top-1% income derives from capital.
- Smith-Zidar-Zwick (SZZ) assign the wealthy an interest rate higher than in datasets where both income and wealth are observed, biasing top wealth shares downward; capitalizing at the rate of top interest earners (r-bar) creates a first-order downward bias.
- Capitalizing equities almost only with dividends (90/10) makes SZZ dramatically underestimate billionaire wealth relative to Forbes; with business-wealth fixes, the SZ-SZZ gap closes.
- Auten-Splinter (AS) classify untaxed top-1% business profits (notably excess depreciation) as tax evasion and allocate evasion mostly to the bottom 99% from an erroneous reading of random-audit data; this explains about half of the AS-PSZ level gap and most of the trend difference.
- The revised series show a rise of inequality similar to Saez-Zucman (2016) and Piketty-Saez-Zucman (2018), and are consistent with the Fed's Distributional Financial Accounts.
- SZZ's 2001-2016 trends are inconsistent with SCF and DFA evidence, while the original SZ method is consistent with them (Section 2.1 consistency check).
Every problem found
1majorlikely · previously known · overclaiming
The paper says one group of critics mishandles income that never appears on tax returns, and that this explains about half of why their numbers differ. Checked against the government's accounts and the IRS audit study it relies on, the effect more plausibly explains a quarter to a third of the difference. The critics' method does still seem to understate the top 1%'s income somewhat, so the rebuttal points the right way but is roughly twice as strong as the evidence allows.
Technical detail and evidence
Why it matters, in detail. The 2.4 pp combines (i) an unsupported 60-70% top-1% share applied to a residual (7.2 - 4.3 = 2.9% of NI) that is mostly wage-coverage and S-corp/farm/rental gaps rather than partnership excess depreciation, and (ii) a comparison of the Johns-Slemrod gross 27% with AS's net 9%. Holding SZ's other inputs, a 20-30% residual share gives 1.1-1.4 pp. Using the net-equivalent 17.8% for the misreporting part gives about 1.0 pp. AS's own Table 4 attributes 1.2 of the 4.3-pp 1979-2014 trend gap (28%) to this item, so 'most of the difference in the rise' is also overstated. The BEA vintage behind '4.3%' (misreporting plus evaded wages was $764bn, about 4.9% of NI, in Splinter's figures; 4.44% in the 2026 vintage) could not be settled.
Evidence. Paper pp. 26-29, fn. 36-42, Figs. 17-18. Johns & Slemrod (2010), author manuscript, Table 3: 7+20=27, 2+3=5; Table 5: 3.8+14.0 = 3.7+14.1 = 17.8. Splinter (17 Nov 2020, 'Reply: Trends in US Income and Wealth Inequality: Revising After the Revisionists'), pp. 2-5 and fn. 9. NIPA, Sept-2026 vintage (scripts/04), 2015: Table 7.14 misreporting $604bn, revisions -$414bn, non-farm CCAdj $223bn = 1.42% of NI; Table 7.18 misreporting $92bn, 'other' $235bn. AS (Dec 2019) Table 1: adding underreported income moves the 2015 top-1% share from 17.2% to 16.4%. AS Table 4: the underreported-income item is 2.2 of the 5.9-pp 2014 level gap and 1.2 of the 4.3-pp 1979-2014 trend gap. Top-1% share of partnership income in the public Sept-2020 DINA 2018 file (scripts/08): 68.4% positive, 77.9% net. Sensitivity (bridge = 2.9 x s + 4.3 x 0.27 - 0.6): s = 0.30 gives 1.43; s = 0.20 gives 1.14; with the net-equivalent 17.8% for misreporting and s = 0.30, 1.04; with a 2.3-pt residual (misreporting 4.9%, Splinter's BEA figures), 1.41.
Why this severity. The quantitative core of C4 (2.4 pp, about half the 2015 level gap, and 'most' of the trend gap) falls to roughly 1.0-1.9 pp under defensible alternatives. That is about 20-40% of the gap, a reduction of the claimed magnitude by 20-60%. The qualitative point survives: AS's procedure puts less underreporting in the top 1% than an audit-consistent allocation would.
The authors' response. No direct reply to Splinter (2020) located. In later comments, as summarized in corpus us-income-10/-12, PSZ (2024; 2026) maintain that partnership excess depreciation and S-corporation exempt income are top-concentrated and explain 41% of the 1979-2019 trend gap. They attribute only about 5% of that gap to evasion.
2moderateconfirmed · previously known · overclaiming
To estimate how much the richest own in bonds and deposits, researchers divide their interest income by an assumed interest rate, so the answer depends heavily on that rate. The paper says the evidence clearly shows the very rich earn only slightly more than average. The authors' own survey figures for the richest 0.1% are mixed, and in the key year they point the other way. The paper's logical objection to the rival method holds, but how much it matters is uncertain: anywhere from a small fraction of what the paper claims up to roughly the full amount.
Technical detail and evidence
Why it matters, in detail. With beta ~ 0.25 and SZZ's effective r-bar = 3.0%, the first-order factor 1 - beta(1 - r/r-bar) gives a 2016 bias of +2.2 pp with the estate rate (1.44%), +1.0 pp with the SCF concentration-adjusted rate (2.22%) and +0.3 pp with the uniform-corrected rate (2.75%). For the revised series, moving the top premium from 1.4x to 2x lowers the recent top-0.1% share by about 0.9 pp (simulation on the public 2018 micro-file, aggregates held fixed), roughly 30% of its +2.9-pp 2001-2016 rise. At SZZ's 3.5x the reduction is about 1.8 pp (about 60%). The series also shows low-rate amplification: top-0.1% interest-bearing wealth swings by +3.6 then -2.2 pp of total wealth over 2000-2016.
Evidence. Workbook DataF6, rows 2007-2019 (cols O-T) and notes A35-A46; DataF14 D25 = 2.48%; DataF4 cols I-J; Fig. 6 caption; fn. 15; pp. 7-21, 46. SZ 'Comments on SZZ' (14 May 2020): r/rm = 1.57 for the SCF top 0.1%. SZZ reply (7 Oct 2021) pp. 1-2, 19-20: SCF 2016 top-0.1% rate 2.3% -> 3.9% (s.e. 1.0) after removing mutual funds; top/average ratio ~3.5 (CI 2.8-4.3). scripts/11 (beta = 0.25, r-bar = 3.0%): r = 1.44% gives +2.1 pp, 2.22% +1.0, 2.33% (estates 2015) +0.85, 2.75% +0.30. Codex's sensitivity on the public 2018 DINA micro-file: raising the premium from 1.4x to 2x lowers the top-0.1% share by 0.94 pp (18.11% -> 17.17%); 3.5x lowers it by 1.84 pp. DataF24: top-0.1% interest-bearing assets +3.62 pp of total wealth over 2000-2012 (total share +2.88), -2.18 pp over 2012-2016. IRS SOI 2016 Table 2.4 (scripts check): cash $641.2bn of $877.4bn.
Why this severity. The conceptual critique stands: capitalizing at top interest earners' rate biases shares down, and SZZ later dropped interest-ranking. The 1.4x choice is close to the 2008-2016 estate average (1.35x). But the claimed SZZ bias ranges from 0.3 to 2.2 pp across the authors' own sources and years. A top-0.1% premium near 2x would lower the revised recent top-0.1% share by about 1 pp and cut its 2001-2016 rise by about a third. That is material for recent trends but small against the long-run rise.
The authors' response. SZ (2022, NBER WP 30396) argue that SZZ's high top yields come from extrapolating a small, unrepresentative sample of investment funds, and that SZZ underestimate top interest-bearing assets by a factor of 1.6 (hedge-fund filings). SZZ (2021 reply; QJE 2023) statistically reject SZ's 1.4 ratio. Unresolved.
3moderateconfirmed · newly identified here · selective presentation
The paper checks a rival's numbers against Federal Reserve survey data and finds them far off. Its own revised numbers would also fall noticeably short of that test since 2001, though by much less. Over the longer period since the late 1970s, the old and revised series really are similar; it is the recent-decades comparison that is presented too favorably.
Technical detail and evidence
Why it matters, in detail. Against the paper's own benchmarks, SZZ's draft reaches 18-27% of the 2001-2016 rises, and the revised SZ series reaches about 58-68% (top 0.1%, vs SCF and SCF+Forbes) and 69-93% (top 1%, vs SCF/harmonized DFA and raw DFA). By 2021-22 SZZ's revised rise (2.1 pp, per SZZ 2021) and SZ's (2.9 pp) were close, as SZ (2022) acknowledge: 'close in both levels and trends except for the top 0.01%'. The long-run claim is unaffected.
Evidence. Workbook DataF1, DataF2, DataF24, DataF25, DataF26 (scripts/12); public SCF summary extracts 2001/2016 (scripts/06); DFA May and Sept 2020 vintages (scripts/10); paper pp. 2-5, 46-47; notes to Figs. 1, 25, 26.
Why this severity. The long-run similarity holds: the equal-split top-0.1% rise over 1978-2016 is +11.3 pp, and the income rise is 90% of the original. But over the 2001-2016 window central to the SZZ debate, the revision moved about 30% of the way toward SZZ and falls 30-40% short of the benchmark rises. The paper does not acknowledge this while using the same benchmarks against SZZ.
The authors' response. SZ (2022, NBER WP 30396, abstract and p. 1) state that SZ (2020) and SZZ (2022) are close in levels and trends except for the top 0.01% (9.3% vs 7.1% in 2016).
4moderateconfirmed · newly identified here · documentation gap
A key number in the paper is that about half of private-business profits are a return on the business's machines, buildings and other assets rather than pay for the owner's work. It comes from a study the authors have only ever released as presentation slides. The paper's own simpler checks do suggest the rival's one-quarter is too low, but readers cannot verify the one-half.
Technical detail and evidence
Why it matters, in detail. Moving the pass-through capital share between about 25% and 50% shifts the 2014 top-1% capital share between about 41% (SZZ-type assumptions, per Kopczuk & Zwick 2020) and 56%. The claim that 'a majority of income at the top derives from capital' therefore depends on undocumented estimates; top income shares themselves change little (Fig. 22).
Evidence. Author homepage (gabriel-zucman.eu, cached 25 Oct 2026): 'Labor and Capital in American Businesses (with Emmanuel Saez), preliminary slides, November 2020'; SaezZucman2020LK.pdf metadata; OpenAlex search; workbook DataF20 (source note cites the unreleased paper); paper pp. 34-42, fn. 59.
Why this severity. The qualitative claim that 25% implies implausibly low returns is supported by reproducible arithmetic. But the specific ~50%, and the rule behind the 56% capital share of top-1% income and the business-wealth concentration in the revised series, cannot be checked six years on.
5moderateconfirmed · previously known · methodology
The paper treats its own revised figures as precise, and as if they could only be too low, while rejecting its critics' figures. It never shows how much its numbers would move under reasonable alternative assumptions. Those alternatives shift the recent figures by about as much as the disagreements being judged.
Technical detail and evidence
Why it matters, in detail. Single plausible changes (top premium 1.4 -> 2x; owner-linked CCAdj key) each move recent top shares by about 1 pp. That is comparable to the residual 2016 top-0.1% gap with SZZ's published series (18.4% vs 15.7%) and to individual AS-PSZ reconciliation items. Without bands, the claim that rival estimates lie outside a plausible range is not demonstrated.
Evidence. Workbook DataF1, DataF2, DataF24; paper abstract, pp. 2-3, 45-53; F2 and F4 sensitivities; corpus us-income-15 (Iselin & Reck 2025).
Why this severity. Unreported uncertainty of the same order as the SZ-SZZ and SZ-AS gaps does not overturn the long-run rise, but it undercuts the strength of the comparative verdict.
The authors' response. SZ (2020, pp. 2-3) flag private-business valuation and evasion as areas of future revision and call for better public statistics; no uncertainty bands are provided in this or later SZ series.
6moderateplausible · previously known · speculative assumption
Tax rules let businesses deduct investment costs early, which makes their taxable profits look smaller than their real profits. The paper assigns nearly all of this hidden profit to the richest 1% because they own most partnerships. The figure it cites (90%) is higher than the authors' own public data show (about 70-78%), and the owners taking these deductions may not be those with high reported partnership income. This matters for how much top incomes grew after 2000, but little for the long-run rise.
Technical detail and evidence
Why it matters, in detail. The CCAdj grew by about 1.5% of NI over 2000-2018. Allocated like reported partnership income (top-1% share 68-78%), it adds about 1.0-1.2 pp to the top-1% pre-tax share; at a linked-owner share of roughly 10-30% it would add about 0.15-0.45 pp. The difference (about 0.6-1.0 pp) is 35-60% of the revised 2000-2018 rise. It also underlies the claim that this channel explains 'most' of the AS-PSZ trend difference. Over 1978-2014 it is about 6-11% of the rise.
Evidence. scripts/08 on usdina2018-sept2020.dta and usdina2016.dta (July 2019); Cooper et al. (2016) Tax Policy and the Economy 30, p. 94; NIPA B047RC/A032RC (scripts/04); workbook DataF17; DataF2 (revised top-1% income share +1.65 pp, 2000-2018); corpus us-income-12 (AS 2026 Table 1: partnership excess depreciation adds +1.2 pp to the 2019 top-1% share under the PSZ key vs +0.16 pp with AS's linked owners).
Why this severity. The allocation is untested, and the stake is large relative to post-2000 trends: about 0.6-1.0 pp of the revised +1.65-pp 2000-2018 rise. It is small relative to the long-run headline (about 0.5-0.9 of the 8.5-pp 1978-2014 rise). The true distribution remains unresolved.
The authors' response. As summarized in corpus us-income-12, PSZ (2024; 2026) argue that BEA and IRS aggregates imply slightly more than 100% of non-corporate excess depreciation belongs to partnerships. They add that AS's original key (depreciation on individual returns) is 95% sole proprietorships.
7minorconfirmed · newly identified here · arithmetic or transcription
A quick sum meant to show that a rival's estimate, properly corrected, matches the authors' counts one piece twice. It reaches about 45% rather than 50%. The broader point that the rival's 25% is too low still stands.
Technical detail and evidence
Why it matters, in detail. Corrected reconciliation: 25 + 11.5 + 8 = 44.5% (versus 50%), closing about 78% of the gap to the ~50% estimate. Mixing firm-count and dollar weights adds further imprecision.
Evidence. Paper pp. 38-40; IRS SOI 2014 Form 1120S Table 2.4 (scripts/09: share of returns 0.19%, share of ordinary business income 20.9%).
Why this severity. An illustrative reconciliation is off by 5.5 pp. The conclusion that the capital share is well above SYZZ's 25% is unaffected, and the revised series does not use this calculation.
8minorconfirmed · newly identified here · data error
A chart meant to show that a rival method undervalues real-estate companies uses data from the wrong year and counts one income item twice, inflating the 'true profit' bar by about a third. Corrected, the rival's measure is still several times too low, so the point survives, but the chart as published is inaccurate.
Technical detail and evidence
Why it matters, in detail. Corrected 2014 values: SZZ-style EBITD about $6.1bn vs actual about $34.3bn (ratio 5.7). The 8825/rental omission alone gives 4.2x, and the 75% ordinary-income reclassification is about 31% of the gap. The headline 'factor of 5' is therefore roughly right in total, but not as a pure measure of the omitted schedule, and the published bar heights are wrong.
Evidence. Workbook DataF16; IRS SOI Form 1120S 2014 and 2016 Tables 6.1, 7, 8 (scripts/01, 02, 03; output work/fig16_corrected.json); identity total net income = ordinary business income + portfolio + real-estate rental + other rental, checked on all industries; SZZ draft (24 Apr 2020), pp. 18-19: 'Modified EBITD equals interest plus depreciation plus 25% of profits'; SZZ reply (7 Oct 2021) ch. 10.
Why this severity. This affects one illustrative figure for one component of a business-wealth adjustment that the paper itself puts at 0.5-1 pp of the top-0.1% share. The underlying omission is real, and SZZ conceded a related real-estate coding problem.
9minorconfirmed · previously known · misrepresentation of source
The paper says one technical choice by its critics lowers the measured share of the top 1%. The critics' own paper shows that this particular step slightly raises it. The amount involved is tiny, so the bigger picture does not change.
Technical detail and evidence
Why it matters, in detail. The size-ranking step raises AS's top-1% share by about 0.8-1 pp. Against PSZ's unit, AS's unit choice explains at most 0.2 pp of level gaps, so it cannot help explain the AS-PSZ gap in the direction the paper implies.
Evidence. AS (20 Dec 2019) p. 13, fn. 18 and Table 4 ('Groups by individuals/size-adjusted incomes': 0.0, 0.1, 0.2 pp in 1962, 1979, 2014); Splinter (2020) p. 7 and fn. 14; paper pp. 32-33.
Why this severity. The mechanism is misdescribed, but the quantity at stake is about 0.2 pp of the 5.9-pp 2014 gap and has no effect on trends.
10minorconfirmed · previously known · overclaiming
The authors point out that their old method came close to Forbes' total wealth for billionaires, as a sign it worked. In their own data files, that match came from counting about three times too many billionaires, each holding far more in bonds than real billionaires do. The closeness was largely a coincidence. Their new figures simply set billionaire wealth equal to the Forbes-based total, so they cannot confirm it either.
Technical detail and evidence
Why it matters, in detail. The original method's billionaire total is right only on average: 1,668 units at an average of $2.0bn, versus about 540 Forbes billionaires at $4.4bn, with fixed income at 39% of the total. The aggregate match does not validate the dividend/capital-gains weighting at the very top, and the revised series cannot test Forbes because it is calibrated to it.
Evidence. scripts/07 on usdina2016.dta (July 2019 vintage): 1,668 billionaire tax units, $3.27tn, composition hwfix 39.4%, hwequ 57.3%; at or above $50m, $12.12tn (paper: 12.31); workbook DataF15 (2.9804 = 2.9804); paper pp. 22-23, 50; Fig. 15 notes; corpus us-wealth-taxes-11.
Why this severity. This weakens one supporting argument (that the original method is validated by Forbes) without affecting the evidence that SZZ's dividend-weighted method misses non-dividend billionaires. The revised series' matching is disclosed in Section 5.1.
The authors' response. SZ (2022, NBER WP 30396, Table 1) use SEC ownership filings to show the SZZ method values the five largest listed-firm stakes at $37bn versus $263bn actual (e.g., Buffett $0 vs $72.1bn), supporting reliance on Forbes-type values for such founders.
11minorconfirmed · newly identified here · internal inconsistency
Several numbers in the text do not match the authors' own published spreadsheet, and there are some small typos. None of them changes what the paper concludes.
Technical detail and evidence
Why it matters, in detail. Mostly stale or rounded descriptive numbers; the plotted series and released data are internally consistent.
Evidence. Workbook DataF25 B33/C33, DataF27 rows 59 (cols K-O), DataF17; NIPA B047RC/A032RC; paper pp. 1, 3, 9-10, 21, 27, 42-43, 47, 53, Figs. 5, 6, 18; SYZZ (2019) NBER WP 25442, p. 5.
Why this severity. None changes a conclusion; (a) and (b) bear on the claimed consistency with the DFA and on the description of top incomes.
12minorconfirmed · newly identified here · overclaiming
The authors say their revisions barely change their earlier tax-rate estimates for the very rich, but for the most recent year the rate goes up by about 2.4 percentage points. That is modest, not trivial.
Technical detail and evidence
Why it matters, in detail. The revision raises the 2018 top-0.1% rate by 2.4 pp and reduces the long-run decline by 14%; 'trivial' fits most years but not 2018, the headline year of Saez & Zucman (2019b).
Evidence. Workbook DataF28 (scripts/12); paper p. 53, Fig. 28.
Why this severity. Not a core finding of this paper; the qualitative decline in top tax rates since the 1950s remains.
13minorlikely · newly identified here · arithmetic or transcription
Two comparison figures from Federal Reserve survey data seem to mix dollars from different years, which makes the rise in the share of the very top look about 0.4 points bigger than it was. The rival's numbers still fall well short.
Technical detail and evidence
Why it matters, in detail. The corrected SCF+Forbes benchmark is +4.6 pp for the top 0.1% (not 5.0). This slightly narrows the shortfall of the revised SZ series (+2.9) from 42% to 37%.
Evidence. scripts/06 on rscfp2001.dta and rscfp2016.dta (Fed CPI-U-RS factors 2614/4789 and 3548/4789); Forbes 400 totals per CBS News (Sept 2001, citing Forbes) and the Forbes press release (Oct 2016), as cached by auditor A; paper p. 5.
Why this severity. The benchmark rise is overstated by about 0.4 pp (top 0.1%) and 0.2-0.3 pp (top 1%). SZZ's +0.9/+1.4 remain far below it.
14minorplausible · newly identified here · speculative assumption
To total up billionaire wealth, the paper adds an estimate for 'smaller' billionaires that is about three times what Forbes itself lists. Forbes probably misses some, so the true figure is uncertain. The rival method's undercount of billionaires is real but looks somewhat smaller than stated: roughly 30-40% rather than 43%.
Technical detail and evidence
Why it matters, in detail. SZZ capture 57% of the paper's $2.98tn, about 65% of a Forbes-enumeration total of about $2.6tn, and 71% of SZZ's own Pareto $2.4tn. The revised top-0.01% level inherits a calibration uncertainty of up to about 0.5% of household wealth.
Evidence. scripts/05 on Forbes World's Billionaires JSON (2016: 541 US billionaires, $2.40tn; band $1-1.7bn: 155, $196.7bn); scripts/06 (public SCF 2016 at or above $1bn: $0.583tn, 40 records, 504.7 households); SZZ draft (24 Apr 2020) fn. 52; workbook DataF3, DataF15; paper p. 22, fn. 32, pp. 25, 50.
Why this severity. Depending on the benchmark, SZZ's $1.7tn misses 29-43% of billionaire wealth. The paper's 43% is at the top of that range. The direction of C3 is unaffected, and the reconciliation term moves from 1.7 to at most about 1.2 pts.
15minorplausible · newly identified here · internal inconsistency
One chart compares totals built on two differently sized estimates of national wealth. Lining them up shrinks the rival's apparent shortfall a little, but it remains large.
Technical detail and evidence
Why it matters, in detail. The shortfall factor of 1.7 depends on using the SCF's higher aggregate; under proportional alignment it is about 1.5.
Evidence. scripts/06 (SCF 2016 total net worth $86.87tn nominal); workbook DataF3, DataF13 (C7 = 76.476); paper p. 5, Fig. 3, p. 19.
Why this severity. The scaling is crude, and the SCF-FA gap is concentrated in business assets that the paper argues the FA undervalue. The ranking of SZZ below both SZ and SCF+Forbes is unaffected.
16minorplausible · previously known · speculative assumption
How retirement income is split across the income ladder still differs slightly between the two research teams. It moves the top 1%'s share by a few tenths of a point at most.
Technical detail and evidence
Why it matters, in detail. The remaining dispute is about 0.3 pp in the recent top-1% pre-tax share; the 10% weight is calibrated to SCF + defined-benefit pension wealth of the top 1%.
Evidence. Paper pp. 31-32 (fn. 47), 51-52; workbook PSZ(TB2b) W59 = 0.22%; AS (2019) Table 4 retirement row (0.5, 0.1, 1.3); Splinter (2020) p. 6; corpus us-income-11.
Why this severity. At most about 0.3 pp of recent top-1% income shares; the main rollover error was acknowledged and largely corrected.
The authors' response. SZ credit AS for the rollover point ('We are grateful to AS', p. 31) and calibrate the 10% weight to SCF plus defined-benefit pension wealth of the top 1%.
Claims examined and rejected
Problems one auditor raised that did not survive checking.
- Strong form of auditor A's I1: the SCF top-0.1% rates in the authors' workbook 'contradict' the paper, exceeding the 1.4x premium 'in every SCF wave from 2007 to 2019'; with same-year SCF ratios the revised 2001-2016 top-0.1% rise becomes -2.3 pp; rated major. — This holds only for the uniform pass-through correction (cols O-Q of DataF6). The same sheet's second correction variant (cols R-T) gives top-0.1% premia of 0.74, 0.63, 1.01, 2.07 and 0.86 (2007-2019; mean 1.06). The estate data for >$20m (about the top 0.1%) average 1.35x. The -2.3 pp scenario relies on the 2001 same-year SCF ratio (0.58) and is not a credible estimate. The SCF evidence is noisy and mixed rather than contradictory, so the issue is retained at moderate severity as overclaiming (F2).
- The '~$3tn' billionaire benchmark is inflated by about $0.4tn, so the revised top-0.1% share is overstated by about 0.3-0.4 pp. — The arithmetic is right (Forbes' own $1-1.7bn band is $197bn), but an overstatement is not established. Forbes' enumeration implies a local Pareto coefficient below 1, which is implausible and suggests Forbes undercounts lower-tier fortunes, and the public SCF independently gives $583bn. Retained only as a minor benchmark-uncertainty and labelling issue (F10).
- Using the paper's definitions, the S-corporation capital share of value added is 9.9% overall and rises to about 16% for firms with $100m or more of assets, contrary to Fig. 20 and to the claim that it 'does not rise with assets'. — Not established. The authors' DataF20 shows 7.0-9.8% with no rise by size. The discrepancy reflects the auditor's own mapping of ambiguous SOI definitions (portfolio income, officers' compensation, net gains), and no definitional error by SZ was identified.
- Figure 16: the Form 8825 omission alone is only a factor of 3.1, and 48% of the gap is the 25%-vs-100% labor-share convention (raised identically by codex). — This quantification inherits DataF16's mislabelled input: the 'ordinary business income' of $23.9bn is total net income, which already contains rental and portfolio income. With ordinary business income backed out, the omission alone is about 4.2x, and the reclassification is about 31% of the gap (if SZZ's base is ordinary business income). The attribution point survives in weakened form within F8.
- The misdescription of AS's size-adjusted ranking is a moderate issue for the AS rebuttal. — The misdescription is confirmed, but AS's own Table 4 shows the unit/ranking choice accounts for 0.0-0.2 pp of the level gaps with PSZ and does not affect trends, so it is graded minor (F9).
- The large-firm reconciliation double count (25+17+8) is a moderate issue. — The arithmetic error is confirmed, but it affects an illustrative reconciliation that does not feed the revised series, and the conclusion (capital share well above 25%) is unchanged. Graded minor (F7).
- SZ misquote SYZZ: the paper's 52.1% should be 52.9%. — Unverified rather than rejected: 52.9% appears in NBER WP 25442, but SZ cite the QJE pages, and the published QJE text was not available. Listed in F14 as unconfirmed.
Audit notes
What reproduced. Reproduced exactly from the authors' workbook (SaezZucman2020.xlsx): the headline top-1% pre-tax income rise (1978-2014) of +8.47 pp revised vs +9.43 pp original; the top-0.1% wealth share (tax units, 1978-2012) of +12.60 vs +14.94 pp; the revised equal-split top-0.1% share of 15.50% (2001) and 18.42% (2016); the top-1% capital share of 56.1% vs 58.9% (2014) and +10.1 pp over 2000-2018; and the arithmetic of Figs. 13, 15 and 16 and the theory examples (1.010/0.903; 1.040/0.8125; bias factors 0.845 and 0.867). Reproduced independently from primary data: the public SCF top-0.1% rise 2001-2016 (+4.29 pp); the Fig. 3 SCF amounts ($10.23tn; $0.583tn); the DFA top-1% rise (+4.7 to +5.4 pp across 2020 vintages vs 5.2); Johns-Slemrod 27%/5% and 17.8%/17.8%; the IRS S-corp size shares (0.19%, 20.9%); the CCAdj in 2018 (2.06% of NI; 31% of proprietors' income); Forbes 400 and World's Billionaires totals; and the SZZ draft numbers quoted by SZ (14.3%, $1.7T, 640 billionaires, 'half as large'). Discrepant: the SCF+Forbes rises (+4.62 vs 5.0; +6.50 vs 6.7); Fig. 16 (2014 inputs labelled 2016; total net income used as ordinary business income); the CCAdj in 2015 (1.42-1.46% vs 1.6%); the partnership concentration (68-78% vs 'about 90%'); the 25+17+8 reconciliation (44.5% vs 50%); and the text vs workbook composition figures. Not reproducible: the full DINA estimator (no code posted), the SZZ-replication results (2.48% top-0.1% rate; '2.0 points'), the Compustat-based Figs. 7, 8 and 19, the DFA fixed-income total in Fig. 13, the AS workbook figures ($1,135bn; $99bn), and the companion 'Labor and Capital' estimates.
Limits of this audit. The authors' Stata code for the September-2020 revision was not posted, so the effects of individual assumptions (premium, CCAdj key, capital shares, Forbes matching) were assessed with the paper's first-order formulas, workbook data, public micro-files and auditor simulations, not by re-running the DINA program. AS's 2019-2020 Excel files were not obtained, so the $1,135bn aggregate and the $99bn top-1% figure (fn. 42) and the exact before/after construction could not be inspected, and the decomposition of the AS residual relies on Splinter (2020) cross-checked against current-vintage NIPA tables. The 2020 NIPA vintage was not obtained, so whether SZ's '4.3%' includes evaded wages is unresolved. The SCF interest rates were taken from the authors' own output (DataF6), not recomputed from full public SCF files; the construction of its second correction variant (cols R-T) is undocumented, and reading it as concentration-adjusted is an inference. SZZ's confidential information-return evidence, the estate microdata, Compustat and the detailed DFA fixed-income categories were unavailable. The SYZZ QJE text and the NTJ version of Johns & Slemrod were not checked. Forbes March world lists and the September Forbes 400 differ in timing and coverage. The public DINA micro-files used for the partnership-concentration and portfolio checks are the authors' public versions and may differ from internal files. Conventions in post-tax income (government consumption, deficits, incidence) and the Social Security wealth concept were examined against the critique corpus but not scored, as they are conventions rather than errors and do not drive C1-C6.