← All papers

Gabriel Zucman · World Inequality Lab Technical Note 2020/07, November 2020 (last update September 2020)

US Distributional National Accounts: Updates

19 problems found: 8 moderate · 11 minor

What the paper claims

Every problem found

1moderateconfirmed · newly identified here · internal inconsistency

The note says its revised numbers now match the Federal Reserve's own wealth statistics almost exactly. The authors' own spreadsheet shows they still differ: in 2018 the richest 1% hold about 25% of their wealth in bonds and deposits in the authors' series versus 22% in the Fed's, and the two estimates of the top 1%'s share of all wealth differ by about 2 percentage points in most years before 2017. The series are close, but 'almost identical' and 'fully consistent' overstate the match.

Technical detail and evidence

Why it matters, in detail. On the authors' own harmonised basis (SZ 2020 Fig. 26: wealth excluding durables and unfunded DB pensions, FA totals by asset class), the SZ-DFA top-1% wealth-share gap averages +2.0 pp in 2010-16 and ~0 only in the projected years 2017-19 (see F5). With 2018 levels matched but a fixed-income portfolio share 3.1 pp higher (25.1% vs 22.0%), the SZ top 1% still holds roughly $1T (~1.1% of household wealth) more fixed income than in the DFA: the benchmark the note invokes points toward a larger, not smaller, top-end interest premium (F2). 'Reconciles' overstates a housing gap that was narrowed (9.2% -> 12.0% vs 14.4%, public-file estimate) rather than closed; the pension statement cannot be verified with public DFA categories.

Evidence. Note Sec. 1.1.1 p.2, Sec. 1.1.6 p.5, Sec. 1.1.7 p.6. SaezZucman2020.xlsx DataF25 row 33 (2018): DFA 21.96%, SZ (hwfix - miscw)/hweal = 25.12%; 1995: 29.51% vs 29.16%; mean |gap| 1989-2019 1.76 pp, >=2 pp in 13 of 31 years, +2.28 pp mean 2013-19 (outputs/dfa_compare.txt). DataF26: SZ-DFA top-1% share gap +1.21 to +3.19 pp in 2010-16 (mean +2.03), +0.07 pp in 2017-19; 1989-2019 rise SZ +6.8 pp vs DFA +9.4 pp. Public 2018 micro-file: top-1% fixed-income share 25.08% (outputs/micro_sensitivity_2018.txt). SZ (2020, NBER w27921) p.3 itself says 25% for 2018 while p.46 (copied into the note) says 22%-23%. Housing (Claude auditor, public 2018 file): top-1% share of owner-occupied housing 12.0% after vs 9.2% before the revision, DFA real estate (owner-occupied incl. vacant land) 14.4%. DFA 'pension entitlements' exclude IRAs while SZ pension wealth includes IRAs and life insurance (DFA definitions file).

Why this severity. The note's main external-validation claim is overstated and one of its two illustrative numbers is contradicted by the authors' own data; the direction (the revision moved the series much closer to the DFA) holds, and the DFA shows an even larger 1989-2019 rise, so the long-run conclusion is not threatened.

The authors' response. SZ (2020, w27921) report 25% for the top-1% interest-bearing share in 2018 in their introduction (p.3) and disclose the DFA harmonisation in the Fig. 26 note (adding that converting the DFA to equal-split adults would raise the DFA share slightly). The 22%-23% statement was not corrected in the February 2022 log.

2moderateconfirmed · previously known · version instability

The note lists each change but never says what all the changes together did to the main results. Comparing the old and new data shows the revisions reduced the measured increase in the top 0.1%'s share of wealth since 1989 by about a quarter, and since 2001 by more than a quarter, while leaving the longer-run increase since the late 1970s largely intact. Readers are not told this, and no margin of error is given.

Technical detail and evidence

Why it matters, in detail. The revision raised 1980s levels (property-tax phase-in, Forbes from 1982, business-wealth split; F4) and lowered post-2008 levels (interest premium; F2), shifting the measured rise of concentration earlier: for the top 0.1% the 2001-16 rise falls from +4.05 to +2.92 pp and the 1989-2016 rise from +8.13 to +6.35 pp, while 1978-2016 falls only from +12.09 to +11.34 pp. The SZZ (2023) characterisation of 'SZ20' (+7.7/+6.3/+4.1 pp for top 1/0.1/0.01%, 1989-2016) reproduces exactly from these tables. No decomposition by component or uncertainty band is provided, the pattern Kopczuk (2019) criticised.

Evidence. Table II-E1/II-B1 in the Nov 2017, Oct 2019, Sept 2020 and Feb 2022 archives (outputs/vintages.txt). Oct 2019 -> Sept 2020: top-0.1% wealth rise 1989-2016 +8.13 -> +6.35 pp (-22%), 1998-2016 +5.17 -> +3.66 (-29%), 2001-2016 +4.05 -> +2.92 (-28%); top-1% 1989-2016 +10.34 -> +7.73 (-25%); top-1% level 2010 37.80 -> 34.21; top-0.01% 1989 3.98 -> 5.16; 1978-2016 top-1% rise +15.12 -> +13.93 (-8%), top 0.1% +12.09 -> +11.34 (-6%). SZ (2020) Fig. 1 data (tax units) vs SZ (2016) QJE: top-0.1% rise 1989-2012 +10.5 -> +7.1 pp, 2001-2012 +6.3 -> +3.6 pp. Pre-tax top 1% 1980-2014: +9.53 (PSZ 2018) -> +8.53 pp. SZ (2020) pp.2-3 ('track each other closely').

Why this severity. The revisions do not overturn the long-run rise (-6% to -8% of the 1978-2016 increase) but reduce the post-1989/post-2001 rises of top wealth shares by 22%-42% depending on window and baseline vintage, which a reader of the note cannot learn from it.

The authors' response. SZ (2020, Figs. 1-2) plot original and revised series and describe them as tracking 'closely'; per the critique corpus, SZ (2022, NBER w30396) acknowledge that 'real uncertainty remains about the top 0.01%'.

3moderateconfirmed · previously known · overclaiming

In October 2019 the authors changed how they assign corporate taxes, putting all of them on owners of corporate shares; the note calls this 'the only meaningful way', although official agencies and many economists split the tax differently. This choice, together with the other October 2019 changes, more than doubled the measured fall in the tax rate of the richest 0.1% between 1950 and 2014 (from about 9 to about 21-23 percentage points). The note does not say how big the effect is, and the companion paper's chart comparing old and new tax rates uses the wrong old series, which hides it.

Technical detail and evidence

Why it matters, in detail. The 1950-2014 decline in the top-0.1% average tax rate goes from -9.3 pp (PSZ 2018) to -23.4 pp (Oct 2019) and -20.9 pp (Sept 2020); the corporate component alone goes from -7.0 to -24.5 pp, so moving incidence from all capital to C-corp equity only is the dominant driver (adding pure capital gains to denominators lowers recent top rates further, e.g. top 0.1% in 2000: 46.8% -> 37.9%). For the top 0.01% the decline goes from -21.6 to -37.8/-39.2 pp. The only published quantitative comparison (SZ 2020, Fig. 28) uses the wrong PSZ 2018 series (62.4% in 1950 instead of 49.1%), so it shows the revision lowering the 1950 top-0.1% rate when it raised it by about 7 pp.

Evidence. Note Sec. 2.1.2 p.8. taxrates sheets (outputs/taxrate_vintages.txt): top-0.1% average tax rate 1950: 49.1% (Nov 2017 = PSZ 2018) -> 56.4% (Oct 2019) -> 55.7% (Sept 2020); 2014: 39.8% -> 33.0% -> 34.8%; corporate component 1950: 19.5% -> 34.6% -> 31.6%; 2014: 12.5% -> 10.1% -> 9.5%. Top 0.01%: 1950 62.4% -> 68.5% -> 73.9%; 2014 40.8% -> 30.7% -> 34.7%. PSZ2018MainData.xlsx 'Data' col. 'Avg rate top 0.1%' = 0.491 (1950). SaezZucman2020.xlsx DataF28 'PSZ 2018' column (titled 'Top 0.1% effective tax rate') equals the Nov 2017 top-0.01% series to 6 decimals (max |diff| 0.000000; vs top-0.1% max |diff| 0.133) (outputs/fig28_check.txt); SZ (2020) p.53: the series 'differ somewhat from the original PSZ (2018) series'. Table I-S.A12 (Claude auditor): 1950 corporate tax $17.7bn vs federal individual income tax $17.4bn.

Why this severity. The convention is disclosed and its direction acknowledged, so the note's documentation is not wrong; but it presents a contested convention as uniquely valid and leaves unquantified that it (with the other Oct 2019 changes) more than doubles the measured decline of top tax rates. For downstream tax-progressivity claims, which this note does not itself make, the choice is decisive for magnitude.

The authors' response. The note argues that existing corporate taxes are paid by shareholders and that wage effects of rate changes belong to reform analysis (citing Saez & Zucman 2019b); per the critique corpus (us-wealth-taxes-13) the argument is expanded in Saez & Zucman, 'Distributional Tax Analysis in Theory and Practice: Harberger Meets Diamond-Mirrlees' (NBER w31912, 2023).

4moderateconfirmed · newly identified here · internal inconsistency

For years before 1962 the authors estimate tax rates for each income group separately, and the pieces do not add up: the groups together pay up to about 5% more or less tax than the national total, and in the 1930s the rate shown for the top 0.1% is about 7 points higher than its two sub-groups imply because a cap applied to the very top was not carried through. The 1950s figures are only slightly affected.

Technical detail and evidence

Why it matters, in detail. Applying different affine transformations of raw shares to overlapping and disjoint groups cannot preserve partition identities unless the 1962 DINA and raw shares coincide; the residual equals taxall x (sum of disjoint fractions - 1). Signs are systematic (positive 1918-43, negative 1944-61). The 1930s cap repairs only the top-0.01% cell: the published top-0.1% (and higher aggregates) still reflect an implausible ~94% top-0.01% rate in 1932, so the 1930s top-end rates are overstated by up to ~7 pp at the top 0.1%. The qualitative mid-century pattern of high top rates is unaffected.

Evidence. taxrates.do ll.127-133 and 147-157 (frac = 1 - (1 - raw)(1 - m1962)/(1 - raw1962) per group), ll.170-172 (1.3 cap; PSZ2017Programs/taxrates.do l.155 used 1.2). Table II 'taxrates' and 'avgpeinc' (Sept 2020), income weights = population share x group average / overall average (outputs/tax_partition.txt): bottom 90% + top 10% imply taxes of 15.95% vs 15.22% of NI in 1932 (+0.73 pp, 4.8% of total taxes); |residual| > 0.01 pp in 48 of 49 years 1913-61 (mean 1.9% of taxes); -0.2 to -0.44 pp of NI in 1944-61; top 10% 19.55% vs 18.22% implied by P90-95, P95-99 and top 1% (1932). Nesting (outputs/tax_nesting.txt): 1932 top 0.1% published 51.46% vs 44.62% implied by P99.9-99.99 (36.95%) and the capped top 0.01% (66.90%), i.e., the published top-0.1% rate embeds an uncapped top-0.01% rate of about 94%; 1940: 41.98% vs 39.56%. Post-1962 identities hold (|residual| <= 0.06 pp except interpolated 1963/1965).

Why this severity. Confined to pre-1962 secondary series; the 1950s-60s discrepancies are small (about 1-2% of total taxes, top-0.1% nesting within 0.2 pp), but the 1930s top-group rates are materially inconsistent (about 7 pp at the top 0.1% in 1932).

5moderatelikely · previously known · speculative assumption

To turn reported interest income into an estimate of bonds and savings, the authors must assume what interest rate the rich earn. The note picks one number (1.4 times the average rate for the whole top 1%) and describes it as a small, well-established premium, but the evidence is mixed and the authors' own survey data suggest the very richest earn even higher rates. Because this single assumption moves the estimated wealth share of the top 0.1% by about a percentage point, how fast their share grew after 2001 is quite uncertain, although the long-run rise is not.

Technical detail and evidence

Why it matters, in detail. Because the premium switches on in 2003/2008, it scales the post-2001 top-0.1% trend almost one-for-one: the Sept 2020 rise of +2.92 pp (2001-16) would be about +4.2 pp with homogeneous returns and about +2.0 (+1.6) pp with r/rm = 2.0 (2.5), using 2018 effects as a proxy. Applying 1.4 to the entire top 1% is supported by SCF top-1% ratios but not by $10-20m estates (0.93), while the SCF top-0.1% ratios (1.5-2.6 in 2010-19) point to a higher premium at the very top, the direction argued by SZZ (2023). The 1.15 used for 2003-07 has no support in the cited estate (1.03) or SCF (0.99 in 2004) evidence and lowers 2003-07 top shares by perhaps 0.5-1 pp. The dispute (rank by wealth vs by interest income) is unresolved.

Evidence. build_usdina.do ll.697-757 (premium 1.15/1.4, top 1% by wealth, one iteration). SaezZucman2020.xlsx DataF4: estates $20m+ r/rm 2008-16 mean 1.351, median 1.287, ratio of means 1.326; 2003-07 mean 1.034 (1.30, 1.02, 1.10, 0.90, 0.85); $10-20m estates 2008-16 mean 0.926. DataF6 (SCF): top-1% r/rm 0.99 (2004), 1.59 (2007), 1.28-1.42 (2010-19); top-0.1% 1.49 (2010), 1.48 (2013), 2.56 (2016), 1.91 (2019); the sheet's own note says the ratio is 'about 1 from 1989 to 2004, and around 1.4 from 2007-2016' and that top-0.1% SCF interest is 'very noisy' (outputs/premium_evidence.txt). Public 2018 micro-file: replacing hwfix by hwfix_homo raises top 1/0.1/0.01% shares by +1.78/+1.29/+0.72 pp (+1.67/+1.10/+0.48 with billionaire target re-imposed) (outputs/micro_sensitivity_2018.txt); Claude auditor's rescaling to r/rm = 2.0 lowers them by about 1.50/0.94/0.52 pp. SZZ (2020 WP, pp.35-37; 2023 QJE Table I): top 1%/0.1% 33.7%/15.7% in 2016 vs 36.3%/18.4% for the Sept 2020 series (36.27/18.42 in Table E1).

Why this severity. The parameter is disclosed and moves the series toward the critics, and the long-run rise is unaffected; but within the range supported by the authors' own SCF evidence it shifts the 2018 top-0.1% share by about 1 pp and changes the 2001-16 top-0.1% rise by a third or more.

The authors' response. SZ (2020, NBER w27921, Sec. 2) argue the correct capitalization rate is the rate of the wealthy ranked by wealth, not by interest income, and that observed yields of wealthy entities (estates, SCF, foundations, listed firms, S-corp holding companies) were about 1.0-1.7% in 2016; per the critique corpus, SZ (2022, NBER w30396) argue that the remaining gap is concentrated in the top 0.01% and that SZZ underestimate billionaire equity. Unresolved.

6moderatelikely · speculative assumption

Because the method misses billionaires whose shares pay no dividends, the authors force the total wealth of all 'billionaires' in their data to match an outside target. After 2006 that target was a statistical extrapolation about 25-30% larger than Forbes's own count of American billionaire wealth, and in 2010 the same rule produced an impossible result: the richest ~2,000 adults were assigned less wealth than the Forbes 400 alone. The authors later dropped this step. It matters for the very richest 0.01%, not for the broader top 1%.

Technical detail and evidence

Why it matters, in detail. Re-imposing Forbes's own March-2018 US billionaire total ($3.10T) instead of the coded $3.918T, with the difference reallocated pro rata to other units' non-S-corp equity as the code does, lowers 2018 top 1/0.1/0.01/0.001% wealth shares by -0.40/-0.65/-0.83/-0.93 pp. This is an upper bound: the authors' Feb 2022 method (match the Forbes 400 only) gives the same 2018 top-0.001% share (4.75%), so much of the excess is allocated near the top by capitalization anyway. The threshold-based group definition produces the 2010 anomaly (top 0.001% = 0.86x the Forbes 400 vs 1.25-1.50x in every other year 1982-2019), a ~1.6 pp dip at the top 0.001% that the Feb 2022 vintage removes.

Evidence. parameters.xlsx (ParametersStata, col. tomatch): $2,958bn (2016), 3,598 (2017), 3,918 (2018), 3,968 (2019), 4,365 (2020 = 2019 x 1.1); 2006 and 2010 targets equal the Forbes 400 totals ($1,260bn, $1,370bn). build_usdina.do ll.968-1008. Forbes World's Billionaires (March lists, US citizens; Claude auditor's capture, counts consistent with Forbes's published totals): $2.40T, 2.76T, 3.10T, 3.11T, 2.95T for 2016-20 (ratios 1.23, 1.31, 1.26, 1.28, 1.48). Table II-E1 (Sept 2020): top-0.001% share 4.22% (2008), 3.19% (2009), 2.24% (2010, = $1,162bn of $51,972bn), 3.88% (2011), vs Forbes 400 $1.37T in 2010; Feb 2022 vintage: 3.87% in 2010. Public 2018 file: tax units with wealth >= $1bn (dweght) = 1,269 holding exactly $3,918bn (outputs/billionaire_target_cf_2018.txt). SZ (2020) fn. 32; Feb 2022 log Sec. 1.1.

Why this severity. Affects levels and year-to-year dynamics of the top 0.01%-0.001% (secondary series) and the interpretive claim in footnote 6; the effect on the top 1% is a few tenths of a point, and the authors abandoned the method in the next update.

The authors' response. SZ (2020, fn. 32) justify about $0.6T of non-Forbes-400 billionaire wealth in 2016 by a Pareto extrapolation ($567bn) and by the public 2016 SCF ($583bn). The Feb 2022 log (Sec. 1.1) replaced the procedure with matching the Forbes 400 only, noting there is 'no a priori reason' to prefer Pareto interpolation over the capitalization method.

7moderatelikely · newly identified here · speculative assumption

The authors raise the housing wealth of the richest 1% because, in survey data, they pay lower property-tax rates. That survey evidence exists only from 1989, yet the code silently assumes the gap did not exist before 1979 and grew through the 1980s, citing California's Proposition 13. This unstated assumption alone adds roughly one percentage point to the measured rise in the top 1%'s wealth share during the 1980s.

Technical detail and evidence

Why it matters, in detail. The full 0.75 discount adds +0.87 pp to the 2018 top-1% share; by construction the effect is zero in 1978 and complete from 1989, so on a 2018-sized proxy about 0.9 pp of the 1978-89 top-1% increase comes from the phase-in. Together with the 1982 Forbes switch-on (about +0.4 pp on the top 0.01% from 1982), the Sept 2020 revisions raised the 1978-89 rise by +1.4 pp (top 1%), +1.0 (top 0.1%) and +0.9 (top 0.01%) relative to Oct 2019. Applying the 1989 differential to all years, or any differential that pre-dated Prop 13, would flatten the measured 1980s rise; the California reform is an arbitrary anchor for a national adjustment.

Evidence. build_usdina.do ll.839-846 (discount schedule) and ll.921-966 (Forbes 400 matching 1982-2005). Note Sec. 1.1.4 p.5 ('We make no correction before 1982'), Sec. 1.1.6 p.5. Public 2018 file: homogeneous property-tax rate lowers top 1/0.1% shares by 0.87/0.25 pp (outputs/micro_sensitivity_2018.txt). Table II-E1 Sept 2020 minus Oct 2019 vintage: top-1% gap +0.74 to +1.02 pp in 1975-82 vs +1.5 to +2.4 pp in 1983-91; top-0.001% share 0.98% (1981) -> 1.42% (1982); top-0.01% gap +0.28 (1981) -> +0.69 (1982) -> +1.06 (1983) (outputs/break_1980s.txt). 1978-89 rise: top 1% +4.77 -> +6.20 pp, top 0.1% +3.96 -> +4.99, top 0.01% +1.94 -> +2.80 (outputs/vintages.txt).

Why this severity. Roughly 0.9 pp of the revised 1978-89 rise in the top-1% wealth share (about 14% of +6.2 pp) is assumed rather than measured and the assumption is not disclosed in the note; the long-run rise is unaffected.

8moderatelikely · newly identified here · internal inconsistency

The last three years of the series are partly projected rather than based on actual tax data. At the switch from real data to projections (2016 to 2017), several published shares jump by unusual amounts: the bottom half's share of taxable income falls by a quarter in one year, and the top 10%'s wealth share drops during a stock-market boom. These jumps are probably artifacts of the projection method, and the close match with Federal Reserve data appears only in these projected years.

Technical detail and evidence

Why it matters, in detail. The re-weighting of filers to 19 AGI bins (multipliers capped at 2) while keeping non-filers from the older projection plausibly explains a one-year 25% fall in the bottom-50% fiscal share and fiscal top-share jumps about twice those in the official Piketty-Saez series. In wealth, the 2017 fall of the top 10% (-1.58 pp, the largest annual move of the decade, in a year of strong equity gains) runs against the DFA. Part of the post-2016 decline in SZ top wealth shares, and hence of the 2017-19 convergence with the DFA used in F1, may therefore be methodological rather than economic.

Evidence. Table II-D1 (tax units): fiscal bottom-50% share 6.70% (2016) -> 5.04% (2017), vs |change| <= 0.35 pp in every other year 2010-19; DINA fiscal top 10%/top 1% +2.10/+1.90 pp in 2017 vs +1.04/+1.26 pp for the official Piketty-Saez series in the same table's memo columns. Table II-B1: pre-tax bottom-50% share +0.51 pp in 2017 (largest move 2010-19). Table II-E1: wealth top 10% 73.37 -> 71.79 and top 1% 36.27 -> 35.56 in 2017, while the harmonised DFA top-1% share rose 0.34 pp (DataF26), so the SZ-DFA gap falls from +1.30 to +0.24 pp. 2019 fiscal distribution identical to 2018 by construction. Feb 2022: fiscal bottom 50% 6.76 -> 4.87 (outputs/seam.txt). projection.do ll.224ff.

Why this severity. The years are flagged as provisional, but the jumps (0.5-1.7 pp in several published series) are probably artifacts, they persisted into the next vintage, and they coincide with the years that carry the note's claimed agreement with the DFA.

The authors' response. The note states that distributional series for 2017-2019 are provisional and subject to revision (Sec. 1, p.1).

9minorconfirmed · newly identified here · arithmetic or transcription

The computer code that splits big business owners' profits into 'work' and 'capital' parts does not follow the rule the authors describe in print: above about $4 million, it counts every extra dollar as capital, whereas the published rule counts 75%. Fixing it lowers the estimated wealth share of the top 0.1% by about 0.3-0.4 percentage points in 2018, a real but small error.

Technical detail and evidence

Why it matters, in detail. Relative to the documented rule, the code over-weights large active pass-through owners in the partnership wealth key (and hence partnership capital income, allocated by partw) and in the S-corp wealth key (S-corp income is allocated by total S-corp profit and is unaffected); the aggregate capital keys fall by 6.0% (partnerships) and 9.3% (S-corps) when corrected. Because the public file assigns zero passive share to every large unit, these 2018 magnitudes are an upper range; with positive passive shares the active amount, and the error 0.25(A - T), is smaller. Effects in earlier years should be smaller still, as less active income exceeds T.

Evidence. SZ (2020, w27921) p.41. build_usdina.do ll.362-374 (partnerships) and 411-423 (S-corps): partpinc_k = 0.75 s P + 0.25 A + 0.75 (A - T) for A >= T; pre-1987 0.35 A + 0.75 (A - T) exceeds A when A > 7.5T. Public 2018 file: the capital keys reconstruct exactly from partw/flmil_p and scorw/fkequ_s (relative residual ~1e-16); T = $4,270,775 (wage P99.99 = $4,270,273); all 106 large partnership records have labour part exactly $3,203,081 = 0.75T; 106 partnership and 169 S-corp records (about 15,400 tax units) are above T, all with zero passive share in the public file (outputs/recon_keys_2018.txt, recon_T_2018.txt). Documented rule: top 1/0.1/0.01% wealth -0.23/-0.42/-0.44 pp (-0.14/-0.28/-0.26 with billionaire target re-imposed); pre-tax top 1/0.1% -0.03/-0.06 pp; capital share of top-0.01% pre-tax income 83.0% -> 81.7% (outputs/bus_formula_cf.txt).

Why this severity. A genuine coding error that contradicts the stated method, but its measured effect on headline shares is small (about 0.3 pp at the top 0.1% after the billionaire matching, less than 0.1 pp on income shares) and likely smaller in the internal data; it does not affect conclusions.

10minorconfirmed · newly identified here · internal inconsistency

The note says half of partnership profits are counted as capital income, but the spreadsheets actually raise that fraction step by step from 40% in 1986 to 55.5% in 2017. This does not change how much income the top groups get, only how much of it is labelled capital, and it makes capital look more important in recent decades than a flat rule would.

Technical detail and evidence

Why it matters, in detail. Relative to the original constant 30%, the ramp reclassifies about $118bn of 2018 mixed income from labour to capital (and about $16bn the other way in 1990); relative to the 1986 share, about $109bn (0.6% of NI) of 2019 partnership income. This mechanically raises the capital share of top incomes in recent decades, relevant to SZ (2020, p.53)'s statement that all the rise of the top-1% share since the late 1990s is due to capital income; top income shares are essentially unchanged because total partnership income allocated is fixed.

Evidence. Table I-S.A3b col. AB: 0.40 (<=1986), '=AB(previous)+0.005' (1987-2017), 0.555 (2017-20); col. AC: 0.274 (1980), 0.257 (1990), 0.339 (2010), 0.354 (2016), 0.377 (2018); the r x W-imputed partnership share in the same sheet varies between 0.19 and 0.61 with no smooth trend before about 2004. SZ (2020) p.40 discloses '50% or more in the 2010s, with a rising trend since the 1980s'.

Why this severity. A documentation inaccuracy that affects the labour/capital composition of top incomes, not their level; the rising trend is disclosed in the companion paper.

The authors' response. SZ (2020, p.40): returns of listed firms imply 'a capital share of partnership income of 50% or more in the 2010s, with a rising trend since the 1980s'.

11minorconfirmed · previously known · documentation gap

The note says pension income is spread across people using the same rule as pension savings; the code actually uses a different rule, and the example it gives for 'non-taxable pension payouts' (Roth IRA withdrawals) is a tiny part of that category, which is mostly money moved from one retirement account to another. These are description errors with small effects.

Technical detail and evidence

Why it matters, in detail. The 1/7 (benefits) and 10% (wealth) residual weights still load on rollovers, which are stocks concentrated at the top (79% of pension distributions reported by AGI >= $1m were non-taxable in 2014, Auten & Splinter 2019 WP fn. 31); Auten & Splinter (2026, per corpus) put the remaining effect at about 0.3 pp of the top-1% pre-tax share. Dropping the wage component for benefits while keeping it for wealth is a defensible but undisclosed asymmetry.

Evidence. build_usdina.do ll.1733-1743 (weight = 1/6 x sum peninc / sum penincnt; pre-1984 equal weights), ll.511-550 (wealth: 60/30/10 from 1984, 60/40 before; wagetop50 = wage - median for the top half). Splinter (2020), 'Reply: Trends in US Income and Wealth Inequality: Revising After the Revisionists' (davidsplinter.com), p.6: 2014 Roth distributions 'appear to be less than 0.5% of non-taxable amounts and the rest is nearly all highly concentrated rollovers'. Total Oct 2019 -> Sept 2020 revision of the 2016 top-1% pre-tax share: 19.26% -> 18.61% (-0.65 pp).

Why this severity. Description errors in a revision that genuinely reduces the weight on rollovers; the residual effect of the remaining weight is estimated by critics at about 0.3 pp of the top-1% share or less.

The authors' response. SZ (2020, p.31) accept that PSZ (2018) put too much weight on non-taxable distributions and credit Auten and Splinter.

12minorconfirmed · previously known · documentation gap

To estimate taxes of the 400 highest earners, the authors assume their taxable income is at most 45% of their true income. The code applies this more broadly than the note says, reshuffling incomes even in years when the cap does not bind and spreading the removed taxes over everyone else. Later administrative data (42% in 2018-20) support the 45% figure for recent years, so the practical effect is probably small.

Technical detail and evidence

Why it matters, in detail. Because the cap binds mainly in recent decades, any error in the constant 45% would bias the trend of the top-400 tax rate; the record-level resets in non-binding years redistribute fiscal income within the group and can change group taxes; the proportional reallocation raises everyone else's income taxes by the removed amount. None can be quantified from public outputs (top400.xlsx is not released, and the corrected variables are dropped from the online files).

Evidence. top400.do ll.1-7 ('Assume constant over time'), 44-48, 59-67, 69-92. Note Sec. 2.1.3 pp.8-9. Codex auditor's synthetic two-unit example (group ratio 40%: taxes 32 -> 24 before rescaling) and public-file translation (ratio 45.3% after national rescaling). Balkir, Saez, Yagan & Zucman (2025, NBER w34170) Table 1: fiscal/economic income 42% for the Forbes 400 in 2018-20.

Why this severity. The recent level of the cap was later validated with administrative data, the cap rarely binds in early decades, and the record-level effects are likely small; affects only top-400 tax-rate series, not income or wealth shares.

The authors' response. The note (Sec. 2.1.3) calls the 45% 'the best approximation we could create using public sources' and invites refinement; Balkir, Saez, Yagan & Zucman (2025, NBER w34170) later find 42% for 2018-20.

13minorconfirmed · newly identified here · documentation gap

The way the most recent years were estimated changed in August 2020, but the note still describes the older method. This matters for people trying to understand which numbers are projections, not for the results.

Technical detail and evidence

Why it matters, in detail. Users of the note cannot tell that 2017-18 are AGI-bin re-weightings of a 2012-based file and 2019 a uniform extrapolation of 2018 (identical fiscal distribution); the seam artifacts in F5 arise under the undocumented method.

Evidence. projection.do l.11 ('$old=0') and block from l.224 (19 AGI bins, multipliers capped at 2); runusdina.do l.190 ('August 2020 modified projection to use SOI published tabs for 2017 and 2018'); 2019 = 2018 x 1.045 (nominal) / 1.009 (population). FOMC SEP 26 Sept 2018 medians: real GDP 3.1/2.5, PCE inflation 2.1/2.0 (= 5.2/4.5, Q4/Q4; Claude auditor). TA0 adult-population growth 0.73%/0.68% in 2018-19 (Claude auditor).

Why this severity. A vintage-tracking gap; the implemented method uses more 2017-18 tax information than described, and no level effect is documented.

14minorconfirmed · newly identified here · data error

In the authors' spreadsheet of tax rates by group, the breakdown by type of tax for the very richest groups (including the top 400) leaves out business property taxes, so the pieces add up to less than the total. The totals are right; only someone splitting the richest groups' taxes by type would be misled.

Technical detail and evidence

Why it matters, in detail. For the top 400 in 2018 the displayed corporate component is understated by 6.4 pp of pre-tax income (about 23% of their 27.2% total rate), so any decomposition of top-400 taxes by type from this export (e.g., the corporate-tax share) is biased; whether any published figure used these columns was not checked.

Evidence. taxrates.do ll.177-180 (group list omits top0p001, P99p99_P99p999, top400); outsheet_dina.do builds totals including propbustax. Table II 'taxrates', 2018: total minus sum of six components = 6.10 pp (top 0.001%), 3.75 pp (P99.99-99.999), 6.36 pp (top 400); 0.000 for all other groups; 2.4-7.6 pp across years; Oct 2019 2.4-5.3 pp; Feb 2022 3.0-9.0 pp (outputs/tax_partition.txt, comp_gap_vintages.txt).

Why this severity. Total tax rates are correct and the affected groups do not appear in the composition tables (TG2-TG2c); only users decomposing the very top's taxes by type from the raw export are misled.

15minorconfirmed · newly identified here · speculative assumption

The note assumes a common type of small corporation keeps none of its profits, and even suggests it keeps less than nothing; tax data show it keeps a noticeable share. When the authors later corrected this, the top 1%'s income share rose by about 0.3 percentage points.

Technical detail and evidence

Why it matters, in detail. About 19% of 2014 S-corp total net income was retained (about 3% if only trade-or-business income is compared with distributions), so the zero-retention assumption understates S-corp income, which is more concentrated than C-corp income; the BEA prototype profits imply a larger gap.

Evidence. Note Sec. 1.2.2 p.7, fn. 8. IRS SOI Form 1120S Table 7, tax year 2014: total net income (less deficit) $506.8bn ($420.8bn from a trade or business) (irs-14co07ccr). BEA FAQ 318 table (from SOI): S-corp dividends $409.0bn (2014), $471.0bn (2015). Feb 2022 log Sec. 1.2: S-corp profit $747bn in 2017; +0.3 pp top-1% share.

Why this severity. Explicitly provisional and corrected in the next update; the authors' own later estimate of the effect is about 0.3 pp of the top-1% pre-tax share.

The authors' response. Footnote 8 flags that the assumption would be revisited once BEA decomposes corporate profits; the Feb 2022 update adopted BEA's prototype S-corporation profits.

16minorconfirmed · newly identified here · methodology

When computing tax rates, the authors add part of capital gains to income, but before 1987, when only part of gains was taxable, their rule almost never adds anything, even in years with large gains. This slightly overstates top tax rates in a few pre-1987 years.

Technical detail and evidence

Why it matters, in detail. Pure gains are concentrated in the top 0.1%, so adding 1-1.7% of NI to denominators in 1968 and 1983-85 would lower top-0.1% rates in those years by several points, slightly steepening the measured post-1980s decline relative to a consistent treatment.

Evidence. outsheet_dina.do ll.708-721 (kg_agi = 0.5 x gains before 1978, 0.4 x in 1978-86; pure_kg = max(0, frac - 0.03)/frac x kg_agi; 1986: 50%). Claude auditor (kg_threshold_asymmetry): 1962-86 sum of pure gains 1.74 (rule used) vs 12.28 (total-gains rule) NI-point-years; 1989 taxable gains 3.08% of NI.

Why this severity. Affects effective-tax-rate denominators in a handful of pre-1987 years; the long-run pattern of top tax rates is unaffected.

17minorconfirmed · newly identified here · documentation gap

A number of small factual slips (a mislabelled data release, outdated figures carried over from the 2019 note, ambiguous citations). None affects the results.

Technical detail and evidence

Why it matters, in detail. None affects any series; the retained October 2019 figures understate 2018 consumption taxes by about $40bn and 2018 taxable gains by about $67bn relative to the September 2020 tables.

Evidence. Note pp.1, 5, 6, 8, 9 and fn. 2. BEA, 'The 2020 Annual Update of the National Income and Product Accounts', SCB 100(8), August 2020 ('released its annual update ... on July 30'). 2018 realized gains $966.6bn = 5.46% of NI (Table I-TC1, adjudicator); consumption taxes $889.8bn and the 1982 share and upgrade-ratio figures from Tables I-TA0/I-S.B7 (Claude auditor, not re-run).

Why this severity. Presentation and documentation slips with no bearing on conclusions.

18minorconfirmed · newly identified here · documentation gap

The public data files the authors release do not exactly reproduce their published tables (for example, the top 10%'s wealth share is about 2 points lower in the public 2018 file), and the confidential inputs needed to regenerate the tables are not available. The authors warn about this, and the differences for the main top-1% figures are small.

Technical detail and evidence

Why it matters, in detail. Independent replication is limited to public synthetic files whose top tail differs from the internal series (top-10% wealth -2.2 pp, top-1% -0.65 pp, pre-tax top-1% -0.27 pp in 2018) and whose top record exceeds 400 units, which breaks top-400 analyses (fixed in Feb 2022); component effects stated in the note (pensions, equity, housing) cannot be isolated by outsiders.

Evidence. outputs/micro_vs_published.txt and micro_peek_2018.txt (adjudicator, equal-split = tax-unit totals divided by adults): wealth top 10/1/0.1% 69.42/34.71/18.11% vs 71.60/35.36/18.16% published; pre-tax 44.91/18.66/8.56% vs 45.55/18.93/8.61%; id 0 record weight 51,850,017/1e5 = 518.5 tax units. Codebook (Sept 2020) p.1: public files 'do not exactly reproduce' internal results but 'track ... closely' for most series. Feb 2022 log Sec. 1.1 (aggregate record split).

Why this severity. Disclosed in the codebook and modest for the main top-1% figures; it limits external verification rather than changing results.

The authors' response. The codebook warns that external-use files do not exactly reproduce internal results; the Feb 2022 update splits the aggregate top record so that the top 400 match the Forbes 400.

19minorplausible · previously known · methodology

Sales taxes are assigned to households partly according to what they buy with government benefits, but the tax rate is then computed on income that excludes most of those benefits, which makes poorer households' rates look higher. The 70/30 split between consumers and producers is not sourced. These are known conventions with no effect on top income shares.

Technical detail and evidence

Why it matters, in detail. Consumption financed by transfers (e.g., SSI, refundable credits, in-kind benefits) generates sales taxes that are divided by a pre-transfer income base, inflating bottom-50% tax rates relative to income concepts that include transfers (Splinter 2020, NTJ); allocating 30% by factor income raises top rates. Magnitude not re-estimated here.

Evidence. Note Sec. 2.1.1 pp.7-8. The aggregate breakdown is internally consistent ($400+200+100+100+50 = $850bn; 850/17,690 = 4.8%) and close to NIPA 2018 (about $890bn; Sept 2020 tables $889.8bn, 5.0% of NI; Claude auditor).

Why this severity. A known, disclosed convention that affects bottom-group tax rates, not income or wealth shares; no re-estimate is available.

The authors' response. SZ treat transfers as transfers rather than negative taxes, consistent with national-accounts concepts (per critique corpus, us-wealth-taxes-12).

Claims examined and rejected

Problems one auditor raised that did not survive checking.

Audit notes

What reproduced. Reproduced (adjudicator re-runs or checks): household partnership equity $3,834bn in mid-2016 (23.8% of NI) and 2010-19 averages of 21.3% (partnerships) and 10.9% (sole props + farms) of NI; estate premium for $20m+ estates 2008-16 (mean 1.35, ratio of means 1.33); top-1% fixed-income share 1995 (29.2% vs DFA 29.5%); top-0.1% interest-bearing share in 2016 (about 27%); 2019 headline values (top-0.1% wealth 18.02%, top-1% pre-tax income 18.73%); SZZ (2023)'s description of the Sept 2020 series (+7.73/+6.35/+4.11 pp, 1989-2016); full vintage comparisons Nov 2017/Oct 2019/Sept 2020/Feb 2022; post-1962 tax tables add up; code matches the note for the interest premium, equity method, Forbes matching periods, pension-wealth weights, 45% cap, 3% capital-gains rule and property-tax discount level. In the public 2018 micro-file the business capital keys were reconstructed exactly (residual ~1e-16), allowing counterfactuals for the business formula (F9), the interest premium and property-tax differential (F2, F4) and the billionaire target (F3). Agreed with both auditors (not re-run): SCF $1.9T sole-prop/farm and 45% partnership figures, IRS LLC share 15.6%, FA 2020Q2 vintage checks, consumption-tax and capital-gains averages. Not reproduced: the 2018 '22%-23% in both datasets' figure (SZ is 25.1%); 'fully consistent' with the DFA in 2010-16 (gaps 1.2-3.2 pp); the isolated pension (-1 to -1.5 pts; -0.6 pt), equity ('modest') and housing effects; the published tables from public files (internal IRS data required; public 2018 file differs by up to 2.2 pp at the top 10% of wealth).

Limits of this audit. No Stata and no internal IRS data, so the pipeline could not be rerun and the note's isolated effects (pensions, equity, housing, mortgage reclassification) could not be verified. Only one September 2020 micro-file (the public 2018 file) was used; the 1.37 GB nested archive of the other years was not downloaded, so trend effects of F2, F4 and F9 rely on 2018 proxies and vintage differences. In the public file every large pass-through unit has zero passive income, so the F9 counterfactual is an upper range; the billionaire counterfactual in F3 is an upper bound. Forbes billionaire tallies come from the Claude auditor's capture of Forbes lists and were not independently re-downloaded; DFA instrument detail (IRAs, fixed income) is not public, so the DFA comparisons rely on the authors' own harmonised series. Several primary-source checks by the auditors (FA B.101/B.104 values, IRS Figure K, SCF microdata, FOMC projections, NIPA consumption taxes) were accepted where both auditors agreed or where the figure is not material, rather than re-run. The SZ (2020) Figure 28 series error concerns the companion paper, not the note's text.