← All papers

Emmanuel Saez · Updates of Saez (Pathways Magazine, Winter 2008) posted on Saez's website, several also issued as WTID/WID.world Technical Notes

Striking it Richer: The Evolution of Top Incomes in the United States (annual updates)

17 problems found: 1 major · 6 moderate · 10 minor

What the paper claims

Every problem found

1majorlikely · previously known · methodology

The notes say inequality fell abruptly during World War II and take that as evidence the war itself drove the change. A careful re-estimate from the same old tax records finds that pre-war inequality was lower, and that much of the fall came earlier, during the Great Depression. On that estimate the wartime drop is about a third smaller overall and far less sudden. The long-run 'U' pattern still holds, but the war-shock story is much less certain.

Technical detail and evidence

Why it matters, in detail. Mechanism: numerator uplifts that GMMS show exceed recorded AGI-eligible deductions in most pre-war years; tax-exempt public employees counted in tax units and in the denominator but not in the numerator; and a fixed 80% denominator below the post-war ratio (~83%) and the Goldsmith-reconciled level (~94% of PI less transfers in 1941). Together these inflate pre-war shares, and the biases fade as filing becomes near-universal in the 1940s, which mechanically steepens the measured wartime fall. Magnitude (GMMS, incl. capital gains): 1928 top-1% share 19.5% vs 23.9%; 1928-44 fall 8.6 vs 12.7 pp; 1940-44 fall 1.8 vs 5.2 pp (top 10%: 4.8 vs 12.8 pp). The exact size is uncertain because GMMS also impute (the 1937 public-employee distribution is extrapolated, the PS capital-gains treatment is kept, evasion is not corrected). Not every PS choice pushes the same way: the 1944-45 denominator uses a 20% non-filer imputation (F17), which raises 1944 shares.

Evidence. TabFig2024 Table A3 vs GMMS DataAppendix-6-7-21.xlsx, sheet TopIncomeShares (corrected, cumulative + full capital gains): top 10% 1928 49.29 vs 42.47, 1940 45.29 vs 36.06, 1944 32.51 vs 31.25; top 1% 1928 23.94 vs 19.47, 1940 16.48 vs 12.63, 1944 11.28 vs 10.84. The 1928-44 fall shrinks from 16.78 to 11.22 pp (top 10%, -33%) and from 12.66 to 8.63 pp (top 1%, -32%); the 1940-44 fall from 12.78 to 4.82 pp (-62%) and from 5.20 to 1.79 pp (-66%) (scripts/misc_checks.py section 8; GMMS accepted manuscript pp.2-3: PS 'overstate the magnitude of the observed leveling ... by roughly one third'). The assumed step-down of the deduction uplift (16% -> 9%) alone lowers the 1943 top-1% share by ~0.7 pp, about a fifth of the measured 1941-43 fall (15.79% -> 12.31%). PS's own post-war ratio of the fiscal denominator to PI less transfers averaged 82.7% in 1944-74 (GMMS p.24); in the 2026 NIPA vintage it averages 0.81 in 1946-79 (scripts/deflator_denominator.py), above the 0.80 assumed before 1944. Codex reproduced the GMMS archive arithmetic from group means. OpenAlex lists 29 works citing GMMS (doi:10.1093/ej/ueac020); none is a reply by Piketty or Saez.

Why this severity. The U-shape survives, but on the best available re-estimate the 1928-44 compression is about a third smaller and roughly two-thirds of the 1940-44 drop disappears. That removes most of the timing evidence the notes use to attribute the levelling to the war, a change of well over a quarter in the magnitude of a core finding. The post-1944 series and every recent finding are unaffected.

The authors' response. None located. Piketty & Saez (2003; 2004/2007 longer version, appendix pp.28-35) justify the 80% pre-1944 ratio by the post-war ratio and test Pareto interpolation and reranking. Their working-file notes say the deduction rates are 'reasonable values' taken from composition tables and Stata files. The updates repeat the 2008 text without mentioning GMMS.

2moderateconfirmed · newly identified here · version instability

In March 2024 the notes said the richest 1% took 81% of income growth from 2019 to 2022 and that everyone else's income grew only 1%. Those figures came from early tax-filing counts. With the full IRS data the figures become 66% and 2.4%. The main message, that the top took a disproportionate share, still holds, but the headline overstated it and the next update did not say it had changed.

Technical detail and evidence

Why it matters, in detail. The capture ratio c = 0.01*(change in top-1% mean)/(change in overall mean) has a small denominator (+4.2% preliminary). A 1.4% upward revision of the end-year bottom-99% mean raises the overall change by a third and lowers c from 0.81 to 0.66, while bottom-99% growth more than doubles. The projection under-counted returns (-1.3%) and income excluding gains (-1.9%) and over-estimated realised gains (+4.8%), which skewed the end-year distribution toward the top. Re-projecting with the then-latest (TY2021) ratios reproduces the final aggregates within about 0.1-0.3% and gives ~67% captured, so the error was not inherent in filing-season data.

Evidence. saez-UStopincomes-2022.pdf p.1, and p.5 ('2022 estimates have been updated using filing season tax return statistics posted by IRS in February 2024'). TabFig2022 vs TabFig2024 Table A0, row 2022: returns 159.25m vs 161.34m; income excl. capital gains $13,259.8bn vs $13,511.5bn; incl. $14,560.3bn vs $14,752.4bn. 2019-22 all/top 1%/bottom 99%/capture: 4.22/16.12/1.02/80.87 (TabFig2022) vs 5.59/17.41/2.42/65.83 (TabFig2024) (scripts/check_incidence.py, misc_checks.py section 1). IRS 22inweek47.xls, 23inweek47.xls, 21in14ar.xls, 22in14ar.xls: TY2021 final/FSS ratios 1.0421 (returns) and 1.0487 (AGI). claude/results/fss_projection.csv: 2022 projected with TY2021 ratios gives total income $14,734bn vs final $14,752bn. TabFig2022's 'What's new' is copied from TabFig2021 and does not mention the 2022 estimates.

Why this severity. One dated headline overstated the top 1%'s capture by 15 pp (81% vs 66%) and understated bottom-99% growth by 1.4 pp (1.0% vs 2.4%). The qualitative conclusion survives (the top 1% still took about three times its ~21% income share), and the latest 2019-24 headline is robust to the projection method. The effect on the series' core findings is noticeable but not transformative.

The authors' response. None specific. The March 2024 note discloses on p.5 that its 2022 estimates use filing-season statistics. Earlier notes (Mar 2019 and Feb 2020) state that 'it is possible to project reliable full-year statistics by June of the following year'.

3moderateconfirmed · previously known · selective presentation

The notes' best-known figure is that the top 1% took 95% of income gains in 2009-2012. It is correct for those exact years, but the period starts at the bottom of the recession and ends in a year when the rich had pulled income forward to dodge a tax increase. Adding one more year cuts it to about 69%, and other headline percentages were later revised by up to 15 points without notice. The general point, that the top took far more than its share, holds; the precise percentages are much less solid than they look.

Technical detail and evidence

Why it matters, in detail. With the overall mean rising only 2-7% in early recoveries, a +/-0.5% error or revision in the bottom-99% mean moves c by 5-15 pp, and an end year inflated by tax-motivated retiming (2012) raises it mechanically. Current data: 90.8% (2009-12), 68.9% (2009-13), 85.4% (2010-12); excl. capital gains 98.1% (2009-12). Revisions go both ways (93->103, 121->136, 106->118 up; 95->91, 81->66 down), so there is imprecision the headlines do not convey but no sign of systematic bias.

Evidence. saez-UStopincomes-2012.pdf p.1 and fn 4; 2014.pdf fn 1 ('abnormally high due to retiming'). Current-data windows: 2009-12 90.8% (98.1% excl. capital gains), 2009-13 68.9%, 2010-12 85.4%; over 2007-12 every group lost income (average -11.7%) and the top 1% bore 28.6% of the loss (scripts/misc_checks.py section 5). Composition: scripts/composition.py (2011-12 and 2012-13). Vintage revisions: results/incidence_check.csv. Oct 2019 version: Wayback capture (6 Dec 2019) of saez-UStopincomes-2018.pdf.

Why this severity. The direction (the top 1% capturing several times its income share of early-recovery gains) holds for every window ending in 2012-14 and in every vintage. The 95% is arithmetically correct for its window and robust to excluding capital gains. But defensible window choices move the magnitude by 20-25 pp, which qualifies a headline finding without overturning it.

The authors' response. Saez (Sept 2013, fn 4): 'the exact percentage 95% is sensitive to measurement error ...'. The same note flags 2012 as preliminary and possibly retimed. Later notes call 2012 'abnormally high due to retiming' and report 91%. No direct reply to Rose located.

4moderateconfirmed · previously known · methodology

The figures count only the cash income people report on tax returns, before taxes and without government benefits or employer-paid health insurance, and they treat each tax-filing unit (a single person or a couple) as a 'family'. That yardstick covers a shrinking slice of the income the national accounts record, which makes recent top shares look somewhat higher and the rise somewhat steeper. On the authors' own broader measure, 2021 was not a record; 1928 was.

Technical detail and evidence

Why it matters, in detail. The ratio of the fiscal denominator to PI less transfers fell from ~0.81 (1946-79) to ~0.735 (2010-24). Because the untaxed income is disproportionately middle-class, a fixed-ratio denominator lowers the 2024 top-1% share by ~1.7 pp (20.75% vs 22.45%) and the 1980-2024 increase from 12.4 to 10.7 pp. The 2021 record is driven by realisations (27.6% incl. vs 21.2% excl. capital gains); DINA, which excludes realised gains and distributes all national income to equal-split adults, peaks in 1928. Tax units outnumber households (by 27% in 2004 per the workbook), so bottom-99% 'families' include single adults living in larger households (LMS 2021: ~2 pp on the 2010 top-1% level).

Evidence. saez-UStopincomes-2024.pdf p.6 ('share of total personal income'), p.7 ('overall economic growth'), fn 1. TabFig2024 Explanations rows 7-19 and Table A0 notes. BEA NipaDataA.txt (A065RC - A577RC) vs Table A0 col 4 (scripts/deflator_denominator.py). WID bulk download (15 Oct 2026) sptincj992 p99p100 (scripts/wid_range_fetch.py, wid_compare.py). Larrimore, Mortenson & Splinter (2021, JHR Table 3): 2010 top 1% 15.6% for tax units vs 13.5% for households. Bricker et al. (2016, BPEA p.301): 2012 top 1% 'just under 18 percent' (SCF-based) vs 'nearly 23 percent' (tax data).

Why this severity. The concept is defined in each note, and the authors' national-income series corroborates the long-run rise (+8.6 pp 1980-2019 vs +9.4 pp in the fiscal series excluding gains), so the core trend survives. But the level, part of the post-1980 trend, and the 'record' and 'economic growth' framing depend on a concept narrower than the labels suggest.

The authors' response. Footnote 1 (since 2016): fiscal income per family has grown more slowly than national income per adult, and PSZ (2018) DINA and Realtime Inequality reconcile with national accounts. Feb 2023 fn 1: the surge in tax data is larger than in DINA because it combines rising concentration with surging capital-gains realisations. PSZ (2018) adopted equal-split adults to neutralise the decline in marriage.

5moderateconfirmed · previously known · methodology

To compare incomes across years the notes adjust for inflation with one particular consumer price index. With another widely used index, incomes of the bottom 99% grew about 42% from 1993 to 2024 instead of 27%. The rich still pulled ahead, but 'stagnation' for everyone else looks less severe.

Technical detail and evidence

Why it matters, in detail. The CPI-U-RS is a Laspeyres-type index (geometric means only at the lower level), so it keeps upper-level substitution bias relative to the Fisher-chained PCE index. PCE also gives more weight to health care paid by employers and government, which arguably fits a broader income concept than the cash income measured here. A common deflator scales every group alike: growth is ~12% higher cumulatively for all groups over 1993-2024, capture falls 6 pp, and top-1% growth is ~3.5 times bottom-99% growth (vs ~4.5 times).

Evidence. TabFig2024 Table A0 col 8 (CPI-U-RS factors) and Explanations rows 45-47; BEA NIPA series DPCERG, 2026 vintage (scripts/deflator_denominator.py -> results/deflator_denominator.txt). Both auditors obtained the same figures independently (claude/scripts/sensitivity_growth.py; codex/scripts/final_checks.py).

Why this severity. Statements such as 'bottom 99% grew only by 26.6%' change by more than half under a standard alternative index. But the gap between top and bottom growth and all shares survive, and the choice is disclosed and defensible, so the effect on the core concentration findings is modest.

The authors' response. Disclosed in the Table 1 notes and fn 1 ('We deflate incomes using the Consumer Price Index instead of the National Income deflator'). The workbook's Explanations sheet justifies the CPI-U-RS over the CPI-U.

6moderateconfirmed · newly identified here · overclaiming

The notes say rising inequality mostly reflects soaring pay for top earners. That was true up to about 2000. In the authors' own updated data, the top 1%'s gains since 2000 come from business profits and investment income, while the wage part has shrunk. The wage figures the text cites stop in 2011, and one is misquoted.

Technical detail and evidence

Why it matters, in detail. Since 2000 the whole rise in the top-1% fiscal share excluding gains (+2.0 pp) is non-wage: pass-through business income, dividends and interest. Whether this contradicts a 'labor market' story depends on how much of pass-through profit is labour income. Smith, Yagan, Zidar & Zwick (2019) classify ~75% of it as human-capital income; PSZ/SZ (2020) assume less. The notes' own W-2-based wage series, which could settle the wage component directly, has not been updated since 2011.

Evidence. saez-UStopincomes-2024.pdf p.8; TabFig2024 Table B2 (1970 5.133, 2007 12.200, last year 2011); Table A7 x Table A1 (scripts/composition.py -> results/composition.txt).

Why this severity. The 1970-2007 statement is accurate (wages explain ~61% of the rise), but the present-tense conclusion for recent decades is not supported by the composition data the workbook itself publishes. This affects a secondary, interpretive finding, not the measured shares.

7moderatelikely · previously known · methodology

A 1986 tax law made it worthwhile for business owners to report profits on their personal tax returns instead of keeping them inside corporations. Over 1986-88 the measured top 1% share jumped by about four points, roughly two-fifths of its whole long-term rise, partly because the income moved to a different tax form rather than to different people. The notes never mention this when they describe the trend.

Technical detail and evidence

Why it matters, in detail. Reclassification moves income from the corporate base onto individual returns, raising measured fiscal top shares without any change in economic income. The jump is +4.1 pp in three years against a ~0.25 pp/yr long-run trend. If, say, half of it were reclassification (BFJL imply more, PSZ less), the 1962-2019 rise excluding gains would be ~22% smaller. Counter-evidence: the wage-only top-1% share also rose (5.1% -> 12.2%, 1970-2007), and Smith et al. (2019) show that much post-1986 pass-through income is genuine, largely owner-labour, income.

Evidence. TabFig2024 Tables A1/A3, rows 1980-1991 (scripts/misc_checks.py section 7); WID sptincj992 p99p100, 1985-88; 2024.pdf p.2 fn 2 and p.6; Piketty & Saez (2003, QJE pp.3-4: 'a significant part of the gain is concentrated in 1987 and 1988 just after the Tax Reform Act of 1986'); BFJL (2012, REStat Table 1 p.377: 1986-88 top-1% share +22.1%/yr in PS vs +2.0%/yr in CPS); Auten & Splinter (2024, JPE p.2181).

Why this severity. A large share of the measured long-run rise excluding gains falls in a tax-reform window when income changed the form in which it was reported. How much of the jump is reclassification and how much genuine growth is disputed, but the narrative omits a caveat that bears on roughly 40% of the trend. The direction of the post-1980 rise is not in doubt.

The authors' response. Piketty & Saez (2003, pp.3-4, 31) acknowledged the 1986-88 jump as at least partly fiscal manipulation but argued that long-run trends reflect real change. PSZ (2018) DINA add retained earnings. The updates' fn 2 calls the missing link between C corporations and their owners 'perhaps the most significant measurement gap'.

8minorconfirmed · newly identified here · arithmetic or transcription

One table in the March 2012 note copied the wrong number: the top 1% took about 60% of income growth over 1993-2010, not 52%. The mistake made inequality look slightly smaller, and later notes corrected it.

Technical detail and evidence

Why it matters, in detail. A wrong-row reference: the reported full-period capture is 7.7 pp too low (52.3% vs 59.98%). The error runs against the authors' thesis, so it is not evidence of bias.

Evidence. saez-UStopincomes-2010.pdf p.4 and Table 1 p.7; TabFig2010.xls Table1!E5 and Table_Incomegrowth!E146; recomputed 59.98% (61.3% in current data) (scripts/check_incidence.py; same result in both audits).

Why this severity. A transcription slip in a superseded note. It understates concentration, does not change the conclusion that the top 1% took a disproportionate share, and was fixed in later updates.

9minorconfirmed · newly identified here · arithmetic or transcription

The notes repeatedly said how much of their recession losses ordinary families had won back, but the arithmetic mixed up starting points. The recovery was reported as faster than it was in 2014-2017 and slower than it was from 2018 on. This does not change the story of a slow recovery.

Technical detail and evidence

Why it matters, in detail. The loss is measured on the 2007 base and the gain on the lower 2009 base; the correct ratio is g_t/(1/0.884 - 1) = g_t/0.131, not g_t/0.116. The direction of the error depends on whether the text was updated (2014-17) or carried over (2018-2026).

Evidence. 2014.pdf p.1; 2015.pdf p.1; 2017.pdf p.1; 2018.pdf p.1; 2024.pdf pp.2-3; scripts/misc_checks.py section 4 (own vintages TabFig2014prel, 2015prel, 2017, 2018 and TabFig2024).

Why this severity. Verbal characterisations of the recovery were off by 5-9 pp in 2014-17 (too fast) and by ~10-13 pp from 2018 on (too slow). The picture of a slow, lopsided recovery is unaffected.

10minorconfirmed · newly identified here · documentation gap

The authors' spreadsheet mixes early and final IRS figures in ways its own notes do not describe, and the method used to estimate the most recent years from early IRS counts is not published. The effects on the numbers are small, but outsiders cannot fully reproduce the latest estimates.

Technical detail and evidence

Why it matters, in detail. Mixing a preliminary numerator with a final denominator makes year-to-year changes partly artefactual: the 2022 excl.-gains top-10% share moved 0.9 pp with no change in the measured top incomes. Without the projection inputs, the latest-year shares can be approximated (top 1% 22.5-22.6% for 2024 vs 22.45% published) but not replicated.

Evidence. scripts/misc_checks.py section 2; TabFig2016-2024 Table A0 row 2016 vs IRS 16in14ar.xls (150,272,157 returns; taxable net gain $634.3bn less net loss $20.1bn); 2017.pdf p.5; TabFig2017 and TabFig2024 Explanations; 23in14ar.xls Last-Modified 26 Mar 2026; claude/results/fss_projection.csv; codex/calculations/primary_results.json.

Why this severity. Effects on published shares are at most ~0.4 pp, and on the latest headline at most ~1 pp. The problem is reproducibility and transparency, though the same undocumented projection step produced the material 2022 error in F2.

11minorconfirmed · newly identified here · external inconsistency

The World Inequality Database, which the authors help run, publishes its own version of this series, and it does not match the authors' spreadsheet in about a third of the years, sometimes by several points. Anyone citing 'Piketty-Saez' numbers from that database may be quoting different figures from the ones in these notes.

Technical detail and evidence

Why it matters, in detail. The same nominal series appears in two author-maintained outlets with differences of up to 1.8 pp (top 1%) and 3.6 pp (top 10%) in recent years. The pattern (exact in some years, off in adjacent ones) suggests a patchwork of vintages or constructions; cross-study comparisons that mix the two will be inconsistent.

Evidence. Fresh range-request download of wid_all_data.zip (member WID_data_US.csv dated 15 Oct 2026; scripts/wid_range_fetch.py), identical to the shared cache; scripts/wid_compare.py -> results/wid_compare_fresh.txt.

Why this severity. The notes cite TabFig, not WID, so their own findings are unaffected. The discrepancy matters for users who take the WID series to be the Piketty-Saez series.

12minorconfirmed · newly identified here · internal inconsistency

Older paragraphs are copied into each new edition without being updated, so some statements no longer match the authors' own numbers, for example the claim that growth in 2018 was shared evenly. There are also small typos and wrong dates. None of these change the main findings.

Technical detail and evidence

Why it matters, in detail. Text-data divergence from carrying commentary forward across vintages. In current data the 2017-18 growth of the top 10% (1.90%) exceeds that of the bottom 90% (1.43%), reversing the stated conclusion; the other items are rounding, labelling and date slips.

Evidence. 2024.pdf pp.2-3, 5, 7; 2021.pdf (lines on 1993-2022 and the sub-periods); 2017.pdf p.1; 2018.pdf p.5; Wayback Oct 2019 version p.1 vs TabFig2018prel; 2012-2014.pdf GDP footnote; 2015.pdf p.1; 2013.pdf p.6; results/incidence_check.csv; scripts/misc_checks.py.

Why this severity. Presentation and maintenance errors that do not affect the headline findings, although the 2017-18 'evenly distributed' claim is now reversed by the authors' own data.

13minorconfirmed · newly identified here · documentation gap

The spreadsheet's instructions get wrong which tables include investment gains, one chart label quotes the wrong income cut-off, and two notes link to an older version of the data that does not match their text. These are bookkeeping errors, not errors in the results.

Technical detail and evidence

Why it matters, in detail. Pure documentation defects. The table headers and the computed shares use the right concepts; following the printed links leads to superseded workbooks whose numbers differ by ~1-3 pp.

Evidence. TabFig2024 Explanations rows 15, 22, 41; Thresholds_nominal row 117; 2017.pdf and 2018.pdf p.1 links; TabFig2018prel Table A3; scripts/misc_checks.py; codex/scripts/audit_details.py; claude/results/.

Why this severity. Labelling and documentation errors that hamper reuse but do not affect published shares or conclusions.

14minorlikely · newly identified here · overclaiming

The latest note says COVID did not interrupt a steady rise in top income shares. In fact the top 1%'s share has been roughly flat since 2007, and the top 10%'s share has fallen since 2021. The long-term rise since 1980 is real, but 'steadily increasing' overstates recent years.

Technical detail and evidence

Why it matters, in detail. Using short windows to characterise a long-run trend: since 2007 the top-1% share is flat to down and the top-10% share excluding gains is up 2.6 pp; the 'steady' path holds only on a 1980 base.

Evidence. 2022.pdf p.1; 2024.pdf p.1; TabFig2024 Tables A1/A3; scripts/misc_checks.py section 5; check in results/incidence_check.csv.

Why this severity. Interpretive wording about short windows. The long-run rise since 1980 is not in question, and the 2019-24 numbers themselves reproduce.

15minorlikely · documentation gap

The authors' notes describe one rule for estimating the income of people who did not file tax returns in 1944-45, but the published numbers use another. The difference slightly changes the wartime low point and works against, not for, a dramatic wartime drop. It matters mainly for anyone trying to rebuild the series.

Technical detail and evidence

Why it matters, in detail. Denominator = filer income x (1 + f x non-filers/filers). With non-filers/filers = 0.250 (1944) and 0.182 (1945), f = 0.5 instead of 0.2 raises the denominator by 7.1% and 5.3%. Applying the documented rule would lengthen the 1940-44 fall in the top decile from 12.8 to ~14.9 pp.

Evidence. scripts/misc_checks.py section 3; TabFig2007-TabFig2024 Table A0 rows 1944-47; ps2004.txt (piketty-saezOUP04US.pdf) appendix text; /mnt/zdata/downloads/crit-usincome/PS2003_QJE.txt Table A1 row 1944.

Why this severity. The documentation does not match the series. The published choice makes the 1944 trough shallower, so it works against an exaggerated wartime drop and does not affect conclusions.

16minorplausible · overclaiming

The notes say the 2013 tax increase on top earners made the rich shift income into 2012 but did not lastingly reduce inequality before taxes. That is plausible, but it depends on guessing what would have happened without the tax change, and the short note leaves out the caveat its own companion study states.

Technical detail and evidence

Why it matters, in detail. The medium-run elasticity (~0.25 with the 2009-11 recovery trend; negative with the 1978-2011 trend or constant shares) is pinned down only by the counterfactual. 'No significant impact' is conditional on that choice, and reclassification responses (ASN 2016) would not show up as lower top shares.

Evidence. 2014.pdf p.1; 2015.pdf p.1; 2024.pdf pp.3-4; Saez (2017) pp.73-74, 99-101 (/mnt/zdata/downloads/saez-striking-audit/saezTPE17.pdf); corpus optimal-tax-behavior-21.

Why this severity. The descriptive retiming evidence is strong, the claim is hedged ('seems', 'suggests'), and it holds against the alternative counterfactuals in the cited companion paper. Only the caveat is missing.

The authors' response. Saez (2017, TPE pp.73-74, 99-101) states the identifying assumption and tests it against the recoveries after 1991 and 2001; the June 2026 note cites this paper (fn 8).

17minorplausible · previously known · methodology

To compute income per 'family' the notes need a count of families. For the latest years that count was extended at a fixed growth rate and grew more slowly than the adult population. Using the faster growth would lower ordinary families' income growth a little and make the top 1%'s share of gains slightly larger, so this does not undermine the findings.

Technical detail and evidence

Why it matters, in detail. Per-family averages divide by the CPS-based count of tax units. Undercounting 2024 units by ~0.7% (if units track adults 20+) overstates 2024 average and bottom-99% income by about the same proportion. The non-filer imputation offsets about a fifth of this.

Evidence. TabFig2021/2022/2024 Table A0 col 1; Census nc-est2025-agesex-res.csv; scripts/misc_checks.py section 6; claude/scripts/check_tax_units.py.

Why this severity. The direction of the likely correction would slightly strengthen the headline, and the magnitude is at most ~2 pp of the capture share.

The authors' response. Saez (March 2012 fn 2; Sept 2013 fn 4): the capture percentage is 'sensitive to measurement error, especially the growth in the total number of families'.

Claims examined and rejected

Problems one auditor raised that did not survive checking.

Audit notes

What reproduced. Reproduced by the adjudicator: (i) the printed growth-incidence headline cells of all 11 updates from their own TabFig vintages, exact to rounding except March 2012's 1993-2010 capture (52% printed vs 60.0%), June 2016's 7.7% (7.65%) and March 2019's 2.9% (2.82%) (scripts/check_incidence.py); (ii) the June 2026 Table 1 exactly; (iii) published top-1% shares from IRS SOI Table 1.4 brackets within 0.04 pp for 2019, 2021 and 2022 (soi_top1.py); (iv) the 2022 projection error (-1.3% total income, -1.9% excluding gains) and the counterfactual using TY2021 final/filing-season ratios (AGI +0.3% vs final) (misc_checks.py); (v) the deflator, denominator-drift, TRA86, composition, loss-recovery, tax-unit and window sensitivities; (vi) the WID comparisons on a fresh official bulk download; (vii) the GMMS corrected shares from their data appendix; (viii) the 1944-46 non-filer fractions from PS's working file. Reviewed but not re-run in full: the auditors' bracket-level SOI denominator reconstructions (exact in most years 2008-22, -0.14 to -0.16% in 2007) and Claude's bracket-level filing-season projection. Not reproducible from public materials: the exact projection algorithm for the 2023-24 shares, the shares excluding capital gains (public tables rank by AGI including gains), the pre-1960 series from historical SOI reports, the CPS tax-unit construction, and the BLS CPI-U-RS file (HTTP 403).

Limits of this audit. The pre-1960 series was not re-estimated from historical SOI reports: the GMMS correction is assessed from their data appendix, their text and spot checks of PS's working files, and GMMS's own adjustments (Goldsmith-type denominator, the 1937 public-employee distribution, deductions) were not audited line by line. The 2024 estimates cannot yet be checked against final IRS data (TY2024 Table 1.4 returned 404 on 25 Oct 2026), and the 2025-26 forecast is untestable. The authors' code and interpolation inputs are not public, so the shares excluding capital gains and the exact projection step could not be replicated. The WID.world Technical Note versions (2012/2-2015/6) and the Routledge reprints were not retrieved (assumed identical to the website PDFs), and the June 2018 version of the 2017 note is not archived. The BLS R-CPI-U-RS file was inaccessible (HTTP 403). External re-estimates (Auten-Splinter, BFJL, LMS, BHKS, SYZZ) were used as published. The author was not contacted.