← All papers

Mark W. Frank, Mark Price, Emmanuel Saez, Estelle Sommeiller · WTID / WID.world Technical Note 2015/7

Frank-Sommeiller-Price Series for Top Income Shares by US States since 1917

13 problems found: 7 moderate · 6 minor

What the paper claims

Every problem found

1moderateconfirmed · newly identified here · overclaiming

To make the state figures match the official national figures, every state's numbers were multiplied by the same correction each year. The note calls this correction minor, but for some groups and years it is large: up to 32% for the top 10% in the 1930s, and it more than doubles the top 0.01% in the late 1980s. In those cases much of what looks like each state's own history is really the national trend copied into every state. For the most-used figure, the top 1%, the correction is small.

Technical detail and evidence

Why it matters, in detail. Published share s_{i,t,x} = SP_{i,t,x} x f_{t,x}. For the top 10% in 1917-41, f averages 1.118 (up to 1.316), lifting every state to PS's level (1928: 49.3% vs SP's own 43.1%): half of the median state's 1928-79 top-decile compression (-9.9 vs -4.3 pp) is imposed. For the top 0.1-0.01% in 1986-89, f = 1.57-2.43, so 36-59% of each state's level and essentially all TRA86-era dynamics are national (top-1% 1988-90 change reversed in 44/49 states). Long-run top-1% trends are 86-96% state-data driven.

Evidence. scripts/a1_benchmark.py (work/a1_factors.csv): 11 years with |f-1|>0.2 for the top 0.01%; mean |PS - SP-US| for the top 10% 4.6 pp in 1917-41 vs 0.4 pp after. scripts/a2_factor_trends.py (49 units, excl. AK/HI): top 10% 1928->1979 median change -9.9 pp in the file vs -4.3 pp in SP's own state estimates (statewise median 51% of the decline comes from the factor; sign reversed in 5 states; a decline appears in 44/49 states in the file vs 39/49 in raw SP); top 0.1%/0.01% 1979->2007: 26%/30% of the rise comes from the factor; top 1% 1988->1990: -1.10 pp (file) vs +1.24 pp (SP), sign differs in 44/49 states; top 0.01% 1982->1988: 35/49 sign reversals. For the top 1% the factor contributes 4% of the 1928-79 decline, 14% of the 1979-2007 rise and -7% of the 1979-2012 rise. National top-10% U-shape: SP-US +8.5 pp 1928->2012 vs PS +1.3 pp. Note text: WID PDF p.3 and fn 1.

Why this severity. The method is disclosed and the factor is uniform within each year, so within-year cross-state rankings and log/year-effect panel analyses are unaffected and the headline top-1% series barely depends on it. But the note's description is wrong for the top decile before 1942 and for the top 0.5-0.01% in 1974, 1982-89 and 2002-09. There, a quarter to more than half of state-level changes come from the national series. This qualifies, rather than overturns, the dataset's main content.

2moderateconfirmed · newly identified here · speculative assumption

In some years the IRS published state data only up to '$200,000 and above', so the incomes of the very richest had to be estimated by extrapolation. The size of that estimation error differs from state to state, but the note applies the same national fix to every state. For the richest 0.1% and 0.01%, the state numbers in those years can therefore be off by 10-40% in either direction. The top 1% figures are only slightly affected.

Technical detail and evidence

Why it matters, in detail. With a common factor, the published share equals the true (fine-bracket) SP share divided by r_i = f_i/f_US, where f_i is the state's own needed correction. In simulations, r_i for the top 0.01% has a p10-p90 range of about 0.83-1.14 (extremes 0.77 SD/ND to 1.44 VA). The induced log errors reappear as state-specific breaks at 2001/02 and 2009/10 (r = 0.64-0.77, slopes 1.4-1.5). Cross-state comparisons and within-state trends of very-top shares across 2002-09, and likely 1974 and 1982-89, therefore carry spurious +/-10-40% level shifts.

Evidence. scripts/a6_pareto.py: an independent implementation of SP's rule (EPI 2015 App. Tables A5-A6) reproduces SP's raw 2012 shares for all 51 states + US exactly (x1.0100 AGI add-back), and 2001/2010 state shares within 0.3%. Collapsing IRS classes >= $200k into one open class in fine-bracket years gives a national 'needed factor' for the top 0.01% of 1.14 (2001), 1.39 (2010), 1.36 (2012) and 1.42 (2014); the note's own 2002-09 factors are 1.17-1.36. State/national ratios for the top 0.01% range 0.85-1.16 (2001), 0.81-1.44 (2010), 0.77-1.34 (2012) and 0.77-1.30 (2014), with 27-43% of states off by >10%. Top 0.1%: 0.87-1.21. Top 1%: 0.95-1.04. Top 5%/10%: unaffected. Predicted spurious jumps correlate with the file's state-relative log changes at the regime switches: top 0.01% r = 0.64 (2001->02) and 0.77 (2009->10); top 0.1% 0.59/0.69; top 0.5% 0.60/0.53. Placebo non-switch years give a mean r of about -0.1. scripts/a7_addingup_frank.py: the sum of states' top-0.01% incomes exceeds the national top-0.01% income, which is infeasible, in 1974 and 1984-89 (max 1.022, 1985). Frank (2026) online appendix sections 3.2 and 8: state tables have far fewer classes in 1974, 1982-1988 and 2002-2009; section 1: ('as few as four classes in 1974 and 1982-1983').

Why this severity. Affects only the top 0.5-0.01% series in about 18 coarse-bracket years. The typical state-specific error at the top 0.01% is about +/-15%, up to about +/-30-40%. The top 1% is within +/-4% and the top 5-10% are unaffected, so the main series and the U-shape are not at stake.

The authors' response. Partly anticipated: fn 5 flags 2002-09 as problematic and argues that Pareto extrapolation is preferable; state-specific errors are not discussed.

3moderateconfirmed · newly identified here · methodology

A 'top 1% share' needs an estimate of everyone's income in the state. Before 1929 that estimate was built from the incomes of the few people who filed tax returns, who were mostly the rich. So when the rich got richer, the estimate of everyone's income rose too, hiding the change; Florida's 1920s land boom even shows inequality falling. From 1929 a different source is used that leaves out stock-market gains. That produces impossible results such as Delaware in 1929, where the top 10% get 97% of all income.

Technical detail and evidence

Why it matters, in detail. Before 1929, D_{i,t} = k_i x NI_{i,t} with NI the filers' net income, so a state's share is roughly (top income / filer income) x const. State dynamics are therefore within-filer dynamics: attenuated towards zero, and perverse in booms (FL 1925). The 1929 switch moves state income shares by a median 8.8%; on a 1929 basis, 1928 top-1% levels shift by a median 1.3 pp (max 6.8 pp). BEA excludes realized gains, so gain-heavy state-years (DE 1929, WY 2012) get top shares near or beyond feasibility.

Evidence. scripts/a3_anomalies.py: the cross-state CV of (TotalInc/AGI)_t/(TotalInc/AGI)_1928 is 0.0000 in every year 1917-28. Filers were <10% of tax units in 704 state-years (1917-40), so the denominator of, e.g., a top-1% share is essentially a multiple of the top decile's own income. The median |dlog state income share| is 11.0% in 1925 (filing-threshold change) and 8.8% in 1929 (switch to BEA), vs 2.9% typical (1930-60); 32 states moved >5% in 1929. Re-basing 1928 on 1929 state income shares changes 1928 top-1% shares by a median 1.3 pp (max 6.8 pp, NY 30.2->37.0; 19 states >2 pp). Florida's share of US income goes 0.70% (1923) -> 3.39% (1925) -> 0.83% (1928), so its top-1% share falls from 17.8% to 13.8% at the 1925 land-boom peak. Delaware 1929: top 10% = 96.6% and top 1% = 69.5%, implying a bottom-90% mean of 3.8% of the state mean (US minimum 55%). AGI exceeds TotalInc in 668 state-years (431 since 1987). scripts/a9_denominator_sensitivity.py: after 1942, allocating income by AGI instead of BEA changes top-1% levels by a median 0.3-0.4 pp (rank correlation ~0.95).

Why this severity. Makes pre-1929 state top-share dynamics and a few outliers unreliable, and shifts 1928 levels by about 1 pp (up to 7 pp). But the 1928>1979 decline survives in all 49 states, and post-1942 levels are robust to the denominator.

The authors' response. SP (2015, pp.23-24) document the construction; the note does not describe it.

4moderateconfirmed · newly identified here · data error

For Alaska before it became a state, some published numbers are mathematically impossible; for example, the richest 10% supposedly receive less than 10% of income. This happened because the underlying tax data were mixed with Washington State's and the population and income totals were guessed. These years should not be used, but the note gives no warning.

Technical detail and evidence

Why it matters, in detail. Feasibility requires s_x >= x and non-decreasing group means; 24 state-years fail, 20 of them Alaska. Alaska's counts and denominators are not state-specific in 1921-39 and 1943-54. Benchmarking removed most of SP's 89 violated pairs, but the residual cells cannot be interpreted.

Evidence. scripts/a3_anomalies.py and a11_violation_origin.py: Alaska's top-10% share is 7.7/7.5/8.1% in 1918-20 and its top-5% share 4.7/4.2/4.4%; a top x% group cannot hold less than x% of income. Across 7 disjoint groups there are 27 violated adjacent-mean inequalities in 24 state-years. 20 of those are Alaska: 17 inherited from SP, 3 in the imputed 1983-85 rows (AK 1985: P99.9-99.99 mean is 11.8x the state mean vs 1.9x for the top 0.01%). The rest are small violations in MS 1919 (inherited) and in DC 1931, SC 1952, WV 1952 and AK 1955 (created by fractile-specific benchmarking). Alaska returns/tax units are 3.2-3.6 in 1943 and 1945-54, and AGI is 1.6-2.3x TotalInc; IRS reported Alaska with Washington in 1921-39 and 1943-54 (Frank 2026 App. section 8). Alaska and Hawaii incomes for 1917-49 are backcast from 1950 shares (SP endnote 13), giving Alaska an average income 3.05-3.79x the US in 1917-20. Hawaii returns exceed tax units in 1942-50. Frank's later releases include AK/HI only from 1959; the current WID still shows the impossible 1918-20 values.

Why this severity. Limited to 2 of 51 units before 1959 (about 84 state-years) plus scattered cells. But the values are unusable, the note does not flag them, and they enter any panel analysis that includes these units. National and other-state results are unaffected.

5moderateconfirmed · newly identified here · data error

In 1944 the IRS published only each state's totals, not how income was spread within the state, so there was no real information on that year's state top shares. The published 1944 state numbers drop and bounce back for no real reason (New York's top 1% goes 15%, 11%, 16% in 1943-45).

Technical detail and evidence

Why it matters, in detail. Cross-state ranks in 1944 are uncorrelated with 1943/45 (rho about -0.1 to -0.25, vs 0.88). 1944 state levels therefore deviate from interpolated values by up to about 10 pp (DE), and any state-level analysis using 1944 (e.g. 1928-44 or 1944-49 changes) is contaminated.

Evidence. Frank (2026) online appendix section 8: 'For the year 1944, the IRS only reported each state's total number of returns and total income'. scripts/a3_anomalies.py: Spearman rank correlation of state top-1% shares (excl. AK/HI) is -0.11 for 1943-44 and -0.20 for 1944-45, vs 0.88 for 1943-45 and 0.92 for typical adjacent years; top 0.01%: -0.12/-0.25. SP's raw 1944 values show the same (-0.11), so the defect is inherited. Spurious V-shapes in the top 1% (1943/44/45): New York 15.3/10.6/16.3, Delaware 18.6/9.4/20.9, Florida 13.0/9.3/13.0, Nevada 12.1/8.5/13.9.

Why this severity. One year only, but it is the trough year of the wartime compression and its 51 state values are effectively noise. The note does not flag it, so users comparing states in 1944 or measuring 1940s changes would be misled. Long-run trends are unaffected.

6moderateconfirmed · newly identified here · version instability

The figures now online at the World Inequality Database, which cite this note, are not the ones the note describes. They were recomputed from a later version of the underlying data, include impossible values (Florida's top 10% getting 166% of income in 1925), and quietly undo the note's own Wyoming fix.

Technical detail and evidence

Why it matters, in detail. WID_{i,t} = SP2018_{i,t} x f_t(2015). SP's 2018 vintage changed pre-1929 state denominators (e.g. FL 1924-27), so applying the 2015 factors to the new raw data yields a top-10% share above 100% (FL 1925) and moves 11% of top-1% values by more than 1 pp, without documentation.

Evidence. The adjudicator fetched WID_data_US-FL/DE and the metadata from the WID bulk archive (15 Oct 2026) by range requests (scripts/fetch_zip_member.py). scripts/a8_wid_vintage.py: WID equals SP-2018 x the TN factor to 0.01 pp for FL and DE 1923-30. Across 49 jurisdictions (auditor A's download), WID differs from the note's file by >0.05 pp in 93% (top 10%) and 77% (top 1%) of state-years, and by >1 pp in 23%/11%. Florida 1925: WID top 10% 166.3%, bottom 90% -68.1%, top 1% 63.3% (file 36.3%/13.8%). Delaware 1923-26 top 1%: 15-33% (file 23-39%). Wyoming 2012 top 1%: 68.7% (file 29.7%). Alaska 1918-20 top 10%: below 10%. WID metadata (WID_metadata_US-FL.csv) cites the TN URL. WID also publishes these numbers as 'pre-tax national income' (sptinct992), identical to the fiscal series (auditor A).

Why this severity. Does not affect the note's documented file. But the series users now get under the note's citation are undocumented revisions, with some impossible cells (FL 1925) and a reversed documented correction (WY 2012). Most cells move by less than 1 pp.

7moderatelikely · previously known · external inconsistency

Before World War II only a small minority of households filed federal tax returns, so the top 10%'s share has to be estimated for people who left no records. Researchers using more complete state tax records (e.g. Wisconsin's) find much less concentration than the IRS-based numbers used here. A peer-reviewed reconstruction of the national series also finds that 1920s inequality and its wartime fall were overstated. The basic story (high inequality in the 1920s, lower after the war) survives, but its size in the early years is uncertain.

Technical detail and evidence

Why it matters, in detail. Top-decile state shares in 1917-41 are model-driven. The two national methods (SP vs PS) disagree by 4.6 pp on average, and the factor adopts the higher PS level (+11.8% on average). Both inherit PS's undocumented net-income-to-AGI multipliers and its 80%-of-personal-income denominator, which GMMS criticise. Substituting GMMS's national targets mechanically lowers 1928 state levels by 18.7% (top 1%) and 13.8% (top 10%) and cuts the 1928-44 compression by a third. The Wisconsin state-tax comparison suggests that IRS-based top-decile shares are roughly double the state-record value.

Evidence. scripts/a3_anomalies.py: filers/tax units <10% in 704 state-years 1917-40 (275 below 5%; Mississippi 1931 1.47%); the US ratio is 6.4-9.0% in 1925-35. scripts/a1_benchmark.py: SP-US and PS top-decile shares differ by 4.6 pp on average in 1917-41 (max 10.7 pp, 1931). In 1928, SP's own US top decile (43.1%) is close to the 42.5% corrected estimate of Geloso, Magness, Moore & Schlosser (EJ 2022, accepted ms pp.2-3; us-income-07), but the note's x1.144 factor lifts every state to PS's 49.3%. GMMS's national top-1% share (19.5% in 1928 vs PS 23.9%) would lower every 1928 state top-1% level by 18.7% (e.g. NY 30.2% -> 24.5%) and shrink the 1928-44 national drop from 12.7 to 8.6 pp; this is auditor B's conditional substitution, re-derived. Geloso & Magness (2020, Economic Inquiry 58(2):834-855, abstract): state income-tax records 1919-45 'show lower overall levels of income inequality than IRS records'. Their working-paper chart (Magness 2017 blog) shows Wisconsin top-10% shares of about 16-21% (state records) vs about 29-33% (IRS) in 1919-25; the note's file has 33.6-38.4% (raw SP 32.3-36.8%). The top-0.01% group in small states is a handful of tax units (Nevada 1917: 3.9; Wyoming: 7.5).

Why this severity. Concerns the levels and the depth of the left arm of the U (1917-41). Pre-war state top-decile levels could be overstated by about a seventh (the GMMS-consistent national level) to around a half (the Wisconsin state-record comparison). It does not reverse the 1928->1979 top-1% decline, which holds in all states.

The authors' response. No response by Piketty or Saez located (corpus us-income-07; none among the 29 OpenAlex citations of GMMS 2022). SP (2015, pp.25-26) warn users about interpolation errors; the note does not pass this on.

8minorconfirmed · newly identified here · speculative assumption

IRS figures show a huge jump in the income of Wyoming's millionaires in 2012. The note calls this a typo and replaces it. But the IRS numbers hang together (deductions, taxable income and tax all jump too), so the jump may reflect one or two real, extraordinary tax returns, and the replacement does not reproduce the published number. It is a single state-year, so the rest of the dataset is unaffected.

Technical detail and evidence

Why it matters, in detail. The override removes $8.98bn from the $1m+ class. If that income is genuine, which fits the cell's internal coherence and ZIP-file dominance suppression, the economically consistent estimate is about 48% with an AGI-based denominator, versus 29.7% published and about 27.0% under the stated rule. If it is erroneous, the posted value is still about 2.7 pp above the stated rule. Either way the cell should be flagged rather than presented as a corrected typo.

Evidence. IRS Historic Table 2, 2012 (12in54cmcsv.csv, WY agi_stub 10): 910 returns and AGI $13.38bn, a mean of $14.70m (vs $4.29m in 2011, $5.99m in 2013, $5.21m in 2014; next-highest state in 2012: NV $4.80m). The cell is internally consistent: total income $13.43bn; itemised deductions $3.72bn including charitable $3.54bn (vs $0.24/0.39/0.47bn in 2011/13/14); taxable income $9.65bn; total tax liability $1.86bn (vs $0.61bn in 2011 and $1.18bn in 2013). 53% of the class's total income lies outside wages, interest, dividends, business income, capital gains, pensions/IRAs and partnership/S-corp income, vs 4-6% in other WY years and 4.8% for the US $1m+ class (scripts/a4_wyoming.py, a5_millionaire_means.py). The IRS ZIP-code file shows WY $200k+ AGI of $6.52bn vs $16.10bn, but its documentation (TY2012 ZIP Code Data guide, sections C-D) says returns 'representing a specified percentage of the total of any particular cell' are suppressed and state totals 'may not be comparable'. In 2012 the ZIP $200k+ AGI is below the state table in all 51 jurisdictions (median -5.1%; -12% to -18% in small states such as SD, MT, ND, WV and VT). SP (Jan 2015, endnote 5, p.31) attributed the surge to genuine IRS income missing from BEA. Stated rule: $4,288,371 x 1.12543 = $4,826,264, giving a top-1% share of about 27.0% (both auditors) vs 29.70% posted. The file's WY 2012 AGI remains $29.36bn = 1.45 x TotalInc. Treating the income as genuine with an AGI-consistent denominator gives about 48% (both auditors); the raw benchmarked value (~70%) is a denominator artefact.

Why this severity. One disclosed state-year out of about 4,900. It moves Wyoming's 2012 level and rank (top-1% share somewhere between about 27% and 48%) but no panel-wide result.

The authors' response. The note (p.3) asserts a typo. SP's own report (Jan 2015, endnote 5) gave the opposite diagnosis. Frank's 2025/2026 releases also replace the value (25.5-25.6%), while SP's 2018 vintage keeps it (69.6%).

9minorconfirmed · newly identified here · internal inconsistency

For the last year in the data, 2012, the state figures were matched to an older version of the national numbers than the one in the file's own US row. That makes the 2012 state figures slightly too low, by about 1.6% for the top 1%.

Technical detail and evidence

Why it matters, in detail. s_{i,2012} is scaled by PS_Aug2013/PS_Jun2015 = 0.9965 (top 10%) ... 0.9415 (top 0.01%) relative to the documented rule. The 2012 state rows are therefore inconsistent with the 2012 US row by 1.6% (top 1%) and 5.9% (top 0.01%).

Evidence. scripts/a1_benchmark.py: 2012 state cells equal SP x PS(Aug-2013)/SP-US with zero error. Under the June-2015 vintage the residual is -0.35% (top 10%) to -5.85% (top 0.01%) for every state. PS 2012 top 1%: 22.464 (Aug 2013) vs 22.828 (US row); the PS vintages are identical for 1917-2011. SP's EPI workbook contains a 'United States (Piketty & Saez)' row identical to TabFig2012prel (SP endnote 6), the likely source of the factor. The largest top-1% shortfall is 0.53 pp (CT 33.01 vs 33.54).

Why this severity. A uniform proportional shortfall in one year. Rankings are unchanged, and 2011-12 state growth changes by at most about 0.5 pp (top 1%).

10minorconfirmed · newly identified here · documentation gap

Three years (1983-85) had no state data from the main source and were filled in using another researcher's figures. The note does not explain how, and the filled-in numbers cannot be reproduced, so they should be treated as rough estimates.

Technical detail and evidence

Why it matters, in detail. The very-top 1983-85 cells are essentially rescaled 1982/1986 cross-sections (rank correlation 1.000). They carry no independent state information and should not be used for annual dynamics around TRA86.

Evidence. scripts/a7_addingup_frank.py and a10_imputation_1983_85.py: published 1983-85 values do not equal corrected Frank (none of the 153 state-years exact for the top 10%, 2 for the top 1%; median gaps 0.5-2.3 pp). 1983 top-10/top-1 values closely follow a forward splice of 1982 values using Frank's state growth (R2 0.99/0.97), and the other fractiles follow the imputed top-10/top-1 growth (R2 0.58-0.99). No exact rule was found; auditor A tested five candidates and none reproduced any of the 153 state-years. The top-0.01% cross-state ranking is identical in 1982 and 1983 and in 1985 and 1986 (Spearman 1.000). Alaska 1985 is infeasible (F5). AGI is missing and N_TaxReturn fractional in all 153 rows.

Why this severity. Three disclosed imputed years (153 state-years). Levels are uncertain by roughly 1-2 pp for the top 1%, and more at the very top, with no effect on long-run patterns.

11minorconfirmed · newly identified here · overclaiming

The note says its numbers are close to another researcher's independent series. That is roughly true for the top 1%, but much less so for the top 10%. And since both draw on the same IRS tables, the match does not prove either is right.

Technical detail and evidence

Why it matters, in detail. The median cross-state correlation of the corrected top-decile series is only 0.64 in 1942-81. Part of the gap is conceptual: Frank's 2015 shares were computed within filers (fn 3). Shared IRS tabulations and BEA-type denominators mean that agreement cannot validate common assumptions.

Evidence. scripts/a7_addingup_frank.py (Frank_All_v97.csv, Wayback capture of Jan 2015; corrected by PS/Frank-US): top-1% mean absolute difference 0.55 pp (1942-81), 0.91 (1986-2001), 2.87 (2002-09) and 0.92 (2010-12); pooled correlation 0.92-0.93 outside 2002-09 (0.70 within). Top 10%: mean absolute difference 1.5-2.0 pp; pooled correlation 0.58 (1942-81) and 0.87-0.88 later; maximum state-year gaps 9-22 pp.

Why this severity. A secondary qualitative claim that is broadly fair for the headline top-1% series. It overstates agreement for the top decile and should not be read as validation.

12minorconfirmed · newly identified here · documentation gap

Several labels in the documentation are wrong. Most importantly, the price variable is described in a way that would lead a user to apply it backwards when converting to today's dollars. None of this changes the published shares.

Technical detail and evidence

Why it matters, in detail. Dividing nominal values by CPI2014, as the text instructs, deflates in the wrong direction by a factor of up to about 290 (16.94^2 for 1917). Units in the July-2015 description were off by a factor of 1,000.

Evidence. CPI2014 is an inflator equal to PS Table A0 col. 8 (1917 = 16.943; nominal x CPI2014 reproduces PS real averages). The note nevertheless calls it a 'Consumer price index, CPI-U-RS series, base 1 in 2014' and says values 'need to be deflated' with it (WID PDF p.2, p.4). The variable list says 'CPI14', but the file has CPI2014. 'Sommeiller-Frank' is used three times for 'Sommeiller-Price' (pp.2-3), and 'The correct' appears for 'corrected' (p.3). The July-2015 author version gives TotalInc/AGI in 'current thousand dollars' and names the variable 'AGI_1000'; the file is in $ millions and uses 'AGI' (corrected in the April-2017 WID version). Frank (2014) is said to update the series to 1916-2012, but its title covers 1916-2005. Sommeiller & Price is cited as the Feb 2014 report (1917-2011), but the data are the Jan 2015 update (1917-2012). Fn 5 omits the coarse 1974 and 1982-88 tables. The Version-2 Excel file is not available at the stated locations. SP's own caution about non-monotone thresholds (19 in their text, 18 in Table A2) and 244 inconsistent interfractile means (EPI 2015 pp.25-26) is not passed on. 572 state-years with returns exceeding tax units are unexplained.

Why this severity. No effect on the published shares; affects usability and the risk of misuse (e.g. dividing by the inflator).

13minorlikely · previously known · methodology

Like the national figures they are tied to, the state figures count income as it appears on tax returns. Tax-law changes in the 1980s moved some business income onto personal returns, so part of the measured rise in top shares is an accounting change rather than a real one.

Technical detail and evidence

Why it matters, in detail. Part of each state's 1979-2007 top-share rise reflects TRA86 reclassification and realization timing. The national literature attributes up to about a third of the fiscal top-1% rise to such tax-base changes (AS 2024), and the state series cannot separate them out.

Evidence. PS national top-1% share incl. gains: 12.67% (1985), 15.92% (1986, a realization spike), 12.66% (1987), 15.49% (1988). A TRA86-era level shift of about 2.8 pp is thus present in the gains-inclusive benchmark, contrary to auditor B's reading. The 1986-89 factors (x1.15-1.22 for the top 1%; x2.1-2.4 for the top 0.01%) transmit this pattern into every state (F1). Auten & Splinter (2024 JPE, p.2181) find that one-third of PS's 1962-2019 fiscal top-1% rise occurred in the years around TRA86 (us-income-08); see also Reynolds (2007; us-income-01), Kopczuk & Zwick (2020; us-income-17), and tax units vs individuals (us-income-09). State tax units are allocated shares of PS's national count (households before 1979, CPS after). There is no break at the 1979 source switch (median |dlog unit share| 0.85% vs 0.77% typical).

Why this severity. The note states its concept explicitly (fiscal income including realized gains, tax units). These critiques concern reading the series as economic inequality and apply equally to the national PS series, so they do not undermine the construction.

The authors' response. Piketty & Saez (2003, pp.3-4, 31) acknowledge TRA86 income shifting. Their response to Reynolds (20 Dec 2006, https://eml.berkeley.edu/~saez/answer-WSJreynolds.pdf) points to wage-only and gains-inclusive evidence. PSZ (2018) moved to national income and equal-split adults.

Claims examined and rejected

Problems one auditor raised that did not survive checking.

Audit notes

What reproduced. Reproduced exactly: (i) the US row = PS TabFig2014prel Table A3 for all fractiles 1917-2012; (ii) all non-imputed state cells 1917-2011 = SP x PS/SP-US (max error 7e-15 pp); (iii) SP's raw state shares from IRS Historic Table 2 with an independent implementation of SP's Pareto rule: exact for all 52 units in 2012 (x1.0100), and within 0.3% for states in 2001 and 2010; (iv) auditor A's factor sizes and trend decompositions; the anomaly counts (Alaska, 1944, pre-1929 CV = 0, Delaware 1929); auditor B's 24 infeasible state-years and their origin; the Frank comparison; and the WID drift (FL/DE fetched independently from the WID bulk archive). Extended: the bracket-collapse uniformity test, using adjacent fine-bracket years (2001, 2010) and the regime-switch correlation test; the IRS ZIP-vs-state comparison for all states; and the composition of the Wyoming 2012 $1m+ cell. Not reproduced: the 2012 state cells under the documented June-2015 vintage (they match the Aug-2013 vintage); the posted Wyoming 2012 value from the stated rule (27.0% vs 29.70%); and the 1983-85 imputation (approximate structure only). Not attempted: SP's 1934-86 state capital-gains coefficients, pre-1997 IRS inputs, BEA/Census controls, and GMMS's national reconstruction.

Limits of this audit. The authors' construction code was not available. The Version-2 Excel file and the original July-2015 data posting were also unavailable, so the audited .dta is the May-2017 WID GitHub copy, which matches the April-2017 note. The unpublished 1982-87 IRS state tabulations and the 1974 Small Area Data were not inspected; coarse-bracket years are taken from Frank's 2026 appendix. The true 2012 Wyoming income cannot be determined without IRS microdata. Geloso & Magness (2020) was read at abstract level, with Wisconsin magnitudes taken from a 2017 working-paper chart. GMMS (2022) was not replicated; its figures come from the corpus entry and auditor B's transcription. The uniformity test uses adjacent fine-bracket years as proxies for 2002-09 and cannot test 1974 or 1982-88 directly. WID drift for states other than FL and DE relies on auditor A's download of the same WID release. SP's 1934-86 state capital-gains scaling (EPI 2015 endnote 17), pre-1997 IRS inputs and the BEA/Census controls were not rebuilt.