Facundo Alvaredo, Chiaki Moriguchi, Emmanuel Saez · World Top Incomes Database / WID.world Technical Note 2012/7

Japan Estimates of Top Income Shares: Update up to 2010

10 problems found: 5 moderate · 5 minor

What the paper claims

The Moriguchi-Saez series of top income shares excluding capital gains is extended annually to 2006-2010.
The 2001-2005 estimates are revised because of downward revisions in National Accounts; the nominal income control is 1.4%-2.7% lower.
The 2005 estimates also change because finer income-tax tabulations became available: from 2005 eight intervals above 50 million yen with an open top bracket from 10 billion yen (before: open bracket from 50 million yen, 1972-2005).
New series: shares of the top 10% and top 0.05% for 1947-2010.
New series including reported capital gains for 1947-2010 (top shares, average incomes, thresholds).

Every problem found

1moderateconfirmed · newly identified here · misrepresentation of source

The note's figures are right, but the main public database that hosts them now shows a different set of numbers. Those numbers include one-off profits from selling land and shares, yet are labelled as excluding them, and in 2009 they switch to another team's series. Anyone downloading 'Japan's top 1% share' there gets figures up to about a fifth too high in the 2000s, far higher in the late-1980s bubble, and a fake drop in 2009.

Technical detail and evidence

Why it matters, in detail. Versus the note's ex-gains series, WID sfiinci992 is higher by +0.9 to +1.9 pp (top 10%) and +0.8 to +1.7 pp (top 1%, +9% to +18%) in 2001-2008, by +26% to +96% for the top 0.01%, and by +2.0 to +6.2 pp for the top 1% in 1969-73 and 1987-91 (e.g. 1990: 12.98 vs 8.05). The 2008->2009 step is a concept splice: -1.76 pp for the top 1% on WID vs -0.15 pp in the note's series (and -0.47 pp on a consistent with-gains basis); -47% vs -3% for the top 0.01%. Through the DINA backcast and co-movement, the same gains spikes plausibly enter WID's pre-tax national income shares for Japan.

Evidence. WID bulk file (2026-10-15) and API pull (2026-10-25) identical on all 806 common cells. 1947-2000: max |WID - workbook with-gains| = 0.012 pp vs up to 6.5 pp from the ex-gains series (japanRESTAT.xls, sheet matlab-incomes47-05indiv, cols 17-23 vs 32-38). 2001-2005: WID = with-gains x IC_prev/IC_rev within 0.005 pp. 2005-2008: adjudicator's with-gains reproduction within 0.07 pp in all groups (e.g. 2006 top 10% 42.76 vs 42.75, top 1% 11.28 vs 11.31). 2009-2010 WID top 1% 9.14/9.00 vs 10.36/10.38 on the AMS with-gains basis. Metadata: 'Tax units are individuals aged 20+. Excludes capital gains...' (countries-variables-metadata, sfiinc_p99p100_992_i). The WTID-era archive (data to 2005) labels Japan 'Estimates exclude capital gains' and holds the ex-gains values (e.g. 1990 top 1% 8.05). DINA: correlation of annual changes in sptincj992 with the with-gains fiscal series, 1980-2008, is 0.98 (slope 0.97) for the top 1% and 0.99 for the top 10%, vs 0.75 with the ex-gains series; top 1% 1990->1992: DINA -4.59 pp, with-gains -4.56 pp, ex-gains -0.93 pp. Scripts: wid_vs_workbook.py, dina_vs_fiscal.py, reproduce_ms.py.

Why this severity. The note's own numbers and labels are correct and reproducible. The defect lies in WID's current presentation, introduced after the WTID era, so it does not undermine the findings themselves. It is nonetheless the most consequential caveat for users: what WID offers as this series is 9-18% too high for the top 1% in 2001-2008, up to 6.2 pp too high in boom years, and has a spurious 2008-09 break. Auditor A rated it major; it is graded moderate here because severity is judged against the note's findings.

2moderateconfirmed · previously known · speculative assumption

To strip out one-off profits from selling assets, the method subtracts 80% of them, a rule borrowed from Canada and never tested for Japan. For the richest 0.01%, where such profits are about half of reported income, changing that rule within a plausible range moves the estimate by up to a quarter. For the top 1% the effect is small.

Technical detail and evidence

Why it matters, in detail. Excluding-gains shares = with-gains share x (1 - k s), with k = 0.8 and s the group's reported-gains share. dShare/dk = -with-gains share x s, so at the top 0.01% (s ~ 0.43-0.62) a plausible k range of 0.6-1.0 moves the level by up to a quarter. Over 2006-2010 the sign of the top-0.01% change flips only at k = 1, which MS argue biases shares downward. Validation is circular because k was tuned to make ex-gains shares not co-move with gains.

Evidence. Codex, 2010 (removal 60/75/80/85/100%): top 1% 9.741/9.567/9.510/9.452/9.278; top 0.1% 2.657/2.539/2.500/2.461/2.344; top 0.01% 0.745/0.680/0.658/0.636/0.571. Adjudicator reproduction, 2005-2010, 60% vs 100% removal: top 1% +/-2.3-4.4%, top 0.1% +/-5.9-12.0%, top 0.05% +/-7.6-15.1%, top 0.01% +/-13.0-24.7% (Claude: same ranges). Top 0.01% change 2006->2010: -9.9% (80%), -4.5% (90%), +2.5% (100%), -17.6% (60%). Workbook note: japanRESTAT.xls, matlab-incomes47-05indiv, row 68 (cell A68). MS (2010) pp. 120-121.

Why this severity. The headline top 1% moves only +/-2-4%. The top 0.1% moves +/-6-12%, and the new top 0.05% and the top 0.01% series move +/-8-25%. The limitation was acknowledged in 2010 and left unquantified in the update. It does not threaten the top 10%-1% levels or any claim the note makes.

The authors' response. MS (2010, p. 121) argue that the ex-gains top 0.1% series is 'fairly stable' during the early-1970s and late-1980s asset booms, so the rule 'is fairly adequate', and that micro data would be needed to improve it.

3moderateconfirmed · newly identified here · methodology

In 2005 the tax office began publishing much more detailed figures on the very richest. The update uses the better data from 2005 on but leaves earlier years on the cruder method, which by the authors' own test overstated the top 0.01% by about 15%. So the before-and-after comparison for that tiny group is not like-for-like.

Technical detail and evidence

Why it matters, in detail. Applying the 2005 overlap ratio (0.682/0.797 = 0.856), the 2004 top 0.01% would be about 0.61-0.65 rather than 0.71. The published mixed-basis series (0.71 -> 0.70 -> 0.76 for 2004-2006) therefore understates the 2004-2006 rise, and pre-2005 levels are overstated by roughly 10-17%, more in gains-boom years. The size of the bias before 2005 cannot be measured, because fine brackets do not exist.

Evidence. Workbook 2005 rows (matlab-incomes47-05indiv rows 63-64 in Excel numbering): ex-gains top 0.01% 0.797 (coarse) vs 0.682 (fine), +16.9%; with-gains 1.238 vs 1.326 (-6.6%); other groups within 1.5%. Adjudicator test collapsing the 2005-2010 brackets above 50m yen: ex-gains top 0.01% +16.2/+15.5/+16.1/+13.4/+9.9/+10.5%, top 0.05% 0 to +1.3%, top 0.1% -0.5 to 0% (Claude: +9 to +15%). In coarse years the implied gains share is identical for the top 0.05% and 0.01% (and for the top 0.1% in 1989-91: 75.7-78.3%), whereas fine 2005 data give 60.7% for the top 0.01% vs 45.6% for the top 0.05%. Boom-year spikes in the coarse-era ex-gains top 0.01% (0.44 -> 0.86 in 1972-73; 0.40 -> 0.86 in 1986-90) are consistent with incomplete removal.

Why this severity. Affects the level of a secondary series, the top 0.01% (by about 10-17%), and its comparability across the 2004/2005 seam. It possibly affects the new top 0.05% series in boom years and the top 0.1% in 1989-91. It is partly disclosed (note 3 attributes the 2005 change to finer data) and leaves the 2005-2010 top 10%-0.1% estimates untouched.

The authors' response. MS (2010, p. 118): 'our extrapolation method within the top bracket in fact provides a fairly close (within 5 per cent) estimate for year 2005.' Note 3 attributes the 2005 revision partly to the finer tabulations.

4moderatelikely · newly identified here · internal inconsistency

For the 1950s and early 1960s, a correction that stops wage earners being counted twice was applied to the main series but not to the companion series that includes asset-sale profits. That companion series is therefore about a tenth too high for those years and shows a sudden, fake drop in 1963.

Technical detail and evidence

Why it matters, in detail. With-gains top shares for 1951-62 should be multiplied by about 0.92 (top 10%), 0.91 (top 1%) and 0.96 (top 0.1%) to be consistent with the ex-gains series and with post-1963 values. As released they overstate the 1950s top 10% by about 2.5-3.4 pp and show a spurious fall of 2.6 pp (top 10%) and 0.8 pp (top 1%) between 1962 and 1963.

Evidence. scripts/withcg_consistency.py on japanRESTAT.xls (matlab-incomes47-05indiv; composition block, 'K gains' columns): residual factor 1.000 in 1947-50 and 1963-2005, and the constants above in 1951-62. Workbook top-10% gains share 0.23% (1951), 0.13% (1953), 2.6% (1962). WID top 10% 1962->1963: 34.48 -> 31.73 (-2.75 pp) vs ex-gains 31.09 -> 30.97 (-0.12 pp); top 1%: 10.03 -> 9.01 vs 8.68 -> 8.50. 1950->1951 top 10%: with-gains 31.70 -> 32.90 vs ex-gains 31.66 -> 30.27.

Why this severity. Confined to the companion with-gains series and to 1951-62, so the ex-gains headline series is unaffected. The error is 8-10% at the top 10%-1% (about 2.5-3.4 pp for the top 10%), creates a false 1963 break, and sits in the series WID now distributes as Japan's fiscal-income series.

5moderatelikely · previously known · methodology

Most Japanese employees never file a tax return, so the method stitches tax returns together with an employer survey of wages. The stitching double-counts some high earners and pads the top with assumed civil servants. Fixing this lowers the very top shares by roughly 5% and the top 1% by up to about 3%, a small change that leaves the overall picture intact.

Technical detail and evidence

Why it matters, in detail. Upward channels: the x1.1 uplift applied to the top wage brackets, and the spurious non-filer residual created by subtracting net filer wages from gross survey wages. The downward channel is that filers' wages are measured net of the 5% + 1.7m yen employment-income deduction. Using NTA Table 7 for the filers' gross-salary distribution, a gross income concept and no uplift above 15m yen lowers top-0.1% shares by about 5-7% (2.52 -> about 2.36 in 2010) and the top 1% by about 1.5-3%.

Evidence. Adjudicator reproduction (reproduce_ms.py): residual 149,653 persons / 4.64 tn yen above 20m yen in 2006 (Claude: 150k / 4.6 tn). NTA Table 7: 130,397 taxpaying filers with gross salary above 20m yen in 2006, vs 222,866 in the survey. Variants vs baseline, 2005-2010. No x1.1 above 15m yen: top 1% -2.0 to -2.2%, top 0.1% -1.5 to -1.8%. Filers' wages taken from Table 7 (gross), with filers' incomes grossed up and no x1.1 above 15m: top 1% -1.4 to -3.1% (2009: -6.9%), top 0.1% -5.1 to -6.7%, top 0.01% -2.9 to -7.4% (Claude's equivalent: -1.6% / -6.1% / -4.3%). A consistent gross conversion through rank equivalence instead gives +1.2 to +1.4% (top 1%) (Codex 2010 gross variant: +1.1%). Codex x1.0/x1.2 uplift: top 1% 9.29/9.72 vs 9.51. Trends unchanged in all variants.

Why this severity. Small for the headline top 1% (about 0 to -3%, sign depending on the correction) and the top 10%, but noticeable for the top 0.5%-0.01% (about -3% to -7%) under the most data-based alternative. The direction was acknowledged by MS, and the update did not use the available gross-salary table.

The authors' response. MS (2010, p. 119): the uniform 10% uplift 'probably introduces a slight upward bias', and if the rank assumption fails the method 'would overstate the number of high wage filers ... and hence create small upward bias'. Iguchi et al. (2025), co-authored by Moriguchi, replace both devices.

6minorconfirmed · newly identified here · version instability

The note says the 2001-2005 figures changed only because official national totals were revised. For 2002 they also changed for another, unexplained reason. The change is small and probably a correction.

Technical detail and evidence

Why it matters, in detail. The 2002 revision combines a -2.13% denominator change with a numerator cut of 244-273 bn yen for the top 5%-0.1%, coming from a higher removed-gains share. It does not affect other years or the extension.

Evidence. revision_checks.py: revised minus predicted (pp), 2002: top 10% -0.110, top 5% -0.106, top 1% -0.106, top 0.5% -0.101, top 0.1% -0.093, top 0.05% -0.080, top 0.01% -0.036 (numerators -0.3% to -6.0%); other years within 0.017 pp. Implied gains shares 2002: top 1% 8.9% -> 10.3%, top 0.1% 19.7% -> 24.3%, top 0.01% 23.8% -> 30.4% (2001: 22.3%, 2003: 25.4% for the top 0.1%). WID with-gains 2002 = workbook x IC ratio within 0.005 pp.

Why this severity. At most 0.11 pp in a single year; probably a legitimate data correction; a change-log omission rather than an error.

7minorconfirmed · newly identified here · documentation gap

The note does not list all its ingredients or share its spreadsheets. Others can nonetheless redo the main calculation from public tax tables and get almost the same answers.

Technical detail and evidence

Why it matters, in detail. Exact replication of the controls and of the very-top cells is impossible without the authors' files. The extension is approximately replicable from public NTA tables (top 10%-1% within 0.6%).

Evidence. Note pp. 1-2 (PDF pp. 2-3). Auditors' searches of RePEc/WID/WTID archives (Codex results/material_inventory.json; Claude report section 2). The authors' public 2007 workbook (japanRESTAT.xls) documents 1947-2005 only.

Why this severity. Typical of WTID update notes. Three independent reconstructions from the cited sources nonetheless match within 0.3-3%, so replication is approximate but feasible. Auditor B rated this moderate.

8minorconfirmed · newly identified here · internal inconsistency

The count of adults used to define 'the top 1%' jumps in 2010 because of a census update rather than real growth. The effect on the results is negligible.

Technical detail and evidence

Why it matters, in detail. A 0.6% jump in the 2010 adult base enlarges every top group by 0.6% in headcount, raising 2010 shares by roughly 0.2-0.3% relative to a smooth population series.

Evidence. check_tax_units.py (e-Stat population estimates): 2005 note 103,830k vs official 103,590k; 2006-2010 equal official 1 October estimates within 2k; adults +704k in 2009->2010 vs +60k in 2008->2009 (census rebasing).

Why this severity. The effect on shares is below about 0.3%.

9minorlikely · newly identified here · documentation gap

The 'total income' figure that every share is divided by cannot be rebuilt from the published recipe. Following the recipe would make all the shares about 2% smaller, which is a small difference.

Technical detail and evidence

Why it matters, in detail. All shares scale with 1/denominator. The documented definition with the 2012 vintage would lower every 2001-2010 share by about 2% (top 10% 2010: 40.50 -> about 39.8) and could create a level seam of similar size at 2000/2001 against the older-vintage pre-2001 controls.

Evidence. esri2010-nationalincome.xls, sheet Amount: 2010 = 206,070.6 + 22,904.7 + 1,902.0 + 10,427.6 = 241,304.9 bn yen vs 237,053 (+1.8%); 2005: 255,787 vs 250,478 (+2.1%); 2001: 268,367 vs 263,347 (+1.9%). Household property income 2005: 16,488.7 (FY2005 report) vs 22,976.2 (FY2010 report). The older vintage does not reproduce the previous control exactly either (2005: 254,031 vs 256,749). Codex and Claude searched component combinations: no exact match (best +/-0.5-0.7%).

Why this severity. A proportional level uncertainty of about 2% on all shares (e.g. 2010 top 1% 9.51 -> about 9.34 under the documented definition), with no effect on trends; a narrower concept would be defensible.

10minorplausible · previously known · methodology

Some savings income, mainly interest and many dividends, never shows up on tax returns but is counted in the national total. If the rich receive much of it, their true share is somewhat higher than estimated. The authors judged this effect small.

Technical detail and evidence

Why it matters, in detail. If missing dividends were distributed like reported top incomes, top 1% levels would be understated. The extreme upper bound, assigning all household dividends to the top 1%, is about +2.9 pp in 2006. Survey evidence suggests the realistic effect is far smaller.

Evidence. NTA 2006 Table 2 appendix (dividends 551,981 m yen); ESRI FY2010 report (household dividends 7,194.5 bn yen in 2006); MS (2010) pp. 97-98 (upper bound +2.8 pp on the 1999 top 1%) and Appendix 3D.

Why this severity. This is a known, disclosed limitation. The likely bias runs downward on top-share levels and is probably small given the 1999 survey evidence, but it is unquantified for 2006-2010.

The authors' response. MS (2010, pp. 97-98): even assigning all interest and dividends to the top 1% would raise its 1999 share only from 7.8% to 10.6%, and the 1999 NSFIE suggests these components are 'not particularly concentrated at the top'.

Claims examined and rejected

Problems one auditor raised that did not survive checking.

Iguchi, Ueda, Sato & Moriguchi (2025) find 2008-2010 levels 1-7% below the note, an external check corroborating that the merge procedure biases top shares upward (I3). — Not valid as corroboration. Iguchi et al.'s denominator is 270.2 tn yen (2020 prices) in 2008 against the note's 250.4 tn yen nominal, roughly 4-8% larger, so their implied top-1% numerator (about 24 tn yen) is about the same as the note's. Iguchi et al. (p. 24) themselves warn the levels are not comparable. The merge concern (F5) is retained on other evidence.
The untested 80% rule 'decides the sign of the top-0.01% change between 2006 and 2010' (I4). — Overstated. The sign flips only under 100% removal (+2.5%), which MS argue biases shares downward; at 60-90% removal the decline persists (-17.6% to -4.5%). The note makes no trend claim. The level sensitivity is retained in F2.
The WID mislabelling/splice is a MAJOR issue for this work (I1). — Severity not accepted; the issue itself is confirmed and retained as F1 (moderate). The note's estimates and labels are correct; the error is in WID's later presentation. Under the rubric, severity is judged against the note's own findings, which are not changed.
The current WID mismatch reflects a documented successor vintage (Iguchi et al. 2025), so it is not an issue for this note (checks passed 11; WID values 'differ from the note' before 2009 but labelling not adjudicated). — Rejected for 1947-2008. Those WID values are not the successor's but the authors' WITH-gains series under an 'Excludes capital gains' label (max deviation 0.012 pp from the authors' workbook); only 2009 onward comes from the successor (F1).
Full historical additions (C4) and gains-inclusive outputs (C5) cannot be checked, warranting a MODERATE documentation-gap finding (I2). — Largely rejected. The with-gains series for 1947-2008 can be recovered and checked (WID = authors' public workbook + note's revised controls; the adjudicator's reproduction matches 2005-2008 within 0.07 pp), and the 1947-2005 top-10%/0.05% values are in the same public workbook. These checks were feasible and revealed F4. The residual documentation gaps are minor (F7, F8).

Audit notes

What reproduced. Reproduced: (1) Table 1, decoded independently; its 'previous' columns match MS (2010) and the authors' 2007 workbook exactly. (2) The control revision (-1.38% to -2.72%) and the 2001/2003/2004/2005 revisions, exactly. (3) The 2006-2010 shares, rebuilt from NTA Table 2 (taxpayers by total-income bracket) and the private wage survey (Table 3, full-year employees, x1.1), with rank-equivalence filer wages, Piketty-Saez Pareto interpolation and 80% gains removal. Using the note's controls and adult counts, top 10%/5% are within 0.3% in every year (top 5% exact to two decimals in 4 of 6 years), top 1% within 0.2-0.6%, top 0.1% within 0.4-1.3%, top 0.01% within 0-3.2%. This agrees with Codex (top 1% within 0.003 pp) and Claude (top 1% within 1.4%). (4) The with-gains series for 2005-2008, reproduced within 0.07 pp of what WID shows, which identified WID's series as the AMS with-gains one. Not reproduced: the revised income control itself (the documented definition gives totals about 1.8-2.4% higher; no exact component combination found by either auditor), the 2002 numerator change, and the frozen 2012 WTID release (not recoverable). Sensitivity runs (adjudicator, confirmed against both auditors): gains removal 60-100% moves the top 1% by +/-2-4% and the top 0.01% by +/-13-25%; coarse brackets raise the 2005-2010 top 0.01% by 10-16%; merge corrections move the top 1% by about -3% to +1.4% and the top 0.1% by about -7% to +2%.

Limits of this audit. No author code, spreadsheet or frozen 2012 WTID release exists. The three reproductions follow the published description and match within 0.2-3%, but they are not the authors' implementation, so sub-percent differences at the very top cannot be attributed. The income-control construction remains inferred. The pre-2005 coarse-bracket bias cannot be quantified for 1972-2004 because fine brackets do not exist; it is extrapolated from the 2005 overlap and 2005-2010 tests. The claim that WID's DINA series for Japan inherits the with-gains fiscal series rests on statistical co-movement and metadata, not on WID code. That the 2012 release contained the 1951-62 with-gains values is inferred from the identity of WID's values with the authors' workbook. WID data and metadata were read on 15-25 October 2026 and may change. NTA Table 7 for 2009-2010 was replaced by the NTA in 2013 and looks anomalous for 2009. Iguchi et al. (2025) figures were partly read from charts. The authors were not contacted. No Japan-specific published critique exists in the critique corpus.