Thomas Piketty, Li Yang, Gabriel Zucman · American Economic Review 109(7): 2469-2496

Capital Accumulation, Private Property, and Rising Inequality in China, 1978-2015

16 problems found: 1 major · 9 moderate · 6 minor

What the paper claims

China's national wealth-national income ratio roughly doubled; the rise is 'explained by a combination of high saving rates and a gradual rise in relative asset prices' (saving 50-60% of the rise).
Public property declined sharply; China is a mixed economy with strong public ownership of corporations.
Income inequality rose sharply (pre-tax national income, equal-split adults).
Sharp upward revision of official (survey-based) inequality using tax data and national accounts.
China moved from Nordic-like inequality in the late 1970s toward, but still below, US levels.
Income inequality stabilized after about 2006.
Wealth concentration rose sharply.
Growth incidence 1978-2015: all groups gained, the top much more.

Every problem found

1majorlikely · newly identified here · methodology

The paper says the richest tenth's share of Chinese wealth jumped from 40% to 67% in twenty years. But the early years come from one survey and the later years from a different one, and the correction applied to the surveys was quietly changed from year to year, which made the rise look smooth. Wealth very likely did become more concentrated; how much and when is far less certain than the paper suggests.

Technical detail and evidence

Why it matters, in detail. C7's trend is identified only across the CHIP->CFPS switch. CHIP is internally comparable (+11 pp in the top decile 2002->2013, KLW). Around 2012-13 the two surveys differ strongly (Gini 0.62 vs 0.74). So part of the +17.5 pp 2002->12 rise in the PYZ series is plausibly a survey effect. Translating KLW's change into PYZ's units gives roughly +12 pp (2012 ~61%); Auditor A estimates ~7 pp of overstatement. The undisclosed factor ramp (1.6->2.1 in 2004-09) generates ~90% of the 2003-09 increase and hides the 2010 break. The 1995 factor is ambiguous: 1.0 under the authors' NA-consistency rule, 1.5 per the text. That puts the 1995 level at 35-44%, the 2015 level at 58-72% (factors 1.0-2.0; ~62% on a like-for-like survey basis), and the 1995-2015 rise anywhere from ~18 to ~32 pp, not a point estimate of 26.6 pp.

Evidence. Paper p. 2479 and fn 26-28; appendix B12 pp. 49-50 ('we choose upgrade factor=1.5 for year 1995, 2002, 2010, and 2012'). Gpinter/WealthTabulations/BenchmarkWealthTabulations.xlsx factor cells: '1995'!J11=1.3; '2002','2003'!J11=1.5; '2004'-'2009'!J11=1.6,1.7,1.8,1.9,2.0,2.1; '2010'!J28=1.3; '2011'-'2015'!J32=1.5 (scripts/wealth_year_factors.py). Sheets 2003-2009 rescale the CHIP-2002 tabulation (e.g. '2006'!H5='2002'!H5*T0!L26/T0!L22); the file's ReadMe instead says '2003-2009: weighted average between 2002 and 2010'. Corrected survey mean / national-accounts mean: 1.096 (1995), 1.008 (2008), 1.034 (2009); Variants/WealthTabulations2002Variant.xlsx ReadMe: '2002120 and higher: logically inconsistent, average wealth > national accounts'. Published top-10% shares (gperc_w_China.dta = Data3!AI): 40.8 (1995), 49.0 (2002), 49.0 (2003), 50.6/52.3/53.9/55.8/56.9/58.2 (2004-09), 62.8 (2010), 66.5 (2012), 67.4 (2015); the authors' own 2002 variants (Data8) at factors 1.6-2.0 give 50.6/52.1/53.5/54.8/56.0. 1995 share by factor (adjudicator approximation, validated 40.7 vs 40.8 at 1.3): 35.2% at 1.0 (NA-consistent), 43.8% at 1.5. Knight, Li & Wan (2016, Oxford Econ. WP 816) p. 11 and Table 4: CHIP 2002->2013 national top decile 37%->48%; p. 7: net housing 'based on respondents' reported values (of both owner-occupied and other houses) in each year'. Luo & Chen (2021, China Economic Quarterly International 1(4): 295-307) abstract: wealth Gini 0.619 (CHIP 2013) vs 0.736 (CFPS 2012). PYZ factor-1.0 tabulations: CHIP 2002 39.4% -> CFPS 2012 57.6% (Data8).

Why this severity. Does not affect C1-C5. For C7, defensible alternatives put the 1995-2015 rise (26.6 pp) anywhere between about 18 pp (the stated constant factor plus a like-for-like 2002-12 survey comparison, i.e. ~30% smaller) and about 32 pp (an NA-consistent 1995 factor with the CFPS-based 2015 level). The published 2003-09 path is constructed, not measured. Wealth concentration still rises sharply in every scenario, so C7 is weakened rather than reversed.

The authors' response. None located. In the paper (fn 26; appendix B3, B12) the authors justify the survey choice (CHIP 2008 not nationally representative; CHIP 2013 'has no information on housing values') and call the factor's effect second order (fn 28). Neither text discusses year-varying factors.

2moderateconfirmed · newly identified here · internal inconsistency

The paper says high saving explains over half of why Chinese wealth grew so much faster than income. Its own numbers say otherwise: saving was huge but only kept wealth growing about as fast as China's booming income. The extra rise came from the prices of housing, land and shares rising faster than other prices.

Technical detail and evidence

Why it matters, in detail. With g ~ 8.3% and s ~ 0.35, s/g ~ 4.2 ~ beta1978, so volume accumulation is consistent with a roughly stable beta. Under the additive counterfactual (observed saving flows, zero residual gains) beta2015 = 3.70. Under the multiplicative accounting, saving net of growth contributes -0.79 log points and real gains +1.44. Saving's contribution to the change in beta is therefore ~0% (additive), and at most ~14% under a steady-state s/g counterfactual, against 50-60% claimed. The 'gains' are a residual (relative prices of housing, land and equity, plus measurement error, including farmland valued at 6x output and unlisted equity at book). This puts China's rising beta squarely in the territory of the Stiglitz (2015) wealth-residual and Bonnet et al. (2014) house-price critiques.

Evidence. NA workbook (v2018_na/PYZ2017NationalAccountsData.xlsx) A42!B26:J27: beta1978 3.7222; beta2015 7.1046 = initial wealth 0.1979 + cumulated saving 3.5068 + residual gains 3.3999 (saving = 49.4% of beta2015). A43!B26:F27 (from A47 row 67): g 8.254%, g_w 10.162%, g_ws 5.956%, q 3.969%; E27 = 5.956/(5.956+3.969) = 60.0%. Adjudicator (scripts/macro_checks.py): saving-only beta growth (1.0596/1.0825)-1 = -2.12%/yr; ln(7.105/3.722) = 0.646 = -0.794 (saving net of income growth) + 1.440 (real gains); income-weighted s/g = 0.346/0.0825 = 4.19. Piketty & Zucman (2014, QJE 129(3)) describe the same decomposition as 'accumulation of national wealth' (Sec. V.B, pp. 1285-1286) but say 'about 40% of the rise of beta' is due to capital gains in Sec. VI (p. 1292).

Why this severity. The measured beta path (C1's numbers) is unaffected, but the paper's stated explanation of it is reversed: high saving only kept wealth growing about as fast as 8% income growth, and the doubling of beta is a valuation effect. It is a secondary, interpretive result, so moderate.

3moderateconfirmed · newly identified here · speculative assumption

In 1978 more than half of China's measured wealth is farmland that could not be bought or sold, so its value had to be assumed. Reasonable alternative assumptions taken from the authors' own sources put 1978 wealth at anywhere from 2.9 to 5 times national income instead of 3.7. Wealth rose relative to income either way, but whether it 'doubled' or rose by half depends on this guess.

Technical detail and evidence

Why it matters, in detail. beta1978 = 3.72 + 1.99(m/6 - 1) and beta2015 = 7.10 + 0.51(m/6 - 1), so delta-beta = 3.38 - 1.48(m/6 - 1). Non-farm wealth rises ~3.8x relative to income (1.73 -> 6.59) whatever the multiple. The 'doubling' of total beta mixes that rise with a farmland component mechanically tied to agriculture's falling share of output. 6x is a mid-range, legally anchored choice, so this is uncertainty, not demonstrated bias.

Evidence. Main workbook Data2!D (farmland, x NI): 1.992 in 1978 (53.5% of beta 3.722); 0.513 in 2015 (7.2% of 7.105). Appendix A12 pp. 11-14: Table 2 (2000: UN 8,380bn = 3.4x output; World Bank 24,235bn = 9.9x; Method I 6x; Method II 5.6x) and Table 1 (value/output: USA 7.1, France 3.7, Spain 12.0, Poland 4.3, Ireland 9.5). Adjudicator rescaling (scripts/farmland_multiple.py): beta1978 = 2.86/3.06/3.59/3.72/4.39/5.02 at 3.4/4/5.6/6/8/9.9x; beta2015 = 6.88-7.44; ratio 2.41/2.27/1.97/1.91/1.66/1.48; public share 68.6-69.6% (1978), 31.1-31.8% (2015).

Why this severity. The rise in beta survives in every case, but its size runs from 72% to 119% of the benchmark rise under the authors' own cited alternatives (about +/-15% for a narrower 4-8x band). That is a noticeable magnitude uncertainty for C1, with no known direction of bias.

The authors' response. Appendix A12 argues that 6x lies between the UN and World Bank estimates, criticises both (UN values flat 1990-2008; World Bank implies a 72% land-rent share of farm value added), and cites international value/output ratios of 3.7-12.

4moderateconfirmed · newly identified here · circularity

The paper offers a market-to-book ratio of about 1 for Chinese companies as evidence that its valuations are sound, and even as a sign of strong shareholder rights. But for the great majority of companies, those not on the stock market, the authors simply set that ratio to 1, so it proves nothing. China's total wealth could plausibly be about a tenth higher or lower.

Technical detail and evidence

Why it matters, in detail. Aggregate Q = u*Q_unlisted + (1-u)*Q_listed, with u = 0.83-0.97 and Q_unlisted = 1 by assumption, so comparing China's 'Q' with US/UK Q (mostly market-priced) is not like-for-like. Valuing unlisted equity at a discount to China's high listed Q (e.g. 1.3) gives beta2015 ~ 7.6 and a ~14% larger 1978-2015 rise. If SOE equity is worth less than book (e.g. 0.7), beta2015 ~ 6.6 and the rise is ~14% smaller. The public share in 2015 stays in the 29-34% range.

Evidence. Appendix A14 pp. 17-18 ('For unlisted companies, we assume that the market value of equity is equal to the book value of equity'). NA workbook A26 rows 26 (1993) and 48 (2015), columns P-X: unlisted share of book equity 97.32% / 82.54%; listed Q 2.765 / 1.840; aggregate Q 1.047 / 1.147 (= Data2!AX). Adjudicator sensitivity (scripts/ownership_checks.py), other components fixed: unlisted Q 0.5/0.7/1.3/1.5/1.84 gives beta2015 6.30/6.62/7.59/7.91/8.47 (-11%/-7%/+7%/+11%/+19%) and a public share of 27.1/28.9/33.6/34.9/36.9%.

Why this severity. The argument for Q ~ 1 and its institutional reading are circular. beta2015 is uncertain by about +/-7-19% depending on how unlisted equity is valued, in either direction. The rise in beta and the public-share decline survive.

The authors' response. Appendix A14 discloses the book-value assumption; p. 2483 adds that cross-country differences in Q 'may also reflect data limitations'.

5moderateconfirmed · newly identified here · version instability

The published tables were produced from a later, never-posted version of the data. For broad groups like the top 10% the difference is tiny. But the figures for the very richest (the top one-in-100,000) doubled between versions without explanation and cannot be checked.

Technical detail and evidence

Why it matters, in detail. Above the top 0.1% the income distribution is a generalized-Pareto extrapolation above one tax threshold (~P99.5 of urban adults). It is combined, through a Gumbel copula, with non-fiscal income allocated by the Hurun-anchored wealth distribution. Lowering theta from 3 to 2 means less dependence and should lower top shares, so it cannot by itself double the top-0.001% share. Other undocumented changes must be involved, including a revised fiscal stage (+1.1 pp T1 in 2015). Printing 164,188,917 yuan is false precision.

Evidence. v2018_dist/StataFiles/gperc_y_China.dta 2015 (adjudicator integration, scripts/gperc_shares.py): B50 14.83, M40 43.74, T10 41.43, T1 13.92, T0.1 6.30, T0.01 3.20, T0.001 1.36 = WP Table 2 (14.8/43.7/41.4/13.9/6.3/3.2/1.4; top-0.001% mean 78,833,979). AER Table 1 (p. 2484): 15.3/43.6/41.1/14.5/7.3/4.6/2.8; top-0.001% mean 164,188,917. Main workbook Data3!FC43:FF43 = 15.26/43.61/41.13/14.46; chart sheet F4a plots Data3!B/D/F (= FC/FD/FE). Fiscal stage: Data3!EB43 = 12.48% (AER p. 2484 '12.5 percent') vs gperc_yf_China.dta 11.42% (WP '11.5%'). No posted file among the 359 .dta files with 2015 top-bin rows contains the AER top-0.001% threshold (35.2m) or mean (164.2m) (scripts/scan_top_2015.py). AER Table 2: top 0.001% +2,529% (WP Table 3: +3,817%), non-monotone across the top groups (9.4/9.5/9.2%/yr). WID_data_CN (2026) sptincj992 p99.999p100: 0.32% (1978), 1.37% (2015). openICPSR E116194V1 returned Cloudflare 403; a Wayback snapshot lists a 647.8 MB PYZ2017NationalAccountsData.zip (2019-12-07).

Why this severity. Headline shares barely change between versions (T10 41.4 vs 41.1; B50 14.8 vs 15.3), so C3-C5 are unaffected. The extreme-tail statistics in Tables 1-2 and Fig. 9 (C8's 'multiplied by more than 26') cannot be verified, moved by factors of 2-3 between versions, and conflict with the authors' own WID series.

The authors' response. None located. The AER states theta = 2 (p. 2478); the inaccessible journal deposit may hold the final files.

6moderateconfirmed · documentation gap

The first and last years of the 1978-2015 series are mostly filled in from neighbouring years rather than measured, even though the text says every year's survey data were used. This barely changes the headline numbers, but it means the claim that inequality stopped rising cannot be tested for the final years.

Technical detail and evidence

Why it matters, in detail. The 1978 level combines 1980/81 shapes with the hard-coded urban/rural ratio (F3) and the transplanted tax correction (F9). The 2015 level combines 2012/2014 shapes, the frozen 2006-10 tax factor, a frozen fiscal/national-income ratio and a smoothed urban/rural gap. After 2012 only the wealth-based non-fiscal component and population weights vary, so any post-2012 change in survey-measured inequality is excluded by design.

Evidence. AER p. 2475 and fn 12; fn 33 discloses only the frozen 2011-15 tax correction. Tabulations/rawdata_bysheet_urbanChina_raw.xlsx has sheets 1981-85, 1987-91, 1993, 1995-2014. Its ReadMe row 13 says thresholds 1987-2014 were approximated, and for 2013-2014 'we approximate bottom decile and top decile using year 2012'; Parameters!F47:H48 (2013-14 b*, p*) are identical to 2012. rawdata_bysheet_ruralChina_raw.xlsx has sheets 1980, 1985, 1990, 1995, 1998-2000, 2002-2012. StataFiles/dofiles/do1_gperc_China_urban_yfraw.do lines 61-82 (1978-80 <- 1981; 2015 <- 2014). do2_gperc_China_rural_yfraw.do lines 62-90 (2013-15 <- 2012; 1978-79 <- 1980; interior years interpolated). PYZ2017DistributionalData.xlsx TB2!H41:H43 = H40; TB2!M41:M43 = 3-year averages (3.53/3.50/3.50) vs observed P41:P42 = 3.14/3.08. gperc_ys_China_raw.dta: 2014 and 2015 identical (T10 30.97, T1 6.57).

Why this severity. Endpoint levels change little (2012 revised series 41.3/15.3 vs 2015 41.1/15.3), so C3's numbers stand. But 1978 and 2013-15 are not observed distributions, the stated data coverage is inaccurate, and post-2012 stability (part of C6) holds by construction. That is an undisclosed caveat on secondary results.

The authors' response. Fn 33 discloses the frozen 2011-15 tax correction and warns that the post-2006 stabilisation 'could partly reflect data limitations'.

7moderateconfirmed · previously known · misrepresentation of source

Surveys miss very rich people, so the authors scaled up the wealth reported by the top survey group by 50%. That choice largely sets how much the top tenth owns. Reasonable alternatives give anywhere from about 62% to 69% rather than 67%, so the claim that China is nearly as unequal as the US is less firm than stated.

Technical detail and evidence

Why it matters, in detail. Raising the whole top-bracket mean mechanically raises the top-decile share. The calibration targets are loose: Xie & Jin's household-unit 62.1% is a different estimand from PYZ's equal-split-adult share. Their mean wealth per adult (174k yuan) does support a level correction of about this size, so the factor is not clearly mis-set. The 2015 level is best read as ~62-69% (factors 1.2-1.7) rather than 67%, placing China between France (55% in the authors' own file; 59% in the current WID) and the US (72%).

Evidence. AER p. 2479 and fn 28; appendix B12 p. 50. BenchmarkWealthTabulations.xlsx '2012'!C34 = J32*H34 (bracket from p = 0.9027, threshold 280,000; bracket mean 762,907 -> 1,144,360). Main workbook Data8: 2012 T10 57.6/61.7/63.5/66.5/69.1/72.3% at factors 1.0/1.2/1.3/1.5/1.7/2.0; 2002: 39.4/43.6/45.6/49.0/52.1/56.0%. Variants/WealthTabulations2012Variant.xlsx ReadMe A10 ('2012118 and higher: logically inconsistent'). Xie & Jin (2015, Chinese Sociological Review 47(3)) Table 1 p. 207 (household P90 687k -> 692k yuan; mean 322k -> 422k), Table 2 p. 209 (62.1%, 35.3%), p. 208 (174,000 yuan per adult). PYZ 2012 corrected mean per adult 176,485 ('2012'!B3).

Why this severity. Within plausible factors (1.2-1.7) the 2012/2015 level moves by about +/-5 points, which matters for the comparison 'China 67% vs US 72% vs France 50%'. The 2002-12 rise is stable under a common factor (16.3-18.2 pp), so the trend is less affected (but see F1).

The authors' response. The paper shows variants for factors 1-2 (appendix Figs. B12a-b) and argues the benchmark is consistent with published Chinese estimates (fn 28).

8moderatelikely · newly identified here · speculative assumption

To measure inequality in 1978 the authors had to say how much more city dwellers earned than villagers, and they assumed twice as much. Official statistics say about 2.6 times as much. Using the official figure makes 1978 China somewhat less equal, so the increase in inequality since then is about a fifth smaller, though still large.

Technical detail and evidence

Why it matters, in detail. Raising R from 1.94 to 2.57 widens the gap between the 18%-urban and 82%-rural populations in 1978. National B50 falls by 2.4 pp and T10 rises by 2.0 pp in the fiscal distribution; the final pre-tax series behaves similarly, since non-fiscal income was ~5% of national income in 1978. The headline would read ~25% -> 15% (bottom 50%) and ~29% -> 41% (top 10%). The authors' own later WID vintage moved in this direction (1978 B50 24.7%, T10 28.0%). Bottom-50% real growth becomes larger (~+455% vs +409%). The Nordic comparison survives (Finland 1978: B50 25.2%, T10 27.1%).

Evidence. v2018_dist/PYZ2017DistributionalData.xlsx TB2!P6:P8 = 2 (constants); TB2!M6=AVERAGE(M7:M8) ... M9=AVERAGE(M10:M11), giving 1.940 (1978) and 1.916 (1981) against observed P9 = 2.240 (TB4 urban 500.4, TB5 rural 223.44 yuan per capita); TB5!G13 = 133.57 (rural 1978). TB1 applies one adult/population ratio (0.539) to urban and rural, so the per-adult ratio equals the per-capita ratio. NBS China Statistical Yearbook 2009, Table 9-2 (https://www.stats.gov.cn/sj/ndsj/2009/html/J0902e.htm): urban disposable income per capita 343.4 (1978), 477.6 (1980), 739.1 (1985); rural net income 133.6, 191.3, 397.6. Adjudicator merge of the authors' corrected urban and rural fiscal g-percentiles (scripts/urban_rural_merge.py; validates at B50 27.52 / T10 26.28 vs the authors' 27.57 / 26.11): R = 2.57 gives 1978 national fiscal B50 25.11%, T10 28.26%, T1 6.29%. Current WID China (2026 vintage) 1978 pre-tax: B50 24.73%, T10 27.98%.

Why this severity. The 1978-2015 fall of the bottom-50% share shrinks from ~12.0 to ~9.6 points (-20%) and the rise of the top-10% share from ~14.3 to ~12.3 points (-14%). This is a noticeable but not transformative change; 1978 China stays Nordic-like.

The authors' response. None located; the constant is not justified in the files, and the text (p. 2486) presents the 1978 ratio as a result.

9moderatelikely · previously known · speculative assumption

China only has tax data on high earners from 2006-10, so the authors assumed the rich were under-counted the same way in 1978. For the very richest this implies that a few thousand people in Mao-era China earned hundreds of times the average income, which is hard to believe. The paper's striking growth figures for the tiniest top groups are therefore not reliable; its broader conclusions about rising inequality are not affected.

Technical detail and evidence

Why it matters, in detail. The correction multiplies 1978 top-0.001% fiscal income ~26x (3,040 -> 78,392 yuan). This transplants a b** ~ 2.6-3.0 tax tail from 2006-10 onto a 1978 distribution with no private enterprise. Growth rates for groups above the top 1% (x25-x39 across the AER and WP versions) are driven by this assumed 1978 tail; a 1978 top-0.001% share of 0.3-1.0% is untestable and almost certainly too high. For T10 the constant profile adds +4.4 pp (1978) vs +7.0 pp (2015). The sign of the true trend bias depends on how under-reporting evolved; marketization makes rising under-reporting more plausible, which would make the published rise conservative.

Evidence. AER fn 19 and Figure 4 notes; appendix B11 pp. 45-46 (rural: 'the most justified assumption seems to be to apply the same proportional upgrade factors'); appendix B12 (pre-1995 wealth distribution held fixed). 1978 top g-percentile mean / overall mean from the posted .dta files (adjudicator): raw survey 7x (2,089 yuan); raw fiscal 7x (3,040); corrected fiscal 184x (78,392); non-fiscal income 1,534x its mean (48,809); final pre-tax 317x (202,158); rural final 633x. AER Tables 1-2 imply a 1978 top-0.001% average of 6.25m (2015 yuan) = 984x mean = ~627,000 1978 yuan, a share of 0.98%. Data3 national fiscal T10: raw 21.76 -> corrected 26.20 (1978); 30.71 -> 37.74 (2015). Raw rural T10 exceeds raw urban T10 in every year (1978: 20.9 vs 16.2; 2015: 29.6 vs 24.7).

Why this severity. For the top-10% and bottom-50% trends the assumption is probably conservative (smaller 1978 under-reporting would enlarge the rise), so C3 is not undermined. But C8's very-top growth claims (top 0.001% x26, growth rising toward the top) rest on 1978 levels that are artifacts. That is a moderate qualification of a secondary result.

The authors' response. Fn 19 and appendix B11 justify the constant profile by the stable simulated/actual tax-revenue gap in Piketty & Qian (2009) and note it may under-estimate the rising trend. Appendix B12 warns that pre-1995 wealth distributions are not wealth-history estimates.

10moderateplausible · previously known · methodology

To correct the surveys the authors compare them with tax records of high earners. The tax records count one person's income, while the surveys share income among everyone in a household, and the conversion between the two income definitions adds a 10% pension most high earners do not receive. Both choices probably make the rich look somewhat richer than a like-for-like comparison would, though tax evasion pushes the other way.

Technical detail and evidence

Why it matters, in detail. The upgrade factor at p2 = 0.995 is q_tax/q_survey ~ 1.3-1.6, interpolated linearly down to 1 at p1 = 0.8, so errors in q_tax scale the whole top-decile correction. Dropping the pension add-on alone lowers q_tax by ~10% (r 0.80 -> 0.89). The unit mismatch lowers it by 0-21% depending on spousal income and co-residence. Together, the fiscal-stage 2015 top 1% (12.5%) could be ~1-2 pp lower, and the T10 correction (+6.8 pp in 2015) perhaps 1-2 pp smaller. This is the generic concern of Yonzan et al. (2022): survey-tax gaps shrink once concepts and units are aligned.

Evidence. AER fn 16 (deductions ~20%; fiscal threshold ~150,000) and fn 20; appendix B11 pp. 46-47 (r = 80% benchmark; r = 70%/90% variants with 'relatively limited impact'); appendix B44 pp. 61-62 and fn 68 (contribution base capped at 300% of the provincial average wage). PYZ2017DistributionalData.xlsx TaxableVsFiscalIncome!B5:O5 (2006): 120,000 + contributions 14,084 = 134,084, divided by 0.9 for 'Pension income (Received) 10% fo gross income' = 148,982; ratio 0.805. Medical 2%, unemployment 1% and housing 10% are applied to the full gross wage. XlsSeries/ChinaUrban_yf_raw.xlsx CompUpgradeFactors!AC1 ('See sheet AltUpgradeFactors for an alternative computation'; no such sheet); StataFiles/variants/equalsplit/ is empty. 2006 Self-Declaration Measures (Guoshuifa [2006] 162) Arts. 6-8. Adjudicator Pareto bound (alpha ~ 1.5): the sole-earner case scales the tax quantile at fixed rank by 2^(1/alpha-1) ~ 0.79; the bias vanishes if spouses earn at least ~26% of the taxpayer's income.

Why this severity. A plausible 10-20% overstatement of the tax/survey upgrade at p = 0.995 means roughly 1-2 points on the 2015 top-1% share (14.5%) and on the 41% top-10% share. That is a modest part of C4's ~10-point upward revision, and offsetting biases are of unknown size.

The authors' response. Fn 20 acknowledges that the estimates overstate inequality relative to the equal-split benchmark (and understate it relative to individual units) and defers separate series to future access to micro tax data. Appendix B11 reports r = 70%/90% variants with 'relatively limited impact'.

11minorconfirmed · newly identified here · internal inconsistency

The paper says half of Chinese housing was private in 1978, but it leaves out the large stock owned by state enterprises. Counting that, only about 39% was private, so the move to private ownership was even bigger than reported.

Technical detail and evidence

Why it matters, in detail. Danwei housing held by state enterprises (0.20x NI in 1978) is publicly owned through corporate equity; excluding it overstates the initial private share of housing by ~13 pp. The national public share (C2) is unaffected because corporate housing enters through public equity.

Evidence. NA A27 row 11 (1978): total housing B11 = 0.8159x NI, corporate I11 = 0.2043 (25.0%), private M11 = 0.3146, so 38.56% of all housing vs M11/(B11-I11) = 51.44% = Data2!Y8 (Figure 3A series). Data2!E (Figure 2A 'Housing') = AH+AI+AJ, which excludes corporate housing (Data2!AQ). 2015 (row 48): plotted 98.33%; private/all 96.95% (97.64% including the private share of corporate housing) (scripts/ownership_checks.py).

Why this severity. A labelling/denominator inconsistency whose correction makes the measured privatization larger (39% -> 97% instead of 51% -> 98%); nothing is weakened.

12minorconfirmed · newly identified here · internal inconsistency

A number of figures quoted in the text do not match the authors' own data; for example, the state owns about 55% of company shares, not 'close to 60%'. These are small slips that do not change the conclusions.

Technical detail and evidence

Why it matters, in detail. The two that matter most are the government equity share (54.5% vs 'close to 60%', with no post-2009 rebound by 2015) and the France wealth benchmark (55% vs 50%), which narrows China's distance from Europe in C7.

Evidence. Main workbook Data2!BW, BV, W, D/B; Data3!Q43/Y43 (22.65/18.95) vs AR43/CK43 (24.62/20.29); Data3!FE6/FE43 (26.80 -> 41.13) and N6/N43 (21.76 -> 30.71); Data1!T36 (0.0304); Data3!HK40:HK43 (0.55), plotted by chart sheet F10b. AER pp. 2471, 2472, 2477 (fn 19), 2480, 2482, 2484, 2486, 2489, 2491.

Why this severity. Presentation-level inconsistencies. None changes a qualitative conclusion, although (a) overstates the public equity share by ~5 points and (k) understates France's wealth concentration in the comparison.

13minorconfirmed · newly identified here · version instability

A few numbers in the published paper were carried over from an earlier draft and no longer match its own tables; the differences are small.

Technical detail and evidence

Why it matters, in detail. Theta 3 vs 2 moves 2015 T10 by ~0.6 pp and B50 by ~0.4 pp (Data3 g=3 vs g=2: 41.77/14.89 vs 41.13/15.26).

Evidence. AER fn 35 (p. 2490); Data3!H38 = FF38 = 15.62% vs posted gperc_y_China.dta 2010 T1 15.12%; WP Table 2 (531,771,316); AER Table 3 (p. 2491); WP text ('Gumbel parameter theta = 3'); AER p. 2478.

Why this severity. Cosmetic; differences of 0.5 pp or less in shares. It shows the text was incompletely updated after the revision (F7).

14minorconfirmed · newly identified here · documentation gap

For two of the five years of tax data, the average income of high earners was estimated rather than published. This makes almost no difference to the results.

Technical detail and evidence

Why it matters, in detail. Changes b** in the 2006-10 average upgrade profile by ~1.5%; negligible for shares.

Evidence. PYZ2017DistributionalData.xlsx TB11!H13 = (H12+H14)/2; H15 = AVERAGE(H12:H14); F13 = F14 - 287,573. The five-year mean b** is 2.8997 vs 2.8556 using only the three observed means (-1.5%).

Why this severity. Visible in the spreadsheet formulas and numerically immaterial (~1.5% in the calibration coefficient).

15minorplausible · newly identified here · methodology

The tax figures for high earners include one-off windfalls such as gains from selling property or lottery wins, which national statistics do not count as income. This may slightly exaggerate the incomes of the very top.

Technical detail and evidence

Why it matters, in detail. Non-national-income items in declared income raise b** and, after scaling to NI totals, the extrapolated top 0.1-0.001% shares. The size cannot be quantified from the published tabulations.

Evidence. 2006 Self-Declaration Measures (Guoshuifa [2006] 162) Art. 6(9)-(10) and Art. 8(6)-(7) (gains net of original value; contingent income in full), as quoted in appendix B4.1 pp. 57-58; TB11 b** = mean declared income / 120,000 = 2.64-3.03 (2006-10).

Why this severity. Probably small (individuals' gains on listed shares and on long-held sole homes are exempt; prizes are small in aggregate) and confined to the top 1% and above.

16minorplausible · newly identified here · overclaiming

The authors describe their inequality figures as probably too low. Some of their own assumptions push the figures up and others push them down, so the true numbers could be higher or lower.

Technical detail and evidence

Why it matters, in detail. Without a quantified error budget, 'lower bound' is a prior rather than a result. The omitted offshore wealth the authors cite (Alstadsaeter, Johannesen & Zucman 2017a,b) matters most at the very top; the comparability choices matter most around P99-P99.9.

Evidence. AER pp. 2471, 2477, 2479, 2490-2492; see F6, F9, F10, F14.

Why this severity. A framing claim that changes no estimate, but readers should not treat the shares as bounds.

The authors' response. The paper cites evidence from other countries that evasion biases measured inequality downward (p. 2479) and lists offshore wealth and Hong Kong/Macao as omissions (pp. 2490-2492).

Claims examined and rejected

Problems one auditor raised that did not survive checking.

Check passed: the statement that saving explains 50-60% of the rise in beta is verified by A42 (51%) and A43 (60%). — The cells reproduce, but they decompose wealth accumulation or the 2015 stock, not the rise in beta. By the same cells saving explains ~0% of the rise (additive counterfactual beta2015 = 3.70 < beta1978 = 3.72). Adjudicated in favour of Auditor B (F2).
I5: the time-invariant pre-2006 tax correction 'does not inflate the measured rise: it raises the 1978 level more than the 2015 level'. — Only the rural component is front-loaded (~3.5 pp in 1978 vs ~1.3 pp in 2015). The full constant correction adds +4.4 pp to national fiscal T10 in 1978 but +7.0 pp in 2015 (Data3 DT/DV vs DY/EA), so it raises the measured rise relative to raw data. Whether that biases the true trend depends on how under-reporting evolved (kept in F9 with the corrected statement).
I5: the rural correction makes rural China's top-1% share exceed urban China's 'purely by assumption'. — The raw surveys already show more inequality in rural than urban China in every year (raw fiscal top 10%: 20.9 vs 16.2 in 1978, 29.6 vs 24.7 in 2015; raw top 1% in 2015: 7.2 vs 5.4; Data3 CH/CI vs AO/AP). The ordering is in the data; only the corrected levels depend on the assumption.
I8: because the 1995 wealth upgrade factor is 1.3 rather than the stated 1.5, the 1995-2015 rise in the top-10% wealth share is overstated by ~3 pp (~11%). — The discrepancy is real (merged into F1), but its direction is not established. The 1995 corrected survey mean already exceeds national-accounts private wealth at 1.3 (1.10x) and would at 1.5 (1.16x). Under the authors' own consistency rule the admissible 1995 factor is ~1.0, which gives ~35% and a larger rise. The effect is two-sided.
I9: matching Xie & Jin's 62% requires a factor of ~1.2-1.25, and with factor 1.0 China's 2015 share would be ~58%, comparable to France. — Narrowed. Xie & Jin's 62.1% is a household-unit share, a different estimand from PYZ's equal-split-adult share, and their mean wealth per adult (174k yuan) matches PYZ's corrected mean (176k), which supports a correction of about this size. Factor 1.0 is a bound, not a plausible benchmark (the survey mean is then only ~67% of national accounts in 2012). The broader calibration concern is kept in F10.
I6: the individual-vs-equal-split unit mismatch could overstate the top-end correction by up to ~20%. — Narrowed. This is the polar sole-earner bound (2^(1/alpha-1) ~ 0.79 at alpha ~ 1.5). If spouses earn at least ~26% of the taxpayer's income the quantile bias vanishes, while co-resident non-earning adults push the other way, so the realistic magnitude is probably well below 20%. Kept at reduced magnitude in F6.
I3: the stabilization of inequality since 2006 (C6) is mechanical. — Only partly. 2006-2012 rests on observed inputs (national tax data 2006-10, surveys to 2012), and on those data the revised series shows T10 falling from 42.4% (2010) to 41.3% (2012). Only 2013-15 hold by construction (kept in F8).
I5 (minor): the wealth-calibration problem is only a documentation mismatch (threshold vs bracket-mean wording) with no numerical consequence; no claimed bias should be retained. — Partly rejected. The benchmark files also apply undisclosed year-varying factors (1.3-2.1) that break the authors' own national-accounts consistency rule in 1995, 2008 and 2009, and the factor sets the 2012/2015 level within about +/-5 points. Upgraded and split into F1 (trend, major) and F10 (level, moderate).
Assumption inventory: excluding CHIP 2008/2013 is a disclosed survey selection with stated reasons, not an issue. — The stated reason for excluding CHIP 2013, that it has no housing values, is contradicted by Knight, Li & Wan (2016, p. 7), who build CHIP-2013 housing wealth from respondents' reported values; Luo & Chen (2021) likewise compute a CHIP-2013 wealth Gini. The CHIP-to-CFPS splice materially affects C7 (F1).
Checks passed / assumption inventory: endpoint survey imputations and the time-invariant, rural-extended tax correction are disclosed choices that do not warrant issues. — Partly rejected. The tax-correction transplant and the unit mismatch are disclosed (fn 19, fn 20, appendix B11) and are kept only for their unquantified magnitudes and the 1978 tail artifacts (F6, F9). The copying of 1978-80 and 2013-15 survey distributions and the hard-coded 1978 urban/rural ratio are disclosed nowhere in the paper, which states the opposite (p. 2475) (F3, F8).

Audit notes

What reproduced. Reproduced from the authors' files with my own code: beta and its components (1978/1993/2015); public and private shares; housing and equity ownership series; Tobin's Q construction; every cell of the saving decompositions (A42/A43/A47). Income shares were integrated from 127 g-percentiles: the WP series in the posted .dta files and the revised AER series in Data3 (FC-FF), which match AER Table 1's broad groups, Fig. 4 and the abstract. AER Table 2 broad-group growth rates reproduce, as do Table 3 wealth shares and the raw-survey 2015 benchmarks. Sensitivities run: 1978 urban/rural ratio (own urban+rural merge, validated within 0.05-0.17 pp); farmland multiple; unlisted-equity Q; housing denominator; wealth upgrade factors by year and their national-accounts consistency; 1995 factor (point-mass approximation validated within 0.1 pp). NOT reproducible: AER Table 1 thresholds, averages and top-0.1/0.01/0.001% shares; Table 2 top-0.1/0.01/0.001% growth; the Fig. 9 top-0.001% bar. None of these appears in the 359 posted .dta files with 2015 top-bin rows, and the openICPSR deposit (E116194V1) is behind a Cloudflare challenge. Gpinter/Matlab interpolations were not re-run; wealth counterfactuals use the authors' own variant outputs (Data8) and coarse approximations.

Limits of this audit. (1) The AEA/openICPSR deposit (doi:10.3886/E116194V1) could not be opened (Cloudflare 403). A Wayback snapshot shows it holds a 647.8 MB PYZ2017NationalAccountsData.zip, and it may also hold the final AER g-percentiles; F7 concerns the accessible files only. (2) Gpinter/Matlab interpolations and the copula step were not re-run, and no Stata code was executed. Wealth counterfactuals rely on the authors' variant outputs (Data8) and point-mass approximations; urban-rural merges are validated to within 0.2 pp. (3) CHIP/CFPS microdata were not accessed. The cross-survey comparisons (KLW, Xie & Jin, Luo & Chen) involve unit differences (household vs per-adult; ranking by household total vs per capita), so the size of the survey effect in F1 is approximate. (4) Only abstracts were read for Cao (2020, Revue d'economie politique 130(5): 683-698), Tao, Tang & Chen (2026), Luo & Chen (2021) and Wan & Yu (2023). (5) NBER WP 23368 and CEPR DP 12410 were not compared separately (the WID WP 2017/6 was used). (6) SAT high-income releases were taken from the authors' archive and Auditor B's checks, not re-fetched. (7) The US/France comparator series were not audited. (8) A material convention was not treated as an issue: government capital income plus net production taxes (y_g ~ 26-28% of national income in 1978 and 2015, TB0) is allocated in proportion to income (fn 24). Auditor B reports that allocating government capital income equally per adult would lower 2015 T10 by ~3.9 pp (not re-run here). This would make China look more equal relative to the US. (9) No author correspondence was attempted, and no author reply to these specific points was located.