Tax Evasion and Inequality

Why it matters, in detail. Mechanism: the benchmark hidden-wealth distribution (App. J.1) is the simple average of the HSBC (top-0.01% share 55.3%) and Norwegian amnesty (47.8%) distributions, and the top bin's share of hidden wealth feeds directly into its evasion rate. The HSBC top bin has only 8 households with known values; one Norwegian household (NOK 1.848bn, about $288m) supplies 51.7% of its wealth, 28.6% of the Scandinavian HSBC wealth in J.1 and 88.9% of matched Norwegian HSBC wealth. Dropping it lowers the HSBC top share to 37.5%, the averaged share from 51.6% to 42.6% (-9 pp) and the headline from 24.38% to 21.04% (-3.3 pp, -14%); the HSBC-only variant falls from 25.70% to 18.97%, while the amnesty-only estimate (23.00%) is untouched. The same observation drives the bottom-up corroboration of the macro stock: $54.2bn (1.65% of wealth) -> $38.8bn (1.19%) with Table 2's $105bn, or $48.1bn -> $34.4bn (1.05%) with the text's $118.4bn, implying headlines of 19.7% and 17.9% if used as the stock (15.3% if the household is also removed from the distribution); Norway's own bottom-up stock falls from 2.4% to about 0.3% of wealth. Direction: the undisclosed outlier raises the headline (C3) and the concentration figure (C2), and makes the 'striking' HSBC-amnesty similarity, the 'likely to be robust' stock and the footnote-13 inference look stronger than they are; top-share corrections built on this distribution (C5) would also be smaller (not quantified). The distribution propagates to AJZ (2018a)'s ten-country application and to Guyton et al. (50% weight in the US benchmark, 100% in the high-end case). The qualitative finding survives because the amnesty leg alone gives 23.0%.

Evidence. programs/norway/hsbc_skatt.xlsx, sheet 'Export Worksheet', row p99.99-p100: 1 household with a positive value, mean NOK 1,848,269,805, SD 0. The 22 matched households with positive values total NOK 2,078.8m = $324.1m at 6.41333, matching Table 2's Norway HSBC figure of $0.32bn. App. J.1: R21 = $557.8m, sum of R = $1,008.0m; App. Table E.3: 8 top-bin households with values. Adjudicator re-implementation (scripts/replicate_offshore.py, outputs S1/S2/S5). Dropping the household: HSBC top-0.01% share 55.3% -> 37.5%; averaged share 51.6% -> 42.6%; headline 24.38% -> 21.04%; HSBC-only variant 25.70% -> 18.97%; amnesty-only unaffected at 23.00%. Bottom-up stock $54.2bn (1.65% of wealth) -> $38.8bn (1.19%) with Table 2's $105bn, or $48.1bn -> $34.4bn (1.05%) with the text's $118.4bn. Headline using these stocks: 19.7% and 17.9%. Norway's bottom-up stock: $17.3bn (2.4%) -> about $1.9bn (0.3%). Guyton et al. (draft of 22 Dec 2025) describe the AJZ finding as '52% of offshore wealth in that bank'.

Why this severity. The headline falls by about 3.3 points (-14%) and the concentration figure by 9 points. The corroborating claims (HSBC-amnesty similarity, the bottom-up stock check, footnote 13) are much weaker than presented. The amnesty-only estimate (23.0%) does not involve the household, so the qualitative core finding survives.

Why it matters, in detail. Mechanism: in App. K.3 each price-taking bank chooses customer shares s_i(y); the paper imposes the first-order condition y p(y) = lambda*phi*(y s_i + k_i) and treats the symmetric allocation s_i(y) = 1/B above a cutoff as the equilibrium. At those prices any change in a bank's customer mix changes its profit by exactly -lambda*phi*dS*dK, so replacing poorer clients with fewer, richer ones (dS < 0, dK > 0) strictly raises profit: the Hessian determinant is -c^2(y_h-y_l)^2 < 0 and the candidate is a saddle point, not a maximum. Numerically (a = 1.5, ymin = 1, B = 10, lambda*phi = 1, theta = 0.01; cutoff 11.70), shifting shares gives dS = -4.6e-5, dK = +9.3e-3 and a profit gain of about +4e-7 > 0. Consequently the proposed competitive equilibrium is not an equilibrium, and its comparative statics (including the 'new result' that more banks raise evasion) and the claims that the key intuitions (fn 42) or 'all our results' (appendix fn 19) carry over are unestablished. Magnitude: confined to the secondary theoretical claim C8; the monopoly model (closed form s* = theta/[lambda*phi*(1+a/(a-1))] and the fn 22 second-order condition) is correct, and no empirical estimate depends on K.3.

Evidence. AER fn 42 (p. 2099); App. K.3 eqs. (5)-(9), pp. 27-29. Codex scripts/theory_check.py (Hessian determinant -c^2(y_h-y_l)^2 < 0). Adjudicator scripts/competitive_model_check.py with a=1.5, ymin=1, B=10, lambda*phi=1, theta=0.01: cutoff 11.70, S_i=0.0025, K_i=0.0877. Shares 0.10->0.09 on [11.7,17.5] and 0.10->0.234 on [117,234] give dS=-4.6e-5, dK=+9.3e-3, profit gain +4.2e-7 > 0.

Why this severity. Invalidates a stated secondary (theoretical) result; no bearing on any empirical estimate.

Why it matters, in detail. Mechanism: the evasion rate is measured in taxes, not wealth. In the benchmark the top 0.01% hide 12.1% of their true wealth (J.1 B21), but hidden wealth is assigned a fully taxable 4.5% return taxed at a 52.08% combined capital MTR (about 2.34% of its value a year), whereas their recorded wealth produces reported taxable income, including labour income, of only 2.2% of its value and about 1.0% in tax; so 12.1% of wealth becomes 22.0% of taxable income and 24.4% of taxes owed. Footnote 32 instead describes the top 0.01% as concealing '20 percent' of wealth and the assumed return as only 'slightly higher', which obscures that the yield gap roughly doubles the hidden-wealth share into the evasion rate. Magnitude (wealth tax held fixed per dollar): r = 3.5% (the paper's observed top-1% return) gives 21.6%, 3.0% gives 20.1%, 6% gives 28.2%, and an equal 2.2% yield gives 17.6% (13.6% under the J.4 convention); over a defensible 3-6% range of 2006 taxable yields the headline is about 20-28%. Direction: a true taxable yield on hidden assets below 4.5% means the headline is overstated, and vice versa. The authors' arguments (recorded top wealth is mostly closely held equity with retained earnings; no incentive to under-realize returns in evading accounts) and AJLZ (2022)'s post-disclosure rise in tax payments of about 2% of net wealth a year point against a low yield, but the justification also leans on Norway's post-reform 2006 dividend collapse (its top 0.01% then had 0.235% of national taxable income vs 4.6% of wealth), which makes the 2006 base year matter. The same assumption scales the evaded tax in the effective-tax-rate profile, so C7's regressivity at the very top inherits this sensitivity (not separately quantified).

Evidence. App. J.1 B21 (12.15%); J.2 B21 (22.0% of taxable income hidden); J.7 B21 (average tax rate 45.6%); J.7b C21 (52.08%); J.8 C9:C10; J.9/J.10 B21 and D21. AER fns 31-32 (p. 2093). Adjudicator (replicate_offshore.py S4/S8), holding the wealth tax fixed per dollar: r = 3.5% (the paper's observed top-1% return) -> 21.6%; r = 3.0% -> 20.1%; r = 6% -> 28.2%; equal yield 2.2% -> 17.6% (13.6% under the J.4 convention).

Why this severity. Over a defensible 3-6% range of 2006 taxable yields the headline is about 20-28%, a noticeable but not transformative change. Footnote 32 obscures which assumption drives the doubling from hidden-wealth share to evasion rate.

The authors' response. AER fn 31: the observed return on non-hidden wealth is a lower bound, and there is no incentive to under-realize returns on evading accounts. AJLZ (2022, p. 3) report that disclosers' annual tax payments rose by about 2% of their total net wealth after disclosure, which does not suggest a low yield.

Why it matters, in detail. Mechanism: no effect on the Scandinavian estimates; the issue is external validity and downstream reuse. AJZ (2018a) apply the Scandinavian concentration (top 0.01% about 50% of hidden wealth, top 0.1% about 77%) to ten countries, Guyton et al. give it 50% weight in their US benchmark (100% in the high-end case), and the AER (p. 2075) suggests evasion among the wealthy 'may be even higher elsewhere'. Applying the AJZ method to more than 29,000 Dutch evaders, Leenders et al. (2023) find that the top 0.01% hold about 7% of hidden wealth and evade about 8% of their liability (vs 24.4% here), with 67% held by P90-P99.9, and the top 0.1% wealth-share correction is 8.7% -> 9.3% (+0.6 pp, vs +1.3-1.4 pp in Norway). In US FBAR data, 21% of the $124bn of newly disclosed offshore wealth belongs to the top 0.01% by income. Direction: where such patterns hold, transplanting the Scandinavian distribution overstates top-end evasion and top-share corrections (about 7% vs about 50% of hidden wealth in the Dutch case); Colombia shows higher concentration, so the bias is country-specific rather than uniform. The extensive-margin gradient does generalize (more than 10% of the Dutch top 0.01% used the amnesty), and amnesty-based estimates may miss the most sophisticated evaders, which tempers the Dutch comparison.

Evidence. Leenders, Lejour, Rabaté & van 't Riet (2023, JPubE 217:104785), abstract and Sec. 4.2. Guyton, Langetieg, Reck, Risch & Zucman (draft of 22 Dec 2025), Sec. 3, discussion of Fig. 3c ('Of the $124 billion ... 21% belongs to the top 0.01%'). AER p. 2075 and fn 35.

Why this severity. Does not affect the Scandinavian estimates. It does undermine the suggestion that the results are a lower bound elsewhere, and the downstream reuse of the distribution for other countries, a central channel of the paper's influence on inequality measurement.

The authors' response. AER p. 2075: 'We certainly do not claim that our estimates ... hold everywhere as a universal law'. No reply to Leenders et al. located; Leenders et al. note that amnesties may miss the most sophisticated evaders.

Why it matters, in detail. Mechanism: the WP's trend claims apply the Norwegian offshore correction (App. B.4) to the historical top-share series. In the authors' posted Data-F11 (identical in the 6 Oct 2017 release and the current file; '2010' = 2010-13 average), the correction erases 33.5% of the top 0.1% share's decline since the 1930s and 25.4% since the 1940s, not 'half'; the corrected 2010-13 top 0.01% share (4.85%) is above the 1930s (4.01% corrected, 3.85% uncorrected) but below the 1940s (5.12% corrected). Scaling to the WP's larger 2.3% Norwegian stock (vs 1.94%) still erases only about 40% and leaves the top 0.01% (about 5.03%) below the corrected 1940s level (about 5.15%); the 1940s claim then holds only against the uncorrected 1940s figure (4.96%). The 1970s-2010s rise in the top 0.1% share goes from +1.82 pp in tax data to +2.61 pp corrected (+43%), arithmetically 'marked' but assumption-driven: offshore wealth is backcast with Swiss banks' foreign-wealth series and its distribution is held at 2006 values back to 1930. Direction: the WP overstates the long-run correction in two specific respects and the rest is assumption-driven (rated 'likely' because the WP-era figure data were unavailable). The AER drops the claims and lowers the headline ('close to 30%' total to 'about 25%' offshore; lower bound 'at least 15%' to 12%) without flagging the change, while the NBER abstract remains in circulation. The AER's core estimates are not affected; only the long-run extension of C5 (the B.4 backcast) shares the assumption-dependence.

Evidence. AJZ2017.xlsx Data-F11 (identical in the 6 Oct 2017 release and the current file; row '2010' = 2010-2013 average); scripts/document_checks.py (longrun). Share of the top 0.1% decline erased: 33.5% since the 1930s, 25.4% since the 1940s. Corrected top 0.01% share in 2010-13: 4.85%, vs 1930s 4.01% corrected (3.85% uncorrected) and 1940s 5.12% corrected (4.96% uncorrected). 1970s->2010s rise of the top 0.1% share: +1.82 pp in tax data vs +2.61 pp corrected (+43%). Scaling the correction to the WP's 2.3% Norwegian stock (vs 1.94%) still erases only about 40%, and leaves the corrected top 0.01% (about 5.03%) below the corrected 1940s level (about 5.15%). App. B.4: offshore distribution fixed at its 2006 values back to 1930. WP pp. 2-4, 35-36.

Why this severity. The working paper's headline trend claim, which is part of the work's widely cited abstract, is overstated in two specific respects and otherwise assumption-driven. The AER's core estimates are not affected.

The authors' response. The AER removed the claims; WP p. 36 acknowledged 'a sizable margin of error'.

Why it matters, in detail. Mechanism: every bin-level estimate is scaled to an aggregate Scandinavian offshore stock taken from AJZ (2018a)'s allocation of Zucman's (2013) $5.6tn, which assigns the whole global portfolio asset-liability discrepancy to households (about 1.6% of Scandinavian household wealth, 90% assumed hidden; J.1 B23 = 0.9*50.96/(3,228.7+50.96) = 1.40%). The headline is close to proportional to the stock: one-third lower gives 17.7% (-6.7 pp, about a quarter), halved gives 13.9% (auditor B), and the Swiss-only low end gives 12.0%. Direction: if, as Schmitz (2021) argues (euro-area household offshore assets of 3.5-5% of GDP vs 7-8% in AJZ-type studies), part of the discrepancy belongs to non-financial corporations or institutional investors or reflects reserve-asset and direct-vs-portfolio misclassification, the stock and the headline are overstated. But that evidence concerns the euro area in later years and partly reflects post-2008 transparency, independent global totals are similar, and TJN argues the stock is too low, so the sign is uncertain (confidence 'plausible'). The paper's bottom-up HSBC check does not independently pin the stock down: without the F1 household it implies 1.05-1.19% of wealth rather than 1.65%, and the text and Table 2 disagree on its inputs (F9). Even the 12% low end exceeds audit-detected evasion about threefold, so the qualitative C3 and the contrast with audits (C4) survive.

Evidence. AER pp. 2089-2090 and Table 2. App. J.1 B23 = 0.9*50.96/(3,228.7+50.96), reproduced exactly. Schmitz (2021, 63rd ISI WSC, IPS214, abstract): euro-area household offshore assets 'around 3.5% to 5% of GDP' vs 7-8% in AJZ-type studies, partly from 'allocating unrecorded assets entirely to households rather than corporations'. Adjudicator sensitivity (replicate_offshore.py, S5): stock one-third lower -> 17.7%; Swiss-only -> 12.0%. Auditor B: stock halved -> 13.9%.

Why this severity. A one-third overstatement would cut the headline by about a quarter. But the Schmitz evidence concerns the euro area in later years, partly reflecting post-2008 transparency. Independent global totals are similar, and TJN argues the stock is too low. The authors disclose a 12-40% range whose lower end still exceeds audit-detected evasion threefold.

The authors' response. AER fn 28: the $5.6tn is 'at the low-end of the scale of available estimates'. pp. 2094-2095 report a Swiss-only low-end scenario (12%) and a world-average high-end scenario (40%). No direct reply to Schmitz located.

Why it matters, in detail. Mechanism: Swedish amnesty amounts are unobserved, so Table 1 col. 7, Fig. 3B and the self-selection odds ratios rank Swedish participants by recorded wealth plus a disclosed amount drawn from a lognormal fitted to Norwegian amounts (build_norswe.do), which the main text does not state. Because adding (imputed) hidden wealth moves participants up the ranking, the top-0.01% participation rate is 13.8% for Norway and Sweden with imputed amounts vs 9.7% ranking by recorded wealth; Norway alone, with observed amounts, gives 11.4% (7.7% by recorded wealth) and Sweden 15.6% (9.1%). 'We know that at least 14 percent' therefore overclaims; the supportable figure is about 10-14%. The self-selection odds ratios move from 32.8/20.4 (ratio 1.6) to 28.2/13.2 (ratio 2.1) without imputation, so that test also depends on it. Separately, App. G.1's 'small number' of dropped Norwegian cases is 461 of 1,883 households (24.5%), concentrated at the bottom (46% of P0-50 participants vs 3% in the top 0.01%). Magnitude for the core: none for the offshore-evasion estimates (C3), which use only observed Norwegian amounts; the steep wealth gradient in C1 survives.

Evidence. programs/scandinavia/build_norswe.do ('Impute Swedish amounts based on Norway ...', lognormal draws); App. Table G.2: N+S top 0.01% 13.8% incl. imputed vs 9.7% excl.; Norway 11.4%/7.7%; Sweden 15.6%/9.1%. Odds ratios 32.8/20.4 -> 28.2/13.2 without imputation (ratio 1.6 -> 2.1; scripts/document_checks.py). App. G.1: 461 of 1,883 households dropped; AppendixG.xlsx discloser(NOR2): dropped share 46% in P0-50 vs 3% in the top 0.01%.

Why this severity. Changes a highlighted descriptive number ('at least 14%' becomes about 10-14%) and the self-selection test, not the offshore-evasion estimates, which use only observed Norwegian amounts; the steep gradient survives.

Why it matters, in detail. Mechanism: J.4 varies the taxable return r from 2% to 7% but keeps the combined capital MTR of 52.08%, which embeds the wealth tax expressed per unit of a 4.5% return (e.g., J.7b AA10 = 0.28 + 0.8*0.9/4.5). Evaded wealth tax per dollar of hidden wealth (about 0.79%), which should not depend on r, therefore scales with it, steepening the sensitivity in both directions. Corrected with the wealth tax fixed: r = 2% gives 16.9% (published 12.5%), 3% gives 20.1% (17.7%), 7% gives 30.6% (33.4%); the WP's '35% at 6%' is not reproduced (J.4 gives 30.1%, corrected 28.2%). Direction: the published table overstates return-sensitivity; the correction raises the low-return cases, leaves the 4.5% benchmark unchanged and changes no conclusion.

Evidence. App. J.4 B11 = B27*T.J7b!$C21/(T.J7!$B21*(1-B27)+B27*T.J7b!$C21); J.7b AA10 = 0.28+0.8*0.9/4.5. Published vs corrected (wealth tax fixed at 0.79% per dollar): 2% 12.5% vs 16.9%; 3% 17.7% vs 20.1%; 7% 33.4% vs 30.6% (adjudicator and both auditors agree). NBER fn 31's '35% at 6%' is not reproduced (J.4: 30.1%).

Why this severity. An appendix robustness table. The error makes the result look more return-sensitive than it is; correcting it strengthens the low-return case and changes no conclusion.

Why it matters, in detail. Mechanism: the bottom-up check scales matched Scandinavian HSBC wealth (about $1.01bn) by the ratio of global offshore wealth ($5.62tn) to HSBC's total client assets. Table 2 uses $105bn (the 2006 value in App. Table E.5, although labelled mid-2007 and attributed to HSBC 2015), giving a multiplier of 53.5 and $54.2bn (1.65%, shown as 1.7%), slightly above the $50.96bn benchmark; the text (p. 2077, fn 26) and the cited source use $118.4bn (2007), a multiplier of 47.5 and $48.1bn (1.5%), below it. The NBER WP's Table 2 used 118.4 (48.1 vs a 54.4 benchmark), so the switch moved the corroborating figure from below to above the benchmark without the text being updated. Direction and magnitude: with the cited $118.4bn the bottom-up figure lies about 6% below the benchmark rather than about 6% above it; the benchmark stock and the headline are unaffected, and given F1 neither version is strong corroboration ($34.4-38.8bn without the dominant household). The AER also calls HSBC's 30,412 accounts 'clients'.

Evidence. AER pp. 2076-2077, 2089 fn 26, 2090 Table 2 and note; AJZ2017.xlsx T2!C6 = 105 (Oct-2017 archive formula =[1]T.E5!C10/1000, the 2006 row of App. Table E.5), D9 = D6/(C6/C8) = 54.24; with 118.4 -> 48.10. HSBC (2015), p. 3: 'In 2007, the Swiss Private Bank had 30,412 accounts ... total client assets of USD118.4 billion.' NBER WP Table 2.

Why this severity. Affects only a secondary corroborating calculation; the benchmark stock is unaffected.

Why it matters, in detail. Mechanism: evaded tax at the top equals hidden taxable income times the capital MTR applied. The text (p. 2091) says the top-0.01% capital MTR is 49% and 'is the rate that we apply', but the workbook applies 52.08% (J.7b C21; 48.79% is the adjacent labour rate); applying 49% lowers the headline from 24.38% to 23.27% (-1.1 pp). Denmark's rate on capital income other than share income is approximated as the labour rate minus 8 points (48.0% at the top), below the pre-2010 statutory 59.7% on positive net capital income above the top-tax threshold, so this approximation is conservative; footnote 30's '48 percent for interest at the top' inherits it. Auditor A's alternatives (Danish share-income rate; Norwegian ceiling rate) each give 23.8%, and country rates are combined with adult-population weights. Net: effects of about +/-1 pp that partly offset; no conclusion changes.

Evidence. App. J.3 G21 = J.2 B21 * J.7b C21 (0.5208); J.7b B21 = 0.4879; J.7b column J = I - 0.08; replicate_offshore.py S3: 49% -> 23.27%. Danish Ministry of Taxation, 'Danish Tax Reform 2010', p. 8 (59.7% on positive net capital income above DKK 347,200 under the then-current rules). Auditor A: Danish share-income rate -> 23.8%; Norwegian ceiling rate -> 23.8%.

Why this severity. Effects of about +/-1 point that partly offset; the Danish approximation is conservative.

Why it matters, in detail. Mechanism: App. J.5 combines audit-detected (a) and offshore (b) evasion rates as a(1-b)+b(1-a); if both are measured against the same taxes paid, the exact combination is (a+b-2ab)/(1-ab). At the top (a = 3.67%, b = 24.38%) the exact value is 26.50% vs the published 26.26% (and 25.90% in the workbook's own 'full formula' column), a 0.2-0.6 pp gap; the aggregate is 2.71% vs 2.72%. Direction: under that reading the published combined gap is slightly too low; the offshore-only headline (C3) and the audit-vs-offshore contrast (C4) are unaffected. Negligible.

Evidence. App. J.5 B22 (26.26%) vs exact 26.50%; the workbook's own 'full formula' column S22 gives 25.90%; aggregate 2.71% vs 2.72% (scripts/document_checks.py; codex reproduce.py). Same rule in AddStandardErrorsToFig5A-5B-6.do.

Why this severity. Negligible (about 0.2-0.6 pp at the top); offshore-only headline unaffected.

Why it matters, in detail. Mechanism: B.4 computes corrected group shares as s/(1+h)+d*h, where h (1.94%) is offshore wealth as a share of total rather than recorded wealth; the consistent form is s(1-h)+d*h. The corrected shares therefore sum to 1.00037 (= 1 + h^2/(1+h)), i.e. they are marginally overstated; fixing the formula changes the 2000-09 top 0.1% mean (9.62%) by -0.003 pp. Quantitatively negligible for C5.

Evidence. App. B.4 C107 formula 'T.B2!B107/(1+$B107)+I107*$B107', B107 = 0.019428 (= T2!F13); C107+D107 = 1.00037 = 1 + h^2/(1+h). Effect on the 2000-09 top 0.1% mean: -0.003 pp (codex).

Why this severity. Quantitatively negligible.

Why it matters, in detail. Mechanism: Figure 1 (country shares of HSBC wealth vs of Swiss offshore wealth) supports the claim that HSBC is representative of Swiss banks, which underpins the bottom-up scaling. Its note reports a slope of 0.90 (SE 0.04) and R2 0.75 for '134 countries', which reproduce only as a no-constant regression on all 146 rows of App. Table E.8 (0.904, SE 0.043, uncentered R2 0.75); with a constant the slope is 0.887 (0.047) and R2 0.71 (the WP reported 1.00 and 0.80). The figure's own data give Scandinavians 1.75% of HSBC wealth vs 0.81% of Swiss offshore wealth, not 'about 1 percent' of each, which holds only for matched-account data (0.97%); because the 1.75% uses ICIJ totals that include non-resident, corporate and multi-country accounts, over-representation is not established. Direction: documentation slips that slightly weaken the representativeness argument; the benchmark stock and the headline are unaffected.

Evidence. scripts/figure1_regression.py on App. Table E.8: no constant 0.904 (0.043), uncentered R2 0.75, n = 146; with constant 0.887 (0.047), R2 0.71. E.8 shares: Denmark 0.63%/0.17%, Sweden 0.86%/0.48%, Norway 0.26%/0.16%. AER p. 2078 and Fig. 1 note; AJZ2017.xlsx T2 D27-D29.

Why this severity. Documentation slips that slightly weaken the HSBC-representativeness argument; benchmark unaffected.

Why it matters, in detail. Mechanism: presentation and citation errors with no effect on any estimate. (i) The Norwegian top 0.1% correction quoted as 8.4% -> 9.8% 'over the 2000-2009 period' is the 2010-13 average; for 2000-09 it is 8.35% -> 9.62% (+1.27 pp vs +1.44 pp), the same qualitative message for C5. (ii) Johns and Slemrod dated 2001 in the text vs 2010 in the references. (iii) The 2006 S&P 500 total return is given as 13.4% (fn 29, J.4 note) vs 15.79% actual. (iv) Table 1's note says bootstrapped SEs; reg_evaders.do uses robust SEs (a 'negligible difference' per its comment). (v) The amnesty window is given three ways ('since 2006', '2007 to 2015', 'between 2009 and 2015'). (vi) 'Kauffmann' for Kaufmann. (vii) Appendix fn 19 cites K.2 for the competitive model (K.3). No conclusion is affected.

Evidence. (i) p. 2094 'from 8.4 percent to 9.8 percent on average over the 2000-2009 period': data give 8.35% -> 9.62% for 2000-09; 8.40% -> 9.84% is the 2010-13 average (Data-F11; B.4). (ii) p. 2096 'Johns and Slemrod 2001' vs 2010 in the references. (iii) fn 29 and the J.4 note give the 2006 S&P 500 total return as 13.4%; it was 15.79% (slickcharts.com, us500.com). (iv) Table 1 note says bootstrapped SEs; reg_evaders.do uses robust SEs ('makes negligible difference'). (v) Amnesty window given as 'since 2006' (p. 2080), '2007 to 2015' (Fig. 3) and 'between 2009 and 2015' (p. 2086). (vi) 'Kauffmann' for Kaufmann. (vii) Theory appendix fn 19 cites K.2 for the competitive model (it is K.3).

Why this severity. Presentation and citation errors; no conclusion affected.

Why it matters, in detail. Mechanism: the bootstrap behind the Figure 4B intervals (AddStandardErrorsToFig5A-5B-6.do, section 1D) equalizes weights by expanding HSBC rows by round(N_disc/10) and amnesty rows by round(N_hsbc/10), then draws 100 samples of the combined original size, so each replicate contains about 970 HSBC households vs the 520 observed (about 1.9x). This understates the sampling variability of the heavy-tailed HSBC leg; in particular the probability that a replicate omits the dominant Norwegian household (F1) falls from about 37% to about 15%, and omitting it lowers the estimate (to about 21% in the point-estimate calculation). The bootstrapped statistic is also the pooled-wealth share ('true' value 24.67%) rather than the arithmetic average used for the 24.38% point estimate. Direction: the reported 95% interval for the top 0.01% (19.0%-30.6%, SE 2.95 pp, normal approximation) is probably too narrow, especially on the downside; the size of the understatement was not quantified (code read, not re-run). Point estimates are unaffected.

Evidence. programs/scandinavia/AddStandardErrorsToFig5A-5B-6.do, section 1D (expand weight with round(N_disc/10) for HSBC rows and round(N_hsbc/10) for discloser rows), bootstrap ..., size($samplesize) reps(100), e(ci_normal). AJZ2017.xlsx Data-F1-6b Z18 (true 24.67%), AB18 (SE 2.95 pp), AD18/AC18 (19.03%/30.58%). Sample sizes from App. E.2 (520) and G.1 (1,422). Code read only; not re-run (Stata; Danish tabulation withheld).

Why this severity. Affects the reported precision, not the point estimates or qualitative conclusions.

Why it matters, in detail. Mechanism: J.1 averages HSBC shares binned on Scandinavian wealth (top-0.01% threshold $44.5m) with Norwegian amnesty shares binned on Norwegian wealth (threshold $33.9m), so the amnesty 'top 0.01%' includes households that would fall below the Scandinavian cutoff. Re-binning the amnesty top bin to $44.5m (Pareto a = 1.45 fitted to App. Table G.4; 88.5% of top-bin wealth lies above the threshold) lowers the amnesty top share from 47.8% to 42.3%, the headline from 24.4% to 23.4% (-1 pp) and the amnesty-only estimate from 23.0% to 20.9%; combined with dropping the F1 household the headline is 20.0%. Direction uncertain: whether rank-based mapping across countries (the authors' choice) or absolute-wealth mapping is appropriate is debatable. The AER does not mention the mismatch (the data file labels it).

Evidence. AJZ2017.xlsx Data-F6a headers; App. J.1 col. Q = App. Table G.4 col. F (threshold $33.87m); Data-F3-4-5 threshold $44.51m. Adjudicator re-binning (Pareto a = 1.45 from G.4 threshold and mean; 88.5% of top-bin wealth above $44.5m): amnesty top share 47.8% -> 42.3%; headline 24.4% -> 23.4%; amnesty-only 23.0% -> 20.9%; with F1: 20.0%.

Why this severity. About 1 point on the headline, in a direction that depends on a debatable mapping choice.