Taxing Capital in a Globalized World: The Effects of Automatic Information Exchange

Why it matters, in detail. share = G/(A - B) with B = income_cf/yield: the denominator is a small difference of two large, imprecise numbers, so the ratio is highly non-linear in B and in the yield. Propagating the paper's own standard errors gives a 95% range of roughly 41% to >100% (28% of draws infeasible); a 25% higher yield gives 53%. Any positive CRS response of the >100k group of the size seen in the 2019 data would make the accounting infeasible, so the 'conservative' zero-response assumption (p.22) is not innocuous.

Evidence. Reproduced: (9.6+3.9+3.4)/30.1 = 56.1% and (16.8+8.1+3.8)/30.1 = 95.3%; with repatriation propagated into H: 64-85%. Panel B at its CI bounds: H = 54.2 -> 42%; H = 6.0 -> 380%. Monte Carlo with independent normals from the Table 5 intervals: median 76%, 2.5th percentile 41%, P(share > 100% or H <= 0) = 28%; holding B fixed: 67-85%. Yield x0.9 / x1.1 -> 96% / 64% (fn 36: 96% / 63%); x1.25 -> 53% (actual 47%); x0.8 -> 138%. Exact Fig 7(c) data: 2019 observed 1,022m vs trend 704m (+2.4 residual SDs; SD 135m = ~7bn capitalised); applying the benchmark 'no effect above 100,000 DKK' literally (counterfactual = observed) gives B = 64bn and a share of ~174%; treating the 2019 deviation as a CRS effect gives >130%.

Why this severity. The point estimate is not shifted, but the paper's own reported uncertainty is consistent with closure shares from about 40% to over 100%; 'about 70%' and its 56-95% 'bounds' are not established at the stated precision, and they rest on unvalidated yields and an extrapolated linear trend for a volatile series.

The authors' response. In the paper: Section 8 opens by cautioning that 'the offshore tax gap is known with less certainty... a margin of error is necessarily involved' (p.30); fn 36 reports the +/-10% yield test (63-96%); fn 37 argues omitted margins make the benchmark conservative.

Why it matters, in detail. Bias in the residual denominator: H_true = H_paper + X (plus the proportional non-sending add-on), where X is in-scope reported-but-unmatched direct wealth; equivalently the paper subtracts the unmatched owners' declared wealth from hidden wealth in matched accounts. Variants (numerator 22.8 fixed): T = W/r -> H = 55.5, share 41% (actual 36%); composition-adjusted with Fig 8(b) wealth-matched shares (havens 0.82, OECD non-havens 0.59, non-OECD 0.17) -> 40-51%; adding f x 23.9bn -> 63% (f = 0.25), 54% (0.5), 42% (1.0). The direction is one-signed (share overstated); the size depends on the in-scope share of unmatched reports and the representativeness of the repeat-transfer sample, which the paper itself assumes when applying s to the whole stock. Partial offsets: possible joint-account double counting in matched balances (Table 3 suggests ~1% of income) and repatriation possibly not scaled for the excluded top-5 bank (if scaled, +3-6pp in any variant). The haven-only AJZ benchmark (27-35bn) cannot adjudicate because the correction adds mainly non-haven wealth.

Evidence. Recomputed: r = 1,187/2,006 = 0.592; M = 22,009/30,758 = 0.716; s = 0.173; 53.1*s/(1-s) = 11.1 (Table 4 p.50; Table 5 p.51). Model-consistent stock T = 53.1/r = 89.7bn (53.1 matched + 21.1 reported-but-unmatched + 15.5 not sent) vs 64.2bn in Table 5. Exact Fig 5(c) data: 21,580 of 43,963 self-reporters in 2019 (49%) have no matched CRS report; Fig A9(a): 112m of the 262m declared by the <100k group comes from unmatched self-reporters. Self-reported foreign income (Fig 7a + 7c) was 835m, 1,285m and 928m in 2018-2020, i.e. 1.55-2.38x the 540m of dividends and interest banks reported on all matched direct accounts in 2019 (Table A1; that 540m also includes FATCA income whose balances are absent from the 53.1bn). The paper itself explains part of the 24% residual by the fact that 'some wealth that is being reported cannot be matched to specific taxpayers, ... potentially enabling some residual noncompliance to take place' (p.32), although its stock excludes such wealth by construction. p.9 dismisses unmatched accounts because they are 'predominantly held in non-haven countries', but Table 5 includes non-haven matched wealth (28.0bn) and all taxpayers' non-haven self-reports.

Why this severity. Under the paper's own coverage model the share closed falls from 76% to about 41% (actual enforcement 67% -> 36%); if only a quarter to a half of the unmatched individual wealth belongs to in-scope Danish taxpayers it is 54-63%. Any in-scope fraction above ~0.42 lowers the headline by a quarter or more. The qualitative conclusion (a large reduction) survives.

The authors' response. No external response located. In the paper, p.9 argues that unmatched accounts are 'predominantly held in non-haven countries' so their exclusion 'should not be a major source of concern'; p.28 cautions against extrapolating the repeat-transfer estimates; pp.31-32 present the AJZ benchmark (27.2bn) as independent support for ~30bn.

Why it matters, in detail. The pre-period shows a level shift in 2009 rather than flat parallel paths. If the shift reflects an ongoing enforcement process (cf. F3), the post-2012 effect is overstated; if a one-off step, re-basing on 2009-12 costs ~1bn; linear detrending removes ~25%. Counterweight: one top-5 bank (25% of 2013-19 transfers) is excluded and the 24 banks cover ~90% of sector working capital; if the aggregate is not scaled up (undocumented), repatriation is conservative.

Evidence. Exact vector data, Fig 3(b): 2005 -12.6 (upper CI -3.6), 2006 -14.7 (+1.0), 2007 -23.0 (-11.0), 2008 -21.2 (-3.1); 2009 +10.4, 2010 +8.2, 2011 +13.4. Sum of 2013-19 coefficients 751.6% x implied 2012 base 1.756bn = 13.2bn. Baseline = 2009-11 mean -> 11.9bn; 2009-12 mean -> 12.2bn; linear 2005-12 trend removed -> 9.9bn. Table A2: 7.2 (line 7), 10.5 (line 1, = the 2024 baseline), 11.7, 12.4, 13.2 (baseline), 13.8 (DDD), 14.0 (incl. indirect). The 2024 figures show the same pattern (2007-08 significantly negative). Oct 2024 p.14 ('close to the one obtained in our baseline design') vs 2026 p.14 (claim dropped).

Why this severity. Repatriation's magnitude (C2) is uncertain by roughly -25% to +6% (-45% for the non-haven-control design), but repatriation remains the largest margin in every variant; the headline moves only to 70-76% because repatriation enters both numerator and denominator.

The authors' response. In the paper: the baseline has the longest pre-period; the triple-difference estimate (13.8bn) is close to the baseline; non-haven designs are de-emphasised because non-haven data start in 2010 (p.14); the balanced-bank estimate is 11.7bn (Table A2).

Why it matters, in detail. Overall share = w_d*s_d + (1 - w_d)*s_i with w_d ~ 0.5-0.54 (counterfactual direct share of haven household wealth ~0.49). The indirect evidence measures a level of remaining detectable noncompliance (1.9-5.2bn, Table A6) in a non-random 26-firm sample, not a change relative to a counterfactual; shifting from direct to indirect structures would also inflate s_d.

Evidence. Table 1 (p.47): direct 53.1 of 98.8bn household CRS wealth; haven direct 25.2 of 62.5bn. Abstract; pp.30-31; pp.33-34; fn 38-39; Conclusion p.35. July 2024 text pp.3 and 31 verified; the Oct 2024 version replaced the granularity caveat with 'our analysis is likely to be informative about the response of the overall stock' (pp.3-4); the 2026 text contains neither 'granular' nor 'opaque'.

Why this severity. Extending the headline to all offshore wealth rests on thin and partly flawed evidence (F6); if compliance on indirect holdings improved less, the overall share is below 70%. The paper's own fallback ('more than half') shows the gap is recognised.

The authors' response. In the paper (Section 8.2, pp.33-34): the same reporting rules apply to indirect holdings; shell-owned accounts show large repatriation (Fig A7); holding-company audits suggest detectable noncompliance (~3.5bn) similar to individuals'; fn 39: overturning 'more than half' would require much more haven wealth than AJZ estimate.

Why it matters, in detail. Under size-based risk selection the prefactor (A/N)/(sample mean assets) < 1, so E'' < HT; a correct IPW or Hajek estimate could exceed 2.3 / 1.9bn and even the 5.2bn 'quasi-random upper bound'. With n = 26 of 4,619 and unknown audit years and adjustment types, the indirect-holdings evidence is not informative about relative compliance.

Evidence. Algebra verified (Appendix B, eqs. 5-7, pp.52-53). Exact counterexample (100 firms with a = 1 and 100 with a = 9, half noncompliant with e = a, audit probabilities 0.2 / 0.8): true total 500, Horvitz-Thompson 500, eq. (5) 500, Hajek ratio A*sum(e/p)/sum(a/p) 500, eq. (6) 337.8 (-32%). In Table A6 the risk-weighted estimates are 45% and 37% of the unweighted 5.23bn (2.33, 1.93bn).

Why this severity. Affects the secondary claim (C6) that post-CRS noncompliance on indirect holdings is of the same order as on direct holdings, which underpins extending the 70% to all offshore wealth (F5); the Danish magnitude cannot be corrected without the audit data.

The authors' response. In the paper (p.53): the estimates likely overstate noncompliance because all assets of a noncompliant firm are counted as noncompliant and residual selection on unobservables is positive, so they 'should be seen as upper-bounds'.

Why it matters, in detail. Mixing a 2020 wealth effect, inflated by the 2020 yield drop, with a 2019 stock and a 2019 Panel B counterfactual is internally inconsistent; a consistent 2019 estimate is 4.1-4.9bn. A full 2020 accounting is not possible (CRS data end in 2019) and would raise Panel B to ~61bn at 2020 yields.

Evidence. Exact vector data, Fig 7(b): 2019 observed 13.59 vs trend 9.52bn (deviation 4.08bn, 43%); 2020 16.02 vs 10.01bn (6.01bn, 60%); 6.0/7.56 (2015 trend) = 79%. Digitised Fig 7(a): 2019 income deviation 91.2m, 2020 63.9m (printed 63.2m). 2019-consistent alternatives: 4.1bn (wealth space), 4.7bn (2019 income x 2019 factor), 4.9bn (x Panel B factor 54.2); 2020 income x 2020 factor 4.2bn.

Why this severity. Self-reporting's contribution (C3) falls from 6.0bn (20% of hidden wealth) to 4.1-4.9bn (14-16%), an 18-32% reduction; the headline moves only from 76% to 69-72%.

The authors' response. None; Fig 7's notes state that the treatment effect is the 2020 dummy without explaining its use in the 2019 accounting.

Why it matters, in detail. The DiD identifies CRS effects only if no other shock differentially affected own- vs other-account haven transfers after 2012. MT II-IV audits (2014-2019) targeted exactly such transfers, and the 2009 break shows that an enforcement shock can shift the own/other ratio by about 50% (from ~0.71-0.76 to ~1.13-1.14 of its 2012 level; DiD coefficients jump ~30 points). The direction of bias is ambiguous (MT could induce declared repatriation or deter observable transfers). If a quarter of the 2013-19 excess transfers reflected MT, repatriation would fall to ~10bn (C2 -25%) and the headline to ~73%; the constant-GDP-share aging of AJZ's 2007 stock (fn 35) is also undermined.

Evidence. SKAT (July 2016), 'Undersøgelse af indsatsprojektet Money Transfer II', pp.4-7: MT I collected data from 22 banks + 3 money-transfer firms on 50 countries for 1 Jul 2004 - 30 Jun 2009; audit phase from 1 Jan 2011; by end-2013 1,451 decisions, DKK 1.27bn revenue, 722 penalty referrals, 15 police cases; MT II (data 1 Jul 2009 - 30 Jun 2013 from 'udvalgte' banks; project 1 Feb 2014 - 31 May 2017) aimed to make control visible so that behaviour changes. Skatterådet ruling SKM2020.49.SKTST (28 Jan 2020, https://info.skat.dk/data.aspx?oid=2293315): MT IV data 2017-2019; purposes include showing 'at der er en opdagelsesrisiko'; this ruling labels the 1 Jul 2009 - 30 Jun 2013 'selected banks' window as MT I and MT II's as 1 Jul - 31 Dec 2013 (the two SKAT documents number the projects differently, but both list the same 2009q3-2013q2 selected-bank window). DR News (6 Sep 2010, B. Lambek, https://www.dr.dk/nyheder/penge/1000-milliarder-sendt-i-skattely): SKAT obtained all transactions between Danish accounts and 55 haven countries five years back; the minister called it potentially 'Danmarks største skattesag nogensinde'. Skatteministeriet, Aktuelle skattetal (20 Feb 2014): expected MT revenue ~DKK 1.5bn; MT 2 added 14 countries. These windows match the paper's data: haven transfers from 2005, the 19 smaller banks missing exactly 2009q3-2013q2 (p.11), non-haven data from 2010. Exact Fig A6(d) data (balanced panel of banks): own/other ratio 0.62-0.98 in 2005-08 vs 1.09-1.18 in 2009-11.

Why this severity. A plausible confounder of the attribution of repatriation to the CRS, contradicting fn 17 and the premise of fn 35; its effect cannot be quantified from public data. Because repatriation enters both numerator and denominator of the headline, even a sizeable change in it moves the 70% only a few points.

The authors' response. None located; the paper does not mention Money Transfer. Its design checks (amnesty participants excluded, Fig A6(a)-(b); balanced bank panel, Fig A6(d); non-haven placebo, Fig A7(a); triple difference 13.8bn, Table A2) difference out common shocks but not MT's targeting of own-account haven transfers.

Why it matters, in detail. share = 22.8/(30.1 + E), where E is wealth that left the reportable universe because of the CRS without being repatriated to Denmark; E is unobserved in the paper (transfers to third countries and non-financial assets are not measured), and cross-country evidence documents substitution toward non-AEOI havens and shells.

Evidence. p.8 (real estate, unlisted shares and precious metals excluded; 'portfolio reallocation ... towards offshore real estate'); pp.32-33 and fn 37 (only conservative omissions listed); July 2024 p.31 ('opaque structures'). Bénétrix, Emter & Schmitz (2024), 'Automatic for the (tax) people', Economic Policy 39(120), doi:10.1093/epolic/eiae041 (abstract, checked): AEOI reduced cross-border deposits, but households' deposits shifted to non-AEOI havens and shell-company networks grew. Tax Justice Network (2024), State of Tax Justice, ch.5 (corpus offshore-havens-07): CRS induced shifts into non-reportable equivalent assets.

Why this severity. A one-signed (upward) bias of unknown size in the headline share, e.g. 3bn of escaped wealth lowers 76% to ~69% and 10bn to ~57%; combined with the claim that the benchmark 'might be conservative', the presentation of omitted margins is one-sided.

The authors' response. No direct response located. In the paper: all large financial centres participate and 'all financial assets are covered' (p.6); fn 39 argues an overturn would require more haven wealth than AJZ estimate.

Why it matters, in detail. The 16.3% offshore rate (87.8m/540m, fn 30; 11.6% without the below-threshold extrapolation) should be compared with 14.9-17.7%, so it is not 'in between'.

Evidence. Kleven, Knudsen, Kreiner, Pedersen & Saez (2011), Econometrica 79(3), pp.668-671 (author PDF https://eml.berkeley.edu/~saez/kleven-knudsen-kreiner-pedersen-saezEMA11taxaudit.pdf), checked.

Why this severity. Comparison framing only; on a like-for-like amount basis offshore third-party-reported income is evaded about as much as self-employment income, which strengthens rather than weakens the paper's argument.

Why it matters, in detail. The top-1% share of cumulative repatriation is 76% rather than 'almost all'; the CRS-matched share of the self-reporter excess is 78%; the country-group wealth match rates that drive F1's composition adjustment differ by ~23 points between havens and OECD non-havens.

Evidence. Exact vector data from Figs 3(c), 5(c), 8 and A9(a); JRRSGL (6 Nov 2023 author version, Table 3) at BEA 2018 GDP $20,656.5bn; EC (2019) PDF pp.18, 49 (printed '12') and 106-107 (Box 7.1).

Why this severity. Descriptive overstatements, a transcription error and citation errors; none changes a conclusion.

Why it matters, in detail. Readers cannot interpret the column (5)-(8) coefficients as percentage-point effects on the compliance rate without knowing the transformation.

Evidence. Table A3 (PDF p.69), checked visually at 200 dpi: constant -0.6100*** (0.0347) in column (5).

Why this severity. A labelling or specification problem; the qualitative association between pre-filling and bank reporting (positive and significant in all columns) is unaffected.

Why it matters, in detail. Under standard finite-life amortisation and a dividend/interest-only tax ratio the ratio is ~4.9-7.0; the 'lower bound' label relies on excluding indirect-wealth benefits while also adopting cost conventions that favour a high ratio.

Evidence. Recomputed: B/C = 357.5/46.6 = 7.67; with 6.5/25.9 -> self-report revenue 15.9m and B/C 7.04; implementation principal 1.9m/0.0177 = 107m amortised over 5/10/20 years -> B/C 5.31/6.33/6.99 (4.88-6.42 combined with (i)). EC (2019) Box 7.1 (printed pp.69-70) gives the cited country ranges. Appendix C: 20bn x 1.85% = 0.37bn -> >=3.7x vs 'significantly less than DKK 100 million'; x 4% (Fig 4c) = 0.8bn -> >=8x.

Why this severity. The conclusion that benefits far exceed costs survives every alternative (ratio about 5-7); only the 'lower bound' and 'at least 10 times' characterisations are not established.

The authors' response. In the paper: benefits exclude any compliance gain on indirectly held wealth (an 'extreme assumption', p.34); the scaling and annuitisation are disclosed in the Table A5 notes.

Why it matters, in detail. Without the yield series and dividend/interest splits the 47.4bn Panel B and the 6.0 / 3.6bn capitalised gains cannot be reconstructed; the missing US balances are a second reason the stock (A) and the self-report base (B) cover different account universes.

Evidence. p.8 fn 8; p.21 fn 26; p.32 fn 36; Table A1 (PDF p.67); Table A5 notes (PDF p.71); Fig 7 data.

Why this severity. Prevents verification of the most sensitive input to the headline; the direction of the FATCA omission reinforces F1.

Why it matters, in detail. Either an undescribed ratio/regression estimator, a different confidence level, or an error; a corrected interval for the 3.6bn of detectable wealth would still be narrow (about +/-0.1-0.2bn).

Evidence. Recomputed from Table 2 (p.48): minimum SE with equal positive corrections within strata = 1.07m -> minimum 95% half-width 2.10m vs printed ~1.15m; count SE 82.2 -> (2,690; 3,012). Table 3 (p.49) components.

Why this severity. Point estimates and the audit-based conclusions are unaffected, and the audit component's uncertainty is negligible relative to Panel B's in the headline.

Why it matters, in detail. AJZ's own two estimates give haven-only values of 27.2-34.7bn (closure 66-84% with the paper's numerator) before the 2006 concealment and constant-share aging assumptions are questioned; the benchmark is consistent with both the paper's H and F1's corrected H, so it cannot validate either.

Evidence. AJZ (2019), AER 109(6), Table 2 and text, p.2090 (checked): benchmark $8.4bn, bottom-up $10.7bn (2007); '10 percent ... duly declared ... and 90 percent hidden'. Statistics Denmark NAN1 GDP 2007 1,743.1bn and 2019 2,303.6bn DKK; DNVALA USD 2007 5.4456: 45.7bn DKK -> 2.36% of GDP -> 54.4bn -> 27.2bn (reproduced); bottom-up -> 69.3bn -> 34.7bn.

Why this severity. An auxiliary check whose weakness does not change the estimate but removes the main independent support for the denominator; because F1's correction mostly adds non-haven wealth, the haven-only benchmark cannot discriminate between ~30bn and ~45-55bn.

The authors' response. In the paper (fn 35): the benchmark relies on 'completely different data sources' and takes AJZ's 2007 fraction 'with no modification whatsoever'.