Automatic Exchanges of Information and Offshore Tax Compliance: Evidence from Colombia

Why it matters, in detail. A DiD coefficient identifies the effect on treated units relative to controls. Scaling it by the treated plus control stocks implicitly assigns the full haven differential to the controls. Correcting only the base gives 23.8-24.0% (20.2% with the 2022 coefficient). Beta is an unweighted country-level log effect, and a value-weighted aggregate counterfactual from Fig. 3b could instead give 40-64% (F5). This correction therefore does not pin down the true share. It shows that the published 38% is not the number implied by the paper's own design and identifying assumption.

Evidence. Sec. 4.3, p.12: the coefficient 'increased self-reported offshore financial wealth in havens by about 92%'. Fn 19, p.12: 'Applying the estimated 92% increase to the 2016 stock of 7,600 billion pesos implies that CRS entry accounts for roughly 38%'. Adjudicator digitization of the raster embedded in Fig. 3b, p.20 (scripts/digitize_figs.py): 2016 havens 4,868-4,921 bn, non-havens 2,895 bn constant COP (sum ~7,800 ex-US vs 7,600 for CRS countries). (e^0.646 - 1) x 4,895 / 18,600 = 23.9% (both auditors: 23.8-23.9%); with the 2022 event-year coefficient (0.570), 20.2%. To rationalize the paper's 7,600 base, the non-haven stock would need an unidentified common CRS effect of about +20% (log 0.186), yet aggregate non-haven self-reported wealth fell 35% in 2017 (2,895 -> 1,882 bn). See scripts/adjudicator_checks.py, section 2.

Why this severity. Holding every other choice in the paper fixed, the internally consistent calculation lowers the deterrence share from 38% to about 24% (-37%) and the combined CRS-related share from 60-65% to roughly 46-52%. The existence of a 2017 jump in haven reporting is unaffected.

The authors' response. The paper calls the exercise a back-of-the-envelope calculation and states that it is 'likely conservative if offshore evasion also occurred in non-haven CRS jurisdictions' (p.13). It provides no estimate of a non-haven effect.

Why it matters, in detail. The haven x post-2017 coefficient identifies a haven-specific break, not a CRS effect. VDP disclosures, criminalization and the Panama shocks are all haven-weighted and coincide with the break. The VDP also feeds the outcome mechanically. Its foreign-asset disclosures (about COP 11 tn in 2022 pesos) are of the same order as the jump, although they cover all asset types, ~12% were made by firms, and some were in the US, non-financial or later repatriated. Without additional variation, the CRS-attributable share is bounded only by [0, 38%]. The US series jumped too, consistent with a country-agnostic domestic shock, although it is not a clean placebo because FATCA exchange began in 2016.

Evidence. DIAN Comunicado 217 (4 Dec 2017): VDP (normalization) tax base COP 3,481 bn (2015), 4,642 bn (2016), 12,791 bn (2017, to 10 Nov). Portafolio (25 May 2017), quoting the DIAN director: by 22 May 2017, 8,101 filers (7,547 individuals, who accounted for 88% of the COP 11.3 tn normalized in 2017 to that date) had declared COP 9.2 tn of omitted foreign assets; the finance minister tied the programme to the OECD multilateral exchange and the IRS agreement. Adjudicator: 9.2 tn nominal = ~11.0 tn Jan-2022 pesos (DANE CPI factor 1.1965). That is about equal to the whole 2016->2017 rise in all-jurisdiction self-reported financial wealth rebuilt from Fig. A.3 (13,308 -> 24,940 bn, +11,632), and larger than both the haven rise in Fig. 3b (+8,644) and the paper's deterrence amount (6,900 bn). Disclosed assets concerned holdings on 1 Jan 2017, the reference date of the 2017 F160, so assets kept abroad flow mechanically into the outcome series. Londoño-Vélez & Ávila-Mahecha (2021, AER: Insights 3(2), p.133) describe 'a large spike in disclosures made six months after Colombia criminalized tax evasion ... just before the disclosure window expired'. They present it as suggestive evidence on sanctions, and the main text does not mention the CRS. Their p.135 notes that the Colombia-Panama TIEA provided on-demand exchange from June 2017 and automatic exchange from 2018. Implied self-reported wealth in the US and never-CRS countries (Fig. A.3 total minus Fig. 3b) rose 72% in 2017 (5,518 -> 9,519 bn) while CRS non-havens fell 35%.

Why this severity. The CRS-specific part of the 38% is not identified. The final VDP round alone disclosed foreign assets of the same size as the entire 2017 jump, so the CRS share could be anywhere from small to most of the estimated effect. The combined 2017 transparency-and-enforcement response is real, and anticipation of CRS exchange plausibly contributed, so the qualitative finding survives in bundled form.

The authors' response. Body: the estimates 'should be interpreted as capturing the combined effects of CRS participation and contemporaneous policy changes' (p.9) and 'may partly capture contemporaneous policy changes' (p.13). The caveat is not carried into the abstract, the introduction's summary or the conclusion.

Why it matters, in detail. The imputation W = income / r makes the ratio linear in r. The relevant r is FATCA-reportable cash income over the value of the compared holdings. In 2021 that was plausibly 1-2% or less (0.04-0.14% on deposits, ~1.4% on Treasuries and equity dividends, 2.5% on 10-year corporates), versus the 3.37% used. Holding flows and self-reports fixed, the aggregate ratio becomes ~3.3-6.5x and the top-0.01% ratio ~3.6-7.1x. The paper's 0.6% 'reconciling' rate presumably refers to the gap left after its 'roughly half' indirect-ownership adjustment (3.37%/6 = 0.56%), and such cash yields were ordinary in 2021. The same capitalization drives the US share of firm-held AEOI wealth (70.6% of USD 10.9bn), so the 7.4%-of-GDP total in C1 is, if anything, understated.

Evidence. Fn 4 (p.4) rates 4.79/4.68/4.58/4.14/3.41/3.37% equal the annual means of the HQM 100-year spot rate (adjudicator, scripts/yields_check.py: 4.81/4.64/4.60/4.14/3.41/3.37). 2021 annual means: 3-month T-bill 0.04%, 10-year Treasury 1.45%, 30-year 2.06%, HQM 10-year 2.52%, HQM 30-year 3.23%, S&P Composite D/P 1.38%; FDIC deposit rates 0.05-0.14% (Auditor A). US-Colombia IGA (20 May 2015), Art. 1(1)(bb) and Art. 2(2)(b). Ratio(r) = 11 x r / 3.37: 6.5x at 2%, 4.7x at 1.45%, 3.3x at 1% (top 0.01%: 7.1x, 5.2x, 3.6x). Paper p.8: a 0.6% rate is 'an implausibly low figure'; 'using a 4% rate is more likely to overstate than understate wealth at the top'. F160 instructions (2022 form), box 37: 'el país donde se encuentra registrado el activo'.

Why this severity. With reportable yields in a plausible 1-2% range for a mixed 2021 portfolio, the abstract's 'more than elevenfold' falls to roughly 3-7x, a material change in a headline magnitude. The qualitative claim that FATCA misses much US-held wealth survives on extensive-margin evidence: 38.5% of the top 0.01% self-report US financial assets vs 14% in FATCA records, which does not depend on the rate.

The authors' response. p.8: capitalization errors 'cannot plausibly explain either the magnitude of the gap or its increase with wealth'; reconciling the top-0.01% gap 'would require a capitalization rate of about 0.6%, an implausibly low figure given evidence that rates of return rise with wealth'.

Why it matters, in detail. Numerator and denominator are not the same object. The numerator is cumulative campaign disclosure flows (all assets, current pesos, partly repatriated, partly non-financial); the denominator is the change in a constant-peso stock of offshore financial assets. Aligning asset scope is a confirmed 2.9 pp reduction. The repatriation LATE implies that more than half of induced participants moved at least some assets home, but the value share is unknown, so the further reduction is a scenario. The extrapolation assumes a constant proportional disclosure rate across the gap distribution and across campaigns, which is untested in either direction.

Evidence. Fn 20-21 (p.13). Table A.6 (p.XIII): value of disclosed assets in CRS countries 41.692 (control 17.13); financial assets in CRS countries 36.323 (13.7); financial assets in CRS havens 29.757; repatriation 0.225 (control 0.02); participation 0.398; F160 value in CRS countries in 2020 169.789 (bias-corrected SE 209.648, robust 192.17, not significant). Table A.5 col. 1 event years. Adjudicator (scripts/adjudicator_checks.py, section 5): all-asset level formula 22.1%; financial assets 19.3%; financial assets in havens 15.8%; retaining 75% / 50% of financial disclosures offshore gives 14.4% / 9.6%. The share of induced participants who repatriated is 0.57 on a participation basis, not by value.

Why this severity. The confirmed scope correction and the loss of the log-based upper bound reduce 22-28% to about 19%, a 14-32% cut. Netting repatriated assets would lower it further: about 10-14% if a quarter to a half of disclosed value was repatriated. With F1, CRS-related policies explain roughly 40-45% rather than 'nearly two-thirds' of the increase, before the attribution problem in F2.

The authors' response. p.13: 'These estimates are approximate ... the enforcement estimates extrapolate local treatment effects and omit intensive audits and other follow-up activities' (omitted audits would push the true effect up).

Why it matters, in detail. With few clusters, and few treated among them, conventional clustered SEs are likely too small, so the true interval is probably wider still (e.g., under a wild-cluster bootstrap). The pooled post-2017 average also mixes the 2017 jump with a decaying differential (0.79 -> 0.57); for a 2015-2022 decomposition, the 2022 coefficient is the natural choice. The unweighted country-level log effect is not the wealth-weighted aggregate effect, which is dominated by Panama and Switzerland.

Evidence. Table A.5 (p.XII), col. 1: pooled 0.646 (0.312); event years 2017 0.791 (0.382), 2018 0.692 (0.392), 2019 0.670 (0.373), 2020 0.583 (0.367), 2021 0.498 (0.355), 2022 0.570 (0.359); cols 3, 5, 7: 0.733, 0.622, 0.712. Adjudicator (scripts/adjudicator_checks.py, section 2): normal-approximation 95% CI of the share [1.4%, 102.8%] with the paper's base, [0.9%, 66.2%] with the haven base, and [-3.3%, 67.7%] with the 2022 coefficient and haven base. Value-weighted counterfactuals: 44%/40% with a 2015 base (2017/2022 end year), 56%/64% with a 2016 base. Both auditors computed the same CI; Auditor B did so in its uncertainty discussion rather than as a numbered issue.

Why this severity. Imprecision does not move the estimate in a known direction: value-weighting raises it, the haven base lowers it. But it means the 38% is not statistically pinned down, and readers of the abstract are given no indication that the data are consistent with a share between roughly 1% and 100%.

The authors' response. The calculation is labelled 'back-of-the-envelope' (p.12) and 'approximate' (p.13). Robustness specifications are reported in Table A.5 but not propagated to the share.

Why it matters, in detail. With mass at zero, the log(1+Y) treatment effect conflates the extensive and intensive margins and is not scale-invariant (Chen & Roth 2024, QJE 139(2):891-936). Footnote 20 additionally multiplies a geometric-type effect by an arithmetic control mean. Valid alternatives are the level LATE (41.7m, +243% of the control mean), a two-part model, or a scale-invariant normalization.

Evidence. p.2, p.11, fn 20 (p.13); Table A.6 Panel A. The arithmetic reproduces: e^1.99 - 1 = 631.6%; e^1.089 - 1 = 197%; 17.13 x (e^1.413 - 1) = 53.2m. Level estimates in the same table: CRS-country assets 41.692 on a 17.13 control mean (+243%); penalty 23.21 on 2.5 (+928%, not significant). Adjudicator two-point illustration calibrated to Table A.6 magnitudes (scripts/adjudicator_checks.py, section 5): identical behaviour yields e^b - 1 = 4.3% (Y in bn COP), 311% (m COP), 1,953% (USD) or 18,733% (COP), against a unit-free mean effect of 554%. The binned plot (Fig. 4f) shows the log outcome jumping from ~0.3 to ~1.6, consistent with a coefficient driven by the extensive margin.

Why this severity. The 632% and ~200% figures and the 28% upper end of the enforcement share are not identified. But level estimates in the same table show large positive disclosure effects, and the participation result does not depend on logs. The enforcement range collapses to its level-based lower end, and lower still after F4.

Why it matters, in detail. An insignificant average ATT, combined with failed pre-trend tests, few countries per cohort and a contaminated last-treated control, is weak evidence of a zero entry effect. The value point estimates are positive, and two of three key cohorts show excess entry-year growth in aggregate. Even a true zero would not discriminate between CRS anticipation and the country-agnostic 2017 VDP and criminalization shocks. The spillover nulls are low-powered relative to a 1% base.

Evidence. Table A.4 (p.XI); Fig. A.4 (p.IV); Table A.6 Panel C (p.XIII); LVAM (2021) p.135 on the TIEA. Adjudicator digitization of Fig. A.4b (scripts/digitize_figs.py): the 2017 cohort's value index rose 1.35 log points in its entry year, vs 1.03 for the Switzerland cohort and 0.77 for the Panama cohort. The Panama cohort rose 0.38 in its 2020 entry year, vs 0.09 and 0.18, partly reversed in 2021. The Switzerland cohort shows no excess in 2019. Fig. A.4c reads ~0.17 at k = -3 and ~-0.05 at k = +1, vs 0.238 and -0.111 in Table A.4 col. 2; Fig. A.4d reads ~0.09 and ~0.05, vs 0.227 and -0.010 (both auditors).

Why this severity. These results do not enter the headline numbers. But the timing null is the paper's main evidence that the 2017 jump reflects CRS anticipation rather than domestic shocks (F2), and that evidence is weaker and more mixed than stated.

The authors' response. p.10 notes that the confidence intervals for reported values are wider, and that the estimates 'rule out increases larger than roughly 6% in the number of declarations' (this reproduces: the upper bound is 5.9%). Fn 11 lists the identifying assumptions.

Why it matters, in detail. Measured discrepancies combine evasion, lawful valuation differences (historical TRM, fiscal cost vs market value), attribution artefacts (full balance to each CP or joint holder) and timing/threshold effects. The FY2021 comparisons appear to use the F160 for 1 Jan 2022, filed after the July 2022 instruction change, when the valuation wedge from the 2016-2021 depreciation was large. Over-reporting at the top (38% of the top 0.01%) shows that self-reports also contain items CRS does not capture, so gross under-reporting counts only one side. The paper quantifies none of these components.

Evidence. Paper: p.7 and fn 7; Fig. A.2 (p.II, +/-10% band, F160 threshold); Appendix B.2 (p.XIV); conclusion p.14. Law 1819/2016 art. 116 and art. 114 (ET arts. 269, 267); DIAN F160 FAQ no. 11 ('TRM de su reconocimiento inicial'); DIAN Oficio 907958/2021. DIAN Resolution 1096 (13 July 2022) re-worded the F160 value instruction for tax year 2022 to the ET valuation rules, in view of the art. 269 change. OECD (2017) CRS Standard, Commentary on Section I, para 13; HMRC IEIM402184/402186. TRM 31-12-2016 3,000.71; 31-12-2021 3,981.16. Adjudicator (scripts/table_checks.py): identified non-US CRS household wealth USD 6.45bn; 85% of it is USD 5.49bn, matching 2022 self-reports in CRS countries (22,200 bn / 3,981 = USD 5.58bn); the 15% net shortfall is 0.30-0.39% of GDP.

Why this severity. The individual-level under-reporting shares and the 0.9%-of-GDP figure are secondary descriptive results, and they are likely overstated by an unknown amount. The RD design is unaffected, because valuation noise is smooth at the cutoff. The aggregate 85% match also bounds how large CP double counting can be in the totals.

The authors' response. Fn 7 attributes part of the aggregate gap to the USD 25,000 reporting threshold, and Fig. A.2 uses a +/-10% tolerance. Valuation basis and deduplication are not discussed.

Why it matters, in detail. The scope restriction is stated in Sec. 4.3 but dropped in the abstract, introduction and conclusion. With fn 17's own totals, the CRS-related contributions explain 32-35% of the 2015-2022 increase in all self-reported offshore financial wealth.

Evidence. Abstract; p.3; p.12 and fn 17; p.14. Adjudicator rebuild from Fig. A.3 and Fig. 3b (scripts/adjudicator_checks.py, section 4): all jurisdictions 7,740 -> 42,339 bn; US + never-CRS 4,042 -> 19,839 bn.

Why this severity. The body defines the denominator correctly, and the CRS-country share is the natural one for a CRS effect. But the headline wording invites readers to apply the shares to all declared offshore wealth, which would double them.

Why it matters, in detail. A conversion slip: constant pesos are treated as nominal, and a single-year GDP denominator is used. The underlying peso series and Fig. A.3 are internally consistent.

Evidence. Adjudicator (scripts/adjudicator_checks.py, section 4; scripts/digitize_figs.py): 2015 CRS USD 1.11bn = 0.35% of GDP (printed 1.5bn, 0.6%); 2015 all USD 2.41bn = 0.76% (printed 3.2bn, 1.2%); 2022 CRS 1.86% (printed 2.0%); 2022 all 3.55% (printed 3.8%). Printed USD / USD 270.35bn = 0.55, 2.03, 1.18, 3.85%. Rebuilding fn 17's totals from Fig. A.3 (GDP of t-1, DANE CPI) gives 7,740 and 42,339 bn vs 7,800 and 42,300, confirming the figure and the constant-peso series.

Why this severity. The peso-based decomposition shares are unaffected. Only descriptive USD and %-of-GDP levels in the text are wrong; 2015 levels are overstated by about 35-70%.

Why it matters, in detail. Post-treatment conditioning: selection on an outcome the treatment affects. Identifying an intensive-margin effect for always-participants would require a principal-strata or bounds approach.

Evidence. Fn 16 (p.11); Fig. A.7 (p.VI); Table A.6 participation LATE 0.398. Counterexample in codex/scripts/audit_calculations.py (logical, not an empirical re-estimate).

Why this severity. This is a secondary, explicitly 'suggestive' result in a footnote and an appendix figure, and no core finding depends on it. The weaker reading the paper also offers (letter recipients disclose about what the tax authority knows) is consistent with the composition explanation.

The authors' response. Fn 16 and the Fig. A.7 note call the evidence 'suggestive' and note sensitivity to bandwidth.

Why it matters, in detail. Items (a)-(d) are wording and reporting inconsistencies; (e) is a documentation gap for inference.

Evidence. pp.2, 11-12; Table A.6 (p.XIII); Fig. A.8 (p.VII). scripts/adjudicator_checks.py, section 6.

Why this severity. Presentation only; the RD's qualitative findings and the 39.8 pp participation effect are unchanged.

Why it matters, in detail. Internal labelling inconsistencies. Item (d) understates the 2017 value effect in percentage terms (121% vs '79%').

Evidence. Fig. 2 legend (p.19); Table A.1 notes (p.VIII); Fig. 3 (p.20) vs Table A.5 (p.XII); p.9; Fig. A.1 (p.I); Fig. 1a (p.18). Visual checks by the adjudicator on 110-300 dpi renders.

Why this severity. Documentation and wording errors that do not change any estimate or conclusion.

Why it matters, in detail. Item (a) reverses the cited paper's main finding on timing and mechanism; (b) and (c) are loose citations.

Evidence. Johannesen, Larsen & Riedel (2026), 'Does global financial transparency improve tax compliance in developing countries?', Oxford CBT WP 26-02, abstract and pp.2-3 (https://oxfordtax.sbs.ox.ac.uk/sitefiles/wp26-02-niels-johannesen.pdf). Faye, Godar, Moura & Zucman (2025), 'Global Offshore Wealth, 2001-2023', EU Tax Observatory WP, Sec. 3.1 and conclusion. US-Colombia IGA (20 May 2015): preamble ('reciprocal automatic exchange') and Art. 2(2)(b).

Why this severity. These affect the literature framing, not any Colombian estimate. The South African contrast supports the paper's synthesis that gains require active enforcement, so it matters for interpretation, but the paper's own RD evidence on enforcement stands on its own.