Collecting the Tax Deficit of Multinational Companies: Simulations for the European Union

Why it matters, in detail. For 13 parents the domestic deficit is (r - tau_local) x Pi_local, with Pi_local = national-accounts profits of all non-foreign-controlled corporations, whereas the scenario and the other 10 parents use CbCR home-country profits of >= EUR 750m groups. The base is inflated (Polish MNEs' entire positive worldwide profit is USD 10.5bn vs USD 70bn of local profits), but the economy-wide tau need not equal MNEs' domestic ETR (German MNEs' 2017 CbCR domestic cash ETR is about 11%, against the 22.75% imposed). Direction: upward for Poland (at least EUR 8bn, about 5% of EUR 167.7bn), Spain's domestic part and Greece; downward for Germany and Sweden if their CbCR data are taken at face value (those data are themselves dividend-inflated, F2). Net effect on C1 plausibly 0 to -15%; on C7 country figures large (Poland overstated by a factor of 4-5).

Evidence. Code: 2b_twz_localfirms.do (TableA6 cols K-L; DE ETR = 0.2275) and 3b_imputation.do ll.119-132; report Appendix A (June PDF p.37). Adjudicator re-implementation (scripts/repro_macro.py): component = EUR 39.29bn at 25%, 23.4% of EUR 167.69bn (Poland 10.76, Germany 12.04, Spain 10.35). Poland: OECD CbCR Table IV (fresh OECD API pull) shows that Polish-parented groups' positive-profit sub-groups in all jurisdictions had USD 10.51bn of profit and USD 0.79bn of accrued tax in 2016 (USD 7.27bn in the <10% ETR bracket, 3.24bn in the >=10% bracket), so their 25% deficit cannot exceed about USD 2.2bn on an accrued basis (about EUR 2.25bn in 2021 terms; at most about EUR 2.7bn even with zero tax), against the report's EUR 11.0bn for Poland (124% of Polish CIT), which rests on USD 70.1bn of TWZ local profits at a 10% ETR. Later CbCR data processed with the report's own method (scripts/cbcr_later_years.py, 2021 EUR): Spain 2017 domestic EUR 6.8bn vs imputed 10.35bn (but Spain's total 14.3bn vs the report's 12.4bn); Greece 0.4 vs 1.4; Germany 2017 domestic 26.8 vs imputed 12.0 (total 46.3 vs 29.1); Sweden 2016 7.5 vs 0.7 (dividend-inflated, see F2). In the 10 CbCR countries, MNE domestic profits are 35.5% of TWZ local profits. Sensitivities (scripts/sensitivity.py): local base x0.355 -> EUR 142.3bn (-15.1%); x0 -> 128.4bn (-23.4%); Poland only x0.2 -> 159.1bn (-5.1%); German ETR 20% / 25% / TWZ's 10.57% -> 182.4 / 155.6 / 232.9bn.

Why this severity. The affected component is 23% of the headline, but defensible corrections move the EU total by roughly -5% (Poland alone) to -15% (base scaled to the MNE share at unchanged ETRs), and later CbCR data imply the imputation is too small for Germany and Sweden; the net effect on C1 is below a quarter and of uncertain sign beyond Poland, Spain and Greece. For the country results in C7 (Poland's 124% of CIT, the Eastern-European 'doubling') the problem is major.

The authors' response. No direct response. The authors' October 2021 update (EU Tax Observatory Note 2) replaced the TWZ-based estimates for Germany, Spain and Greece with 2017 CbCR data and called the TWZ-based figures 'rather cautious' (Germany at 15%: about EUR 5bn -> 13bn), but kept Poland at EUR 11.1bn at 25% from the same local-firm imputation.

Why it matters, in detail. Dividends received by parents and holdings enter Pi but not T, so tau_hat = T/Pi is biased down and the deficit max(r - tau_hat, 0) x Pi is biased up on both margins. Applying the authors' own Note 2 factor (0.6028, from NL/SE/UK 2017 adjustments) to the ten parents' domestic cells lowers C1 by EUR 25.0bn; the BEL-NLD artefact adds EUR 8.2bn; jointly with the max-rule (F4) and Poland (F1) corrections the 2016-data headline falls to about EUR 118bn (-30%), or about EUR 126-140bn after re-applying the measured cell-aggregation factor. Belgium's 115%-of-CIT result rests mostly on the single BEL-NLD cell.

Evidence. OECD (2020), 'Important disclaimer regarding the limitations of the country-by-country report statistics', pp.3-4 (Wayback 20200818; adjudicator copy); OECD Corporate Tax Statistics 2023, Box 5.2. Later national tax-administration notes linked by the OECD: Netherlands, 2016 positive CbCR profits EUR 24.8bn -> 15.6bn after removing dividends and other participation results (-37%), warning that neglecting these biases overestimates Pillar Two revenue; Sweden, tax-free dividends about SEK 200bn of SEK 342bn of 2016 domestic profit (58%); UK, GBP 55bn = 49% of 2017 domestic CbCR profit; Italy, 27% of FY2019 positive profits. 2016 domestic cash ETRs of EU parents: Luxembourg 0.9%, Ireland 3.1%, Belgium 4.1%, Austria 4.4%, Netherlands 11.6%, Italy 15.4%, France 16.3%. Adjudicator sensitivities (tax held fixed): the authors' later rule-of-thumb factor 0.6028 on OECD domestic cells -> EUR 142.7bn (-14.9%, -25.0bn); factors 0.9 / 0.8 / 0.7 -> -4.0% / -7.9% / -11.9%; BEL-NLD profit replaced by the authors' later margin-based USD 16.2bn -> EUR 159.5bn (-4.9%).

Why this severity. Documented and flagged by the data provider, but its 2016 size is known only for a few countries; plausible corrections lower the headline by about 5-20% (domestic dividend share 10-40% plus the Belgian cell), a noticeable but not transformative change that the downward aggregation bias partly offsets. Belgium's country figure falls by about 40% from the Belgian cell alone.

The authors' response. Later and partial. Note 2 (Oct 2021) states that included dividends 'can lead to artificially low effective tax rates and thus, an overestimation of potential revenues', adjusts NL, SE and UK domestic profits, proposes a -40% rule of thumb for all headquarter countries (EU 15% estimate EUR 80bn -> 67bn), and flags the 2016 Belgium-Netherlands cell as possibly 'overstating the Belgian revenue gains'. Intertax (2022) states the problem 'applies primarily to domestic observations' and reports EUR 42bn instead of 55bn (15%, long-run carve-outs) under the rule of thumb.

Why it matters, in detail. Cell-level aggregation understates sum_i max(r - tau_i, 0) Pi_i by Jensen's inequality, but Table IV brackets show the effect is only x1.1-1.4 at r = 25% (larger at 15%), whereas dividend inflation, out-of-scope domestic bases and max(OECD, TWZ) selection bias the estimate upward by comparable or larger amounts. The sign of the total error at 25% is not identified (the 2016-based figure is more likely high than low), and data-year volatility (2017 data: +40%) dwarfs the caveats the report discloses.

Evidence. Adjudicator test with OECD Table IV ETR brackets vs Table I cells (scripts/heterogeneity.py, accrued taxes): pooled 2016 EU parents, bracket/cell ratio 1.13 at 25% and 1.61 at 15% (2017-2023: 1.08-1.24 and 1.39-2.59); Auditor A's bracket-bound version gives 1.15-1.36 at 25% and 1.7-2.2 at 15% for 2016. Upward-bias corrections (scripts/joint_scenarios.py): no max rule + BEL-NLD fix + Poland bound -> EUR 142.8bn (-15%); plus the 0.6028 domestic factor -> 117.7bn (-30%); with x1.13 on OECD cells -> 154.0bn and 125.7bn. Data-year sensitivity: the authors' own Oct 2021 update with 2017 CbCR data gives EUR 234.3bn at 25% (+40%; Belgium 36.9, Germany 47.0) and 83.3bn at 15%. Ex post: the same method on actual 2021 CbCR data gives EUR 147.8bn vs the report's 151.6bn projection for the 15 overlapping countries (-2.5%), with country errors from -69% (Belgium) to +258% (Denmark).

Why this severity. It does not change the computed number, but it removes the claimed one-sided protection: with 2016 data the documented upward biases (about 15-30%) are at least as large as the measured downward aggregation bias at 25% (about 10-35%), so the headline is a central estimate with roughly +/-25-40% uncertainty, not a floor. C6 itself is not supported; C1's order of magnitude survives.

The authors' response. No direct response. Later work (Note 2, Oct 2021; Intertax 2022) presents dividend-adjusted variants as 'lower bound' rule-of-thumb estimates rather than revisiting the June 2021 characterisation.

Why it matters, in detail. E[max(X, Y)] >= max(E[X], E[Y]): choosing the larger of two noisy estimators of the same haven deficit biases upward, especially when one of them (the TWZ ownership allocation, F5) is a pro-rata FDI-stock key that can grossly misallocate across owners. For France the key assigns about 90% of the non-haven owners' net Belgium allocation (USD 42.7bn) to French groups, against USD 5.7-7.7bn of French-MNE profit booked in Belgium in CbCR, inflating France's haven deficit by EUR 7.8bn (2021 EUR).

Evidence. Adjudicator: without the rule the EU total is EUR 159.53bn (-8.16bn, -4.9%); France 18.35bn instead of 26.13bn; Denmark -0.36bn (scripts/sensitivity.py; matches both auditors). French CbCR (scripts/france_havens.py, OECD Table I, positive-profit sub-groups): profits in listed TWZ havens USD 20.9bn (2016) and 32.0bn (2017), never above 32bn in 2016-2023 (cash ETR 8-15%); Belgium, when listed separately (2017, 2018, 2020-2023), USD 5.7-7.7bn a year. TWZ's own Table U1 puts all shifted profits booked in Belgium in 2016 at USD 15.2bn, far below the USD 42.7bn its ownership key assigns to French owners there. In 2016 Belgium, Luxembourg and Ireland are hidden in France's 'Other Europe' aggregate, which the method already counts as non-haven, so the substitution partly double counts.

Why this severity. Raises the EU headline by about 5% and France's figure by about 42% over its CbCR-only value: noticeable for C1 and material for C7, not transformative.

Why it matters, in detail. Using allocated shifted profits Pi^s instead of booked haven profits Pi^b understates the base in aggregate (Pi^s < Pi^b), but the FDI-stock key ignores where shifted profits are actually booked and can exceed owners' booked haven profits (France). The '36% vs 40%' agreement is between a shifted share and a booked share and validates nothing. The disputed reconciliation of haven FDI income in TWZ (Blouin & Robinson 2025, corpus offshore-havens-09, unverified) would scale Pi^s further.

Evidence. Formulas checked in the authors' input file: TableC4!N20 = 1000*C105*TableC3!C18/TableC3!$C$91, with C105 = 'Table U1'!$AD$90 = 666.882. Belgium column: positive allocations USD 71.0bn vs net 48.8bn vs TWZ's own estimate of shifted profits booked in Belgium of 15.2bn (Table U1 col AD). Zeroing negatives adds only about EUR 0.9bn to the EU total. TWZ-based haven deficits used: EUR 20.28bn (13 imputed parents 8.93bn; max-rule substitutions 11.35bn, of which France 10.75bn), plus imputed non-haven deficits of 16.83bn built as a multiple (1.883 at 25%) of them: EUR 37.1bn, 22% of the headline, rests on this key. Germany's allocation (USD 38.6bn) is consistent with German 2017 CbCR (haven deficit EUR 7.8bn vs 5.9bn in the report).

Why this severity. Misdescribes the source of about 22% of the headline and removes the claimed cross-validation; the aggregate direction is ambiguous (shifted < booked biases down, while the ownership key can overstate individual countries, e.g. France via F4), so the demonstrated impact is mainly on country allocations.

Why it matters, in detail. Numerator/denominator mismatch: sum(TD)/T_CIT,all firms (52%) vs sum(TD_i)/sum(T_i) for the sampled groups (41%). On the same basis (deficit over own taxes) the macro method gives about 100% for EU CbCR parents, against 22.5% for EU-27 banks and 27.7% for EU-27 non-banks: a factor of about four. The micro data are at best uninformative and, taken literally, suggest the macro cells overstate deficits (consistent with F2).

Evidence. Adjudicator re-implementation of 8_MNE.do / 9_banks.do (scripts/firm_level.py): banks 41.106% (reproduces July), EU-27-headquartered banks 22.52% (32 banks), excluding HSBC 28.41%; non-banks 20.47%, EU-27 non-banks 27.66%. scripts/macro_vs_own_taxes.py: macro TD25 EUR 102.6bn vs own worldwide cash taxes EUR 99.1bn for AT, BE, DK, FI, FR, IE, IT, LU, NL, SI (103.6%; 92% counting only OECD cells). Report: June PDF pp.3-4, 7-8, 32; Table 4 pp.33-34. Note: the UK was an EU member in 2019, so 'EU banks' is defensible for the data year, but the macro comparison is with the EU-27.

Why this severity. Removes the independent support the report claims for C1 (the validation statement in C5 is not supported) but does not show C1 is wrong, since the firm samples are small and selected.

Why it matters, in detail. The key s_pj = UPR_pj / sum_j UPR_pj measures booking location. A destination key would shift collections from EU hubs to market countries and could raise or lower the EU-wide total (EU sales booked in non-EU hubs vs non-EU sales booked in EU hubs). Country-level first-mover gains for IE, NL, LU and BE are therefore overstated relative to the stated policy rule.

Evidence. US-parented groups' unrelated-party revenue booked in Ireland is USD 144.4bn (1.51% of their total), about the same as in Germany, USD 147.5bn (1.55%) (scripts/macro_vs_own_taxes.py); hence Ireland collects EUR 2.5bn from US firms in Table 3, almost as much as Germany (2.6bn), and 5.1bn from non-EU MNEs in Table C1 (Germany 5.7bn); Ireland's EU-only total is 229% of its CIT (June Table C1). IRS Form 8975 instructions group constituent entities by tax residence (Auditor B). Laffitte & Toubal (2022) document sales booked in havens.

Why this severity. Large effect on country allocations in scenarios 2-3 (first-mover gains of Ireland, the Netherlands, Luxembourg, Belgium) but an ambiguous and probably modest effect on the EU totals.

The authors' response. The report notes (PDF p.13) that, in implementation, sales information 'can be cross-checked against VAT filings'; the simulation contains no destination adjustment.

Why it matters, in detail. Because max(r - tau, 0) is convex, cell averaging hides more low-taxed sub-groups when r is close to cell ETRs (15%) than when r is well above them (25%). Bracket data imply the 15% deficit is understated by x1.4-2.6 versus x1.1-1.4 at 25%, raising the implied 15%/25% ratio from 0.29 to roughly 0.4. The policy inference ('unlikely to significantly increase revenues') leans on the most biased number.

Evidence. Adjudicator heterogeneity test: pooled 2016 EU parents, bracket/cell ratio 1.61 at 15% vs 1.13 at 25% (2017-2023: 1.39-2.59 vs 1.08-1.24); Auditor A's bounds 1.7-2.2 vs 1.15-1.36 (2016). The authors' own 2017-data update: EUR 83.3bn at 15% (ratio to 25% of 0.36, against 0.29 in the report). For context, the European Commission JRC (2025, JRC141119) estimates the enacted Pillar Two (15%, carve-outs, QDMTT) at EUR 26bn a year on average over 2017-2021 data (range 7.8-48.3bn), a different and narrower design.

Why this severity. Relative to the 25% figure, the 15% level could plausibly be about 40-60% higher from the aggregation bias alone (the upward biases of F1-F2 also apply at 15%), which materially changes C2's 'a quarter' and 'factor of two' statements; the qualitative point that revenue rises steeply with the rate survives.

The authors' response. The report itself warns that the 15% scenario is especially likely to be a lower bound (PDF p.29). The October 2021 update with 2017 data raised the EU 15% estimate to EUR 83.3bn.

Why it matters, in detail. Separately calibrated scale-ups (intra-EU export shares for scenario 3, extra-EU import shares for scenario 2) applied to different subsets break the monotonicity that the report's coalition argument requires; a common imputed parent-by-destination matrix would be needed for coherent comparisons. Counting regional aggregates as EU revenue and summing mutually exclusive unilateral scenarios inflates the presented EU-wide figures modestly.

Evidence. June Table 3 (PDF p.31) and Table C1 (PDF p.46); 6_unilateral.do ll.132-141; 7_EUcooperation.do ll.59-65. Auditor B's no-scale-up totals: EUR 183.7bn (EU-only) and 203.1bn (first mover). Adjudicator: US column of Table 3 sums to EUR 12.3bn for the EU; median and weighted shares recomputed from Table 3.

Why this severity. Changes C3's non-EU increment by up to about 40% (about EUR 12bn, roughly 6% of C3's 197.5bn) and C4's average by a few points; the conclusions (EU firms matter far more than non-EU firms; first-mover gains are sizeable) are unaffected.

The authors' response. The July revision relabelled Table 3's 'EU total' as 'EU average'. The report notes (PDF pp.8, 30) that a US minimum tax would reduce what the EU can collect from US firms.

Why it matters, in detail. D_nonhaven(15%) = 0.337 x 0.2825 Pi_h = 0.095 Pi_h vs D_nonhaven(21%) = 0.5136 x 0.11 Pi_h = 0.057 Pi_h: non-monotone in r, impossible for a fixed profit and tax distribution. A single rule (observed D21/D25 = 0.687) raises the EU 21% total by EUR 8.2bn.

Evidence. Adjudicator reproduction (component totals): TWZ non-haven deficits 5.665 / 3.365 / 16.826 / 33.653bn at 15 / 21 / 25 / 30%. Applying the observed-ratio rule also at 21% gives EUR 106.1bn (+8.4%) instead of 97.9bn (scripts/sensitivity.py; both auditors).

Why this severity. Only the 21% total moves (+8%); 'about EUR 100 billion at 21%' survives.

Why it matters, in detail. Numerator error only: the 2016 sales-based deficits of EU locations are set to zero in the ratio step; euro amounts are unaffected.

Evidence. July PDF pp.46-47 vs June PDF pp.45-46; 6_unilateral.do ll.70-76 and 7_EUcooperation.do ll.90-102; correct ratios 61.57% and 14.19% (Auditor B), equal to the June figures.

Why this severity. A calculation error in a secondary table that understates the scenario's percentage gains; the revenue amounts (EUR 197.5bn) are correct and the June version is right.

Why it matters, in detail. Firm-level inputs changed between versions and the June inputs are unpublished, so the June firm-level tables cannot be reproduced; the June statement on Shell's home-country ETR was wrong because participation income was included in pre-tax profit.

Evidence. Adjudicator comparison of the June PDF (MD5 8271863f..., Wayback 20210605) and the July PDF (Wayback 20210907; identical body to the HAL and simulator copies per both auditors); June PDF pp.22-23, 30-35 vs July pp.23, 31-35.

Why this severity. Corrections of secondary firm-level figures and wording; the headline is unchanged, but readers citing the widely cited June version get uncorrected firm figures.

Why it matters, in detail. Selective exclusion (-EUR 10-12bn relative to using the data) motivated by the very dividend artefact (F2) that also affects the retained parents' data.

Evidence. Full-text search of both versions; code. Keeping Sweden's data (as an aggregate-only parent) gives EUR 177.9bn (+6.1%; Auditor A: +7.0%). The Swedish tax agency's note puts dividends at about 58% of 2016 domestic CbCR profit, so the exclusion is substantively justified.

Why this severity. The choice lowers the estimate and is well founded, but it is undisclosed and applied to one country only, while the same dividend problem elsewhere is left uncorrected (F2).

The authors' response. Note 2 (Oct 2021, Appendix B) states that the first version based Sweden on TWZ 'because domestic effective tax rates for Sweden were implausibly low in the 2016 CbCR data' and then uses dividend-adjusted Swedish profits.

Why it matters, in detail. A one-year cash-tax snapshot records a documented payment deferral as low taxation; the code already replaces missing or negative bank ETRs by six-year averages but not a near-zero rate known to reflect timing. After-tax equity-accounted income has no matching tax line, biasing the ETR down.

Evidence. HSBC CbCR 2019, notes 3-4; HSBC CbCR 2020, note 3 (texts re-read by the adjudicator). scripts/firm_level.py: the Hong Kong cell is 23.30% of the bank deficit; using the six-year average Hong Kong ETR already in the authors' input (8.38%) gives 37.98% instead of 41.1%; pooling 2019-20 gives about 35-37% (Auditor B); excluding USD 3.1bn of after-tax associate income lowers the 2019 deficit by a further EUR 0.69bn (Auditor B).

Why this severity. Lowers the illustrative bank ratio by about 3-6 points (41% -> about 35-38%); C5 is in any case undermined by F6.

Why it matters, in detail. Most of the revenue comes from raising effective rates on home-country and non-haven profits to 25%, not from havens; this matters for interpreting the estimate as 'collecting what havens do not' and for its sensitivity to F1/F2.

Evidence. Adjudicator components at 25%: domestic 73.75bn (OECD 34.46 + TWZ 39.29), non-haven 36.95, haven 35.72, aggregate 21.27 (reproduces Table A1, PDF p.39). The decomposition is disclosed in Table A1, Figure 1 and PDF pp.25-26. German MNEs' cash ETR in Singapore is 7-13% in 2017-2023 CbCR, so the illustration itself is grounded.

Why this severity. A presentation point; the decomposition is disclosed in the body of the report.

Why it matters, in detail. None affects the implemented arithmetic; the EUR 340bn baseline is undocumented and makes the 'about 50%' and 'more than 50%' statements inconsistent.

Evidence. June PDF pp.3-5, 17, 23-25, 32-34, 36-37, 43, 45; July PDF p.23; input files MNE_cbcr2019.xlsx and CBCR_TABLEI_positive.csv; WEO April 2021 EU GDP USD 17,127.5bn / 1.2041.

Why this severity. Labelling and documentation slips that do not affect the computed results.