← All papers
Javier Garcia-Bernardo, Petr Janský, Gabriel Zucman · IMF Economic Review (2026), Policy Corner
Did the Tax Cuts and Jobs Act Reduce Profit Shifting by US Multinational Companies?
15 problems found: 1 major · 6 moderate · 8 minor
What the paper claims
- The share of US corporations' profits booked abroad declined after the TCJA, by 1-5 percentage points between 2017 and 2020, to about 27% for all US corporations, with agreement across sources (national accounts, BEA survey, IRS country-by-country data, Compustat).
- Part of the decline reflects repatriation of intellectual property to the US by six large firms (Alphabet, Microsoft, Facebook, Cisco, Qualcomm, Nike); these firms 'drive the macroeconomic decline'.
- The share of US multinationals' foreign profits booked in tax havens stayed stable at around 50% over 2015-2020 across sources (profit-type return about 45%, direct investment income about 55%, corrected CbC in between); 'No source suggests a significant immediate change'.
- Haven profits were about 13-15% of total (domestic plus foreign) US corporate profits throughout 2015-2020, a historically high level (below 5% in the 1980s-1990s); about 15% in 2017 (46% x 32%), falling back to 13% in 2020, i.e. a modest decline.
- Secondary: the sensitivity of reported profits to (low) tax rates increased after the TCJA, especially in havens, offsetting the weaker incentive from the rate cut; the TCJA's anti-avoidance measures 'may not be sufficient'.
Every problem found
1majorconfirmed · newly identified here · methodology
The paper says US companies moved 1-5 percentage points of their profits back home after the 2017 reform, but it measures this only against 2017, an unusually good year for profits abroad. Compared with the average of the years just before the reform, the share booked abroad in 2018-2019 was about the same or higher. After allowing for ups and downs in the economy, the fall is about 1-2 points and could be chance. The paper's broader point, that the reform changed little, is if anything strengthened.
Technical detail and evidence
Why it matters, in detail. The 1-5 pp is a raw t-1 difference in a volatile ratio series (SD of two-year changes 5.5-6.8 pp) whose base year sits at the top of its 2014-19 range in NIPA, BEA and Compustat. Conditional on rest-of-world and US growth and the dollar index (the paper's own Table A3 specification), the post-2018 shift in the BEA share is -1.3 to -2.0 pp, with p between 0.03 (HAC) and 0.31-0.33 (classical SE; trend-augmented). Relative to 2015-17 averages, raw 2018-19 changes are +1.9 (NIPA), -0.9 (BEA) and +2.9 pp (Compustat). The defensible descriptive decline is therefore at most the lower end of the range, and its sign within 2018-20 depends on the baseline. A decline relative to 2015-16 is visible only in 2021-23 BEA data (about -3 pp), outside the paper's window and coinciding with other reforms (end of the Double Irish grandfathering, Pillar Two anticipation) and the 2021-23 US profit boom.
Evidence. Adjudicator recomputation (scripts/baseline_cycle.py, country_patterns.py; output/foreign_share_series.csv, baseline_sensitivity.csv) from the authors' GJZ2021Data.xlsx (Data sheet, cols AR, Q, J, G, D, AT) and data/df_share_controls.csv. Change in foreign share, avg 2015-17 -> avg 2018-19: NIPA all-corporation +1.9 pp, BEA -0.9 pp, Compustat (plotted) +2.9 pp, Compustat raw +1.1 pp, Compustat positive-profit +2.8 pp; avg 2015-16 -> avg 2018-19: +3.3, +0.8, +4.6, +2.6, +4.6 pp. Relative to 2017: 2017->2019 -1.4 (NIPA), -4.6 (BEA), -0.9 (Compustat), -4.0 (CbC). Table A3 model re-estimated (const 35.79, ROW growth 2.31, US growth -3.83, R2 0.749, N 29): residual 2016 -6.3, 2017 +0.4, 2018 -0.6, 2019 -1.1, 2020 -1.8, 2021 -1.7 pp. Adding a post-2018 dummy: -1.75 pp (OLS SE 1.70, p 0.31; HAC SE 0.78, p 0.03); sample to 2019: -1.26 pp (HAC p 0.05); with a linear trend: -2.00 pp (SE 2.06, p 0.33). SD of two-year changes in the pre-reform period (start years 2000-2015): 5.8 pp (NIPA), 5.5 pp (BEA), 6.8 pp (Compustat). Later BEA data (outside the paper's window): 26.1% (2021), 27.7% (2022r), 28.8% (2023p) vs a 2015-16 average of 30.5%.
Why this severity. The upper half of the headline 1-5 pp range reflects 2017's cyclical high. The decline that survives the authors' own cycle control is about 1.3-2 pp and is not robustly significant, and with a multi-year baseline the raw 2018-19 shares are flat or higher. This changes the magnitude of C1 by well over a quarter, although the qualitative message that aggregate changes were small survives and is, if anything, reinforced.
The authors' response. Not addressed directly. The paper notes that the foreign share 'fluctuates in the short run' and reports that the cycle-adjusted residual decline is 'about 2 p.p. relative to 2017' (Sec. 3.1; Fig. A2; Table A3), which corresponds to the lower end of the headline range.
2moderateconfirmed · newly identified here · methodology
In 2020 the government paid companies hundreds of billions of dollars through the pandemic Paycheck Protection Program, and the official statistics count this money as US corporate profit. That makes the share of profits earned abroad look smaller in 2020 than it otherwise would. Without it, most of the reported fall for 'all US companies' disappears, although the paper's other sources still show some fall up to 2019.
Technical detail and evidence
Why it matters, in detail. With domestic profits net of corporate PPP subsidies, the 2020 NIPA foreign share is 29.3-31.4% instead of 27.35%. The 2017->2020 change therefore shrinks from -4.3 pp to between -0.2 and -2.4 pp, and haven profits as a share of total profits are 14.1-15.1% in 2020 (vs 13.2%), at or above 2017 (14.5%). In NIPA the COVID bias (subsidy-inflated domestic profit) runs opposite to the one in the BEA survey (collapse of parents' profits), which the paper acknowledges only for BEA. The clean all-corporation comparison is 2017-2019 (-1.4 pp), itself subject to F1.
Evidence. BEA FAQ 1408 ('How does the Paycheck Protection Program impact the NIPAs?'): forgiven PPP loans to businesses are classified as subsidies, and 'Subsidies are a subtraction in the calculation of GDI because of their implicit inclusion in operating surplus (proprietors' income and corporate profits)'. BEA, 'Effects of Selected Federal Pandemic Response Programs on Federal Government Receipts, Expenditures, and Saving, 2022 Third' (released 30 March 2023), annual sheet, line 47 'Corporate business': 285.2 (2020), 113.5 (2021). Recomputation (scripts/ppp_adjust.py) from the authors' workbook: 2020 share 27.35% -> 31.44% with all corporate PPP removed (29.25% with half removed); 2017->2020 change -4.28 pp -> -0.19 pp (-2.38 pp); Fig. 3 2020 value 13.17% -> 15.13% (14.08%), versus 14.48% in 2017.
Why this severity. It removes most of the all-corporation 2017-2020 decline and the 'about 27%' level, and reverses Fig. 3's 2020 fall. But the 1-5 pp range is still spanned by pre-COVID 2017-2019 changes (NIPA -1.4 pp, BEA -4.6 pp), and the adjustment brackets rather than estimates the no-PPP counterfactual. For these reasons it is rated moderate rather than Claude's 'major'.
3moderateconfirmed · newly identified here · methodology
For listed companies, the authors trimmed the biggest losses before adding up profits, but they described a different, two-sided trimming. That choice matters a lot in 2020: without it, listed companies' share of profits abroad rose slightly instead of falling by about 5 points. Up to 2019 the small fall is there whichever way the data are treated.
Technical detail and evidence
Why it matters, in detail. The bottom-1% cap on pre-tax income (pooled threshold) neutralises 2020's large domestic loss-makers. Compustat's 2017->2020 change therefore ranges from +0.8 pp (raw) through -4.9 pp (as plotted) to -7.2 pp (positive-profit firms, which also changes composition). The listed-firm leg of the claimed agreement across sources holds for 2017-19 (-0.9 to -2.0 pp) but is treatment-dependent for 2020.
Evidence. Zenodo package v1: 1a_processing_raw_data.ipynb cell 95 (winsorize(comp_sample_for['pi'],[0.01,0])); 2_figures_paper.ipynb cell 6 (comp['share_fo'] = comp['share_pi_fo_winsor_all']). Authors' aggregates in GJZ2021Data.xlsx (Data sheet, cols D, G, J): winsorized 51.23% (2017), 50.37% (2019), 46.32% (2020); raw 53.61%, 51.74%, 54.42%; positive-profit 48.73%, 46.75%, 41.50%. Fig. 1a note in the published article. Compustat microdata are not public, so a genuinely two-sided winsorization could not be computed.
Why this severity. It affects the 2020 endpoint of one of the four series (the -4.9 pp top of the range) and misdescribes the method. The Compustat decline to 2019 is robust to the loss treatment, and the BEA 2017-2019 change still reaches -4.6 pp. It is rated moderate rather than Codex's 'major' because the unadjusted series is not a superior alternative.
4moderateconfirmed · newly identified here · overclaiming
One of the four data sources was adjusted to match another, so it cannot independently confirm it. After the authors' own adjustment, that source actually shows the share of foreign profits in tax havens falling by about 7 points, which the text does not acknowledge. The government survey data, which need no such adjustment, do support the paper's finding that the haven share stayed stable.
Technical detail and evidence
Why it matters, in detail. The corrected CbC foreign share is mechanically (target foreign)/(target total) from the Compustat-based reconstruction, so it cannot corroborate Compustat. The corrected haven share is (raw haven - allocated deduction)/(raw foreign - total deduction), so its trend is driven by year-to-year swings in the estimated deduction ($43bn, $259bn, $155bn). As published it falls by 7.5 pp (13% of its level) over 2017-19, while BEA PTR and DII-type shares are flat.
Evidence. 1b_fix_double_counting.ipynb cell 5 (foreign deductions 246/457/331 $bn; country-level deductions 43/259/155 after subtracting hard-coded stateless amounts; Ireland excluded) and cell 7 (printed haven shares 56.2/51.1/48.8% of corrected foreign profits $596.0/658.8/602.5bn); Table A2 (IES WP 2025/28, p. 55); IRS SOI Tables 1A 2017-2019 (scripts/irs_cbc.py: raw haven share excl. stateless 58.1/61.0/54.8%). BEA comparison (scripts/bea_haven.py): PTR haven share 45.8/47.8/45.7%; BEA 'income' (DII-type) 57.5/57.7/56.4%. Auditors' alternative allocations (not re-run here): 2019 haven share about 42-56% (Codex); 2017-19 trend between -7.5 and +2.7 pp (Claude).
Why this severity. The level and stability of the haven share (C3) rest securely on BEA PTR and DII data, both of which are stable. The problem is that one of the 'independent' sources adds no independent information on the foreign share and, on the haven share, contradicts the stability claim, so cross-source corroboration is overstated.
The authors' response. The rescaling to Compustat-based totals and the allocation of foreign double counting to holding-company jurisdictions are disclosed (Sec. 2.2.4; Online Appendix B.1 and D.1).
5moderateconfirmed · newly identified here · methodology
The paper's estimate of how much profit six big companies brought back home is roughly right but somewhat too high, mainly because Microsoft changed its accounting rules partway through the period. On consistent numbers these firms moved about $49-56 billion rather than $61 billion. That they moved intellectual property back to the US is well documented.
Technical detail and evidence
Why it matters, in detail. Using restated Microsoft splits alone cuts the six-firm total by $5.75bn (-0.26 pp of $2,194bn); a common 2015-17 pre-reform window cuts a further ~$5-7bn; jointly the figure is 2.23 pp. Because the screen selects on large declines, the sum is a gross, selection-based decomposition rather than a bound. The net aggregate decline (F1) can fall short of the six-firm sum only because other firms moved the other way, which is what the data imply.
Evidence. Adjudicator rebuild from SEC XBRL company facts (IncomeLossFromContinuingOperationsBeforeIncomeTaxesDomestic/Foreign; scripts/six_firms.py; output/six_firms.csv): as-filed vintage, Table A4 baselines $61.327bn = 2.80 pp of $2,194bn (each firm matches Table A4 to within $2m); latest pre-2022 filings (restated Microsoft) $55.675bn = 2.54 pp; common 2015-17 baseline $56.351bn = 2.57 pp; both $48.988bn = 2.23 pp. Codex reports the same figures and the Nike fiscal-year alternative (calculations/sec_counterfactuals.csv). Sec. 3.2; Table A4 notes.
Why this severity. The firm-level IP-repatriation finding stands, but its quantification (2.8 pp, described as a 'lower bound') is overstated by roughly 10-20% on comparable accounting and baseline definitions, and the bound language is loose.
6moderateconfirmed · newly identified here · overclaiming
The paper's statistical test of whether companies became more responsive to low tax rates after the reform does not show a reliable change. The before-and-after differences are within what chance could produce, and they disappear when official tax rates are used instead of rates computed from the same profit data.
Technical detail and evidence
Why it matters, in detail. Point estimates move in the stated direction for the nonlinear and haven-interaction forms, but the post-minus-pre differences are 0.4-1.5 standard errors with country clustering, rest on a handful of haven observations, and use an ETR regressor that shares a denominator with the dependent variable. They are not evidence of a structural change in tax responsiveness.
Evidence. Adjudicator re-estimation on the package's BEA panel (scripts/regressions.py, regressions_statrate.py; output/regressions.csv), sample built as in notebook 2 cell 18: N 197/111 (published 198/112); log term -0.910/-1.735 (published -0.930/-1.453); haven x (1-ETR) 2.72/4.32 (2.66/3.94). Change in the elasticity w.r.t. (1-ETR) at ETR = 1%, clustered by country (52 clusters): log 70.5 (SE 47.2, p 0.14), quadratic 3.6 (3.2, 0.26), haven 1.2 (1.1, 0.28), linear -0.4 (0.8, 0.62); joint Wald p 0.29-0.62. With statutory rates: post interactions p = 0.67-0.76. Both auditors obtained the same results independently.
Why this severity. C5 is secondary (not in the abstract), but the paper draws a policy inference from it (the TCJA's anti-avoidance measures 'may not be sufficient'), and the evidence does not support it.
7moderateconfirmed · previously known · overclaiming
The paper's data describe what happened after the reform; they cannot show that the reform caused it, because other important rule changes, notably in Ireland, happened at the same time. In one place the paper says the reform played a 'central role'; elsewhere it says the observed change is only an upper limit on the reform's effect.
Technical detail and evidence
Why it matters, in detail. A before/after descriptive comparison without a control group cannot separate the TCJA from contemporaneous shocks, and the paper itself treats the observed decline as an upper bound on the TCJA effect. Dharmapala (2024, Sec. 3.3) shows that even this upper bound (3-5 pp in the NBER version) is fully accounted for by the rate cut under conventional semi-elasticities, leaving 'little role for the international provisions proper'.
Evidence. IMF ER Sec. 3.1 (last sentence), Sec. 3.2 (final paragraph), Sec. 4 (penultimate paragraph), Sec. 6; Abstract; NBER WP 30086 Sec. 4.1 (p. 12); Online Appendix Table A4 (Microsoft row).
Why this severity. This is interpretive overreach on the paper's central question (did the TCJA reduce profit shifting?). No number changes, but attributing the decline to the TCJA is not supported by the design and conflicts with the paper's own caveat.
The authors' response. The paper itself concedes (Sec. 3.2) that the observed decline is probably an upper bound for the TCJA's effect because of the contemporaneous Double Irish phase-out.
8minorconfirmed · newly identified here · internal inconsistency
The main chart labels one line 'all corporations' but actually shows a narrower group of companies, so it shows a much higher share of profits abroad (38% in 2020) than the 27% stated in the text. The direction of change is the same, so the conclusion is unaffected, but readers comparing the chart with the text would be misled about the level.
Technical detail and evidence
Why it matters, in detail. The plotted ratio is pre-tax USDIA equity income/(pre-tax USDIA equity income + domestic profits excluding S-corporations). The text/Appendix C ratio keeps S-corporation profits, and the workbook contains a third 'half S-corp' variant (38.5% -> 33.6%). The level comparison across populations in Sec. 3.1 is therefore made against a mislabelled line.
Evidence. 2_figures_paper.ipynb cells 6 and 9; national_accounts_WrightZucman2018Appendix.xlsx, sheet T.A1 (col. S: 42.11% in 2017, 40.70% in 2019, 38.11% in 2020); GJZ2021Data.xlsx, Data col. AR (31.63%, 30.24%, 27.35%); published Fig. 1a image; NBER WP 30086 Fig. 1a (PDF p. 15). Scripts: wz_ta1.py, gjz_workbook.py.
Why this severity. The trend is nearly the same (-4.0 vs -4.3 pp) and the numbers in the text are correct, but the main figure misstates the level by about 10.7 pp and visually contradicts the text's comparison of populations.
9minorconfirmed · newly identified here · arithmetic or transcription
A check meant to strip out the effect of exchange-rate swings was programmed backwards. Fixing it barely changes the result: the fall up to 2019 is still about 4 points in that data source.
Technical detail and evidence
Why it matters, in detail. The inverted adjustment doubles rather than removes translation effects. The correctly signed constant-2017-rate series implies a 2017-19 BEA decline of 4.1 pp vs 4.6 pp raw (authors' version: 4.9 pp), a 2020 change of about zero and a 2021 decline of 9.2 pp.
Evidence. 3a_robustness_data_processing.ipynb cell 5 (exch_rate_2017 = value/value_2017; pi_corrected = pi/exch_rate_2017) with GEM series 'Official exchange rate, LCU per USD, period average'; data/bea_sf_combined.csv; scripts/fx_check.py; output/fx_check.csv.
Why this severity. This is a coding error in a robustness check. Once fixed (with hyperinflation handled), the check still supports the authors' statement that exchange rates do not explain the 2017-2019 decline: FX accounts for about 0.4 pp of it. It is rated minor rather than Codex's 'moderate'.
10minorconfirmed · newly identified here · arithmetic or transcription
A number copied from an older data release was used in one correction step, leaving about $11 billion in tax havens that should have been removed. The effect on the paper's conclusions is small.
Technical detail and evidence
Why it matters, in detail. $11bn too little is subtracted from haven and conduit jurisdictions in 2019. Correcting it lowers the corrected CbC haven share to about 47.2%, which slightly enlarges the 2017-19 decline discussed in F4.
Evidence. 1b_fix_double_counting.ipynb cell 5 ('double_counting_for_19-176*mult'); IRS SOI 19it01acbc.xlsx, row 'Stateless entities and other country', profit column = 164,969,526,720 (whole dollars); Table A2 col. M ($166bn); Codex and Claude both computed the 1.5 pp effect.
Why this severity. It changes the 2019 CbC haven share by about 1.5 pp and does not affect the aggregate foreign share or the BEA-based haven evidence.
11minorconfirmed · newly identified here · internal inconsistency
A few small errors in the case-study table: one company is listed twice and one percentage is wrong. They do not change the paper's numbers on how much profit moved.
Technical detail and evidence
Why it matters, in detail. The count of large listed firms with sustained 20-pp drops is 22 of 219 (10.0%), not 23 (10.5%). The Nike drop is misreported in the table but not in the dollar calculation.
Evidence. Online Appendix Table A4; IMF ER Sec. 3.2 and fn. 3-4; Fig. 1b; scripts/six_firms.py (Nike drop 83.9 pp; Table A4 dollar amounts reproduced); 22-firm total $92.6bn / $2,194bn = 4.22 pp (consistent with counting Newmont once).
Why this severity. These are transcription and labelling slips; the upper-bound arithmetic counts Newmont once and the dollar totals are right.
12minorconfirmed · newly identified here · internal inconsistency
A number of figures quoted in the text were not updated to match the final tables. None of them changes the paper's conclusions.
Technical detail and evidence
Why it matters, in detail. The C4 range is better stated as 11.5-14.6% over 2015-2020 (13.3-14.6% in 2016-2019). The other items are labelling and vintage carry-overs.
Evidence. Online Appendix P0033-P0035, P0355-P0356, P0363, Table A1, Table A2 (IES WP 2025/28 p. 55) and Fig. A10; IMF ER Sec. 1 and Sec. 4; GJZ2021Data.xlsx Data col. BQ (2015 11.52%, 2017 14.48%, 2018 14.64%).
Why this severity. These are presentation and carry-over inconsistencies that do not change any conclusion; the largest affects the stated 13-15% range for 2015.
13minorconfirmed · newly identified here · documentation gap
In several places the computer code does not do exactly what the text says, for example in which places count as tax havens. These differences slightly change some levels but not the paper's conclusions.
Technical detail and evidence
Why it matters, in detail. Under the stated seven-haven list the BEA PTR haven share is 39.6-49.3% over 2015-2020 (42.5/43.8/43.5% in 2017-19), versus 44.7-50.8% with the code's list. The 2020 BEA 'spike' is +4.0 pp over 2019 with revised data rather than +7.7 pp; with revised data, the BEA 2017->2020 change is -0.6 pp.
Evidence. 2_figures_paper.ipynb cell 4 (haven lists); 1a cell 90; 3a cells 2, 5 and 8 (INDICATOR NY.GDP.MKTP.KD); GJZ2021Data.xlsx col. O; data/bea_sf_combined.csv; scripts/bea_haven.py (PTR haven shares by list); README.md.
Why this severity. These are documentation and definition mismatches that shift levels slightly (haven list) but not trends or conclusions; the replication package makes them discoverable.
14minorconfirmed · newly identified here · internal inconsistency
The paper's description of which individual tax havens gained or lost profits does not match the data for the full period. This does not affect the overall finding about havens as a group.
Technical detail and evidence
Why it matters, in detail. The narrative matches CbC 2017->2018 changes, which are distorted by the large 2018 double-counting correction, rather than 2017->2019 changes or the BEA series. Shifts within the haven group (out of Bermuda, Jersey and Cayman; into Ireland and the Netherlands) are larger than the text conveys.
Evidence. scripts/country_patterns.py (output/country_patterns.csv) computed from the authors' corrected panel (combined_dataset_stateless_False_all_False.tsv).
Why this severity. This descriptive detail does not affect the aggregate haven share.
15minorplausible · previously known · methodology
Some profits recorded in places like Ireland or Singapore reflect real business there, so 'profits booked in tax havens' is not the same as 'profits artificially shifted'. Because the paper mainly compares changes over a few years, this caveat matters little for its conclusions.
Technical detail and evidence
Why it matters, in detail. Location shares conflate shifted profit with returns to real activity and intangibles located in havens (the narrow-versus-broad notions of shifting distinguished by Dharmapala 2019). The stable 2015-2019 activity shares limit the bias in trends but not in levels.
Evidence. IMF ER Sec. 1, Sec. 4 and Fig. 2 (panels B-C); corpus entries offshore-havens-10 (Dharmapala 2019, AEA Papers and Proceedings 109: 488-492) and offshore-havens-13 (Forstater 2018, CGD Note).
Why this severity. This is an interpretive caveat. Over a short window with stable real-activity shares, changes in haven profit shares remain informative, and the paper reports the activity shares.
The authors' response. The paper reports real-activity shares in havens alongside profit shares (Fig. 2, Figs. A9-A10) and frames its statistics as profit location; Torslov, Wier and Zucman (REStud 2023) respond to the general definitional critique elsewhere.
Claims examined and rejected
Problems one auditor raised that did not survive checking.
- Reproduction R2: the Fig. 1a 'All corporations (NIPA)' series reproduces exactly (26.8/27.7/31.6/30.9/30.2/27.35% in 2015-2020). — The published Fig. 1a plots a different series that excludes S-corporation profits (Wright-Zucman T.A1: 42.1% in 2017, 38.1% in 2020; notebook 2 cells 6 and 9; the published figure image). Claude reproduced the text/Appendix C series, which is correct but is not the plotted line (see F8).
- Qualcomm's estimated 2020 shift should be about $13.6bn rather than Table A4's $4.7bn, because its pre-fall foreign shares exceed 100%. — Pooled three-year shares (sum of foreign over sum of total pre-tax income), the method in Table A4's notes, reproduce $4.697bn exactly from SEC XBRL data (scripts/six_firms.py); the larger figure comes from averaging annual ratios. The substantive caveat (US losses during the Apple dispute; the April 2019 settlement) is retained in F5.
- I10 (version_instability): the headline and definitions changed between the NBER (2022) and IMF (2026) versions without explanation (3-5 pp -> 1-5 pp; 'across-the-board' -> 'not broad-based'; Puerto Rico dropped from the BEA haven share). — These are working-paper-to-journal revisions that move toward caution and coincide with new data (2020 BEA survey, revised 2019); the published version is internally consistent on these points, and such revisions are not normally itemised. They are recorded as context in final.md, not graded. The one revision-introduced error (the Fig. 1a series swap) is F8.
- Published Table A3 coefficients differ from those produced by the package's data and code. — The package's saved notebook output (3a_robustness_data_processing.ipynb, cell 9) reproduces Table A3 exactly (35.7525, 2.3246, -3.8321, -0.0350). Only re-running on the shipped CSV yields third-decimal differences (35.79, 2.31, -3.83, -0.03), which are immaterial.
- Strong form of I1: 'Measured against the pre-reform average, the foreign share did not fall in 2018-2020 in any series.' — Partly rejected. It holds for raw NIPA and Compustat against 2015-17 averages, but the BEA 2018-19 share is 0.9 pp below its 2015-17 average. The authors' cycle model also implies a 1.3-2.0 pp post-reform decline relative to the full pre-period, with 2016 (not 2017) as the outlier. The weaker form, that the decline is small, fragile and baseline-dependent, is retained as F1.
- Ireland's share in direct-investment-income data fell (-0.3 pp, 2017-2019), contradicting the text's 'consistently across sources'. — Not reproduced with the authors' DII-type series (BoP income + foreign taxes), where Ireland's share rises slightly from 10.64% to 10.76%. The figure likely reflects a different BEA vintage or measure. The broader point that the country narrative is inaccurate stands (F14).
- I1 (major): the lower-tail Compustat adjustment 'reverses the trend', and the unadjusted share is a defensible alternative measure for 2017-2020. — The misdescription and the 2020 sensitivity are confirmed (F3), but severity is reduced to moderate. The unadjusted series is dominated by extreme losses (83.8% foreign share in 2001, 130.8% in 2008), all three loss treatments show a 0.9-2.0 pp decline for 2017-2019, and only the COVID-year endpoint flips.
- I4 (moderate): the currency-robustness calculation 'does not establish what is claimed'. — The inversion is confirmed (F9), but a correctly signed constant-2017-rate series (Venezuela held at nominal USD) still shows a 4.1 pp 2017-2019 BEA decline (raw 4.6 pp), so the stated conclusion that exchange rates do not explain the decline holds for 2018-2019 and 2021. Severity reduced to minor.
- I2 rated major: PPP subsidies make the 2017-2020 decline an artefact. — The distortion is confirmed (F2), but severity is reduced to moderate: the 1-5 pp range is still spanned by pre-COVID 2017-2019 changes (NIPA -1.4 pp, BEA -4.6 pp), and the PPP adjustment brackets rather than estimates the counterfactual 2020 domestic profit.
Audit notes
What reproduced. Reproduced (adjudicator, independently of the auditors' code): the NIPA all-corporation foreign share (31.63% -> 27.35%, 2017-2020) and the Appendix C example; the BEA-survey share (35.52% -> 30.96%, 2017-2019); the Compustat aggregates as supplied (51.23% -> 46.32%); the corrected CbC foreign and haven shares; the PTR and DII-type haven shares; the Fig. 3 haven/total series (14.48% in 2017, 13.17% in 2020); the Table A3 regression; the six-firm $61.33bn and 22-firm $92.6bn totals (from SEC XBRL); and the published FX-adjusted series. Raw IRS CbC tables 2017-2023 and BEA AMNE 2022r/2023p were parsed for out-of-window context. Not reproduced: the published Fig. 1a NIPA line, which is a different (S-corp-excluded) concept from the text's series; the exact Table A5 coefficients and sample (N 197/111 vs 198/112); Nike's 36.4 pp drop (data give about 84 pp). Not reproducible from public inputs: the Compustat firm screen, the missing-foreign-profit imputation and the private-firm (20% of average) imputation behind the CbC targets, and Orbis-based statements. Robustness recomputations: multi-year baselines remove the raw decline in NIPA and Compustat; the cycle-adjusted post-2018 decline is 1.3-2.0 pp (p 0.03-0.33); removing PPP subsidies raises the 2020 NIPA share to 29.3-31.4%; the correctly signed FX adjustment gives a 4.1 pp 2017-2019 decline; and the six-firm total on comparable accounting/baselines is $49-56bn (2.2-2.6 pp).
Limits of this audit. Compustat and Orbis microdata are proprietary, so the Compustat series, the Compustat-based CbC targets, the missing-data imputations and the firm screen could not be rebuilt; the authors' supplied aggregates were used and the six firms were checked with SEC XBRL data. The PPP adjustment brackets (half to full removal) rather than estimates the counterfactual effect of subsidies on 2020 profits, and uses the March 2023 BEA vintage of PPP allocations, while the authors' NIPA data are from an earlier vintage. NIPA/ITA components (portfolio dividends, current-cost adjustment) were taken from the authors' workbook rather than re-downloaded. BEA AMNE 2022r/2023p tables, IRS CbC 2016-2023 files, SEC company facts and BEA PPP files were read from the shared cache (official files downloaded by the Claude auditor) and parsed by the adjudicator; 10-K passages were not re-read by the adjudicator (both auditors verified them). The exact published regression sample could not be recovered. The HAL version of the working paper was not obtained. Which pre-reform baseline is 'right' is a judgment. F1 reports the sensitivity across reasonable baselines and the authors' own cycle model rather than a single corrected estimate. Later data (BEA 2021-2023; IRS CbC 2020-2023), outside the paper's window, show a fall in the foreign share and in the haven share (BEA PTR seven-haven share 35.5% in 2022r and 39.1% in 2023p vs 42.5-43.8% in 2017-19; raw IRS haven share 46-51% in 2021-23 vs 55-61% in 2017-19). These are reported as context and were not used to grade the paper's period-specific claims.