Camille Landais, Thomas Piketty, Emmanuel Saez · Institut des politiques publiques (IPP) methodological guide, January 2011 (HAL halshs-02512804)

Le modèle de micro-simulation TAXIPP - Version 0.0

12 problems found: 2 major · 3 moderate · 7 minor

What the paper claims

TAXIPP 0.0 simulates, for a representative synthetic population, all French compulsory levies — including the often-omitted indirect taxes, corporate tax (IS) and payroll taxes — plus transfers for 2005-2010, with levy totals calibrated to national accounts.
The synthetic source files (about 800,000 virtual adults) were drawn so as to be 'parfaitement représentatifs' of the population, reproducing the joint distributions of surveys and tax files. They give a fine decomposition of the top 10%, shaped within the top 1% by exhaustive high-income return files (Landais 2007, 2009).
Module validation: simulated masses and receipts closely reproduce observed ones.
Documented incidence and imputation conventions determine how levies absent from tax returns are distributed.
Illustrative macro factor tax rates frame the labour-vs-capital debate: primary income is taxed at about 45%, labour at about 50% vs capital at 25-30%. Excluding contributory contributions, labour is at about 35% vs capital at 25% ('plutôt 35-40% vs 25-30%'). The CSG base misses about 50% of financial and 70% of property income.
Downstream result (not printed in the guide, but generated by the documented model and posted results; the reason the note matters): the 2010 total tax rate on primary income by fractile, for full-time workers aged 18-65 (TableSTI5 / book Graph 1), is roughly flat at 48-49% from P50 to P95 and falls within the top percentile.

Every problem found

1majorconfirmed · previously known · speculative assumption

Companies pay corporate tax, and the model has to decide which households ultimately bear it. It spreads the tax over everyone with savings, including holders of tax-free passbook accounts, rather than over shareholders, and it gives no reason for this choice. If shareholders bear it, as the authors themselves have described the tax, the richest group's rate rises from about 33% to 39% and the drop at the top shrinks by about a third. If all owners of capital bear it, homeowners included, the drop grows somewhat.

Technical detail and evidence

Why it matters, in detail. With IS on equity income only (dividends plus retained profits), re-ranked, the top-0.001% rate rises from 33.4% to 39.1%. P99.9-99.99 rises from 38.3% to 42.2% and P99-99.9 from 43.7% to 45.5%. The P90-95-to-top decline falls from 15.5 pp to 9.9 pp (-36%). TAXIPP 0.1's base, which excludes exempt savings, changes little (top 33.9%, decline 15.1 pp). Allocating IS over all capital including net rents lowers the top rate to 31.1% and lengthens the decline to 17.8 pp (+14%). IS enters numerator and denominator one-for-one and leaves disposable income unchanged, so these changes are exact for the IS block (wealth-based taxes held fixed). On the secondary concept, equity incidence cuts the decline from 7.3 pp to 2.3 pp.

Evidence. Guide p.19 and p.90. do_revcap.do l.124-128. Parametres 2010: masse_is_cn 35.0; interest 42.05; dividends 71.41; life insurance 49.97; retained profits 24.17. TAXIPP 0.1 guide §1.4.5 pp.13-14 and fn.8. Bozio, Guillot & Laffeter (2015) pp.44-45. Landais, Piketty & Saez (2012), Revue de l'OFCE 122, p.150. IS formula reproduced to €0.45; re-imputations in final/scripts/03_sensitivity.py (variants IS-1, IS-2, IS-3).

Why this severity. A defensible alternative removes about a third of the top-end decline in the model's main output (three-quarters combined with F1), so the magnitude rests on an untested convention. The uncertainty runs both ways: all-capital incidence including housing steepens the decline. The direction therefore survives, and this is not critical.

The authors' response. The 0.0 guide does not discuss the choice. TAXIPP 0.1 (Nov 2012) calls all-capital incidence a common, CBO-like assumption that may over- or understate the tax's redistributive effect. Bozio et al. (2015) keep it 'en l'absence de consensus'. The authors' 2012 OFCE reply treats IS as a précompte on distributed profits.

2majorlikely · previously known · methodology

To match national statistics, the model adds a large amount of company income that does not appear on tax returns, and it gives most of it to the people who declare dividends, who are mostly the very rich. About two-fifths of that 'dividend' total is really money that owners of unincorporated businesses draw from their firms. France's statistics office and the authors' own later model count it as self-employment income. Correcting this raises the richest group's tax rate by about 4-5 points and shrinks the drop in tax rates at the top by a third or more.

Technical detail and evidence

Why it matters, in detail. D4222 is 38.9% of the D42 benchmark (27.75 bn in 2010). It is allocated through declared dividends, which are concentrated at the very top, so it raises the top's denominator with no IRPP or CSG attached. Moving it to the self-employment (kpro_cn) key, holding other taxes fixed and re-ranking, raises the top-0.001% rate from 33.4% to 38.8%. With IS re-spread on the reduced base the rate is 37.8%; with INSEE's three-quarters convention, 37.4%. The P90-95-to-top decline falls from 15.5 pp to 9.0-10.7 pp, and to 4.1 pp when combined with shareholder incidence of IS (F2). On the secondary-income concept the decline falls from 7.3 pp to 0.7 pp. Side effects: CN16's 'about half of financial income escapes the CSG' becomes 47.8% instead of 55.7%, and the 2010 macro capital tax rate (method a) rises from 29.2% to about 30.6%. Uncertainty: who owns the quasi-corporations is unobserved. The remaining dividend gap (only ~37% of national-accounts dividends appear in tax data) is also allocated proportionally, and the direction of that bias is ambiguous.

Evidence. ComptesNationaux.xls CN15 col. K (code 'D42+D43 reçu', 70.674 / 71.412); raw-insee-income.xls T_3401 rows 36-38 (2009: D42 70.674 = D421 43.211 + D4222 27.463); do_revcap.do l.98 and l.125-126; guide pp.19-21, 59-60. INSEE definition of D42 (insee.fr/fr/metadonnees/definition/c2267). Bellamy, Consales, Fesseau, Le Laidier & Raynaud (2009), INSEE DESE working paper G2009/11, annex A10.1 ('20 Md€ sur les 43 Md€'; 'ne conserver qu'un quart ... le reste est basculé en revenus des non salariés'). Bozio, Guillot & Laffeter (2015), Économie et Statistique 481-482, p.47 and Table 2. INSEE T_7401 (base 2020). Re-imputations in final/scripts/03_sensitivity.py (variants D-1, D-2, D-3) and 08_macro_tables.py.

Why this severity. The misclassification inflates untaxed imputed income at the very top, so its effect runs one way. Correcting it with INSEE's own convention, or the authors' successor model's, removes 31-42% of the P90-95-to-top-0.001% decline in the model's main output. The decline survives, so this is major rather than critical.

The authors' response. The guide (p.21) argues that proportional imputation of retained profits probably understates very high capital incomes, because large fortunes keep profits in holding structures (simulated fortunes peak at about €2 bn). No response on the D4222 component was found. The successor model TAXIPP 0.3 reclassified D422 as mixed income.

3moderateconfirmed · previously known · methodology

Pension and unemployment contributions mostly buy workers their own future benefits, and the richest pay relatively little of them, so whether they count as taxes changes the picture a lot. The guide offers both ways of counting. Counted as deferred pay, the drop in tax rates at the top is about half as large, about 7 points instead of 15, though it does not disappear.

Technical detail and evidence

Why it matters, in detail. The contributory wedge enters only the numerator. It is capped and not levied on capital income, so it is 15.5-18 pp of the rate for P0-P90 (14.5 pp at P90-95) but only 0.4 pp at the top 0.001%. Moving to the secondary concept lowers P0-10 from 40.5% to 27.4% and the P90-98 peak from about 49% to 40.5%, while the top 0.001% barely moves (33.4% to 33.2%). The decline shrinks from 15.5 pp to 7.3 pp; combined with F1 or F2 it nearly vanishes (0.7-2.3 pp).

Evidence. Guide pp.61-64 and p.96 (fn.197-198); do_autres_impots.do l.155-158 and l.227-229. Recomputed on the headline sample, ranked by y_seca (final/scripts/02_reproduce_sti5.py): P0-10 27.4%; P90-95 40.2%; P97-98 40.5% (peak); P99-99.9 39.0%; P99.9-99.99 36.5%; P99.99-99.999 34.2%; top 0.001% 33.2%. That is a 7.3 pp decline from the peak, against 15.5 pp on the primary concept. Sterdyniak (2012), Revue de l'OFCE 122, pp.110-111, writes that without contributory contributions P98-99 and the top 0.001% are both taxed at 33% and 'la prétendue régressivité disparaît'. His P98-99 figure is wrong. Numerator-only removal gives 47.11 - 11.26 = 35.9% (STI5 workbook), and consistent removal gives 40.4%, against 33.0-33.1% for the top 0.001%. Codex examined this point as a critique check and confirmed its economic importance without listing it as an issue.

Why this severity. The alternative halves the top-end decline, but the guide states and defends its convention and supplies the alternative concept, and a fall of about 7 pp within the top 1% survives. The reversal of the bottom-versus-top comparison bears on the companion book's claims rather than on this note.

The authors' response. Guide p.63, remark ii: contributions are pure taxes for those who end up on the minimum pension. The 2012 OFCE reply answers the transfers point (the book's scope is levies only) but not this convention.

4moderateconfirmed · newly identified here · internal inconsistency

The model assumes the richest save about half of their income, more than the 40% the guide describes. Saving more means spending less and so paying less VAT. Using the guide's own stated figure would raise the top group's tax rate by about 1 point, or about 2 points if the extra top-end saving were dropped.

Technical detail and evidence

Why it matters, in detail. Capping saving rates at the described 40%, with the aggregate re-normalised, raises the top-0.001% rate from 33.4% to 34.2% and cuts the decline from 15.5 pp to 14.7 pp. Removing the top doubling altogether gives 35.1% and 13.9 pp (-10%); combined with equity IS (F2), the decline is 8.7 pp. The top-end gap is insensitive to the consumption-incidence share α itself (15.4-15.7 pp for α in [0,1]).

Evidence. Guide pp.40-41; do_autres_impots.do l.170-188 (reproduced to €25); final/scripts/03_sensitivity.py (variants S-1, S-2 and the saving-rate table).

Why this severity. The bias is one-directional but modest, 5-11% of the top-end decline; this is at the low end of moderate.

The authors' response. Guide p.40: a flat 25% rate within the top 5% 'n'est pas crédible'. TAXIPP 0.1 fn.58 concedes the assumption is arbitrary for lack of studies.

5moderateconfirmed · newly identified here · documentation gap

Everything rests on a synthetic population of 821,815 adults built from confidential data, and the recipe was never published. Its outputs look right where they can be checked, but nobody outside can rebuild or audit it.

Technical detail and evidence

Why it matters, in detail. Nonlinear taxes and the composition of fractile incomes depend on joint distributions (income by type × household structure × hours) that aggregate matches do not identify. Because the 2010 file is a rescaled 2006 file, the 2010 top-end composition (e.g. 27.5% realised gains in the top 0.001%) is inherited from 2006. External validation is limited to margins and top shares.

Evidence. Guide pp.7-15; posted program and file listings (original-index.html, programs-index.html); do_mise_a_jour.do. Output checks: counts exact (final/scripts/04_validation_numbers.py). 2006 top shares of declared income excluding capital gains are 32.9 / 8.70 / 2.49% for the top 10 / 1 / 0.1% (06_top_shares_2006.py), against about 32.3 / 8.3 / 2.2% for 2005 in Landais (2007, Fig. 3).

Why this severity. The model's distinctive input, the joint distribution behind the fine top-decile decomposition, cannot be reproduced or audited. Its outputs look plausible where they can be checked.

The authors' response. The guide stresses that the files are virtual so as to respect statistical secrecy, and it lists the primary sources (pp.7-11).

6minorconfirmed · newly identified here · selective presentation

The guide sums up its national figures as labour taxed at about 50% and capital at 25-30%. That fits 2010, a year when corporate tax receipts were depressed, but not 2005-2008, when capital was taxed at 32-34% by the guide's preferred method. Once pension contributions are set aside, labour and capital were then taxed at nearly the same rate.

Technical detail and evidence

Why it matters, in detail. Including contributory contributions, the labour-capital gap is 15-17 pp (method a, 2005-08) and 19-21 pp (2009-10), against an implied 20-25 pp. Excluding them, it is 1.4-4.9 pp under method a and 5-9 pp under method b, against ~10 pp. Reclassifying D4222 (F1) would move the 2010 method-a rates to 47.6% for labour and 30.6% for capital.

Evidence. ComptesNationaux.xls CN20a rows 11-16 and CN20b rows 8-13 (final/scripts/08_macro_tables.py); guide pp.61-64; Parametres (masse_is_cn).

Why this severity. This is a secondary, explicitly illustrative result. The sign of the contrast holds in every year and under both methods, and the hedged statement fits 2010; only the size of the gap is overstated for the earlier years.

7minorconfirmed · newly identified here · documentation gap

Some of the 2010 data files posted for replication were left by the program's last step in a 'what if wages rose 10%' state. They show, for example, income tax of €72 bn instead of about €52 bn. The published tax-rate results come from a different, correct file and reproduce exactly.

Technical detail and evidence

Why it matters, in detail. This is a replication-hygiene problem. The 2010 module-level outputs carry a 10% + €100 wage increment and no national-accounts calibration. The baseline could be regenerated in Stata from the posted 2006 originals and parameters, but not from the posted 2010 files.

Evidence. final/scripts/01_posted_file_state.py: for 79.0% of single wage-only declarants earning €15-60k, taxable wages equal 0.9×(1.1×wage + 100) to the euro; none equal 0.9×wage. IRPP is €71.8 bn (2006 file: €54.1 bn); reduc_ds is 0 (guide: €7.0 bn); ts = csp_fac = 0 in cotsoc; PPE is €2.65 bn net and €1.53 bn refundable (guide: €4.5 / 3.2 bn). TableSTI5_2010.dta has ts = 0 and components that do not sum to the total (P50-60: 46.6% vs 48.1%). The published SimulationsTousImpots.xls is internally consistent; the all-tax rates of the two tables differ by at most 0.154 pp.

Why this severity. The published rates are unaffected and reproduce exactly. The defect makes the guide's 2010 validation figures unverifiable from the posted files and would mislead anyone using them.

8minorconfirmed · previously known · methodology

The model estimates wealth from investment income, and it treats a one-time stock-market gain as if it were a permanent income stream. A few people therefore look far richer than they are and are assigned more wealth and inheritance tax. Because the same people are at the top of the income ranking, the error makes the top look slightly more heavily taxed, not less.

Technical detail and evidence

Why it matters, in detail. Removing capitalised gains from wealth, with DMTG recomputed exactly and net ISF re-assigned by rank, lowers the top-0.001% rate from 33.4% to 32.1% and P99.99-99.999 from 34.9% to 34.2%. The convention thus narrows the decline by about 1 pp. This is the classic uniform-yield capitalisation problem, here applied to realised gains.

Evidence. final/scripts/05_capitalised_gains.py. The DMTG rule is reproduced to €11; ISF re-ranking is approximated by weight-aware rank matching, which preserves the net ISF mass (3.499 -> 3.505 bn).

Why this severity. The totals are calibrated, net ISF plus DMTG is 5.1% of top-0.001% income, and the convention raises rather than lowers measured top rates. It is therefore conservative for the top-end decline.

The authors' response. Guide p.38: the ISF program is 'relativement rudimentaire', and the shape of the wealth distribution should be improved.

9minorconfirmed · previously known · overclaiming

The '2010' picture is really the 2006 population scaled up to forecasts of 2010 totals. The guide says so, but its claim that the files are 'perfectly representative' is too strong.

Technical detail and evidence

Why it matters, in detail. The headline-year profile is a 2006 joint distribution scaled to forecast 2010 masses. Realised gains, 27.5% of top-0.001% income, are assumed to rebound to their 2006 level (€19.8 bn).

Evidence. Guide pp.7-8, 12-13, 48-50, 65; Parametres (masse_rfin_pv_*; masse_sal_irpp); final/scripts/00_params.py and the mass check in final.md.

Why this severity. The projections and the rudimentary updating are disclosed, and the effect on fractile rates is unquantified and probably small.

The authors' response. Disclosed in §2.2.2 ('très rudimentaire') and §4.2.1 (projections).

10minorconfirmed · newly identified here · misrepresentation of source

The guide says the new 2010 business tax exempts reinvested profits; the tax rules of the time show it does not. This weakens the guide's argument for treating that tax like VAT, but it has almost no effect on the results.

Technical detail and evidence

Why it matters, in detail. A gross-product-type value-added tax burdens capital more than a consumption-type VAT does, so the VAT analogy understates the capital share of TP/CVAE incidence; TAXIPP 0.1 concedes this for the pre-2010 TP. Shifting the whole TVA+TP block between consumption and factors (α from 0 to 1) moves the top-end decline by only ±0.2 pp.

Evidence. Guide p.57 fn.20; BOI 6 E-1-10 §29 and §38 (bofip-archives.impots.gouv.fr, cached by codex); TAXIPP 0.1 guide §1.4.4.

Why this severity. This is explanatory text. The model allocates aggregate receipts with the 70/30 split, and the top-end gap is insensitive to that split.

11minorconfirmed · newly identified here · arithmetic or transcription

A handful of figures in the text are slightly off from the authors' own files or from public sources; none changes any conclusion.

Technical detail and evidence

Why it matters, in detail. All discrepancies are small (≤ about 2 bn or a few percent). The fortune understatement (5-15 vs 17-23 bn) reinforces, rather than weakens, the guide's own point that the model misses the very top.

Evidence. final/scripts/04_validation_numbers.py, 08_macro_tables.py, 00_params.py; Parametres (masse_reduc_ds, masse_reduc_irpp); press reports of Challenges' 2010 list (La Dépêche/AFP and Le Figaro, 7 July 2010, cached by claude).

Why this severity. These are rounded orders of magnitude in a technical note, and none affects a conclusion.

12minorconfirmed · newly identified here · documentation gap

The guide has typos, wrong labels, broken cross-references, and a few places where the code does something slightly different from what the text says. Someone rebuilding the model from the guide alone would make mistakes, but the posted code resolves them.

Technical detail and evidence

Why it matters, in detail. Anyone re-implementing the model from the guide alone would mis-code marital status, the TVA aggregate, the DMTG and ISF imputations, and the incidence base of the factor part of TVA/TP. They could also not reconstruct the current-system tax-rate variables without reading do_autres_impots.do and do_nouvel_ir.do.

Evidence. Guide pp.16, 39, 43, 53-56, 69-75, 81, 90, 92, 96-99; do_autres_impots.do l.53-77, l.161-229; do_isf.do (non-payers' imputation); do_nouvel_ir.do l.124-128; ComptesNationaux.xls CN8!C6 and CN10 row 5; indiv_demo_2010.dta (mat, marie).

Why this severity. These are documentation problems. The posted code resolves them, and none changes a result.

Claims examined and rejected

Problems one auditor raised that did not survive checking.

No major issue: the D4222 ('other distributed income') dividend-calibration problem is moderate because ownership of the other distributions is not established. — Rejected on severity. The classification error is confirmed by INSEE's definition and the vintage table T_3401. INSEE's own 2009 micro-macro reconciliation moves three quarters of the component to self-employment income, and the authors' successor model treats it as mixed income. Even under INSEE's convention the top-end decline falls by 31% (by 42% under full reclassification), and the effect is one-directional. Graded major/likely (F1).
Corporate-tax incidence is one element of a combined 'top-tail robustness' issue of moderate severity (I2). — Partly rejected. IS incidence is a distinct mechanism whose effect (-36% of the decline under shareholder incidence, +14% under all-capital incidence) is as large as F1's, so it is graded on its own as major (F2). The saving increment is treated separately (F4).
The posted 2010 module files being in the post-MTR state (I3) is a moderate issue. — The finding is confirmed (79% exact matches to the incremented-wage rule, 0% to the baseline rule), but downgraded to minor (F7): the published rates come from a different, uncontaminated file and reproduce exactly.
The effect of capitalising realised gains in wealth on the fractile profile has an ambiguous sign (I7, moderate/plausible). — Corrected. Removing capitalised gains from wealth (DMTG recomputed exactly, ISF re-ranked approximately) lowers the top-0.001% rate by 1.3 pp, so the convention is conservative for the top-end decline. Downgraded to minor (F8).
Under method a the macro capital tax rate is 29.2-34.0% in 2005-2010 (I9). — Corrected: the 2009 method-a rate is 27.2%, so the range is 27.2-34.0%. The underlying point is retained and developed in F6.
Imputed income at the top (I4) is a separate moderate/plausible issue with uncertain direction. — Merged into F1. The imputed share (46% of top-0.001% income) is confirmed, and its demonstrably misclassified component (D4222) makes the issue major/likely. The direction of the residual proportional-allocation bias remains ambiguous, as A said.
The guide's IS convention contradicts the claim that all assumptions minimise top-end regressivity (I1). — Not held against this work. The 'toutes les hypothèses ... minimiser la régressivité' claim is made in the companion book (p.54), not in the guide. Several of the guide's documented choices are in fact conservative for the decline: capitalised gains in wealth (F8), DMTG concentrated above P95, and the exclusion of offshore wealth.

Audit notes

What reproduced. Reproduced exactly from the posted 2010 micro-files: the headline profile (TableSTI5 / book Graph 1) for 20.222 m full-time workers, matching TableSTI5_2010.dta to four decimals and the published workbook within 0.154 pp (174 synthetic records in the top 0.001%). Also reproduced: the file counts (821,815 records; 50.366 / 35.560 / 32.219 / 28.899 m); the 2010 income masses against their calibration targets; the IS formula (to €0.45), the saving/consumption block (€25) and the DMTG rule (€11); the 2006 module figures (other reductions €3.50 bn, tax reductions €5.24 bn, PPE €5.74 bn / €4.06 bn, 10.11 m beneficiaries); the CN5 levy totals (835.1 / 790.0 bn), with INSEE's 793.0 bn for 2009 verified in the PLF 2011 RPO; and the CN20a/b macro rates. Not reproducible from what is posted: the generation of the 2006 base (no program published), and the 2010 module validation figures (the posted 2010 module files are in the post-MTR state). Small discrepancies: QF parts 1.725 vs 1.74; PPE mean €568 vs €400; D291 24.41 vs 22.5 bn (2008); 2010 'other reductions' 7.5 vs 7.0 bn. Not verified: INSEE's 834.4 bn for 2008 (the RPO 2010 could not be retrieved). Every sensitivity result is a partial-equilibrium re-imputation on the posted outputs; the Stata pipeline was not re-run.

Limits of this audit. No Stata was available, so the model was not re-run from the 2006 originals. All alternatives are partial-equilibrium re-imputations on the posted 2010 outputs, holding wealth-based taxes fixed except in F8, where ISF is re-ranked approximately and the bouclier's income dependence is ignored. Ownership of the D4222 quasi-corporations is unobserved, so the F1 reallocations are defensible alternatives, not proven corrections. Later national-accounts vintages (base 2014/2020) are used only to document the reclassification. The following were not audited in detail: INSEE's 2008 levy total (834.4 bn), the PPE budget figures and PLF forecasts, the Enquête Emploi/ERFS imputations, and the transfers, TH/TF and MTR modules. The Carbonnier (2007) journal version was not read. The 2006 top-share comparison with Landais (2007) is approximate (2005 values read from figures). The successor-model (TAXIPP 0.3) and IPP 2023 comparisons involve different years, tax laws, concepts and data. The downstream profile (C6) is not printed in the guide; it is treated as core because the guide is its methodological documentation.