← All papers

Emmanuel Saez, Joel Slemrod, Seth H. Giertz · Journal of Economic Literature 50(1): 3-50

The Elasticity of Taxable Income with Respect to Marginal Tax Rates: A Critical Review

15 problems found: 3 moderate · 12 minor

What the paper claims

Every problem found

1moderateconfirmed · newly identified here · internal inconsistency

The paper says that if top earners respond twice as strongly to taxes, the hidden economic cost of raising each extra dollar from them doubles, from about 34 cents to 68 cents. By the paper's own formula it roughly triples, to about $1.03. The paper's main figures are fine, but this one 'what if' check makes higher top tax rates look less costly than the paper's own method implies.

Technical detail and evidence

Why it matters, in detail. MEB(e) = eK/(M-eK) is convex. With eK/M = 0.253 at e=0.25, doubling e gives an MEB of 1.03, a threefold rise, not 0.678. The correct figure means that at e=0.5 each marginal dollar raised from the top 1% carries more than a dollar of excess burden. The welfare cost of top-rate increases is therefore much more sensitive to the top-earner ETI than the footnote suggests. The main-text MEBs ($0.339, $0.195) are approximately reproducible from public 2005 inputs under eq. (6) and are not affected.

Evidence. JEL p. 42 and fn. 72; eq. (6), p. 8. If $0.339 follows eq. (6), then -dB/dM = 0.339/1.339 = 0.2532 and MEB(0.5) = 0.5064/0.4936 = $1.026 (scripts/formulas.py). The adjudicator checked that $0.339 does follow eq. (6). Inputs: the 2005 Piketty-Saez Pareto a at P99 on income excluding gains (1.602, TabFig2008 Table A4_pareto) and tau = 35% federal + 5.9% deductible state = 38.8%. With these, eq. (6) gives $0.341 at e=0.25 and $1.035 at e=0.5. The across-the-board MEB computed from IRS SOI 2005 Table 3.4 (returns by top marginal rate) is $0.151 with federal rates and $0.185 when a 5% deductible state rate is added, against the paper's $0.195 (scripts/meb_2005_check.py). The alternative reading, that $0.339 is the revenue-loss share -dB/dM, would require a top-1% marginal rate of 46-48%. That is above any 2005 federal+state rate, so this reading is implausible. If the welfare numerator counted state revenue losses while the denominator counted only federal feedback, the consistent value would lie between $0.68 and $1.03. Either way, $0.678 is too low under any net-revenue denominator. (Inputs a=1.5 and tau=35% happen to give $0.677 at e=0.5, but the same inputs give $0.253, not $0.339, at e=0.25. The word 'doubles' and the exact factor of 2 point to linear scaling.) The MEB figures and fn. 72 are new in the JEL; the 2009 WP has neither.

Why this severity. This is the only sensitivity analysis offered for the conclusion's welfare-cost figures, and it understates the top-1% MEB at e=0.5 by about a third ($0.68 vs about $1.03). But it sits in a footnote and does not affect the framework, the identification discussion, or the e=0.25 figures in the main text.

2moderatelikely · previously known · speculative assumption

The paper's headline example says the top tax rate could go as high as about 73% before revenue starts to fall. That rests on assuming top earners respond fairly little to taxes. The paper's own evidence suggests they respond more. With its higher estimates, the revenue-maximizing rate is closer to 54-62%: still above today's rate, but not 'far' above it. So the conclusion that there is lots of room to raise top rates is more fragile than it looks.

Technical detail and evidence

Why it matters, in detail. tau* = 1/(1+ae) and MEB = eatau/(1-tau-eatau) are steep in e. Using the taxable-income centre of the cited 'best' estimates (about 0.30-0.40) instead of 0.25 lowers tau* to 62-69% and raises the top-1% MEB to $0.44-0.68. Using top-group estimates (0.5-0.6) gives tau* of 53-57% and an MEB of $1.0-1.4. The benchmark therefore biases the reported revenue-maximizing rate upward and the welfare cost downward, by roughly 15-25% of the tau* level within the paper's own evidence. Some evidence points lower: Saez's (2017) medium-run estimate of 0.26, the predicted-MTR estimates of 0.14-0.24, and the upward bias of trend-controlled time-series estimates. The true top-earner long-run ETI is thus genuinely uncertain, and that is the point: 72.7% is one point in a wide band.

Evidence. JEL pp. 8-9, 15-16, 19, 39, 42 and fn. 72; WP pp. 5-7 and 58. Sensitivity with a=1.5: tau* = 72.7% (e=0.25), 62.5% (0.40), 57.1% (0.50), 53.9% (0.57), 44.8% (0.82). tau* falls to 42.5% only at e of about 0.90 (0.85 with a=1.6). Top-1% MEB, scaled consistently from $0.339 via eq. (6): $0.68 (e=0.40), $1.03 (0.50), $1.37 (0.57). Pareto a at P99 (TabFig2008): excluding gains 1.602/1.591/1.575/1.606 (2005-08), 1970s mean 2.43; including gains 1.452/1.435/1.413/1.479, 1970s mean 2.15 (scripts/pareto_a.py, scripts/formulas.py). Neisser (2021, EJ; ECONtribute DP 067) meta-analysis: after-deduction ETI mean 0.403 (median 0.353), before-deduction mean 0.287 (median 0.185). She finds selective reporting, with 'a slight incline to report values between 0 and 0.4', and quotes SSG p. 42. Later evidence points both ways. Lower: Saez (2017, TPE 31), medium-run top-1% elasticity 0.26 for the 2013 reform. Higher: Weber (2014, JPubE), 0.858 with a mean-reversion-robust estimator (from her research-page abstract).

Why this severity. The framework and the range itself are unaffected. The qualitative claim that the current top rate is below the revenue-maximizing rate also survives for any top-earner ETI below about 0.9. But within the paper's own evidence, tau* falls by 11-19 pp (to 54-62%) and the top-1% MEB doubles or triples, so the policy calibrations are materially less certain than presented. The issue would be 'major' if 72.7% were treated as the paper's headline. It is graded moderate because the abstract makes no numerical claim, the authors label the calculation illustrative ('proceeding mechanically'), and fn. 72 flags higher top-earner ETIs.

The authors' response. No direct response located. The paper itself concedes (fn. 72) that the top-earner ETI 'could conceivably be higher than 0.25' and calls the calculation mechanical. Saez (2017, Tax Policy and the Economy 31) later reports a medium-run top-1% elasticity of 0.26 for the 2013 reform.

3moderatelikely · previously known · overclaiming

The paper concludes that when top earners report less income after tax increases, it is always because they shift income between years or into other forms, not because they actually work or produce less. Its own figures show that only about a third of the big 1986-88 jump in top incomes was such shifting. It also leaves out an earlier study finding that wives in top-earning households worked more after the 1986 tax cut. The safer conclusion is that real effects were not proven, not that they were absent.

Technical detail and evidence

Why it matters, in detail. The policy argument (pp. 42-43) that large responses call for base-broadening rather than lower rates requires the response to be avoidance (e2) rather than real (e1). If even part of the persistent post-TRA86 wage response, or of high-income secondary earners' labor-supply response, is real, then base-broadening does not remove the welfare-relevant elasticity for rate changes, and the e1/e2 split is uncertain. On magnitudes: about a third of the TRA86 top-share increase is demonstrably shifting, and about 25-30% (the persistent wage component) is unattributed. Eissa's labor-supply elasticity of about 0.8 for top-1% wives is a direct counterexample that was available but not weighed. None of this is decisive: each counter-study faces the identification problems (diverging trends, mean reversion) that the paper itself emphasises.

Evidence. JEL pp. 35, 37, 38, 42; WP pp. 58-59. Figure 2 decomposition from the Saez (2004) workbook (data-Fig7 and data-Fig8; identical to JEL Figure 2 for 1960-98). Top 0.01%, 1986 to 1988 (+1.00 pp): wages 35.3%, S-corp 22.5%, partnership 8.5%, interest 12.1%, dividends 11.8%, other 7.6%, sole proprietorship 2.2%. Top 1%: wages 40.1%, S-corp 20.0%, partnership 14.7% (scripts/fig2_decomp.py). Eissa (1995): NBER w5023 abstract. Carroll et al. (2001): executive summary and Section 4. Mertens & Montiel Olea: NBER WP 19171 (2017 revision), introduction and Section 5. A text search finds no citation of Eissa (1995) or Carroll et al. (2001) in either version.

Why this severity. The defensible core survives: the clearly identified top-end responses to post-1980 reforms are dominated by timing and reclassification, and real responses are not convincingly established. But the categorical 'in all cases' claim, one of the paper's most-cited conclusions, is not demonstrated by the paper's own decomposition and omits relevant contrary evidence. C5 should therefore be read as 'not established' rather than 'absent'.

The authors' response. No direct response located. Mertens & Montiel Olea themselves note that large short-run top elasticities 'do not contradict more modest long run responses' and that the main conclusions of Saez et al. (2012) on responses outside the top 1% 'remain intact'.

4minorconfirmed · newly identified here · internal inconsistency

The paper says one-third of the 1986-88 surge in top incomes came from business profits moving onto personal tax returns (S-corporations). Its own chart shows about one-fifth; one-third is reached only if another category of business income is added. The broad point, that much of the surge was income being reshuffled on paper, still stands, but less strongly.

Technical detail and evidence

Why it matters, in detail. The share of the TRA86 top-share increase attributable to corporate-to-individual shifting is about 22% (S-corp), not about 33%. Adding partnership loss-closing, which is a shelter change rather than a shift to another tax base, gives about 31%. This also undercuts fn. 15's s=0.5 (F9). Wages (35-40%) remain the largest component.

Evidence. Saez (2004) workbook TPE04.xls, data-Fig8. Total top-0.01% share 0.00991 (1986) to 0.01991 (1988). S-corp component 0.002360 to 0.004609 (+0.002250, i.e. 22.5% of the rise); partnership +0.000852 (8.5%). The JEL Figure 2 vectors give 22.4% (both auditors). The statement is unchanged from the WP. scripts/fig2_decomp.py.

Why this severity. This misstates the paper's own figure, overstating the S-corporation share by about 11 pp (about 50% relative), but the figure is not used in any calculation. The qualitative point is correct: a one-time S-corporation surge was a large component of the TRA86 response. A likely innocent source also exists (S-corp plus partnership, or the 1986-87 change). Its main consequence is to strengthen the overclaim in F3.

5minorconfirmed · newly identified here · replication failure

One column of a results table, for the group just below the top 1%, was evidently computed with the wrong variable. Corrected, the numbers are still close to zero, so the paper's point that this group barely responds to taxes is unaffected.

Technical detail and evidence

Why it matters, in detail. The trend-controlled next-9% coefficients flip from slightly negative to slightly positive (+0.03 to +0.09), and the HAC standard errors roughly double. The 'no correlation' reading for the next 9% (p. 19) survives.

Evidence. Adjudicator reproduction (scripts/table1.py; data/log_table1.txt). Inputs: Saez (2004) Tables B1/B2 for 1960-99, and an independent extraction of the JEL Figure 1 vectors for 2000-06. The extraction matches the workbook to within 0.0001 in shares and 0.0007 in MTRs. Both auditors' independent digitisations give the same numbers.

Why this severity. This is a coding slip in a secondary column. With the correct regressor, the next-9% elasticities are still small and statistically insignificant (the largest is +0.09, t of about 1.9), so the paper's contrast between the top 1% and the next 9% is unchanged. Fn. 43's 'e = 0.01, not significant' also holds, with a standard error of 0.24 rather than 0.13.

6minorconfirmed · newly identified here · methodology

The shortcut formula the paper uses to turn changes in the top 1%'s share of income into a responsiveness number slightly understates responsiveness, by roughly a tenth for the top 1%. This makes no difference to the paper's conclusions.

Technical detail and evidence

Why it matters, in detail. Share-based ETIs are attenuated by the factor (1-p): about 0.85-0.91 for the top 1% over 1981-94, and about 0.75 for the next 9%. Correcting this makes top-1% elasticities larger, which reinforces rather than weakens the sensitivity message. The same attenuation applies to share-based ETIs elsewhere in this literature.

Evidence. Algebra confirmed. On Saez (2004) data, the log-odds version raises the top-1% episode estimates by 9-15% (0.60 to 0.65, 1.36 to 1.53, 0.45 to 0.52, -0.39 to -0.45) and the next-9% estimates by about a third (0.21 to 0.28, -0.20 to -0.26) (scripts/table1.py).

Why this severity. This is a conservative (attenuating) approximation error of 9-15% for the top 1%, in illustrative estimates whose identification problems the paper itself stresses. It changes no conclusion.

7minorconfirmed · newly identified here · arithmetic or transcription

The paper says the comparison group got about half the tax cut that top earners got in 1986-88; its own data show about a third. This weakens one of its explanations for why early studies found huge responses, though the explanation still points the right way.

Technical detail and evidence

Why it matters, in detail. With e_C=0 and a control-group change in log(1-tau) equal to 0.31 of the treatment group's, the DD estimator inflates e_T by 1/(1-0.31) = 1.44 rather than 2. The mechanism therefore explains less of Feldstein's large TRA86 elasticities than implied.

Evidence. Saez (2004) Table B2, with the next 9% computed as the income-weighted average of P90-95 and P95-99. Top 1%: 45.34% to 28.59%; next 9%: 34.83% to 29.27% (scripts/table1.py). The same sentence appears in the WP.

Why this severity. This overstates the magnitude of one illustrative bias mechanism (applied to Feldstein's 1995 TRA86 estimates). The qualitative point, that tax changes in the control group bias DD estimates, stands.

8minorconfirmed · newly identified here · documentation gap

To show how income moving into other taxes softens the cost of a tax increase, the paper uses an example it says it will justify later, but never does. Its own data suggest a smaller amount of such shifting. The general point is right; the example is a bit generous.

Technical detail and evidence

Why it matters, in detail. The illustrative offset from shifting is roughly halved: 23% of revenue lost instead of 18%, an MEB of 0.30 instead of 0.22, and tau*_s of 74.6% instead of 76.8%.

Evidence. Full-text reading of JEL Sections 3-5; Figure 2 decomposition (scripts/fig2_decomp.py). With s=0.225 and t=0.3, the Section 2.2 illustration gives 23.3% of revenue lost (vs 17.9%), an MEB of 0.30 (vs 0.22) and tau*_s=74.6% (vs 76.8%) (scripts/formulas.py). The same footnote appears in the WP.

Why this severity. This affects only the size of an illustrative example of the (correct) fiscal-externality point; the conceptual result is unaffected.

9minorconfirmed · newly identified here · internal inconsistency

After the authors changed how one table was computed between drafts, a few sentences describing it were not updated. The table's message, that results swing a lot with method, is unchanged.

Technical detail and evidence

Why it matters, in detail. The positive Panel A estimates fall by 65-72% between versions. The illustrative point, that estimates vary in sign and size with specification, is if anything strengthened, but readers comparing text and table find mismatched numbers.

Evidence. JEL Table 2 (p. 24), text p. 25, fn. 49; WP Table 2 and text ('e = -0.404(0.089)', '1.329(0.107)'); scripts/version_checks.py.

Why this severity. These are editorial inconsistencies in an illustrative table; the disclosed switch and the sensitivity message are unaffected.

10minorconfirmed · newly identified here · arithmetic or transcription

Two formulas in the theory section have a minus sign in the wrong place. The worked examples next to them are correct, so a careful reader is not misled about the results.

Technical detail and evidence

Why it matters, in detail. Taken literally, (i) would double rather than eliminate the burden in the s=1, value=tau example (0.554 per mechanical dollar instead of 0), and (ii) is impossible with u_c > 0 and u_y < 0.

Evidence. JEL pp. 15, 17; WP equivalents. Re-derived by codex (scripts/audit_calculations.py) and checked by the adjudicator.

Why this severity. These are local typographical sign errors. The surrounding examples and fn. 34 make the intended results clear, and no number depends on them.

11minorconfirmed · newly identified here · arithmetic or transcription

A handful of typos and mislabels: wrong section numbers, a tax increase called a reduction, a wrong technical term. None affects the findings.

Technical detail and evidence

Why it matters, in detail. None on the estimates. The 'qualified' slip could mislead readers about the income concept behind Figure 2's wage component. Because disqualifying dispositions of qualified options can generate wages, it does not prove a data misclassification.

Evidence. JEL pp. 8, 17, 28, 33 and the WP equivalents (scripts/version_checks.py).

Why this severity. Presentation errors resolved by context; no effect on any result.

12minorconfirmed · newly identified here · data error

A few historical top tax rates in the appendix table are slightly off. Nothing in the analysis uses them.

Technical detail and evidence

Why it matters, in detail. None for the analysis: the MTR series in Figures 1-2 come from TAXSIM microsimulation, not from Table A1.

Evidence. IRS SOI Historical Table 23, rows 1952-1981 and footnotes 18, 19 and 29 (histab23.xls). The remaining 55 of 58 years match (Auditor A).

Why this severity. Appendix reference table; not used in any estimate.

13minorconfirmed · newly identified here · overclaiming

The paper's example of how much higher top tax rates could go if loopholes were closed uses two very uncertain estimates, without saying how uncertain they are. The general idea is still supported by other evidence.

Technical detail and evidence

Why it matters, in detail. The illustrated 26-pp gain in tau* from base-broadening is within sampling error for the top group. The full-sample estimates (0.400 (0.144) vs 0.120 (0.106)) support the direction.

Evidence. Gruber & Saez (2002, Journal of Public Economics 84) Table 9: 0.567 (0.298) and 0.171 (0.240); scripts/formulas.py.

Why this severity. Presentational: the conceptual base-broadening point rests on theory and on the more precise full-sample contrast (0.40 vs 0.12), not on these two numbers.

14minorconfirmed · newly identified here · misrepresentation of source

When summarising an earlier study, the paper says its results held up under every variation, but one variation actually found no effect. A small omission that, if anything, supports the paper's own warnings.

Technical detail and evidence

Why it matters, in detail. The IV-with-linear-trend null suggests the top-1% time-series elasticity is less robust to the trend specification than the summary says, which reinforces the paper's own caution about time trends.

Evidence. Saez (2004) workbook TPE04.xls, sheets Table2 and Table3: OLS 1.58/0.85/0.62/0.68; 2SLS 1.70/-0.02/0.59/0.61.

Why this severity. One of eight specifications is omitted from a one-sentence summary; most specifications do give elasticities of 0.5 or more.

15minorlikely · newly identified here · documentation gap

For the most recent years in its main chart, the paper's tax-rate figures look as if they were filled in by assumption rather than calculated. Two different data versions were also joined without saying so. This shifts some of the paper's illustrative estimates a little but does not change its conclusions.

Technical detail and evidence

Why it matters, in detail. If the 2003-06 MTRs were carried forward from 2003, the final observations add identifying variation that is partly imputed. Dropping them moves the top-1% no-trend estimate from 1.71 to 1.57 and the cubic-trend estimate from 0.58 to 0.64. The vintage splice adds a level break of about 0.45 pp in 2000.

Evidence. Adjudicator's own extraction of the Figure 1 vector paths (scripts/digitize_fig1.py): the top-1% MTR polyline has 41 vertices for 44 observations (missing 2001, 2004, 2005), and the next-9% polyline 43 (missing 2005). TabFig2007/2008 vs TPE04 comparison in scripts/tabfig_compare.py. Sensitivity of the Table 1B top-1% coefficients: 1.57/0.84/0.68/0.64 for 1960-2002, and 1.51/0.86/0.60/0.68 using only Saez (2004) data for 1960-2000, vs published 1.71/0.82/0.74/0.58.

Why this severity. This changes the illustrative time-series coefficients by up to 0.2 (no trend) and 0.1 (cubic trend). The 'significant and above 0.5' conclusion survives in every variant.

Claims examined and rejected

Problems one auditor raised that did not survive checking.

Audit notes

What reproduced. Reproduced exactly: - all Section 2 arithmetic in both versions (JEL: 27.7%, $1.38, $0.38, 72.7%, 68.4%, the 42.5% combined rate; 17.9%, 22%, 76.8%; 54%, 80%. WP: 43.1%, $1.76, $0.76, 55.6%; 25%, 32%, 62.2%); - Table 1 Panel A (0.60, 1.36, 0.45, -0.39; 0.21, -0.20) from Saez (2004) Tables B1/B2; - the Table 1 Panel B top-1% coefficients and Newey-West SEs (1.71 (0.31), 0.82 (0.20), 0.74 (0.06), 0.58 (0.11)), from the workbook plus the adjudicator's own extraction of the JEL Figure 1 vectors (validated against 1960-99 published data); - the Figure 2 composition statements (dividends over 60% in the early 1960s, Kennedy MTRs, 61% S-corporation share of the 1981-84 rise). Approximately reproduced from public inputs: the conclusion's MEBs, $0.339 (adjudicator: $0.341) and $0.195 (adjudicator: $0.15-0.185). Not reproduced: - the Table 1 Panel B next-9% column under the stated specification (it reproduces only with a log(MTR) regressor); - 'about one-third' S-corporation share (data: 22.5%); - fn. 72's $0.678 (consistent value about $1.03); - 'about half' next-9% rate cut (data: one-third). Not reproducible: Table 2, which uses a confidential 1991-97 SOI panel and CBO's internal tax calculator, and the authors' exact 2005 microdata calculations behind $0.195 and $0.339.

Limits of this audit. - Table 2 rests on a confidential 1991-97 SOI panel and CBO's internal tax calculator and could not be reproduced; only its internal and cross-version consistency was checked. - The conclusion's MEBs were reproduced only approximately, from public aggregates (SOI 2005 Table 3.4 and Piketty-Saez Pareto coefficients, with an assumed 5-5.9% state rate), not from the authors' 2005 microdata. - The post-2000 values in Figure 1 were recovered from the PDF's vector paths, validated against the 1960-99 published data. Figure 2's post-2000 MTRs rely on Auditor A's extraction. - Not obtained or unreadable: Slemrod (1992, 1996) (scanned or garbled text); the Moffitt-Wilhelm (2000) figures; Giertz (2009); Weber's 2010 draft; the Kleven-Schultz (2011) draft; the Fack-Landais (2010) full text; and the published REStat version of Auten-Carroll (1999). - Not verified: the itemization statistic (25% fewer itemizers in 1993 than in 1986) and Kopczuk's 0.42/0.36 itemizer ETIs. - The non-U.S. literature reviewed in the WP was not audited, and no new ETI estimation was attempted. - Judgments about which elasticity is 'appropriate' for policy calibration are inherently contestable, and later evidence points in different directions: Saez (2017), Mertens & Montiel Olea (2018), Weber (2014), Neisser (2021).