Generalized Social Marginal Welfare Weights for Optimal Tax Theory

Why it matters, in detail. With quasi-linear utility, a weight that varies along an individual's (c, z) indifference curve (with t = z - c, or with c alone) allows utility-preserving, revenue-neutral policy paths along which the sign of the same small transfer's evaluation flips. That yields T0 < T1 ~ T2 < T3 ~ T0. Local optima defined by Proposition 1 with libertarian (elastic-earnings) or poverty-line weights are therefore not maxima of any transitive ordering. A consistent rendering of the same values requires a behavioral-response term in the weights (Bergstrom-Dodds eq. 24), which changes Gbar(z) and hence T'(z) by amounts not yet quantified. 'Unification' thus shrinks to generalized utilitarianism, integral of F(U_i, x_i) (Sher Theorem 1), unless weights are allowed to depend on T'' or on population statistics. The effect is qualitative (an impossibility result); no US number in the paper depends on it, but the shapes in Proposition 7 and Proposition B2 do.

Evidence. Paper: abstract p. 24; Introduction pp. 24-25; Definition 5 p. 33; Proposition 7 pp. 41-43; Online App. B3 p. 10 and B4 p. 11. Sher (2024), 'Generalized Social Marginal Welfare Weights Imply Inconsistent Comparisons of Tax Policies', AER 114(11): 3551-3577, https://doi.org/10.1257/aer.20211025: abstract and pp. 3552-53; Proposition 4 pp. 3564-65; Theorem 3 p. 3566; Section V.D pp. 3570-71; Section VII.A pp. 3573-74, which links the result to Kaplow-Shavell and quotes Fleurbaey-Maniquet 2018, p. 1059. Bergstrom and Dodds (2026), 'Consistent Non-Utilitarian Welfare Weights', Tulane Economics WP 2609 (September 2026; also CESifo WP 12978), http://repec.tulane.edu/RePEc/pdf/tul2609.pdf: Introduction pp. 1-4, Section 5.2 eq. (24), conclusion p. 31. Adjudicator reproduction of Sher's example (scripts/06_theory_checks.py, block A): along the path, revenue = 0.25 and both groups' money-metric utility = 0 (|du| < 1e-16), while taxes paid swap from 1/3 vs 1/6 (theta0 = sqrt(1/3)) to 1/6 vs 1/3 (theta1 = sqrt(2/3)). The same transfer eps = 0.001 from group A to group B has dW = -1.07e-4 at theta0 and +1.07e-4 at theta1 (with g = exp(t)), which closes the cycle. Both auditors reproduced the same construction.

Why this severity. It reaches the paper's signature claim. The extra flexibility beyond generalized utilitarianism (weights depending on taxes paid or on consumption relative to a poverty line) is exactly what produces inconsistent rankings, so two of the three alternatives named in the abstract are not coherently represented as specified. It is not critical: formulas (2)-(3) remain correct conditions under the paper's own local definitions; every structurally utilitarian application survives (including two-type EOp and fixed incomes); the authors disclaim global rankings; and a repaired weight concept exists (Bergstrom-Dodds 2026, unrefereed).

The authors' response. No published reply by Saez or Stantcheva located (OpenAlex lists 14 works citing Sher 2024 as of 2026-10-25, none by the authors). The paper itself states that the theory 'is fundamentally local and hence cannot rank optima when they are multiple' (p. 31) and concedes (fn. 3) that the standard approach can rank multiple optima.

Why it matters, in detail. In comparisons 3 and 5, S = 0 for both alternatives whatever (dT, dc), which attenuates both OLS slopes roughly proportionally (factor about 0.6) and pulls the intercept to 0.31. Corrected coefficients are 1.5-1.8 times the published ones (column 1: b_T 0.00294, b_c -0.00736). The ratio is nearly invariant: corrected alpha is 0.400/0.538/0.658/0.679/0.188/0.806 vs published 0.37/0.58/0.65/0.64/0.17/0.77, and MTRs are 71.4/65.0/60.3/59.6/84.2/55.4% vs 73/63/61/61/85/57%. The liberal tax coefficient becomes significant (naive t = 3.3, respondent-clustered t = 2.6, vs 1.79 published).

Evidence. Adjudicator re-analysis (scripts/01_survey_audit.py, output/01_survey_audit.log). test3's answer is in the display5 pair 100.0% of the time but in its own display3 pair 5.86%; test5's is in display3 100.0% and display5 5.94%; the chance rate is 2/35 = 5.71%. Pairs 1, 2 and 4 match their own display 100%. Swapping test3 and test5 leaves zero invalid answers. Dominance check (one family has weakly lower net income and weakly higher tax): respondents choose the dominant family 94.3/95.5/94.2% of the time in pairs 1/2/4, against 5.7/5.6% in pairs 3/5 under the do-file mapping and 95.1/98.0% after the swap. The do-file mapping reproduces the published column 1 (b_T 0.00171, b_c -0.00456, alpha 0.374, MTR 72.8%, N 11,450) with intercept 0.307; the corrected answered-only intercept is 0.500, as a complete binary-choice design implies. Claude and Codex found the same swap independently, with identical counts.

Why this severity. A confirmed error affecting 40% of the observations behind a published table and its coefficients by 50-80%. However, alpha moves by at most 0.05 and the implied MTRs by at most 2 pp, and every stated survey conclusion survives; the survey is a secondary, appendix result.

Why it matters, in detail. The OLS projection of F(X beta) onto X preserves coefficient ratios only under special designs (e.g. elliptical X, the Stoker/Brillinger result). Here identification is dominated by $500K-$1M pairs in the saturated region of F. Applying a fitted logit to the actual design and projecting by OLS returns 0.409 from a true index ratio of 0.619 (Codex check), so the bias is mechanical. Implied fixed-income MTRs are 71-73% under the LPM vs 58-63% under logit/probit/maximum score. The liberal-conservative gap is 29 pp under the LPM vs 14 pp under the logit. The much better fit of a log index (log-likelihood gain of about 576) implies social indifference curves that are not linear in dollars, so no single constant MTR is identified.

Evidence. Adjudicator, corrected mapping, one observation per answered pair, respondent-clustered (scripts/01_survey_audit.py, scripts/02_survey_functional_form.py). Logit alpha = 0.619 (SE 0.044), MTR 61.8%; probit 0.579, MTR 63.3%. Logit by subsample: 0.707/0.740/0.717 (MTR 58.6/57.5/58.2%). Liberals: 0.492 (MTR 67.0%) vs LPM 0.193 (83.8%); conservatives: 0.895 (52.8%) vs LPM 0.819 (55.0%). Log-likelihood: dollar-linear logit -2,859.5 vs log-index logit -2,283.6, with the same number of parameters. Codex obtains identical logit values. Claude's distribution-free maximum-score estimate is alpha = 0.70 (90% CI 0.34-0.86); this was not re-run by the adjudicator.

Why this severity. It materially changes the survey's quantitative results: the full-sample MTR falls from 73% to about 62%; the subsample decline shrinks from 12 to about 4 pp; and liberals are no longer 'relatively close to the utilitarian polar case'. The qualitative conclusions (both c and T matter; conservatives weigh T more) are intact, and the exercise is an appendix 'first attempt'.

The authors' response. The paper calls Table A3 'a first attempt' and states that 'more complex and detailed survey work in this spirit could help calibrate the weights more precisely' (App. pp. 19-20).

Why it matters, in detail. The top asymptotic rate (1 - Gbar)/(1 - Gbar + a e) depends on the representation index of the weighted set. Under Roemer's worst-off-type rule, finer types lower Gbar_top from 0.334 to 0.20-0.26. That raises the EOp top rate from 47% to 49-51% (the gap to the utilitarian 56% shrinks from about 10 to 5-7 pp), and raises P25-P50 rates to 54-73%; with quintile or decile types these are at or above the utilitarian 58-67%. Under the paper's own below-rank-mean indicator with fine types, Gbar rises in the upper-middle (0.66 vs 0.48 at P90), cutting the P75/P90 EOp rates to 36%/25% (a gap of 15-17 pp). Combined with the undocumented utilitarian column (F6) and the elasticity (F9), the 'ten points' figure is uncertain by at least +/-5 pp at most percentiles.

Evidence. Chetty et al. (2014) Online Data Table 1, national 100x100 matrix (scripts/04_table2_eop.py, 05, 08). Column 1 reproduces within 0.004 (0.442/0.371/0.299/0.240/0.167 vs 0.443/0.373/0.303/0.236/0.170). Roemer worst-off-type rule, P25-P99.9 EOp MTRs (P99.9 held at the P99 value): types = parents <= P33 give 62.8/53.5/46.3/37.8/48.4/49.1%; <= P20 give 68.7/58.3/49.8/40.1/49.8/50.6%; <= P10 give 73.1/61.5/52.0/41.4/50.4/51.2%. Published: EOp 53/45/40/34/46/47% and utilitarian 67/58/51/42/54/56%. The paper's own indicator rule with 100 parental-centile types gives a weight-1 share of 55% and Gbar = 0.868/0.754/0.671/0.662 at P25-P90, i.e. MTRs of 56.5/44.0/35.5/24.8%; the P99 value is not reliable because the public matrix has a single bin above P99. Claude's Roemer-partition numbers reproduce exactly.

Why this severity. A secondary, explicitly illustrative calibration. Its headline comparison (a gap of about 10 pp, EOp less progressive) changes by a quarter or more, in either direction, under defensible alternative implementations of the same principle, and this is not disclosed. The qualitative point survives: EOp weights fall with income for reasons unrelated to marginal utility, and Gbar stays positive at the top.

The authors' response. The paper calls the simulations 'illustrative at best' (p. 40) and notes (fn. 31) that in a less meritocratic society Gbar at the top would be smaller and the top rate closer to the utilitarian one; the type partition is not discussed.

Why it matters, in detail. The discontinuous, allocation-dependent binary weights become uniform exactly at tau = 1. The self-fulfilling 'high-tax optimum' therefore exists only as the limit of the re-evaluated-weight fixed-point map (RHS(tau) -> 1), not as a local optimum under frozen weights, where cutting tau yields a surplus gain of d(q) - v(d(q)) > 0. The Alesina-Angeletos analogy survives only as dynamics under repeated re-evaluation of weights. Genuine multiplicity needs an elasticity profile that yields a stable interior root above 0.5.

Evidence. Paper: Online App. A1, Definition A1, pp. 1-2; App. B2, eq. (A5), pp. 6-10. Fleurbaey and Maniquet (2018), 'Optimal Income Taxation Theory and Principles of Fairness', JEL 56(3): 1029-1079, pp. 1050-51. Adjudicator (scripts/06_theory_checks.py, block B): the share with weight 1 at tau = 1 is 1.000; cutting tau to 0.99/0.95/0.90 raises surplus by +1.1e-8/+6.8e-6/+1.0e-4. The formula-(3) map changes sign at tau of about 0.055 and 0.53, with RHS > tau on (0.53, 1). Codex gives an analytical counterexample (luck U[0.5,1.5], d = 48 q^4) in which the frozen-weight gain is 48 q^4 (1 - 4q/5) > 0.

Why this severity. A heuristic online-appendix illustration whose multiplicity the authors disclose and present as a feature. The error removes the claimed 100% optimum, and with it the multiplicity in standard parameterizations, but it does not touch any main-text result.

The authors' response. The authors call the discussion 'heuristic' and note that 'the low tax can well Pareto dominate the high tax optima. The tax reform approach is inherently local' (App. pp. 8-10). No reply to Fleurbaey-Maniquet located.

Why it matters, in detail. Gbar_util at the bottom depends heavily on consumption near zero and on the schedule T(z) used. Plausible constructions put the P25 utilitarian MTR anywhere from 59% to 70% (P50: 50-63%), so the low- and middle-income half of the 'about ten points higher' claim is not pinned down. At P99-P99.9, Gbar_util is about 0.02-0.08 and rates are 53-56% in every construction, so the top comparison is unaffected.

Evidence. Spreadsheet formulas and constants (adjudicator inspection); ReadMe text. The public column 1 matches within 0.004 (F4). Auditor A built plausible reconstructions of the utilitarian column from IRS SOI bracket data; the adjudicator verified only the formula-(2) arithmetic. A self-consistent fixed point gives 58.8/49.5/45.3/38.7/52.7/55.9%; a revenue requirement of 20% of AGI gives 63.6/53.8/48.4/40.9/53.5/56.2%; actual 2008 taxes with a $15K consumption floor give 69.7/63.0/54.6/44.4/54.6/56.4%. Published: 67/58/51/42/54/56%.

Why this severity. All twelve rates reproduce given the inputs, and the EOp side can be checked from public data. The gap concerns the benchmark column, which is uncertain by roughly 5-10 pp at P25-P50 but robust (53-56%) at the top.

Why it matters, in detail. Respondent-level dependence (up to five choices, ten mirrored rows) overstates precision; cluster-robust SEs for alpha are 1.3-1.7 times the naive ones. The *** on the liberal tax coefficient contradicts the do-file's own star convention (p < .01) under any SE choice on the posted data. After correcting the pairing, the liberal coefficient is significant at about the 1% level with clustering, so the qualitative statement survives.

Evidence. Adjudicator (scripts/01_survey_audit.py): on corrected data, clustered-to-naive SE ratios for alpha are 1.31-1.71. On the posted (do-file) data, column 5's clustered t is 1.60 (p about 0.11). On corrected data, every tax coefficient is significant with clustering (column 5: t = 2.58-2.65). Published alpha SEs (0.06/0.06/0.07/0.09/0.12/0.16) compare with clustered delta-method SEs of 0.079/0.090/0.094/0.108/0.103/0.186 on the do-file data.

Why this severity. Standard errors ignore within-respondent dependence (cluster-robust SEs are 1.3-1.7 times the naive OLS ones), one published significance claim is wrong, and two columns are not reproducible from the posted code. But after the data fix (F2) every survey conclusion is significant under clustered inference.

Why it matters, in detail. Given Gbar, T' = (1 - Gbar)/(1 - Gbar + alpha e) falls with e. Moving e from 0.5 to SSG's midpoint of 0.25 raises the top EOp/utilitarian rates from 47/56% to 64/72%. The choice biases levels down, not toward high taxes; the EOp-utilitarian gap stays at 7-11 pp (8-14 pp at e = 0.5).

Evidence. SSG (2012) p. 42 and fn. 72 (shared-cache text). Adjudicator sensitivity (scripts/04_table2_eop.py): at e = 0.25, EOp rates are 69.1/61.9/56.9/50.7/62.7/63.7% and utilitarian 80.3/73.1/67.3/59.2/70.1/72.1%; the asymptotic rates are 64.0% vs 72.7% instead of 47.1% vs 57.1%. The EOp < utilitarian ordering holds at every percentile for e = 0.12-1.0.

Why this severity. Every Table 2 level shifts by 13-17 pp within SSG's range, but the ordering is unchanged, the paper flags the uncertainty ('illustrative at best'), and the high e lowers rather than raises the computed rates.

The authors' response. The paper notes that 'because of uncertainty in the level of e, the simulations should be considered as illustrative at best' (p. 40).

Why it matters, in detail. Using 1/c instead of 1/(c - v(z)) understates the weight of high earners, whose v(z) is large in absolute terms, so the benchmark is not the utilitarian Gbar of the model being optimized. The direction of the effect on bottom rates depends on v(z) at the bottom and is not identified from the published inputs.

Evidence. Paper pp. 26, 40 and Table 2 notes. Codex's constructed example (e = 0.5, tau = 0.4, same allocation): income-weighted normalized weight 0.742 with 1/c vs 0.782 with 1/(c - v). The adjudicator checked the logic, not the example's numbers.

Why this severity. A labelling and consistency problem with the comparison benchmark; its effect on Table 2 cannot be quantified without v(z) and is likely smaller than the documentation uncertainty in F6.

Why it matters, in detail. T'/(1 - T') = (1 - Gbar)/(alpha e) has no solution with T' < 1 when (1 - Gbar)/(alpha e) < -1. The welfare derivative then has a constant sign, pushing T' to a corner where the local formulas no longer apply. The sign claims hold, but the stated closed forms are invalid in part of the parameter space.

Evidence. Adjudicator (scripts/06_theory_checks.py, block C): at the paper's own P25 alpha = 0.2035 and e = 0.5, a negative rate requires H < 0.092; H = 0.10/0.20/0.25 gives T' = 11.9/1.69/1.44. Application 1 with gbar0 = 0.5: e0 = 0.1/0.3 gives T'(0) = 1.25/2.5. These match Claude's values. Codex observed above-100% rates in a B4 test but did not count them, since a proposed density need not be an equilibrium density.

Why this severity. The qualitative conclusion (subsidize work at the bottom) survives in a stronger, corner form; only the closed forms are incomplete.

Why it matters, in detail. Each share moves by about 0.2 pp; average ranks (1.375/1.593/3.021/3.506) and the freeloader ordering are unaffected.

Evidence. Adjudicator (scripts/07_tableA1_panelC.py). Literal do-file: first rank 57.57/37.35/2.67/2.50, last rank 2.37/2.92/25.05/70.84 (sum 101.2). Intended logic: last rank 2.16/2.71/24.84/70.64. Codex and Claude report the same.

Why this severity. Changes the shares by about 0.2 pp; the ranking of the transfer recipients is unaffected.

Why it matters, in detail. None of these slips changes any computed quantity: Table 2 uses 2 x column 1, and Proposition 5's formula and the zero-tax libertarian optimum are unaffected by the B3 sign typo and the A4 regularity statement.

Evidence. Published text, pp. 40, 43-44 and Online App. pp. 4, 10, 17, 19-20, 23-25; item counts in output/01_survey_audit.log. Items (i) and (vi) also appear in the NBER revisions of Oct 2014, Mar 2015 and Jul 2015.

Why this severity. Typographical or presentational; none changes a computed result or conclusion.

Why it matters, in detail. Ideology strongly moderates alpha (liberals 0.19-0.49 vs conservatives 0.81-0.90 across estimators). Liberals are twice their population share in the sample, which biases pooled alpha down and the implied MTR up by about 3-6 pp. Framing effects of the 'tax break' wording cannot be identified with these data; excluding the $10K families (where the break can exceed taxes paid) leaves alpha essentially unchanged (columns 3 vs 4).

Evidence. Adjudicator (scripts/01, 03). Self-reported ideology: 22.0% conservative, 32.4% moderate, 45.6% liberal (N = 1,225), against Gallup's 2012 national 38/36/23 (F. Newport, 'Alabama, North Dakota, Wyoming Most Conservative States', 1 February 2013, https://news.gallup.com/poll/160196/). Obama supporters are 77.9% of Obama+Romney supporters, against an actual 2012 two-party share of 51.96% (FEC, 'Federal Elections 2012': Obama 65,915,795, Romney 60,933,504). Median age is 28. Reweighting to Gallup ideology shares raises LPM alpha from 0.410 to 0.537 (MTR 70.9% to 65.1%) and logit alpha from 0.619 to 0.701 (MTR 61.8% to 58.8%). Capozza & Srinivasan ('Who Should Get Money? Estimating Welfare Weights in the U.S.', 18 Sept. 2026 version, fn. 3) note the non-representative MTurk sample and hypothetical choices, and show that incentives matter.

Why this severity. Shifts the survey-implied MTR by a further 3-6 pp and limits external validity, but the qualitative patterns hold within both liberal and conservative subsamples, and the survey is presented as exploratory.

The authors' response. The appendix presents the survey as a first step and reports the liberal/conservative splits; it acknowledges the moral framing of frugality (App. p. 17).