Wealth Taxation and Wealth Accumulation: Theory and Evidence from Denmark

Why it matters, in detail. The 30-year effect is governed by alpha, which the near-linear 8-year TOT path does not identify. alpha is fixed by imposing a zero change in the 80-90 log-wealth slope for a cohort treated at 61, and the before/after cross-sections offered as support have no power against alternatives: given actual exposures, alpha=0 to 10 predicts the same slope change. Plausible ranges for the long-run elasticity wrt the net-of-tax return are therefore about 0.2-0.9 (couples) and 0.35-1.45 (ceiling), against the reported 0.77 and 1.15; under alpha=1 they are 0.41 and 0.35.

Evidence. Independent re-implementation of the Section V model (scripts/jjkz_model.py) reproduces Table III exactly. scripts/identification_sensitivity.py: 8-moment RMSE 0.0143-0.0170 (couples) and 0.0299-0.0307 (ceiling) across alpha in [0,10]; 30-year effect 0.000 (alpha=0), 0.161/0.200 (alpha=1; Auditor B's numbers reproduced), 0.537/0.950 (alpha=10); alpha=sigma benchmark (p. 369) 0.336/0.819, elasticities 0.86/1.45; shifting the restricted 10-year slope by -0.05/+0.05 gives couples elasticities 0.60/0.96 and ceiling 1.05/1.26 (+/-0.10: 0.43/1.14 and 0.95/1.36; Auditor A's numbers reproduced). scripts/latelife_profile_test.py: predicted after-minus-before change in the 80-90 log slope is -0.037 to -0.045 (couples) and -0.073 to -0.114 (ceiling) for every alpha tested (published calibration -0.040/-0.080); the plotted data change is roughly -0.06 to -0.10 (OA Fig A.XVIII). Ceiling SSE: 0.00737 at sigma=6.65 vs 0.00654 at sigma about 34 (output/identification_sensitivity.txt).

Why this severity. The abstract's long-run claim is determined by an assumption the data cannot test; defensible alternatives (log bequest utility alpha=1, or +/-0.5% a year of extra late-life growth) move the 30-year effect by -47% to -69% or by +25-45%, although it stays positive for any alpha>0.

The authors' response. fn 36 and OA Fig A.XXI concede that the TOT moments alone do not robustly identify sigma and alpha; the paper argues that the end-of-life flatness requirement 'nails both parameters' and cites unchanged 80-90 profiles before and after the reform (pp. 376, 378-379).

Why it matters, in detail. For the same comparisons, the published magnitudes are not robust across the authors' versions. Income-decile-by-year controls removed 77% of the couples-vs-singles effect in the WP design, and equity-share controls removed the ceiling effect in the 2017 design. Combined with F3 and F6, this indicates that the counterfactual trends of couples vs singles (income, age) and of bound vs unbound households (portfolio, W/Y) are not well pinned down, so the 8-year magnitudes in C1-C2, and the C4 targets built on them, are materially uncertain.

Evidence. NBER WP 24371 Table 1 (p. 54): couples vs singles ITT 0.109 / TOT 0.198 (plain) vs ITT 0.025 / TOT 0.045 (controls); couples vs couples TOT 0.100/0.104; ceiling TOT 0.189 (plain) / 0.322 (controls). NTA 2017 draft, https://ntanet.org/wp-content/uploads/proceedings/2017/NTA2017-315.pdf (dated 23 April 2017; md5 06dbafc3d74a7e6ac6090a7890e50250): abstract; pp. 12-14; Fig 7 (no post-reform divergence, raw or with equity-share controls). Text search of the QJE article and Online Appendix finds no 'portfolio', 'decile' or 'reweight' in this context. An illustrative recalibration that applies the WP's controlled/plain ratio to the published couples path gives a 30-year effect of 0.077 and an elasticity of 0.20 (sigma = 0.60, below the authors' grid floor of 1.01).

Why this severity. The authors' own earlier preferred specifications imply core effects 77-100% smaller. This does not prove the published estimates wrong, but it shows they sit at the top of the range of specifications the authors themselves used, and the reader is not told.

The authors' response. None located; the published paper does not discuss the earlier specifications or results.

Why it matters, in detail. About 30-37 of the 41 pp of control-group switching by 1996 is reform-induced, judged against natural churn of 4-5% observed outside the window. Those households received a marginal-rate increase, not a cut, so substitution effects, and the income effects of a lower average rate documented by Ring (2024), imply their accumulation should fall or stay flat. With no response, TOT_1996 = ITT/P_T = 0.172; recalibrating the model (paper's restriction) gives a 30-year effect of 0.384 instead of 0.651, an elasticity of 0.68 instead of 1.15 and a long-run behavioural effect of +0.02 instead of +0.38. A negative response would mean a larger bias. The upward bias plausibly lies between zero and 45% or more of the headline, most likely well above zero.

Evidence. Exact Fig VII-A series recovered from the PDF vector graphics (scripts/extract_vector_series.py): the control unbound share is 0.066, 0.245, 0.302, 0.330, 0.311, 0.359, 0.344, 0.411 in 1989-96, i.e. jumps of +0.066/+0.178/+0.057/+0.067 in the rate-cut years 1989/90/91/96 against +0.043 net over 1992-95 (Auditor A's digitisation confirmed). Only 4.0-4.7% of the same households were unbound in 1980-81, one or two years outside the assignment window at a constant 2.2% rate, against 24-41% two to eight years after the reform; the treated group's churn is symmetric (1.5% bound one year either side). Rate schedule: scalars.do. Denominators (scripts/ceiling_denominator_check.py): printed elasticities imply D = 0.0090; coding switchers as receiving the cut gives 0.0076-0.0098; coding them as unaffected 0.0112-0.0142 (elasticity 7.2-9.1); counting their rate increase 0.0141-0.0179 (5.7-7.2). Counter-evidence: in the authors' own variations the ITT rises as controls are restricted to more strongly bound households (2018 WP p. 17 and Fig A.V: ITT 0.191 vs 0.159, TOT 0.330 vs 0.322; the 2017 draft, apparently without the 20% exclusion, showed no post-reform divergence). That pattern fits the authors' coding, but it also fits more extreme controls having weaker counterfactual growth.

Why this severity. If reform-unbound controls do not respond, the very-wealthy TOT falls by about 45% (0.312 to 0.172) and the recalibrated long-run elasticity from 1.15 to 0.68, eliminating the long-run behavioural effect and the 'larger at the top' ranking; the qualitative positive ITT is unaffected.

The authors' response. None. The paper states that bound households' marginal rate was 0 'before and after the reform' (p. 345), attributes switching to 'wealth and income shocks' (p. 360), and says results are 'robust to alternative cuts' of the bound group (p. 358) without showing them; the 2018 WP (p. 17, Fig A.V) reported that strongly bound controls give the same TOT.

Why it matters, in detail. The trend term contributes 0.085 of the 0.159 ITT (54%) in 1996. If the pre-period gap reflects selection within the 1982-88 assignment window rather than a persistent growth differential, as the placebo samples and the window dependence suggest, the adjusted estimator overstates the effect. The defensible range for the 8-year ceiling TOT before the F2 correction is about 0.14-0.31, and the paper reports and calibrates only its top.

Evidence. Exact raw and adjusted ITT paths recovered from Fig VI-B vectors: raw 1989-96 = -0.009, 0.021, 0.043, 0.080, 0.062, 0.069, 0.120, 0.074 (mean 0.057); adjusted = raw - 0.007 + 0.0115*(t-1988) (mean 0.102). Pre-trend slopes of -0.0115 (1985-88), -0.0115 (1982-88), -0.0101 (1982-87) and -0.0099 (1980-88) log points a year. Recalibration (scripts/alternative_targets.py): raw path gives a 30-year effect of 0.343 and elasticity 0.61 (Auditor A: 0.345/0.61), with a long-run behavioural effect of -0.04; raw path with non-responsive switchers (F2) gives 0.219 and 0.39. Tazhitdinova & Vazquez-Bare (2023), NBER WP 31063, pp. 20-21 and 24-28, discuss this design explicitly.

Why this severity. The preferred very-wealthy estimate more than halves without the extrapolation, and the long-run elasticity falls from 1.15 to 0.61 (0.39 with F2); the direction of the bias is suggestive rather than proven, because a raw DD with a visible pre-trend is not clean either.

The authors' response. The paper shows raw and adjusted estimates side by side (Table II; Fig VI), flags the extrapolation assumption (fn 3, fn 20), and prefers the long assignment window because it 'ensures better pretrends in the raw data and is relatively conservative' (p. 362).

Why it matters, in detail. No specification-invariant ordering of responsiveness between the two populations is supported. Downstream readings that the very rich are more elastic, consistent with avoidance (p. 384; Kopczuk 2019, corpus optimal-tax-behavior-16), rest on the trend-adjusted, switcher-inflated ceiling estimate.

Evidence. Table II elasticities wrt (1-tau)R-1: raw ceiling 0.219 vs raw couples 0.345; adjusted 0.393 vs 0.232 (couples within vs below 0.217); preferred 0.393 (0.038) vs 0.345 (0.048), z about 0.8 with the printed SEs and about 0.35 with SEs consistent with the ITT (F11). Per percentage point of tax cut, the 8-year TOTs are 0.215 (ceiling, 1.45 pp) vs 0.186 (couples, 1 pp). Under F2 the ceiling elasticity wrt 1-tau is 5.7-9.1 instead of 11.3. Recalibrated long-run elasticities under the F2/F3 targets (0.39-0.68) are below the couples' 0.78; under alpha=1 both are about 0.35-0.50.

Why this severity. A qualitative claim in the abstract is not robust to the paper's own alternative specifications; the existence of responses is unaffected.

The authors' response. The paper notes the elasticity is 'about 0.4' for the ceiling vs 'just above 0.2' for the pretrend-adjusted couples designs (p. 364) and that larger elasticities at the top 'are consistent with larger avoidance at the top' (p. 384).

Why it matters, in detail. For the Saez-Stantcheva sufficient statistic invoked on pp. 333-334, the relevant object is the elasticity of the aggregate tax base. Within the model that is about 0.60/0.90 for ages 61-90, before mortality weighting and without the excluded younger cohorts.

Evidence. output/steady_state_check.txt: log of the aggregate wealth ratio over ages 61-90 is 0.232 (couples) and 0.510 (ceiling), elasticities 0.60 and 0.90 vs 0.77 and 1.15; the simple age-average is 0.228/0.482 (Auditor A's numbers reproduced); peak effects are 0.322 and 0.708 at year 25-26.

Why this severity. A more policy-relevant aggregation changes the long-run elasticity by about 22%; the issue is definitional and the paper describes its experiment accurately.

The authors' response. The paper defines the 30-year effect as end-of-life wealth (p. 379) and notes that aspiration effects on younger cohorts are not captured (fn 6; conclusion).

Why it matters, in detail. Whether eq. (3) counted reform-unbound controls' rate increases (F2), and how the elasticity SEs were computed (F11), cannot be determined from the package. The posted table code would not reproduce the published Table II.

Evidence. DD_tables_estimation.do: denominators deltatau_* and deltatauR_* never defined; SE lines use an undefined deltautau; ceiling 'ITT' rows filled with TOT locals (2e_X80tot_*), unlike the published table. DD_mechanical_effects_wealthcuts.do: undefined allowance_ajd_*/allowance_mar_*, 'foreach yr in 1988/1996', keeps mec_effect_form, which is never generated. DD_reg_pretrend.do: year_treat never constructed. DD_reg_placebo_reforms.do: gen group_stub = group_2e_X80 (missing for non-main households, which evaluates as true). DD_balance_and_treatment_status.do: treatmt_c1f built on c1d, skipping the 1982-83 checks. wealthshares_figs.do imports a non-existent 'Top Wealth Shares_DK.xlsx' and US columns that are absent from the posted workbook. Hard-coded I:\Workdata paths. Calibration: delta hard-coded although p. 377 says it is calibrated; the sigma grid cannot go below 1.01. All checked by the adjudicator; the calibration code itself is complete and reproduces exactly.

Why this severity. Prevents verification of the elasticities, their SEs and the mechanical split, which are exactly the quantities at issue in F2 and F11; there is no evidence that the published numbers are wrong because of it.

The authors' response. The replication statement (p. 386) points to the Dataverse package (doi:10.7910/DVN/PFQU4R); microdata are confidential.

Why it matters, in detail. The Wald estimator with current status identifies the effect of continuous exposure only if effects depend on current status alone, which is inconsistent with the cumulative-stock model used for calibration. For the ceiling this is an alternative to the F2 correction, not additive to it (under F2 the denominator is P_T, which is nearly constant). Together they span a 23-45% overstatement of the 1996 TOT.

Evidence. Exact first-stage paths (Figs V-A, VII-A vectors). A linear cumulative-exposure correction, ITT_t / mean over s<=t of (P_T - P_C), gives 1996 continuous-exposure effects of 0.136 for couples (vs 0.186, -27%) and 0.239 for the ceiling (vs 0.312, -23%). Recalibration (scripts/alternative_targets.py) gives a couples 30-year effect of 0.243 and elasticity 0.62 (vs 0.304/0.78), and a ceiling 30-year effect of 0.526 and elasticity 0.93 (vs 0.651/1.15). Auditor B's stylised illustration gave 41-42%; the exact paths give 31-36% (TOT relative to the continuous-exposure effect).

Why this severity. About a quarter off the 8-year effects and a fifth off the long-run elasticities under the paper's own saving model; this is borderline major, rated moderate because the correct dynamic scaling depends on how reversible the responses are.

The authors' response. The paper notes that the Wald scaling enhances 'the dynamically growing effect', as 'an implication of the gradual reduction in persistence' (p. 353), but does not address carry-over of past exposure.

Why it matters, in detail. Conditioning on age, which is a weaker and more credible parallel-trends assumption given age gradients in wealth growth, implies an 8-year couples TOT of at most about 0.14 and more plausibly about 0.13, against 0.186. Combined with F5 the figure is about 0.10, and the recalibrated long-run elasticity is 0.49 (vs 0.78).

Evidence. OA Tables A.I-A.II, couples vs singles: ITT 1996 0.113 (young) vs 0.021 (old); TOT 0.193 vs 0.049 (pretrend-adjusted old: -0.038). With household and year FE in a balanced panel, pooled DD = w_T*D_young + (1-w_T)*D_old + (w_T-w_C)(dC_young - dC_old). The young half holds 49% of households (51,425/104,839 obs) and couples are 75%, so w_T <= 0.654 and the maximum feasible weighted ITT is 0.081 < 0.096. Composition is therefore at least 0.015 (15% of the ITT; about 23% of the TOT at the bound and about 30% with realistic sorting). The top-1% age profile (Calibration_Data.xlsx) predicts +0.008 differential growth 1982-88 but +0.031 over 1988-96 for these mean ages. Recalibration with the TOT scaled by 0.77 / 0.69 gives elasticities of 0.61 / 0.55. Ceiling: pooled raw ITT 0.074 is below both age-specific estimates (0.102, 0.126).

Why this severity. Lowers the moderately-wealthy headline by at least 15-23% (probably about 30%), with no effect among older households; it does not bias the ceiling estimate upward.

The authors' response. Age splits are reported in OA Tables A.I-A.II (fn 26) without discussion of composition.

Why it matters, in detail. The calibrated sigma and alpha absorb reporting and avoidance margins, so the 30-year numbers extrapolate reported-wealth responses with a saving model's transition dynamics. They are not estimates of the steady-state capital-supply elasticity that enters optimal capital-tax formulas, and policy uses that treat them as such overstate what was identified.

Evidence. Reproduced consumption change at t=0 (scripts/reproduce_table3.py): -33.2% (couples vs singles), -49.7% (within vs below), -81.8% (ceiling); -24.9/-37.5/-58.4% at R=1.07. The cut is also visible in the authors' stored output (Working_Data/plot_R05_ceiling.csv: cons_model_pre 1.453 vs cons_model_post 0.265 in year 0). (delta*q)^sigma = 1.0048 vs 1.1081 for the ceiling. External evidence: Best, Cloyne, Ilzetzki & Kleven (2020, REStud) put the EIS around 0.1; Ring (2024, REStud, doi:10.1093/restud/rdae100) finds Norwegian wealth taxes raise saving, with income effects dominating; Brulhart et al. (2022, AEJ:EP 14(4)) attribute at most about 6% of Swiss responses to saving; Advani & Tarrant (2021, p. 520).

Why this severity. Undermines the structural and capital-supply interpretation of C4-C5, not the reduced-form estimates.

The authors' response. The paper states its elasticities 'represent upper bounds on real wealth accumulation responses' and that the implied EIS is much larger than existing estimates (p. 384).

Why it matters, in detail. Unless an unreported estimator explains it (for instance a 2SLS on realised tax status, which is itself a function of the outcome), the printed elasticity SEs understate uncertainty. This weakens the cross-design comparison in F7 and any meta-analytic weighting of these elasticities.

Evidence. scripts/elasticity_se_check.py / output/elasticity_se_check.txt. Table II relative SE of the average ITT vs the elasticity: (1) 0.140 vs 0.134; (2) 0.237 vs 0.138; (3) 0.357 vs 0.264; (4) 0.094 vs 0.060; (5) 0.368 vs 0.143; (6) 0.324 vs 0.100. For the preferred ceiling column, a consistent SE is about 3.66 instead of 1.138 (wrt 1-tau) and about 0.127 instead of 0.038 (wrt net return). OA Table A.II col 6: ITT average 0.151 (0.089), t=1.7, yet elasticity 20.663 (2.049), t=10. Auditor B noted the ceiling column as an unresolved diagnostic.

Why this severity. Overstates the precision of secondary results (the elasticities) by up to about 3x in the preferred very-wealthy column; point estimates are unaffected.

Why it matters, in detail. The 'very wealthy' estimate is silent on the top 0.01%. Applying it to billionaire-level taxes, as downstream revenue scores have done, extrapolates beyond the estimation sample.

Evidence. DD_balance_and_treatment_status.do lines 195-223 (main windows) and 278-294 (placebo windows): 'replace treatmt_*/control_* = 0 if form_pctile19xx > 0.9999'. No disclosure found by text search of the article or Online Appendix. PWBM (2019) fn 34 builds its -20.8 semi-elasticity from the WP ceiling estimate.

Why this severity. Affects external validity at the extreme top; the effect on the estimate is unknown and probably small.

Why it matters, in detail. All of these conventions understate the paper's size claims. Users who need level effects or local elasticities should convert, but no conclusion changes.

Evidence. Exact level changes: 20.5%, 36.6%, 35.3%, 64.7% and 91.7%. Log-log elasticities: 0.92 / 0.81 / 1.45 vs 0.77 / 0.58 / 1.15. Ceiling mechanical share: 42% under the paper's residual convention, 50% in levels, 58% as log(W_fixed-c / W_base) (output/decomposition_convention.txt; Auditor B's figures reproduced).

Why this severity. Conservative presentation conventions with no effect on conclusions; tables and figures are correctly labelled in log points.

The authors' response. Tables and figures are labelled as changes in log wealth.

Why it matters, in detail. None of these slips alters a reported estimate. The 'per capita wealth' mislabel overstates the age-90 wealth multiple: the arithmetic-mean ratio would be well below 18.

Evidence. QJE pp. 331, 335, 339 fn 12, 345 fn 18, 346, 353, 358, 381-384; OA p. 23; calibration_alpha_sigma.m (netOfTauAfterReform = 0.9925); Top Wealth Shares_USDK.xlsx (2012 top 1% = 0.19011); Calibration_Data.xlsx row for age 90 (2.9112); NBER WP p. 34 (1.56 pp).

Why this severity. Presentation slips; none affects the estimates.

Why it matters, in detail. These inputs are immaterial for the 30-year effects, which are pinned by the end-of-life restriction (F1). The sensitivity of sigma to the income assumption is further evidence that the reported EIS (C5) is weakly identified.

Evidence. Numeraire factor exp[(logY - logYn) - (logW - logWn)] = 1.146 (Auditor B; Auditor A's 0.873 inverts it). Recalibration (output/calibration_input_checks.txt): ceiling 30-year effect 0.651 -> 0.643 (units) and 0.654 (ages 61-90); with zero income, 0.678 (ceiling) and 0.306 (couples). The calibrated sigma moves a lot, however: ceiling sigma 6.66 -> 13.9 with zero income.

Why this severity. Tested: under 2% change in the long-run effects (under 5% even with zero income).

Why it matters, in detail. The 'after eight years' magnitude for couples partly rests on one noisy year with confounding policy changes. A 7-year horizon gives 0.132, and the calibration impact is -7%.

Evidence. scalars.do (allowance_1995 1,592,000; allowance_1996 2,400,000; tax_rate_1996 0.007); Lex.dk 'Formueskat' (P. Schou): 'I forbindelse med et forlig om finansloven for 1996 blev formueskatten saenket til 0,7 % i dette aar og helt afskaffet fra og med 1997'; OA fn 3. Couples ITT 0.074 -> 0.096 and TOT 0.132 -> 0.186 from 1995 to 1996 (SEs 0.016/0.029). Dropping the 1996 moment gives a couples 30-year effect of 0.281 (vs 0.304); the ceiling is unchanged (output/drop1996_check.txt; Auditor A's numbers reproduced).

Why this severity. The 1996 jump could account for up to about 0.05 of the quoted 0.186 if it were spurious; it is negligible for the long-run results.

The authors' response. OA fn 3 notes the 1996 rate cut and threshold increase and excludes them from the mechanical-effect calculation.

Why it matters, in detail. Normalising on a base year whose measurement date follows enactment builds any anticipatory under-reporting at end-1988 into all post-reform DD coefficients. The evidence for this is weak.

Evidence. Exact Fig V-B and VI-B series: couples ITT 1986 0.011, 1987 0.012, 1988 0; ceiling raw 1985 0.037, 1986 0.031, 1987 0.028, 1988 0. The dips are about 12% of the couples TOT and about 18% of the raw ceiling TOT when scaled by the 1996 first stages, but they are within the year-to-year noise of the pre-period.

Why this severity. Speculative; if real, a 10-20% upward shift in the 8-year effects.

The authors' response. fn 16: 'there was essentially no room for anticipation effects'.