Who Benefits from Tax Expenditures on Capital? Evidence on Capital Income and Wealth Concentration

Why it matters, in detail. Capitalized fixed-income wealth of group g is W_g = Y_g x (W/Y). If the top's true yield is q times the aggregate yield r, its holdings are overstated by the factor q. - With q about 1.6 for $20m+ decedents in 2008-11 (about 1.2 in 2007), the top-0.1% taxable-bond component falls from 8.5% to 5.2% of total wealth in 2012. - The top-0.1% share then falls from 22.0% to 18.7%, leaving a 2007-12 rise of +1.8 pp instead of +4.3 pp. - Milder corrections give 2007-12 rises of +2.9 to +3.3 pp: the income-ranked Treasury-yield variant, and the authors' 2020 revision (q = 1.4 since 2008, wealth-ranked, with bond and money-market funds reclassified as dividend-generating). Net effect: the post-2007 trend is biased upward, and the 2012 level of the top-0.1% share is overstated by roughly 2-3 pp (about $1.3tn at the 2012 denominator under the heterfix variant).

Evidence. wealth_baseline_soi, 2007 -> 2012: - top 1% +5.87 pp, of which fixed claims +5.95 pp; - top 0.1% +4.34 pp, of which fixed claims +4.02 pp; - top 10% +5.60 pp, of which fixed claims +6.28 pp. The top-0.1% fixed-claims component rose from 3.31% (2000) to 9.45% (2012) of total wealth. Underlying interest data: - Taxable interest fell from $270bn to $112bn, while the top-1% share of interest (ranked by interest) rose from 56.2% to 74.8% (intcomp_soi). - The aggregate capitalization factor (1/rintinctax) rose from 34.2 (2007) to 86.7 (2011). - The $20m+ estate premium (retint20/rintinctax) was 1.22 (2007), then 1.68, 1.54, 1.60 and 1.63 (2008-11). Authors' own variants: - QJE Online Appendix Table B41c (top-0.1% taxable bonds divided by the $20m+ premium) reproduces exactly from the IRS workbook: 16.86% (2007), 16.71% (2008), 18.70% (2012). Rise 2007-12: +1.84 pp vs +4.34 (-58%). Rise 2008-12: +1.99 vs +3.03 (-34%). - Workbook sheet wealth_heterfix_soi applies the 10-year Treasury yield to the top 1% ranked by income (confirmed in the replication code build_usdina(SZ).do). Top 0.1%: +3.33 pp (-23%) for 2007-12 and +2.09 (-31%) for 2008-12. Top 1%: only -8% and -12%, because income rank does not pick out wealth holders. - Authors' current series (WID 2026 vintage, tax units, shwealt992): 2007-12 top 0.1% +2.91 pp (-33%), top 1% +3.41 (-42%); 2012 is a local peak. - The adjusted SCF series plotted in the paper's own Fig 7 (QJE Table C4b) shows the top 0.1% at 15.3% (2007) and 17.2% (2013). Text: - The QJE (2016, pp. 549-551) and the QJE manuscript dated 20 Oct 2015 report the 18.7% variant and call the premium 'not negligible in relative terms'. - The IRS revision of Sept 2015 says only 'a modest taxable interest rate premium' (pp. 10-11) and 'the top wealth shares have increased sharply from 2007 to 2012' (p.5). Scripts: scripts/f1_fixed_income.py, scripts/wid_series.py.

Why this severity. The post-2007 surge is the paper's distinctive new empirical result. Variants built from the authors' own data cut the 2007-12 rise in the top 0.1% share by 23-58% and the 2008-12 rise by 31-34%, above the rubric's one-quarter threshold. A rise survives in every variant and the long-run 1979-2012 increase keeps its direction, so the issue is major rather than critical.

The authors' response. QJE 2016 (pp. 549-551) reports the estate-yield variant: the top 0.1% share reaches 18.7% in 2012 against 22.0% in the baseline. The authors retain the baseline because only a few hundred non-married individuals die each year with estates above $20m, while conceding 'perhaps leading to some overestimation of top wealth shares since 2008'. Saez & Zucman (2020, NBER w27921) and the PSZ update of February 2022 (pp. 3-4) adopt a premium of 1.4 from 2008 (1.15 in 2003-07) for the top 1% ranked by wealth. They also reclassify bond and money-market funds as dividend-generating, saying this reduces top fixed income 'especially in the post-2008 period'. The critique is thus partially accepted.

Why it matters, in detail. Flat yields across estates of $0.5m to $20m+ say nothing about the ratio of top yields to the rest, and that ratio is what determines bias. Where it can be observed: - interest: 1.05-1.15 even before the crisis, implying top fixed income over-capitalized by 5-15%; 1.5-1.7 at the very top after 2008; - dividends: about 1.4; - dividends plus realized gains: about 0.7, consistent with step-up incentives. This matches the heterogeneous-return evidence of Bricker, Henriques & Hansen (2018) and Smith, Zidar & Zwick (2023).

Evidence. Sources: return_lag0 and QJE Tables C6/C6b; adjudication log. Yield of estate classes relative to the aggregate yield: - Interest: every estate class at 1.05-1.12 times the aggregate in 1997 (the paper's own example: 4.1-4.3% vs 3.9%); classes of $5m+ at 1.2-1.7 times in 2008-2011. - Tabulated 2007-decedent panel (2006 income): interest 3.09% vs aggregate 2.69% (1.15 times); dividends 3.49% vs 2.48% (1.41 times); dividends plus gains 8.56% vs 12.21% (0.70 times). - Micro data: dividend yields 0.7-2.6 times the aggregate by class and year; munis mostly 0.83-1.00 times. Text, pp. 10-11: 'within-asset-class returns appear constant across wealth groups'; 'rates of return are similar across the wealth distribution'; munis 'similarly negligible'. Caveat: the estate categories ('publicly traded stock'; 'fixed claim assets' including cash, notes and bond funds) differ from Financial-Accounts classes. Level comparisons are therefore imperfect, which itself limits these data as a validation.

Why this severity. This weakens the main validation offered for the capitalization method behind all wealth shares. For interest, the quantitative consequence is the one already measured in F1 and is not counted twice. For equities the deviations run in both directions with no demonstrated net bias. Hence moderate.

The authors' response. QJE 2016 (p. 549) cautions that decedents' wealth-return profile may differ from that of the living, and leaves mortality-weighting of the matched data to future research. Saez & Zucman (2020) accept 'the presence of a small interest rate premium at the top'.

Why it matters, in detail. The incidence of each preference depends on the distribution of the preferred base, the marginal-rate differentials, eligibility and itemization. From 2008 to 2012 the bases' concentration was flat: - equities +1.0 pp; - pensions +0.3 pp; - owner-occupied housing +0.4 pp; - mortgage debt -0.3 pp; - munis -1.0 pp. All of the rise in wealth concentration is in taxable fixed income, which carries no tax preference. The 2008-12 rise in capital-income concentration is the 2012 retiming spike in gains and dividends.

Evidence. Share of each asset held by the top 1% (ranked by wealth), 2008 -> 2012 (scripts/f_regressivity.py): - equities 68.3% -> 69.3%; - pensions 16.2% -> 16.5%; - owner-occupied housing 9.1% -> 9.5%; - owner mortgages 6.0% -> 5.6%; - munis 68.7% -> 67.7%; - taxable bonds and deposits 47.8% -> 64.2%. Top 0.1%: munis 35.5% -> 32.6%; taxable bonds and deposits 28.4% -> 41.5%. Over 2007-12, equities rose 65.4% -> 69.3% for the top 1%. Capital income: the top-0.1% share was 38.2% (2008) and 37.0% (2011) before the 2012 retiming spike (see F7). Codex gives a numerical counterexample in which a group's asset share rises from 20% to 30% while its benefit share falls from 20% to 14.6%.

Why this severity. The policy implication that answers the paper's title question is unsupported for 2008-2012. The only rising component is a non-preferred asset, and its rise is itself largely an artifact (F1). The distributional series themselves are unaffected. Hence moderate.

The authors' response. The abstract and conclusion present the series as 'a first step' and 'the key input needed' for a distributional analysis of tax expenditures. That tempers, but does not withdraw, the implication drawn on p.2.

Why it matters, in detail. Capital gains and dividends are highly elastic to anticipated rate changes; the paper notes elsewhere that retiming saves about 10 points of marginal tax on capital income. The jump in 2012 alone exceeds the entire 2007/08-2012 change in capital-income concentration. Equity wealth, capitalized from dividends plus gains, inherits part of the spike. A 2013 observation, or a 2011 endpoint, would show no rise in capital-income concentration after 2007.

Evidence. kinc_kg_soi, top-0.1% capital-income share including gains: 38.32% (2007), 38.20% (2008), 34.31% (2009), 38.48% (2010), 36.95% (2011), 42.30% (2012). - The 2011 -> 2012 jump of +5.35 pp breaks down as realized gains +5.28 and dividends +1.42. - Excluding gains: 29.77% -> 33.45%. - Passive capital income: 35.07% -> 41.35%. - Aggregate realized gains rose from $375bn to $621bn, dividends from $195bn to $260bn. - Top-0.1% share of dividends (ranked by dividends): 41.7% -> 50.0%. Wealth: the top-0.1% share rose from 20.33% to 22.01% in 2012. Equities account for +0.58 pp of that in the baseline and +1.23 pp in the dividends-only variant. WID 2026: 2012 is a local peak (top 1% 39.1% -> 38.0% in 2013). Saez, 'Striking it Richer' update of 3 Sept 2013: 'part of the surge of top 1% incomes in 2012 could be due to income retiming ... Retiming should be most prevalent for realized capital gains'. Update of 25 Jan 2015: 'This retiming inflates 2012 top income shares and depresses 2013 top income shares.' Script: scripts/f_retiming.py.

Why this severity. This removes the support for 'capital income concentration has grown substantially from 2008 and 2012': from 2008 to 2011 the share including gains fell 1.25 pp. Wealth shares are affected only by roughly 0.6-1.2 pp of the 2012 top-0.1% level, and long-run trends are untouched. Hence moderate.

The authors' response. Not discussed in this paper. Saez's own top-income updates flagged the retiming from September 2013 onward; from the July 2015 update they state that 'top income shares in 2012 were abnormally high due to retiming of income from 2013 to 2012'.

Why it matters, in detail. The '25%' appears to transcribe 25.3 million households as a percentage. With the cited 21.5%, the elderly-to-average wealth ratio is about 1.6-1.7 in the 1960s and about 1.7-1.8 in 2010-12. It dipped to about 1.4 around 1979-84 and has risen since. The comparison also mixes tax units (wealth shares) with households (Census), an unresolved unit inconsistency that the paper notes only for the definition of 'elderly'.

Evidence. US Statistical Abstract 2012, Table 62 (the paper's footnote 4): 117.5m households in 2010, of which 13.2m had a householder aged 65-74 and 12.1m aged 75+, i.e. 25.3m or 21.5%. Census HH-3: 18.8% (1962), 21.5% (2010), 22.2% (2012). Elderly share of total wealth (QJE Table B4; workbook wold0/wsu0): 30.6% (1962), 33.0% (1965), 33.1% (1970), 36.8% (2010), 37.8% (2011), 37.9% (2012). Resulting ratio of elderly to average wealth: 1.63, 1.73, 1.68 in the 1960s vs 1.71, 1.77, 1.71 in 2010-12. The paper's 40% figure is 0.368/0.25 = 1.47. The 2011 elderly share of top-0.1% wealth is 38.6% (36% in the text); in 1962 it was 46.4%. Script: scripts/f_elderly.py.

Why this severity. Correcting the denominator reverses a stated secondary finding: elderly relative wealth did not decline. No core finding depends on it, so moderate.

The authors' response. Footnote 4 notes that the Census defines elderly families by householder and the tax data by either spouse; there is no other response.

Why it matters, in detail. The paper's relative mortality is R = m_top/m_filers; against the population it is R_pop = R x (m_filers/m_pop). Between 1979-83 and 2004-08: - m_filers/m_pop rose from 0.74 to 0.80 for men aged 65-79 (coverage 63% -> 82%) and from 0.68 to 0.79 for women (48% -> 73%); - men top 1%: R_pop 66 -> 48 (change in log -0.32, vs -0.39 in Table 1); - men top 10%: 70 -> 61 (-0.14 vs -0.21); - women top 10%: 64 -> 60 (-0.07 vs -0.22). The KS factors at 65-79 (78 for men, 82 for women) already exceeded the population-benchmarked top-10% rates in 1979-83. Estate multipliers therefore understated top wealth from the start, while the trend in the bias is smaller than Table 1 suggests.

Evidence. Coverage = weighted filer counts in mortality_big divided by Census population (5-year sums): - ages 65-79: men 0.63 (1979-83), 0.73, 0.75, 0.73, 0.75, 0.82 (2004-08); women 0.48 -> 0.73; - women 80+: 0.26 -> 0.58; - all ages 30+: men 0.86-0.88, women 0.76-0.85. The internal TY2007 file holds 153.6m units vs 143.0m returns published by SOI. It therefore includes about 10.6m returns 'filed by individuals to only receive the economic stimulus payment and who had no other reason to file' (SOI Pub. 1304, 2007), which SOI excludes from its published counts. This coincides with the 2004-08 jump in elderly coverage (men 65-79: 0.75 -> 0.82; women 80+: 0.47 -> 0.58). Filer vs population mortality, ages 65-79 (5-year horizon; Human Mortality Database m_x, validated against NCHS life expectancy at 65: 1980 men 14.11 vs 14.1, women 18.32 vs 18.3): - filer mortality as a share of population mortality: men 0.74 -> 0.80, women 0.68 -> 0.79. Re-benchmarked to the population, ages 65-79, 1979-83 -> 2004-08 (Table 1 in brackets): - men top 1%: 66 -> 48 (88 -> 60); - men top 10%: 70 -> 61 (95 -> 77); - women top 1%: 62 -> 54 (92 -> 69); - women top 10%: 64 -> 60 (95 -> 76). Share of Table 1's log decline due to reference-group drift: men top 1% 19%, men top 10% 34%, women top 1% 52%, women top 10% 69%. Comparison with Kopczuk-Saez factors: against the population, men in the top 10% in 1999-2008 have relative mortality of about 33, 46, 63 and 78 in the four age bands. The KS factors are 58, 62, 78 and 96. So Fig 4's 'very close to the mortality correction used by Kopczuk and Saez' is an artifact of the benchmark. Consistency with Fig 5 (Pareto mapping k^(1/a), a = 1.5): Table 1's benchmark implies decedent/full top-1% ratios of 1.00-1.03 for 1979-98, against 0.73-0.84 observed in Fig 5. The population benchmark implies 0.79-0.81. Scripts: scripts/f_mortality.py, scripts/f_table1_fig5.py.

Why this severity. Against the population, the wealthy's mortality advantage is larger in level. That strengthens the first half of C6 and the case that estate multipliers understate top wealth. But the widening 'since 1979' shrinks by about a fifth to a third for men aged 65-79, and by about half to three-quarters for women. The Fig 4 comparison with KS reverses. The men's headline trend survives, and the computation relies on external life tables and approximations. Hence moderate.

The authors' response. None in this paper. The QJE (2016, p. 572) labels the benchmark correctly, as 'the average mortality rates of male tax filers aged 65-79' (60% vs 90% in 1979-83), but does not adjust for filer selection.

Why it matters, in detail. Reweighting decedents by 1/(m_pop x KS) reproduces the living distribution of capitalized wealth only if mortality conditional on capitalized wealth equals the assumed rates. A failure therefore reveals mortality gradients in capitalized-income space. These equal gradients in true-wealth space only if decedents' income-to-wealth ratios near death resemble those of the living. Selection on health (Kopczuk 2015, pp. 61-62) and estate-tax planning such as trusts and gifts would both make the measured gradient steeper. The share of the 1979-2012 estate-capitalization gap that is due to mortality is therefore not identified.

Evidence. Fig 5 data (wealth_baseline_soideath vs wealth_baseline_soi): - top 1% share 1979 -> 2012: decedents 20.7% -> 24.1%; full sample 24.2% -> 41.8%; - decedent/full ratio by period: 0.84, 0.81, 0.84, 0.73, 0.72, 0.59. The Table 1 gradients, mapped through k^(1/a) with a = 1.5, account for only part of that fall: predicted 1.03 -> 0.87 with the filer benchmark, 0.81 -> 0.70 with the population benchmark (scripts/f_table1_fig5.py). QJE 2016, p. 572: the mortality rates 'not only capture differential mortality by lifetime income, but also estate tax planning and any effect that the approach of death may have on real wealth. We leave to future research the difficult task of identifying what part...'. Counter-evidence: the divergence persists for passive income (top 0.1%: living 13.4% -> 41.4%; decedents 11.2% -> 19.7%) and when capital gains are excluded.

Why this severity. The descriptive failure of the KS weights is robust. What is overstated is C7's causal attribution ('fully', 'because of'). The passive-income check limits the alternative channels but does not eliminate them. Hence moderate.

The authors' response. The paper reports that the divergence persists when realized gains are excluded and for passive income only (pp. 11-12). The QJE (2016, p. 572) acknowledges that estate-tax planning and approach-of-death effects are mixed into the measured gradients and leaves the decomposition to future research.

Why it matters, in detail. R_g = sum_a p_ga x m_ga / sum_a p_all,a x m_all,a with p_g different from p_all is not an age-standardized ratio. Because top groups are older, R_g overstates relative mortality by up to about 20 index points at ages 30-49, but by at most about 3 points at 65-79, where the trend is unaffected.

Evidence. Recomputed from mortality_big (scripts/f_mortality.py), common all-filer weights vs Table 1: - men 30-49, top 1%, 1994-98: 56 vs 76; - women 30-49, top 1%, 1999-2003: 45 vs 59; - men 65-79, top 1%: 87 -> 58 vs 88 -> 60; - women 65-79, top 1%: 90 -> 67 vs 92 -> 69. Codex obtains the same values (0.556 vs 0.760; 0.866 -> 0.585).

Why this severity. The procedure is disclosed, and the 65-79 trends change by only 1-3 points. The bias works against the paper's 'mortality declines with wealth' claim, so that claim is understated rather than overstated. The issue affects only level comparisons at younger ages and the KS comparison. Codex rated it moderate; minor is more accurate.

Why it matters, in detail. The robustness check cannot be verified for 2009-2012. For 1979-2008 the unit effect is 0.1-2.1 pp on the top 1% share, small relative to the gap with estate-based estimates.

Evidence. wealth_indiv_soi equals wealth_baseline_soi in 69 of 70 cells for each year 2009-2012; only wm3 differs. In 1979-2008 the individual-unit top 1% share is on average 1.13 pp lower (maximum 2.10 pp; 2008: 36.03% vs 38.13%), so the copy creates an artificial +1.8 pp jump in 2009. QJE Table B1b instead extrapolates post-2008 individual shares using the growth of the family series, and discloses this (2012: top 1% 39.52%, top 0.1% 20.78%). Script: scripts/f_indiv_ira.py.

Why this severity. Where data exist, the 'negligible difference' claim holds: about 1-2 pp against an estate-capitalization gap of about 13 pp. No headline number uses this sheet, so minor.

The authors' response. QJE Table B1b discloses that its post-2008 individual-unit series is an extrapolation; the IRS workbook does not.

Why it matters, in detail. With the workbook yields, the $20m+ premium exceeds 1.4 in 1998, 2000, 2003, 2005 and 2008-11. The premium then looks persistent rather than post-crisis, so pre-2008 top fixed income may also be over-capitalized. The B41c-style correction would lower the 2000 top-0.1% share by 0.8 pp (15.19% vs 15.89%).

Evidence. $20m+ class: 5.81% (1998) and 6.15% (2000) in return_lag0, vs 4.36% and 4.34% in C6b. $10-20m class in 2000: 5.43% vs 3.86%. QJE Table B41c uses the C6b values (premium 1.10 and 1.04, vs 1.47 and 1.48 from the workbook). Scripts: scripts/f1_fixed_income.py and the C6b comparison in the adjudication log.

Why this severity. The effect on levels is small. But under the workbook values a $20m+ premium of about 1.5 also appears in 1998 and 2000, which undercuts the text's dating of the premium to '2003, 2005 and since 2008'.

Why it matters, in detail. Shares of IRA wealth held by pension recipients and by above-median earners do not pin down the weights of the two allocation keys. The split is a calibration to SCF pension concentration, disclosed only in the QJE.

Evidence. fracira: 2011 IRA wealth $5,349bn (28% of pension wealth). 2004-11 averages: 63.9% held by filers with pension income; 47.7% by filers with wages above P50; 9.6% by filers with no pension income AND wages below P50 (Explanations row 169). The text on p.6 says 'or'; the literal union would be about 78% in 2011. The 60/40 split is calibrated to the SCF top-10% pension share (QJE 2016 fn 24), not to these statistics. IRAs are more concentrated than imputed pension wealth (top 1%: 36% vs 24-26%), as QJE fn 25 notes. Script: scripts/f_indiv_ira.py.

Why this severity. These are wording and justification problems in a secondary calibration. Pension wealth is about 5% of top-0.1% wealth (QJE p. 537), so the effect on top shares is negligible.

The authors' response. QJE 2016 fn 24-25 explain the SCF calibration and note that IRAs are more concentrated than total pension wealth.

Why it matters, in detail. The capital share of selected individual income-tax expenditures is 49% as computed by the paper. It is 41% under OMB's treatment of preferential rates on gains and dividends, and 54% if the TPC-omitted capital items are added. Interaction effects are not separable for capital items alone. Business and corporate tax expenditures, which are largely capital-related, are outside the total.

Evidence. Burman, Toder & Geissler (2008, TPC Discussion Paper 31), Table 1, p.7: - $701.8bn is the sum of selected individual income-tax expenditures without interactions; with interactions it is $760.5bn (+8.4%). - Capital items: life insurance 18.3 + retirement 126.8 + tax-exempt bonds 12.2 + capital gains 83.7 + dividends 11.1 + mortgage interest 92.4 = $344.5bn, or 49.1%. - The $1,020.7bn is the model's 'baseline individual income tax revenues' under 2007 law before the AMT patch (pp. 3, 7), not actual collections. - The provisions cover about 90% of nonbusiness individual tax expenditures and exclude business and corporate items. - Omitted provisions include two capital preferences: the exclusion of gains on home sales ($37.0bn) and step-up of basis at death ($32.6bn) (p.5, fn 9). - Note 2: OMB does not count the lower rates on capital gains and dividends as tax expenditures. Resulting capital share: 41.1% if the gains and dividends preference is excluded; 53.7% if the two omitted capital items are added.

Why this severity. 'About half' holds within roughly 41-54% under the source's own alternative definitions, and the paper cites the exact source column. Hence minor; Codex rated it moderate.

The authors' response. The paper notes that the Tax Policy Center excludes imputed rent, and that Treasury (OTA) estimates are more comprehensive (p.1).

Why it matters, in detail. Mislabelled data sources, wrong years and unflagged infeasible estimates (top-1% estate shares after 2003) are presentation defects. The absence of sampling uncertainty matters most for the $20m+ estate yields and the top-1% mortality cells, which are non-monotonic in several bands.

Evidence. Missing or mis-described sources: - Sheets cited for Table 1 ('mortality_raw', 'Kopzuk-Saez_mortality') and Appendix Tables C4, C4b, C6 and C6b (Figs 6-7) are not in the companion workbook; they are in the QJE appendix workbooks. - Fig 7 bottom: the text (p.11) and notes say decedents' estate returns were matched to income returns since 1979. The plotted series is the income-file decedent sample (kinc_kg_soideath); internal matched estate-income files cover only 1996-2011. - Eller et al. (2001) is summarized as finding 'about 2-4% for audited returns', omitting their estimated 13% reporting gap and their conclusion that estate-tax noncompliance is likely larger than income-tax noncompliance (pp. 9-10). Wrong years and dates: - 'In 1976 ... about 15% in both': Fig 7 bottom starts in 1979 (16.2% vs 13.5%). - 'Remarkably similar in both level and trend' for 1916-1976 holds for 1946-76 (mean gap 0.56 pp) but not for 1916-45 (mean 3.1 pp; 1932: 22.4% vs 14.7%). - The divergence is dated 'after 1976', yet the gap between capitalized and estate-based top-0.1% shares fluctuates between -0.4 and +2.1 pp in 1981-86 (1986: 9.2% vs 9.6%) and widens persistently only from 1987-88 (1988: 11.6% vs 9.0%; 2000: 16.0% vs 9.1%) (adjudicator; cf. Kopczuk 2015, p. 60, on the Tax Reform Act of 1986). Labels and typos: - Table 1 'Period 1994-1993' should read 1989-1993. - Fig 4 labels the first period '1979-1984' (1979-83 in Table 1). - The Fig 4 note says 90% where Table 1 has 88% (men) and 92% (women). - The text says 'the top 1% ... live less long than the top .1%', but no top-0.1% mortality is estimated. - 'Kopczuk and Saez (2014)' should be (2004). Workbook inconsistencies: - Estate coverage after 2010 is given as '.25%'; the workbook's coverage2 is 0.156% (2011) and 0.154% (2012). - wm0 is documented as total wealth divided by all tax units including non-filers, but it equals total divided by filers (2007: 153.6m). - The KS-extension top-1% thresholds are implausible or negative from 2004 (w3 = $126k in 2004, negative from 2006), yet sh3 is reported without a flag; the paper itself plots only the top 0.1%. - 'PUF and SOI estimates are always very close' holds on average (mean absolute difference 0.08-0.25 pp) but not in 1984 (top 10%: -2.18 pp). The 1984 SOI file shows $0.6bn of estate and trust income, against $6.5bn and $9.7bn in adjacent years. Uncertainty: no standard errors are reported for any mortality cell or estate yield. The workbook says only that each mortality cell has 'at least 30,000 individuals', and the QJE notes that only a few hundred $20m+ decedents die per year. Scripts: scripts/f_docs.py, scripts/f_tra86.py.

Why this severity. None of these changes a headline number. Together they impair verification and slightly overstate the precision and coverage of the estate-based evidence.

Why it matters, in detail. Relative age (the top's elderly share divided by the overall elderly share) fell from 1.41 to 1.03 over 1979-2012. The whole decline comes from the denominator: the top's own elderly share stayed between 32% and 41%. Age is in any case a weak proxy for inherited versus self-made wealth, because longevity and the age at which people inherit both rose.

Evidence. Elderly share of top-0.1% wealth: 46.4% (1962), 45.5% (1970), 39.7% (1979, PUF; 41.5% in the SOI file), 33.6% (1984), 36.1% (1990), 32.4% (2000), 35.6% (2007), 39.1% (2012). Over the same years the top-0.1% wealth share rose from 7.9% (1979) to 22.0% (2012). The overall elderly wealth share rose from 28-30% (1979) to 37.9% (2012). Script: scripts/f_elderly.py.

Why this severity. The descriptive claim (relative age fell since 1979) is literally true, and the causal reading is hedged ('This suggests'). The issue is confined to the interpretation of a secondary result, so minor. Claude rated it moderate; Codex considered it and did not charge it.

The authors' response. The claim is hedged ('This suggests'), and the paper cites Edlund & Kopczuk (2009) on the falling share of widows among top wealth holders as consistent evidence.