# 4. Data and measurement

;; 4.01
The household data are the 2022 Survey of Consumer Finances. The SCF is the only public United States microdata in which top-percentile municipal holdings are measurable: it oversamples wealthy households from an IRS-derived list frame, and its income questions cover calendar 2021, so every tax computation in the paper uses 2021 law. The survey's five implicates are pooled with weights the Board distributes already divided across implicates; the pooled weights sum to 131.3 million households.

;; 4.02
Households are not tax units, and the conversion is not rebuilt here. We use the NBER SCF-to-TAXSIM files — produced by Kevin Moore's Federal Reserve Board conversion program and distributed by NBER — which split each SCF household into filing units with incomes, deductions, and dependents assigned by the program's rules. The analysis keeps the primary economic unit's tax units — {{N:units_total:,}} of them for 2022, representing 148 million weighted units — and excludes the financially independent co-residents whose assets the survey does not separate. That universe is smaller than the roughly 180 million units in return-based tables, which include dependent and other filers; levels here are therefore not directly comparable to published schedules, while differences within the paper are.

;; 4.03
Statutory rates come from TAXSIM (Feenberg and Coutts 1993; version 35), which computes 2021 federal law including the net investment income tax and the alternative minimum tax. TAXSIM's default marginal rate is a wage-margin rate, and for municipal analysis it is wrong in both directions: a household's marginal rate on interest differs from its rate on wages whenever an extra dollar of interest crosses a threshold the wage margin does not. We therefore compute the interest margin directly, running every tax unit twice — once as reported, once with $1,000 of additional taxable interest — and differencing the liabilities. A Massachusetts example from the state calculations illustrates the stakes: a couple whose default combined marginal rate reads {{N:ma_example_default_pct:.1f}} percent has a true interest margin of {{N:ma_example_true_interest_pct}} percent, because the added interest also pushes capital gains across the net investment income tax threshold.

;; 4.04
The reconstruction validates against its source. Our as-reported liabilities match the federal tax computed in the NBER files for {{N:taxsim_validation_within1_pct}} percent of units to within one dollar; the residual traces to TAXSIM revisions since those files were produced, concentrated in 2021's refundable credits, not to input mapping. Both of our runs use the same calculator on the same day, so the differenced margins are internally consistent. Every submission to the calculator carries a pinned self-check case whose hand-computed liability must return exactly, and a run aborts if it does not (Appendix A1).

;; 4.05
The municipal income measure is reported tax-exempt interest, which the SCF asks directly: ${{N:agg_taxexempt_interest_bn}} billion in aggregate for 2021, held by {{N:units_with_munis_pct:.1f}} percent of tax units. The survey question spans interest received directly and through tax-exempt funds, matching the household-channel concept the aggregate anchors use. The alternative — imputing income as holdings times a current curve yield — is rejected because reported income captures the coupon flow on which the household actually bore the implicit tax, at the yields prevailing when its bonds were bought; imputation at 2021 curve yields, which were far below the stock's average coupon, would understate that flow by more than half — though reported coupons on premium bonds also exceed economic income, since part of each coupon amortizes the premium, a bias running the other way. The market-value construction — τ* times the 2021 average ten-year Treasury yield of {{N:dgs10_2021_avg_pct}} percent times the ${{N:scf_muni_mv_tn}} trillion of survey holdings (direct bonds of ${{N:scf_direct_munis_bn}} billion plus tax-exempt funds of ${{N:scf_taxexempt_funds_bn:,}} billion, the survey counterpart of the Financial Accounts household channel) — gives ${{N:burden_mv_2021_bn}} billion, the opportunity-cost burden at 2021 market terms. The two are not a bracket on one quantity. They measure different estimands — the give-up embedded in the coupon flow the household actually received, against the give-up on the stock it actually held at that year's market yields — and the survey cannot order them, because it records neither acquisition prices, nor vintages, nor premium amortization, nor maturity composition. Both appear in Section 5, each labeled with its own estimand, year and asset base. [WRDS robustness to be added: an imputation built on holdings-weighted vintages once maturity composition is available.]

;; 4.06
The administrative record disagrees with the survey. IRS Statistics of Income tabulations report ${{N:soi_2021_taxexempt_interest_bn}} billion of tax-exempt interest for tax year 2021, on 6.6 million returns; the SCF aggregate exceeds it by a factor of {{N:scf_soi_ratio}} — and the gap has a second dimension: the survey has roughly half the holders ({{N:scf_holder_units_m}} million units against {{N:soi_returns_m}} million returns) holding almost three times as much per holder (${{N:scf_per_holder_dollars:,}} against ${{N:soi_per_holder_dollars:,}}), so the survey's distributional shape may overstate concentration as well. Candidate explanations run in both directions: non-filing holders and under-reporting on line 2a deflate the SOI figure, while respondent rounding and gross reporting of fund distributions inflate the survey's. Rather than adjudicate, the paper scales: every aggregate burden number has an SOI-scaled sibling, computed by multiplying by the ratio of the administrative to the survey base, and the distributional results — which depend on where tax-exempt interest sits, not on its level — are presented with the survey's shape in both cases.

;; 4.07
Tax units are ranked by expanded income: cash income plus tax-exempt interest at the observed yield. The baseline effective rate is remitted federal individual income tax over expanded income; the corrected rate adds the implicit tax to both sides, per Section 2. The corrected object is the federal individual income tax schedule — not an all-federal-tax schedule of the kind CBO publishes; mapping the correction into a schedule that includes payroll and corporate taxes would dilute it by roughly the individual income tax's share of federal taxes at the top. Ranking is not driving the results: re-ranking units by corrected income moves the top-percentile correction from {{N:top1_delta_pp_headline:.2f}} to {{N:top1_delta_pp_reranked:.2f}} points.

;; 4.08
The public SCF has no state identifiers, and state taxes matter twice: they raise holders' statutory rates on taxable interest, and Section 3's identified market rate embeds them. The paper's headline is federal-only by design. As robustness, every municipal-holding unit is run through all fifty-one TAXSIM state calculators, and the results are combined two ways. A population-weighted expected state rate adds {{N:exp_state_mtr_dollarwtd_pct}} percentage points to the dollar-weighted holder's marginal rate. Population weighting understates the add-on if municipal clienteles concentrate in high-tax states holding in-state paper — a pattern documented directly by Babina, Jotikasthira, Lundblad, and Ramadorai (2021) — so an upper bound is also reported: assigning every holder her maximum-state rate adds {{N:exp_state_mtr_hightax_bound_pct}} points, putting the dollar-weighted holder at {{N:combined_hightax_pct}} percent, above the identified market rate. Section 5 returns to what the pair establishes, and to why only the maximum-state figure is a bound in the literal sense.{{FN: TAXSIM's state law is frozen at 2021, which matches the income year exactly, so no overlay of later state changes is needed.}}

;; 4.09
Aggregate anchors come from the Financial Accounts. As of the first quarter of 2026, ${{N:z1_muni_total_bn:,}} billion of municipal securities are outstanding, of which {{N:z1_household_channel_share_pct}} percent is held through channels that belong to taxable households: {{N:z1_share_households_pct}} percent directly, {{N:z1_share_mutual_funds_pct}} percent through mutual funds, {{N:z1_share_cef_etf_pct}} percent through closed-end funds and exchange-traded funds, and {{N:z1_share_mmf_pct}} percent through money market funds. The SCF's direct-holdings aggregate of ${{N:scf_direct_munis_bn}} billion undercounts the Financial Accounts household line by roughly half on our 2022 comparison; Bergstresser and Cohen (2015) document the same undercount pattern in earlier waves. The paper therefore uses the Financial Accounts for stock-based aggregate constructions and the SCF for everything distributional.

;; 4.10
Every number in the paper regenerates from public sources by one command. The replication package pulls the SCF, the NBER tax-unit files, the curves, the Financial Accounts and Treasury series, the SOI tables, and the Census state populations; runs the TAXSIM submissions with their pinned self-checks; and rebuilds every table, every figure, and a registry of every number quoted in the text, against which the prose is mechanically compiled. Appendix A5 gives the details.

;; 4.11
The data cannot yet support three things. The public SCF's missing geography means state rates are expected rates, not actual ones; the restricted file would fix this. The household data contain no maturity composition — Section 5 states what that costs. And the curve history before 2001 is naive; [WRDS robustness to be added: author-constructed curves from the MSRB transaction feed, and MMD scale history, when institutional access arrives]. Each appears again in Section 8's forward design.
