Methodology

Every number in this report traces to a named source and a documented method. State election administrators, federal reviewers, foundation officers, and academic reviewers will find the trail below.

The question and the approach

The report answers two layered questions:

1. Diagnostic. Why don't young people vote — by race, gender, and age — and what can change, at what cost, to close the gap?^[1]
2. Strategic (the Second Mile thesis). The participation-equity movement has built voter registration in priority groups. It has not yet built the vote-completion infrastructure for registered young Black and Hispanic adults aged 21-32. Which interventions close that second-mile gap? See Finding 4, Finding 5, and Recommendations.

The seven findings below address question 1; the Second Mile thesis is the strategic frame they serve.

Approach. Pair one large probability survey (1.64 million Census Current Population Survey (CPS) respondents, 2000-2024)^[1] with voter-file-validated state turnout benchmarks (McDonald voting-eligible population (VEP))^[2], state election-law and redistricting-institution data (National Conference of State Legislatures (NCSL)^[6] plus commission-adoption events^[8]), and a peer-reviewed intervention catalog with effect sizes and per-contact costs^[13][14].

Descriptive, not causal. Where the report makes causal claims — specifically around independent-redistricting-commission adoption — it uses within-state pre/post comparison with the identifying assumptions stated.

Scope. U.S. federal general elections, November 2000-2024 (13 cycles), 50 states plus DC, all four major racial and ethnic groups. Not in scope: primaries, off-year elections, sub-state analysis, district-level individual-voter analysis, partisan realignment.

Findings index

The seven empirical findings produced by this methodology. Every numeric claim on each page traces back to a source and method documented below.

• Finding 1 — The persistent gap — CPS 13-cycle senior-youth gap; race × gender within-youth spread
• Finding 2 — The midterm amplifier — cycle-type sensitivity (presidential vs midterm)
• Finding 3 — Who didn't register, and why — CPS VOYNOTREG, access barriers dominate
• Finding 4 — Registered, didn't vote, and why — CPS VOWHYNOT, the Second Mile diagnostic (50% logistical, 23% engagement, 9% access)
• Finding 5 — Method preference: the ROI spine — CPS VOTEHOW × state availability; revealed-not-stated preference
• Finding 6 — State policy as the lever — NCSL × CPS calibrated to VEP; same-day registration + mail expansion
• Finding 7 — Institutional structure — redistricting method × competitiveness × youth turnout
• Recommendations — group-level intervention ROI catalog (23 interventions × effect range × cost range × evidence quality)

Deep-dive interactive explorers: State Youth-Gap Explorer, Race & Demographics.

The five sources that drive every ROI claim

1. CPS November Voting Supplement (via IPUMS)^[1]

• What it is: Census Bureau biennial survey asking adult citizens whether they voted, with full demographic breakouts. IPUMS harmonizes variable codings across all 13 cycles.
• Our extract: 1.64 M respondents, 2000-2024.
• How we measure turnout: per Census P20 methodology — only explicit "voted yes" (VOTED=2) counts as voter; all other codes (didn't vote, refused, don't know, no response, not-in-universe) count as non-voter. Denominator is the full citizen adult population. This is the method Census uses in its own published Voting and Registration tables.
• How we measure why people didn't vote: two CPS variables (VOWHYNOT, VOYNOTREG) ask non-voters and non-registrants for their reason directly. We harmonize individual codes into five policy-relevant categories:
  • engagement — not interested, vote wouldn't matter, didn't like candidates
  • access — registration problems, missed deadlines, didn't know how
  • logistical — too busy, out of town, forgot, transportation, weather
  • personal — illness, ineligible, language
  • other — residual
• Headline finding from this diagnostic: across 2000-2024, youth 18-29 non-voters cite logistical barriers 50% of the time, engagement concerns 24%, access barriers 9%. The dominant causes are policy-addressable.
• Self-report bias: CPS over-reports voting by 3-8 pp vs voter-file validation.^[11] This is disclosed and calibrated against McDonald VEP (source 2).

2. McDonald United States Elections Project (VEP turnout)^[2]

• What it is: state-by-state turnout as a share of the voting-eligible population — citizens 18+, excluding ineligible felons, plus eligible overseas voters. This is the voter-file-validated benchmark Census itself cites.
• Coverage: 1980-2022, all 50 states (v1.2, October 2024, hosted at the University of Florida Election Lab).
• 2024 extension: state totals from MIT Election Data Science Lab (MEDSL) official certified returns^[4] ÷ ACS CVAP denominators^[5].
• 2020 backfill: MEDSL moved its 2020 dataset behind a paywall; we re-sourced state margins from FEC-certified state totals and cross-checked against four well-known anchor margins (GA +0.24pt D, FL 3.36pt R, AZ +0.31pt D, WY 43pt R) — all within 0.02pt of expected.
• How we use it: every state-level turnout claim is compared to matching VEP. Where CPS−VEP residual exceeds 3 pp, we disclose it. Across 50 states × 2020 + 2022 (102 state-years), the CPS mean residual is +2.6 pp.
• Hur-Achen adjusted alternative: we also publish turnout_pct_adj — Hur & Achen (2013) state-level post-stratification that matches CPS rates to VEP by construction.^[16] At cohort='all' state level the match is exact. Published alongside the Census-P20 value; neither is "the right answer" in all contexts. Note: Hur-Achen weighting does not fully remove racial over-report — see Ansolabehere-Fraga-Schaffner caveat.^[17]

3. NCSL state election law tracker^[6]

• What it is: the National Conference of State Legislatures' authoritative compilation of state election-administration laws. We coded six policies directly relevant to youth turnout and registration: Automatic Voter Registration, Same-Day Registration, Online Voter Registration, pre-registration for 16-/17-year-olds, no-excuse absentee voting, and universal vote-by-mail.
• Coverage: 155 state-policy rows, 6 policies, 48 states + DC, with effective dates and source URLs. Effective dates backfilled from state statute citations or NCSL historical summaries. Where a policy has been in continuous effect since before our 2000 window, we use a conservative "1990" default and flag it in the notes.
• How we use it: per state per year, we know whether each policy was in effect. Sum is policy_score (0-6). Correlations with state youth turnout are suggestive, not causal. For the two policy reforms where the literature supports causal attribution (same-day registration, mail-ballot expansion), we cite peer-reviewed estimates directly.^[13][14]

4. Independent redistricting commission adoption events^[7][8]

• What it is: the canonical list of U.S. states that moved congressional redistricting authority from the legislature to a commission, dated by the first redistricting cycle the commission was in effect (the basis used throughout this analysis): Arizona 2002, California 2012, Colorado 2022, Michigan 2022, New York 2022, Ohio 2022, Utah 2022, New Mexico 2022, Virginia 2022.
• How we use it: natural-experiment design. We compute state turnout in the first redistricting cycle after the commission became effective, and compare against the same state's turnout before the event plus matched comparison states. This is the closest we get to a causal claim in the structural analysis.
• Framing: we describe the institutional change (legislature-drawn → commission-drawn), never its partisan consequences. "District non-competitiveness correlates with depressed turnout" is the language; "gerrymandering suppresses votes" is not. The finding is equally valid regardless of which party benefits.

5. Intervention effect-size catalog

• What it is: 23 specific election interventions across six categories — policy reform, administrative improvement, programmatic get-out-the-vote (GOTV), civic infrastructure, institutional reform, and (explicitly) failed-or-overclaimed. Primary source: Gerber & Green, Get Out The Vote (4th ed., Brookings 2019)^[13], supplemented by Brennan Center cost-benefit analyses^[14] and peer-reviewed follow-ups including Results for America evidence-based policy infrastructure^[15].
• Each row contains: intervention, effect-size range (low-high, with CI where available), youth-specific multiplier where documented, per-contact cost, per-marginal-voter cost, evidence-quality rating, and time horizon.
• Ranges not point estimates. GOTV literature has genuine uncertainty, and this report is more credible for acknowledging it than pretending otherwise. A peer-to-peer text campaign might move turnout 0.5-1.5 pp per contact at $40-$100 per marginal voter — not "1 pp at $70/voter" with spurious precision.
• Includes a "do-not-fund" section. We name interventions where the evidence does not support the claims: celebrity "go vote" campaigns, untargeted digital ads, generic SMS broadcasts without peer-to-peer, and brand-awareness GOTV. Credibility here depends on being willing to say what doesn't work.

Standard errors and confidence

Every published rate carries a 95% confidence interval. We use Census Bureau generalized variance function parameters (CPS November 2022 technical documentation, Tables 8-11)^[12]:

SE(p) = sqrt(b × p × (100 − p) / y)

where p is the rate, y the weighted base, b the Census-published parameter for the relevant geography (state, region, division, national). The applied national parameter is b = 5,949 (the Census "Voting and registration" value); per-state, division, and region parameters live in data/external/cps_variance/, so any standard error on the site can be reproduced from the formula above. This is the same method Census uses for its published P20 tables. CPS does not publish replicate weights for the November voter supplement, so Fay's BRR is unavailable — the generalized variance function is the defensible alternative.

Sample size and suppression. Small samples get suppressed or flagged — but the floor depends on the analytical unit, because a single-cycle state cell and a pooled-13-cycle national cell do not carry the same reliability at the same n. Rather than publish one number that fits none of them, we publish the rule for each surface:

The provenance chip and the cell-level rule use the same 100 / 400 thresholds, so a number a chip marks "confident" is one we would publish as a rate. Two chips report a panel-unit count rather than a respondent sample — the state-policy panel's 153 state-years, and the state-gap explorer's count of states shown. Those are marked "confident" against panel completeness, not against the respondent floor, and the chip names the unit so the distinction is visible.

Trust-but-verify — provenance chips on every cell-level claim

Across the analytical pages, cell-level numeric claims carry a small provenance chip ⓘ — hover or focus to reveal the underlying source, unweighted sample size (n), generalized-variance-bounded standard error (SE), year range, and suppression status. This is the operational form of the cell-level disclosure rule above. Example, in context:

Suppression auto-classifies from n, matching the cell-level rule above: n < 100 flags the claim red (suppressed), 100–400 marks it indicative-only, and 400+ renders the disclosure in muted text (confident). The rule is "every untraceable cell-level claim does not ship" — the chip is what makes the rule machine-checkable.

What we tried that didn't deliver

Credible methodology includes admitting the paths that didn't pan out.

• American National Election Studies (ANES) Cumulative Data File as a group-level turnout cross-validator.^[9] For the single well-anchored group (2020 White 18-29), CPS and ANES agreed within 0.6 pp — a tight validation. But for race-disaggregated youth groups, ANES sample sizes (n=56 for Black youth 2020; n=79 for Hispanic youth 2020) produce noisy rates that don't correlate meaningfully with CPS at the group level (r = −0.16 across 73 groups). ANES remains valuable for national-level attitudinal trends (efficacy, trust, political interest) but is not a reliable turnout cross-validator for the cohorts we care about. We disclose this rather than paper over it.
• Election Assistance Commission (EAC) Election Administration and Voting Survey (EAVS) state registration rates vs CPS registration.^[3] CPS runs ~22 pp lower than EAVS roll totals in 2020-2022. This is a known phenomenon — EAVS rolls include inactive records, federal-employee registrations (DC reports 124% of CVAP), and pre-removal lag — not a CPS problem. For citizen-population rate analysis we use CPS; for roll-health analysis we use EAVS. We don't mix them.
• Harvard Institute of Politics (IOP) pre-election "definitely vote" as a turnout indicator.^[10] IOP intent correlates strongly with actual turnout (Pearson r = 0.944 across five cycles), but stated intent overstates actual turnout by 5-9 pp in recent cycles. We use IOP as a directional leading indicator only, not as an absolute turnout estimate.
• CPS VOTEHOW = 3 (mail / absentee) cohort-level breakout.^[1] The IPUMS extract powering this site (cps_00002, 2000-2024) contains zero rows where VOTEHOW = 3; the harmonization in scripts/process_cps.py is correct (code 3 → "By mail (absentee)"), but the source data lacks those rows — almost certainly because the VOTEHOW variable definition was revised across CPS cycles in this window and the current extract did not pull the harmonized mail-ballot code. The user-visible effect is that Finding 5's cohort-over-time and race × gender charts show only Election Day and Early-voting series. The state-availability cut on the same page — universal-mail states vs. all others — still captures the mail-voting story at the aggregate level; the cohort-level mail-uptake breakout is what's missing. This is a standing limitation of the current extract, disclosed here so no reader mistakes the absence of a cohort-level mail series for a finding that young voters don't vote by mail.

Nonpartisan framing commitment

Institutional findings — redistricting, commission adoption, state policy variation — are expressed in terms of competitiveness and governance structure, not partisan advantage. This is a condition of 501(c)(3) compatibility and a necessary posture for a report pitched at Secretaries of State from both parties, federal administrators, and funders across the civic-philanthropy spectrum.

• "District non-competitiveness correlates with depressed turnout, particularly among voters 18-29" — yes.
• "State redistricting reform through independent commission adoption correlates with turnout increases in the first post-adoption cycle" — yes.
• "Gerrymandering suppresses votes" — no.
• "Party X benefits from these reforms" — no.

The empirical findings are equally valid regardless of which party benefits. The framing is deliberate.

Reproducibility

Every number can be regenerated from the documented pipeline and data inventory, which we share on request for review:

• Pipeline: scripts/process_cps.py, scripts/build_state_year_panel.py, scripts/aggregate_eavs.py, scripts/build_warehouse.py
• Build-time precompute: scripts/precompute_findings.py + declarative manifest at scripts/precompute/manifest.py. Every chart on every finding page reads its data from JSON precomputed against the parquet sources at build time, eliminating any browser-side SQL execution. JSON files live under src/data/findings/{slug}/.
• Validation harness: scripts/validate_panel.py — eight checks, all passing.
• Render-test harness: scripts/test-render.mjs — Playwright headless smoke across every demo-path route; asserts SVG count + zero console errors.
• Cross-validation outputs: data/external/cross_validation/{V2,V3,V4}_findings.md
• Full data provenance: DATA_INVENTORY.md

All underlying source datasets are publicly available and free — CPS via IPUMS, McDonald VEP via the University of Florida Election Lab, NCSL's election-law tracker, EAC EAVS, ACS CVAP, and the rest are listed on the sources page. Raw IPUMS and ANES microdata require free user registration at cps.ipums.org and electionstudies.org respectively; otherwise unrestricted. Our own pipeline code, data inventory, and methodology are not in a public repository at this time; they are available on request for review.

Interactive group exploration. Free-range multi-dimensional group exploration (arbitrary age × race × region × education selection) ships in Phase 1 of the platform roadmap. The v1 demo focuses on the group-targeted findings indexed above; intermediate-depth interactive exploration is available now through State Youth-Gap Explorer and Race & Demographics.

Funding and affiliations

Ballot Bridge Initiative operates as movement infrastructure for the participation-equity ecosystem — open-methodology, foundation-funded, free-tier-by-default group-level intelligence for civic 501(c)(3)s, Secretaries of State, foundations, and academic partners working to close the participation gap. We succeed when the orgs we power achieve group-level lifts they couldn't measure or target without us.

This v1 report is produced as pro bono public-interest research with no external sponsor at time of publication. Ballot Bridge Initiative, Inc. is in formation as a 501(c)(3); the founder operates pro bono. Should scope later be extended under a funded engagement — state Secretary of State office, foundation strategic-infrastructure grant, academic partnership, or movement-partner co-brand — that relationship will be disclosed here and in the commit history.

Comparable mission-infrastructure model precedents include ProPublica, OpenElections, Pol.is / The Computational Democracy Project, Bridgespan Group, and Results for America — foundation-funded public-interest infrastructure organizations sustained by grants and earned revenue rather than subscription SaaS.

Methodology open-source commitment. The intervention catalog, group-targeting framework, suppression rules, and measurement methodology are published under a permissive license alongside the code. Defensibility comes from integration, delivery, and intelligence — not from closed-source IP.

Quick reference — what to use when

Anything not on this table is either not in this report or requires an explicit methodological disclosure in the section that uses it.

Last updated: 2026-06-22. This methodology page reflects the current analytical state — seven group-level findings + recommendations, with build-time precomputation, so every numeric claim on every finding page traces to deterministic JSON regenerable from the manifest at scripts/precompute/manifest.py. Subsequent refinements (additional CVAP vintages, broader race-disaggregated validation, updated effect-size literature, Phase 1 interactive group-explorer) will be noted in commit history and reflected here as they land.