Sources & References
Every empirical claim on this site traces to one of the sources below. Where the original data is publicly available, the link goes to it. Where the source is a peer-reviewed paper, we cite the canonical reference.
Each finding page carries numbered citation markers — [1], [2], etc. — that map directly to the entries below.
Primary microdata
[1] U.S. Census Bureau, Current Population Survey (CPS), November Voting and Registration Supplement, 2000–2024. IPUMS-CPS extract cps_00002. 1,641,940 raw rows; 1,195,957 citizen-adult observations with valid voter-supplement weights. cps.ipums.org
Turnout benchmarks
[2] Michael P. McDonald, United States Elections Project — Voting-Eligible Population (VEP) Turnout Rates, 1980–2022 v1.2. October 2024. election.lab.ufl.edu
[3] U.S. Election Assistance Commission, Election Administration and Voting Survey (EAVS), Sections A / C / D / F + Policy Survey, May 2025 time-series release. 2004–2022 jurisdiction-level data. Scope: 61,262 county/jurisdiction rows × 125 variables per section (DATA_INVENTORY.md §EAVS). eac.gov/research-and-data/eavs-retrospective
[4] MIT Election Data and Science Lab (MEDSL), state-level presidential / senate / house returns, 1976–2024. github.com/MEDSL
[5] U.S. Census Bureau, American Community Survey Citizen Voting-Age Population (CVAP) Special Tabulation 2019–2023. State × race/ethnicity. Scope: 677 state-level rows × 13 race/ethnicity categories (DATA_INVENTORY.md §ACS CVAP). census.gov/data/datasets/2023/dec/rdo/2019-2023-CVAP.html
State policy layer
[6] National Conference of State Legislatures (NCSL), state election policy taxonomy — automatic voter registration, same-day registration, online voter registration, pre-registration at 16/17, no-excuse absentee, universal vote-by-mail. Current as of April 2026. Scope: 155 state-policy rows across 6 policy dimensions (DATA_INVENTORY.md §NCSL). ncsl.org/elections-and-campaigns
Institutional structure
[7] State congressional redistricting method taxonomy, derived from NCSL Redistricting Commissions page and Ballotpedia. Independent commission / advisory commission / backup commission / legislative-nonpartisan-staff / legislature classifications across all 50 states + DC. Scope: 51 state × redistricting-method classifications (DATA_INVENTORY.md §Redistricting).
[8] Commission-adoption event dataset, 1968–2021. Compiled from Ballotpedia ballot initiatives, NCSL records, and state constitutional histories. 13 commission-adoption events spanning the analytical window.
Attitudinal complements
[9] American National Election Studies (ANES), Time Series Study — external efficacy, political trust, political interest series, 1960–2024. electionstudies.org
[10] Harvard Institute of Politics (IOP), Youth Poll — pre-election waves, 2016–2024. 18–29 cohort intent, trust, and country-direction measures. iop.harvard.edu/youth-poll
[11] Pew Research Center, political engagement cross-tabs, 2020–2023. Voter-file-validated turnout by age and race; always-vote habit measures. pewresearch.org/politics
Standard-error methodology
[12] U.S. Census Bureau, CPS Generalized Variance Parameters, November 2022 Supplement, Technical Documentation Tables 8–11. Used in place of replicate weights (not published for the November supplement). Scope: 51 state parameter rows (50 states + DC) (DATA_INVENTORY.md §GVF). census.gov/programs-surveys/cps/techdocs
Intervention effect-size catalog
[13] Donald P. Green & Alan S. Gerber, Get Out the Vote: How to Increase Voter Turnout, 4th ed. Brookings Institution Press, 2019. Foundational meta-analysis of GOTV interventions across canvassing, phone banking, mail, digital, and peer-to-peer modalities.
[14] Brennan Center for Justice, cost-benefit analyses of election reforms — automatic voter registration, same-day registration, mail-ballot expansion, polling-place access. brennancenter.org
[15] Results for America, evidence-based policy infrastructure for voter participation. results4america.org
Methodology adjustments
[16] Aram Hur & Christopher H. Achen (2013). "Coding Voter Turnout Responses in the Current Population Survey." Public Opinion Quarterly, 77(4), 985–993. CPS over-reporting adjustment baseline.
[17] Stephen Ansolabehere, Bernard L. Fraga, & Brian F. Schaffner (2022). "The Current Population Survey Overstates Turnout, but Race-Specific Estimates Are Robust." Caveat on Hur-Achen weighting for groups disaggregated by race.
Mobile- and internet-voting research (background)
Cited where the site discusses mobile or internet voting as a candidate intervention.
[18] Mihkel Solvak & Kristjan Vassil (2018). E-voting in Estonia: Technological Diffusion and Other Developments Over Ten Years (2005–2015). Republic of Estonia. Foundational empirical evidence on Estonia's i-voting system.
[19] Nicole Goodman & Leah C. Stokes (2020). "Reducing the Cost of Voting: An Evaluation of Internet Voting's Effect on Turnout." British Journal of Political Science, 50(3), 1155–1167.
[20] Daniel Stockemer (2024). Reanalysis of Goodman & Stokes (2020). Policy & Internet. Methodological reanalysis finding the Goodman-Stokes lift effect was largely a novelty effect plus spatial non-randomness.
[21] Michael A. Specter, James Koppel, & Daniel Weitzner (2020). "The Ballot is Busted Before the Blockchain: A Security Analysis of Voatz, the First Internet Voting Application Used in U.S. Federal Elections." USENIX Security Symposium.
[22] Andrew W. Appel et al., Princeton Center for Information Technology Policy (CITP), 2025–2026. Critique of Tusk Philanthropies' VoteSecure SDK, including the January 2026 statement: "Internet voting is insecure and should not be used in public elections."
Analytical samples used per finding
Each finding's central numbers come from the pooled CPS Voter Supplement microdata [1]. The specific analytical samples that drive the visible numbers on each finding page:
| Finding | Analytical sample size (n) | Source file in this repo |
|---|---|---|
| 1. The persistent gap | Per-cycle: youth (18–29) n = 12,303–18,524; seniors (65+) n = 14,759–21,642 across 13 federal cycles | src/data/findings/finding-01/gap_by_year.json |
| 2. The midterm amplifier | Same per-cycle samples as Finding 1, split by presidential vs midterm cycle | src/data/findings/finding-02/by_cycle.json |
| 3. Who didn't register, and why | n = 47,011 pooled youth non-registrants — access barrier 20,556; engagement 10,300; other 9,889; personal 6,266 | src/data/findings/finding-03/reason_national.json |
| 4. Registered, didn't vote, and why | n = 34,031 pooled registered youth non-voters — logistical 17,216; engagement 7,769; other 4,795; access 2,992; personal 1,259 | src/data/findings/finding-04/reason_national.json |
| 5. How young voters actually vote | Youth voters' voting-method shares by state regime + cell; cell-level n in source JSON (range ~480–12,236 by race × gender pooled) | src/data/findings/finding-05/method_cell.json, …/state_regime.json |
| 6. State policy as lever | State × year × youth panel; per-state-year n ranges 93 (ME 2020) to 1,326 (CA 2020); suppression threshold n < 75 | src/data/findings/finding-06/policy_cross.json |
| 7. Institutional structure | 51 states × 13 federal cycles aggregated by redistricting method (state-level analysis; no per-respondent n) | src/data/findings/finding-07/method_youth.json |
| Supplement — efficacy (Does my vote matter?) | ANES [9] national-level efficacy / trust series. Race-disaggregated youth cells too small for stable per-cell rates (n = 56 Black youth 2020; n = 79 Hispanic youth 2020 — see methodology page) | src/finding-efficacy.md |
Cell-level samples (race × gender × age) are exposed inline on each finding page via the provenance chip ⓘ — hover or focus reveals n, generalized-variance-bounded SE, year range, and suppression status. Representative examples from the precomputed cell JSONs:
- Finding 4 — Black female registered non-voters: n = 1,924 · White male registered non-voters: n = 11,958 · Hispanic male registered non-voters: n = 1,987 (pooled 2000–2024)
- Finding 5 — Black male youth voters: n = 1,562 · White female youth voters: n = 12,236 · Asian female youth voters: n = 480 (pooled across all methods + years)
- Finding 6 — CA 2020 youth respondents: n = 1,326 · MA 2024: n = 267 · ME 2020: n = 93 (cells below n = 75 are suppressed)
Sources for cell-level n: src/data/findings/finding-04/reason_cell.json, finding-05/method_cell.json, finding-06/policy_cross.json. The SQL that produces these is in scripts/precompute/manifest.py.
A note on external sample sizes
References [2], [4], and [9]–[22] point to external data products and peer-reviewed papers (McDonald VEP, MIT MEDSL, ANES, IOP, Pew, Green & Gerber, Brennan, etc.). Their sample sizes are documented in each source's own technical materials — codebooks, methodology statements, or the papers themselves. We don't restate those n's here because we don't carry the per-respondent microdata for those sources in this repo (we use them as aggregates or as cited findings). For verifiable n, follow the link to the source.
The one exception we do document: ANES race-disaggregated youth cells (n = 56–79 for 2020 Black/Hispanic youth) are noted on the methodology page because we explicitly tested ANES as a per-cell cross-validator of CPS and disclosed why it doesn't work at that disaggregation.
Data provenance and reproducibility
The full data inventory — every acquired asset, its vintage, scope, known limitations, and analytical role — is maintained internally in DATA_INVENTORY.md and is available on request. Pipeline and analysis code are similarly maintained internally; methodology decisions are fully documented on the methodology page.
For methodological choices — the CPS coding decision, the Hur-Achen adjustment, the generalized-variance approach, the group-level cross-validation work that did and didn't pan out — see the methodology page.