Arizona election results, systems and finality

Observed and calculated: Arizona's nine regular cycles contain 45 contests for governor, secretary of state, attorney general, state treasurer and superintendent of public instruction. Republicans won 33 and Democrats won 12. That aggregate advantage is real, but it is uneven. Republicans won every treasurer contest. Attorney general divided five to four. Democrats won three of five offices in 2022. The office and cycle matter as much as the statewide total.

Observed: the administrative system also changed. The 1994 record describes punch-card capture with computerized readers at counting centers. Arizona moved through county-specific optical-scan transitions and, by the 2004 general election, official guidance said every voter used optical scan. In 2022 the official county matrix showed ES&S equipment in 13 counties, Dominion in Maricopa and Unisyn in Yavapai. Arizona did not have one statewide machine vendor.

Inferred with limits: neither of those histories supplies a simple explanation for the timing or closeness of results. The series contains sub-one-point margins in 1990, 2002, 2018 and 2022. Many documented outcomes resolved on election night or the next day, while some close modern contests took longer. Certification intervals follow administrative schedules and are not measures of when the public knew the winner.

Findings

The party summary represents ultimate office control in each regular cycle. The 1990 governor entry needs special handling. Fife Symington led the initial election by 4,293 votes, but no candidate achieved the legally required majority. The election therefore continued to a February 1991 second election, which Symington won by 44,401 votes. Because the same candidate led both rounds, the party-control total is unchanged. The regular-cycle table labels the initial plurality; the second election is treated separately and is not counted as a sixth office or tenth regular cycle.

Three cycles were Republican sweeps: 1994, 2010 and 2014. No cycle was a Democratic sweep. The strongest Democratic cycle was 2022, when Democratic candidates won governor, secretary of state and attorney general. That pattern weakens any description of the period as uniform one-party control, even though the aggregate count strongly favored Republicans.

The close-race cluster is modern, but not exclusively modern

Six regular contests finished below one percentage point. Four occurred in 2018 or 2022, which creates a genuine recent cluster. Two older counterexamples matter: the initial 1990 governor election and the 2002 governor election. Close results therefore predate the current vendor mix, current election-management software and several modern election laws.

The 2022 attorney-general recount is the narrowest result in the series: 280 votes, 0.0111 percentage points, or about 0.108 votes per 1,000 statewide ballots. The 2022 superintendent result was also inside the new automatic-recount threshold. Both recounts confirmed rather than reversed the initial winner. That is evidence about finality and verification, not evidence that the original count was perfect or defective.

Master contest table

Contest participation is candidate votes divided by statewide ballots. Roll-off is statewide ballots minus contest votes. Third-party share is the portion outside Democratic and Republican candidates, including independent, no-party and minor-party votes. An uncontested race has no meaningful runner-up margin, so the table reports it as not applicable rather than inventing a zero-vote opponent.

Arizona's estimated resident population rose from 3,679,056 in 1990 to 7,359,197 in 2022, almost exactly 100.0 percent. The estimated voting-age population rose 115.9 percent. Registered voters rose from 1,858,941 to 4,143,929, or 122.9 percent. Ballots cast rose from 1,094,735 to 2,592,313, or 136.8 percent. These are Census July 1 annual estimates and official election totals, not an Election Day population census.

The endpoint comparison indicates that the voting system served a faster-growing electorate than population alone suggests. It does not explain why. Migration, age structure, registration rules, administrative cleanup, political mobilization and election salience are all plausible contributors. The 1998 to 2002 registration decline while population increased is a counterexample to treating registration as a smooth proxy for population.

Turnout among registered voters fell to 45.82 percent in 1998, recovered to 60.47 percent in 2006, fell again to 47.52 percent in 2014 and reached the series high of 64.85 percent in 2018. The adjacent 2014 to 2018 increase is 17.33 percentage points. A nine-cycle sequence is too short and institutionally varied to support a claim of a stable long-run turnout trend.

Registration as a share of voting-age population moved from 69.55 percent in 1990 to 71.82 percent in 2022. The intervening low was 56.17 percent in 2002. This measure is deliberately labeled as a VAP rate. It includes noncitizens and other residents who may not be eligible, so it is not an eligible-voter registration rate.

Why CVAP is not silently inserted

Citizen voting-age population is closer to an eligibility denominator than total VAP, but a comparable annual state series does not span the full report. The Census annual special-tabulation sequence begins with the 2005 to 2009 American Community Survey five-year product, while the earlier product is Census 2000. Combining them into a seamless 1990 to 2022 line would hide survey and product changes. CVAP is therefore unknown in the main series. A separate ACS sensitivity analysis is a legitimate future extension.

Roll-off and participation vary by office

Governor has the highest median contest participation at 98.08 percent and the lowest median roll-off at 1.92 percent. Treasurer has the lowest median participation at 92.87 percent and the highest roll-off at 7.13 percent. This is consistent with ballot position and voter attention, but the retained data do not isolate those mechanisms. The same pattern also interacts with uncontested or weakly contested races.

Third-party share is episodic rather than uniform. The median is zero for three offices because many cycles have only two material party nominees and small write-in totals. The 2002 governor contest is the largest among the close-race set at 8.59 percent outside the two major parties. A third-party share can change the two-party geometry without proving which major candidate would have received those votes in its absence.

The phrase "how long it took to decide" has at least four meanings: a news projection, a candidate concession, the official canvass or certification, and completion of a mandatory recount or court process. Treating those as one timestamp creates false comparisons. This report keeps them separate.

A defensible public-resolution date was located for 20 of 45 regular contests. Fourteen resolved on election day or the next calendar day. Two took more than seven days: the 2018 secretary-of-state contest, where the valid milestone was Steve Gaynor's concession ten days after Election Day, and the 2022 superintendent contest, resolved politically after nine days. The invalid election-night 2018 projection for Gaynor is expressly excluded.

Twenty-five regular contests do not have a sourced call or concession date in the retained evidence. That missingness is not random. Older down-ballot races are less likely to have accessible contemporaneous digital reporting. Comparing only the 20 known rows may therefore overstate or understate historical speed. The correct result for the other 25 is unknown.

The official schedule displays a separate pattern. The retained canvass dates are 20 days after Election Day for the regular 1990 through 2002 rows and 27 days after Election Day for 2006 through 2022. That shift describes the official process in this dataset. It does not prove that vote counting or public knowledge became seven days slower.

Reporting technology is not tabulation technology

A ballot can be paper, read by an optical scanner, aggregated through an election-management system, transmitted through a county reporting workflow and displayed on a public website. Each is a different layer. The evidence is incomplete on historical election-night transmission, so this report does not infer internet connectivity from the word electronic and does not equate faster public reporting with faster counting.

Older elections also had computerized components. The 1994 procedures describe readers or scanners, data processing, accuracy certification and report generation. Calling that period purely analog would be inaccurate. The more defensible distinction is between ballot-capture methods, tabulation location, software support, network or transmission pathways, audit controls and public reporting.

Arizona administers elections through counties, so the correct unit is often the county and election cycle, not the state. The evidence does not support assigning one machine, one vendor or one configuration to every county in a year. Exact 1990 and 1991 equipment remains unknown. The 1998 and 2002 statewide county matrices are partial. Those limits are part of the result.

From punch cards to optical scan

The 1994 official manual establishes a punch-card framework with computerized card readers or scanners at counting centers. It includes data-processing, accuracy-certification and procedural control boards. This was a paper or card voter interface combined with computerized tabulation. Maricopa County's documented move to ES&S optical scan in 1996 is one county transition, not proof of immediate statewide conversion.

The 1998 and 2002 elections fall inside a mixed county transition. A state-commissioned 2004 assessment later described nine punch-card counties moving to Diebold optical scan and six counties that had already selected Diebold or ES&S systems. The assessment helps explain the transition but cannot be back-assigned as an exact November 2002 deployment matrix. Official 2004 voter guidance then stated that every Arizona voter in that general election used optical scan.

Later systems kept paper ballots while adding accessible voter-facing devices, election-management software, central-count equipment and vendor support. Official guides list accessible Premier, ES&S, Sequoia and Unisyn devices across different counties and years. Those devices served accessibility requirements; they are not automatically the scanners used for all paper ballots.

The 2022 county matrix

The official 2022 equipment list provides the clearest complete modern snapshot. Thirteen counties listed ES&S equipment, Maricopa listed Dominion and Yavapai listed Unisyn. The same matrix names models, firmware, election-management software, central-count components and maintenance providers. The number 13 describes county presence, not 13 identical installations.

Controls also changed

The official record includes logic-and-accuracy procedures, software escrow, physical controls, hand-count audits and post-election reconciliation. Arizona's statutory hand-count process was in place for the 2006 included cycle, and official hand-count result pages document later use. A control can detect some classes of error without proving that every possible failure is impossible.

The interaction of paper ballots and electronic tabulation matters. Paper creates a physical record that can be sampled or recounted. Electronic tabulation makes high-volume counting practical. Election-management software configures contests and aggregates results. The evidence supports describing this as a layered sociotechnical system, not a choice between paper and computers.

The law timeline covers access, registration, ballot return, signature cure, audits, recounts and office structure. The clean analytical distinction is between a plausible administrative pathway and a measured outcome effect. The sources establish the former much more often than the latter.

Changes capable of affecting valid-ballot totals

Registration eligibility, documentary requirements, ballot-return restrictions and signature-cure procedures can affect whether a person registers, receives a ballot, returns it through a permitted channel or cures a questioned signature. Those are plausible mechanisms that can affect valid-ballot totals. The historical aggregates do not identify how many net ballots each rule added or removed, nor which candidate would have received them.

Changes that mainly affect verification or finality

Audit and recount rules act later in the process. Senate Bill 1008 is the strongest direct example. By raising the automatic-recount threshold to one-half of one percent, it caused the 2022 attorney-general and superintendent contests to proceed through recount finality. The recount adjusted totals but confirmed both winners. The law explains the existence of the recount, not the closeness of the underlying election.

Explicit non-effects

Three changes are excluded from causal discussion in the included results. Senate Bill 1485's early-voting-list removal cycle had not matured by November 2022. House Bill 2492 became effective in January 2023. Proposition 131 first changes the ticket in 2026. Chronology prevents all three from explaining a certified result through 2022.

Supported observations

• Republicans won a majority of the included offices, with substantial variation by office and cycle.
• Close contests occurred in multiple periods, and 2022 contains the largest cluster.
• Registered voters and ballots grew faster than resident population between the endpoints.
• Registered-voter turnout moved sharply between cycles rather than following a steady trend.
• Arizona's equipment and vendor environment was county-specific and changed over time.
• Public calls, official canvasses and recount finality measure different stages.
• Election law changed access, processing, verification and finality pathways.

Defensible inferences

The administration burden grew by more than population alone because ballots increased 136.8 percent while population increased 100.0 percent. Down-ballot roll-off differs consistently enough by office to justify office-specific comparisons. Mandatory recount law can extend official finality when margins qualify. These are restrained inferences rooted in observed counts and explicit procedures.

Unsupported conclusions

The series does not show that electronic tabulation caused a party to win, that internet use explains a margin, that commercial vendors controlled statewide outcomes, that a legal change caused turnout to rise or fall, or that a longer count indicates fraud or incompetence. It also does not prove the opposite of each claim. The necessary causal, security or forensic evidence is outside this historical aggregate dataset.

No statistical test can repair missing design. With nine regular cycles, overlapping law and technology changes, county heterogeneity and nonrandom missing call times, a regression would produce precise-looking coefficients without credible identification. The report intentionally stops at descriptive statistics, sensitivity checks and counterexamples.

For a public-sector topic, "where to play" does not mean partisan opportunity. It means the records, transparency and operational services that would let administrators, journalists, researchers and the public answer routine questions without reconstructing decades of incompatible archives.

Ground that is already crowded

Where demand is real and supply is thin

1. Versioned historical result data. Preserve certified totals, recount revisions, source scans and machine-readable fields together. The 2006 historical-table typo demonstrates why a source hierarchy and conflict ledger matter.
2. Public-resolution timelines. Store projections, concessions, canvasses, recounts and court events as separate timestamps. Twenty-five of 45 regular contests remain without a defensible public-resolution date.
3. County equipment history. Publish election-specific vendor, model, software, maintenance and count-location records rather than only current certification lists. The 1998 and 2002 matrices remain partial.
4. Contractor role registers. Distinguish printing, mailing, programming, maintenance, poll books, audit support and reporting. A vendor name without a role creates more heat than insight.
5. Comparable denominator packages. Release registration cutoff, status definition, ballots, contest votes, VAP and CVAP with source vintage and formulas. This prevents unlike rates from being compared.
6. Audit and recount data standards. Publish machine-readable pre-audit totals, sample selection, discrepancies, escalation and final totals. This would support verification without converting a control into a slogan.

The report began with a fixed scope and one structured row per contest. Results were extracted from official canvasses, then checked mechanically for candidate sums, margins and turnout arithmetic. Separate ledgers covered population denominators, public decision timing, systems, vendors and law. A slow verification phase reopened primary sources and visually matched the material claims. Citation loop 1 logged 52 checks with no failures.

Definitions

Contest votes are the sum of official candidate votes in an office. Ballots cast are the statewide election-level total. Contest participation divides contest votes by ballots. Roll-off subtracts contest votes from ballots. Third-party share is the vote outside Democratic and Republican candidates. Margin points divide the winner-runner-up vote difference by contest votes. Registered turnout divides ballots by registered voters.

The 1991 second election is an institutional outlier and is not inserted into regular-cycle trends. An uncontested race has no manufactured runner-up. Public-resolution dates remain null when no defensible contemporaneous call or concession was found. Census population values are July 1 estimates from named vintages, not Election Day counts.

Confidence notes

Certified contest totals and the 1994, 1998, 2002, 2006, 2010, 2014, 2018 and 2022 election metrics were visually matched to official canvasses. The 2022 recount summary was also visually checked. The 1990 normalized registration denominator differs from the canvass registration count, producing 58.89 percent rather than 58.86 percent turnout; both are disclosed. The certified 2006 canvass controls over an inconsistent current summary page.

Equipment confidence is uneven. Exact 1990 and 1991 configurations are unknown. The 1998 and 2002 matrices are partial. Accessible-device lists from 2006 through 2018 do not identify every paper-ballot scanner. The 2022 county matrix is complete for the fields the state published. Gartner authored the 2004 state-commissioned assessment, so its role and date are disclosed.

Twenty-five regular contests lack a defensible public call or concession date. Older digital news archives are incomplete, which may bias known timing toward recent and higher-profile contests. Exact clock-time comparisons are avoided when sources provide only dates.

Citizen voting-age population is not included in the main series because a uniform annual 1990 to 2022 product is unavailable. The Census-vintage sensitivity check changes the 2010 VAP registration rate by 0.07 points. Legal mechanisms are described, but no law is credited with causing a winner, party, margin or turnout change.

DO-NOT-PRINT exclusions applied

• No canvass date is relabeled as an election call.
• The invalid 2018 secretary-of-state election-night projection is excluded.
• No later equipment inventory is back-cast to an earlier election.
• No statewide vendor is inferred from one county or accessible-device list.
• No 2026 result is included as final.
• The 2006 summary-page ballot typo is not used.
• No fraud, suppression, manipulation, motive or causal outcome claim is inferred from correlation.