19 — Statistical Inference from AV Results

A natural question follows every Actual Vote analysis: when the audit finds nothing wrong, what does that tell us about the places we didn’t audit? America Counts answers that question in a dedicated technical paper, What Can We Infer? A Framework for Statistical Reasoning About Actual Vote analysis Results (working draft v3, June 2026). This chapter summarizes its conclusions. The full paper — with all derivations, tables, and worked examples from the 2024 North Carolina Primary analysis — is a working draft that isn’t yet ready for publication. We’re happy to discuss the ideas on an individual basis: contact us.

19.1 — The Honest Framework, in Brief

For audited precincts, no statistics are needed. Every comparison is direct arithmetic: the tape said X, the official report said Y, and they match or they don’t. This is the bedrock claim, and it involves no inference at all.

Two exact conclusions need no extrapolation either. Where coverage includes every precinct of a contest, that contest’s reporting is fully verified — each analysis reports its count of fully verified contests. And where a winner’s verified margin within audited precincts exceeds the total ballots cast in unaudited precincts, the outcome is reporting-robust no matter what the unaudited reports contain. These are arithmetic certifications, not statistical estimates, and they feed directly into the materiality analysis in the Legal chapter.

For unaudited places, AV is deliberately modest. Volunteer coverage is not a random sample, so AV does not offer poll-style margins of error and does not compute a statistical risk limit. The paper shows exactly how much — and how little — a clean audit shifts the evidence under a range of openly stated assumptions: in the NC Primary example (zero discrepancies across 52 precinct-equivalents), the evidence that per-precinct discrepancy rates are low is moderate-to-strong, but its precise weight depends on one’s prior, and the paper says so plainly.

Even missing tapes are measurements. Tapes not posted, not legible, or not retained are observations of compliance with posting and retention law — including the federal requirement that election records be preserved for 22 months (52 U.S.C. § 20701). Missingness patterns also inform where future audits matter most.

Future deployments can support much stronger claims. The paper’s design-of-experiments analysis shows that breadth beats depth — covering one precinct in each of 30 counties teaches far more about a state than 30 precincts in one county — and gives concrete volunteer targets (roughly 80 well-placed volunteers for 80% power against a 2% statewide problem; about 200 for a genuine statewide audit). It also describes how AV can measure its own error rates with seeded-error proficiency testing, how official results change between first posting and certification (and why documenting that baseline counters misinformation), and exactly what to report and request when a discrepancy survives investigation.

The discipline throughout is the point: claiming exactly as much as the data supports — and no more — is the foundation of AV’s credibility, in public and in court.