Survey methodology

Stealth Analytics designs and fields political surveys for campaigns, caucuses, and ballot committees. This page is the public methods statement we stand behind for that work. It describes the sampling frame, modes, weighting, and disclosure standards we apply on every release, so clients, journalists, and peer researchers can evaluate our work on the merits.

It covers the research side only. The voter file work - target universes, voter scores, walk and phone lists, chase and curing, signature verification, mail files and appends - runs on its own standards and is not documented here.

Sampling frame & population

Our target population is defined per study, typically likely voters or registered voters within a specified geography and primary or general electorate. Samples are drawn from the state voter file with vote-history filters appropriate to the race (for example, registered Republicans with at least one of the last two GOP primary participations for a Republican primary instrument). The frame, eligibility filters, and the fielded sample size are disclosed in every release.

Survey modes

We field as a mixed-mode probability sample combining SMS-to-web, live caller, and IVR where appropriate. Blending modes widens reach across age, region, and device preference, balances coverage of voters who would otherwise self-select out of any single channel, and reduces the single-channel bias that has compromised polling accuracy in recent cycles. The mode mix actually achieved is reported with each release.

Weighting

Responses are weighted using iterative proportional fitting (raking), a standard technique used by leading polling organizations. Weighting variables typically include gender, age range, region, and education, calibrated to the target population derived from the voter file and Census benchmarks.

Weights are bounded between 0.2 and 4.0 to prevent any single respondent from exerting excessive influence on the weighted estimates.

Precision & margin of error

We report the margin of error at the 95% confidence level, computed against the effective sample size after accounting for the design effect (DEFF) introduced by weighting. Every release also discloses the effective sample size (Kish), the weighting efficiency, and the overall MOE, so consumers can judge the precision of subgroup estimates rather than relying on the unweighted n alone.

Quality controls

• Raking is run to a strict convergence threshold with a hard cap on iterations; non-converging runs are flagged for review rather than published.
• Weight-bound saturation is monitored. If too many respondents hit the floor or ceiling, the targets and frame are reviewed before the topline is released.
• Response and cooperation rates (AAPOR RR3 / COOP3) are computed for each release and disclosed in the technical appendix.

AAPOR transparency

We adhere to the disclosure standards of the AAPOR Transparency Initiative. With every published release, we provide the population definition, sampling frame, sampling method, mode mix, field dates, weighting variables and method, weight bounds, response metrics, and overall margin of error. If you cannot find a disclosure item you need, email us and we will provide it.

The NavX™ pipeline

Our deliverables (the topline document, thermal crosstabs, and the index deck) are produced by NavX™, our proprietary pipeline. NavX™ ingests the raw response file the moment fielding closes, applies the weighting plan disclosed above, computes statistical significance per cell (p<0.05) for the crosstabs, and renders the executive deck and topline in hours rather than days. Every deliverable ships with a methods note that mirrors this page, scoped to that specific study’s frame, modes, and weights.

Updates to this page

Methodology evolves. When we change something material (a weighting variable, a frame definition, a disclosure standard), we revise the “Last updated” date at the top of this page and note the change in the next release that uses it.

Questions

Methods questions, replication requests, or transparency disclosures? Email admin@stealthanalytics.net or use the contact form.

Election data: the ABEV tracker and the precinct maps

The early ballot tracker is built from the counties’ own ABEV files, read through a per-county adapter and aggregated to county, congressional district, legislative district and precinct. District registration is transcribed from the Secretary of State’s published figures rather than apportioned from county totals, because an apportioned denominator quietly moves every rate built on it.

Nothing publishes unless it reconciles. Each rebuild runs an audit that checks identity, that returned never exceeds issued and issued never exceeds registered, that district registration reconciles to the state source, that the published file agrees with what the engine produced, and that the roll-ups agree with their own parts. A failure aborts the run, so the site keeps yesterday’s numbers rather than showing numbers that do not add up.

Where sources disagree, precedence is fixed and not negotiable per case: the certified state canvass outranks a county statement of votes cast, which outranks a live feed. A gap is left open and labelled rather than closed by fitting a number to the feed.

The precinct result maps are assembled from certified county canvasses and reconciled against the Secretary of State’s post-canvass results feeds. Comparisons against 2022 are a turnout-rate difference in percentage points, each cycle measured against its own registration and aligned on days before election day rather than calendar date - calendars do not line up between cycles, election days do. A missing denominator on either side prints as a dash, never as a substitute.