Public trust standard
Research Methodology Standard
This standard governs public datasets, research reports, quantitative findings, charts, and election analysis. A specific report may use a narrower method, but it should disclose any material departure from these principles.
Research question and scope
Each research release should begin with a defined question, geography, time period, unit of analysis, and intended use. The method should be chosen before interpreting the result. We avoid retrofitting a method merely because it produces a more interesting political conclusion.
A report should identify whether it describes official results, survey evidence, campaign operations, public digital activity, a forecast, or an illustrative model. These categories are not interchangeable.
Source provenance
Primary official records are preferred for election administration and results. Every downloadable dataset should provide, at row or table level where practical, the source URL, issuing body, document or table identifier, retrieval date, and status such as preliminary, revised, or final.
We preserve the source’s definitions and units. When a source changes, disappears, or is replaced, we record that change instead of presenting the later version as if it were the original input.
Collection and transformation
Research notes should describe how records were collected, filtered, joined, recoded, aggregated, or excluded. Derived fields must be distinguishable from source fields. Percentages should disclose their denominator, and totals should state whether postal ballots, uncontested seats, missing stations, or other material categories are included.
Where manual transcription or optical character recognition is used, a second check should compare the extracted values with the source document. Automated extraction is a workflow aid, not independent verification.
Surveys, estimates, and models
A survey-based claim should disclose the population, sample size, field dates, sampling approach, collection mode, weighting, question wording where material, sponsor, and known sources of bias. A model or forecast should disclose its inputs, assumptions, specification, validation approach, and an uncertainty range when one can be estimated responsibly.
Model outputs are estimates, not official results. Small differences should not be described as decisive when they fall within the method’s uncertainty or rest on incomplete data.
Reproducibility package
For a substantial public data release, we aim to publish a versioned data dictionary, machine-readable files, source list, transformation notes, limitations, citation guidance, and analysis code or a reproducible workbook when licensing and confidentiality allow.
Files with incompatible schemas are separated into named configurations or tables rather than combined in a way that breaks public viewers. Published files receive stable version labels so later corrections can be traced.
Privacy, consent, and research ethics
Public research should use aggregated or legitimately public information wherever possible. We do not publish private voter contact details, confidential client strategy, or personally identifying field records merely to make a dataset appear more complete.
When research involves interviews, surveys, or non-public operational data, the release should describe consent, confidentiality, and any restrictions that materially affect interpretation. Protected characteristics should not be inferred or presented as fact without a lawful, ethical, and well-supported basis.
Limitations, review, and corrections
Every report should place material limitations near its findings, not hide them in a remote footnote. The author and reviewer should test headline claims against the tables and source records before release.
Confirmed errors are handled under our Editorial Standards and Corrections. Corrected datasets receive a new version or documented revision; previously published values are not silently overwritten when the change affects interpretation.