# Preregistration: Columbia Daily Spectator news-coverage audit

**Frozen:** 2026-08-10 15:35 UTC, before treatment scoring or outcome calculation  
**Collection cutoff:** 2026-08-10 (inclusive through the last page successfully retrieved that day)  
**Study window:** 2023-10-01 through 2026-08-10  
**Status:** Confirmatory design for the available corpus; collection coverage and any deviations will be reported.

## Question and hypotheses

The directional research hypothesis is that, within Spectator news reporting about Columbia institutional conflict and governance, University administrators and trustees receive more favorable or less skeptical treatment than students, workers and unions, protesters, faculty critics, and comparable challengers to institutional power.

The primary estimands are actor-level differences in (a) skeptical/evaluative framing, (b) sourcing and placement, and (c) independent testing of claims. Tone, source balance, framing, and dependence on official material are separate outcomes; none alone will be labeled “bias.”

Competing explanations include topic and event severity; the administration's legal duty or practical ability to respond; deadline and breaking-news constraints; story genre; article length; availability and safety of nonofficial sources; source nonresponse; legal-risk editing; author and leadership period; later updates; and genuine differences in the evidentiary support for competing claims. Contrary and null evidence will be retained.

## Population, collection, and sampling

The target population is all Columbia Daily Spectator news articles published from 2023-10-01 through 2026-08-10 that substantially concern campus protest or Palestine; student discipline; policing or Public Safety; labor or union bargaining; University governance; presidential or trustee decisions; federal pressure; campus access; speech; institutional policy; or a major administrative announcement. University News is the principal section, but qualifying news articles in other news sections are eligible. Opinion, editorials, columns, letters, sponsored content, newsletters, photo-only pages without reported text, event listings, and duplicate URLs are excluded from quantitative analysis.

The collection will attempt a complete relevant corpus using section/archive pages, sitemaps or feeds where permitted, site search, and search-engine discovery. Robots.txt, access controls, and reasonable rate limits will be honored; inaccessible full text will not be bypassed. Discovery method and gaps will be logged.

A neutral comparator will be drawn from University News articles in the same calendar quarters that do not center institutional conflict or an allegation against Columbia. Eligible comparator topics include routine academic programs, awards, facilities, research announcements, appointments, and student-life administration. Comparators must satisfy the same article-type and access rules. They will be sampled without reference to tone or named author.

If near-complete enumeration is infeasible, the following fallback is fixed in advance: create strata by calendar quarter and topic family; within each stratum, sort eligible URLs by publication timestamp and stable URL, assign a deterministic SHA-256 rank using seed text `spectator-audit-2026-08-10`, and select up to 12 conflict/governance articles plus up to 4 neutral-comparator articles. If a stratum contains fewer than the cap, include all. Oversample rare eligible topic families only to a minimum of five articles across the study, assign inverse-probability weights for population estimates, and publish both weighted and unweighted results. Articles discovered only because they were criticized in the input documents enter the sampling frame but receive no automatic inclusion unless needed for a predeclared topic minimum; they may be analyzed separately as case studies.

## Units

The **article unit** is one canonical news URL/version as accessible at cutoff. A live article that was updated remains one article; the latest accessible version is coded, with original publication and update information recorded when available. A materially separate follow-up URL is a separate article. Syndicated or duplicate copies are deduplicated by canonical URL and text hash.

The **actor unit** is an actor category as treated in one article, not each individual mention. Categories are: University administration/trustees; students/student organizations; protesters; workers/unions; faculty critics; government officials; and other institutional/community actors. A person may occupy more than one real-world role, but is coded once per article according to the role relevant to the quoted action. Ambiguous roles are flagged and excluded from pairwise actor-category tests.

## Variables

Article metadata: stable ID, title, subtitle, author(s), date, update/correction status, section, URL, retrieval date, length, article type, topic(s), leadership period, access/coding status, discovery method, and content hash where text was retrieved privately.

Actor-level variables: presence; headline and lede framing; manual evaluative tone (-2 to +2); automated actor-context tone where feasible; attribution verbs; source count/type; first actor/source; direct-quote word count and placement; reliance on University statements/releases; independent testing of official assertions; use of documents, records, filings, data, or prior reporting; affected-person consultation; skepticism applied to administrative and challenger claims; power/history/consequence context; foreseeable identification risk; and corrections/follow-up indicators. Article-level controls include topic, date/period, length, type, and author.

Primary confirmatory outcomes are: (1) manual skepticism score toward each actor (0 none, 1 mild/implicit, 2 explicit testing/context, 3 strong documentary or adversarial testing); (2) actor is quoted/appears first; (3) direct-quote share; (4) independent-testing indicator; and (5) source diversity. The primary contrast is administration/trustees minus pooled challengers (students/organizations, protesters, workers/unions, and faculty critics). Government and other actors are reported separately. Headline/lede tone and attribution-verb markedness are secondary outcomes.

## Planned analysis

Report counts, denominators, means/proportions, and 95% confidence intervals. For unadjusted actor comparisons, use article-clustered bootstrap intervals (10,000 deterministic resamples) and paired within-article contrasts when both administration and challenger actors appear. For binary outcomes, report risk differences and odds ratios; for ordered outcomes, report mean differences plus an ordinal-model sensitivity analysis when cell sizes allow.

If there are at least 100 coded articles and adequate outcome variation, fit multilevel models with topic, log word count, time/leadership period, article type, and actor category as fixed effects and repeated author as a random intercept (or author-clustered robust standard errors if the mixed model fails). Interactions will be limited to actor-by-topic and actor-by-period and labeled exploratory. If sample size or separation makes a model unstable, omit it rather than simplify opportunistically; emphasize stratified estimates.

Multiplicity will be handled by designating the skepticism contrast as primary, reporting all other outcomes transparently, and using Benjamini-Hochberg adjusted q-values for the family of secondary actor comparisons. Statistical significance is not treated as substantive importance. No causal claim about editors or intent will be made from observational associations.

Robustness checks: alternate skepticism dichotomies (>=1 and >=2); weighted versus unweighted fallback sample; conflict-only versus conflict-plus-comparator; excluding breaking-news briefs; excluding articles under 400 words; excluding input-document case studies; leave-one-topic-out estimates; and, where possible, article fixed effects for within-story actor comparisons.

## Manual validation and reliability

At least 15% of the final quantitative corpus, with a minimum target of 30 articles when the corpus permits, will receive full manual audit. The audit sample is stratified by topic, quarter, actor mix, and preliminary automated classification and is selected before comparing automated and manual scores. A second coder will independently code at least 20% of the manual-audit set (target minimum 20 actor units when feasible). Report Cohen's kappa for nominal/binary variables, weighted kappa for ordinal variables, and ICC or Krippendorff's alpha for continuous/count outcomes as appropriate, with raw agreement and denominators. Disagreements are preserved for reliability estimates, then adjudicated for final descriptive coding.

Automated sentiment is validation support, not ground truth. Models can mistake negation, quotations, legal allegations, institutional titles, conflict vocabulary, and criticism reported neutrally; can attribute one actor's words to another; and may encode social or political bias. Actor-window scores will be compared with manual judgments and will not substitute for manual coding when reliability is poor.

## Corrections, versions, and missing data

Corrections and editor's notes are retained as accountability outcomes. The latest accessible version at cutoff is the scored version; earlier versions are described only when independently archived and lawfully accessible. A correction does not erase the original error for qualitative discussion, but corrected text is not scored as if still present. Retrieval timestamps and hashes document the accessed version.

No missing outcome is imputed. `not_applicable`, `not_observed`, `inaccessible`, and `unclear` are distinct values. Denominators will be outcome-specific. Missing actor text excludes that actor unit from tone measures; inaccessible full text excludes an article from treatment scoring but not from the collection ledger. Unknown author or leadership period remains missing and is not guessed.

## Method-change rule

Changes are permitted only for an external collection barrier, a demonstrable coding ambiguity affecting at least 5% of a pilot, a reliability statistic below 0.60, model nonconvergence/separation, or a factual error in this preregistration. Every change must be timestamped in `methodology.md`, state the reason, be made without consulting the affected outcome contrast where possible, and preserve the original specification as a robustness analysis. New analyses prompted by observed results are exploratory.

## Interpretation rule

Evidence will be described as consistent with, mixed on, or contrary to the hypothesis. The study will not claim to have “proved bias.” Ethical evaluation will distinguish accuracy, truth-seeking, harm minimization, independence, and accountability under the complete SPJ Code of Ethics. Examples of strong reporting and evidence that weakens the campaign's case will be included.