Does Local News Stay Local?: Online Content Shifts in Sinclair-Acquired Stations

AI summary

Overview

Research area: Computational social science / natural language processing applied to media studies, specifically computational analysis of local news content on YouTube.

Technical level: Intermediate. The paper is written accessibly, but interpreting it requires familiarity with topic models (LDA/STM), log-odds ratios, and word embeddings.

Scope: A text-analysis study comparing YouTube transcripts from 8 local news stations before and after their acquisition by the Sinclair Broadcast Group, using Fox News and CNN as national comparison outlets.

What This Paper Is About

Local news stations are widely seen as more trusted and less partisan than national outlets, but the Sinclair Broadcast Group has acquired many of them. This paper asks whether a station's online content changes once Sinclair buys it, and whether that content becomes more national and more politicized. The authors answer this by comparing the same stations before and after purchase, and by comparing them against CNN and Fox News.

Key Contributions

1. A new data source for studying media ownership: The authors build a dataset of automated YouTube closed captions from 8 geographically diverse Sinclair-acquired local news stations plus CNN and Fox News, covering sixteen years of publishing. They argue this is distinctive because prior work focused on broadcasts or station websites, while Americans increasingly consume digital local news.

2. A within-station before/after design combined with cross-outlet comparison: Because they collected data spanning each purchase date, they can compare the same station to itself and also compare Sinclair-affiliated stations against non-affiliated stations and national outlets at a given point in time.

3. A multi-method text analysis: They combine Fightin' Words log-odds ratios, Structured Topic Models (STM) with covariates, and Word2Vec embedding analyses to examine both word-level and topic-level shifts.

4. A tightly controlled paired analysis (appendix): Two Sinclair-acquired stations are compared with two nearby never-Sinclair stations in the same regions, with matched subsampling of CNN and Fox data for the same time periods.

Main Findings

• Shift from local to national and political topics: Across all text analysis methods, the authors find consistent evidence that acquisition by Sinclair is associated with increased coverage of national and political news, often at the expense of conventional local topics such as cooking or local sports.

• Fightin' Words (Table 2): Sinclair-purchased stations used words such as "president," "government," and "federal," while non-Sinclair/pre-purchase stations used local words such as "snow," "downtown," and "school." The authors note that Sinclair-owned stations in 2015 were more likely to discuss the national election than non-Sinclair stations were in the election year of 2016, which they say means purchase timing alone cannot fully explain the shift.

• Topic model results (Figure 2): Topics most prevalent on stations not Sinclair-affiliated included local community topics such as school, family, local events, weather, animals, health, football, other sports, and cooking. After acquisition, topics shifted toward presidential candidates including Clinton and Trump, the FBI, ISIS, and terrorism. Some local coverage remained after purchase (family/church, community, city/mayor information, football).

• Topics that did not track the purchase: Police and crime were prevalent both before and after purchase and were particularly aligned with CNN. Weather discourse appeared both before and after acquisition.

• Smaller polarization signal along the CNN/Fox axis: The authors report that the difference in topic proportion is less pronounced on the CNN/Fox axis, and that topics on pre-purchase stations tend not to align clearly with either CNN or Fox. Variation on the CNN/Fox axis is more visible after Sinclair purchase, for example Israel-Hamas conflict reported more by CNN, and Trump, ISIS, Cuba, immigration, and presidential candidates reported more by Fox.

• Nearest-neighbor embeddings (Table 3): After purchase, nearest neighbors to query words became more politically charged. Before purchase, "bias" was associated mostly with cooking terms ("chiffonade," "mince"), plausibly from the phrase "cutting on the bias" in cooking videos; after purchase it was associated with "implicit," "prejudice," and "racism." Before purchase, "white" and "black" were associated with colors and patterns ("red," "stripes") or items that might be that color ("chardonnay," "roses"); after purchase, "black" was associated with "africanamerican," "racism," and "movement," and "white" was associated with "supremacist"/"supremacy" as well as presidency-related terms ("house," "obama").

• Time confounds acknowledged in the embeddings: The post-acquisition model associated "climate" with "emissions" and "pollution" rather than "growth" and "economy," which the authors say may reflect increased discussion of global warming or simply increased discussion of climate change in recent years. Similarly, pre-acquisition "equality" neighbors suggested discussion of Obergefell v. Hodges (2015), while post-acquisition neighbors suggested reproductive justice terms ("prolife," "unborn"), possibly reflecting Dobbs v. Jackson Women's Health Organization (2022).

• Embedding similarity to national outlets (Figure 3): Using data restricted to 2014–2016 and aligned with a Procrustes transformation, the authors found a general trend of increasing similarity between the local station embeddings and both Fox News and CNN after purchase. Only three words ("freedom," "welfare," and "bias") decreased noticeably in similarity to one national broadcaster, and each increased in similarity to the other. They observed no shift toward either national outlet in particular.

• No clear right-wing slant detected: Unlike prior work, the authors state their results do not consistently show Sinclair ownership is associated with more similarity to Fox than CNN.

• Paired analysis (appendix): Topic modeling on the matched station pairs showed the prominence of pandemic coverage, which dominated topic proportions for dates after purchase and was disproportionately covered by Sinclair-owned channels. On data before 2020, the same shift away from local coverage for Sinclair-owned stations appeared, especially in topics covering football and pets.

Methodology in Plain English

The authors identified Sinclair-acquired stations from a public list, keeping only those that had a YouTube channel and started posting before purchase. They downloaded YouTube closed captions for all videos on each channel and cleaned them by lowercasing and removing non-speech tokens such as "[music]," "[applause]," and "uh."

Three analysis strategies were used:

1. Word-level comparison (Fightin' Words): A log-odds ratio with a Dirichlet prior identifies words overrepresented before versus after purchase. Words appearing fewer than ten times in every station's transcripts were filtered out, and the comparison was stratified by year (2014, 2015, 2016) to limit the confound of general news change over time.

2. Topic modeling (STM): The Structured Topic Model, an extension of LDA that incorporates document metadata as covariates, was used to look at clusters of co-occurring words. Five models were trained: four used news affiliation (Before Sinclair purchase, After Sinclair purchase, CNN, Fox) and date as covariates, differing in the data subset (all dates, 2014, 2015, 2016 respectively); a fifth used Sinclair-affiliated versus non-Sinclair-affiliated as a topical content covariate to see how language within a topic differs. The authors removed the 1% most sparse and common words and used 30 topics, which they found to be the most coherent.

3. Word embeddings: Separate Word2Vec models were trained on before-purchase and after-purchase transcripts, plus models for CNN and Fox News. They used window=50, min_count=10, seed=42, workers=16, and vector_size=100. For a curated keyword list, they examined the 10 nearest neighbors by cosine similarity before versus after purchase. For the similarity analysis, they restricted data to 2014–2016, aligned embedding spaces and vocabularies with the Procrustes transformation, and computed cosine similarity between vectors for the same word.

The appendix adds a paired analysis: KECI/KCFW/KTVM (@NBCMontana) against KPAX-TV (@kpaxmissoula) in western Montana, and WCYB (@wcyb5) against WJHL-TV (@WJHLtv11) on the Virginia-Tennessee border, with matched numbers of videos and sub-sampled CNN/Fox data.

Why This Matters

Impact on research: The paper extends the small body of quantitative work on Sinclair ownership into a new medium — station YouTube channels — and shows that the shifts observed in broadcast and website data appear in social media content as well. It also demonstrates a methodological template for studying media ownership effects with topic models and embeddings, and it reports a null result on right-wing slant that differs from some prior work.

Real-world applications:

• Media policy and regulation: Regulators and commentators considering merger and ownership rules can point to evidence that ownership changes coincide with content changes in local stations.
• Journalism and newsroom practice: Journalists and editors can use the finding that local topics such as cooking, local sports, and community events decline after acquisition as a measurable indicator of local coverage loss.
• Media literacy and audience awareness: Viewers can be informed that the station they watch may have shifted toward national and politicized content, even if the branding stays local.
• Platform and dataset research: The work shows YouTube captions are a usable, scalable data source for large-scale media analysis, which is relevant to researchers studying social platforms.

Industry relevance: The paper touches on the economics of local news, including the documented decline of local news organizations and growing "news deserts," and on industry-wide incentives to attract views and engagement. It suggests that a shift to national, politicized content may be partly driven by engagement pressures, which matters to broadcasters, advertisers, and platforms.

Future Directions

• Measuring framing directly: The authors found some evidence of vocabulary shifts within topics (Table 11) but say future work targeting framing specifically is needed to explore trends that agenda-setting analysis cannot capture.
• Studying priming effects on viewers: The paper does not measure priming, and the authors propose using comments on YouTube videos as a way to directly examine viewers' responses to specific content, possibly crossed with other social media sources such as what links are shared on other platforms.
• Investigating right-wing slant with better tools: The authors suggest word embeddings may not capture slant because of limited context window size, and that CNN and Fox have been measured as less polarized before 2020. They call for methods that can disentangle whether the absence of a Fox-ward shift reflects content choices on YouTube, limited polarization of the comparison outlets in the study period, or methodological limitations.
• Separating confounders in the paired design: The paired analysis raises a possible interference problem, since local news stations may copy one another and a Sinclair purchase could affect content at nearby non-Sinclair stations. The authors note that addressing this would strengthen causal claims, which remain limited because the study is observational.

Target Audience

This paper is most useful to computational social scientists and NLP researchers interested in media analysis, to communications and journalism scholars studying ownership effects, and to media policy researchers and analysts tracking the decline of local news. Readers with a general interest in US media and politics will also find the framing accessible, though the methodological sections assume some familiarity with topic modeling and word embeddings.

Authors’ abstract

Local news stations are often considered to be reliable sources of non-politicized information, particularly local concerns that residents care about. Because these stations are trusted news sources, viewers are particularly susceptible to the information they report. The Sinclair Broadcast group is a broadcasting company that has acquired many local news stations in the last decade. We investigate the effects of local news stations being acquired by Sinclair: how does coverage change? We use computational methods to investigate changes in internet content put out by local news stations before and after being acquired by Sinclair and in comparison to national news outlets. We find that there is clear evidence that local news stations report more frequently on national news at the expense of local topics, and that their coverage of polarizing national topics increases.