The Space of Newspaper Policy Preferences

• In Event: Party Ideology and the Economy

Abstract

Representative democracy is predicated on having a public sufficiently informed to hold leaders accountable and to make sensible choices when voting. In the United States and elsewhere, the media is the primary institution responsible for keeping the public informed. But decades of elite criticism of the press and economic turbulence within the industry have undermined its institutional credibility and left the public deeply skeptical of partisan bias in the press. Accordingly, a large body of research attempts to quantify media’s partisan slant – generally implicitly assuming media outlets have ideological or partisan biases that can be mapped cleanly onto the traditional unidimensional spectrum.

In this paper, we challenge this assumption and ask: (1) To what degree can American newspapers’ endorsement behavior be considered ideological in the liberal-conservative sense of the word? (2) Insofar as newspapers are not ideological, what considerations drive their endorsement behavior?

Our approach is to study newspaper editorial board endorsements on statewide ballot measures, leveraging the fact that interest groups and political parties also endorse on and citizens ultimately vote on these same ballot measures. This allows us to locate newspapers relative to these groups and voters on the same scale. Across four states with a sufficient number of ballot initiatives and newspapers – California, Florida, Oregon, and Washington – we have to date collected 32,888 endorsements on 842 ballot measures from 196 newspapers and 242 groups between 1988 and 2022. We have also scraped more than 76,000 newspaper pages to examine the full text of newspaper endorsements.

To answer our first research question, we apply three different scaling techniques to the endorsement data: optimal classification, principal components analysis, and W-NOMINATE. We measure goodness-of-fit using the average proportional reduction in error (APRE), a measure that is well-suited to our data because it accounts for the fact a naïve model could accurately predict a sizable share of endorsements by simply assuming all endorsers side with the majority.
Preliminary results show two interesting findings. First, U.S. newspapers are less ideological than previously assumed, in the sense that their behavior is not well-explained by a simple unidimensional spatial model. A one-dimensional model fits interest groups relatively well (APRE around 60-70%) and newspapers relatively poorly (APRE around 20-40%). Second, focusing on California, where the data is richest, we find that newspapers have become more ideological over time.

To answer the second question, we exploit recent developments in computer vision and natural language processing. First, we scraped Newspapers.com for all newspaper page scans that mentioned a statewide ballot proposition and then trained an n-gram classifier on the concatenated optical character recognition of the page to discover which pages were editorial pages. Now, alongside data from NewsLibrary.com, we are in the process of performing text detection, recognition, association, and ordering on newspaper opinion pages to extract articles and classify which of those articles are editorial board endorsements of statewide ballot propositions. Next, we plan to fine-tune a language model to extract spans of text that represent arguments in those articles. We then classify and cluster arguments to ask whether editorial boards’ arguments can be grouped into meaningful bins or whether journalists’ considerations are so diverse and idiosyncratic that they essentially amount to random noise.

Given the ongoing threats to American media – and the centrality of journalism’s watchdog role in fostering a healthy democracy – it is imperative that scholars challenge our assumptions about the American press and accurately model and measure its behavior. We believe this project helps to do so and offers a worthwhile contribution to the field of political communication.

Authors

• Antonio Camara

• James M. Snyder, Harvard University

• David Beavers, Harvard University