Large-scale quantitative evidence of media impact on public opinion toward China

The study investigates whether and how mass media shape Americans’ attitudes toward China. In the context of intensifying U.S.–China rivalry and limited direct experience most Americans have with China, media portrayals are a primary source of information. Competing theoretical perspectives argue either that media exposure influences public opinion or that audience preferences shape media content. Given mixed evidence and often small-scale tests in prior research, the authors pose a big-data question: does elite media sentiment about China predict subsequent American public opinion toward China? Focusing on The New York Times (NYT) as an influential, elite U.S. outlet, the study aims to quantify the linkage between NYT reporting on China and longitudinal U.S. public opinion, acknowledging that the analysis is not causal but informed by a causal framework in which media influence can lead public attitudes.

The paper reviews two broad strands: (1) media effects research suggesting that media exposure shapes public opinion (e.g., Baum and Potter; Iyengar and Kinder); and (2) demand-driven views where consumer preferences and polarization influence media content (e.g., Gentzkow et al.; Mulligan and Shleifer; Jacobs and Shapiro), compounded by social media feedback loops. It also notes micro- and meso-level processes affecting opinion formation, such as social networks and life-course influences. Specific to China, prior surveys show Americans often hold unfavorable views of China’s government and power but are more positive toward Chinese people and culture; stereotypes place Asians as high-competence, low-warmth out-groups. The NYT is highlighted for its influence and elite readership, potential biases, and evolving dynamics in the digital and social media era. Prior related work used time-series models linking news to public sentiment; this study extends that tradition with a much larger corpus and modern NLP to assess NYT coverage of China across decades.

Data: The authors collected 267,907 NYT articles (1970–2019) mentioning China-related keywords (China, Chinese, Beijing, Shanghai) via the NYT API, extracting titles and dates. Public opinion data comprise a longitudinal series built from 101 cross-sectional U.S. surveys (1974–2019) from Roper Center, NORC (GSS), and Pew, harmonized into a yearly attitude measure relative to a 1974 baseline (−1 to 1 scale). NLP and topic-sentiment modeling: Using BERT (pretrained by Google and fine-tuned by the authors), articles were assigned to eight topics—ideology; government & administration; democracy; economic development; marketization; welfare & well-being; globalization; culture—and classified for sentiment (positive, negative, neutral/irrelevant). Performance was assessed with article-level accuracy and paragraph-level AUC; topic-wise accuracies reported high performance (e.g., article-level accuracies commonly 77–91% across topics; paragraph assignment AUCs ≈0.90–0.97). Yearly fractions of positive and negative articles per topic constitute media sentiment features. Modeling approach: The authors regress annual public opinion H_t on current and lagged media sentiment features F_{k,s,t−j} (topic k, sentiment s ∈ {positive, negative}, lag j), allowing for inertia in opinion and potential delayed media effects. They search over lag structures (primarily t−1 and t−2) to maximize explained variance (R^2). To guard against overfitting, constraints include: (a) non-negative coefficients after reverse-coding negative-sentiment variables (implying positive articles have zero or positive impact; negative articles have zero or negative impact on opinion); and (b) sparsity with at most one non-zero coefficient per topic (L0 constraint). They evaluate nested models increasing the number of topic predictors to assess gains in explained variance. Descriptive analyses include time trends of public opinion and NYT volume/sentiment by topic.

- The reporting of The New York Times on China in year t−1 explains approximately 54% of the variance in American public opinion toward China in year t (R^2 ≈ 0.539 using two predictors). - The best two predictors are: (1) the prior-year fraction of positive NYT articles on Chinese culture, and (2) the prior-year fraction of negative NYT articles on Chinese democracy, together explaining 53.9% of variation in public opinion. - Expanding the predictor set with more topics increases explained variance up to about 0.659 in the most inclusive nested model. - NYT attention to China has been consistently high (≥3,000 articles per year across the period). - Topic sentiment patterns diverge: negative sentiment is consistently more common in ideology, government & administration, economy, and welfare; positive sentiment is more common in economic development, globalization, and culture. Standard errors of yearly topic shares are very small (<1.55%). - U.S. public opinion toward China fluctuated markedly: lowest around −24% in 1976, peaking around +73% in 1987, and ranging between +10% and +48% for much of the subsequent three decades. - Media sentiment toward China has become more polarized over time, aligning with broader patterns of media polarization.

The findings indicate that elite media sentiment, as captured from NYT coverage of China, precedes and predicts a substantial share of variance in subsequent U.S. public opinion toward China. This temporal lead-lag pattern supports the thesis that media can shape public attitudes, addressing the study’s core question. The strongest signals come from cultural positivity (which correlates with more favorable public opinion) and democratic negativity (which correlates with less favorable public opinion), suggesting specific content dimensions through which media may influence perceptions. The authors emphasize a reduced-form, non-causal interpretation: the observed relationships aggregate many underlying processes, including news selection, agenda setting, and public reception dynamics. They discuss potential confounders and pathways—elite communications, exogenous geopolitical events, and social media amplification—that could modulate or mediate these effects. Nevertheless, the robustness of the lagged association over decades and across topics supports the relevance of media sentiment as a leading indicator of public attitudes toward foreign countries.

Reporting on China in The New York Times has grown more polarized over time, and media sentiment in specific domains (notably culture and democracy) strongly predicts subsequent American public opinion toward China. The study contributes large-scale, quantitative evidence linking elite media sentiment to public attitudes in foreign policy. Future research directions include: experimental or quasi-experimental designs to better identify causal effects of media exposure; richer modeling of mediating mechanisms and heterogeneous effects; extending the corpus to other influential outlets (e.g., The Guardian) and social media; and updating analyses to include post-2020 dynamics, including the COVID-19 era.

- Non-causal design: results are correlational/reduced-form and cannot establish causality. - Potential omitted variables: geopolitical events, elite cues, and policy changes may drive both media coverage and public opinion. - Media source bias: The NYT has an elite, left-leaning audience and specific editorial slants that may not generalize to the broader media ecosystem. - Simplified modeling: restricted coefficients and sparsity constraints may omit complex interactions, non-linearities, or multi-year dynamics. - Scope of media: reliance on a single outlet excludes heterogeneity across other newspapers, TV, radio, and digital-native platforms. - Temporal coverage: analysis ends before major shifts associated with COVID-19 and recent U.S.–China developments. - Micro/meso processes unmodeled: mechanisms of opinion diffusion via social networks and individual-level selection into media exposure are not directly analyzed.