Public opinion in Japanese newspaper readers’ posts under the prolonged COVID-19 infection spread 2019–2021: contents analysis using Latent Dirichlet Allocation

The study addresses how public opinion evolved in Japan during the prolonged COVID-19 pandemic, focusing on perspectives not well captured by social networking services (SNS). Prior research largely used SNS data in early pandemic phases and for short periods, which may underrepresent older demographics and emphasize acute or specialized topics. Given Japan’s extended pandemic period and reliance on voluntary compliance without strict lockdowns, the authors aimed to: (1) extract and characterize long-term, everyday public opinion topics from hardcopy newspaper readers’ posts; and (2) examine how these topics changed across four government-declared emergency periods (2020–2021) using topic modeling. The premise is that newspapers reach older audiences and provide identifiable contributor information, enabling insights different from SNS, and that long-term trends in public opinion are important for planning public health communication and policy responses.

The Introduction synthesizes work using questionnaires, interviews, and predominantly SNS (Twitter, Facebook, Weibo) to analyze public opinion during COVID-19, covering attitudes, emotions, and specific issues such as lockdowns, vaccines, and general pandemic discourse. SNS-based studies identified themes such as symptoms, case origins, impacts, and preventive measures, and highlighted utility for stakeholders in designing interventions. However, two limitations are noted: demographic bias toward younger users and anonymity effects that may encourage unsocial or extreme postings; and predominantly short-term/early-phase data collection, missing longer-term trends. Newspapers, with higher usage and trust among older demographics in Japan, and stable archival databases, offer a complementary source to capture everyday, longer-term public opinion beyond acute topics commonly seen on SNS.

Design: Document analysis of readers’ posts published in three major Japanese newspapers (Yomiuri, Mainichi, Asahi) during four COVID-19 state-of-emergency periods in Japan: Apr 7–May 25, 2020; Jan 8–Mar 21, 2021; Apr 25–Jun 6, 2021; Jul 12–Sep 30, 2021. Data sources and sampling: Newspaper databases (Asahi Shimbun Cross-Search; Mainichi Maisaku; Yomiuri Yomidas Rekishikan) were searched using the term “coronavirus.” Only posts published during the four emergency periods were included. Total included posts: 1910 (First: 412; Second: 521; Third: 458; Fourth: 519). Distribution by paper: First—Yomiuri 123, Mainichi 123, Asahi 166; Second—Yomiuri 164, Mainichi 165, Asahi 192; Third—Yomiuri 139, Mainichi 151, Asahi 168; Fourth—Yomiuri 133, Mainichi 192, Asahi 194. Average posts/day ≈ 7.2 (range 0–19). Mean characters per post 376.7 (SD 107.3). Contributor age was reported in 996 cases (52.2%), median 62 (range 8–98). Preprocessing: Analysis focused on nouns and adjectival nouns. The corpus contained 12,951 unique nouns with 100,927 total occurrences. Notational variants were normalized (e.g., “world” variants unified). Extremely frequent generic pandemic terms (“COVID-19” and “Corona-Ka/コロナ禍”) were excluded from topic modeling to prevent dominance and improve topic discrimination. Topic modeling: Latent Dirichlet Allocation (LDA) implemented via KH Coder (v3 Beta 04a) with R-based morphological analysis was used. The number of topics K was tuned using multiple indices via the ldatuning package: Arun (2010), CaoJuan (2009), Deveaud (2014), Griffiths & Steyvers (2004). Considering minima (Arun, Cao, perplexity) and maxima (Deveaud, Griffiths), K=10 was selected (10–11 indicated; 10 chosen due to index behavior). KH Coder produced topic scores; the top 10 words per topic guided interpretation and naming by author consensus. Statistical analysis of trends: For each post, LDA topic proportions (topic ratios) sum to 1. To assess monotonic trends across the four emergency periods, the Jonckheere–Terpstra trend test was applied to topic ratios. Effect sizes were quantified using Kendall rank correlation coefficients (τ) between period order and topic ratios. Analyses were conducted in IBM SPSS Statistics 27. Given stable public health measures and low infection/mortality rates in Japan during 2020–2021 and large sample sizes, τ ≥ 0.10 in magnitude was considered a meaningful effect.

- Ten topics were extracted and grouped into three themes: Life (Family; Daily life in the COVID-19 disaster; Education in the COVID-19 disaster; The importance of humanity; Daily life unrelated to COVID-19), Awareness of the emergency (Awareness of being a party to an emergency; Concerns about the medical environment), and Policy (Domestic and foreign policies; Opposition to hosting the Tokyo Olympics; Criticisms of the Japanese government). - Topic trend significance (Jonckheere–Terpstra p-values): Family 0.004; Daily life in COVID-19 0.044; Education in COVID-19 0.017; Importance of humanity <0.000; Daily life unrelated to COVID-19 <0.000; Awareness of being a party to an emergency 0.000; Concerns about the medical environment 0.046; Domestic and foreign policies <0.000; Opposition to hosting the Tokyo Olympics <0.000; Criticisms of the Japanese government 0.011. - Effect sizes (Kendall τ) across periods: Awareness of being a party to an emergency showed the only notable effect, decreasing over time (τ = −0.189). Other topics showed small effects: Opposition to hosting the Tokyo Olympics τ = 0.094; Criticisms of the Japanese government τ = 0.044; Domestic and foreign policies τ = −0.058; Daily life unrelated to COVID-19 τ = −0.060; Importance of humanity τ = −0.057; Family τ = −0.050; Education in COVID-19 τ = −0.041; Daily life in COVID-19 τ = 0.034; Concerns about the medical environment τ = −0.034. - Contributor characteristics: Age reported in 52.2% (median 62, range 8–98). Mean post length 376.7 characters (SD 107.3). - Overall, everyday topics persisted with limited change, while explicit emergency-awareness content declined over successive emergency declarations.

Newspaper readers’ posts yielded everyday, human-centered topics (family, humanity, daily life) that were not prominent in SNS-based studies, likely due to demographic differences (older readership), non-anonymous posting, longer allowable text, and the editorial context of print media. Conversely, acute or highly specialized topics common on SNS (e.g., outbreak origins, rapid scientific updates) were not salient in the newspaper posts, suggesting SNS may be better suited for rapidly evolving, technical discourse. Longitudinally, only the topic "Awareness of being a party to an emergency" showed a meaningful decline (τ = −0.189), consistent with pandemic fatigue, wherein motivation to adhere to preventive behaviors wanes over time. Mentions of other topics changed little, indicating that everyday concerns (family, education, medical environment, general policy sentiments) persisted throughout the prolonged pandemic. While some statistical trends appeared due to large sample size, effect sizes suggest stability in most domains of public opinion. These findings imply that, during extended crises, policymakers should not only address acute, time-sensitive sentiments (often captured on SNS) but also sustain responses to enduring everyday concerns that remain salient over long periods.

Analyzing readers’ posts in major Japanese newspapers across four COVID-19 emergency periods revealed persistent everyday public opinion topics and a notable decline only in emergency awareness, indicative of pandemic fatigue. Newspaper-based analyses complement SNS by capturing long-term, day-to-day concerns among older demographics. For comprehensive understanding and effective response to public opinion during pandemics, stakeholders should integrate multiple data sources (SNS and traditional media) and combine short-term/acute with long-term/everyday perspectives. Future research should continue multi-modal, longitudinal monitoring and refine topic modeling approaches to enhance robustness and generalizability.

Generalizability is limited because the dataset consists of published readers’ posts selected by newspaper editors; unpublished submissions and editorial policies may bias the sample. Findings cannot be generalized to all Japanese society, all newspaper readers, or all submitters. Topic modeling (LDA) outcomes depend on data quality/quantity, parameter choices (including K), and subjective interpretation of topics, which may introduce variability. Further studies are needed to validate and extend these results.