Proximity-diversity paradox: was spatiotemporal proximity to news events associated with greater diversity in news topics?

1 Introduction

News reporting is of paramount importance in people's daily lives, for it serves not only as a crucial channel for information acquisition but also as a key mechanism shaping people's perception of the real world and the formation of their worldviews (Happer and Philo, 2013; Zaklama, 2025). In the process of public perception shaping, the diversity of news topics plays a significant role (Joris et al., 2020). Therefore, research on news topic diversity has attracted increasing attention from scholars (Hendrickx et al., 2022). The diversity of news topics could influence the breadth of public perception of the real world (Huang, 2010), and help media attract larger audience in the competitive markets, making it a key factor affecting media survival (Gentzkow et al., 2014). Hence, research on news topic diversity could reveal the operational mechanisms of news media and their influence on the public, providing insights for effectively leveraging media to disseminate information and shape social reality.

Currently, research on news topic diversity primarily employs comparative approach to analyze the similarities and differences in coverage topics of relevant news across media. This type of research mainly revealed 2 patterns: the topics of news concerning a specific subject were significantly different (Freudenthaler and Wessler, 2025) or roughly the same (Hendrickx et al., 2022) among different media. However, not many scholars probed into topic homogeneity of news on a specific subject in different media across different countries. In addition, theoretically, domestic media coverage of national affairs might be more varied than foreign media coverage of the same affairs, and the former might exhibit greater topic diversity than the latter. This was because the former enjoyed superior spatial proximity, convenience, and accessibility in news gathering, which might yield more comprehensive and nuanced information about domestic affairs. However, the media of a country often had resident reporters in many overseas countries. At the same time, the Internet made information disseminate faster across countries around the world, and facilitated the reuse of news contents among media from various countries. These channels of news production might cause news topics across different media to converge toward homogeneity. In that case, domestic media coverage of national affairs might not necessarily exhibit greater topic diversity than foreign media coverage of the same affairs. Therefore, news production might display the proximity-diversity paradox, i.e., domestic media that were spatiotemporally close to national news events did not present greater topic diversity in news than more distant foreign media. In other words, the spatial and temporal proximity of news media to news events might not lead to an increase in topic diversity in media coverage.

Guangdong Province was one of the fastest-growing economic provinces in China and an epicenter of U.S.-China geopolitical and trade tensions, therefore, theoretically, it is a crucial target for news reports of China's and American newspapers. Methodologically, Guangdong Province's high newsworthiness in both countries’ media systems ensured an adequate supply of comparable news data. In sum, the comparison of the English news on Guangdong Province between China's and American newspapers was theoretically informed and methodologically sound. This study would employ the dependency parsing (Imrényi and Mazziotta, 2020), Bayesian hierarchical Bernoulli logistic regression model (Gelman et al., 2014), and Bayesian hierarchical negative binomial regression model (Gelman et al., 2014), to verify the homogeneity of topic nouns in news on Guangdong in China's and American newspapers, so as to examine whether the observed evidence was consistent with the proximity-diversity paradox. This study defined a topic as a surface-level lexical domain activated by nouns in core syntactic positions (subjects, objects, predicatives, and roots). It was distinct from a theme, which involved the deep-level semantic contents obtained through propositional interpretation. Accordingly, 2 sentences might activate the same lexical domain (topic) while addressing entirely different themes. The current study could reveal a new relation between media-source proximity and news topic diversity, introduce a new methodology to the studies of the homogeneity of news topics, and enhance the research on news texts by leveraging modeling analysis.

2 Research on the topic diversity among different media

Currently, comparative analyses of cross-media news coverage are the main approach to uncover the news topic diversity. This type of research involved the news reports on a specific subject among different media and mainly identified 2 patterns: the topics of the news concerning a specific subject were significantly different or roughly the same among different media. Regarding the research on the differences in the topics of the news on a specific subject among different media, scholars disclosed there were significant differences in the topics of news released by different media, so as to verify certain media conveyed more diversified or different topics in the news. For example, when reporting on the conflict between large carnivores and humans in France, the French regional newspaper Nice-Matin and the national newspapers Le Monde presented different topics in the reports, the former focused on factual issues and displayed these issues from a human-centered perspective, while the latter concerned general management issues and showed these issues from a viewpoint of carnivores, the different topic representation might influence humans’ perception of carnivores (Chandelier et al., 2018). In the comparison of the topic diversity of the news on U.S.-led multinational military operations in Libya, Iraq, Afghanistan and Kosovo in different newspapers in 116 countries, it was found that the topics of the newspapers co-owned by a media enterprise were less diverse than those of newspapers which were not co-owned, and the competitive market pressures could alleviate the problem of topic homogeneity (Baum and Zhukov, 2019). In addition, comparison of 12 U.S. and U.K. outlets’ coverage of the Paris Agreement negotiations (6 per country) revealed that the U.K. media system exhibited higher external topic diversity (greater differentiation across outlets), while U.S. outlets showed marginally higher internal diversity (greater variation within each individual outlet), this suggested that public funding mechanisms like the BBC's, which provided autonomy from market pressures, might contribute to external diversity in news topics, though such mechanisms could reduce individual outlets’ internal diversity in the topics (Neff, 2020). In the transregional news coverage in multilingual country of Switzerland, both publicly-oriented media (public service and subscription) and commercially-driven media (tabloid and commuter) presented a lot of news about different language regions of Switzerland, but they focused on different topics, specifically, the former mainly concerned the news on politics, while the latter mainly involved the news on sports, this divergence arose from their funding models: the former, funded by public or subscribers, fulfilled a civic mandate to cover political news, whereas the latter, reliant on advertising revenue, targeted mass audiences with popular sport news (Vogler and Udris, 2021). When reporting on refugee policy in Germany, the mainstream online news outlets displayed more diverse topics than the alternative online news outlets did, but the latter embodied more internal topic diversity, this was because the topics of the alternative online news outlets were restricted by their partisan stances and organizations’ work (Freudenthaler and Wessler, 2025). In other words, partisan orientation and operation models would influence the news topic diversity of media (Bohr, 2020). In sum, this type of research revealed there were significant differences in topic diversity of news on a specific subject among different media, however, since media organizations might adopt operation models from one another and information sharing among them across different countries is perhaps becoming more frequent, cross-national differences in the diversity of news topics on a specific subject might be narrowing. Therefore, some scholars probed into the similar patterns in the topics of the news on a specific subject among different media.

As for the research on the homogeneous topics of the news on a specific subject among different media, scholars mainly examined the similarity in the news topic distribution of different media, so as to prove some media did not display more diverse topics than other media. For instance, 7 Indian dailies showed significant homogeneity in news topics, and the homogeneity was caused by uniform news selection principles and shared standards of news values across the newspapers (Haque, 1986). Meanwhile, the similarity in international news topics between the morning New York Times and the evening news programs of ABC, NBC, and CBS might be partially explained by inter-media agenda-setting effects, where the Times’ coverage appeared to influence subsequent television news selection (Golan, 2006). Furthermore, the comparison of 9 major newspapers in Belgium displayed that the news contents of these newspapers had become more homogeneous over time, and news contents demonstrated greater homogeneity among newspapers under common corporate ownership than among their independently owned counterparts (Beckers et al., 2019). A content analysis of the print and online editions of 4 Mediahuis newspapers showed that article recycling among these newspapers was on the rise, significantly eroding topic diversity of their news output, particularly on their news websites (Hendrickx, 2020). Sharing a content management system among different newspapers in a media group could accelerate the copy-pasting of news articles across these newspapers, which might make news contents more homogeneous (Hendrickx and Ranaivoson, 2021). In addition, ownership of media might give rise to the topic homogeneity of the news in the media, specifically, the ownership concentration of media might cause the news to be more homogeneous (Hendrickx et al., 2022). The collaboration among different media companies during the production of news might also make the news in these media more homogeneous (Pakvis and Hendrickx, 2025). In sum, this type of studies mainly disclosed the homogeneity of news topics and its causes, however, not many scholars investigated the topic homogeneity of news on a specific subject produced by different countries’ media. As most newspapers now have online editions, media in different countries could more easily share news contents with each other, which might lead to topic homogeneity in their coverage of a subject. This homogeneity could reflect how different media synchronized narrative framing and how public across countries developed similar perception of the subject, making research on this phenomenon theoretically and practically significant.

In summary, firstly, the research on the differences in the topics of the news on a specific subject among different media revealed that, there were significant differences in topic diversity of the relevant news. However, as information sharing among media across different countries becomes increasingly easy and frequent, cross-national differences in news topic diversity on specific subjects are likely narrowing. Secondly, the studies of the homogeneous topics of the news on a specific subject among different media disclosed the homogeneity of news topics and its causes. However, this type of studies mainly focused on cross-media comparison within a country or an individual media conglomerate, not many studies probed into the topic homogeneity of the news released by the media across different countries. Thirdly, Guangdong Province is China's economic powerhouse and a primary locus of U.S.-China geopolitical and trade tensions, hence, theoretically, it is a critical site for news reports of China's and American newspapers. Methodologically, Guangdong Province's high news values in both countries’ media systems ensured that there would be enough parallel news data for comparison. In addition, China's newspapers are more proximate to Guangdong's news events than American newspapers across spatial and temporal dimensions, the news on Guangdong would be the ideal data for examining the proximity-diversity paradox. Therefore, the empirical comparison of the English news on Guangdong Province between China's and American newspapers was theoretically and methodologically grounded. Accordingly, this study would verify the lexical topic homogeneity of the English news on Guangdong Province from the 2 countries’ newspapers, and provide preliminary evidence consistent with the proximity-diversity paradox.

This paradox was not a rebranding of media homogenization. Media homogenization and global media logic (Altheide and Snow, 1979; Hallin and Mancini, 2004) predicted universal convergence driven by transnational professional norms and shared news sources, an effect that might apply uniformly to some western news agencies. By contrast, the proximity-diversity paradox identified a specific condition: it predicted that despite varying spatiotemporal distances from news events, Eastern and Western news agencies exhibited no significant lexical topic divergence in their coverage of Guangdong. Agenda-setting and institutional constraints explained why content converged through organizational routines or peer influence among journalists (Guo and Vargo, 2017; Harder et al., 2017), while this paradox transcended the explanatory scope of organizational routines and peer influence among journalists, revealing lexical topic homogenization in global news coverage of Guangdong. Thus, the contribution of this study lied not in observing similarity between China's and American newspapers, but in demonstrating that various spatiotemporal distances from news events might not prevent news agencies from producing homogenized news topics, i.e., an implicit tension that existing theories did not explicitly theorize.

3 Research design

3.1 Research questions

This study compared the lexical topics of the English news on Guangdong Province between China's and American newspapers so as to examine the topic diversity of the news in China's newspapers, and provided preliminary evidence consistent with the proximity-diversity paradox. The greater spatial and temporal proximity of China's newspapers to Guangdong-related news (relative to their American counterparts) was a prerequisite for the proximity-diversity paradox. This prerequisite was treated as an a priori condition, and it was not subjected to empirical analysis. Specifically, this study would probe into 2 research questions: (1) Were there any significant differences in the lexical topic diversity of the news on Guangdong Province between China's and American newspapers? (2) Were there any significant differences in the lexical topic frequency of the news on Guangdong Province between the 2 countries’ newspapers?

3.2 Data

On June 13, 2025, this study selected the research data from Beijing Review (BR), China Daily (CD), Global Times (GT), People's Daily (PD), Shanghai Daily (SD1), Shenzhen Daily (SD2), Chicago Tribune (CT), Journal of Commerce (JC), Los Angeles Times (LAT), Pittsburgh Post (PP), The Associated Press (AP), and The New York Times (NYT) in LexisNexis corpus (http://www.lexisnexis.com/ap/academic/). This study exhaustively retrieved the English news with the title containing “Guangdong” from the above newspapers, and obtained 7, 1,161, 153, 387, 188, 256, 1, 1, 4, 1, 27, and 4 news texts, respectively. These news texts were cleaned by deleting the exact duplicates and texts with images but no textual contents. Finally, 7, 876, 149, 382, 182, 246, 1, 1, 4, 1, 26, and 4 news texts were retained from the aforementioned newspapers, respectively. Among them, the news texts from CT, JC, LAT, PP, AP, and NYT were compiled into the American Corpus (AC), which contained 37 relevant news texts and 13,650 English tokens. In order to balance the corpus size of AC and the China's Corpus (CC), CC was constructed via systematic sampling from BR, CD, GT, PD, SD1, and SD2 by selecting every 48th news article (i.e., sampling 1 news text after every 47 texts), beginning with the first, and yielding 39 news texts (13,459 English tokens). This study would employ the dependency parsing (Imrényi and Mazziotta, 2020), Bayesian hierarchical Bernoulli logistic regression model (Gelman et al., 2014), and Bayesian hierarchical negative binomial regression model (Gelman et al., 2014) to compare lexical topic diversity and lexical topic frequency between CC and AC.

3.3 Research tools

In this study, dependency parsing was performed to extract topic nouns, and Bayesian hierarchical Bernoulli logistic regression model was exploited to compare the diversity of topic nouns in CC and AC, in addition, Bayesian hierarchical negative binomial regression model was employed to compare the frequency of topic nouns in the 2 corpora, so that preliminary evidence consistent with the proximity-diversity paradox would be provided. Specifically, this study utilized RStudio to invoke spacyr package to perform dependency parsing (Imrényi and Mazziotta, 2020) on the news in CC and AC, respectively, and lemmatize all the English words in these 2 corpora. The results were restored in 2 CSV documents. Microsoft Excel was exploited to extract the lemmatized nouns that functioned as subjects, objects, predicatives, and roots. These nouns were considered as topic nouns (Perfetti and Goldman, 1974).

Firstly, this study examined the differences in the occurrence probabilities of topic nouns between CC and AC, therefore, it was not viable to conduct ANOVA (Analysis of Variance) or Simple Linear Model Analysis, etc. to compare the occurrence probabilities of the topic nouns in CC and AC. Bayesian hierarchical Bernoulli logistic regression model (Gelman et al., 2014) was employed in RStudio to analyze the differences in occurrence of these nouns between CC and AC, so that the differences in diversity of lexical topics in CC and AC could be testified. This study initially examined the differences in lexical topic diversity (i.e., richness of topic nouns), therefore, the occurrence of topic nouns was binarized: each topic noun was coded as 1 (present) if it appeared in CC or AC (regardless of how many times it occurred in a corpus), and 0 (absent) otherwise. This transformation was applied exclusively to the calculation of lexical topic diversity. Since the analysis of lexical topic diversity was based on the binary data (presence/absence of topic nouns) and the occurrence of topic nouns might be correlated within each corpus (CC or AC) for the reason that the occurrence of them might be influenced by similar contexts or stylistic norms, hence, the analysis was based on a hierarchical Bernoulli distribution, as shown in Equation (1):

occurrence∼Bernoulli(p)(1)

In this distributional assumption, p referred to the occurrence probability of a topic noun in CC or AC. This study established a Bayesian hierarchical Bernoulli logistic regression model to analyze the influence of different countries and specific topic nouns on the occurrence probabilities. The model was presented in Equation (2):

logit(pi)=β0+β1⋅countryi+vnoun[i](2)

In this model, logit(pᵢ) referred to the log-odds of occurrence of the ith topic noun in a specific corpus (CC or AC), indicating its tendency to appear within that corpus, β₀ denoted average log-odds of topic nouns occurring in the baseline corpus CC, β₁·countryᵢ represented the log-odds differences for AC relative to CC (i.e., AC−CC), and v_noun[*i*] reflected the random effect (random intercept) associated with the ith topic noun, capturing word-specific deviations from the overall mean log-odds of occurrence predicted by the fixed effects (i.e., the average tendency of all topic nouns within each corpus). By exploiting the model, this study could testify whether there were significant differences in the occurrence distribution of topic nouns of the English news on Guangdong Province between CC and AC.

Secondly, this study examined differences in the lexical topic frequency between CC and AC. Since the frequency of the topic nouns in CC and AC was sparse (i.e., many topic nouns appeared exclusively in CC or AC, respectively), it is not viable to conduct ANOVA or Simple Linear Model Analysis, etc. to compare the frequency of the topic nouns in CC and AC. Bayesian hierarchical negative binomial regression model (Gelman et al., 2014) was employed in RStudio to analyze the differences in frequency of these nouns between CC and AC, so that the differences in topic frequency in CC and AC could be evaluated and the proximity-diversity paradox could be examined. Since the variance of topic noun frequency was greater than the mean of the topic noun frequency in CC and AC, respectively, the comparison of the frequency of topic nouns in the 2 corpora was suitable for the negative binomial distribution assumption, as shown in Equation (3):In this assumption, μ referred to the expected frequency of a topic noun in CC or AC, and ϕ denoted the dispersion parameter of the model. This study established a Bayesian hierarchical negative binomial regression model to analyze the influence of different countries and specific topic nouns on the frequency. The model was presented in Equation (4):In this model, log(μ_i) referred to the logarithm of the expected frequency of the ith topic noun in CC or AC. Log(total) was the logarithm of the cumulative frequency of all the tokens in CC or AC, which served as an offset in the model. β₀ was the global intercept, β₁⋅countryᵢ denoted the fixed effect of a country, v_noun[*i*] indicated the random effect associated with the ith topic noun. This model employed log(total) as an offset of the model, it could effectively estimate the relative frequency of each topic noun after controlling for the total numbers of tokens in CC or AC. This method could make sure that comparison of topic noun frequency was not influenced by the differences in the corpora size of CC and AC. Therefore, by exploiting the model, this study could testify whether there were significant differences in the frequency distribution of topic nouns of the English news on Guangdong Province between CC and AC, and hence examining whether the evidence was consistent with the proximity-diversity paradox.

frequency∼NegBinomial(μ,ϕ)(3)

log(μi)=log(total)+β0+β1⋅countryi+vnoun[i](4)

3.4 Research procedures

This study employed dependency parsing, Bayesian hierarchical Bernoulli logistic regression model, and Bayesian hierarchical negative binomial regression model, to assess whether the evidence was consistent with proximity-diversity paradox. Firstly, the English news on Guangdong from China's and American newspaper were collected and cleaned. In order to balance the comparative corpora, CC consisted of a sample of 39 news texts from China's newspapers, while AC comprised all 37 cleaned texts from American newspapers. Secondly, this study exploited RStudio to invoke spacyr package to conduct dependency parsing on the news texts and lemmatize the English words in CC and AC, respectively. Subsequently, Microsoft Excel was utilized to extract the lemmatized nouns that functioned as subjects, objects, predicatives, and roots, and these nouns could be the topic nouns of the English news on Guangdong (Perfetti and Goldman, 1974). Thirdly, Bayesian hierarchical Bernoulli logistic regression model and its robustness tests were exploited to compare the diversity of topic nouns in CC and AC, so that the differences in the lexical topic diversity of the news on Guangdong Province between China's and American newspapers could be assessed. Fourthly, Bayesian hierarchical negative binomial regression model and its robustness tests were employed to analyze the differences in the frequency of the topic nouns between CC and AC, therefore, the differences in the lexical topic frequency of the news on Guangdong Province between the 2 countries’ newspapers could be evaluated. On the basis of examining whether lexical topic diversity and lexical topic frequency differed, the evidence of proximity-diversity paradox could be assessed.

4 Data analysis

4.1 The comparison of topic diversity

With the help of dependency parsing and lemmatization of CC and AC, 2,240 and 2,290 topic nouns that functioned as subjects, objects, predicatives, and roots were extracted from these 2 corpora respectively. The frequency of the top 20 topic nouns in CC and AC was displayed in Table 1.

Table 1. The frequency of the top 20 topic nouns in CC and AC.

The standardized frequency was weighted by 10,000.

According to Table 1, generally speaking, CC and AC showed few differences in the standardized frequency of their top 20 topic nouns, and they shared 8 words in this word list, such as province, year, percent, yuan, people, city, case, and government. The top 20 high-frequency topic nouns in CC and AC mainly involved the issues of government and public administration (such as office, policy, government, mainland, and tariff), economy, finance, and development (e.g., yuan, development, investment, bank, company, stake, report, consortium, investor, worker, bid, share, and deal), society and demography (such as people, case, couple, divorce, and cat), regional and urban development (e.g., province, region, city, area, and country), and statistical indicators (such as year, percent, number, and time). In sum, the high-frequency topic nouns in CC and AC concerned similar issues related to Guangdong.

This study employed Bayesian hierarchical Bernoulli logistic regression model (Gelman et al., 2014) to assess the differences in lexical topic diversity between CC and AC. Prior to analysis, topic nouns were dichotomously coded for 1 (presence) versus 0 (absence) in CC or AC, disregarding their frequency of occurrence. In the model, the log-odds of occurrence of a topic noun in CC or AC were modeled as the sum of average log-odds of topic nouns occurring in the baseline corpus CC (β₀), the log-odds differences for AC relative to CC (β₁·country), and the random effect (random intercept) associated with the topic noun [v_noun(*i*)]. In other words, this study considered the topic nouns themselves (i.e., each topic noun's inherent variation of occurrence probabilities) might influence the occurrence of the topic noun in CC or AC. The model incorporated 4 Markov chains, each of which was executed on a separate core to enhance sampling efficiency. Each chain performed 2,000 iterations, with the initial 1,000 iterations designated for warm-up to ensure convergence and stability in sampling. Additionally, a high adaptation parameter (adapt_delta = 0.99) was set to minimize the likelihood of divergent transitions. The maximum tree depth was specified as 15 (max_treedepth = 15) to enable the sampler to thoroughly explore the data.

In this model, all Rhat values were below 1.01, and both Bulk-ESS and Tail-ESS values exceeded 400, indicating good convergence. The model also showed zero divergent transition. Overall, the model achieved satisfactory goodness of fit and prediction performance.

Using the Bayesian hierarchical Bernoulli logistic regression model and the credible intervals derived from the Bayesian posterior distribution, this study examined the differences in the occurrence probabilities of topic nouns between CC and AC, and the results were displayed in Table 2.

Table 2. The comparison of the occurrence of topic nouns between CC and AC.

According to Table 2, the Bayesian credible interval of the results of the comparison between CC and AC contained 0 (in logistic regression, 95% Credible Interval containing 0 indicated non-significant results), which suggested that the differences in the occurrence probabilities of topic nouns between CC and AC were not significant.

To address the concern that non-significant results might reflect insufficient statistical power rather than genuine equivalence of the occurrence probabilities of topic nouns between CC and AC (Lakens, 2017), the current study conducted 3 robustness checks for the Bayesian hierarchical Bernoulli logistic regression model. Firstly, to facilitate substantive interpretation, this study converted the log-odds coefficients from the Bayesian hierarchical Bernoulli logistic regression model to the probability differences (i.e., the absolute differences in predicted occurrence probabilities of topic nouns between CC and AC) and conducted a ROPE (Region of Practical Equivalence) analysis (Kruschke, 2013; Kruschke and Liddell, 2018). A ± 0.05 threshold was chosen on the probability difference scale because a 5-percentage-point difference was substantively negligible in media effect research. Meanwhile, Cohen's h was computed using a ± 0.1 ROPE, corresponding to half of Cohen’s (1988) small effect benchmark for proportion. The analysis showed that 96% of the posterior mass fell within the equivalence region [median = −0.017, 95% Credible Interval (−0.054, 0.022)]. Additionally, Cohen's h exhibited consistent results [median = −0.035, 95% Credible Interval (−0.112, 0.047), with 95.2% of the posterior mass within a ± 0.1 ROPE]. These results provided strong evidence that the true differences in the occurrence probabilities of topic nouns between CC and AC were very small and negligible. Secondly, Bootstrap resampling (n = 30) confirmed that the 95% credible interval crossed zero in 100% of iterations, which indicated that the observed similarity of the occurrence probabilities of topic nouns between the 2 corpora was robust to sampling variation and not the result of idiosyncratic topic noun distributions or particular data points. Thirdly, a prior sensitivity analysis (diffuse, weakly informative, and concentrated priors) yielded virtually identical results on the probability difference metric (median range: −0.0165 to −0.0160; maximum absolute difference: 0.0005), indicating that the findings were not driven by prior specifications but rather by the empirical features of the data itself. Therefore, it was safe to conclude that there was not significant difference in the lexical topic diversity of the news on Guangdong Province between China's and American newspapers.

4.2 The comparison of topic frequency

This study constructed Bayesian hierarchical negative binomial regression model (Gelman et al., 2014) to analyze the differences in the frequency of the topic nouns between CC and AC. The logarithm of the cumulative frequency of all the tokens in CC or AC [i.e., log(total)] was incorporated into the model as an offset, thereby, the comparison of topic noun frequency was not influenced by the differences in corpora size of CC and AC. In other words, the standardized frequency of the topic nouns was compared in the model. Meanwhile, the model incorporated the fixed (average) effect of countries on frequency (β₁⋅countryᵢ) and the random effects of individual topic noun on frequency [v_noun(*i*)], the latter indicated the topic nouns’ inherent frequency variation might influence usage frequency. To enhance sampling efficiency, the model ran 4 Markov chains concurrently on separate cores. Each chain generated 8,000 iterations, with the initial 4,000 served as warm-up to achieve sampling stability and convergence. Moreover, to minimize divergent transitions, this study set adapt_delta = 0.999, and specified max_treedepth = 15 to enable the sampler to thoroughly explore the data.

Model diagnostics showed strong convergence (Rhat < 1.01, Bulk-ESS and Tail-ESS > 400) with no divergent transition, indicating satisfactory fit and predictive performance. By exploiting Bayesian hierarchical negative binomial regression model and the credible intervals derived from the Bayesian posterior distribution, this study evaluated the differences in the frequency of the topic nouns between CC and AC, and the results were shown in Table 3.

Table 3. The comparison of the frequency of topic nouns between CC and AC.

According to Table 3, AC exhibited a slightly lower topic noun incidence rate than CC (IRR = 0.956, 95% Credible Interval = 0.864–1.058). However, as the Bayesian credible interval contained 1 (in negative binomial regression, 95% Credible Interval spanning 1 indicated non-significant results), the difference in the frequency of topic nouns between CC and AC was not statistically significant.

For the purpose of addressing the issue that non-significant results might embody insufficient statistical power rather than genuine equivalence of the frequency of topic nouns between CC and AC (Lakens, 2017), this study performed 3 robustness tests for the Bayesian hierarchical negative binomial regression model. Firstly, the main Bayesian hierarchical negative binomial regression model estimated the corpus effect on the log-count scale, yielding a median value of −0.045 [95% credible interval (−0.148, 0.059)]. In the ROPE robustness check, the same posterior was transformed to the incidence rate ratio (IRR) scale [median = 0.956, 95% credible interval (0.862, 1.061)]. Using an IRR-based ROPE of [0.9, 1.1], which captured relative frequency ratios within a ± 10% range around a ratio of 1, 87% of the posterior distribution fell within the equivalence region. This result indicated that the differences in the frequency of topic nouns between CC and AC were probably negligible in practical terms. Secondly, in the alternative-operationalization analysis, all 3 measurement specifications produced credible intervals spanning the null value: raw count analyzed with the Bayesian hierarchical negative binomial regression model (identical to the main model) [median log-ratio = −0.044, 95% credible interval (−0.148, 0.057)], normalized frequency per 1,000 tokens analyzed with a Gaussian model [median = 0.001, 95% credible interval (−0.016, 0.018)], and binary occurrence analyzed with a Bernoulli model [median log-odds = −0.072, 95% credible interval (−0.234, 0.096)]. Because these estimates were reported on their respective original scales, they were not directly comparable in magnitude. Nevertheless, all 3 intervals include the null value. This suggested that no substantive differences in the frequency of topic nouns were found between the 2 corpora under any of the above 3 measurement approaches, therefore, the observed frequency similarity in topic nouns between CC and AC was not attributable to any particular measurement approaches. Thirdly, a prior sensitivity analysis comparing diffuse, weakly informative, and concentrated specifications returned consistent IRR estimates (median IRR: 0.9559–0.9562; maximum difference: 0.0003), with 95% credible intervals largely overlapping across all 3 priors. This result demonstrated that the observed equivalence of topic noun frequency between CC and AC was not a result of prior choices but instead reflected the empirical features of the data. Therefore, it was safe to judge that there was not significant difference in the lexical topic frequency of the news on Guangdong Province between China's and American newspapers.

In sum, the above robustness tests showed that CC and AC did not merely show no detectable difference in topic noun occurrence and frequency, but exhibited substantive equivalence within the predefined regions of practical equivalence, with 96% of the posterior mass for occurrence probabilities and 87% for frequency ratios falling within their respective ROPEs. China's newspapers are more proximate to Guangdong's news events than American newspapers across spatial and temporal dimensions, however, newspapers of the 2 countries exhibited substantive equivalence in the diversity and frequency of topic nouns in the news on Guangdong Province. Therefore, this study revealed an observational pattern consistent with the proximity-diversity paradox. While one country might enjoy greater spatial and temporal proximity to news events than another country, such advantages might not necessarily convert into greater lexical topic diversity in its relevant news coverage.

5 Discussion

This study employed the English news on Guangdong Province from China's and American newspapers, and utilized dependency parsing, Bayesian hierarchical Bernoulli logistic regression model, and Bayesian hierarchical negative binomial regression model, to assess whether the evidence was consistent with proximity-diversity paradox. The study showed that, although China's newspapers are more proximate to Guangdong's news events than American newspapers across spatial and temporal dimensions, the former did not exhibit greater lexical topic diversity and lexical topic frequency than the latter in coverage of Guangdong Province. Therefore, this study revealed an observational pattern consistent with the proximity-diversity paradox.

The present study departed from conventional null-hypothesis significance testing, where a non-significant p value merely indicated absence of evidence. Instead, the Bayesian hierarchical Bernoulli logistic regression model and the Bayesian hierarchical negative binomial regression model, coupled with ROPE-based equivalence testing, bootstrap resampling, prior sensitivity analyses, and alternative operationalizations of the outcome variable, provided positive evidence for the null hypothesis (i.e., evidence of no differences): the posterior distribution of cross-national differences was concentrated within a pre-defined region of practical equivalence, suggesting that CC and AC exhibited substantively similar topic noun patterns, a finding that reflected genuine convergence rather than merely a failure to detect differences.

This study revealed that there were no significant differences in lexical topic diversity and lexical topic frequency in the English news on Guangdong Province between China's and American newspapers, the results were similar to the findings of the research conducted by Golan (2006), Hendrickx (2020), Hendrickx and Ranaivoson (2021), etc. There was a trend of content convergence among media, which might be caused by shared sources, global news values, and mutual observation in the international media environment. However, this study analyzed lexical topic homogeneity in news across media from different countries, which might reflect the topic similarity in cross-national media coverage of a specific subject.

China's newspapers might enjoy greater spatial and temporal proximity to the sources of the news concerning Guangdong than American newspapers, intuitively, these advantages might give China's newspapers better access to richer and more detailed news resources. However, China's newspapers were in a controlled media environment (Guo, 2019), and Chinese government would strictly review the contents of the newspapers. The centralized control of news might exacerbate the lexical topic homogeneity of the news in China's newspapers. In the U.S., newspapers were mainly driven by the markets, they followed the commercial logic and strived to maximize profits (Pickard, 2019; Sparks et al., 2016). Therefore, they might produce news with diverse range of topics, so as to attract a large number of readers and earn substantial revenue (Katona et al., 2017; Li and Thorson, 2015). In sum, the spatial-temporal proximity advantages of China's newspapers might co-occur with conditions conducive to greater lexical topic diversity in Guangdong coverage, yet this potential relation might be tempered by the homogenizing tendency of a controlled media environment. Conversely, American newspapers, despite lacking such proximity advantages, might display diversification patterns aligned with market demands. The balancing of these contrasting dynamics might correspond to the observed absence of significant differences in lexical topic diversity between the 2 countries’ newspapers.

China's newspapers are more proximate to news resources concerning Guangdong Province than American newspapers across spatial and temporal dimensions, however, the former did not show greater lexical topic diversity than the latter in the news on Guangdong Province. Therefore, this study provided preliminary evidence consistent with the proximity-diversity paradox. In other words, people living in the area where news events occurred might not receive more information about those events than people living elsewhere. This paradox might reflect that in the networked age, media institutional barriers to news dissemination might outweighed those posed by spatial and temporal factors. The paradox might also embody news was not an objective representation of spatiotemporal events, but a reconstruction and reframing of reality under institutional power.

The proximity-diversity paradox could provide inspiration for the research on the concentration tendency in news coverage in the controlled media environment. It disclosed that despite media was proximate to the news resources across spatial and temporal dimensions, the news contents might remain relatively concentrated. This phenomenon might reflect the problems of news homogenization or viewpoint concentration in the media. In spite of living in an environment of news explosion, the topics people accessed remained relatively constant. This might cause the pseudo-diversification of news. Different news agencies might collaborate and share information with each other, and the news they obtained might originate from the common source materials, which led to content homogenization. In addition, news coverage of a specific subject might reach topic word saturation after a certain number of reports. Readers might not encounter new topics after consuming a certain volume of relevant news coverage.

The proximity-diversity paradox might reflect that the influence of spatial and temporal factors on the production and dissemination of news might be increasingly limited in networked era. However, the production and dissemination of news might be more influenced by the social factors, such as media institutional systems, controlled media environments and market demands, etc. News topics were the major aspects of reality. Since the diversity of news topics were influenced by social factors, these factors played a primary role in shaping reality. In sum, the proximity-diversity paradox might provide evidence for Constructivism (Jonassen, 1991), i.e., there was no absolute objective reality in the world, and media might function as a filtering lens that constructed specific reality for various communicative purposes.

In journalistic practice, the proximity-diversity paradox might demonstrate the media proximate to the news resources did not gather sufficiently diverse information to present the multiple dimensions of a news event, thereby limiting public understanding of the event. Hence, this paradox might make news professionals aware of their responsibility to broaden public cognitive boundaries. The extent of news contents about a subject might set the cognitive boundaries for public understanding of that subject. Therefore, news agencies should increase funding to encourage news professionals to gather information from diverse sources and perspectives on a given subject, so as to provide public with comprehensive information about that subject. In addition, newsrooms could establish news monitoring mechanisms for topic concentration and content diversity metrics, track the homogenization of news contents in real-time, reduce duplicate reporting, and avoid wasting editorial resources and audience attention.

This study investigated the relation between the proximity to news resources and diversity of the lexical topics of news, and revealed an observational pattern consistent with the proximity-diversity paradox. It might provide enlightenment for research on the topic concentration and content homogenization of news reports, and offer suggestion to the practice of solving the proximity-diversity paradox. The absence of significant differences in lexical topics between China's and American newspapers in Guangdong-related coverage was consistent with media homogenization, and this did not support the expectation that news agencies closer to the news events would produce topically distinct coverage from those operating at a distance.

However, there were some limitations to this study. The current study acknowledged that the spatial-temporal proximity of China's versus American news outlets to Guangdong was treated as a macro-geographic premise rather than operationalized through micro-level indicators, thereby precluding a direct empirical test of proximity's effects on lexical topic diversity. Meanwhile, it should be acknowledged that this cross-national comparison displayed the proximity-diversity paradox as an observational pattern (i.e., the lack of significant topic noun differences between China's and American newspapers despite their presumed spatial-temporal proximity disparities), rather than establishing a causal effect of spatial-temporal proximity on news topic diversity. In this study, the empirical scope constrained the generalizability of the reported paradox. With a sample of 39 Chinese and 37 American news texts examining a single region (Guangdong), where the American news corpus was skewed toward The Associated Press news, the findings offered an initial proof of concept derived from a specific bilateral comparison rather than a foundation for a universal theoretical claim. This paper therefore urged caution in generalizing the research results beyond the present empirical boundary and called for future replication across multiple regions, event types, national media systems, and longitudinal design. Additionally, analyzing topic diversity through dependency-parsed nouns might capture lexical variation rather than fully reflecting conceptual or semantic diversity, future research could triangulate the current findings through semantic embedding analysis, topic modeling, human coding, or complementary diversity indices to ascertain whether the observed lexical convergence corresponded to deeper conceptual convergence. Moreover, China's spatial-temporal proximity advantages in news sourcing might be counterbalanced by many institutional constraints, such as shared international news agency sources, translation practice, convergent editorial routines, governance structures, ownership models, and journalistic norms, whereas American newspapers might face opposing pressures from market dynamic and regulatory environments. These institutional factors were not operationalized or empirically tested in the present study, which might be good domains for future inquiry. Finally, public-sphere consequences of the observed lexical topic homogeneity remained theoretically extrapolated from textual data in this study, and awaited future audience-level empirical investigation.

6 Conclusion

This study utilized dependency parsing, Bayesian hierarchical Bernoulli logistic regression model, and Bayesian hierarchical negative binomial regression model, to analyze the differences in lexical topics of the English news on Guangdong Province between China's and American newspapers, so as to provide preliminary evidence consistent with the proximity-diversity paradox. The 2 research questions have been answered. Firstly, there were no significant differences in the lexical topic diversity of the news on Guangdong Province between China's and American newspapers. Secondly, there were no significant differences in the lexical topic frequency of the news on Guangdong Province between the 2 countries’ newspapers. Since China's newspapers are more proximate to Guangdong's news events than American newspapers across spatial and temporal dimensions, yet there were no significant differences in the lexical topic diversity and lexical topic frequency of the relevant news between the 2 countries’ newspapers, the current study offered preliminary evidence consistent with the proximity-diversity paradox. This study might provide a new methodology for the studies of topic concentration and content homogenization of news reports, and offer suggestion to the solution of the proximity-diversity paradox in journalistic practice.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Pradeep Nair, Columbia University, United States

Reviewed by: Muharman Lubis, Telkom University, Indonesia

Muhamad Fuat Asnawi, Universitas Sains Al-Qur’an, Indonesia