Vulnerable voices: using topic modeling to analyze newspaper coverage of climate change in 26 non-Annex I countries (2010–2020)

We selected lower- and upper-middle-income countries using World Bank (2022) income classification data, as they have traditionally lacked significant resources to address climate change. We did not select low-income countries due to limited data access throughout the study period (Okoliko and de Wit 2020). Using the Factiva and LexisNexis databases, and following prior studies (McAllister et al 2021), we searched for the terms 'climate change' or 'global warming' in 50 unique print news sources across 26 non-Annex I countries from 01 January 2010 to 31 December 2020. To select the 50 newsprint sources, we drew from the Media and Climate Change Observatory (MeCCO) 'World Table' (Boykoff et al 2023) and the Reuters Institute Digital News Report 2022 (Newman et al 2022). MeCCO has been cited extensively in the literature (e.g. Brüggemann and Engesser 2017, Kristiansen et al 2021, O'Neill 2020, Schäfer and Painter 2020, Thackeray et al 2020, Hase et al 2021). From these lists, we sampled the highest circulating sources that also included access to the entire print article. We focused on English-language media, which have been viewed as international drivers of public discourse and agenda-setting (Sonwalkar 2002, Khatri et al 2016). Using English-language print sources also enabled us to manually and qualitatively cross-check the findings (see topic modeling section below). Figure S1 and table S1 of the supplementary materials provide more details on our sample.

We used Latent Dirichlet Allocation (LDA) to uncover the major themes in our dataset. LDA is a probabilistic modeling technique for discovering the 'latent' topics in a large and unstructured collection of documents (Blei et al 2003, Griffiths and Steyvers 2004, Blei 2012), such as news articles, journals, blogs, and annual reports, among others. LDA assumes that a document is generated in a process that includes hidden distributions (e.g. the topic structures). Specifically, there are two distributions to be inferred. The first is the word distribution within a topic; the second is each document's topic distribution. The algorithm treats each word in a document (i.e. a news article in our study) as created through two generative probabilistic steps: (1) a topic is chosen among the set of topics with a priori distribution, and (2) a word is chosen from the selected topic, which is a distribution over a set of words. The topic distribution for each news article and the distribution of the words are estimated based on the set of documents. We manually read the top 100 articles (based on model-assigned topic probabilities) from each topic to verify the model's veracity and cross-validate the topic labels. More details on the models implementation and tuning metrics are in the supplementary materials (see methods S1, table S2, and figure S2).

After identifying the main topics, we examined the relationship between a country's vulnerability to climate change and the diversity of topics discussed by its media. We joined our data to annual measures of a country's vulnerability to climate change from the Notre Dame Global Adaptation Initiative (e.g. Chen et al 2015, Kling et al 2021). To measure our outcome variable (i.e. the diversity of topics discussed by a focal news article), we computed a Shannon entropy index ($H = - \mathop \sum \nolimits_{i = 1}^{i = 11} {p_i}*(\ln {p_i})$) and a reverse coded Herfindahl–Hirschmann concentration measure (HHI = −$\mathop \sum \nolimits_{i = 1}^{i = 11} p_i^2$), where p represents the probability that a given news article belongs to a topic, and i is one of the 11 topics in our LDA model (also see articles S1–S2 in the supplementary materials). We then executed multivariate regression analyses using (1) ordinary least squares estimates at the news article level and (2) a fixed effects specification with an unbalanced panel, organized by newspaper source and the month that an article was published as the time identifier (132 time periods, i.e. 11 years × 12 months). We also included controls at both the news article (word count, sentiment, average sentence length, and % of unique words) and country (readiness to adapt to climate change (Chen et al 2015), population, gross domestic product (GDP) per capita, and human development index (HDI)) levels. Standard errors were clustered at the news source level.

Figure 1 (panel (A)) shows the average probability (as a percentage) of the occurrence of a topic in a news article. The three most commonly occurring topics are international governance & development (14.45%), the economics of energy transitions (10.86%), and the impacts of climate change (10.70%). These three topics account for over one-third (36%) of the sample topic distribution. The overall distribution is relatively uniform, albeit with a high level of variability as shown by the error bars. Exemplar excerpts in the supplementary materials (see table S3) qualitatively illustrate these topics.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Figure 1 (panel (B)) shows an intertopic distance map, with two distinct clusters. The first 'socioeconomic' cluster (in green) focuses on energy economics, agriculture, and activism & collective action. The second 'sociopolitical' cluster (in blue) focuses on international governance & development, national governance, and education-related events about climate change. Discussions about climate science (impacts of climate change) co-occur more often with socioeconomic vis-à-vis sociopolitical topics. Four topics are spatially isolated (national political elites, culture, US politics, and health) indicative of their more peripheral and/or discrete (vis-à-vis interconnected with other topics) media coverage.

Countries in table 1 are organized by five geographic regions, using World Bank definitions. Topics in table 1 are organized from left to right (same order as figure 1), and the three most prevalent topics are highlighted in grey. One-way ANOVA tests at both the country and region level showed that mean differences were statistically significant. We observed that the media in Sub-Saharan Africa pays a significantly higher level of attention to agriculture relative to other regions. The media coverage in South Asia, which has the most vulnerable countries in our sample (standardized average = 0.83; last column of table 1), also has a relatively unique focus. It emphasizes the impacts of climate change, educational events, and national political elites (likely due to the comparatively low focus in India on international governance).

The inter-topic distance map in figure 2 uses the country-level topic averages in table 1. We observe that the coverage in South Asia and East Asia is more homogenous relative to the middle East or Sub-Saharan Africa (i.e. the South Asian and East Asian countries form smaller clusters). Interestingly, coverage in South Asia is more similar to that in Sub-Saharan Africa, relative to that in East Asia & Pacific. Note that the eight most vulnerable countries in our sample were Bangladesh, Pakistan, Kenya, Tanzania, Zimbabwe, India, Nepal, and Nigeria (all in the highest quartile of the ND-gain global index, >75% relative to the rest of the world). We observe that they cluster together (shown in red), indicative of an association between media coverage and country climate change vulnerability. We also note that South Africa and Vietnam appear as 'outliers' (i.e. distant from other countries in their region). This is consistent with their disproportionately high focus on topics such as energy economics and collective action respectively (see table 1).

Zoom In Zoom Out Reset image size

The average vulnerability for each country (see last column of table 1) is positively moderately correlated with the diversity of topics in an article (r = 0.44). In table 2, we examined this correlation more systematically through multivariate regressions. Model 1 and Model 2, use the entire data sample (N = 95 216). A 1 standard deviation increase in country climate change vulnerability is associated with ∼0.25 standard deviation increase in the topic diversity of the average news article in that country (β_{model 1} = 0.27, p = 0.01; β_{model 2} = 0.23, p = 0.01). In Models 3 and 4, we repeated the analyses after collapsing (averaging) the data and creating an unbalanced panel (N = 5155) at the news source-observation month level. Again, we find a positive and statistically significant relationship (β_{model 3} = 0.31, p = 0.01; β_{model 4} = 0.20, p = 0.06) between a country's climate change vulnerability and the diversity of topics discussed by its media. We carried out several sensitivity analyses (e.g. removing the bottom 15 sources which had less than 500 articles, excluding India, restricting data from India to one or two sources, excluding 'outlier countries' such as Ukraine, Uzbekistan, and South Africa, omitting highly correlated (with country vulnerability) control variables such as GDP per capita and/or HDI, using a 1 year time lag in the models) and found robust results. We thus find a positive association between country-level climate change vulnerability and the diversity of climate topics covered in relevant news articles (see article S1 and article S2 in the supplementary materials for illustrative examples).

We also carried out a supplemental analysis (see supplementary materials, table S4) where we regressed country vulnerability on the probability of occurrence of each topic as separate dependent variables (instead of computing a single entropy measure). A country's climate change vulnerability is significantly negatively associated with a focus on international governance (β = −0.43, p = 0.01), and US politics (β = −0.49, p = 0.03). It is significantly positively associated with coverage of national governance (β = 1.5, p = 0.004) and national elites (β = 0.62, p = 0.04). Thus, we find tentative evidence that an increase in a country's vulnerability to climate change is associated with a shift in media coverage towards domestic (or internal) issues vis-a-vis external matters. We provide exemplar articles to illustrate this pattern (see supplementary materials articles S3–S6).

Our study supports and extends findings of recent analyses of media reporting on climate change. For example, Hase et al (2021) found that Global South media coverage focuses on the societal dimension of climate change, including increased reporting on economics, climate politics, and human impacts. Vu et al (2019) found that poorer countries report more on 'international relations and the natural aspects of climate change' (p 7). Using a much larger sample of countries from non-Annex I countries, we found that the three most discussed themes were international governance & development, the economics of energy transitions, and the impacts of climate change. We also find significant cross-regional differences (e.g. between South Asia, sub-Saharan Africa, and East Asia). The importance of international politics and the impacts of climate change, such as sea level rise and flooding, are thus highlighted across all studies. These findings are also consistent with the fact that many non-Annex I countries often depend upon resources from Annex I countries to mitigate and adapt to the consequences of climate change (IPCC 2022).

Interestingly, both our study and Vu et al (2019)'s findings do slightly differ from those of Hase et al (2021), who did not find an increase in extreme weather events in their analysis. One potential reason for these differences could be the vastly different sample sizes; both Vu et al (2019) and our study (N = 26 non-Annex I countries) have a significantly larger sample size than Hase et al (2021), which only examined four countries from the Global South. The studies also used marginally different topic labels. For example, Hase et al (2021) categorize rising sea levels and flooding as 'Impacts on humans', but distinguish these from extreme weather events. Lastly, both our study and Hase et al (2021) highlight the importance and interconnectedness of economic issues in more vulnerable countries, unlike Vu et al (2019). More generally, the comparison of our findings with Vu et al (2019) and Hase et al (2021) reinforce the importance of studying the intersections between different themes. Our study highlights how non-Annex I countries are taking a comprehensive approach to climate change media coverage, interweaving topics such as international governance, economic transitions away from fossil fuels to cleaner energy sources, and impacts (e.g. flooding, extreme weather events) of climatic changes (see figure 1, panel B). This finding also aligns with scholarly calls to explore systemic institutional solutions to alleviate climate related challenges (Kalyango et al 2017, Nassanga et al 2017, Chattopadhyay 2019).

We also show that there is a significant positive association between countries' vulnerability to climate change and the diversity of topics discussed in their media coverage (table 2). Many studies discuss the interrelated nature of climate change impacts; our study systematically quantifies this phenomenon in non-Annex I countries. Our study thus adds nuance to the literature by highlighting the localized impacts of climate change across the Global South (Ejaz et al 2022). We also highlight the need for more partnerships and knowledge transfer between South Asia and Sub-Saharan Africa, as these regions share similar climate coverage and climatic challenges, relative to East Asia & Pacific.

Most importantly, our findings demonstrate that in the most vulnerable countries, the media explicitly reports on the complexity and multifaceted nature of climate change. By speaking to the social, economic, scientific, and political aspects of climate change, news coverage in the most vulnerable countries is highly sophisticated. Our findings call out a Western bias and colonial lens by refuting a reoccurring narrative in the climate media literature that journalism in the Global South is less robust due to resource constraints. The Western bias is further reflected by the dearth of scholarship on the media coverage of climate change in less developed countries. For example, Okoliko and de Wit (2020) show that scholarship on media coverage of climate change has only been examined in nine African countries. Our study shows that, even while developing countries may have fewer resources or training to inform climate coverage (Ajaero and Anorue 2018), reporting in the most vulnerable countries is highly nuanced and distinctive from common Western conceptualizations of climate change (Olausson and Berglez 2014).

Intuitively, the implications of our findings make sense as we see that as the most vulnerable countries experience the urgency and worst impacts of our changing climate; the co-occurrence of issues and interconnectedness of climate change is forefronted and reflected in media reporting through the diversity of topics covered in its articles. In the most vulnerable countries, climate change information is not siloed, but addressed systemically. This also tracks the objectives of developmental journalism as vulnerable countries work on developmental, interventional, and educational objectives in the context of lived climate impacts (Kalyango et al 2017). This contrasts with global media coverage, which still struggles to make explicit connections between two topics, such as climate change and health (Romanello et al 2022). Our findings emphasize the opportunity for the Global North to learn from media coverage of climate change in the Global South where the problem is most acute.