Corporate funding and ideological polarization about climate change

To empirically examine these data in relationship to the research questions above regarding the influence of corporate funding on textual content, the study employs a combination of social network analysis and a form of large-scale computational text analysis called latent dirichlet allocation (LDA) (30) topic modeling. Generally stated, topic modeling is a computer-assisted content analysis procedure whereby a set of texts are coded into substantively meaningful themes called “topics.” These topics are not given to the machine beforehand, but emerge inductively as algorithms learn the hidden patterns underlying a collection of texts. The model assumes a relational theory of meaning by measuring the patterns of co-occurrence of words in individual texts and across the entire corpus. Thus, in addition to providing a way to conduct reliable content analysis on massive collections of text that are too big to code by hand, topic models allow researchers to use machine learning to discover patterns and relationships that may have otherwise been missed by hand coding methods. Recent research has shown that unsupervised methods like the ones used here have performed as well as human coders on the same set of documents (31).

More specifically, this paper uses a recently developed approach to topic modeling called Structural Topic Modeling (STM) (31, 32), because it enables the discovery of topics and their prevalence based on document metadata, such as the year written, or important organizational attributes examined here, such as corporate funding. The inclusion of metadata such as the year in which a document was published is especially useful given that these text data cover such a broad period (1993–2013), and given that the climate change discourse during this period was susceptible to thematic change. Building on the standard LDA model (30), Roberts et al. (31) explain, importantly, that:

As in LDA, each document arises as a mixture over K topics. In the STM, topic proportions (θ) can be correlated, and the prevalence of those topics can be influenced by some set of covariates X through a standard regression model with covariates $\theta \sim LogisticNormal\left(\text{X}\gamma ,\mathrm{\Sigma }\right)$. For each word (w) in the response, a topic (z) is drawn from the response-specific distribution, and conditional on that topic, a word is chosen from a multinomial distribution over words parameterized by β, which is formed by deviations from the baseline word frequencies (m) in log space ( ${\beta }_{\text{k}}\hspace{0.17em}\alpha \hspace{0.17em}exp\left(\text{m}+{\kappa }_{\text{k}}\right)$)…Thus, there are three critical differences in the STM as compared to the LDA model…(1) topics can be correlated; (2) each document has its own prior distribution over topics, defined by covariate X rather than sharing a global mean; and (3) word use within a topic can vary by covariate U. These additional covariates provide a way of ‘‘structuring’’ the prior distributions in the topic model, injecting valuable information into the inference procedure.

The STM method is applied to the totality of text data described above to (i) map the entire thematic landscape of content produced by the climate contrarian movement, and (ii) plot covariate interactions using text metadata to examine how corporate funding ties influence topic prevalence and change over time. It is important to note that while these are automated computational methods, they require a deep human understanding of the corpus that guides the number of topics that are estimated, and a process of recursive interpretation of the model’s results based on prior findings (see SI Appendix for important details about topic model estimation and validation).

Taken together, these analyses develop a quantitative account of all public information produced by the climate change countermovement. Most importantly, they assess the substantive content of information within the organizational structure of the contrarian network, providing important analytical leverage to understand the ways that private corporate funding influences the actual content and prevalence of polarizing discourse.

To begin, Fig. 1 provides a broader bird’s-eye-view context with which to interpret the main text analysis findings below. The graph describes which organizations (nodes) produced texts and whether the organization received corporate funding. For example, a dark green circle node is an advocacy or think-tank organization that produced texts about climate change and received corporate funding. It is a one-mode network constructed from the original bipartite social network data, where ties between organizations are a function of a tie they share with individuals (see SI Appendix for details). To give a sense of change over time, the graph includes cross-sections from each decade that restricts the nodes and ties to organizations founded before that year. This approach means that 2013 is the final graph that includes all nodes and ties in the data.

There are two related findings that emerge from this network analysis and serve to inform the main text analysis below. First, organizations who produce climate contrarian texts (versus those who did not) are significantly (P $<$ 0.006) more likely to have higher betweenness centrality scores, which is a reliable indicator that they exert more organizational influence in the network (29). One can spot this trend in the figure by recognizing that these nodes tend to cluster toward the center of the network. Second, organizations who produced texts and received corporate funds were also significantly more likely (P $<$ 0.002) to have higher betweenness centrality scores. Taken together, these results suggest that organizations within the movement who made an effort to produce textual discourse about climate change are the most central to the movement itself, providing them more influence over the transfer of information. This finding aligns with prior knowledge about resource mobilization (33), discursive influence (34), and corporate mobilization (22). The fact that the production of discourse was more likely to be corporately funded organizations at the core of the network provides an important backdrop for the novel analytic focus of this article—namely, assessing via computational text analysis the influence of corporate funding on the actual thematic content and ideological polarization of the discourse itself.

Turning to the results of the computational text analysis, Fig. 2 summarizes the results of the structural topic model. This corpus level visualization displays all of the topics along with the top 15 words associated with each topic. This figure also includes the proportion of the corpus that belongs to each topic. It is important to note that these topics and the words should be interpreted in light of the purpose of the texts and the larger goals of climate change contrarianism described above (8–15). Broadly, one can see that these organizations focused their climate change discourse on people’s knowledge of the issue, Intergovernmental Panel on Climate Change science, debates over temperature trends, international politics such as the Kyoto Treaty, and related energy issues such as fuel, oil, and cap and trade bills.

How are all of these topics related to one another, and do they cluster together in meaningful ways? Fig. 3 plots the layout of all topic correlations using the Fruchterman–Reingold algorithm. A tie between two topics indicates a positive topic correlation, meaning that both topics were more likely to be discussed together within a document. Node sizes correspond to the topic proportions from Fig. 2. The dashed ellipses were overlaid after graphing to draw attention to four thematic clusters: First, toward the bottom left (in green) is a cluster of topics relating to questions about scientific evidence and disputes over the real causes and effects of CO₂, long-term temperature cycles, global cooling, and who has scientific authority. Adjacent (in yellow) is a second cluster pertaining to public knowledge and the role of news media and Al Gore’s film. Just above this (in blue) is a cluster of topics pertaining to federal and state-level bureaucratic politics related to climate change. Lastly, toward the top (in black) is a cluster of topics related to energy industry concerns, and alarm about the economic and political costs of enacting climate change policy. These four broad clusters provide a comprehensive profile of all discourse produced by climate contrarian organizations from 1993 to 2013.

Lastly, Fig. 4 examines how the actual prevalence of these topics might be influenced by corporate funding. It focuses on four topics that have been particularly salient to the recent history of climate contrarianism (for all topic plots, see SI Appendix, Fig. S1). These interaction plots allow the prevalence of the topics to vary by time (by year, 1993–2013) and whether an organization received corporate funding during this same timespan (Red Line = Yes, Black line = No). For example, the plot in the top left demonstrates that temperature trends became much more salient between 2007 and 2013 for organizations who received funding, whereas they became less salient for organizations who did not. Similarly, the issue of energy production (top right), the positive benefits of CO₂ (bottom left), and climate change being a long-term cycle (bottom right) all showed distinctly positive trends over time for organizations who received corporate funding, whereas those who did not receive corporate funding remained more constant over the same time period.

On the whole, these findings show that despite being part of the same contrarian network, the organizations who received corporate funding actually produced discourse that was qualitatively different from organizations that did not receive such funding, and these differences tended to revolve around energy production, debates about the effects of CO₂, and questions about the scientific veracity of long-term climate change.

The main empirical and theoretical contribution of this analysis is that corporate funding influences the actual language and thematic content of polarizing discourse. These effects were visible over time, as the prevalence of certain thematic content shifted within the 20-year span of textual data. A secondary finding of this analysis is that organizations that received corporate funding were more likely to have written and disseminated contrarian texts. Taken together, these findings are especially important because they contribute comprehensive empirical evidence to confirm what has widely been thought to be the case about climate change knowledge and politics, but has heretofore not been demonstrated scientifically with robust data.

While these findings are fundamental to understanding the past and present of climate change policy in the United States, they also have much broader implications for understanding ideological polarization itself. It is well understood that polarization is an effective strategy for creating controversy and delaying policy progress, especially around environmental issues. However, what is less understood are the complex organizational and financial systems that affect the creation of ideological polarization in the first place. Thus, a broader contribution of this paper is to uncover empirically the actual social arrangements within which large-scale scientific (mis)information is generated, and the important role private funding plays in shaping the actual ideological content of scientific information that is written and amplified.

A notable limitation of these findings concerns the furtive nature of modern foundation funding and political action committees (8, 9, 11, 22), which creates a causal dilemma about whether corporate funding leads to the increased production of discourse, or whether the organizations already creating discourse attracted corporate funding. Of course, the strong correlational evidence provided here cannot untangle with certainty all of the microlevel causal processes, but it can provide traction to reject competing explanations. Most importantly, supplementary analysis revealed few discursive differences between organizations that received money before producing discourse (e.g., “front groups”) versus organizations who received it later (e.g., established think tanks). Thus, the observed effects of money on discourse cannot be explained by recourse to one static temporal model or another, but funding influenced organizations who were already writing texts, as well as organizations who were established with the help of such funding. The organizations who were already in operation also created new texts, established and strengthened a relationship to promulgate shared interests, and ultimately amplified the messages that were central to the countermovement. Investigating how this happens at the microlevel is beyond the scope of this paper, and the data cannot address such questions. Future research should continue to develop these lines of inquiry about the multiple ways in which funding operates. Continued social-scientific research is also needed to further develop the scholarly understanding about the sources and influences of polarization on sustainability efforts—both presently with climate change, but also for predicting and preempting future socio-political conflicts from becoming fraught with uncertainty, skepticism, and misinformation.

Lastly, this paper provides a novel analytical model for future research to rigorously capitalize on the rise of the Internet and the unprecedented flood of available social scientific data, much of which is (i) textual and (ii) networked in some way. Unfortunately, much of the early analyses of large amounts of naturally occurring and digitized text data were first restricted to experts with the technical expertise to collect and organize it (e.g., computer scientists), resulting in “big” data content analyses that lacked theoretical depth or meaningful focus (35). This paper provides a model that focuses simultaneously on the fine-grained discursive content of texts and at the same time the ways in which these texts are embedded and produced within larger, complex, and densely connected socio-political networks. This analytical approach, as modeled here, allows researchers to examine how discursive and ideological content varies within and across organizations in a relational field, with a specific focus on the influence of covariates such as external funding and other organizational attributes (e.g., assets, mission focus, location, board interlocks). Thus, combining cutting-edge machine learning techniques with the long history of network science approaches, results in a robust empirical framework that allows researchers to situate fine-grained discursive results within the larger web of political and economic relationships that will continue to influence science and policy into the future.

Comprehensive information about the data collection, cleaning, and statistical analyses are provided in SI Appendix.