Literature review

Big Data analytics techniques have been widely used to evaluate environmental management and sustainability problems in order to uncover hidden trends and patterns in public opinions. Climate change denial often hinges on disputing temperature data, with some arguing that Earth’s climate has historically fluctuated naturally [10].

Climate change is a pressing global issue with significant implications for society, and social media platforms have become instrumental in facilitating discussions around it [11]. Scholars have recognized the role of social media in shaping public perceptions, attitudes, and behaviors towards climate change [12]. Platforms like Twitter, Facebook, and Instagram serve as spaces where individuals share information, express opinions, and engage in dialogue about climate-related topics [13,14].

Studies have highlighted the potential of social media to amplify climate change discourse and foster public engagement [15–17]. For instance, research by [18] found that Twitter discussions on climate change often include a diverse range of voices, from scientists to activists to policymakers. Furthermore, social media enables rapid dissemination of climate-related news and information, allowing users to stay informed and connected to developments in real-time [14].

However, scholars have also identified challenges associated with climate change discourse on social media, including the spread of misinformation, polarization of opinions, and echo chambers [19]. This is further advanced by other scholars who have explored polarization in digital media environments such as climate change debates on Twitter [20–26].

Skeptics may question the significance of current temperature increases or attribute them to cyclic patterns unrelated to human activity. Cherry-picked data or misinterpretation of climate models is common tactics to cast doubt [27]. Such denial obstructs efforts to address climate change, despite overwhelming scientific consensus on its reality and human causation [28]. Recognizing and countering denialism is crucial for implementing effective climate policies and mitigating its far-reaching impacts.

Despite methodological and skeptical challenges, social media platforms remain critical spaces for engaging diverse audiences in discussions about climate change and mobilizing collective action towards mitigation and adaptation efforts [14].

A majority of the data used in these analyses comes from the microblogging service such as Twitter [29]. Several aspects of climate change through human opinions have been studied over the past years by researchers. The early studies of climate change on Twitter by [30,31] considered whether important events affect the discussion, and identified three groups of human stance towards climate change, namely supportive, unsupportive and neutral. [32] reported correlation between human sentiment and both hot and cold temperatures, underlining the importance of temperature as an important variable in climate change sentiments. Skeptics may question the significance of current temperature increases or attribute them to cyclic patterns unrelated to human activity. Cherry-picked data or misinterpretation of climate models is common tactics to cast doubt [27]. Such denial obstructs efforts to address climate change, despite overwhelming scientific consensus on its reality and human causation [28]. Recognizing and countering denialism is crucial for implementing effective climate policies and mitigating its far-reaching impacts.

Resource overconsumption and human intervention profoundly influence the climate change debate, often exacerbating contradictions between deniers and believers [33]. Overconsumption, particularly of fossil fuels, accelerates greenhouse gas emissions, fueling climate change. Deniers, often backed by vested interests, may downplay these impacts to protect economic interests [34]. In contrast, believers emphasize the urgent need for sustainable resource management and emissions reductions to mitigate climate change’s adverse effects. However, societal dependence on resource-intensive lifestyles and industries complicates efforts to enact meaningful change, perpetuating the contradiction between recognizing climate change’s reality and addressing its root causes. The differences in opinions of climate change believer and denier groups exist across many variables including gender. Gender is reported to play a role in attitude towards climate change. While both genders use very similar language, female tweeters had a convinced attitude and male tweeters presented a skeptical stance [35] towards impact of climate change. Finally, another aspect of concern for many researchers is the topics discussed in climate change related tweets [36].

Twitter serves as a crucial platform for political discussions on climate change, and the data derived from it is frequently utilized in the analysis of non-state climate action and public opinion on social media.

Therefore, there exists significant potential to enhance our understanding of how actors within the climate governance complex engage on Twitter, interact with the general public, and are influenced by these interactions. A deeper comprehension of climate debates on Twitter can contribute to climate governance research and advance theories on the impact of social media, including norm diffusion, opinion leadership, and the formation of citizen and elite opinions, on climate governance.

Data and methods

I have used twitter data available on Kaggle [37]. It is claimed to be one of the most extensive dataset available on Twitter that relates to human attitudes and climate change. With over 13 years of temporal coverage, over 15 million tweets geographically dispersed worldwide. Each tweet is analyzed across seven informational dimensions: geolocation, user gender, aggression, climate change attitude and sentiment, deviations from historical temperatures, topic modeling, and data on environmental disaster occurrences. The data used by [37] suggests that these dimensions are the result of testing and analyzing numerous cutting-edge supervised and unsupervised machine learning algorithms and techniques.

The previous studies have heavily relied on text analysis and descriptive techniques of countrywide variation of sentiments and tweets about climate change and have produced useful information. These studies have used text analysis approaches to uncover the pattern and variations of climate change discussion across the globe. In order to have significant insights across the vital opposite groups (believer and denier) of climate change, a robust analytical model is required. Given the three stances, suitable model choices are ordered logit or multinomial. Respecting the opinions of individuals, it will be highly appropriate to not to rank the groups i.e. to make any order of three groups. In this case, it is useful to go for the multinomial model that does not consider any order between the outcome categories. Therefore, in this way, adding to the contributions discussed in the literature review section, I will use multinomial logistic regression model to assess the relationship between the dependent variable (stances: believer, neutral or denier) and the key independent variables. Where the independent variables are temperature, gas emission, impact of overconsumption, tweet language (aggressiveness) and gender. These variables are given in the dataset and are categorized based on the expression of individuals on the twitter. Thus, these categorized expressions will help understanding the likelihood of individuals taking their stance (believer, or denier, or neutral) for the climate change. While all the aforementioned independent variables are important, for the purpose of this paper, I have maintained a focus on gender during detailed interpretation of result.

To assess the relationship between the qualitative outcome variable (stances of climate change by individuals via tweets), I have used the multinomial logistic regression represented by the following equation.

The logistic model is a generalization of the binary outcome of standard logistic regression comparing each category of the outcome (climate believer and denier in this case) to a referent category (neutral). The equation assumes, there are J total categories of the outcome, indexed by the subscript j, and the number of comparisons is then J– 1. The equation is written in terms of the logit of the outcome, which is a comparison of a particular category to the referent category, both denoted as π_j in the equation. The j subscript on both the intercept, A_i, and slope, B_j, indicate that there is an intercept and a slope for the comparison of each category which is a generalized case of binary logistic model.

The predicted probability for each predictor can be computed by generalizing the above equation for standard logistic, using the following equation.

The model is used by earlier studies on climate change in the context of agriculture [38] and is the best choice for this data as the outcome has three unordered categories- believer, neutral and denier and there are sufficient numbers of cases across the categories over such a large dataset (millions of tweets). Conceptually, the model selection is also based on the assumption that no one is considered superior in holding their stance towards climate change and hence the stances cannot be ordered.

Before running the model, I omitted the year 2006 as it contained only four valid observations after dropping the bank cells and also dropped the ‘undefined’ category of gender. After the cleaning, the values in data were left with the year 2006 to 2017. A descriptive summary of variables used in the analysis is given in Table 1 ahead. I have used RStudio software and nnet package [39] to conduct multinomial regression.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Displays the descriptive of key variables summary of the categorical variables used in the model.

https://doi.org/10.1371/journal.pone.0303007.t001

The table indicates low number of cases in two categories (ideological positioning on global warming and impact of resource over consumption) of variable ‘Topics’. In all the other variables, at least 10% cases are present within a variable. The % distribution of cases across the variable categories suggest sufficient number of cases to run the multinomial model with no blank cells as those were dropped from the data in cleaning process. The first category of each variable was held as referent to interpret the results. The model converged in 20 iterations and the significant chi-square test (33936, P<0.00) shows the statistical significant of the model [40]. The results are discussed in the following section.

Limitations

The study is not without limitation. One of the most widely accepted critique of twitter data is the identity of tweeting individuals. We are not 100% certain about the correct information about a bot, or an individual on twitter. So, we cannot be sure if the gender data is correctly provided to the twitter by the users. However, given the large scale data context, we assume this would not create a big difference on results. Therefore, we may hope that the conclusions are generalizable. The data only contained binary gender. So, the results are limited to male and female only. The other limitation was not using the location data as it would generate a long list of countries. This would add complexity in model that would create distraction and distortion in the interpretation of the results of the existing multinomial egression being used and hence, using the location data is beyond the scope of this study.