Utilizing big data to measure key connections between gender and climate change¹

5.1.1Logistic regression analysis

The study considered five gender-related indicators [26] as dependent variables and seven climate change related factors as key independent variables (refer Table 1). Variables were selected based on existing literature, relevance, and availability of comparable data across the five countries. Since many other socio-economic factors are likely to influence gender outcomes, eight control variables were included in the model.

All dependent variables were constructed as dichotomous categorical, and they all have a negative connotation (e.g., increases in the prevalence of each phenomena represent worsening of women’s wellbeing). Binomial logistic regression analysis was thus determined best suited for this analysis. Multivariate logistic regression was used to observe the effect of all independent variables, including climate change related and control variables, on gender-related outcomes.

The final model included the control variables in addition to the climate-related variables. This resulted in the adjusted odds ratios. The Residual Deviance and Akaike Information Criterion values were compared for these models to evaluate the goodness of fit [27].

5.1.2Random forest models

Random forest predictive algorithm (a machine learning technique) focuses on prediction instead of association. These models are thus powerful when dealing with complex and non-linear associations. Random Forest is a tree-based learning algorithm that uses a collection of decision trees to perform classifications tasks, with the primary objective of increasing the accuracy of the predicted outcomes, given the input data. In the context of this analysis, random forests compute multiple decision tress for classifying women into high or low risk of experiencing child marriage, adolescent births, IPV, lacking basic water and clean fuels, given their exposure to different climate variables. As a result, the model identified the climate variables most important in predicting gender-equality outcomes (measuring the mean decrease in the accuracy of the model if each of the climate variables was removed).

5.1.3Limitations

Due to the localized nature of environmental events (e.g. different magnitudes, duration, and intensity) and various degrees of coping capacity among exposed population (depending on asset ownership, existing infrastructure, suddenness of events), making cross country generalizations is difficult. Furthermore, the choice of variables included in the models was restricted by data availability, so future research should control for the effects of asset ownership, quality of infrastructure, or whether weather events took place out of season, to ascertain the sole impact of climate variables.

Findings from the random forest and logistic regression models do not always pinpoint to the same climate variables as the most relevant for explaining changes in gender outcomes. This is because they each serve a different purpose. While logistic regression identifies the odds of an increase in a gender-related outcome given a one unit increase in a climate variable, random forest highlights the reduction in the model’s accuracy if a climate related variable was removed from the model. In addition, while logistic regression performs well in models where explanatory variables have large variances, random forest returns higher true positive rates for datasets with numerous noise variables [28].

An additional limitation has to do with data timeliness [29]. In most countries, the reference year for the latest GIS data and survey microdata differ, and thus it is impossible to study direct impacts of select weather events on a population group. Rather, the analysis aims to ascertain whether geographical areas and population groups that typically suffer from slow-onset or recurrent effects of climate change, are more likely to experience changes in gender outcomes overall. Also related to timeliness, since some climate related phenomena change rapidly, more recent data would provide a more accurate picture of the current situation.

5.2.1Limitations

Since this analysis utilizes online VAWG-related searches only, values should be interpreted with caution. While search trends are expected to mirror trends in actual incidence of violence to a certain extent, they cannot be used to assess the incidence of VAWG with accuracy. The overall number of searches, in addition, is highly dependent on varying degrees of internet penetration across the countries considered, which hampers comparability in terms of total search volume. Furthermore, in contexts where internet penetration is low, it is likely that those with access to the internet are the more empowered women, and thus VAWG cases identified through this analysis are likely severely underestimated.

5.3.1Quantitative analysis: Detecting gender references

COP speeches available in English were first screened for gender references. The identification of gender references was based on a list of keywords developed by the authors. Words such as ‘mother’ and ‘ggi’ [30] are relevant both from a gender and a climate change perspective but were found in the speeches only in the context of climate change. Hence, they were not considered as gender-relevant to avoid overestimation. Algorithm allowed for words such as ‘gender’ as well as other words that appear within the former (e.g., ‘engender’) were all considered. Furthermore, certain words that would only be of relevance to gender equality and women’s empowerment when used in combination with other words (e.g., ‘birth’ to ‘adolescent birth rates’), were also considered only as word strings.

The frequency of sentences with gender references was calculated, instead of counting the number of gender-related words. This avoids inflating the results of speeches that use more gender references rather than utilizing gender references more consistently.

5.3.2Qualitative analysis: Topic modelling and sentiment analysis

To assess the degree to which each of the speeches covered gender issues substantively, two methods were used. Firstly, through topic modelling, a set of abstract topics were identified built on frequent words that seemingly point to certain topics. After identifying the sentences with gender references, these were processed to convert into lower case and remove stopwords [31]. Latent Dirichet Allocation (LDA) technique was used to generate speech-word joint probabilities, that is, the Bayesian probability for each word to fall in each of the abstract topics was calculated, based on the prior distributions introduced over speech-topic and topic-word distributions. The words with the highest probabilities within each topic were shortlisted and grouped into clusters. Based on the group of words produced by the model, topics were identified normatively.

Secondly, to assess the sentiment of the gender related sentences, R software packages Bing and Loughran were utilized. The analysis of the content is ensured to contain stopwords, to detect negative sentiments.

6.1.1Logistic regression analysis

Findings from the logistic regression analysis indicate that increases in the frequency of drought episodes, rises in temperatures and increasing relative aridity exert a statistically significant and negative effect on many gender outcomes (Fig. 3). In particular, temperature rises are associated with more prevalent child marriage and adolescent births across most countries [32]. More frequent drought episodes are linked with lack of access to basic drinking water sources (Bangladesh, and Cambodia), and lack of access to clean fuels for cooking fuels (Bangladesh, and Philippines) [33]. Increases in relative aridity (low aridity index) [34] also correlate with worsening gender outcomes (child marriage and adolescent birth rates in Bangladesh and Nepal), although these effects are smaller in size in all countries except for Timor-Leste (partner violence, access to basic water sources and use of unclean fuels increase substantially when relative aridity increases).

Figure 3.

Adjusted odds ratios (p< 0.000) for the effect of drought, aridity (index), and day land surface temperature across gender outcomes, by country. Note: For brevity, the graph represents adjusted odds ratios for associations significant at p< 0.000. The dotted horizontal line represents an odds ratio of 1 (to represent no association). The 3-digit ISO country codes stand as follows: KHM (Cambodia), BGD (Bangladesh), NPL (Nepal), PHL (Philippines), and TLS (Timor-Leste).

Drought episodes: Increases in the frequency of droughts enhance the likelihood of negative gender outcomes in most countries. Differences across locations can be explained by factors such as whether the region is arid or semi-arid, whether drought episodes are short or long-term, whether the effects are localized or widespread [35, 36], and the overall dependence on agriculture [37]. The strongest associations for drought episodes, across countries, were found with access to basic drinking water sources and clean cooking fuels. This carries gendered consequences as it may exacerbate women’s time burdens for fetching water, cooking and providing hygiene for the household. Although in Nepal the relationship appears inverse (more frequent drought episodes correlate with higher likelihood of accessing basic water sources (AOR = 0.67*⁣**)), geospatial analysis reveals that drought-prone clusters are largely located in or near Kathmandu, where water provision infrastructure is better established. Further analysis with controls for availability of water infrastructure would be useful to establish the sole effect of droughts.

In Bangladesh and especially the Philippines, the associations between frequent droughts and lacking access to clean cooking fuels see some of the largest coefficients in this study (Bangladesh AOR 1.28*⁣**; Philippines AOR 1.63*⁣**). Fuel cost, among other factors, may contribute to this association. Water scarcity limits energy production by hydropower and coal plants, which are highly dependent on water for cooling, which may rise energy prices. Further, drought-driven economic stresses from reduced agricultural yields may push households to rely on cheaper (unhealthy) cooking fuels. In Cambodia the inverse association (AOR 0.79*⁣**) is likely influenced by provincial differences in electricity prices, driven by grid development, availability of high-voltage transmission connections, existence of hydropower dams and other factors influencing provincial tariffs [38].

Relative aridity: Relative aridity [39] is also significantly associated with gender-related outcomes, with smaller effects across countries except in Timor-Leste, where clear associations exist with the prevalence of intimate partner violence (IPV), access to basic water sources and use of clean cooking fuels. This is a concern given that Timor-Leste already sees the highest prevalence of IPV among the five countries considered (an estimated 34.6 per cent of women experienced physical and/or sexual violence at the hands of their partner in the 12 months preceding the survey). A prolonged history of conflict, widespread poverty and stress associated with bride price, may contribute to the high IPV prevalence in the country [40, 41]. Although further analysis is necessary, it is possible that aridity-related economic stresses may be worsening poverty, increasing pressures for meeting bride price, and ultimately contributing to rises in IPV (Fig. 4).

Figure 4.

Geographical distribution of clusters with high rates of intimate partner violence in the past 12 months, by level of relative aridity, Timor-Leste. Key for interpretation: The black markers represent clusters with higher prevalence of intimate partner violence (top 25% of cluster values). The map indicates that arid clusters are more likely to see higher rates of intimate partner violence in the past 12 months.

Associations seen between increases in relative aridity and lack of clean water [42] and fuels align with existing literature on the topic, as changes in aridity and humidity typically affect the availability and quality of aquifers and other water sources, as well as the cost and affordability of fuel and water. Although the effects are smaller, relative aridity is also associated with the availability of basic water sources in Cambodia, where reliance on rainwater makes some groups of women more vulnerable.

Cambodia’s two marked seasons (dry and wet) affect water consumption patterns in areas where basic water infrastructure is not available. For instance, while tube wells (an improved source) are the most common drinking water source in arid clusters throughout the year, many women in humid clusters rely on rainfall (an improved source) during the wet season but shift to unprotected wells and open surface water during dry seasons. As a result, increases in aridity are seen to lower the likelihood of women lacking basic water sources during the dry season but increase it during the wet season (AOR = 1.03*⁣** during the dry season, AOR = 0.95*⁣** during the wet season). This highlights the risks posed by climate-change, as shifts in relative humidity are likely to further alter the predictability of rainwater availability, affect water quality, and increase related collection and treatment burdens on women.

Day land surface temperature: Increases in average day land surface temperature typically present along land degradation [43], sea level rise, reduced agricultural yield, forest transition, biodiversity loss and human disease [44]. They may also take place in the context of rapid urbanization and urban heat islands. Analysis shows that temperature rises are strongly connected with most of the gender-related outcomes considered across almost every country, and although the strength of associations varies, in most cases the effects are large.

In Cambodia and Nepal, increases in day land surface temperature enhance the odds of child marriage significantly. This is likely linked to their effects on agricultural yield and other natural resources, which contribute to economic insecurity and the use of child marriage as a coping strategy [45, 46]. In Nepal, this association is stronger (AOR = 1.05*⁣**), and heavily concentrated across warm clusters in the south-eastern Terai region, where prevailing social norms and traditions are overall supportive [47] of child marriage practices. Controls for wealth, education and other known drivers of child marriage do not change this finding, indicating that indeed temperature may have an effect, particularly in areas where the practice is already prevalent.

In Bangladesh, Cambodia and the Philippines, increases in temperatures correlate with better access to basic drinking water sources. In areas where improved water infrastructure is not available, the effects of temperatures on freshwater availability and contamination may be contributing to these associations. Temperature rises typically present simultaneously with fresh water source depletion (e.g. wetland drying, conversion of freshwater wetlands to agricultural land) and land degradation (e.g. dry forest clearing, salinization) [48], which may result in poor rain absorption and thus runoff draining into waterways and polluting the water supply [49]. Consequently, populations living in high temperature clusters without access to piped water, may shift from drinking river and stream water to improved sources such as bottled or rainwater.

Conversely, in Nepal, increases in temperatures correlate with increased difficulty accessing basic water sources, but disparities in the availability of water infrastructure appear to be affecting this association. Across the Hill and Mountain regions of the country, water supply systems are mostly spring-related sources reliant on precipitation. Especially in the west and central Mountain and Hill regions, many of those who lack piped water rely on unprotected wells and springs. However, across the southernmost provinces (along the border with India), which are further from mountain springs and see higher temperatures, tube wells and hand pumps (improved sources) are widely available.

In the Philippines, as expected, increases in day land surface temperatures also heighten the risk of using unhealthy fuels. Beyond major cities such as Metro Manila and Cebu (which have better fuel infrastructure and LPG availability), clusters reliant on clean cooking fuels are more likely to be located in cooler areas. Economic strains associated with land degradation and heat-driven reductions in agricultural yield and biodiversity loss, coupled with increased electricity and gas prices as a result of heat [50], may be prompting people in hot clusters to shift to cheaper but dirtier fuels.

Unexpectedly, regression shows that the odds of accessing clean fuels are higher in high temperature clusters in Bangladesh and Nepal. In Bangladesh, this association is affected by fuel infrastructure and availability (e.g. people living near natural gas fields, which are warmer areas, have better access to this type of clean fuel); while in Nepal, the country’s topography makes it difficult for residents of mountainous (cooler) areas to access clean fuel infrastructure or purchase LPG. Further analysis controlling for the availability and cost of clean fuels could provide a clearer picture of the associations between clean fuel use and temperature alone.

6.1.2Random forest model

Besides droughts, temperatures and aridity, results from the random forest model highlighted additional variables of importance for gender outcomes, such as proximity to lakes, proximity to coastlines, and rainfall.

Child marriage: Across all countries, proximity to water (lakes or coastline depending on the country) is among the top five most important variables for classifying women into high or low risk groups for experiencing child marriage (Fig. 5). Living near water can have both positive and detrimental effects on poverty, livelihoods and consequently child marriage, depending on the context. For instance, sea level rise (and soil salinization) may be detrimental to the livelihoods of populations living near coastlines, but proximity to coasts may be advantageous for those engaged in fishery-related livelihoods. Relative aridity is also important to assess the risk of child marriage in Bangladesh, Nepal and the Philippines, likely due to its effects on agricultural yield, food security and other poverty-related indicators. Strongly connected with relative aridity, rainfall also appears important in models in Nepal and, to a lesser extent, Cambodia, Philippines, and Timor-Leste, reinforcing this finding.

Figure 5.

Mean decrease in accuracy (in event of variable removal) for child marriage rates, by country. Note: Environment-related variables are shown in darker shade of blue for ease of interpretation.

Figure 6.

Mean decrease in accuracy (in event of variable removal) for lack of access to basic drinking water sources, by country. Note: Environment-related variables are shown in darker shade of blue for ease of interpretation.

Adolescent birth: Although logistic regression only showed mild associations between the different climate-related variables and adolescent births, random forest highlights that climate-related variables do have influence on the risk of adolescent births. The prediction accuracy of adolescent birth models decreases when environmental variables are removed, but less so than in other models. Factors such as culture, religion, limited access to family planning and poor reproductive health services provision, are all known to contribute to adolescent births [51]. Because no data was available to control for these variables at the time of the analysis, these interactions may explain the weaker explanatory power.

Distance to water (nearest lake or coastline) was highlighted as the most important climate-related factor for determining the risk of adolescent births across countries, in line with findings from the child marriage model. The importance of relative aridity to explain the risk of adolescent births was also highlighted in Bangladesh and Nepal, an expected result as it correlates with agricultural yield, poverty, and thus child marriage and adolescent births in these countries, where social norms enable child marriage practices. Environment-related variables have higher explanatory power on adolescent births in Bangladesh and Nepal compared to other countries. As the practice is overall more prevalent in these two countries, climate change may be acting as an amplifier of currently existing practices.

Intimate partner violence: Among the climate-related factors considered, average annual rainfall is one of the variables with relevance across countries (Cambodia, Philippines and Timor-Leste), with model accuracy dropping by 7 per cent in Cambodia and 11 per cent in Philippines and Timor-Leste if the variable is removed. The model also highlighted the importance of proximity to lakes, coastlines and protected areas, with mild effects across countries. Existing literature indicates that the drivers of IPV are multidimensional and highly localized [52], with social norms playing a significant role. As such, although rainfall, proximity to water, and protected areas may contribute to healthy natural resources, decreased economic stresses and thus reduced IPV, the reasons behind these associations are difficult to establish. Further research, including controlling for other known drivers of IPV, could help better understand these reasons.

Access to basic drinking water sources: Climate related variables are powerful for classifying women into groups with high or low risk of lacking access to basic water sources (Fig. 6). Distance to nearest lake, in particular, is among the top 5 most relevant variables in all countries considered. However, living near a lake can imply easier access to livelihoods and infrastructure, but also increased exposure to water contamination. For instance, freshwater lakes are valuable resources for travel, trade, fishing and irrigation and thus, living near lakes may facilitate income generation, and consequently, affordability of improved water sources and infrastructure. However, declines in freshwater quality and increases in water stress may be putting access to basic water sources at stake. Accelerated by rising temperatures, use of pesticides and pollution, cyanobacterial blooms in lakes have increasingly become a challenge as they can impair the safety of drinking water, fishing and irrigation, and pose difficulties for agriculture, fisheries and water treatment plants [53]. If cyanobacteria reach wetlands and groundwater reservoirs near lakes, they may further affect the quality of the water in nearby areas [54].

Day land surface temperature also plays an important role, as expected, given its connections with aridification and extreme weather. This is notorious across all countries, especially the Philippines, where the power of the model would decrease by 31 per cent if the variable was removed. In Bangladesh and Cambodia (dry season) (figure 26), “probability of flood in a 50-year period” is also important for classifying the risk of lacking access to basic water sources. Given the cyclical nature of severe flooding in Bangladesh, it is well documented that drinking water sources are contaminated during floods, as water sources are infiltrated by faecal matter and other debris [55]. To prevent water-borne disease, population groups that typically rely on open surface water for drinking, switch to bottled water and other improved sources during floods. A similar shift takes place in Cambodia, where 11 per cent of women shift away from drinking open surface water during rainy periods.

Access to clean cooking fuels: Of all gender related outcomes considered for this analysis, climate related variables are best at classifying the risk of lacking access to clean cooking fuels. Average day land surface temperature, for instance, is of relevance across all countries considered (in Bangladesh model accuracy would decrease by as much as 38% if the variable was removed, in Nepal and the Philippines by 22%). As noted earlier, different drivers may take place in different countries. Warmer temperatures are often accompanied by land degradation, erratic precipitation patterns, deforestation and reduced agricultural yield, which may lower the affordability of clean fuels (prompting people to shift to unclean). Fuel infrastructure and prices are also key determinants, as noted earlier in the case of natural gas in Bangladesh, but no related data was available to assess their classification power.

Aridity index, rainfall and distance to water bodies are also of relevance across countries. Previously noted relationships between these variables and agricultural yield, economic strains and affordability of fuels may explain these connections. In some countries, rain and proximity to water sources may also affect the process of drying wood and shrubs, thus prompting shifts to other fuels. Random Forest is clear in showing that climate-related variables are important to classify the risk of using unclean fuels, but logistic regression demonstrated that other variables are also at play, such as fuel cost and infrastructure. Repeating this analysis with fuel cost and infrastructure data would be important to understand whether their classification power is higher or lower than that of climate-related variables.

Due to limitations associated with data availability, the localized nature of climate-related phenomena and the multiplicity of factors at play, further analysis is needed to ascertain the causes of the connections between climate and gender outcomes, and the differentiated effects across different population groups. As literature highlights that disadvantaged groups may suffer from climate change disproportionately [56], further analysis with multiple disaggregation by sex and age, or sex and ethnicity, for instance, could provide additional insights.