We have divided our work into two distinct phases: Research and Implementation. During the first phase, we will construct a model to help us predict where environmental impacts are most likely to lead to sociopolitical instability and conflict globally, as well as build relationships with individual domain experts and organizations working in environmental conflict prediction and prevention. Once we have created a predictive model, we plan to use it as an action plan for the next stage of our work, in conjunction with the strategic advice we have received during the process.
Our goal over the next few months is to create a global model predicting future sociopolitical instability and related violence at the city level, as well as in which locations environmental peacebuilding work is likely to make the greatest difference in preventing this conflict. To do so, we plan to improve upon our existing approach in the following ways.
Scale and Scope of Data
We will expand from an analysis of Syria to a global scale. For this reason, all satellite and events data we use will be global in context. Our next analysis will include our original data sources on rainfall, drought severity, NDVI, cropland percentage, population density and population growth rates.1 To this, we will add the following:
- National and subnational governance data on regime type and stability, as well as indicators of well-being including infant mortality, GDP and GDP growth;16
- Environmental data including temperature, soil moisture content, and predicted climate impacts such as rainfall variability and flood and drought occurrence;
- Known conflict risk factors including ethic composition and fractionalization, presence of petroleum and high value mineral resources, and past incidence of conflict;16 and
- Events and media reports from the GDELT Project indicating occurrence of violent events, media tone, and keyword mentions of political and environmental concerns.17
Algorithms and Approach
Because standard regression models are often poor at predicting conflict compared to more recently developed machine learning algorithms,18 we plan to improve our forecasting ability through application and comparative testing of three to four machine learning models. To do so, we will select the random forest and similar supervised learning models, train them on a subset of available data related to past conflicts, and then cross-validate them using the remaining available past conflict data. After fine-tuning each model, we will select the one with the highest predictive accuracy. This approach has advantages over typical regression models in that, with sufficient data, it allows for very high accuracy.18 We believe decision tree based approaches like the random forest model will be especially helpful because they can provide some indication of the potential causal role each of our variables played in spurring conflict, and can be trained using a smaller set of data points.19,20 These last two characteristics are particularly helpful to our study both because conflicts are relatively rare events21 and because we plan to investigate what role environmental factors played in each conflict.
In order to construct the best possible model and ensure thoroughness in providing input data, we plan to consult practitioners with expertise in the following domains:
- Conflict prediction and prevention,
- Disaster and conflict early warning systems,
- Climate change impacts and adaptation in developing nations,
- Environmental peacebuilding methods and impact evaluation,
- International governance and state stability, and
- Machine learning in social science contexts.
These experts will be able to provide practical insights from their past experience, help us refine the questions we are posing, guide our analytics in the right direction, and help evaluate the effectiveness of our results.22 With the help of our team members and advisers, we expect to have this process completed within 4-6 months. Once our algorithm has been validated, we plan to begin the application phase of our work while iterating our research process to further improve our model.
We plan to engage in a cyclical research process (illustrated above) beginning with data gathering and process validation through interviews and discussions with practitioners. We will use this step to refine our questions and ensure we have acquired all available related data. After feeding this data into a multiple test models, we will select and refine our top performing model, interpret the results, and review them in collaboration with practitioners. A few iterations of this research process over the course of 4-6 months will provide a solid foundation for the application phase of our work and continue to support our intervention-focused work thereafter.
As a delay often exists between advances in research and their application, we hope to speed along this process by developing organizational partnerships that will allow us to ensure our research benefits local communities as soon as possible. We are looking at options ranging from advocacy within larger organizations, to launching a pilot study to better determine the impact of this work in climate and conflict vulnerable locations, to raising funding and support for local advocates and organizers in vulnerable locations. Because the UNEP has identified impact assessment as a major gap to be filled for targeted environmental peacebuilding work to gain widespread institutional support, we believe launching a pilot study in partnership with a larger organization and assessing its impact will be our most important next step.11 This pilot study would likely include a subset of the intervention strategies proposed above, tailored to the local context and centered around empowerment of local community organizers. We expect our long-term strategy, however, to involve a multi-faceted approach shaped by the relationships we build and the advice we are able to obtain during our research-focused phase over the next few months. Our ultimate goal is to help individuals and organizations working to address climate change and conflict focus their efforts as effectively as possible, and to ensure that those in vulnerable locations are able to receive the funding and resources they need.