# Network analysis of hazard interconnections

posted in: Uncategorized | 3

This post is the first in a series focused on using network analysis to analyse disaster risk.

Disasters beget disasters. The 2010 Haiti earthquake for example, killed hundreds of thousands on its own, but it also led to a cholera outbreak that has, to this date, infected over 700,000 people and killed 9000. Knowing how these interconnections work can be useful for all elements of emergency planning. The knowledge that earthquakes can trigger vector borne disease outbreaks, for example, can be used to guide mitigation (e.g. by ensuring health and sanitation facilities are up to scratch in earthquake prone areas) preparedness (stocking medicine and investing in disease surveillance) and response (incorporating CLTS or other behavioural change programmes focused on disease following earthquakes).

Understanding these interconnections can, however, be challenging. Causal processes are often complex, and interconnections are not always apparent. Network analysis can be used to sort out these complexities. Network analysis is used to understand how actors, events, or other processes relate to each other. It often has a mathematical component, but can also be a tool for visualising interconnections and seeing how processes interrelate. This post applies network analysis visualisation tools to help understand hazard interconnections.

### Network analysis of natural hazards

This is one of the maps I made using network analysis.  I’ll show you how I produced it and explain its implications.

The first step is to develop a table showing how natural hazards are connected. Each row is of a different hazard incident, and each column is a hazard event that can be ‘triggered’ by the initial hazard. The first row, drought, for example, shows that drought can trigger disease, erosion, fire, and pest infestation. The hazards in this table are from the CRED database, a global database of natural disasters. I have ranked their connections: a 0 indicates no connection, a 1 small and infrequent connection, a 2 is a medium and somewhat frequent connection, and a 3 is a large and common connection.  The table is my estimate of these connections. If you have different opinions on any of these connections please let me know in the comments!

This table already has some utility. For example, it can clearly be seen that disease does not trigger any other hazards, meaning that you don’t need to worry about other natural hazards in preparedness projects focus solely on disease. Floods, however, often result in disease, erosion, and landslides, so flood risk reduction should consider incorporating elements of disease, erosion, and landslide management.

This table is what is known as an adjacency matrix. Adjacency matrices show the connections, called edges, between components, called nodes. The nodes in this instance are the hazards, the edges the number.

I uploaded this agency matrix into Gephi, an open source programme for network visualisation, to create a few maps.

##### In-degree: hazards that are commonly triggered by other hazards

This first map highlights the hazards that are commonly triggered following a disaster event:

The size and colour are tied to the ‘in-degree’, the number of edges coming into a node. Nodes whose in-degrees larger are more likely to be triggered by hazard events. The map makes it clear that landslides, erosion, fire, and disease are common byproducts of other disasters. As such, you might want to incorporate these hazards if developing a generic preparedness plan. Volcanic eruptions, earthquakes, and drought, however, are not triggered by other disasters.

The relative location of nodes is also telling. I used ForceAtlas to spatialise the map. ForceAtlas is a force directed layout that simulates a physical system to spatialise a network. Under this configuration, nodes repulse each other while edges attract their nodes, similar to a spring linking two bipolar magnets (Jacomey et al., 2014). ForceAtlas uses a model of asymmetrical repulsion, which takes into account the degree of nodes, not just the weight of each edge, making each node interdependent on the other nodes for its position in space (Jacomey et al., 2014). In this map landslide, erosion, tsunami, and flooding are close to each other, meaning that they have close connections. They are, however, further away from fire, which shows that fire is less related.

##### Out-degree: hazards that trigger other hazards

The next map shows the hazards that trigger other hazards:

Size and colour are tied to ‘out-degree’, the number of edges coming out of a node, so nodes with higher out-degree are more likely to trigger other hazards. The map shows earthquake and drought to have a number of knock-on effects. As such, if you work in an area where earthquakes and drought are primary hazards, you should prepare yourself to respond to a number of other disasters when those hazards are realised.  Again, the map was spatialised using ForceAtlas.

##### Degree Centrality and Eigenvector Centrality: hazard drivers

Finally, the last two maps show hazard drivers.

This map is a degree centrality map. Degree centrality is simply a combination of in degree and out degree. As such it shows the hazards with the most general connections to other hazards, both as triggers and as byproducts. Erosion, landslides, and floods are largest in this map.

Hazards with the highest degrees can be considered ‘disaster drivers’, the things that lead to disasters. To stop disasters from occurring you’d want to reduce these drivers.

A degree map is a basic method of identifying drivers that is probably good enough for risk management decision-making. However, disasters pathways are sometimes more complex than a simple two step process: flooding, for example, can trigger erosion which in turn can trigger landslides.

Eigenvector centrality is designed to capture these broader interconnections

Eigenvector centrality rests on the premise that each node’s centrality is the sum of the centrality of the nodes it is connected to. Eigenvector centrality can therefore be a more precise measure than degree centrality for identifying hazard drivers. Compared to the degree centrality map erosion and disease are much larger drivers, while drought, earthquake, and tsunami much smaller.

### Conclusion

These maps provide a mechanism for understanding how hazards are interconnected. Using a simple adjacency matrix, it is possible to use in-degree to identify hazards that are commonly triggered by other hazards, out-degree to identify the triggering hazards, and degree and Eigenvector centrality to identify hazard drivers. Such knowledge can help improve preparedness, mitigation, and response, strengthening risk management as a whole.

Besides providing practitioners with a deeper sense of hazard interconnections, the results of this analysis have clear implications and policy and theory. Specifically, by making very clear just how interconnected hazards are to each other, the analysis provides a strong argument for an ‘all hazards’ approach to disaster risk management. Since one hazard is often triggered by another, risk management practices cannot afford to focus on a single hazard at the expense of others. Additionally, the analysis moves beyond ideas that drivers are specific and few, relegated primarily to social issues like inequality and poverty: hazards are themselves drivers and should be viewed as such. These ideas in turn promote a move away from traditional reductionist approaches to risk, often expressed in equations like Risk=Probability x Impact, toward more complexity oriented risk approaches, such as those found in the resilience paradigm and expressed in systems models like the Sustainable Livelihoods Framework.

There is, however, still a long way to go in understanding hazard interconnections. The model presented in this post is an idealised proof of concept that assumes exposure to all natural hazards. It represents sort of ‘hazard hypermarket’ scenario akin to if Afghanistan was located on the coast or if Haiti had volcanoes. While the model provides a clear method for analysing hazard interconnections, it needs to be field tested in a series of empirical cases.

### References

Joe Gill and Bruce Malamud’s 2014 paper on natural hazard interactions: http://onlinelibrary.wiley.com/doi/10.1002/2013RG000445/abstract

Gephi website: https://gephi.org/

Jacomy et al‘s paper on ForceAtlas: http://webatlas.fr/tempshare/ForceAtlas2_Paper.pdf

CRED database: http://www.emdat.be/database

Information on the Sustainable Livelihood Framework: http://practicalaction.org/disaster_approaches_livelihoods

### 3 Responses

1. ##### Joel Gill

It’s interesting and very positive to see more work being done on hazard interactions, thanks for sharing.

In 2014 I published this review, characterisation and visualisation of the interaction relationships between 21 different natural hazards – using a similar matrix form. We also collated case studies and relevant literature for many of these. The results are published in this open-access Reviews of Geophysics article (http://onlinelibrary.wiley.com/doi/10.1002/2013RG000445/abstract).

Colleagues and I are currently in the process of organising a second workshop on the theme, and will be posting details here soon if you are interested (http://www.interactinghazards.com/).

2. ##### Network analysis of hazard interconnections &nd...

[…] This post is the first in a series focused on using network analysis to analyse disaster risk. Disasters beget disasters. The 2010 Haiti earthquake for…  […]

3. ##### Rick Gebethner

Hi Aaron,

Very nice. I’m a big fan of network analysis. Here are some thoughts about relationships that may be worth further analysis – disclaimer being that I am no expert on catastrophes (some people close to me may not agree). Pest infestation leading to disease – whether it’s bubonic plague, lyme, zika, etc. Volcanic events leading to floods where the volcanoes are in glaciated areas. Keep up the good work.