Research presented by Linus Bengsston’s Flowminder Foundation to the recent Netmob 2013 held at MIT’s Media Lab showed that our movements after conflicts and disasters are highly predictable. Analysis of mobile phone data from the 2011 civil war in Cote D’Ivoire (CIV), showed that population movements were up to 88% predictable, an accuracy that was consistent with data collected after the 2010 earthquake in Haiti. In fact, we become more, rather than less, predictable in crises.
Major disasters and conflicts can displace whole populations of people, yet we have no way track these movements with any accuracy. This makes it difficult for relief agencies to respond with water, food, shelter, sanitation, and medical aid in a timely way. Timeliness is key, because long-term effects can be more catastrophic than the original cause of displacement. For instance, up to 50 times more people die in cholera outbreaks among displaced populations, and once infected, populations travel with their diseases, potentially infecting new areas.
“NGOs make guesstimates. You get reports from some city saying ‘We estimate that we have received this many people,’” says Bengtsson. Hundreds of thousands of people fled in panic from the Haitian capital Port-au-Prince after the 2010 earthquake. Where did they all go? Bengtsson tracked their flight via mobile phone data. “Nobody had thought about using this kind of data for this purpose before,” he says. The Haitian National Civil Protection Agency (NCPA) estimated population movements after the earthquake by counting ship and bus traffic, information they distributed to relief agencies. These data provided an accurate count of the number of people moving, but not where they went. For instance, the NCPA’s estimates for the number of people moving to the Departments of Sud (South) and Ouest (West), were less than a third of the actual figures.
With mobile phone location data from 1.9 million SIM cards, Bengtsson and his research team at Flowminder Foundation compared 42 days of pre-earthquake data against 158 days of data after the event. Their analysis included calculations for the radius of gyration (a measure of the size of trajectories) to estimate how much a subscriber moved, and entropy measures to define the disorder and predictability of an individual’s movements.
“We were surprised at the regularity of people’s movements, which we expected would be much more chaotic and unpredictable. The patterns of movement were quite similar to usual times, but the absolute levels of people’s movements were much larger. People were just as predictable as before, but they moved more. People go to places where they have their social support structures,” says Bengtsson.
In fact, people leaving the city mostly went to the same places they had visited for Christmas and New Year’s where they were likely to have friends and family. The geographical distribution of the population obtained from SIM movements closely matched that reported by a United Nations Population Fund household survey performed several months after the earthquake.
One limitation of mobile phone monitoring is that the most vulnerable groups, such as pregnant women, the elderly, and the poorest people are least likely to have phones. “Children and elderly people are often not alone, so they are usually moving with phones, but still there will be fewer phones per person than on average. But that’s a long research process going forward,” says Bengtsson.
Cholera was unknown in Haiti, but 10 months after the earthquake, Haiti suffered the worst epidemic of the disease in recent history. The U.S. Centers for Disease Control and Prevention estimates at least 8,060 Haitians died and hundreds of thousands were hospitalized. Bengtsson’s team started working with Digicel, the mobile phone carrier in Haiti, immediately after the earthquake, but it took a couple of months to sort out the legal issues and get access to the anonymized data. “It was really during the cholera outbreak that we could use the data in a rapid manner,” he says.
The researchers tracked the movements of 138,560 SIM cards for 8 days immediately after the outbreak and distributed their analyses within 12 hours of receiving the latest data from Digicel, considerably faster than existing methods.
“The mobile phone estimates were a vast improvement. The International Organization for Migration, which concentrates on population movements, said that these were absolutely the best estimates they could get and really informed the need across the country.”
Cote D’Ivoire’s 2010 presidential elections, the first in 10 years, ended with both candidates claiming victory. Negotiations failed, and in 2011, the opposition candidate mounted an armed assault against the former president. Conflict turned to disaster as civil war broke out in Cote D’Ivoire and up to a million people fled the violence.
As part of the Data for Development challenge, mobile operator Orange released anonymized mobile phone data for 500,000 subscribers in Cote D’Ivoire covering the period between December 1, 2011 and April 28, 2012. Rather than analyzing the immediate effect of the war’s outbreak, Bengtsson’s research group was studying longer-term after-effects. “The data is from 6 months after the war finished. We wanted to understand how people were moving back,” he says.
Mobile phone location data tracks subscribers by subprefecture, one of 255 administrative sub-units in Cote D’Ivoire. In addition, each connection, a call or message transmitted through a cellular tower, provided more precise location data. 77% of subscribers had at least one location update per day during two-thirds of the tracking period, but there were large variations in how much people moved. 16% of subscribers never left their home subprefecture, while a few outliers visited over 50 subprefectures.
Bengtsson’s team did a statistical analysis of the data using a Markov chain, a discrete modeling method where the next state depends only on the current state and not on the sequence of events that preceded it. So given a subscriber’s location, the Markov chain should be able to predict where the subscriber will go next. In fact, the model proved 90% accurate. “People’s movements are highly regular and predictable even after disasters (Haiti) and conflicts (Cote D’Ivoire),” concludes Bengtsson.
Despite their success, not everyone is easily convinced. Persuading mobile operators to release their data requires getting past legal obstacles, and persuading relief agencies to trust predictions based on the data and then act has not been easy. But if anonymized mobile phone location data can accurately predict where people will go in a crisis, it can improve relief efforts and help prevent disease outbreaks. One day your smartphone really could save your life.
[Image: Flickr user Andrew Mager]