We found that an estimated 630,000 people left PaP during the first 19 d following the earthquake. The geographic distribution of people who moved out of PaP was highly nonuniform, and was similar to estimates from a large retrospective population-based UN survey but very different from the official government estimates which were widely used during the relief operations. We additionally illustrate that it was feasible to produce and disseminate data on movements of SIM cards from an area with an infectious disease outbreak within hours of receiving the data. These results suggest that the speed and accuracy of estimates of population movements during disasters and infectious disease outbreaks may be revolutionized in areas with high mobile phone coverage.
There has been no method available to provide timely and accurate estimates of population movements following many disasters. Large-scale surveys and censuses can give detailed information on people's movement history but are not feasible to implement for monitoring purposes during the acute phase of a disaster. Eyewitness reports may carry severe bias when referring to population movements occurring over time and across areas.
There are several limitations to this study. Mobile phone use is low in several population groups, including children, the elderly, the poorest, and women 
. If these groups have substantially different movement patterns than groups with high mobile phone use, results will be biased. However, the proportional geographic distribution of people moving away from PaP agreed well with the results of a large retrospective population-based survey by UNFPA. The similarities can be due to a widespread use of mobile phones, that nonmobile phone users, e.g., children and the elderly, moved together with mobile phone users, and also that users and nonusers of mobile phones had similar movement patterns. The UNFPA study was however not implemented for validation purposes and future population-based studies in other settings are needed. SIM movements will however unequivocally provide a valuable lower bound on population movements if ownership of multiple SIMs from the same company is uncommon. In the present study the SIMs would then have tracked a minimum of 31% of the PaP population (809,000 persons). If for example children and the elderly, who do not use mobile phones, accompany people who do, the number of persons who are directly tracked by the method would be substantially higher.
The method may be less suitable in areas where mobile phone use and mobile radio coverage is low. Geographic localisation is less precise in areas with low tower density. When relying on call data, locations of infrequent callers are updated less often than those of frequent callers. As well, in some countries, companies put large amounts of airtime on new SIMs to attract new customers. This can lead some customers to keep one stable SIM for receiving calls while new SIMs are continuously bought and used for making calls. This was not the case in Haiti but needs to be taken into account in other contexts. Additionally, there are settings where many people have one SIM card from each company, which needs to be taken into account if data from more than one company is analyzed.
Mobile phone networks are relatively resilient to external shocks. However, major disasters can affect power supply, destroy towers, and cause a complete loss of functionality. Limited possibilities for people to charge their mobile phones can cause bias and might have done so in this study. In Haiti, power cuts were common before the earthquake and existing electrical generators seem to have supplied considerable charging capacity. Other partially compensating factors for power cuts include long stand-by times of many simple inexpensive phones as well as the habit of people in places with insufficient power supply to routinely turn off phones to conserve battery.
Although we did not have access to other data types, network registries contain more detailed data that may be analyzed. Such data include for example all calls (as compared to once daily in the present dataset), text messages and data downloads (but not the content of these), data on expenditure and the size of prepaid refills, as well as SIM location data that are regularly registered by the system without SIMs making calls. Some of these data are difficult to rapidly retrieve and some companies regularly erase certain data types. Discussion with the specific company in each case is important. Data should always be made anonymous before analysts access it, as was done in this study.
With software development and in cooperation with network operators, the described approach can provide data in close to real time. Information on postdisaster population distributions can potentially enable improved distribution of water, food, shelter, and sanitation. Needs assessment surveys can potentially be improved through increased validity of population estimates. These denominator estimates would be important both when constructing sampling frames for needs assessment surveys and when generalizing survey data to overall population needs. For example, survey data indicating that 50% of the people in an area needs shelter are tremendously more informative if the area has an accurately estimated population size. Daily estimated changes in the number of displaced persons can be generated for specified areas, which can signal important on-going developments. Estimates of mortality can potentially be derived from the number and geographic distribution of nonresponding SIMs. Estimation of buried but alive persons following an earthquake is another potential area for development. Network data can provide even richer information when combined with information on the ground and potentially also with data from mobile phone surveys 
Our approach may also be useful in nondisaster contexts. Human mobility is extremely important for the spread of communicable diseases 
, and early containment of epidemic outbreaks is often a key factor in preventing spread 
. Rapid data on population movements can potentially inform outbreak preparedness and response for infectious diseases. In addition, the coarse data on diagnosed cholera cases that are presently available 
show interesting similarities to the mobile phone network data (). More detailed comparisons will be fruitful to perform if data on the location of diagnosed cases become available. Text messaging that targets specific areas is another potentially valuable use of mobile phone network data. If these refer to advice relating to population movements (e.g., advice regarding evacuation or the localisation of relief supplies), results could be directly evaluated with the use of network data. During the cholera outbreak, Digicel sent, on the basis of the presented analyses, text messages with health information to all mobile phones that passed through the outbreak area. This project was then expanded to include an automatic voice dialler and a total of seven text messages per mobile phone subscriber.
We found that routinely collected data on the movements of all active SIM cards in a disaster-affected nation could, with potentially high validity, be used to provide estimates of the magnitude, distribution, and trends in population displacement. With pre-earthquake census data, the method could also provide estimates on area-specific population sizes, which could lead to important improvements in the allocation of relief supplies and the quality of needs assessment surveys. Further, we found that the method was feasible to use for close to real-time monitoring of population movements during an infectious disease outbreak. We recommend establishing relations with mobile phone operators prior to emergencies as well as implementing and further evaluating the method during future disasters.