11, 13?9, 21?3]. These studies are able to show such anomalies by comparing behaviors to (-)-Blebbistatin web events where the character, time, and place of these events are already known. We build on these studies, but develop a blind system that is closer in nature to an actual event detection system. Instead of starting with the time and location of an event, then ZM241385 web looking for anomalous calling behavior, we develop a behavioral anomaly detection system that identifies days with unusual calling or mobility behavior, as well as the location and geographic extent of these disruptions. Our detection system is scalable as it is able to efficiently process years of country-wide mobile phone records. For illustration we use mobile phone records from a single cellular services provider from Rwanda. We connect the identified anomalous days and locations with extensive records of violent and political events and natural disasters. Results of this exercise reveal that some days with anomalous increases in calling and mobility behavior match well with several different kinds of events. In other cases, days with decreases in calling and/or mobility match with events. These cases were surprisingly more numerous than events matched with increases in calling and mobility. In still other cases, we do not find good event matches for days with anomalous behavior and we also find cases where emergency events occurred without resulting in anomalous behavior that our system could detect. Notably, we learn as much from the unmatched events and behavioral anomalies as from the matched cases. We argue that further quantitative and qualitative research into the exact and possibly multi-dimensional nature of human response to emergency events is needed. In this regard, our careful analysis of both the matched events and the events and instances of anomalous behavior that do not match reveal some key insights into further developments needed to better understand human response to emergency events. In fact, it is this outcome, namely the demonstration that human behavioral responses to emergency events are much more complex than previously assumed, that is the most important contribution of this paper. Future research must address this complexity and can benefit from using existing social and psychologicalPLOS ONE | DOI:10.1371/journal.pone.0120449 March 25,2 /Spatiotemporal Detection of Unusual Human Population Behaviortheories of behavioral response to threat. We conclude this article by setting out a clear pathway of research aimed at the goal of a creating an effective system of identifying emergency events in real-time (or close to real time) from mobile phone data.Materials and Methods Measuring human behavior with mobile phone dataCellular service providers continuously collect mobile phone records for billing purposes and to improve the operation of their networks [24?6]. Every time a person makes a voice call, sends a text message or goes online from their mobile phone, a call detail record (CDR) is generated which records time and day, duration and type of communication, and an identifier of the cellular tower that handled the request. We analyze anonymized CDRs provided by a major cellular phone service provider in Rwanda. These data comprise all mobile phone activity in the provider’s network between June 1, 2005 and January 1, 2009 [27, 28]. Many of the existing methods for emergency event detection rely on call volume, at either the cellular tower level or at individual.11, 13?9, 21?3]. These studies are able to show such anomalies by comparing behaviors to events where the character, time, and place of these events are already known. We build on these studies, but develop a blind system that is closer in nature to an actual event detection system. Instead of starting with the time and location of an event, then looking for anomalous calling behavior, we develop a behavioral anomaly detection system that identifies days with unusual calling or mobility behavior, as well as the location and geographic extent of these disruptions. Our detection system is scalable as it is able to efficiently process years of country-wide mobile phone records. For illustration we use mobile phone records from a single cellular services provider from Rwanda. We connect the identified anomalous days and locations with extensive records of violent and political events and natural disasters. Results of this exercise reveal that some days with anomalous increases in calling and mobility behavior match well with several different kinds of events. In other cases, days with decreases in calling and/or mobility match with events. These cases were surprisingly more numerous than events matched with increases in calling and mobility. In still other cases, we do not find good event matches for days with anomalous behavior and we also find cases where emergency events occurred without resulting in anomalous behavior that our system could detect. Notably, we learn as much from the unmatched events and behavioral anomalies as from the matched cases. We argue that further quantitative and qualitative research into the exact and possibly multi-dimensional nature of human response to emergency events is needed. In this regard, our careful analysis of both the matched events and the events and instances of anomalous behavior that do not match reveal some key insights into further developments needed to better understand human response to emergency events. In fact, it is this outcome, namely the demonstration that human behavioral responses to emergency events are much more complex than previously assumed, that is the most important contribution of this paper. Future research must address this complexity and can benefit from using existing social and psychologicalPLOS ONE | DOI:10.1371/journal.pone.0120449 March 25,2 /Spatiotemporal Detection of Unusual Human Population Behaviortheories of behavioral response to threat. We conclude this article by setting out a clear pathway of research aimed at the goal of a creating an effective system of identifying emergency events in real-time (or close to real time) from mobile phone data.Materials and Methods Measuring human behavior with mobile phone dataCellular service providers continuously collect mobile phone records for billing purposes and to improve the operation of their networks [24?6]. Every time a person makes a voice call, sends a text message or goes online from their mobile phone, a call detail record (CDR) is generated which records time and day, duration and type of communication, and an identifier of the cellular tower that handled the request. We analyze anonymized CDRs provided by a major cellular phone service provider in Rwanda. These data comprise all mobile phone activity in the provider’s network between June 1, 2005 and January 1, 2009 [27, 28]. Many of the existing methods for emergency event detection rely on call volume, at either the cellular tower level or at individual.