Introduction

Opportunity theories of crime assume that most crimes take place while the victims and the offenders are involved in non-criminal routine activities. Even crimes that are premeditated and that are not perpetrated during the offenders’ routine activities, are nevertheless supposed to be informed by what offenders heard, saw, read or picked up otherwise during non-criminal daily routines. In their introduction of routine activity theory, Cohen and Felson (1979) summarize this position by stating that “illegal activities feed upon the legal activities of everyday life” (p. 588). In outlining crime pattern theory, Brantingham and Brantingham (2008) take the same position by stating that “what shapes non-criminal activities helps shape criminal activities” (p. 79). The present study aims to test this key assumption of opportunity theories. To do so, it puts front stage the daily mobility flows of the urban population as measured by the tracks of mobile phones and assesses whether these mobility flows help predict where individual thieves commit crimes. The key question is thus one that precisely describes the criminal location choice approach (Bernasco and Nieuwbeerta 2005): Given the home location of an offender motivated to commit a crime and given the locations of all potential targets, can we understand and predict where the offender commits the offense? Whereas prior crime location studies have included only the home-target distance (and occasionally also measures of physical and social barriers) to explain crime location choice (Ruiter 2017), the present study explores whether the explanation can be improved by considering the movement patterns of the general population as a ‘spatial template’ for the journey to crime.

Background

In trying to understand and predict where individual offenders go to commit crimes, it is important to consider their journey to crime, which links their home location to where they commit crime (Rengert 2004). Two empirical observations seem undisputed. The first observation is that the frequency of crime displays distance decay: it decreases with the distance from the offender’s home (Rossmo 2000). The median distance between the offender’s home and the crime location is no more than a mile, with some variation observed across crime types and offender categories (Townsley and Sidebottom 2010). Research in human geography has demonstrated that distance decay also characterizes non-criminal routine activities like working, visiting friends or shop**. The distribution of daily human travel distances, including presumably those of offenders, follows a power law function: Most of our trips are over short distances whereas occasionally we take longer trips (Gonzalez et al. 2008). Thus, legal and illegal trips both display a distance decay pattern. Although distance from home is not an explicit element of opportunity theories, its empirical relevance is evident.

The second undisputed observation about crime locations is that crime is geographically concentrated at and around crime generators, micro-places where many people converge on a regular basis to pursue similar or complementary activities (Brantingham and Brantingham 1995). Crime generators include transit stations, shop** centers, schools, sports venues, entertainment areas and other types of facilities. Crime generators produce crime for two reasons. First, because they are popular destinations for many legal activities, they are known to a large proportion of the population, including potential offenders. They are an element of their activity space. Second, these micro-places often provide specific opportunities for crime, especially when they are being visited by large crowds of people. For example, shops provide opportunities for shoplifters, distracted passengers at subway stations are easy targets for pickpockets, and intoxicated patrons leaving bars are vulnerable for robbery (Bernasco and Block 2011). In sum, crime generators produce crime because they are widely known and because they provide abundant opportunities. Empirical evidence thus demonstrates that offenders generally commit crimes around busy places located within at most a few miles from their homes, depending on the crime type, the size of the city and the convenience of available transportation.

These two observations on distance decay and crime generators, however, do not determine with much precision where individual offenders perpetrate their crimes. The area within one mile of an offender’s home already covers 3.14 square miles. The area within two miles covers no less than 12.57 square miles. Distance does not tell us in which direction offenders prefer to go. Moreover, most urban landscapes contain a multitude of busy places within a few miles from anywhere.

Therefore, understanding where offenders commit crimes requires more than the two ‘stylized facts’ about distance decay and crime generators. As emphasized in the opening lines of this article, opportunity theories of crime assert that most crimes are perpetrated during, or are informed by, routine activities pursued without explicit criminal intentions. Thus, to better understand where offenders commit crimes, we need measures to inform us about their whereabouts during their legal daily routine activities.

Unfortunately, systematic information about the places that offenders visit during non-criminal daily routines is not easily available. Only two small-scale studies have addressed the daily mobility of offenders. Rossmo et al. (2012) used data of an electronic monitoring system that tracked continuously in real-time the whereabouts of parolees and offenders on bail under community supervision. Griffiths et al. (2017) used mobile phone data records of terrorists to track their visits to different locations and to examine their visiting frequency distributions in the months preceding the attack they were involved in.

Fortunately, however, the mobility patterns of the general population provide a template for the daily mobility of offenders, and as a direct consequence for their journeys-to-crime (Boivin and D’Elia 2017). The mobility patterns of the population could thus help predict where offenders commit crimes, one of the main aims of this study. We measure the whereabouts of a large fraction of the population of a huge city in China by geo-tracking their mobile phones during a full day. This allows us to measure daily mobility flows of the general population as a proxy measure of the daily mobility of the offender population, and to include this measure as a covariate in a model of crime location choice, together with distance and the presence of crime generators.

Research on crime and law enforcement in China receives increasing scholarly attention internationally. The cultural distance between China and western industrialized countries presents an excellent opportunity to test whether theories developed in western industrialized countries can be generalized to other contexts, in particular China. This opportunity also applies to environmental criminology, because its main theoretical frameworks—routine activity theory and crime pattern theory—have been developed and tested almost exclusively in a western context. A recent exception is a study by Long et al. (2018), who demonstrated that urban street robbers in China tend to return to locations of their prior offences to commit subsequent robberies, confirming prior findings from the United Kingdom and The Netherlands. Another exception is a study showing that Chinese burglars who live in communities with a higher average rent, a denser road network and a higher percentage of local residents, commit burglaries at shorter distances from their homes (** literature.

In the present paper, we focus exclusively on ‘theft from the person’ or TFP, a non-violent property offense committed in public or semi-public places, excluding burglaries and thefts from vehicles. Although simple theft is a common crime, it has mostly been ignored in the crime location choice literature.

Research Question

In sum, the main research question of the present study is whether daily mobility flows of the general population, as measured by tracked mobile phone data, help us explain and predict where individual offenders commit thefts from the person. If the answer is positive, an additional question is whether the inclusion of mobility flows adds to the effects of the distance and crime generators, or rather replaces them. In other words, we are interested in the extent to which the effects of mobility flows substitute or complement the effects of distance and of crime generators.

Our findings not only add to the testing and development of opportunity theories, they also contribute to a growing body of research that aims to test in other cultures and on other continents crime theories that have been developed in the USA, Europe and Australia. With its high pace of urbanization and economic, cultural and social change, China seems like an excellent context to test and further develop opportunity theories of crime.

Theory and Prior Findings

The present section will provide a brief overview of the role of distance decay, awareness space and crime generators in the literature on crime location choice. It will subsequently address the potential role of the daily mobility of the urban population in modeling and understanding where offenders commit thefts from the person.

Crime Location Choice

Crime location choice refers to an offender’s choice of where to perpetrate a crime (two recent reviews are Bernasco and Ruiter 2014; Ruiter 2017). The discrete spatial choice model was introduced by Bernasco and Nieuwbeerta (2005) as a method to analyze crime location choices using individual-level offender data. The model is rooted in a formal theory of discrete choice known as random utility maximization in micro-economics that is consistent with rational choice as applied to offending criminology (Bernasco et al. 2017b; Cornish and Clarke 1986). The theory asserts that offenders choose crime locations rationally, selecting those locations that maximize their expected rewards and minimize their expected costs and risks. For example, residential burglars would be expected to target accessible areas that are affluent and where formal and informal guardianship is low (Bernasco and Nieuwbeerta 2005), and street robbers would by hypothesized to seek select crime locations nearby their own homes where they are likely to encounter vulnerable victims who carry cash or other valuables (Bernasco et al. 2013). These and other factors assumed to affect crime location choice will be discussed below, in sections “Distance and Awareness Space” to “Mobility Flows”, with a special focus of offenders committing theft from the person.

Distance and Awareness Space

Of all factors that affect crime location choices, distance is by far the most influential. All sixteen crime location choice studies reviewed by Ruiter (2017) included a strong and significant negative effect of distance on crime. The theoretical meaning of distance is disputable, as it could be a merely proxy measure for other concepts, such as the cost of travel.

In support of the argument that the time and money costs of travel are more fundamental factors than distance, some studies have demonstrated the crime-reducing effects of physical barriers other than distance. Clare et al. (2009) studied the role of barriers and connectors on destinations chosen by Australian burglars and found them to have a strong influence on offenders’ decisions on where to commit burglaries. Peeters and Elffers (2010) assessed the extent to which highways, railroads, parks and canals reduce the number of crime trips across these physical barriers, but found only a small effect. Townsley et al. (2015) demonstrated that for adult offenders, but not for minor offenders, the accessibility of a target via the street network increased its likelihood of becoming a target of theft from a vehicle.

Distance might also be considered a proxy measure of the offenders’ awareness space. Because all human mobility, both legal and illegal, displays distance decay, their routine activities help shape their awareness space (Brantingham and Brantingham 2008). Offenders are generally more familiar with places nearby their residence than with more distant places. As a consequence, they must be more knowledgeable of criminal opportunities in nearby places than in distant places. Indeed, some location choice research has suggested that offenders are much more likely to offend within their own awareness space than elsewhere. Arguing that for various reasons offenders might still be familiar with their former areas of residence, it has been demonstrated that they are more likely to offend near their former homes than elsewhere (Bernasco and Kooistra 2010; Lammers et al. 2015; Menting et al. 2016). It has also been established that they are more likely to offend near the homes of family members (Menting et al. 2016).

Offenders may also be familiar with places of prior criminal experiences. They have a tendency to return to places where they previously committed crimes (Bernasco et al. 2015; Lammers et al. 2015), a finding congruent with previous explanations of repeat victimization and near repeat victimization based on victimization data only (Johnson et al. 2007; Townsley et al. 2003).

Establishing a reduced burglary risk in properties along cul-de-sacs, Johnson and Bowers (2010) suggested that their reduced permeability might make cul-the-sacs unfamiliar places for non-residents, including most potential motivated offenders. Frith et al. (2017) use the physical layout of the street network to generate alternative measures of offender spatial awareness, finding the reduced accessibility in the street network reduces the likelihood of street segments being targeted.

In sum, prior research has empirically demonstrated that offenders prefer nearby locations and locations that are part of their awareness space, and thus both proximity and spatial awareness are important criteria in offenders’ location choices.

Crime Generators

Crime generators and crime attractors are concepts in crime pattern theory that potentially play an important role in crime location choices. They are facilities that have elevated levels of crime because they bring together crowds either continuously or at certain moments during the day or the week (Brantingham and Brantingham 1995; Kinney et al. 2008). Crime generators are places and facilities where all kinds of people go to perform legal daily activities. Crime attractors have been defined as places and facilities that have a reputation of providing opportunities for crime, which makes them particularly attractive to motivated offenders (and unattractive to some potential victims) (Brantingham and Brantingham 1995). Because crime attractors share many features with crime generators, and because the reputational element in their definition may be difficult to evaluate objectively, especially in the context of a non-western culture, in this paper we will subsume crime attractors under the more general label of crime generators.

Although it may be impossible to enumerate all types of facilities that can operate as crime generators, some stand out as being particularly attractive for some types of crime, and they will be discussed in the present section. We describe elements common to certain classes of facility, although even amongst the same class (e.g. bars) there exists considerable heterogeneity making some of them risky places but others not (Eck et al. 2007).

High schools have been associated with elevated crime rates in part because of the assumption that the students are more likely than people of younger or older ages to be involved in crime as offenders or victims. A concentration of adolescents around high schools would possibly generate a concentration of crime before and after school hours (Roncek and Faggiani 1985; Roncek and LoBosco 1983). Near elementary schools, opportunities for TFP are abundant when it is crowded because parents, grandparents and other caretakers drop off the children in the morning, and also when they pick them up later in the afternoon.

Research on street robbery has demonstrated that proximity to small scale businesses, including grocery stores, corner stores, gas stations, barber shops and fast-food restaurants, increases the risk of street robbery (Bernasco and Block 2011; Bernasco et al. 2017a; Haberman and Ratcliffe 2015). To the extent that these are the types of businesses where cash payments are more common than non-cash payment modes, customers may be likely to carry cash. Because street robbery and TFP are similar offenses in terms of location (public space) and preferred items to steal (concealable, removable, available, valuable, enjoyable and disposable, see Clarke 1999; Wellsmith and Burrell 2005), these businesses might also function as crime generators for TFP. For similar reasons, banks and ATMs have been shown to have the same effect (Haberman and Ratcliffe 2015).

The criminogenic effects of bars and other types of alcohol outlets have been demonstrated repeatedly to generate not only violent crimes but also property crime, in particular robbery, and mostly likely because of intoxicated patrons might be easy targets for robbers and thieves (Bernasco and Block 2011; Conrow et al. 2015; Groff 2011; Roncek and Maier 1991).

Especially during rush hours, transit stations (bus stations, subway stations) are crowded places that offer opportunities for theft, in part because travelers may be distracted and not vigilant in their attempts to find the way to their destinations. Many studies have confirmed that crime rates are elevated in and around transit stations (Bernasco and Block 2011; Block and Block 1999; Block and Davis 1996; Clarke et al. 1996; Haberman and Ratcliffe 2015; Qin and Liu 2016).

In sum, the extant literature provides empirical evidence for the fact that locations where people congregate, often in large crowds, are more likely to be selected as crime sites by motivated robbers and thieves. This is particularly true for locations with functions that potentially indicate larger benefits or the presence of vulnerable and non-vigilant victims.

Social Disorganization

Social disorganization has been identified as a community attribute that can apply to geographic areas of various sizes (e.g. street blocks, neighborhoods or districts). Although not a true crime generator (crime generators are facilities with a specific function), it is an observable attribute of communities that attracts offenders, because it provides them with opportunities to offend under the veil of anonymity and with low risks of bystander intervention.

In the extant literature, which mostly covers cities in the United States and in other western industrialized countries, social disorganization is defined as the inability of local communities to realize the common values of their residents or solve commonly experienced problems, such as local crime (Kornhauser 1978), and has theoretically and empirically been associated with concentrated socio-economic disadvantage, residential mobility, and ethnic heterogeneity (Bursik 1988; Sampson et al. 1997). However, in the Chinese context, social disorganization and concentrated disadvantage in cities have been associated primarily with concentrations of non-local migrants (Zhang et al. 2007b), who have recently migrated into the city and lack a Hukou status (Hou 2010). Hukou is a household registration system in China that separates locals and ‘nonlocals’ in the allocation of public resources (Song et al. 1970; Miller 2005). Research in social geography has demonstrated that patterns of mobility are strongly related to personal attributes. Hanson and Hanson (1981) used a transportation survey to study the relationship between socio-demographic characteristics and travel patterns, and found that gender, education, and income affected individual travel patterns. Their findings have been replicated in subsequent research. Schwanen et al. (2008) used activity diaries from 420 adults in 270 households in Columbus (Ohio, USA) and found that women face more constraints on their routine activities than men do. In response to questions about how easily they could change their activities (e.g. laundry, cooking, paid labor) to another location or time of day, women reported their activities to be less flexible than men, both in time and space. Zhou et al. (2015) compared the out-of-home activity spaces of low and high income groups in Guangzhou (China), and concluded that the segregation between both groups not only applies to where they live, but also to the places and areas they visit during daily routine activities.

Recently, the accessibility of large volumes of geo-referenced mobile phone data has provided researchers with promising opportunities to study daily mobility patterns (Kwan 2016). Song et al. (2010b) used data of phone users and showed daily mobility to be very predictable. Gonzalez et al. (2008) also used tracking data of phone users, and found that human trajectories show a high degree of temporal and spatial regularity, each person being characterized by an individual travel distance and a high probability to return to a few highly frequented locations.

Similar patterns seem to characterize the mobility of offenders. Griffiths et al. (2017) used mobile phone data records of offenders involved in four different UK-based terrorist plots, in the months preceding their attacks. They found that the offenders visited only a very limited number of places, and pursued their activities mainly close to their home or safe house location. This pattern resembles the mobility pattern of the general population. Ratcliffe (2006) demonstrated that spatiotemporal patterns of property crime are strongly influenced by the temporal constraints experienced by the young people who committed these crimes, and that these temporal constraints are very similar to the constraints that non-offenders are subjected to.

These findings are in line with the opportunity theories of crime (Brantingham and Brantingham 2008; Cohen and Felson 1979), which also assert that offenders’ crime journeys are not very different from the journeys they make when not perpetrating crimes. This seems not only plausible for crimes that took place in response to unexpected opportunities or provocations, but also for crimes that were premeditated. In case of premeditated crimes, one would expect the offenders to use their pre-existing spatial knowledge to select the target and to escape the crime scene after the offence.

Two studies addressed the relation between the mobility of offenders and the mobility of the general population. Felson and Boivin (2015) formulated a ‘funneling hypothesis’ asserting that daily mobility flows of an urban population determine where crimes take place. They used a large-scale (N = 165,000) transportation survey to estimate the daily numbers of people visiting each of the 506 neighborhoods of a large city in Eastern Canada for purposes of work, shop**, recreation and education. The volumes of both violent and property crime in a neighborhood were strongly associated with the number of trips (for work, shop** or recreation) having the neighborhood as their destination.

An even more convincing argument is provided by another study that used the same transportation survey in the same study area. Rather than modeling crime volumes per neighborhood, Boivin and D’Elia (2017) modeled the crime trip volumes between all pairs of the 506 neighborhoods (i.e., how many crimes were perpetrated in neighborhood j that were committed by offenders living in neighborhood i, for all 506 i and j). One of their predictors was an estimate of (non-crime) trip volumes between the 506 neighborhoods. Their results indicated that both for violent and property crime and both for crime committed by youths and by adults, the volume on non-crime trips between two neighborhoods was significantly positively associated with the number of crime trips between these two neighborhoods. This finding suggests that the mobility patterns of general population provide a template of ‘funnel’ for the mobility of offenders, including their journeys-to-crime.

Building on this finding, in the present study we investigate whether, in addition to distance and crime generators, daily mobility flows of the urban population help predict where individual thieves commit crimes. As described in the next section, daily mobility flows are measured using anonymized tracks of mobile phone users in the general population.

Data and Methods

Study Area and Spatial Unit of Analysis

The data for the present research originates from ZG City,Footnote 1 a city located in southeast China. ZG City has a total population in excess of 5 million, and is one of the largest and most developed cities of China. It belongs to the subtropical zone and the temperature fluctuates between 15 and 32 °C, so there is no extreme cold or hot weather and seasonal variations do not affect mobility.

The study area is located in the central part of ZG City, and covers more than 3000 km2. The daily mobility pattern of the urban population is a key variable in the present research. Because commuting volumes and distances in China are strongly and positively related to size and prosperity of cities (Engelfriet and Koomen 2017), the choice of ZG city assures that we study a city with a highly mobile population.

The census unit was chosen as the spatial unit of analysis because census units have a size similar to those in comparable studies (Bernasco and Block 2009; Clare et al. 2009; Menting et al. 2016), because census units are approximately equally sized, and because in the study area they are relatively homogenous in terms of population composition. In addition, a practical advantage is that the use of census units does not force us to estimate spillover effects (effects of attributes of nearby units on crime in a focal unit) which is a requirement if small units of analysis are used, such as census blocks or street blocks (Bernasco et al. 2013; Groff and Lockwood 2014).

Using census units as spatial units of analysis implies that we aggregate numbers of crime generators to the census unit level, and thus sum the numbers of schools, restaurants, bus stops and other facilities per census unit. This is not just a practical necessity because all relevant theoretical constructs must be aligned to the same spatial resolution, it is also legitimized by the observation that crime generators produce mobility flows and presence of people in the streets where they are located and in their wider environments.

The size of census units is 1.62 km2 on average (standard error 2.99 km2, minimum 0.02 km2, maximum 31.30 km2). The mean population is 5956 (standard error 4706, minimum 245, maximum 51,450). The study area comprises 1891 census units in total, of which 275 are not included in the analysis because the lack mobility data due to limited coverage of the GSM mobile phone network, leaving 1616 census units in the analysis. In the section on mobile phone data, we explain why we use census units to locate the cell towers.

Distances between census units are the Euclidian distances in kilometers. For use in the analysis, per offence we calculated the distances between the census unit where the offender lived and all other census units in the study area. Following Menting et al. (2016) and other prior crime location choice studies, the distance to the offender’s own census unit of residence was defined as the average distance between two random points in the census unit, which approximately equals half the square root of the size of the area (Ghosh 1951). A total of 595 incidents (17.3%) were committed in the same census unit as where the offenders lived. The average distance of the 3436 crime trips was 5.02 km and their standard deviation is 6.43.

Crime Data: Theft from the Person

Registered crime data were obtained from the police authorities of ZG City. All crime data are recorded incidents of theft from the person (TFP) in which at least one offender was arrested and prosecuted. It includes the date of the incident and the home addresses of the offenders and the location where the offence was committed.

TFP is a very common offence in China and elsewhere. It covers theft and attempted theft of items (e.g. cash, phone) directly from the victim without the use or threat of physical force. TFP includes both snatch theft—in which the victim is aware of the act—as stealth theft—in which the victim is unaware of it (e.g. pickpocketing). TFP does not include thefts with ‘break and enter’, like residential or commercial burglary, or theft from closed vehicles. Theft from the person targets people rather than buildings, vehicles or other objects. It is therefore a very appropriate crime type to explore the relation between the mobility of the general population and the target choice of offenders.

Because the home location of the offender is an essential element of the model, the data used in this study contain only detected cases, i.e. cases where a suspect was arrested, identified and prosecuted. The clearance rate of theft from the person in ZG City was 5.2% in 2014, which implies that approximately 1 in every 20 thefts leads to the arrest of an offender.

In the original dataset, there were 10,276 cases of TFP committed by 12,670 offenders between June 1st, 2014 and May 30th 2016. The month of February in both 2015 and 2016 was excluded, because during this the Chinese Spring Festival takes place, and a large proportion of the population leaves ZG city to celebrate the lunar new year with their families in their hometowns. This temporary change might alter the regular mobility patterns of both offenders and potential victims.

Because data on the daily mobility patterns of the urban population was only available on weekdays, only TFPs on weekdays were selected for the analysis. Unfortunately, as in many other studies (e.g., Vandeviver et al. 2015) a substantive percentage of cases had to be excluded from the analysis for lack of geographic detail. In 6118 cases either no offender address or no TFP location was available. After selections, the remaining 4019 offences committed on weekdays and with valid location information were committed by 4798 offenders in total. The fact that offenders outnumber offences implies that some TFPs are committed by co-offending groups.Footnote 2 Because estimation results might be biased if offences committed by multiple co-offenders would incorrectly be analyzed as if they were independent observations, and in line with recommendations in the crime location choice literature (Bernasco and Nieuwbeerta 2005; Townsley et al. 2015) only solitary offending was selected for inclusion in the analysis,Footnote 3 leaving for the analysis 3436 cases of TFP.Footnote 4

Crime Generators

The list of potential crime generators is sheer endless, but included are those facilities that seem most relevant based on the extant literature and taking into account the Chinese context: the numbers of primary schools, middle schools, hospitals, basic stores, markets, supermarkets, restaurants, cinemas, bars, banks, subway stations and bus stops per census unit. All of them are so-called ‘point of interests’ (POIs) that were purchased from Daodaotong, a map** and navigation company, which provides the original map data for China’s largest search engine (Zheng and Zhou 2017). These POIs in ZG city were collected in 2014.

Primary schools include kindergartens and elementary schools. Potential TFP victims are not the pupils but the adults (usually parents or grandparents) who gather near the gates of schools around the school exit hour to pick up their children. Most middle school students are not picked up by their parents and would be potential TFP victims themselves, as the middle schools are also very crowded at school exit hours, creating opportunities for TFP. As presented in the descriptive statistics in Table 1, there are 1.28 primary schools and 0.36 middle schools on average per census unit.

Table 1 Descriptive statistics and correlations of census unit characteristics (counts of crime generators and % nonlocals)
Full size table

Hospitals refer to general hospitals, which in China are always busy places. Basic stores include shops for daily supplements and for clothes. With an average of 5.22 per census unit they are quite common. Markets comprise open air markets as well as ‘integrated markets’ that mainly sell food. Banks include both banking business offices and automated teller machines (ATMs). Restaurants, cinemas and bars are popular entertainment facilities that attract crowds, and subway stations and bus stops attract large volumes of people who are in transit. As for public transportation, each census unit has 3.19 bus stops but only 0.07 subway stations on average.

In addition to counting crime generators that are facilities, we also calculated the population heterogeneity for each census tract. In the Chinese context, segregation does not follow racial or ethnic lines as they do in many western countries. In Chinese cities, the main attribute that distinguishes locals from non-locals is their Hukou status. Non-local communities where few residents have local Hukou status are generally characterized by high levels of anonymity, weak social attachments among residents and serious social disorder problems. Across all census units the mean proportion of non-locals is 0.34.

Table 1 also contains a correlation matrix of the crime generator variables. The correlations between the crime generator counts are low. The only one stronger than 0.50 is the correlation between the number of restaurants and the number of supermarkets. As a consequence, there is no reason for concern about collinearity issues.

Mobile Phone Data

To measure population mobility across the study area, we use the tracked data of mobile phone users. There is a burgeoning literature using mobile phone location tracking to study the movement patterns and travel behavior of human populations (Asakura and Iryo 2007; Yang et al. 2014). An advantage of mobile phone data over transportation surveys is that mobile phone data are not forced into a specific origin–destination format that restricts the measurement to start and end locations. Mobile phones monitor the location of the population continuously without assuming origins, destinations or intentionality. Another advantage is that the measures are taken automatically and unobtrusively and do not require active participation of research subjects.

The data of the locations of mobile phone users in this research was provided by a major mobile phone service provider in China that has a market share of 22.5% (Chong et al. 2015). Song et al. (2018), and they spend more than 80% of their time in their 10 most frequented locations (Gonzalez et al. 2008; Song et al. 2010a). In a subsequent aggregation procedure that we ourselves implemented, each phone tower location was assigned to the census unit in which it was located. The result is a data file that records for each mobile phone user and per hour in which of the 1616 census units they were located.

Calculation of Mobility Streams

The resulting information was subsequently used to construct a measure of the volume of mobility flows between all pairs of census units. The measure is defined as the number of unique mobile phone users that visited both census units on the same day. Mobility flows were calculated by summing phone users over all pairs of census units. The result is a square matrix of census units containing the number of mobile phone users that visited both census units on a single day.

Our use of mobile phone locations diverges from how prior studies on crime in London (Bogomolov et al. 2014) and in Osaka, Japan (Hanaoka 2018) have used such data in research on crime. These prior studies used mobile phone users’ locations as a measure of how many people were present at a specific location at a specific day or time, which has been referred to as the size of the ambient population or the actual population at risk. Other studies have measured the ambient population with other data sources, such as the Landscan Global Population database (Andresen 2006; Andresen and Jenion 2010), geo-references Twitter messages (Malleson and Andresen 2015), or public transportation transaction data (Song et al. 2017).

Third, only the data of 4G users of one mobile phone network provider is available in our research. Although the 4G network is very popular in China, the sample of users whose phones were tracked may not be fully representative of the population of ZG City.

Fourth, only theft from the person was studied here. This is a ‘street’ crime characterized by a low rate of reporting to the police (Zhang et al. 2007a). To what extent population mobility flows impact the spatial behavior of those who commit other types of crimes, including crimes against static targets, such as burglary, remains an open question. The approach adopted here, in particular the use of cell phone data to measure population mobility, might prove useful in in future research that addresses this question.

The abovementioned caveats should not dwarf the contributions of the present study. In addition to applying a new mobility measure based on tracking the locations of individual cell phones, this study contains two other new elements to the discrete crime location choice framework. It has been the first to apply the framework in the Chinese context, and it has been the first to apply it to theft from the person. The findings demonstrate that Chinese offenders committing theft from the person decide on target areas in ways that resemble the choices made by burglars and robbers and other types of offenders elsewhere in the world. This may be viewed as tentative evidence that the underlying opportunity theory could be applicable not only in the industrialized world, but also in other countries.