Modeling the Movement of Homicide by Type to Inform Public Health Prevention Efforts

Abstract

Objectives. We modeled the spatiotemporal movement of hotspot clusters of homicide by motive in Newark, New Jersey, to investigate whether different homicide types have different patterns of clustering and movement.

Methods. We obtained homicide data from the Newark Police Department Homicide Unit’s investigative files from 1997 through 2007 (n = 560). We geocoded the address at which each homicide victim was found and recorded the date of and the motive for the homicide. We used cluster detection software to model the spatiotemporal movement of statistically significant homicide clusters by motive, using census tract and month of occurrence as the spatial and temporal units of analysis.

Results. Gang-motivated homicides showed evidence of clustering and diffusion through Newark. Additionally, gang-motivated homicide clusters overlapped to a degree with revenge and drug-motivated homicide clusters. Escalating dispute and nonintimate familial homicides clustered; however, there was no evidence of diffusion. Intimate partner and robbery homicides did not cluster.

Conclusions. By tracking how homicide types diffuse through communities and determining which places have ongoing or emerging homicide problems by type, we can better inform the deployment of prevention and intervention efforts.

There were more than 14 000 homicides in the United States in 2012—a rate 50% below that in 1993.¹ As is well documented, homicide is not randomly distributed across people, places, or time. Segments of the US population are nearly immune to homicide victimization (e.g., rural, White, elderly women), whereas the risk of falling victim to homicide for others (e.g., urban, Black, young males) is a real possibility.² Common explanations for the disproportionate risk of victimization focus primarily on macrostructural factors that are slow to change and not readily malleable through intervention. Such factors do not help explain the dynamic nature of crime, including homicide, whereby rates of violence are neither stable within communities experiencing near constant rates of disadvantage nor similar across communities with analogous macrostructural profiles. More fully explicating whether and how homicide clusters systematically across space and time is, therefore, critically important.

There are important theoretical reasons to believe that homicides both cluster together in space and diffuse through communities in a contagion-like process.^3,4 This contagion of violence may not apply to all homicide types. Investigations of spatial or temporal diffusion have generally analyzed total homicide⁵ or focused on specific types of homicide to the exclusion of others.^4,6–9 Missing from the literature is an investigation of the spatiotemporal diffusion of homicide by motive type to determine which cluster or diffuse and whether these patterns differ in important ways.

Homicide is a complex crime involving multiple interactions between actors and is often the unintended result of other criminal offenses.¹⁰ These precipitating interactions characterize the etiology of specific homicide types and serve as clues to whether particular homicide types are susceptible to processes of contagion. Fagan et al.¹¹ and Fagan and Wilkinson¹² provided a theoretical process to explain why homicide might diffuse and emphasized how the presence of firearms affects social interactions.

The first step in this process is the “contagion of fear,” whereby a real or perceived increase in the prevalence of weapons in a neighborhood changes the expected utility of acquiring firearms, thus increasing the number of guns in the community.¹¹ Second, the presence of firearms becomes a prominent feature of the neighborhood, and they are used as status symbols denoting power. In the third step, the pervasiveness of firearms motivates individuals to adopt violent identities for self-protection, as alternative modes of self-presentation are viewed as dangerous. Lastly, when groups, such as gangs, are implicated in violence, “diffusion and contagion of attitudes, scripts, and behaviors are clearly visible.”^11(p693)

Although this framework was supported in the analysis by Fagan et al.¹¹ of violence in New York City, the parallels between their framework and research on gang violence are striking. Substituting “gang” for “firearm” in the description would do well to summarize how Klein¹³ implicated gangs in the spread of violence throughout Los Angeles, California. Evidence suggests that gangs may indeed play a unique role in the diffusion of homicide.^4,6 Cohen and Tita used police data from 1991 to 1995 to detect spatial diffusion of drug- and gang-related homicide in Pittsburgh, Pennsylvania.⁴ Their method compared changes in homicide levels of local–neighbor pairs of census tracts to determine if

transitions that are compatible with diffusion . . . are more likely than expected based on the prevalence of other transitions to the same local–neighbor outcome.^4(p474)

Unexpectedly, there was no evidence that drug-related homicides spread throughout the community. Evidence suggested, however, that “particular patterns of rivalries among gangs might underlie the spread of gang-related homicides among apparently disjoint spatial units.”^4(p491)

Some types of homicide may not be affected by processes of social contagion. For example, there is little theoretical or empirical research to suggest that intimate partner or other familial homicides are contagious. That is, knowledge of an intimate partner homicide may not signal an increased risk of homicide to community members and thus may not alter how individuals interact more broadly. Rather, intimate partner homicides are generally preceded by a history of intimate partner violence and are often precipitated by separation.¹⁴

To more effectively prevent homicide, it is imperative to understand whether homicide types cluster or diffuse differentially. Without an understanding of the characteristics of the various homicide types, the most efficient prevention initiatives cannot be achieved if different types of homicides present unique policy problems to local practitioners.¹⁵

We addressed this issue by examining the clustering and diffusion of distinct types of homicides through Newark, New Jersey, from 1997 through 2007. Homicide types analyzed by motive include escalating disputes, revenge, intimate partner, nonintimate familial, drug, gang, and robbery homicide. On the basis of previous research and theory,^8,9,16–18 we hypothesized that homicide clustering and diffusion patterns would differ on the basis of homicide type.

Because intimate partner, nonintimate familial, and robbery homicide do not align with the 4-step process, we hypothesized that these types of homicides are unlikely to diffuse. We hypothesized that drug-motivated homicides may cluster and diffuse because of the criminal networks involved and the cultural mechanisms surrounding these homicides. Finally, we hypothesized that escalating dispute, revenge, and gang-motivated homicides may diffuse through cultural mechanisms, social networks, and high rates of firearm carrying.

METHODS

Newark, located in northern New Jersey, is a city of approximately 270 000 people. Roughly 53% identify as non-Hispanic Black and 31% identify as Hispanic.¹⁹ In 2007, 24% of Newark’s population lived below the poverty level,²⁰ more than the 2007 state (9%) and national (13%) averages.²⁰ Newark is above the national average in rates of serious crime and violence and experienced a peculiar trend in homicide relative to the national decline over the past 2 decades. During the study period, the annual rate of homicide in Newark increased from 19.4 per 100 000 in 1997 to 32.7 per 100 000 in 2007.

We obtained data on homicides that occurred from 1997 through 2007 from the Newark Police Department Homicide Unit’s investigation files (n = 816). We developed a data collection protocol to ensure capturing information relevant to homicide motive in a valid and reliable way. After extracting relevant information from the investigation files for each homicide, researchers followed the protocol to code motive according to a strict schema that did not allow overlap between motive types.

We coded motive into 1 of 7 categories. We grounded classifications of intimate partner and nonintimate familial homicides on the relationship between victim and suspect. Drug-motivated homicides were those that occurred as a result of sales or distribution of illegal drugs. We coded homicides that were directly related to furthering the economic, social, or territorial interests of a gang, including internal conflict within a gang, as gang motivated, consistent with the “Chicago” definition of gang homicide.⁶ We coded homicides that resulted from an argument or physical altercation that immediately escalated into lethal violence as escalating dispute homicides. Revenge homicide occurred as a result of the victim’s or offender’s intention to retaliate for a past event. Finally, robbery homicides resulted from an incident in which the perpetrator attempted to take or successfully took money or material possessions from the victim by force.

In the few cases in which we gleaned multiple motives, we coded the primary reason the offender acted against the victim. For example, we classified a case in which a suspect killed a victim to obtain drugs as a drug-related homicide, not a robbery homicide, because the primary reason the suspect acted against the victim was to procure drugs.

We extracted data from investigation files from March 2000 through October 2008, 10 months after the end of the study period. Importantly, not all cases were resolved at the time of initial data collection. Therefore, researchers reexamined open cases every 6 months to assess if there were any pertinent changes.

We retrospectively detected space–time clusters of homicide by motive on the basis of census tract months using the cluster detection software SaTScan version 9.2.²¹ We geocoded the address where each homicide victim was found and assigned the appropriate census tract identifier. We then modeled homicides by type using discrete Poisson models; we assumed the number of homicides per census tract to be Poisson distributed.

We determined the underlying population at risk in a census tract month with decennial census data and linearly interpolated it for years in between.^22–24 We assumed the expected number of cases in each census tract to be proportional to its population size. We specified the at-risk population for each homicide type on the basis of a demographic analysis of the typical populations victimized in that homicide type in Newark, which were intimate partner (Black females aged 18–64 years); nonintimate familial (Black males and females aged 0–70 years); drug motivated (Black males aged 16–74 years); gang motivated (Black males aged 15–38 years); escalating dispute (Black males aged 15–67 years); revenge (Black males aged 16–63 years); and robbery (Black males aged 15–84 years).

To detect homicide clusters, we used SaTScan to scan an ellipse window across the centroids of all census tracts (n = 90), calculating the number of observed and expected homicides inside the window at each location. The expected number of cases was calculated as

where h was the observed number of homicides, p was the population in the census tract, and H and P were the total numbers of homicides and population, respectively. We then calculated a relative risk of homicide for each census tract month by dividing the number of observed homicides by the expected number of homicides. The alternative hypothesis was that the risk of homicide inside the scanning window was higher than that outside the scanning window. Under the Poisson assumption, the likelihood function for a specific window was proportional to

where H is the total number of homicides, h is the observed number of homicides within the window, and E[h] is the expected number of homicides within the window under the null hypothesis of there being no difference. Because the analysis is conditioned on the total number of cases observed, H − E[h] is the expected number of cases outside the window. I() is an indicator function, with I() = 1 when the window has more cases than expected under the null hypothesis and 0 otherwise.²⁵ The likelihood function was maximized over all window locations and sizes, and the one with the maximum likelihood constituted the most likely cluster. We obtained P values through 999 Monte Carlo simulations. We calculated a test statistic for each random replication as well as for the real data set. If the latter was among the 5% highest, the test was significant at the .05 level.²⁵

We also conducted a spatiotemporal analysis. Modeling parameters included scanning for clusters of geographic sizes that would capture between 0.0% and 6.1% as the upper limit to represent the populations at risk for each homicide type. We also scanned for temporal clusters across 50% of the study period that were at least 1 month long, while adjusting for temporal trends in the data.

Our space–time clusters were constrained within a cylindrical window with an elliptical base of 0.65 miles for nonintimate familial homicides, 1.20 miles for drug-motivated homicides, 1.50 miles for intimate partner violence homicides, 1.60 miles for robbery homicides, 1.80 miles for escalating dispute and revenge homicides, and 1.90 miles for gang-motivated homicide. We derived these parameters from the distance that homicide victims were found from their homes. Thus, the potential at-risk locations in our study spanned the distances from the census tract centroid being scanned. The clusters detected could overlap in space; however, we did not allow pairs of clusters to be contained within each other’s centers to assess the potential for homicide diffusion by type in space and time. We have presented only clusters with a P value of less than .05.

RESULTS

During the study period, there were 560 homicides for which law enforcement gained enough information to classify by motive. These included 126 escalating dispute homicides, 120 drug-motivated homicides, 107 revenge homicides, 75 robbery homicides, 48 intimate partner homicides, 42 nonintimate familial homicides, and 42 gang-motivated homicides. The majority of homicides (70.36%) were committed with firearms. The homicide type involving the highest percentage of firearms was gang motivated (97.62%), followed by revenge (90.65%), drug motivated (87.50%), robbery (80.00%), escalating dispute (53.97%), nonintimate familial (35.71%), and intimate partner (16.67%).

We did not detect any significant clusters for intimate partner homicide and robbery homicide. We did find significant clusters for nonintimate familial, escalating dispute, revenge, and drug-motivated homicides, although there was no evidence of diffusion across census tract months. For example, the 2 nonintimate familial homicide clusters are geographically distinct and occur more than 5 years apart, whereas the 2 escalating dispute homicide clusters are also separated by space and more than 3 years in time (Figure 1).

FIGURE 1—

Space–time clusters of nonintimate familial and escalating dispute homicides: Newark, NJ, 1997–2007.

Although drug-motivated homicides clustered in 2 contiguous areas, they were separated by a period of more than 2 years (Figure 2). These 2 clusters occurred on the western border of Newark and in Newark’s western arm. There are also 2 clusters of revenge homicide (Figure 3) that are separated by geography and time. Notably, 1 of these clusters, located in the central area of Newark by the western arm, inhabits some of the same space and time as both of the drug-motivated homicide clusters.

FIGURE 2—

Space–time clusters of drug-motivated homicides: Newark, NJ, 1997–2007.

FIGURE 3—

Space–time clusters of revenge homicides: Newark, NJ, 1997–2007.

Gang-motivated homicide shows evidence of diffusion (Figure 4). There are 4 contiguous clusters that move roughly clockwise through time from July 2002 through December 2005. The first and second clusters begin in July 2002, but the first cluster only lasts 1 month whereas the second lasts through May 2005. In May 2005, the fourth cluster emerges. The third cluster provides a geographic link between the second and fourth and coincides in time with the second cluster but ends before the fourth cluster begins.

FIGURE 4—

Space–time clusters of gang-motivated homicides: Newark, NJ, 1997–2007.

There is an overlap between the 2 revenge homicide clusters and the 3 gang-motivated homicide clusters. The first revenge homicide cluster has an overlap of 2 census tracts with the first gang-motivated homicide cluster, which occurs entirely within the period of revenge homicide cluster 1, and all the census tracts in gang-motivated homicide cluster 2, which coincide from July 2002 through March 2004. The second revenge homicide cluster shares 4 census tracts with gang-motivated homicide cluster 4, which occurs entirely within the period of the revenge homicide cluster. The second drug-motivated homicide cluster shares much of the same space and time as gang-motivated and revenge homicide clusters near the western arm of Newark.

DISCUSSION

We used spatiotemporal analytic techniques to track the movement of homicide by type over census tract months in Newark, New Jersey, from 1997 through 2007. Knowledge of the clustering and diffusion of homicide by type may help target prevention efforts. Although there were significant clusters of nonintimate familial, escalating dispute, revenge, and drug-motivated homicides, gang-motivated homicide was the only type to show evidence of spatiotemporal movement.

The spatiotemporal overlap of gang-motivated, drug-motivated, and revenge homicide clusters suggests the possibility of common precipitating or facilitating factors. Of the 7 homicide types examined, these 3 had the highest percentage of firearm homicides (87.5% and higher). Firearm carrying may be more common in these geographic areas as people attempt to protect themselves from becoming victims of violence stemming from the perception that gun carrying is prevalent in the area, consistent with the 4-step process.¹¹

It may also be that gang activity in those clusters increased the likelihood of revenge and drug-motivated homicide. For example, 11 of the 20 revenge homicides in cluster 1 involved gang members, and 6 of the 7 homicides in which gang membership was determined in revenge cluster 2 involved gang members. Six of the 9 drug-motivated homicides in drug cluster 2 for which gang membership could be determined involved gang members. Researchers have consistently found that gang members have an increased risk of involvement in severe violent crime,²⁶ and there is some suggestion that periods of increasing rates of homicide involving gang members may lead to increasing rates of homicide involving youths who are not in gangs.⁴ One pathway for this is the contagion model: research suggests that youths who live in areas with high rates of firearm violence engage in increased gun carrying, which increases the risk of firearm violence.²⁷

Another possibility is that norms for behavior in these neighborhoods after noted homicide incidents makes homicide more likely, consistent with Anderson’s code of the street thesis.²⁸ Individuals who reside in these areas may adopt violent responses to perceived disrespect as a form of protection. Even otherwise law-abiding residents in these areas may be willing to resort to violence in an effort to avoid future victimization.

As predicted, intimate partner, nonintimate familial, and robbery homicides did not diffuse. Relative to gang, drug, and revenge killings, perpetrators of these offenses were less likely to use firearms, 1 of the main hypothesized sources of homicide diffusion. These homicides are also less likely to spark additional homicides through retaliation, perhaps because they are more likely to occur in a context in which people are comfortable involving law enforcement for redress.

The lack of spatiotemporal clustering of intimate partner and robbery homicides suggests that specific interventions for these homicide types need not be targeted geographically. One of the more promising strategies for the reduction of intimate partner homicide is reducing access to firearms for domestic violence perpetrators.^29,30 However, the percentage of intimate partner homicides committed with firearms in Newark is surprisingly low compared with national statistics, in which more than 50% of intimate partner homicides are committed with firearms.³¹ This suggests a need for additional responses to intimate partner violence.³²

Currently, there exist promising intervention efforts for many of these homicide types, but interventions for 1 type may not prevent those emanating from alternative motives. For example, an intervention focused on drugs and gang homicide may do little to prevent robbery homicide, nonintimate familial, or intimate partner homicide. By tracking how homicide types diffuse through communities and which places have ongoing or emerging homicide problems by type, we can better inform the deployment of prevention and intervention efforts. For example, targeted strategies, such as focused deterrence^33,34 or Cure Violence model programs,³⁵ which focus on individuals most at risk for violence in geographic areas with homicide clusters, might be effective in preventing drug and gang homicides but not escalating dispute and intimate partner and family incidents.

Interestingly, only 20% of the total observations were included in significant clusters. The percentage of observations that significantly clustered for gang homicides, the only motive type that showed patterns of diffusion, was much greater than that of the other homicide types. To illustrate, 48% of gang homicides significantly clustered, compared with 29% of revenge homicides, the type with the second highest percentage of clustered observations. These results emphasize the importance of treating homicide types differently, as they suggest that most homicide types do not diffuse. Instead, the focus may best be put on place. Results suggest that there are identifiable places in which certain types of homicide are likely to occur; however, most homicides may be placed randomly.

Our ability to detect diffusion at smaller geographic levels is limited by our use of census tracts as the geographic unit. There are homicide clusters in which the homicides occur in a specific segment of a census tract but the whole tract is implicated. We chose census tracts as the spatial unit to keep the data manageable: the use of neighborhoods or street segments would have dramatically increased the number of space–time units in examining the whole of Newark, New Jersey.

Two related violence interventions occurred in Newark during the study period. The Greater Newark Safer Cities Initiative launched in March 2000 and, by May 2005, evolved into an Operation Ceasefire model program. These initiatives did not have a significant effect in reducing crime in the target areas and city.^36,37 These initiatives, however, created greater awareness in the city regarding gangs and helped improve the investigation practices of the Newark Police Department. For example, a task force was created to more thoroughly investigate homicides in the target area, which coincided with some of the gang homicide clusters. This task force sought to improve investigations and, hence, may have resulted in better quality and more reliable data in case investigative files. This may have increased the validity of coded motives post-2000. The impact of this on study results is difficult to ascertain.

We were also limited by the results of police homicide investigations. Because our focus was homicide motive, we were unable to analyze the spatiotemporal location of the 31% of homicides in which there was not enough information in the investigative files for us to assign a motive. We cannot speculate on how lacking motive data for those homicides may have influenced our results. However, research conducted in Newark found that unsolved cases are more likely to involve victims enmeshed in criminal lifestyles.³⁸ Because of the homicide types that were more likely to diffuse involved individuals in deviant lifestyles, future studies should examine whether cases with unknown motives have similar diffusion patterns. As we continue to refine and improve the method and begin to conduct spatial regressions aimed at uncovering factors important to the beginning, length, and movement of clusters, we will move closer to predictive models for future homicide. Although we are not yet there, identifying which types of homicide may cluster together and move across a city provides us important information on which to focus future analyses.