Utilizing Incident-Based Crime Data to Inform Strategic Interventions

Theoretical Orientation

The current analysis is grounded in a public health approach to understanding the patterning of violence to shaping prevention efforts. Although criminology has historically claimed violence under its domain of expertise (Akers & Lanier, 2009), recent decades have seen interpersonal violence increasingly cast as an issue affecting the public health (Krug, Mercy, Dahlberg, & Zwi, 2002; Prothrow-Stith & Weisman, 1991; Weisheit & Wells, 2013). That is, the incidence and prevalence of violence in American society are described as an epidemic resulting in negative population health outcomes (Christoffel, 2007; Mercy, Rosenberg, Powell, Broome, & Roper, 1993; Moore, 1995). This perspective depicts violence as being similar in nature to other health risks affecting population mortality, such as illnesses and accidental injuries (Gabor, Welsh, & Antonowicz, 1996). Such a view on the occurrence of violence stands in contrast to classical criminological offender-based approaches, which treat violence as an outcome of individual offender motivations (Moore, 1995; Prothrow-Stith & Weisman, 1991).

Examining violence through a public health lens draws attention to the concept of risk for describing its pattern of occurrence and the selection of targets for intervention. Although frequently used in the literature, risk is a term which often goes without definition. Broadly, risk represents the probability of experiencing an event (Burt, 2001; Last, 2001), and in the context of violence, it represents the probability of experiencing a violent victimization. Differential risk for violence across individuals, populations, and geographies is driven by variation in the presence of risk factors—characteristics or structures which increase the risk of violent victimization for individuals and communities (Last, 2001; Moore, 1995). In this sense, individuals and communities can be described as being at differential risk for violence based on the distribution of risk factors for violent victimization. These risk factors can represent static characteristics that cannot be changed through intervention, or dynamic characteristics, capable of change with time and intervention. Greater exposure to risk factors for violence places some individuals or groups within communities at higher relative risk of victimization than groups with lower exposure.

Recent examples of criminological research have specifically incorporated theory, method, and practice from public health in order to understand and respond to violence risk at microlevels. The Indianapolis Violence Reduction Partnership, modeled after a “pulling levers” intervention, implemented in Boston in the mid-1990s (Braga, Kennedy, Waring, & Piehl, 2001), sought to deliver a deterrence message and community outreach to groups at high risk of involvement in violence (McGarrell, Chermak, Wilson, & Corsaro, 2006). Corsaro and McGarrell (2010) categorized the intervention as public health oriented because it specifically identified groups at high risk of involvement in violence through shared information between law enforcement, community corrections, and community actors. Efforts at communicating violence risk were targeted toward networks of individuals with high exposure to violence risks (e.g., gang-related activities), with specific attention given to target neighborhoods where risk factors were highly concentrated. The intervention decreased homicides involving 15- to 24-year olds by more than 50%, with the strongest impact among Black males in that age-group (Corsaro & McGarrell, 2010). A similar focused deterrence strategy was deployed in Cincinnati as the Cincinnati Initiative to Reduce Violence (CIRV). In line with urban public health interventions, the CIRV inductively identified and targeted gang members at high risk for violence, providing both messages of deterrence and support. An evaluation observed a 40% reduction in gang-involved homicides after 3.5 years of program operation (Engel, Tillyer, & Corsaro, 2013).

Highlighting the concentration of risk factors in communities raises the distinction that in addition to being “at risk of violence,” some populations may be classified as more vulnerable than others. Vulnerable populations are at higher risk of violent victimization due to their exposure to risk factors in combination with their lack of economic and social resources (Aday, 2001). Populations “at risk” are defined by higher exposure to a specific risk factor; vulnerable populations are groups of people who, because of their position in the social structure, are “at risk of risks” because they are commonly exposed to contextual conditions that increase their exposure to risk factors (Frohlich & Potvin, 2008). Recognizing the nested nature of violence risk exposure is important as critics have claimed that the term “vulnerable populations” is potentially stigmatizing (McLaren, McIntyre, & Kirkpatrick, 2010) in so far that it emphasizes a state without identifying causes, that is, that vulnerability is a characteristic of people rather than the result of a process (Popay et al., 2008). Instead, a World Health Organization commission has advanced the concept of “exclusionary process,” which originates in the unequal distribution of material, cultural, social, and political resources (Popay et al., 2008).

To this extent, criminological and public health violence prevention interventions which identify and target groups and contexts at high risk for violent victimization are contrasted against the alternative. One approach consists of punitive, reactive, sweeping efforts to reduce crime and violence through general deterrence (e.g., mandatory minimum sentences, three strikes; Corsaro & McGarrell, 2010; Golembeski & Fuilliove, 2005). Such an approach makes no effort to address the causal mechanisms underlying the incidence of violence. In contrast, recent years have witnessed highly focused, data-driven, and problem-solving criminal justice interventions that use analysis to identify risk and target interventions to risk (Klofas, Hipple, & McGarrell, 2010). Examples include pulling levers gun violence interventions (e.g., Braga et al., 2001; McGarrell et al., 2006), Los Angeles’ Operation Los Angeles’ Strategic Extraction and Restoration Program (LASER) (Uchida & Swatt, 2013), problem solving at crime hotspots (Braga & Weisburd, 2010), and Hawaii Hope (Hawken & Kleiman, 2009). In contrast to these focused efforts, public health interventions can also address the causes of violence through a population-based strategy which seeks to lower the mean level of risk factors, shifting the entire distribution of risk exposure downward (Rose, 1985).

Although there is an ongoing debate within public health concerning the merits of population-level and targeted interventions (Frolich & Potvin, 2008; McLaren et al., 2010; Rose, 1985), in the context of violence prevention, targeted interventions are highly relevant due to the dense concentration of violence. For instance, Papachristos, Wildeman, and Roberto (2015) examined the concentration of nonfatal gunshot injuries across networks of high-risk individuals in Chicago. A co-offending network consisting of less than 6% of the city’s population made up nearly three quarters of nonfatal gunshot wounds. Further, they observed that dense concentrations of gunshot victims in an individual’s social network increased the probability of that individual sustaining a gunshot injury (Papachristos, Wildeman, & Roberto, 2015). In this sense, particularly high-risk networks that disproportionately contribute to violent victimization and offending can be identified as targets for multimodal interventions.

Regardless of whether an intervention is population based or targeted toward those at highest risk of violence, the public health approach identification of at-risk groups and geographies is decidedly inductive and empirical, rather than deductive and theoretical (Mercy & Hammond, 1999; Moore, 1995). This aspect of the public health approach accentuates the matter of how patterns in violence risk are to be appropriately identified for the purposes of violence prevention intervention. The next section will highlight the use of state-level incident-based crime data systems for similar purposes when survey victimization estimates are unavailable.

Methodological Value of Incident-Based Crime Data

In order to produce the greatest impact on reducing violence, empirical evidence suggests that enforcement, intervention, and prevention efforts should be highly focused and data driven (National Research Council, 2004, 2005). In the context of allocating block grant resources, the specific means by which agencies should or could identify groups or localities at the greatest risk of violent victimization is less clear. In the ideal scenario, victimization surveys administered on a regular basis could be utilized to produce estimates of victimization rates for a variety of sociodemographic groups and localities, providing insight into differential risk for violence. Victimization surveys hold several benefits over data from law enforcement record systems, particularly the ability to estimate victimization rates incorporating information from victims who never reported the offense to the police (Cantor & Lynch, 2000). Unfortunately, there is little systematic implementation of victimization survey methodologies for local law enforcement. The NCVS is a valuable resource for gaining insight into victimization risk for subgroups at the national level (see Lauritsen & Heimer, 2010), but the NCVS contains limited geographic identifiers, restricting the ability to produce subnational estimates of victimization (Maxfield, 1999; Roberts, 2009), even at the state level (Fay & Diallo, 2012; Rand, 2009). Although efforts are under way to produce subnational NCVS estimates for large states and cities through sample boosts (Rand, 2009) and model-based strategies (Fay & Diallo, 2012), providing coverage to states and enabling victimization estimates for even smaller localities (i.e., counties, small cities) would come at great expense in terms of both time and resources.

Without local-level survey estimates, incident-based crime data from NIBRS presents a valuable resource toward the identification of demographic and geographic patterns in risk across geographies at the state level. NIBRS captures a wide breadth of information for understanding the nature and extent of violent victimization through the use of relational tables corresponding to different units of analysis (Vasquez, Stohr, & Purkiss, 2005). For any given incident of reported crime, NIBRS data tables are capable of describing multiple victims, multiple offenders, multiple offenses, and multiple arrestees, representing a significant improvement over summary-based reporting systems (Bierie, 2015; Roberts, 2009). In addition to providing detail on the characteristics of victims, offenders, and offenses, NIBRS data fields identify the reporting agency (through the Originating Agency Identifier [ORI]), as well as the county and city that the incident occurred in, allowing insight into geographic variation in victimization and enabling contextual information (e.g., community socioeconomic indicators) to be linked to NIBRS estimates. ²

Unfortunately, NIBRS data systems are lacking widespread implementation. Despite the Federal Bureau of Investigation’s (FBI) initial suggestion of fully implementing NIBRS in only a small sample of agencies and a reduced version for the remaining agencies, the NIBRS rollout advanced with the goal of implementing the full NIBRS among all law enforcement agencies (Addington, 2008; Barnett-Ryan & Swanson, 2008). This ambitious goal has fallen short of being achieved. Addington (2008) observed that in 2002, just under half of the participating UCR agencies were also participating in NIBRS, with the lowest participation among agencies serving large populations (i.e., greater than 250,000). In 2012, there were 6,115 law enforcement agencies submitting data to NIBRS, covering approximately 30% of the U.S. population and 28% of all crime reported to the UCR (NIBRS, 2012), representing a slow rate of growth in participation. There are several explanations for the lack of NIBRS implementation. As Bierie (2015) notes, the burden for producing software platforms to support NIBRS is placed on law enforcement agencies, and the cost of building and maintaining NIBRS data systems is a nontrivial expense for these agencies. Compounding this cost is a perceived lack of direct value for local police departments (Bierie, 2015; Roberts, 1997). That is, although NIBRS allows for enhanced scientific research on crime (i.e., utility to researchers), it lacks (or is perceived to lack) immediate support for law enforcement functions (Bierie, 2015; Roberts, 1997).

Because NIBRS relies on data generated from crimes reported to the police, even relatively complete NIBRS data will represent an undercount of victimization relative to victimization surveys. However, it can still provide a valuable resource for understanding relative patterns in violent crime and victimization as well as estimates of relative differential risk across demographic subgroups and localities. This is particularly the case when state agencies create their own local NIBRS systems (LIBRS) to continue to access incident-level data after it has been submitted to the FBI (Bierie, 2015). This is the case for the state detailed in the current analysis, demonstrating the utility of statewide NIBRS coverage for informing data-driven violence prevention interventions.

Data

The current analysis utilizes data from the Michigan Incident Crime Reporting (MICR) system. MICR is an incident-based crime reporting system maintained by the Michigan State Police (MSP), representing Michigan’s contribution to the NIBRS. All law enforcement agencies in the state are required to submit incident-level crime statistics to MSP. Beginning in 1989, MSP shifted reporting requirements from a summary-based system (i.e., such as the aggregate data contained UCRs) to an incident-based system (MSP, 2014). Although Michigan law requires submission of data on a minimum of a monthly basis, currently local agencies are able to submit data to MSP electronically on a continuous basis. Michigan is currently 1 of 15 states submitting all of their crime data to the FBI via NIBRS. In 2012, Michigan’s NIBRS submitting agencies covered 9,633,716 persons or approximately 97% of the state population. To this extent, MICR represents an LIBRS (Bierie, 2015) covering an entire state.

The availability of the incident-based MICR data covering the entire state represents a valuable resource for law enforcement, researchers, and policy makers for understanding crime patterns and planning prevention, intervention, and enforcement strategies to reduce crime and violence. For the current effort, members of the research team worked with the Criminal Justice Information Center at MSP to access the MICR data files for the most recent year available (2013). For this year of data, 529 Michigan law enforcement agencies were equipped to submit incident data to MICR. Of these agencies, 462 (87.3%) submitted a full 12 months of data, while another 36 (4.9%) submitted less than 12 months of data. A total of 498 (94.1%) Michigan law enforcement agencies were either fully or partially represented in the data.

Complex file structure and sample derivation

The MICR consists of numerous file segments that were linked together based on unique incident, victim, offender, and offense identifiers. In 2013, the MICR contained data on 744,223 unique criminal incidents across the state, where an incident is defined as “one or more offenses committed by the same person or group of persons acting in concert, at the same time and place” (MSP, 2014, p. 1). Within MICR, each incident is identified by a unique identifier, where in any given incident, there may be a single victim or multiple victims, a single offender or multiple offenders, and involve the commission of a single offense or multiple offenses by said offenders against said victims. This complex file structure has been noted as a limitation in the analysis of NIBRS data (Maxfield, 1999), as separate data sets may need to be created depending on the desired unit of analysis (i.e., victims, offenders, or offenses). To proceed with analyses, unique identifiers were created for each victim, offender, and offense occurring within each incident. Based on these identifiers, victim, offender, offense, and incident data were appended into a single file, where each row represented a Victim × Offender × Offense triad. Given that individual victims, offenders, or offenses may be represented in multiple rows of the data, the “dplyr” grammar for data manipulation (Wickham & Francois, 2014) was utilized, as implemented in the R statistical computing environment (R Core Team, 2015). Given particular grouping variables of interest (i.e., victim or offender age, sex, and race), dplyr allows for the counting of unique victims, offenders, or offenses within these groupings, avoiding complications arising from the same victims, offenders, and offenses being represented in multiple rows of the complex file structure.

With the current inquiry’s focus on violent crime, the data were reduced to incidents involving a reported homicide, aggravated assault, or robbery against a victim that was an individual (i.e., a person, excluding businesses and the government as victims). With these criteria in place, the current analysis consists of 32,056 unique violent incidents, which included 37,681 unique victims, 35,978 unique offenders committing 32,183 unique offenses (see Table 1). ³ Of these offenses, 13.1% involved multiple victims (homicides 14.2%, aggravated assaults 13.2%, robberies 12.6%) and 22.2% involved multiple offenders (homicides 17.3%, aggravated assaults 14.6%, robberies 42.0%).

OPEN IN VIEWER

Table 1.

Total Unique Victims, Offenders, and Offenses Across 2013 MICR Incidents.

	Homicides	Aggravated Assaults	Robberies	Total
Units	n	n	n	N
Unique victims	629	26,668	10,390	37,681
Unique offenders	510	23,995	11,473	35,978
Unique offenses	584	22,617	8,982	32,183

Note. n = 32,056 unique incidents. MICR = Michigan Incident Crime Reporting.

Variation in Victimization Risk

LIBRS systems provide a data-driven means to identify subgroups at risk of violence. Density curves describing the distribution of victim age across offense types, gender, and race are displayed in Figure 1. Separate lines are presented for the victims of each offense type, and the vertical lines represent the average age of victims for that respective offense, with the top and bottom 5% of victim ages trimmed. The density plots in Figure 1 suggest differential victimization risk across demographic subgroups—there were higher densities of young Black victims and the average age of Black victims of each offense tended to be younger than White victims. This was particularly the case with homicides. For male and female White homicide victims, the age distribution was relatively uniform, centered at 40 and 39 years old for males and females, respectively. For Black homicide victims, the distribution was substantially denser prior to 25, with average ages of 32 and 30 for males and females, respectively.

OPEN IN VIEWER

Figure 1.

Homicide, aggravated assault, and robbery victim age density by race (Black, White) and sex.

Utilizing data from the 2010 Census, population counts were obtained for the demographic subgroups of interest—victim age, sex, and race. For the purposes of these analyses, “young victims” were defined as those victims between the ages of 15 and 24, when the risk for violent victimization has been noted to be significantly higher (Truman & Langton, 2014), particularly from the mid-1980s onward (Dahlberg, 1998). Specifically, estimates are produced comparing the violent victimization rate per 10,000 residents for all Michigan residents, young men and women; race and gender combinations; and then age, race, and gender combinations.

In general, Table 2 suggests an interaction between age, race, and gender in producing variation in violent crime victimization rates. The overall homicide rate per 10,000 residents for the State of Michigan is 0.64. The homicide victimization rate for young males is 3 times the rate for all residents (1.81 per 10,000), and the rate for young Black males is 15.5 times that (9.33 per 10,000). These estimates are consistent with national trends, placing homicide as the leading cause of death among Black males age 15–24 (Centers for Disease Control, 2014). This pattern manifests across homicides, aggravated assaults, and robberies. For each of the violent crime types, the victimization rate for young Black males is the highest among all age, race, and gender combinations. The aggravated assault rate for young Black women is an exception, as it is nearly equivalent to that of young Black males (197 vs. 200 per 10,000).

OPEN IN VIEWER

Table 2.

Violent Crime Victimization Rates (per 10,000) by Demographic Subgroups.

		Homicides		Aggravated Assaults		Robberies
Subgroups	Population	Victims	Rate	Victims	Rate	Victims	Rate
State of Michigan	9,883,640	629	0.64	26,668	26.98	10,390	10.51
Young females	690,625	34	0.49	3,710	53.72	1,336	19.34
Young males	718,046	130	1.81	4,420	61.56	2,411	33.58
Black females	735,148	77	1.05	6,910	93.99	2,401	32.66
White females	3,957,000	63	0.16	4,627	11.69	1,417	3.58
Black males	665,214	373	5.61	7,701	115.77	3,600	54.12
White males	3,846,120	94	0.24	6,615	17.20	2,600	6.76
Young Black females	121,119	24	1.98	2,389	197.24	887	73.23
Young White females	510,526	9	0.18	1,240	24.29	411	8.05
Young Black males	123,277	115	9.33	2,466	200.04	1,360	110.32
Young White males	533,784	14	0.26	1,814	33.98	952	17.83

Note. “Young” refers to individuals between the ages of 15 and 24.

Geographic Variation in Victimization Risk

One possible utility of the incident-level data contained in a LIBRS system like MICR is the capability to examine regional variation in violent crime victimization. It was possible to count the number of unique violent crime victims across the Michigan law enforcement jurisdictions, allowing for the calculation of violent crime victimization rates for each locality as well as demographic subgroups of interest within each locality. ⁴ As the MICR data contain incidents reported by law enforcement agencies at multiple levels of nested geographic units (i.e., local, county, state), these analyses focus on the 333 local law enforcement agencies reporting for the 2013 MICR, covering 89% of all unique victims in the data.

Homicides were particularly concentrated, with 261 jurisdictions (78%) reporting zero homicides and 325 (97.6%) reporting five or fewer homicides. The city of Detroit accounted for the largest proportion of homicides (329, 57%, 4.7 per 10,000), yet Saginaw and Flint have similarly high homicide rates (4.8 and 5.7, respectively), yet are much smaller cities, each with a population one tenth and one fifth as large as Detroit. Across all three offense types, there are five jurisdictions that appear in the top 5% of victimization rates for each—Detroit, Flint, Inkster, Muskegon Heights, and Saginaw. Of these cities, Detroit, Flint, and Saginaw are particularly noteworthy for their historical levels of violent crime extending back into the 1980s (Matthews, 1997). For instance, in the 2013 MICR data, these three cities comprise 65% of all homicides, 42% of all aggravated assaults, and 59% of all robberies—further highlighting that violent crime is heavily concentrated in particular regions that may be higher priority targets for resource allocation (Messner et al., 1999).

Differential Victimization Risk Within High-Rate Jurisdictions

The current analysis of MICR data indicates that among Michigan residents, violent crime victimization is more common among young people, African Americans, and men. Additionally, across all Michigan cities, the violent crime victimization rate was found to be substantially higher for particular combinations of these subgroups, particularly young Black males. Demonstrating the utility of these data to inform resource allocation for strategic violence interventions, the following examines variation in violent crime victimization among high-risk groups in the high victimization rate cities—Detroit, Flint, Inkster, Muskegon Heights, and Saginaw—in comparison to the victimization rates for the same groups across the entire state. ⁵

Figure 2 displays homicide, aggravated assault, and robbery victimization rates for all residents, males, young males, and young Black males in the entire state and in the high victimization rate cities. As the resolution moves from larger to relatively smaller units, there is considerable variation in the nature of violent victimization risk. Homicide victimization risk for young males (age 15–24) is nearly twice that of all Michigan males, and homicide victimization rate for young Black males across Michigan is more than 9 times higher. When considering specific high victimization risk areas, the homicide rates in Flint and Muskegon Heights—relatively smaller cities with extremely high rates of violence—were more than double the homicide victimization rate among young Black males across all of Michigan and up to 50 times that of male residents in general.

OPEN IN VIEWER

Figure 2.

Variation in high-risk group victimization across high violence cities.

Comparing victimization rates across crime types and high-rate cities, the results suggest distinct crime problems for specific agencies. For instance, Muskegon Heights and Saginaw displayed the highest aggravated assault and homicide rates for young Black males, particularly relative to the much larger city of Detroit. However, the robbery victimization rate among the same cities is 60% lower than Detroit. Indeed, with the exception of robbery, the city breakdown suggests considerably higher victimization risk across all crime types concentrated among young Black males in relatively smaller cities (i.e., 10,000–100,000 residents). These findings are consistent with research suggesting that the association of race and violence is magnified within urban areas (Lauritsen, 2001).

Correlates of Victimization Risk Across Jurisdictions

The previous analyses utilize incident-level data to better define subpopulations at risk and identify where the risk of violence is greatest. In order to gain a better understanding of the underlying factors which contribute to variation in violent victimization across the state, we estimate a series of two-stage least squares regression models. The dependent variable in these models is the excess risk of homicide, aggravated assault, or robbery victimization, operationalized as the difference between the jurisdiction victimization rate and the state victimization rate. Positive values of excess risk suggest that the city residents experience disproportionate violent victimization, comparing actual victimization rates to expected rates (Socia, 2016). The regression models examine the relationship between excess risk for homicides, aggravated assaults, and robberies and the distributions of a set of covariates in order to state how those covariates contribute to variation in excess risk.

The jurisdiction-level covariates were constructed using 2013 5-year estimates from the American Community Survey, and the UCRs, linked by the Law Enforcement Agency Identifiers Crosswalk file (National Archive of Criminal Justice Data, 2007). These covariates are described in Table 3. The variables population density and urbanicity were included to capture variation in violent victimizations across urban and rural locations. The variables racial and ethnic heterogeneity, ⁶ concentrated disadvantage, and residential instability represent traditional measures of social disorganization which have been found to be ecological correlates of violent victimization (Land, McCall, & Cohen, 1990; Lauritsen, 2001; Pratt & Cullen, 2005) and are included here to capture potential fundamental causes of risk. Police density is included as a measure of law enforcement strength, including totals of local and state officers allocated to the jurisdiction.

OPEN IN VIEWER

Table 3.

Jurisdiction-Level Covariates Used in Multivariate Modeling.

Covariate	Description	Source
Excess risk (outcome variable)	Difference between the crime-specific victimization rate in the jurisdiction and in the state as a whole. These measures are logged in the regression models.	MICR (2013) ACS (2013)
Population density	Total population divided by square miles of land area.	ACS (2013)
Urbanicity	Rural jurisdiction = 1; urban = 2; metropolitan = 3.	USDA (2013)
Racial/ethnic heterogeneity	Herfindahl index for Black, Asian, American Indian, Hispanic, and White residents.	ACS (2013)
Concentrated disadvantage (α = .91)	Weighted factor score extracted from % in poverty, % unemployed, % single parent households, median household income, and % receiving supplemental security income.	ACS (2013)
Residential instability (α = .81)	Weighted factor score extracted from % moved in past year, % renter occupied households.	ACS (2013)
Police density	Number of law enforcement officers per 1,000 residents.	Uniform Crime Reports (LEOKA, 2013)
Spatial lag	Average of the untransformed crime-specific excess risk in spatially proximate areas. Logged in regression models.	MICR (2013) and ACS (2013)

Note. MICR = Michigan Incident Crime Reporting; ACS = American Community Survey; USDA = US Department of Agriculture; LEOKA = Law Enforcement Officers Killed and Assaulted.

Additionally, we include a spatial lag covariate to account for spatial dependence. Using an inverse squared distance matrix of jurisdiction centroids, Monte-Carlo simulations of global Moran’s I statistics indicated positive spatial autocorrelation for homicides (I = .11, p = .026), aggravated assaults (I = .10, p = .038), and robberies (I = .63, p < .001). These tests suggest that excess risk in a given jurisdiction is significantly influenced by the excess risk of nearby areas, biasing estimates of the association between covariates and excess risk (Kubrin & Weitzer, 2003). Crime-specific spatial lags were estimated as the average excess risk in spatially proximate jurisdictions and included as a model covariate. Because the excess risk spatial lag would be correlated with the error term in a standard ordinary least squares regression (i.e., creating endogeneity), we follow the example of Socia (2016), utilizing a two-stage least squares regression, with spatial lags for racial/ethnic heterogeneity, concentrated disadvantage, residential instability, and police density as instrumental variables predicting the excess risk spatial lag. The models were implemented using the “ivreg” function in R (Kleiber & Zeileis, 2008) and use heteroscedasticity consistent (HC3) robust variance estimators.

Table 4 displays two-stage least squares regression models examining the relationship between the covariates and the excess risk of homicide, aggravated assault, and robbery victimization. Using a square root of the variance inflation factor larger than 2 as an indicator of multicollinearity (Fox, 1991), no covariates exceeded this threshold in any model, and additional diagnostics indicated that the instruments were sufficiently strong (p < .001 in all models). Across each of the crime types, racial/ethnic heterogeneity and concentrated disadvantage were the most consistent predictors of excess risk—the higher the level of these indicators, the higher the excess risk for violent victimization. Other covariates were not consistently associated with variation in excess risk across all crime types. Robbery excess risk was also correlated with indicators of population density and metropolitan areas. Residential instability and police density were uncorrelated with excess risk in any model.

OPEN IN VIEWER

Table 4.

Two-Stage Least Squares Regression Models of Violent Victimization Excess Risk on Jurisdiction Covariates.

Dependent Variable: ln Excess Risk	Model 1: Homicides	Model 2: Aggravated Assaults	Model 3: Robberies
Covariates	Estimated (RSE)	Estimated (RSE)	Estimated (RSE)
√Population density	0.01 (0.01)	0.01 (0.00)	0.02 (0.00)***
Rural (ref = urban)	0.36 (0.35)	0.03 (0.20)	−0.30 (0.26)
Metropolitan (ref = urban)	0.36 (0.24)	0.12 (0.17)	0.46 (0.20)*
Racial/ethnic heterogeneity	2.60 (0.63)***	1.18 (0.39)**	2.65 (0.44)***
Concentrated disadvantage	0.39 (0.11)***	0.53 (0.05)***	0.37 (0.06)***
Residential instability	0.06 (0.08)	0.09 (0.05)	−0.07 (0.06)
ln police density	−0.24 (0.17)	0.18 (0.11)	−0.02 (0.11)
ln spatial lag	0.28 (0.30)	−0.02 (0.25)	−0.11 (0.24)
Intercept	−3.28 (0.50)***	2.12 (0.76)**	−0.68 (0.32)*
Wald’s test (df)	11.72 (8)***	24.83 (8)***	35.48 (8)***
Weak instruments (df)	11.06 (4)***	20.36 (4)***	14.12 (4)***
Adj. R 2	.15	.37	.45

Note. n = 333. RSE = heteroscedasticity consistent robust standard error.

*p < .05. **p < .01. ***p < .001.

Limitations

There are several limitations which merit consideration. The current inquiry utilized incident-level crime data to estimate differential violent victimization risk. Due to the nature of the data generation, these analyses can only claim to represent violence reported to the police, excluding offenses that never came to the attention of law enforcement. This introduces bias into estimation of victimization rates from incident-level data, downwardly biasing estimates relative to summary statistics (Addington, 2008) and victimization surveys (Lynch & Addington, 2007). It is worth noting that violent offenses, particularly the serious violent offenses of interest here, have tended to be reported to the police more reliably than other offenses. The most recent data from the NCVS suggests that 68% of robbery victims nationally report the incident to the police, as do 64% of aggravated assault victims (Truman & Langton, 2014). These figures stand in comparison to 36% of all property crime victims nationally reporting their victimization to the police (Truman & Langton, 2014).

Unit nonresponse is a validity threat for research utilizing NIBRS and similar incident-level data sets (Addington, 2008). Fortunately, in the context of the current inquiry, approximately 94% of Michigan law enforcement agencies submitted full or partial data to MICR, resulting in significant coverage of the state population. Due to the nature of these data, it is recognized that the current inquiry can only describe relative variation in violent victimization risk across jurisdictions and subgroups, not absolute victimization risk. With the relatively high reporting rate of violent offenses and the broad agency participation and population coverage in the MICR, it is unlikely that these limitations systematically bias the capability of the current analysis to examine such relative variation.

Further, we sought to demonstrate the utility of NIBRS data for informing focused interventions through a single year of data from MICR, utilizing individual agencies as the unit of analysis. It may be useful in future endeavors to draw on multiple years of data in order to analyze trends or assess the impact of interventions by creating measures of victimization prior to and following implementation. Additionally, depending on the level of detail maintained in LIBRS systems, it would be possible to utilize smaller, potentially more informative units of analysis. Although not feasible in the current inquiry, as long as geographic identifiers are maintained, it would be possible to link incidents and the characteristics of incidents to neighborhood or street segments (Braga et al., 2010), allowing additional granularity to describing at risk populations and geographies to inform problem-oriented strategies.

We sought to demonstrate the utility of incident-based crime data for problem analysis in the explicit absence of local area surveys. We recognize that this approach limits the available risk measure to mostly static, individual features (i.e., age, race, gender) and omits salient correlates such as victim–offender overlap (Berg, Stewart, Schreck, & Simons, 2012), and gang membership (Taylor, Peterson, Esbensen, & Freng, 2007), and best measured through surveys.

Finally, the current inquiry uses Michigan as a case study in the application of a relatively complete LIBRS system to a problem analysis of violent victimization. It is unclear whether the specific patterns in victimization risk and correlates of that risk will generalize to other states, but the purpose of a public health approach to violence is to treat such determinations as empirical questions. To this extent, as long as states build NIBRS capacity to encompass the majority of their law enforcement agencies, then similar analyses as those reported here can be pursued.

Policy Implications

The findings suggest implications for expanded use of NIBRS systems toward an informed distribution of state-level resources for violence prevention. The analyses presented here, as well as in previous research (e.g., Vasquez et al., 2005), demonstrate the utility of NIBRS-related data systems to allow for much more precise analyses of violence patterns within states than is available in the UCRs system. Further, given the current inability of the NCVS to produce subnational estimates (Fay & Diallo, 2012), NIBRS and LIBRS systems can be utilized as substitutes or to complement research goals. Together, these implications support recent calls for the expansion of NIBRS systems. Increased emphasis on crime analysis, strategic problem solving, and data-driven decision-making should logically incentivize the expansion of NIBRS systems (Griffith, 2014; Klofas et al., 2010). Bierie’s recommendations for simplifying data collection and retrieval would similarly support expanded adoption (Bierie, 2015).

The analyses also demonstrate the ability to cover large geographic areas, in this case a state, while also examining subunits such as counties or cities/towns. Although not addressed in the current study, augmenting LIBRS systems with geographic identifiers also allows for analyses at the local level such as a particular city with the ability to analyze risk at very fine place-based units such as street segments or specific addresses. Indeed, recent research has suggested that processes at micro units such as street segments drive substantial variation in patterns of violence over time (Braga et al., 2010). Thus, NIBRS can support research following Taylor’s (1997) concept of the “Cone of Resolution,” whereby researchers and analysts can move from larger to smaller units of analysis (nation, region, state, county, city, neighborhood, street segment, address), particularly when local and state agencies enhance data collection with detailed geographic identifiers. Such enhanced NIBRS data systems can thus support decision-making at national and state levels while also enabling law enforcement agencies, public health agencies, and various community partners to move to ever smaller units of analysis (e.g., neighborhoods or street segments) to support focused, data-driven efforts.

Finally, the results suggest the need and the opportunity to develop comprehensive criminal justice and public health interventions that include both long- and short-term violence prevention (see also Braga & Weisburd, 2015; Weisburd, Davis, & Gill, 2015). In the case of Michigan, the extreme risk differential for young Black men in several of the state’s communities calls for targeted, evidence-based violence prevention. As an example, Michigan’s Secure Cities Partnership focuses state enforcement and economic development resources on select cities experiencing high levels of violence. ⁷ The use of MICR data informs problem-solving approaches that move down the cone of resolution and analyze these violence patterns at the neighborhood levels in cities such as Detroit, Flint, Inkster, Muskegon Heights, and Saginaw. Complete state-level NIBRS data are also useful to state public health agencies—the Michigan Department of Health and Human Services is currently utilizing MICR data to formulate resource allocation for Victims of Crime Act funding, pertaining to victim services for child abuse. Coupled with evidence-based interventions, using LIBRS data to inform resource allocation, identify, and monitor communities with the highest victimization risk holds promise and the greatest likelihood for deploying limited resources to reduce violent victimization.

At the same time, the finding that the risk levels follow structured patterns of concentrated disadvantage suggests that enforcement-oriented interventions, although necessary, are not sufficient. The findings suggest that enforcement alone will miss fundamental causes of variation in violent victimization risk and will not address underlying drivers of victimization risk (Sorg, Haberman, Ratcliffe, & Groff, 2013). To this extent, even highly focused enforcement-based strategies will have a temporary effect since there will continue to be high-risk populations unless the underlying causes of risk are addressed. Among these are the social and economic resources available to at-risk groups. Phelan and Link (2005) argue that flexible social and economic resources matter in two ways. First, they shape individual violence risk behaviors, allowing those with resources to avoid risks and to minimize the consequences of violence. Second, resources shape access to geographies like neighborhoods and workplaces with lower risk and higher protective factors. Thus, the unequal distribution of social and economic resources can account for widespread violent victimization disparities by race and socioeconomic status (Phelan & Link, 2005). Recognizing the importance of these underlying sources of risk, policy responses pairing enforcement with a variety of social and economic approaches have shown promising results, including disorder reduction (Braga & Bond, 2008), enhancing collective efficacy (Banyard, Moynihan, & Plante, 2007), and revitalizing business districts (Hoyt, 2005).

The need for such comprehensive approaches is also suggested by criminological research demonstrating the links between fear, disorder, and neighborhood decline (Skogan, 1990). As fear, crime, and disorder result in withdrawal from community life, these factors produce a downward spiral of further fear, crime, and disorder, with the resulting decline inhibiting economic investment. Thus, short-term enforcement/prevention is necessary to halt the downward cycle but must be complemented with longer term prevention that will allow economic and social investment necessary to address structural patterns of concentrated disadvantage. A contemporary example of such a comprehensive strategy is the Byrne Criminal Justice Innovation program that supports local and national programs combining highly focused enforcement with neighborhood revitalization through law enforcement–community and researcher–practitioner partnerships (see Griffith, 2014). Ongoing programs target police problem-solving to simultaneously reduce violence and increase perceptions of police legitimacy have shown promising results (Bureau of Justice Assistance, 2015). The continued growth of this programming will require means to continually update target for intervention and a means to evaluate performance. Building and enhancing the capacity of local and state agencies to draw on detailed incident-based crime data systems will better enable agencies to implement data-driven, public health-oriented strategies.