Serious Assault Victim and Incident Characteristics Associated With Police Notification and Treatment in Emergency Rooms in Rural,Suburban,and Urban Areas

Abstract

Two common sources of violence rates are police data and hospital data on injuries. It is unclear, however, if violence rates calculated using these sources have the same meaning across location types like rural, suburban, and urban areas. We know that characteristics of assault victims and incidents are associated with whether or not the victim reports the incident to the police. If community type moderates this help-seeking behavior—for example, if the impact of sex, age, race, victim–offender relationship, or injury severity on reporting to the police or being treated in an emergency room (ER) is different in rural relative to urban areas—then this selection bias is problematic for comparisons across location type and for including different location types in the same sample. The goal of this study was to determine which victim and incident characteristics are associated with police notification and ER treatment and to see if these characteristics are the same across rural, suburban, and urban areas. We used National Crime Victimization Survey data on reporting serious assaults to the police and being treated in an ER as a result of a serious assault. Results revealed that (1) police notification is much more likely to occur than ER treatment, (2) a majority of victim and incident characteristics are significantly associated with police notification, (3) few victim and incident characteristics are significantly associated with ER treatment, and (4) there is only chance moderation by location type. Findings suggest that both police and hospital data on serious assaults can be used to make comparisons across rural, suburban, and urban communities without being limited by disproportionate inclusion or exclusion of incidents associated with victim and incident characteristics. However, comparisons between the two data sources should be made with caution due to the different victim and incident characteristics associated with police notification and ER treatment.

Keywords

North America comparative crime/justice quantitative methods Other criminal victimization

Introduction

Understanding violence data sources is an important first step in determining both our trust in those sources to provide valid estimates of violence levels and trends and our ability to use those sources to test criminological theory. If victim or incident characteristics are associated with inclusion in a data source independent of true violence rates, then we must understand better how this selection process works and caution should be taken in relying on the source. Likewise, if victim or incident characteristics are differentially associated with inclusion in a data source based on rural, suburban, and urban location independent of true violence rates, then crime rate comparisons across these community types must be made with extreme caution and with an understanding of the selection process that drives inclusion or exclusion. Two measures of violence, police data from the Uniform Crime Reporting (UCR) Program and outpatient data from hospital records, have undergone different levels of scrutiny, but little research has made direct rural, suburban, and urban comparisons of these data sources and no research has directly compared them with each other. In this study, we use data on victims of serious assault from the National Crime Victimization Survey (NCVS) from 1996 to 2005 to determine whether victim and incident characteristics are associated with police notification and victim treatment in the emergency room (ER). That is, we test to see if victim or incident characteristics are partially responsible for the incident showing up in police or hospital data, and we do this across rural, suburban, and urban areas to see if the selection process operates similarly across these areas.

Background

Measurement of Violence

Depending on the research questions being asked, violence can be measured using a number of different data sources. Common sources include crimes reported to the police and arrest records, victimization surveys, hospital outpatient data, criminal offending surveys, and vital statistics (to measure homicides). In this study, we are interested in measuring violence for the purpose of estimating relative community-based violence rates. That is, the goal is to measure a rate of violence in communities of a certain size (e.g., rural areas) relative to rates in communities of other sizes (e.g., suburban and urban areas). Among other things, such rates are often used to test structural theories of crime in rural communities, using counties as the unit of analysis. For example, in testing social disorganization theory in nonmetropolitan counties, Osgood and Chambers (2000) used official police arrest data to measure youth violence and Kaylen and Pridemore (2011) used hospital violent victimization data.

No criminological data source is free from limitations. Understanding these limitations is a crucial part of criminological research and is vital to deciding which data to utilize. This section offers an overview of methodological explanations for differences in violence rates when using police data and then provides examples of how measurement can affect theory testing. We then provide an overview of the existing literature on two important aspects of police and hospital data: reporting to the police and presenting to the ER.

Methodological explanations for differences in violence rates

When measured using different data sources, violence rates will never be exactly the same. Even when the same data source is used across multiple locations, the validity and reliability of the data vary. Police data is one of the most common sources of information about crimes. Whether or not a crime comes to the attention of the police is an important factor in deciding whether an incident is included in the data. In an analysis of nonlethal violent victimization data between 1973 and 2005, violent incidents were reported to the police only 40% of the time (Baumer & Lauritsen, 2010). More recently, between 1992 and 2000, aggravated assaults were reported to the police approximately 55% of the time (Hart & Rennison, 2003). Many factors affect decisions to report to the police, as will be described subsequently. After a crime comes to the attention of the police, multiple considerations influence whether or not an arrest is made. If no arrest is made, the crime is not counted in arrest data (though it may still be counted in data sets containing crimes reported to the police). If an arrest is made, the police officer must decide what charge to record, whether or not to record it in the UCR system, and then must enter the charge correctly. In some instances, the police department then decides whether or not to send the data to the state clearinghouse and the state then must decide to send the data to the Federal Bureau of Investigation (FBI). In other instances, police departments can send their data directly to the FBI. At every decision point, a mistake or a decision not to include an incident could prevent it from being included in the data.

Effects of measurement on theory testing

Violence rates differ by data source. Descriptively, these differing rates are interesting and it is important to understand why they occur. More importantly, however, in tests of criminological theories, different data sources can lead to different conclusions due to the causes of the differences in the data sources. Kaylen and Pridemore (2011, 2013) illustrated this in a pair of articles in which they found that conclusions about the generalizability of social disorganization theory to rural communities differ based on the measure of the dependent variable. In two different samples, the conclusions drawn were different when youth violence was measured with UCR data (youth arrests for aggravated assaults) and hospital data (youth victims of serious assaults with injuries treated at ERs). Other articles have documented county-level criminological and policy studies that have been invalidated due to measurement errors (e.g., Chu, Rivera, & Loftin, 2000; Loftin & McDowall, 2003; Maltz & Targonski, 2002, 2003; Rivera, Chu, & Loftin, 2002). Thus, a better understanding of factors affecting data sources that measure violence and are used to test criminological theories is required.

Reporting to the Police

Rural, suburban, and urban rates

Few studies look at variations in police reporting across geographic areas. Laub (1981) used the National Crime Survey to study whether urban, suburban, and rural victims report to police at different rates and whether reasons for reporting behavior differ geographically. He found rates do not differ by location but that seriousness of offense was a differentiating factor. Rennison, Dragiewicz, and DeKeseredy (2013) studied situational variations in reporting to police among female victims and found that situational factors—victim–offender relationship, type of violence, victim’s marital status, weapon presence, and victim injury—related to police reporting vary across rural, suburban, and urban areas. Using conjunctive analysis, they found combinations of victim and incident factors form situational contexts that affect reporting. For instance, an incident with an injured female victim and a weapon is most likely to be reported to police in urban areas, but this combination of factors does not affect reporting in suburban or rural areas. Across all situational contexts, victim–offender relationship was consistently related to police reporting in rural, suburban, and urban areas.

Correlates of police notification

There is empirical literature on correlates of victims reporting to the police, much of it utilizing the unique questions asked in the NCVS. The survey asks respondents about their victimization experiences in the previous 6 months, followed by specific questions about all incidents including whether or not the police were notified. This data set provides information on both those who reported their victimizations to the police and who did not. A number of victim, offender, and incident characteristics have consistently been found to be related to victim reporting behavior while others show mixed results.

Victim characteristics commonly studied as they relate to reporting behavior include age, sex, race, income, and marital status. Bosick, Rennison, Gover, and Dodge (2012) found that the rate of reporting crimes to the police increases with age. For aggravated assaults in particular, the percentage reported to police is lowest among 12 to 15-year-olds and increases until it reaches a peak of 64% for 35 to 49-year-olds, where it then decreases slightly to 57% among those aged 65 and older. Other research confirms the findings that young individuals are less likely to report than older individuals (e.g., Hart & Rennison, 2003; Kaukinen, 2002).

The literature on reporting behavior of victims often finds women are more likely than men to report, but there are mixed findings for race (e.g., Catalano, 2006; Hart & Rennison, 2003; Kaukinen, 2002; Skogan, 1984). Bosick et al. (2012) recently confirmed the female–male finding in reporting violent crimes among victims up to age 50, but they found very few race differences. Rennison (2007), on the other hand, found Hispanics are less likely to report the most serious crimes to police than non-Hispanic Whites, but reporting differences between Hispanics and other victims were minimal. Several other studies found minorities are more likely to report than Whites (e.g., Avakame, Fyfe, & McCoy, 1999; Catalano, 2006). Finally, Xie and Lauritsen (2012) found that victims are most likely to report to police in instances of Black offender–Black victim compared to all other Black/White offender/victim combinations.

Income and marital status are two more victim characteristics often associated with reporting a crime to the police. Hart and Rennison (2003) found that victims of violence (with the exception of robbery) are less likely to report to the police as income increases. Police reporting decreases from 60% for those with household income below US$7,500 to 50% for those with household income US$75,000 or more for aggravated assaults. Simple assaults tell a similar story, decreasing from 41% to 30% for those same income categories. Married victims of aggravated assaults are most likely to report victimization to the police, while those who have never been married are much less likely to report victimization to the police than those who are divorced, widowed, or separated (Hart & Rennison, 2003). This marital status finding is consistent with other research (e.g., Avakame et al., 1999).

Victim characteristics are most prevalent in the reporting literature, but offender and incident characteristics are also important. Perhaps the most consistent offender characteristic associated with whether or not a victim reports to the police is the victim–offender relationship. Research has consistently shown that when the victim and offender know each other, the incident is less likely to be reported to the police than when it occurs between strangers (Felson, Messner, & Hoskin, 1999; Kaukinen, 2002; Kruttschnitt & Carbone-Lopez, 2009).

Incident characteristics have historically been important factors in whether or not a crime is reported to the police (Laub, 1981; Skogan, 1984). One of the most consistent incident characteristics associated with police reporting is seriousness of the crime (Gottfredson & Gottfredson, 1988). Crime seriousness can be operationalized in a number of ways including presence of weapon and victim injury. Research has found that weapon use and victim injury increase the likelihood of police notification (Davies, Block, & Campbell, 2007; Gartner & Macmillan, 1995; Hart & Rennison, 2003; Kaukinen, 2002; Xie & Lauritsen, 2012; Xie, Pogarsky, Lynch, & McDowall, 2006). Felson, Messner, and Hoskin (1999) created a seriousness scale based on the Sellin-Wolfgang scale, with points assigned for intimidation and injury characteristics. Crime seriousness was positively associated with victims reporting to the police. Recently, Bosick et al. (2012) found that whether or not a third party was present and whether or not the incident occurred at the victim’s home are consistently associated with reporting violent crimes to the police across all ages. Specifically, between 1992 and 2000, incidents of aggravated assault were reported to police 48% of the time by the victim and 22% of the time by “someone else” such as a bystander (Hart & Rennison, 2003). Planty (2002) likewise found that third-party presence increased reporting to police.

Alternative Measure of Violence Rates: Hospital Data

Kaylen and Pridemore (2011) proposed using hospital outpatient data to measure violence in rural communities after research showed serious measurement errors associated with police data, especially in sparsely populated rural counties (Lott & Whitley, 2003; Maltz & Targonski, 2002, 2003). Unlike official crime data, hospital data are not reliant on police notification or arrest. About half of all violent victimizations are not reported to the police, so selection bias likely has an effect on these data. Criminologists have utilized hospital data on occasion and the use of these data is common among epidemiologists who study violence (e.g., Fabio, Li, Strotmeyer, & Branas, 2004; Fabio, Sauber-Schatz, Barbour, & Li, 2009; Freisthler, Gruenewald, Ring, & LaScala, 2008; Gruenewald et al., 2010; Gruenewald, Freisthler, Remer, LaScala, & Treno, 2006). Despite the use of these data, much is still unknown about which victims of violence present to the ER and what effect selection bias has on hospital data.

Two articles by Kaylen and Pridemore (2011, 2013) examined the use of hospital data to measure violence at the county level. First, they completed a partial replication of Osgood and Chambers’s (2000) highly cited test of social disorganization theory in rural areas. Instead of relying on arrest data, Kaylen and Pridemore (2011) utilized hospital outpatient data. While Osgood and Chambers concluded the theory generalized to rural areas, Kaylen and Pridemore concluded it did not. In their companion piece, Kaylen and Pridemore (2013) tested three hypotheses about why the conclusions from these two studies were so different. Specifically, they tested if the differences were due to measurement of the dependent variable (arrest vs. hospital data), different samples (four states vs. one different state), or spatial autocorrelation (tested and controlled for in the partial replication). They concluded that measurement of the dependent variable accounted for the differences in the two studies, though this conclusion does not necessarily mean hospital data are the best measure of community-level violence.

A number of characteristics of hospital data could potentially threaten their validity as a measure of violence. The first consideration is selection bias, namely, whether or not victims of violence actually receive medical treatment in the ER when they are injured. If estimates of the actual levels of violence are of interest to the researcher, this could potentially be a problem. If relative rates (e.g., comparing communities) are sought—as in tests of community-level theoretical models—this would only be a problem if rates of presenting to the ER for injuries resulting from violence vary nonrandomly by community size. That is, it would be problematic if a factor related to likelihood of ER medical treatment in one community type is not related to ER medical treatment in another community type.

In the first study of its kind, Kaylen and Pridemore (2015) directly examined whether victims of serious violence receive medical treatment in the ER at the same percentages in rural, suburban, and urban areas. They concluded ER treatment is independent of location but a significant association exists between ER treatment and the interaction of location and reporting to police. That is, the relationship between location and ER treatment is different for those who do and do not report to the police. Victims who reported to the police also received treatment in the ER 35% of the time in urban areas, suburban 33%, and rural 38%. Victims who did not report to the police did receive medical treatment in the ER 17% of the time in urban areas, suburban 13%, and rural 9% (Kaylen & Pridemore, 2014).

Kushel, Perry, Bangsberg, Clark, and Moss (2002) pointed out that injuries from violent crimes are suitable for emergency care because they require urgent attention, often occur where other treatment is not available, and in serious instances require services that are not available in nonemergency sites. This immediate need for medical treatment will likely outweigh factors that prevent victims from reporting to the police (Wood, 2010). Even offenders report going to the hospital 90% of the time to seek medical treatment when they are seriously injured (May, Hemenway, & Hall, 2002). Research has not examined whether victim and incident characteristics that affect the decision to report to the police likewise affect ER treatment. That is, there may be systematic underenumeration of incidences of violence based on these characteristics.

Aside from whether or not the victim presents to the ER, the other potential limiting factor of hospital data is how the incident is recorded. Just like police data go through a filtering process, so do hospital data. Once a victim presents to the ER, they are assigned an external cause of injury code (E-code). Research suggests injuries are correctly classified. In over 90% of the cases, E-code classifications applied at the time of the event and subsequent reviews by experts are in agreement (LeMier, Cummings, & West, 2001; Meux, Stith, & Andra, 1990). The next step in the official filtering process is the compilation of the data in the hospital recording system and any subsequent state recording systems. Kaylen and Pridemore (2013) report hospital data are not collected by a single entity like the FBI’s UCR Program for police data, but many states have statewide collection systems. In some instances, state hospital associations collect these data and in other instances state health departments perform this task. Thus, it appears the main concern about the validity of hospital data as a measure of violence is whether some victims are systematically more or less likely to receive medical treatment in the ER.

Research Questions

The purpose of this study is to examine a potential source of measurement error—nonrandom inclusion—in police and hospital serious assault data. This goal is achieved by asking the following questions: Which victim and incident characteristics are associated with police notification and with treatment in the ER? And are these characteristics the same for rural, suburban, and urban victims of serious violence?

Data and Method

Data

The data for these analyses came from the NCVS. Specifically, we used concatenated incident extract files from data years 1996–2005. Combining 10 years of data is necessary to have a large enough sample for the analyses. These files combine incident, person, and household-level files, such that each reported incident has the associated incident and victim variables attached to it. These data files are publicly available and were obtained from the National Archives of Criminal Justice Data website and produced by the Interuniversity Consortium for Political and Social Research. The exact files used here are Study # 4699, NCVS, 1992–2005: Concatenated Incident-Level Files.

The NCVS offers a number of unique features that make it suitable for our study. Specifically, it includes data about crime incidents both that were and were not reported to authorities. The NCVS includes crime incidents that overlap with police and hospital data, but it also includes crime incidents that are not known to law enforcement or medical officials and thus are not included in other data sets. Further, the NCVS links victim and incident characteristics to crime incidents and personal characteristics to those who were not victimized. The combination of these unique features of the NCVS makes it the most appropriate data set for our study.

Sample

The sample for this study included any incidents that could potentially be listed as aggravated assaults in police data and/or assaults in hospital outpatient data. As a result of these selection criteria, some incidents included in the sample would not be included in aggravated assault police data (e.g., robberies with injuries) and some would not be included in hospital outpatient assault data (e.g., attempted aggravated assault with weapon). Given the unique nature of sexual assaults (attempted and completed rape and sexual attack), as they relate to police reporting and medical treatment, we excluded them from our analysis. Using the Type of Crime variable (V4529), we included the following categories of crime in the sample: completed robbery with injury from serious assault, completed robbery with injury from minor assault, attempted robbery with injury from serious assault, attempted robbery with injury from minor assault, completed aggravated assault with injury, attempted aggravated assault with weapon, and simple assault completed with injury. These sample selection criteria yielded a sample of 6,808 assault incidents. After dropping cases with incomplete data on current marital status and race (discussed subsequently), the final sample included 6,692 incidents.

Dependent variables

The dependent variables for this study were two responses to a violent victimization incident: if the police were notified about the incident and if the victim was treated in an ER or clinic. Variable V4399 was used to measure if the police were notified about the incident. Police notification was measured as the respondent saying yes, the police were informed or found out about the incident. Variable V4133 was used to measure whether or not the victim was treated in an ER/clinic. Respondents were first asked if they were injured (variable V4111). If they responded yes, they were injured, they were asked if they received medical care for the injury (variable V4127). If they responded yes, they received medical care for the injury, they were then asked a series of questions about where they received treatment. ER/clinic treatment was measured here as the respondent saying that they received medical treatment in an ER or emergency clinic (variable V4133).

Independent variables

The independent variables of interest in this analysis were broken down into two categories: victim characteristics and incident characteristics. These variables have been established in the police reporting literature as being important.

Victim variables included age, sex, race, household income, and marital status. Age was a continuous variable, created using variable V3014, the respondent’s allocated age. Female was a binary variable used to measure sex, with female being 1 and male being 0. This variable was created using variable V3018, the respondent’s allocated sex. Race was measured using a dummy variable for Non-Hispanic White (with all other races as the reference category). This measure was created using a series of variables from the survey. First, variable V3023 is the respondent’s allocated race. This variable was discontinued after 2002, Quarter 4. Variable V3023A replaced V3023 at the beginning of 2003, Quarter 1, and included more race categories and allowed for multiple races to be selected. Finally, V3024 is the respondent’s Hispanic origin. Non-Hispanic White includes those who responded they were White (V3023) or only White (V3023A) and non-Hispanic (V3024). Cases in which race was not listed (n = 74) were dropped. Household income was measured with variable V2026 and is on an ordinal scale with 14 categories ranging from US$0–US$5,000 to greater than US$75,000. We operationalized household income with a series of dummy variables for lowest income (less than US$25,000), low income (US$25,000–US$49,999), middle income (US$50,000–US$74,999), high income (at least US$75,000), and income unknown. Finally, marital status was measured with variable V3015, the respondent’s current marital status. Marital status was operationalized as married if the respondents said they were married at the time of the survey and not married if the respondents said they were widowed, divorced, separated, or never married. Due to the small number of respondents who did not report marital status (n = 42), these cases are excluded from the analysis.

Incident variables included series victimization, weapon, third-party presence, incident location, unknown offender, robbery, and geographic location. Series victimization was measured using variable V4019, the last question in a set of questions used to determine if an incident should be classified as a series. An incident is coded as a series victimization if the respondent reported six or more similar incidences occurred in the previous 6 months, and the respondent could not describe each in detail. Weapon use is one measure of crime severity and was operationalized with three dummy variables using variable V4049. The first variable, weapon, is coded as 1 if the respondent reports that the offender used a weapon.¹ The second variable, weapon unknown, is coded as 1 if the respondent does not know if the offender used a weapon or if the respondent did not answer the question. The third variable, no weapon, is the reference variable and is coded as 1 if the respondent reported the offender did not use a weapon. Third-party presence was measured using variable V4184, which asks if anyone other than the victim and offender were present during the incident. Third party was a dummy variable with 1 being a third party present. Incident location was measured using variable V4024, which asks the respondent where the incident happened. Private property was a dummy variable with 1 being the incident happened at or near the respondent’s residence or a friend’s/relative’s residence. “At or near” includes detached buildings on the property; the yard, sidewalk, or driveway; an apartment hall, storage area, or laundry room; or the street immediately adjacent to the residence.

Offender status was operationalized with three dummy variables and was measured using variables V4241 (single offender) and V4256 (multiple offenders). The first variable, unknown offender, was coded as 1 if the single offender was a stranger or if all the offenders were strangers. The second variable, known offender, was the reference category and was coded as a 1 if the single offender was known or if any of the multiple offenders were known. The third variable, offender status unknown, was coded as 1 if the respondent reported being unsure whether or not the offender or offenders were known or if the respondent did not answer the question.² Robbery was a second measure of crime seriousness. According to the UCR hierarchy rules, robberies are a more serious crime than aggravated assaults. The robbery variable was created with variable V4529, the Type of Crime variable.

Geographic location was operationalized with three dummy variables and was measured using variable V2129, the Metropolitan Statistical Area of the respondent. These designations are determined by the U.S. Office of Management and Budget and define central city, outside central city, and nonmetropolitan. The urban, suburban, and rural terms are commonly used with these data (e.g., Rennison, Dragiewicz, & DeKeseredy, 2013) and are used here. Rural was coded as 1 for rural and 0 for suburban and urban, suburban was coded as 1 for suburban and 0 for rural and urban, and urban was coded as 1 for urban and 0 for rural and suburban.

Method

We estimated two logistic regression models to address the question of whether rural, suburban, and urban victims of violence have the same victim and incident characteristics associated with police notification and treatment in an ER. The first model has police notification as the outcome variable and the second model has treatment in an ER as the outcome variable. The independent variables described above were included in the models, as were the interactions between two of the geographic location dummy variables (suburban and urban) and each of the victim and incident variables. Due to the large number of variables, we estimated the main effects as the first step of each model. We then individually add groups of interactions effects (e.g., Suburban × Female and Urban × Female) to the models.

The NCVS employs a stratified multistage cluster sample design, and we took steps to adjust for this. Traditional statistics that assume a simple random sample cannot be used for calculating standard errors (SE) with these data. The clustering of households results in variances that are smaller than they would be using a simple random sample design. The NCVS data include two variables that can be used to help adjust for the complex sample design: V2117 is a pseudo-strata variable and V2118 is a pseudo-primary sampling unit (PSU) variable. Stata includes survey commands (svyset) to adjust for complex sample designs and survey weights. We specified the complex sample design of the NCVS using V2117 and V2118, selecting Taylor Series approximation for variance estimation, and indicating data should be centered when a stratum had only one PSU in it. The survey command also allows for survey weights to adjust the NCVS sample of victims in a given year to correspond to the population from which they were drawn (persons aged 12 years and older in U.S. households). The person weight (V3080) used in the current analyses was generated by the Census Bureau and accounts for nonresponse and under coverage of certain groups (Rennison & Rand, 2007).

Model Diagnostics

One of the assumptions of logistic regression is that the model is properly specified. Stata has a command (linktest) that can be used to detect specification error (StataCorp, 2013). Linktest uses the linear predictor value from the model and the linear predictor value squared as predictors to rebuild the model (e.g., regress the dependent variable on the linear predictor and line predictor squared). This test is based on the work of Tukey (1949) and Pregibon (1979). A properly specified model generally has a statistically significant linear predictor value and a nonstatistically significant squared predictor value. A misspecified model generally has statistically significant predictor and squared predictor values. For the police notification model, the linear predictor was significant (b = 1.10, SE = 0.08, p < .01), while the squared value was not (b = −0.12, SE = 0.05, p = .16). Therefore, we can reject the hypothesis that the model is completely misspecified. For the ER model, the linear predictor was nearly significant (b = 0.65, SE = 0.39, p < .10), while the squared value was not (b = −0.11, SE = 0.11, p = .35). Therefore, we can reject the hypothesis that the model is completely misspecified.

To assess model fit and identify outlying and highly influential observations, we calculated standardized Pearson residuals, deviance residuals, and leverage values. We plotted these values against the predicted probabilities of the models. These calculations were done on models using the raw data because they cannot be done with the survey commands. That is, these values were calculated with models that do not take into account survey weights or adjustments for survey sample design. The police notification model exhibited no pattern in the residuals and none of the observations stood out as potential outliers or as highly influential. The ER model exhibited a slight pattern of larger standardized Pearson residuals at lower predicted probabilities than at higher predicted probabilities. This pattern is minor and not cause for concern. None of the observations stood out as potential outliers or as highly influential.

Results

Descriptive Statistics

Table 1 presents the descriptive statistics for the variables used in the logistic regression models. Rural areas had the fewest raw incidents (n = 821), followed by urban (n = 2,658) and suburban (n = 3,213). The percentage of incidents resulting in police notification was similar for rural (60.05%), suburban (58.54%), and urban (59.82%) areas. The percentage of incidents in which the victim received treatment at the ER was much lower than police notification, and rural (13.64%), suburban (13.23%), and urban (15.46%) areas had a little more variation in their percentages.

Table 1.

Frequency and Percentage of Each Variable Among Serious Assault Incidents, by Location, 1996–2005.

	Rural	Suburban	Urban
	n = 821	n = 3,213	n = 2,658
Response
Police notification	494 (60.05%)	1,886 (58.54%)	1,598 (59.82%)
ER treatment	112 (13.64%)	428 (13.23%)	414 (15.46%)
Victim variables
Female	399 (48.60%)	1,451 (45.16%)	1,201 (45.18%)
Male	422 (51.40%)	1,762 (54.84%)	1,457 (54.82%)
Age: mean, SD	28.99, 13.89	29.17, 14.08	29.99, 14.10
White non-Hispanic	666 (81.12%)	2,455 (76.41%)	1,518 (57.11%)
Non-White	114 (13.89%)	430 (13.38%)	692 (26.03%)
Lowest income	388 (47.26%)	955 (29.72%)	1,223 (46.01%)
Low income	206 (25.09%)	871 (27.11%)	635 (23.89%)
Middle income	78 (9.50%)	492 (15.31%)	236 (8.88%)
High income	41 (4.99%)	486 (15.13%)	202 (7.60%)
Income unknown	108 (13.15%)	409 (12.73%)	362 (13.62%)
Married	169 (20.58%)	749 (23.31%)	455 (17.12%)
Not married	652 (79.42%)	2,464 (76.69%)	2,203 (82.88%)
Incident variables
Series incident	35 (4.26%)	133 (4.14%)	93 (3.50%)
Not series incident	786 (95.74%)	3,080 (95.86%)	2,565 (96.50%)
Weapon	333 (40.56%)	1,276 (39.71%)	1,182 (44.47%)
No weapon	457 (55.66%)	1,805 (56.18%)	1,312 (49.36%)
Weapon unknown	31 (3.78%)	132 (4.11%)	164 (6.17%)
Private location	428 (52.13%)	1,364 (42.45%)	1,177 (44.28%)
Not private location	393 (47.87%)	1,849 (57.55%)	1,481 (55.72%)
Third party	554 (67.48%)	2,169 (67.51%)	1,698 (63.88%)
No third party	267 (32.36%)	1,046 (32.45%)	957 (35.88%)
Unknown offender	154 (18.76%)	921 (28.66%)	882 (33.18%)
Known offender	638 (77.71%)	2,165 (67.38%)	1,629 (61.29%)
Offender status unknown	20 (2.44%)	73 (2.27%)	99 (3.72%)
Robbery	62 (7.55%)	334 (10.40%)	418 (15.73%)
Not robbery	759 (92.45%)	2,879 (89.60%)	2,240 (84.27%)

Note. ER = emergency room; SD = standard deviation.

We found a number of rural, suburban, and urban differences in victim characteristics in the sample. A lower percentage of urban and suburban serious assault victims were female (45.18% and 45.16%, respectively) compared to rural (48.60%). The average age of serious assault victims was between 29 and 30 in all three locations. In terms of race, White non-Hispanics made up a much larger percentage of victims of serious assaults in rural (81.12%) and suburban (76.41%) than urban (57.11%) areas. A greater proportion of suburban victims had high incomes than rural and urban victims. Finally, a lower percentage of urban (17.12%) victims were married compared to rural (20.58%) and suburban (23.31%) victims.

We also found several rural, suburban, and urban differences in incident characteristics. A greater percentage of urban serious assault incidents involved a weapon (44.47%) than rural (40.56%) and suburban (39.71%) incidents. The majority of rural incidents happened in private locations (52.13%), while suburban (42.45%) and urban (44.28%) incidents were less likely to happen in private locations. Third parties were present in a smaller percentage of urban (63.88%) incidents than rural and suburban incidents. Unknown offenders represented a higher percentage of offenders in urban (33.18%) incidents than rural (18.76%) and suburban (28.66%) incidents. Finally, the percentage of violent victimizations characterized as robberies was higher in urban (15.73%) compared to rural (7.55%) and suburban (10. 40%) areas.

How to Interpret the Interaction Terms

Before describing the specific results of the logistic regression models, it is important to clarify how to interpret the interaction effects in the models. Victim and incident characteristics are of primary interest in this study, while the location variables (rural, suburban, and urban) are moderating variables. The interaction terms were used to test if the independent variable-–dependent variable relationships were moderated by location. For example, if the interaction between female and urban was positive and significant in the police notification model, it means the sex–police notification relationship is different for urban and rural areas (rural is the reference category). That is, the slope of the female variable is different for urban and rural residents. When a significant interaction is included in the model, the significance of and the coefficient for the simple main effect of the variable (e.g., female in the example above) is not interpreted. Instead, this simple main effect is now a conditional coefficient in the model.

Police Notification Model

Figure 1 displays the predicted probabilities of police notification based on the logistic regression model. The observed average probability of police notification is 0.59. Most of the predicted probabilities fall above 0.25 and below 0.75, with the distribution around 0.59 being fairly bell shaped. This figure reflects the rural, suburban, and urban percentages of police notification described above and presented in Table 1.

Figure 1.

Predicted probabilities of police notification from the logit model.

Table 2 presents the results of the logistic regression model for police notification. The following victim variables all have significant main effects on police notification: Female, age, White, middle income, high income, and married. The odds of police notification increase by a factor of 1.33 for females compared to males (p < .01). To facilitate better understanding of these significant variables, we calculated discrete change values.³ For an average serious assault victim, the probability of police notification increased by 6.75% for females compared to male, with the female probability being 64.32% and the male probability being 57.57%. A 1-year increase in age increased the odds of police notification by a factor of 1.02 (p < .01). For an average serious assault victim, the probability of police notification increased by 9.24% when age increased from 20 to 40, with the 20-year-old probability being 56.23% and the 40-year-old probability being 65.47%. The odds of police notification decreased by a factor of 0.71 for a White compared to non-White victim (p < .01). For an average serious assault victim, the probability of police notification decreased by 7.94% for Whites compared to non-Whites, with the White probability being 58.20% and the non-White probability being 66.14%. The odds of police notification decreased by a factor of 0.81 for middle income compared to lowest income (p < .05). For an average serious assault victim, the probability of police notification decreased by 5.23% for middle income compared to lowest income, with the middle income probability being 56.09% and the lowest income probability being 61.32%. The odds of police notification decreased by a factor of 0.74 for high income compared to lowest income (p < .01). For an average serious assault victim, the probability of police notification decreased by 7.36% for high compared to lowest income, with high-income probability being 54.13% and lowest income probability being 61.48%. Finally, the odds of police notification increased by a factor of 1.28 for married compared to nonmarried victims (p < .01). For an average serious assault victim, the probability of police notification increased by 5.70% for married compared to nonmarried victims, with a married probability of 65.20% and the nonmarried probability being 59.50%.

Table 2.

Police Notification Logistic Regression Model.

Independent Variables	b	SE	p Value	(e ^b )
Step 1: main effects
Victim variables
Female (male reference)	0.28	0.06	<.01	1.33
Age	0.02	0.00	<.01	1.02
White non-Hispanic (non-White reference)	−0.34	0.07	<.01	0.71
Low income (lowest income reference)	−0.07	0.08	.42	0.93
Middle income (lowest income reference)	−0.22	0.09	.02	0.81
High income (lowest income reference)	−0.30	0.10	<.01	0.74
Income unknown (lowest income reference)	0.10	0.06	.33	1.10
Married (not married reference)	0.24	0.09	<.01	1.28
Incident variables
Series incident (not series reference)	−0.49	0.13	<.01	0.61
Weapon (no weapon reference)	0.11	0.06	.05	1.11
Weapon unknown (no weapon reference)	0.05	0.15	.73	1.05
Private location (public location reference)	0.60	0.07	<.01	1.83
Third party (no third party reference)	0.31	0.06	<.01	1.36
Unknown offender (known offender reference)	0.36	0.08	<.01	1.44
Offender status unknown (known offender reference)	0.39	0.20	.05	1.48
Robbery (not robbery reference)	0.53	0.08	<.01	1.70
Suburban (rural reference)	0.05	0.09	.56	1.05
Urban (rural reference)	−0.10	0.09	.27	0.90
Step 2: interaction effects^a
Interactions
Suburban × Female	−0.18	0.17	.47	0.94
Urban × Female	0.09	0.16	.35	1.18
Suburban × Age	−0.01	0.01	.74	1.00
Urban ×Age	−0.01	0.01	.15	0.99
Suburban × White Non-Hispanic	0.01	0.26	.87	0.96
Urban × White Non-Hispanic	−0.13	0.26	.59	0.88
Suburban × Low Income	0.44	0.26	.14	1.49
Urban × Low Income	0.36	0.25	.19	1.41
Suburban × Middle Income	0.04	0.28	.98	0.99
Urban × Middle Income	0.04	0.30	.97	1.01
Suburban × High Income	0.54	0.40	.29	1.54
Urban × High Income	0.35	0.42	.46	1.37
Suburban × Income Unknown	0.04	0.28	1.00	1.00
Urban × Income Unknown	0.20	0. 29	.52	1.21
Suburban × Married	−0.15	0.30	.46	0.78
Urban × Married	0.02	0.30	.82	1.08
Suburban × Series Incident	0.83	0.43	.04	2.46
Urban × Series Incident	0.51	0.46	.32	1.59
Suburban × Weapon	0.13	0.17	.60	1.10
Urban × Weapon	0.21	0.17	.14	1.30
Suburban × Weapon Unknown	−0.12	0.50	.75	0.83
Urban × Weapon Unknown	−0.25	0.51	.84	0.89
Suburban × Private Location	−0.38	0.18	.17	0.75
Urban × Private Location	−0.22	0. 18	.23	0.77
Suburban × Third Party	0.17	0.17	.77	1.05
Urban × Third Party	0.06	0.18	.72	0.94
Suburban × Unknown Offender	0.31	0.17	.35	1.24
Urban × Unknown Offender	0.03	0.19	.92	0.98
Suburban × Offender Status Unknown	0.41	0.72	.67	1.40
Urban × Offender Status Unknown	−0.74	0.71	.31	0.45
Suburban × Robbery	−0.38	0.39	.29	0.64
Urban × Robbery	−0.45	0.39	.35	0.67

^aInteractions added to main effects one group at a time (e.g., Suburban × Female added with Urban × Female).

The following incident variables had significant main effects on police notification: weapon, private location, third party, stranger, offender status unknown, and robbery. The odds of police notification increased by a factor of 1.11 when the incident involved a weapon compared to no weapon (p < .05). For an average incident, the probability of police notification increased by 2.62% for a weapon compared to no weapon, with the weapon probability being 62.22% and the no weapon probability being 59.60%. The odds of police notification increased by a factor of 1.83 for incidents occurring in a private location compared to nonprivate location (p < .01). For an average incident, the probability of police notification increased by 14.20% for a private compared to nonprivate location, with the private location probability being 68.36% and the nonprivate location probability being 54.17%. The odds of police notification increased by a factor of 1.36 when a third party was present at the incident (p < .01). For an average incident, the probability of police notification increased by 7.45% when a third party was present, with the third-party probability being 63.18% and the non-third-party probability being 55.73%. The odds of police notification increased by a factor of 1.44 when the offender was a stranger (p < .01). For an average incident, the probability of police notification increased by 8.53% when the offender was a stranger, with the stranger probability being 66.66% and the nonstranger probability being 58.13%. The odds of police notification increase by a factor of 1.48 when the victim–offender relationship is unknown (p < .05). The probability of police notification increased by 8.85% when the victim–offender relationship was unknown, with the unknown relationship probability being 69.28% and the known relationship probability being 60.43%. The odds of police notification increased by a factor of 1.70 for robberies compared to assaults (p < .01). The probability of police notification increased by 11.98% for robberies, with the robbery probability being 71.13% and the nonrobbery probability being 59.15%.

Finally, suburban location moderates the series incident effect. For suburban compared to rural areas, being a series incident compared to a nonseries incident increased the odds of police notification by a factor of 2.46 (p < .05). For an average serious assault victim in a suburban area, the probability of police notification increased by 17.42%, with nonseries incident probability being 60.30% and series incident probability being 77.72%.

ER Medical Treatment Model

Figure 2 displays the predicted probabilities of ER treatment based on the logistic regression model. The observed average probability of ER treatment was 0.14. As seen in the figure, the probability of ER treatment was below 0.50 for all incidents, and most were below 0.25. This figure reflects the low rural, suburban, and urban ER treatment percentages described above and presented in Table 1.

Figure 2.

Predicted probabilities of emergency room treatment from the logit model.

Table 3 presents the results of the logistic regression model for ER treatment. Female, age, White, high income, stranger, and robbery all have significant main effects on ER treatment. The odds of ER treatment decreased by a factor of 0.89 for females compared to males (p < .01). For an average serious assault victim, the probability of ER treatment decreased by 2.73% for females, with a female probability of 12.30% and a male probability of 15.04%. A 1-year increase in age increased the odds of ER treatment by a factor of 13.49 (p < .01). For an average serious assault victim, the probability of ER treatment increased by 3.71% when age increased from 20 to 40, with a 20-year-old probability of 12.08% and a 40-year-old probability of 15.79%. Being White non-Hispanic compared to non-White decreased the odds of ER treatment by a factor of 0.47 (p < .01). For an average serious assault victim, the probability of ER treatment decreased by 4.34%, with a White non-Hispanic probability of 12.52% and a non-White probability of 16.85%. The odds of ER treatment decreased by a factor of 0.72 for a victim with high income compared to lowest income (p < .05). For an average serious assault victim, the probability of ER treatment decreased by 3.51% for high compared to lowest income with a high-income probability of 10.64% and a lowest income probability of 14.16%. The odds of ER treatment increased by a factor of 1.20 when the offender was a stranger (p < .05). For an average incident, the probability of ER treatment increased by 2.26% when the offender was a stranger, with a stranger probability of 15.37% and a nonstranger probability of 13.10%. A robbery compared to a nonrobbery increased the odds of ER treatment by a factor of 1.61 (p < .01). For an average serious assault victim, the probability of receiving medical treatment in the ER increased by 6.44% for robberies compared to nonrobberies with the robbery probability being 19.50% and the nonrobbery probability being 13.06%.

Table 3.

Emergency Room Treatment Logistic Regression Model.

Independent Variables	b	SE	p Value	(e ^b )
Step 1: main effects
Victim variables
Female (male reference)	−0.23	0.08	<.01	0.89
Age	0.02	0.01	<.01	13.49
White non-Hispanic (non-White reference)	−0.35	0.07	<.01	0.47
Low income (lowest income reference)	−0.09	0.10	.38	0.92
Middle income (lowest income reference)	−0.24	0.15	.12	0.79
High income (lowest income reference)	−0.33	0.15	.03	0.72
Income unknown (lowest income reference)	−0.02	0.13	.89	0.92
Married (not married reference)	0.00	0.12	.99	1.00
Incident variables
Series incident (not series reference)	−0.40	0.23	.08	0.67
Weapon (no weapon reference)	−0.02	0.08	.85	0.98
Weapon unknown (no weapon reference)	−0.08	0.19	.67	0.92
Private location (public location reference)	0.07	0.08	.34	1.08
Third party (no third party reference)	0.08	0.09	.34	1.09
Unknown offender (known offender reference)	0.19	0.09	.04	1.20
Offender status unknown (known offender reference)	0.16	0.23	.48	1.18
Robbery (not robbery reference)	0.48	0.10	<.01	1.61
Suburban (rural reference)	−0.05	0.13	.72	0.95
Urban (rural reference)	−0.01	0.13	.97	0.99
Step 2: interaction effects^a
Suburban × Female	0.15	0.20	.47	1.16
Urban × Female	−0.09	0.21	.67	0.91
Suburban × Age	0.00	0.01	.81	1.00
Urban × Age	0.00	0.01	.54	1.00
Suburban × White Non-Hispanic	0.32	0.24	.17	1.38
Urban × White Non-Hispanic	−0.04	0.23	.85	0.96
Suburban × Low Income	−0.01	0.26	.97	0.99
Urban × Low Income	−0.58	0.26	.03	0.56
Suburban × Middle Income	0.84	0.74	.26	2.30
Urban × Middle Income	0.65	0.73	.38	1.91
Suburban × High Income	−0.18	0.59	.75	0.83
Urban × High Income	−0.52	0.57	.38	0.60
Suburban × Income Unknown	−0.27	0.55	.54	0.76
Urban × Income Unknown	−0.10	0.41	.82	0.91
Suburban × Married	−0.27	0.30	.36	0.76
Urban × Married	−0.39	0.31	.20	0.68
Suburban × Series Incident	0.03	0.53	.95	1.03
Urban × Series Incident	0.30	0.52	.57	1.34
Suburban × Weapon	0.05	0.22	.81	1.06
Urban × Weapon	0.20	0.21	.35	1.22
Suburban × Weapon Unknown	−0.07	0.59	.91	0.93
Urban × Weapon Unknown	−0.02	0.56	.97	0.98
Suburban × Private Location	−0.01	0.25	.97	0.99
Urban × Private Location	−0.02	0.24	.92	0.98
Suburban × Third Party	−0.22	0.33	.51	0.80
Urban × Third Party	−0.18	0.33	.59	0.84
Suburban × Unknown Offender	0.12	0.25	.62	1.13
Urban × Unknown Offender	0.00	0.22	.99	1.00
Suburban × Offender Status Unknown	2.31	1.05	.03	10.04
Urban × Offender Status Unknown	2.19	1.03	.04	8.94
Suburban × Robbery	0.06	0.42	.88	1.07
Urban × Robbery	−0.29	0.42	.49	0.75

^aInteractions added to main effects one group at a time (e.g., Suburban × Female added with Urban × Female).

Two significant moderating effects were found in this model. First, urban location moderated the relationship between low income (compared to lowest income) and ER treatment. For urban areas, the odds of the victim being treated in the ER decreased by a factor of 0.56 for those with low incomes compared to those with the lowest incomes (p < .05). For an average serious assault victim in an urban area, the predicted probability of ER treatment decreased by 5.72% for low income compared to lowest income, going from 8.57% for low-income urban victims to 14.29% for lowest income urban victims. To further illustrate the moderating effect of urban location, one can look at this relationship across all locations. For an average serious assault victim, the predicted probability of ER treatment increased by only 1.70% going from low income to lowest income. Thus, urban location not only increased the magnitude of the low-income effect, but it also changed the direction. Second, suburban and urban locations moderated the effect of unknown victim–offender relationship on ER treatment. For suburban areas, the odds of the victim being treated in the ER increased by a factor of 10.04 for unknown victim–offender relationships compared to known relationships (p < .05). For an average serious assault victim in a suburban area, the predicted probability of ER treatment increased by 13.38% for unknown relationships compared to known relationships, going from 60.80% for unknown relationships to 47.72% for known relationships. For urban areas, the odds of ER treatment increased by a factor of 8.94 for unknown victim–offender relationships compared to known relationships (p < .05). For an average victim in an urban area, the predicted probability of ER treatment increased by 44.54% for unknown relationships compared to known relationships, going from 57.84% for known relationships to 13.30% for unknown relationships. By comparison, across all locations, the predicted probability of an average victim receiving treatment in the ER decreased by 11.99% for unknown relationships compared to known relationships. Once again, the location moderating effect not only changed the magnitude of the relationship between victim–offender relationship and ER treatment but it also changed the direction of the relationship.

Discussion

The goal of this study was to determine, in cases of serious assault, which victim and incident characteristics are associated with police notification and medical treatment in the ER and whether these characteristics are the same for rural, suburban, and urban areas. In short, differences in factors associated with police notification and ER treatment exist, but very few factors are moderated by community type (i.e., rural, suburban, and urban). Incident characteristics tend to be significantly associated with police notification more often than ER treatment, and victim characteristics tend to be significantly associated fairly equally with ER treatment and police notification. Consistent factors in both models include female, age, White, high income, stranger, and robbery.

Relative Frequency of Police Notification and ER Treatment

Incidents of serious assaults are much more likely to result in police notification than ER treatment. Kaylen and Pridemore (2014) reported similar results in their study of police notification and ER treatment among rural, suburban, and urban victims of violence, though their sample selection criteria were different than that used in this study. A number of substantive and methodological reasons explain why ER treatment appears less often than police notification in the current study.

First, not all assault victims require medical treatment, while all assault incidents could in theory be reported to the police. In this sample specifically, some incidents did not involve injuries, so medical treatment would not have been sought. For instance, attempted assaults would not have injuries but would still be included in the sample because these assaults could still be reported to the police as (attempted) aggravated assaults.

Second, some incidents in the sample did not involve reported injuries, so the respondent would not have been asked about medical treatment. The skip pattern in the crime incident report is as follows. First, the respondent is asked if she or he was injured in the incident. If the respondent says she or he was not injured, questions about medical treatment are skipped. If the respondent says she or he was injured, questions about medical treatment are asked. To capture the full spectrum of incidents that could be included in police aggravated assault data and/or ER assault data, we decided to include assault incidents in which injuries did not occur.

Third, assault victims requiring medical treatment have a number of alternatives to the ER, while limited viable alternatives to police notification exist. The NCVS lists the following categories for locations at which the respondent could report receiving medical treatment as the result of an injury sustained during an incident: at the scene; home, neighbor’s, or friend’s; health unit at work/school, first aid station, at a stadium/park; doctor’s office/health clinic; ER at the hospital/emergency clinic; hospital (other than ER); and other location. The severity of the injury, presence of someone to administer medical treatment at the scene, distance to an ER, and financial considerations (e.g., health insurance and deductible costs) all can play a role in whether an assault victim receives medical treatment at the ER.

Further Examination of Police Notification Results

Results of the police notification model are somewhat consistent with previous research on reporting to the police. Our study found that for all areas (rural, suburban, and urban), sex, age, White, middle income, high income, marital status, weapon, private location, third-party presence, stranger, unknown victim–offender relationship, and robbery are associated with likelihood of police notification. Previous research similarly found sex, age, household income, incident location, and third-party presence are associated with increased police reporting (Bosick, Rennison, Gover, & Dodge, 2012; Planty, 2002). Hart and Rennison (2003) and Bosick et al. (2012) found that robberies are reported to the police at a higher percentage than aggravated and simple assaults, consistent with our robbery finding. This robbery finding is also consistent with the use of robbery as a measure of crime seriousness. As Gottfredson and Gottfredson (1988) point out and other research has found, crime seriousness is an important factor in whether or not a crime is reported to the police (Davies et al., 2007; Felson et al., 1999; Gartner & Macmillan, 1995; Hart & Rennison, 2003; Kaukinen, 2002; Xie et al., 2006; Xie & Lauritsen, 2012).

Further Examination of ER Treatment Results

Research comparable to the present ER treatment analysis does not exist, so comparisons with previous research are not possible. Nevertheless, potential explanations for some of the findings can still be considered in the context of other literature.

Sex has a significant effect on ER treatment for serious assault injuries. As Kushel et al. (2002) point out, injuries from violent crimes are suitable for emergency care because they require urgent attention, often occur where other treatment is not available, and sometimes require services that are not available in nonemergency sites. The latter point suggests males have a higher probability of ER treatment than females. Past research has found that men are less likely than women to be injured during an assault, but when they are injured their injuries are more severe (Craven, 1997; Felson & Cares, 2005). On the other hand, Felson and Cares (2005) speculate that at the same level of injury, women are more likely to seek medical treatment than men. These points also suggest ER data may be affected by selection bias because sex serves as a filter for the seriousness of crimes. Females receiving treatment in the ER will have a broader range of injuries than males who will tend to have only the most serious injuries.

We found that the effects of two income categories on ER treatment were moderated by geographic location. First, the effect of low income (compared to lowest income) on ER treatment was moderated by urban location. That is, those in the low-income group are less likely to receive medical treatment in the ER than those in the lowest income group in urban areas compared to rural areas. This finding is consistent with past research that suggests ERs are a major source of medical care for inner-city poor families (Deval, Alpert, & Howard 1996; Orr et al. 1991). Once again, Kushel et al.’s (2002) reasons for ER treatment among violent crime victims may apply to this finding. Specifically, medical treatment at other locations may not be available to or may not be suitable for injuries sustained by the urban poor. Second, the effect of middle income (compared to lowest income) was moderated by suburban location. Those in the middle-income group are more likely to receive medical treatment in the ER than those in the lowest income group in suburban areas compared to rural areas. The income structure of suburban victims is different than that of rural and urban victims, with a higher proportion of suburban victims in the middle- and high-income categories compared to rural and urban victims. This may suggest more variation in victimization severity in suburban areas. If more suburban middle-income victims have severe injuries than suburban lowest income victims, there may be an increase in suburban middle-income victims using the ER.

The Lack of Rural, Suburban, and Urban Moderating Effects

One of the main goals of this study was to determine if factors associated with police notification and ER treatment are moderated by geographic location (rural, suburban, and urban). With a few exceptions, this was not the case.⁴ Substantively, rural, suburban, and urban differences in victim and incident characteristics associated with police notification and ER treatment may not actually exist. This study was exploratory in that this question had not been previously addressed and theory did not suggest what the findings might be. If location differences in factors associated with these two outcomes do exist, those factors may not have been included in this study.

Methodologically, the lack of significant findings may be related to the sample size and the number of terms in the models. In some cases, the raw sample size for a variable attribute is small. When creating an interaction between this variable and one of the location variables, the counts get smaller. For example, the rural offender status unknown sample is 20 and the rural weapon unknown sample is 31 (see Table 1). The sample sizes continue to decrease, as the sample is further divided into the outcome categories (i.e., yes or no to police notification and yes or no to ER treatment). Further, the inclusion of interaction terms in the model increases the degrees of freedom, which in turn increases the size of the test statistic. As the test statistic increases, larger effects are needed in order to be considered significant.

The Role of Sample Selection Bias

Sample selection bias is a potential concern in our analysis. The sample consists of individuals who (a) chose to participate in the NCVS after they were selected, (b) were victims of serious assaults, and (c) chose to report that they were victims of serious assaults. It could be that the same factors that predict or are associated with being a victim are also associated with the outcomes of interest, police notification, and ER treatment. Further, those who choose to participate in a victimization survey and who choose to disclose their victimizations also may be more likely to exhibit help-seeking behavior in the form of police notification and ER medical treatment compared to those who choose to not participate in the survey. As a result, the sample of violent victimization incidents in the present study may not be a representative sample of all serious assault incidents, and thus the results may not generalize to all incidents.

Researchers have suggested a number of methods for addressing sample selection bias in criminological research, but the methods have not been fully established in the field and have important limitations. If experimental design with random group assignment is not possible, Bushway, Johnson, and Slocum (2007) suggest utilizing Heckman’s (1976, 1979) two-step correction by adding an additional exogenous variable to the model that predicts sample selection but is independent of the outcome model. Likewise, Bushway and Apel (2010) recommend instrumental variable techniques to handle exogenous independent variables and measurement error when sample selection bias is a concern and important variables are omitted. This type of variable is difficult to find, so Bushway et al. (2007) implore criminologists to at least critically evaluate the possible effects of sample selection bias on their models. A second proposed way to address selection bias is through propensity score matching (Guo & Fraser, 2009). Propensity score designs are limited in two important ways. First, propensity score matching results in the loss of data due to dropped cases that are not matched between the treatment and control groups (Porter & Vogel, 2014). Second, propensity score methods are limited by the strong assumption that assignment to treatment can be ignored because selection bias is accounted for with the propensity scores (Guo & Fraser, 2009; Porter & Vogel, 2014; Rosenbaum & Rubin, 1983; Rubin, 2007). Methods for addressing sample selection bias continue to evolve and are becoming more common in criminological research.

Implications for Future Research

Our findings have a number of implications for future research. The primary impetus for this study was to guide future researchers on the decision to use police data or hospital data to measure violence in community-level tests of theory. Of particular interest was whether or not these types of data can be used in comparisons across rural, suburban, and urban communities.

The ability to make precise estimates of the absolute level of violence is not the focus of this type of research, rather the ability to estimate relative rates of violence (relative to other areas for the sake of comparison) is of importance. A primary concern in such studies is to ensure the data represent the population. That is, it would be problematic if some parts of the population are systematically underrepresented in the data. For instance, if males are much more likely to be victims of violence than females across rural, suburban, and urban areas, but males are considerably less likely to report crime to the police than females in only rural areas, police violence data in rural areas would show comparatively lower overall violence rates than suburban and urban police violence data. It is important to emphasize that the finding that victims of serious assaults receive medical treatment at a lower rate than police are notified about these incidents does not automatically mean that police data should be used in place of hospital data. In fact, the similar rural, suburban, and urban percentages of incidents resulting in each outcome (police notification and ER treatment) are promising for both police data and hospital data, at least in terms of making the types of comparisons across community type described here. In short, at least for aggravated assault, our results suggest that both police and hospital data can be used to compare units (e.g., counties) of varying population size and metropolitan status.

Different factors tended to be significantly associated with a serious assault incident coming to the attention of the police and the victim receiving treatment in the ER. The factors that were significant in both models were sex, age, White, high income, stranger, and robbery, and all but sex had the same direction of effect on the outcomes. Researchers utilizing hospital assault data should realize that robberies are likely included in their data. If a robbery is different than an assault as it relates to the research question, caution should be taken in utilizing such hospital data. Further, if comparing theoretical models that use hospital assault data and police assault data (i.e., aggravated assaults), the issue of robbery compared to nonrobbery may be relevant. Researchers doing such studies might want to consider combining police data on aggravated assaults and robberies.

A few other noteworthy results might affect data decisions in future research. The proportion of incidents that occurred in private locations and/or had third parties present may be higher in police data than in the true underlying distribution because these factors have higher probabilities of police notification than their counterparts. Further, White non-Hispanics are less likely than non-Whites to receive ER treatment, with approximately 4.5% difference in the predicted probability (for an average victim). This race difference in help-seeking behavior (ER treatment) is consistent with Xie and Lauritsen’s (2012) finding that victims exhibit help-seeking behavior (police reporting) at the highest rate when both the victim and offender are Black followed by Black offender–White victim, White offender–White victim, and White offender–Black victim. Future research on the relationship between race and victim medical treatment should examine the races of both victims and offenders to see if this aspect of race affects inclusion in hospital data. Researchers utilizing hospital assault data should be aware of the race differences in ER treatment presented in our study. If incident location, third-party presence, and victim race are not of theoretical importance to the research, this disproportionate representation in the data may not be a problem. Regardless of theoretical importance, though, researchers should consider the implications of these findings in their discussion of potential sample selection bias.

Of particular interest in our study was whether or not factors that affect the outcomes differ by rural, suburban, and urban community. Few differences were found. In urban areas, ER treatment is more likely among those in the lowest income group compared to the low-income group. In both urban and suburban areas, ER treatment is more likely when the victim–offender relationship is unknown compared to known. These findings should be further explored to see why differences in ER treatment exist.

Conclusion

The goal of this study was to examine victim and incident characteristics associated with police notification and ER treatment and to see if these are the same across rural, suburban, and urban areas. This research was motivated by the need to better understand sources of violence data, particularly police assault records and hospital data on assault victims, and how any differences may influence tests of theory. Analysis revealed that police notification is much more likely to occur than ER treatment. Few victim and incident characteristics are significantly associated with ER treatment, a majority of victim and incident characteristics are significantly associated with police notification, and very few characteristics are moderated by location. These results suggest that either police or hospital data on serious assault incidents can be used to make comparisons across rural, suburban, and urban communities without being limited by disproportionate inclusion or exclusion of incidents associated with victim and incident characteristics. Comparisons between the two data sources—police and hospital data—should be made with caution due to the different victim and incident characteristics associated with police notification and ER treatment.

Footnotes

Acknowledgment

We thank Lynn Addington and Chunfeng Huang for their comments on and critiques of earlier drafts of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Notes

References

Avakame

E. F.

Fyfe

J. J.

McCoy

(1999). ‘Did you call the police? What did they do?’ An empirical assessment of Black’s theory of mobilization. Justice Quarterly, 16, 765–792.

Baumer

E. P.

Lauritsen

J. L.

(2010). Reporting crime to the police, 1973–2005: A multivariate analysis of long-term trends in the National Crime Survey (NCS) and National Crime Victimization Survey (NCVS). Criminology, 48, 131–185.

Bosick

S. J.

Rennison

C. M.

Gover

A. R.

Dodge

(2012). Reporting violence to the police: Predictors through the life course. Journal of Criminal Justice, 40, 441–451.

Bushway

Apel

R. J.

(2010). Instrumental variables in criminology and criminal justice. In Piquero

A. R.

Weisburd

(Eds.), Handbook of quantitative criminology (pp. 595–612). New York, NY: Springer.

Bushway

Johnson

B. D.

Slocum

L. A.

(2007). Is the magic still there? The use of the Heckman two-step correction for selection bias in criminology. Journal of Quantitative Criminology, 23, 151–178.

Catalano

S. M.

(2006). Criminal victimization, 2005 (Bureau of Justice Statistics Bulletin, NCJ 214644) . Washington, DC: U.S. Department of Justice, Office of Justice Programs.

Chu

Rivera

Loftin

(2000). Herding and homicide: An examination of the Nisbett–Reaves hypothesis. Social Forces, 78, 971–987.

Craven

(1997). Sex differences in violent victimization, 1994 (Bureau of Justice Statistics Special Report, NCJ 164508) . Washington, DC: Department of Justice, Office of Justice Programs.

Davies

Block

C. R.

Campbell

(2007). Seeking help from the police: Battered women’s decisions and experiences. Criminal Justice Studies, 20, 15–41.

10.

Deval

S.-C.

Alpert

J. J.

Bauchner

(1996). Care-seeking patterns of inner-city families using an emergency room: A three-decade comparison. Medical Care, 34, 1171–1179.

11.

Fabio

Strotmeyer

Branas

C. C.

(2004). Racial segregation and county level intentional injury in Pennsylvania: Analysis of hospital discharge data for 1997–1999. Journal of Epidemiology and Community Health, 58, 346–351.

12.

Fabio

Sauber-Schatz

E. K.

Barbour

K. E.

(2009). The association between county-level injury rates and racial segregation revisited: A multilevel analysis. American Journal of Public Health, 99, 748–753.

13.

Felson

R. B.

Cares

A. C.

(2005). Gender and the seriousness of assaults on intimate partners and other victims. Journal of Marriage and Family, 67, 1182–1195.

14.

Felson

R. B.

Messner

S. F.

Hoskin

(1999). The victim-offender relationship and calling the police in assaults. Criminology, 37, 931–948.

15.

Freisthler

Gruenewald

P. J.

Ring

LaScala

E. A.

(2008). An ecological assessment of the population and environmental correlates of childhood accident, assault, and child abuse injuries. Alcoholism: Clinical and Experimental Research, 32, 1969–1975.

16.

Gartner

Macmillan

(1995). The effect of victim-offender relationship on reporting crimes of violence against women. Canadian Journal of Criminology, 37, 393–429.

17.

Gottfredson

M. R.

Gottfredson

D. M.

(1988). Decision making in criminal justice (2nd ed.). New York, NY: Plenum Press.

18.

Gruenewald

P. J.

Freisthler

Remer

LaScala

E. A.

Treno

(2006). Ecological models of alcohol outlets and violent assaults: Crime potentials and geospatial analysis. Addiction, 101, 666–677.

19.

Gruenewald

P. J.

Freisthler

Remer

LaScala

E. A.

Treno

Ponicki

W. R.

(2010). Ecological associations of alcohol outlets with underage and young adult injuries. Alcoholism: Clinical and Experimental Research, 34, 519–527.

20.

Guo

Fraser

M. W.

(2009). Propensity score analysis: Statistical methods and applications. Thousand Oaks, CA: Sage.

21.

Hart

T. C.

Rennison

(2003). Reporting crime to the police, 1992–2000 (Bureau of Justice Statistics Special Report [NCJ 195710]) . Washington, DC: U.S. Department of Justice, Office of Justice Programs.

22.

Heckman

J. J.

(1976). The common structure of statistical models of truncation, sample selection and limited dependent variables. Annals of Economic and Social Measurement, 5, 475–492.

23.

Heckman

J. J.

(1979). Sample selection bias as a specification error. Econometrica, 47, 153–162.

24.

Kaukinen

(2002). The help-seeking decisions of violent crime victims: An examination of the direct and conditional effects of gender and the victim-offender relationship. Journal of Interpersonal Violence, 17, 432–456.

25.

Kaylen

M. T.

Pridemore

W. A.

(2011). A reassessment of the association between social disorganization and youth violence in rural areas. Social Science Quarterly, 92, 978–1001.

26.

Kaylen

M. T.

Pridemore

W. A.

(2013). The association between social disorganization and rural violence is sensitive to the measurement of the dependent variable. Criminal Justice Review, 38, 169–189.

27.

Kaylen

M. T.

Pridemore

W. A.

(2015). Measuring violent victimization: Reporting to police and presenting to emergency rooms by rural, suburban, and urban victims, 39, 239–246.

28.

Kruttschnitt

Carbone-Lopez

(2009). Customer satisfaction: Crime victims’ willingness to call the police. Ideas in American Policing, 12, 1–15.

29.

Kushel

M. B.

Perry

Bangsberg

Clark

Moss

A. R.

(2002). Emergency department use among the homeless and marginally housed: Results from a community-based study. American Journal of Public Health, 92, 778–784.

30.

Laub

J. H.

(1981). Ecological considerations in victim reporting to the police. Journal of Criminal Justice, 9, 419–430.

31.

LeMier

Cummings

West

T. A.

(2001). Accuracy of external cause of injury codes reported in Washington state hospital discharge records. Injury Prevention, 7, 334–338.

32.

Loftin

McDowall

(2003). Regional culture and patterns of homicide. Homicide Studies, 7, 353–367.

33.

Lott

J. R.

Whitley

(2003). Measurement error in county-level UCR data. Journal of Quantitative Criminology, 19, 185–198.

34.

Maltz

M. D.

Targonski

(2002). A note on the use of county-level UCR data. Journal of Quantitative Criminology, 18, 297–318.

35.

Maltz

M. D.

Targonski

(2003). Measurement and other errors in county-level UCR data: A reply to Lott and Whitley. Journal of Quantitative Criminology, 19, 199–206.

36.

May

J. P.

Hemenway

Hall

(2002). Do criminals go to the hospital when they are shot? Injury Prevention, 8, 236–238.

37.

Meux

E. F.

Stith

S. A.

Andra

(1990). Report of results from the OSHPD reabstracting project: An evaluation of the reliability of selected patient discharge data July through December 1998. Sacramento, CA: Patient Discharge Data Section Office of Statewide Health Planning and Development.

38.

Orr

S. T.

Evan

John

Barbara

(1991). Emergency Room Use by Low Income Children with a Regular Source of Health Care. Medical Care, 29, 283–286.

39.

Osgood

D. W.

Chambers

J. M.

(2000). Social disorganization outside the metropolis: An analysis of rural youth violence. Criminology, 38, 81–116.

40.

Planty

(2002). Third-party involvement in violent crime, 1993–1999 (Bureau of Justice Statistics Special Report [NCJ 189100]) . Washington, DC: U.S. Department of Justice, Office of Justice Programs.

41.

Porter

Vogel

(2014). Residential mobility and delinquency revisited: Causation or selection? Journal of Quantitative Criminology, 30, 187–214.

42.

Pregibon

(1979). Data analytic methods for generalized linear models (Unpublished PhD Dissertation). University of Toronto, Canada.

43.

Rennison

C. M.

(2007). Reporting to the police by Hispanic victims of violence. Violence and Victims, 22, 754–772.

44.

Rennison

C. M.

Dragiewicz

DeKeseredy

W. S.

(2013). Context matters: Violence against women and reporting to police in rural, suburban and urban areas. American Journal of Criminal Justice, 38, 141–159.

45.

Rennison

C. M.

Rand

(2007). Introduction to the National Crime Victimization Survey. In Lynch

J. P.

Addington

L. A.

(Eds.), Understanding crime incidence statistics: Revisiting the divergence of the NCVS and the UCR. New York: Cambridge University Press.

46.

Rivera

Chu

Loftin

(2002). Do strong communities increase homicide? An evaluation of the Nisbett–Cohen thesis. International Journal of Sociology and Social Policy, 22, 109–121.

47.

Rosenbaum

P. R.

Rubin

D. B.

(1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70, 41–55.

48.

Rubin

D. B.

(2007). The design versus the analysis of observational studies for causal effects: Parallels with the design of randomized trials. Statistics in Medicine, 26, 20–36.

49.

Skogan

W. G.

(1984). Reporting crime to the police: The status of the world research. Journal of Research in Crime and Delinquency, 6, 113–137.

50.

StataCorp. (2013). Stata 13 base reference manual. College Station, TX: Stata Press.

51.

Tukey

J. W.

(1949). One degree of freedom for non-additivity. Biometrics, 5, 232–242.

52.

Wood

D. S.

(2010). The validity of injury surveillance system measures of assault: A lesson from the study of violence in Alaska. Journal of Interpersonal Violence, 25, 219–241.

53.

Xie

Lauritsen

J. L.

(2012). Racial context and crime reporting: A test of Black’s stratification hypothesis. Journal of Quantitative Criminology, 28, 265–293.

54.

Xie

Pogarsky

Lynch

J. P.

McDowall

(2006). Prior police contact and subsequent victim reporting: Results from the NCVS. Justice Quarterly, 23, 481–501.