Gender and Injury Risk in Incidents of Assaultive Violence

Abstract

This study investigates the situational characteristics that determine the presence and severity of injury in incidents of assaultive violence. The analysis uses merged data from the National Crime Victimization Survey and the Supplementary Homicide Reports for the years 1992-2008, in order to model the determinants of victim injury. The analysis includes all incidents of attempted or completed, non-sexual assault against victims 12 years of age or older. Injury severity is classified into one of four possible levels: no injury, minor injury, serious injury (requiring doctor, hospital, or emergency room care), and lethal injury. Special attention is given to the way in which gender modifies the influence of situational elements on the presence and degree of victim injury. While the results suggest that the situational determinants of injury are by and large uniform for male and female victims, important gender differences are observed in the salience of relational distance.

Keywords:

• injury
• NCVS
• SHR
• gender
• relational distance
• heteroscedastic logit model

Notes

1. Relational distance also tends to covary with other structural features of interpersonal disputes, namely status equality, functional interdependence, and immobility (see Black, 1990), which can independently exacerbate the use of injurious violence against intimate partners. Status equality refers to the relatively equal social standing of disputants. Functional interdependence refers to the presence of long-standing emotional or familial ties that bind disputants together. Immobility refers to the fact that disputants share social and physical space and are therefore unable to completely avoid one another.

2. This assertion may be true only for assaultive violence, however, as indicated by studies which include broader offense types than assault. Offender gun use in robbery appears to lower injury risk, presumably because of the overwhelming coercive power that such possession entails (Tark & Kleck, 2004). Similarly, the risk of injury tends to be inversely related to the lethality of the weapon used in incidents of sexual assault (Skogan & Block, 1983).

3. Other hypotheses concerning gender symmetry/asymmetry in the situational determinants of injury are possible to make, but would be less strongly grounded in existing theories. We thus focus our attention here on relational distance.

4. NCVS staff treat the first interview of each newly sampled household as a “bounding interview” and exclude it from the data. Thus for all intents and purposes each household is interviewed six times in three years.

5. “Series” crimes are repeat victimizations that are of the same type and that occur six or more times during the six-month reference period. For series crimes, only the details of the most recent incident are reported by the victim, as victims often have difficulty distinguishing the characteristics of discrete incidents. Series crimes constitute only 5% of the NCVS incidents chosen for inclusion in the analysis.

6. Note that our non-lethal injury coding scheme departs from that used by NCVS staff. The NCVS classifies as serious injuries: (1) all completed rapes; (2) all incidents that result in gunshot or knife wounds, broken bones, loss of teeth, internal injuries, or loss of consciousness; and (3) all injuries that require two or more days of hospitalization.

7. Beginning with the 2003 survey, the NCVS adopted an enhanced coding scheme that allowed victims to self-identify multiple racial/ethnic categories. For these years, we coded race in a non-mutually exclusive manner, allowing respondents to claim affiliation with multiple racial groups. We included a separate dummy variable flagging these respondents in the empirical models. Only a small proportion of our NCVS victims (3.2%) identify themselves as multiracial.

8. The heteroscedastic logit model is in the larger class of heterogeneous choice models (see Alvarez & Brehm, 1995). To estimate these models, we rely on the Stata protocol “oglm” (Williams, 2006). Interested readers may consult Williams (2009) for a recent application.

9. We also estimate a variety of other statistical models for sensitivity purposes. First, we estimate a model designed to take into account the fact that our sequential logit models are censored at each subsequent step, where the censoring is determined by the level of injury incurred by the victim. These are sample selection models in which a function of the fitted probabilities from the first step (the inverse Mills ratio) is included as a covariate in subsequent steps. While the model predicting at least serious injury (Model B) qualitatively differs (e.g. the only significant correlates of at least serious injury are African-American offenders and adolescent victims), the model predicting lethal injury (Model C) yields findings that are very similar to those reported in Table 2. In our case, however, this approach is severely limited by the absence of exclusion restrictions, which can yield biases that are worse than models without the selection correction. Second, we estimate a multivariate probit model, which models the sequential injury outcomes simultaneously. The results were substantially similar to the sequential logit models. Third and finally, we estimate multinomial logistic regression, which models the determinants of injury across all levels and also provides contrasts of predictors by injury level. We provide results from the multinomial logit model for all respondents in Appendix B. The multinomial logit models produce qualitatively similar results to the sequential logit models.

10. Subtracting one from an odds ratio and multiplying by 100 gives the percentage increase (decrease) in the odds of injury given a unit increase in the regressor. So for example, an odds ratio of 4.3 implies that the odds of injury at the hands of a spouse or ex-spouse are 330% higher than the odds of injury at the hands of a stranger.

11. Interaction terms in logit models are less straightforward to interpret than in linear models. They are complicated by the fact that the sign, magnitude, and significance of the product term do not necessarily correspond directly to the sign, magnitude, and significance of the marginal effect of the interaction. In fact, the product term and marginal effect can actually be of opposite sign. Ai and Norton (2003) present a method of estimating and interpreting the nature of interaction effects in non-linear models (see also Norton, Wang, & Ai, 2004). We apply this method to our data, in order to ensure that we have correctly interpreted the victim-assailant gender interaction term.

12. Missing cases for relational distance, victim demographics, multiple victims, and weapon use were distributed fairly evenly across the NCVS and SHR data sets (e.g. SHR 18.8% vs. NCVS 16.1% for relational distance). On the other hand, missing data were much more prevalent for multiple offenders and offender demographics in the SHR compared to the NCVS (e.g. SHR 14.3% vs. NCVS 1.9% for multiple offenders). An anonymous reviewer was concerned that the prevalence of missing data for relational distance in our SHR data set appears to be much lower than in other studies using the SHR. We would note that our attention is limited exclusively to homicides with non-sexual, assaultive circumstances, which have more complete data. For example, when we examine the remaining incidents in the SHR, 38.5% of cases have missing data on relational distance.

13. The user-written Stata program “ice” was employed to impute missing cases (Royston, 2005a, 2005b).