[Skip to Content]
Access to paid content on this site is currently suspended due to excessive activity being detected from your IP address Please contact the publisher to request reinstatement.
Sign In
Individual Sign In
Create an Account
Institutional Sign In
OpenAthens Shibboleth
[Skip to Content Landing]
August 2005

Violent Behavior by Girls Reporting Violent VictimizationA Prospective Study

Author Affiliations

Author Affiliations: From the Departments of Society, Human Development, and Health (Drs Molnar and Buka and Ms Cerda) and Health Policy and Management (Dr Browne), Harvard Youth Violence Prevention Center, Harvard School of Public Health, Boston, Mass.

Arch Pediatr Adolesc Med. 2005;159(8):731-739. doi:10.1001/archpedi.159.8.731

Objective  To assess the relationships between individual victimization and neighborhood-level violence on subsequent violent perpetration by adolescent girls in a community-based sample.

Design  Longitudinal, multilevel analysis of data collected by the Project on Human Development in Chicago Neighborhoods. Three in-home interviews were conducted approximately 24 months apart between November 1995 and January 2002 with youth and their caregivers. Community-level data also were collected in 1995 from a random sample of Chicago residents. Hierarchical regression models and propensity scores were used.

Setting  Families and neighborhoods in Chicago.

Participants  Population-based sample of 637 girls, ages 9 to 15 years at baseline, and the neighborhoods in which they resided. This sample is diverse in race/ethnicity, socioeconomic status, family structure, and neighborhood characteristics.

Main Outcome Measure  Self-reports of violent behavior in the 12 months before the third interview.

Results  At baseline, 38% of the girls reported perpetrating at least 1 violent behavior in the prior 12 months, 28% reported past year violent behavior at the first follow-up interview, and 14% reported past year violent behavior at the third interview. The odds of violent behavior were 2.2 times higher among girls who reported prior violent victimization, after prior confounding factors and baseline violent behavior were controlled (95% confidence interval, 1.3- 4.4). Homicides and concentrated poverty in girls’ neighborhoods also were associated with aggression by girls.

Conclusions  Improving safety in communities and homes may reduce rates of violent perpetration by adolescent girls. Study results suggest that, to facilitate identification of and healing among adolescent survivors of violence, practitioners should recognize perpetration of violence as potential sequelae of prior violent victimization.

Violence experienced and perpetrated by adolescents is a serious public health problem, affecting individuals, families, and communities. Adolescents are exposed to a great deal of violence in their homes and communities; this exposure has been associated with their own violent perpetration as reported across numerous studies.1,2 Rates of sexual assault3,4 and victimization occurring at home5,6 are typically higher for girls than for boys, whereas rates of nonsexual violent victimization outside the home are lower for girls.3,7

Two types of data provide prevalence estimates of youth violence: arrest data and self-report data. Although both sources indicate a decrease in rates of violence among youth since the mid 1990s, levels of involvement remain high. National arrest data showed a peak in arrests of teenagers for violent crimes in 1994 of just over 500 arrests for every 100 000 youth aged from 10 to 17 years, and a subsequent decline of 47% by 2002 to 276 arrests for every 100 000.8 However, in contrast to declining rates for boys, rates of arrest for violent crimes by adolescent girls either increased, or decreased less, across categories of violent offenses.8 This trend gave rise to concern that girls were becoming more violent.912

Arrest rates can be difficult to interpret. Large increases in arrest rates for girls are, in part, a result of low base rates.10 Even with increases in arrest rates, girls accounted for just 24% of juvenile arrests for aggravated assaults in 2002.8 Arrest rates also measure societal reactions to behavior, not necessarily behavior itself; growth in arrests may represent changes in the application of laws by gender or in the willingness of the public to report crimes.8 Thus, while there is a popular belief that girls’ violence is growing, the true extent of violence by girls is unknown.

Self-report data by teens show a decrease in violent behavior over the past decade for both girls and boys.13 The percentage of students reporting “being in a physical fight in the past 12 months” in the nationally representative Youth Risk Behavior Surveillance Survey, a school-based survey of 9th through 12th graders, declined from 1993 through 2003. Specifically, the percentage of girls involved in a physical fight dropped from 31.7% in 1993 to 25.1% during 2003 (a decrease was reported by boys from 51.2% to 40.5%).4,14 Still, that one quarter of 9th through 12th grade girls nationwide reported physically fighting with others in a 12-month period underscores the need for further investigation of potential causes.


Exposure to violence in the home includes being the victim of child physical or sexual abuse by intimates, witnessing violence between caregivers, and violent altercations with family members, including siblings. Each type of victimization has been shown to be a risk factor for adolescent aggression as well as for delinquency and involvement with the criminal justice system.1520 Additionally, studies indicate that girls who perpetrate violence are significantly more likely to have experienced sexual violence than nonaggressive girls.1,17,2123

Research consistently identifies exposure to violence in the neighborhood as a precursor to adolescents’ violent behavior, typically using individual reports of neighborhood exposures.1,3,21,24,25 Examining the effects of aggregated neighborhood conditions in addition to individual reports is a burgeoning area of scientific inquiry in public health,26 especially as the statistical methods for analyzing hierarchical data have evolved.27 As an example of the effect of neighborhood conditions on adolescent behavior, higher levels of physical and social disorder and having fewer safe places for children to play both were associated with a higher likelihood of youths carrying concealed firearms, even when individual risk factors for gun carrying were controlled.28

The current inquiry focuses on individual-level victimization and neighborhood violence and their relationships to the subsequent perpetration of violent behaviors by adolescent girls. Although most studies focusing on girls’ violent behavior have used special populations (eg, clinically referred or juvenile justice samples) or cross-sectional surveys, this study examines girls’ violent behavior prospectively among a representative, diverse, community-based sample. Aims of the inquiry were to describe violent behaviors by girls across 3 waves of data and to identify both individual- and neighborhood-level risk factors for violence perpetration, in addition to more traditional factors such as age, family composition, family socioeconomic status, and baseline rates of violent behavior.


Data for this study of girls’ violence came from the Project on Human Development in Chicago Neighborhoods (PHDCN). The PHDCN is a multidisciplinary, longitudinal study of the development of antisocial behavior, prosocial behavior, and mental health among youth and their primary caregivers, as well as a study of their neighborhoods over time. Detailed methods are published elsewhere.29


Informed consent for each interview was obtained from a parent or guardian of the subject in the longitudinal cohort study. Assent was obtained from the youth respondents prior to each assessment. Procedures were approved by 2 institutional review boards at Harvard University, Boston, Mass.

The design of the PHDCN consists of a longitudinal cohort study imbedded within a community study. Neighborhood-level data for PHDCN participants were obtained from (1) a community survey of 8782 randomly selected Chicago residents conducted in 1995, (2) data on violent crime in 1995, obtained from the Chicago Police Department, and (3) US Census data from 1990. Individual- and family-level data were collected by a longitudinal cohort study completed between November 1994 and January 2002, in which overall 6226 children and their caregivers were enrolled.


Chicago neighborhoods were initially grouped into 343 neighborhood clusters, each with approximately 8000 residents. Stratified probability sampling was then used to select a sample of 80 neighborhood clusters diverse in race/ethnic and socioeconomic composition for intensive study.

For the longitudinal cohort study, dwelling units were enumerated and approximately 35 000 randomly sampled households were screened for children and adolescents of eligible ages within the 80 selected neighborhoods. Children who were within 6 months of 7 target cohort ages (0 [in utero through the age of 6 months], 3, 6, 9, 12, 15, and 18 years) were invited to participate. The participation rate was 75%. A primary caregiver was enrolled for all children except the 18-year-old cohort. Three in-home interviews were conducted at approximately 24-month intervals between 1995 and 2002. Individual-level data for the current analyses come from girls enrolled in cohorts 9, 12, and 15 years and their primary caregivers.


Table 1 displays the instruments used at each of the 3 interviews and how each measure was used in the present analysis. As noted in Table 1, the individual-level confounders and neighborhood-level variables described below were measured at wave 1 (the first interview), individual-level victimization exposures were measured at wave 2, and outcomes were measured at wave 3. Additionally, 3 waves of self-report violent behavior data were used to describe prevalence across the age range of study participants.

Table 1. 
Temporal Ordering of Variables Included in the Study Analyses of Data From the PHDCN, Waves 1 Through 3, Ages at Enrollment: 9, 12, and 15 Years
Temporal Ordering of Variables Included in the Study Analyses of Data From the PHDCN, Waves 1 Through 3, Ages at Enrollment: 9, 12, and 15 Years

Questionnaires were administered in respondents’ homes by trained interviewers. Based on a scale adapted from an earlier instrument by Elliott et al,30 perpetration of 1 or more of the following 8 past-year violent behaviors with intent to cause injury were examined at each of the 3 interviews: (1) hitting someone outside of the home; (2) attacking with a weapon; (3) using weapons or force to rob someone; (4) picking pockets or purse snatching; (5) throwing rocks and bottles; (6) setting fires to buildings, cars, or property; (7) carrying a hidden weapon; and (8) participating in a gang fight. The outcome measure was developed based on third-interview responses by girls on this scale. Items were selected because they were aggressive acts against people and because item-response analysis31,32 of their psychometric properties showed that, as a scale, they had good reliability and validity. An item measuring hitting people within the home did not fit well and was dropped.33

At the third interview, girls reported 6 of 8 original violent act items (at baseline they had reported all 8). This led us to base the outcome measures for the statistical model solely on the 6 outcomes that were reported by at least 1 of the respondents—hitting someone outside the home; attacking with a weapon; throwing rocks or bottles; setting fires to buildings, cars or property; carrying a hidden weapon; and participating in a gang fight. (Using weapons or force to rob someone and picking pockets or purse snatching were not reported in the third interview.)


The main individual-level exposure variable for this study was violent victimization, as reported at the second interview. Measures included whether girls were attacked with a weapon, beaten up, chased, shot at, or threatened with serious harm in the past year, as well as reports of past-year sexual assault (molestation or rape). These measures come from My Exposure to Violence, an instrument developed by the PHDCN.34 Psychometric study of the scale of violent victimization showed good item fit, good reliability (Cronbach α = .83), and good construct validity (the scale relates to behavioral and emotional symptoms in predictable ways). It was initially determined that sexual assault was better treated as a separate item since it did not correlate as highly with the others and the Cronbach α for the nonsexual victimization scale did not decrease with its deletion. However, given that a high degree of overlap existed between those who were physically and sexually victimized and the low prevalence of past-year sexual assault in our sample, we constructed one combined measure of violent victimization, indicating experiencing at least one of those forms—sexual and/or nonsexual—of violent victimization in the past year.


All individual and family confounders were measured at baseline. Using the same measures of violent events as was used for the study outcome measure, self-reported violence is a sum of violent behaviors by the girls in the year prior to the baseline interview. Family socioeconomic status is a composite of 3 variables: parental income, parental educational level, and parental occupational code, created using principal components analysis. The maximum value was used for each of the 3 variables if there were 2 parents in the household. Deviance of peers consists of the degree to which girls’ peer groups engaged in risky and/or prosocial behaviors. Given the large number of indicators of peer deviance, cluster analysis helped us to first identify groupings that represented substantive dimensions of interest. Principal components analysis then allowed us to create uncorrelated weighted linear composites with maximum variances. The original 30 variables measured the number of peers who displayed the following specific characteristics in the past year: cluster 1, involved in risky activities (skipping school, taking a vehicle without permission, selling drugs, using illegal substances, and/or having sexual intercourse); cluster 2, displaying negative relations with others (not getting along with teachers, disobeying school rules, getting in trouble in school and/or at home, lying and/or disobeying adults, and hitting someone to hurt them); cluster 3, displaying positive characteristics (considered good citizens, good students, and/or generally honest); cluster 4, involved in criminal activities (stealing, breaking and/or entering, and/or attacking with a weapon); and cluster 5, involved in prosocial activities (nonsport school activity, school sports, or community and/or religious activities). Illegal substance use is the number of days of use of alcohol, marijuana, cocaine, crack, and inhalants in the past year. Family structure was classified as whether the youth respondent was living with 1 biological parent, 2 biological parents, 1 nonbiological parent, or 1 biological and 1 nonbiological parent. Age, race/ethnicity, and parents’ marital status (married vs not married) also were included.


For each neighborhood cluster, neighborhood-level measures were estimated using aggregated data from 3 sources: the US Census, the 1995 PHDCN Community Survey, and the Chicago Police Department. These data sources provided estimates of neighborhood conditions independent of data obtained from individual girls and their caregivers. Violent victimization at the neighborhood level was measured by 1 dichotomous item from the 1995 community survey (aggregated to the neighborhood cluster level) asking whether anyone in the household had been the victim of a mugging, fight, or sexual assault since they began living in that neighborhood. The Chicago Police Department provided data to estimate the 1995 homicide rate for each neighborhood cluster. This rate was calculated by taking the number of homicides and dividing by the 1990 population count, using the formula log{[(homicides/population count)×100 000] + 1}. Additionally, a neighborhood-level covariate was used to control for potential confounding: a scale of concentrated poverty was derived from 1990 US Census measures of (1) percentage living below the poverty line, (2) percentage unemployed, and (3) percentage on public assistance, using principal components analysis.


To describe the patterns of violent behavior across the 3 waves of data collection, we constructed age-violence curves delineated by individual-level victimization. Coefficients from a repeated-measures marginal model with a binomial distribution were used to create the curves. Violent behaviors and history of victimization at all 3 waves were in the model, as well as family socioeconomic status. Since there is a trend for violence perpetration to desist during late adolescence,35 it was expected that violence would decrease across the years of the study. Age-violence curves were created for the whole sample, as well as separately for girls exposed and unexposed to violent victimization. We hypothesized that girls exposed to violence would have higher rates of violent perpetration and perhaps desist more slowly during the study period.

Next, we estimated the likelihood of perpetrating violence by using a multivariate, 3-level random intercept Rasch measurement model with wave 3 violence items at level 1; individual-level predictors, a propensity score in quintiles, and random effects at level 2; and neighborhood-level covariates at level 3. Raudenbush et al33 described and demonstrated the validity of this method.


Rather than include a large set of potential confounders of the main relationship of interest in the multilevel models, a propensity score derived from the baseline data was used to represent the confounders in a single composite.36 Propensity score analysis has been shown to produce estimates that are less biased, more robust, and more precise than logistic regression estimates when there are 7 or fewer exposed subjects per confounder.37 Propensity score analysis uses a preliminary regression model to identify predictors of victimization (the exposure in this study) from observed background characteristics that preceded the victimization. It then uses the estimated probability of being victimized to represent all of the background characteristics. By dividing the propensity score into quintiles, and examining the data to see whether a balanced number of victimized and nonvictimized girls fall into each quintile, we were able to ensure that we had sufficient data to compare exposed and unexposed subjects who had similar values on the confounding variables. Additionally, use of this method avoided potential overparameterization of the model, given the large number of confounders and the low prevalence of violent behavior by the third interviews.

Confounders in the propensity score (all described above) included age at enrollment, race/ethnicity, parental marital status, family structure, and socioeconomic status, in addition to behavioral indicators found to have a significant association with being in fights as either perpetrator or victim: peer deviance,3841 illegal substance use,40 and previous perpetration of violence.41 Missing confounder variables were given “0” if categorical and the mean value if continuous, and dummy variables indicating the presence of missing observations for each specific covariate were included. This imputation of missing confounder variables affected 5.0% of the sample. The one neighborhood-level variable included as a potential confounder, concentrated poverty, was entered separately into the 3-level analytic models, rather than as a neighborhood-level propensity score.


Models first estimated the association between victimization and propensity to perpetrate violence, controlling solely for age (centered at the age of 14 years, since it was the mean age to commit violent acts among girls as noted in the literature) and age squared at level 1. The models then introduced the propensity score to adjust for individual-level confounders at level 2. The propensity score was divided into quintiles and entered as 4 categories (quintile 5 as the reference) to avoid assumptions of linearity of the association between the propensity score and the outcome. In addition to its inclusion in the propensity score, additional adjustment for baseline reports of perpetration of violence was initially introduced in the models. However, since such adjustment did not change the magnitude or significance of the associations and baseline violence was not significantly associated with future perpetration after adjustment for the propensity score, this variable was deleted from the models. Models then estimated the association between neighborhood-level violence and perpetration by introducing neighborhood-level victimization and the 1995 neighborhood homicide rate at level 3. Further adjustment was made for neighborhood-level confounding by introducing concentrated poverty. Propensity scores were estimated using SAS, version 8,42 and multilevel models were estimated using HLM, version 5.43


Of the 1164 girls who had completed the first interview, 82% were retained at the second interview, and 68.6% were retained at the third interview. Of the 799 respondents who participated in all 3 interviews, 5 persons had missing observations on all the victimization exposure items at wave 2, and 1 person had missing data on all the violent perpetration items at wave 3. This led to a sample of 793 subjects. To maximize the longitudinal order of the data, we then removed 156 subjects who had been victimized prior to wave 2. We had initially considered using the sample of 793 subjects and controlling for victimization prior to wave 2. This led to an imbalance in the propensity score whereby 157 subjects did not have a “match” of unexposed in terms of control for confounding and would have to be removed from the analysis. The final sample thus consisted of 637 girls who (1) completed all 3 interviews, (2) answered at least 1 of the items of interest on violent perpetration and 1 of the items of interest on victimization exposure, and (3) had not been victimized prior to wave 2.

The final analytic sample differed in 3 important ways from those who dropped out of the study or were removed from the study because of missing observations or prior victimization: the removed subjects were more likely to report deviant peers, less likely to have married caregivers, and more likely to have a higher rate of baseline violence.


The Figure illustrates the pattern of girls’ violent behavior across the 3 waves of data. Looking at the curve for the full sample, there was a peak in violent behavior for all girls at the age of 14 years, with an initial increase with age, then a later decrease, as expected during adolescence.35 By displaying stratified curves (victimized at any wave vs not victimized at any wave vs all respondents), these curves give preliminary indications that the prevalence of violent behavior is higher among girls who report violent victimization than among girls who did not. Additionally, the acceleration of violent behavior was steeper for the victimized girls, with a peak at the age of 17 years. More than one third (38%) of the respondents reported perpetrating at least 1 type of violent behavior in the 12 months prior to the baseline interview; more than one quarter (28%) of the girls reported perpetrating violent behavior in the 12 months prior to the first follow-up interview, and one-seventh (14%) of the girls reported violence at the third interview.

Age-violence curves by history of victimization and for total sample: Project on Human Development in Chicago Neighborhoods waves 1 through 3, for cohorts aged 9 through 15 years. VCT indicates violent victimization.

Age-violence curves by history of victimization and for total sample: Project on Human Development in Chicago Neighborhoods waves 1 through 3, for cohorts aged 9 through 15 years. VCT indicates violent victimization.

Table 2 lists the prevalence of girls’ perpetration of violent acts by history of exposure to violent victimization. Girls who had been violently victimized in the past year (n = 184) were more likely to report perpetration of violence at wave 3 (24.5%) than girls who had never been victimized (11.3%) (Table 2). As summarized in Table 3, girls who had been violently victimized also were more likely to report associations with deviant peers and using illegal substances, and were in the oldest age cohort at baseline. They were less likely to report having peers who exhibited positive characteristics and more likely to report black or other race/ethnicity and having unmarried parents (Table 3).

Table 2. 
Girls’ Violent Behavior, by Violent Victimization in the Past Year, PHDCN Wave 3, Ages at Enrollment: 9, 12, and 15 Years*
Girls’ Violent Behavior, by Violent Victimization in the Past Year, PHDCN Wave 3, Ages at Enrollment: 9, 12, and 15 Years*
Table 3. Girls’ Sociodemographic Characteristics, by Violent Victimization in the Past Year PHDCN Waves 1 and 2; Ages at Enrollment: 9, 12, and 15 Years*
Table 3. Girls’ Sociodemographic Characteristics, by Violent Victimization in the Past Year PHDCN Waves 1 and 2; Ages at Enrollment: 9, 12, and 15 Years*

The propensity score model, which denotes the propensity of individuals to be victimized, had a c-statistic of 0.72. This indicated that the propensity score model had an adequate capacity to predict exposure to victimization; thus, by including the propensity score in the main study model, it was possible to adequately control for selection.

Table 4 lists the crude and adjusted models estimating the association between past-year victimization experiences reported at wave 2 and the probability of reporting perpetrating violence in wave 3. It also reports on the association between neighborhood violence and the probability of violent behavior by girls. Model 1 indicates that, controlling only for age at wave 3, girls who had been victimized by wave 2 had 2.9 times higher odds (95% confidence interval [CI] = 1.7-5.1) of reporting having perpetrated violence at wave 3 than girls who had not been victimized. Adjustment for the propensity score resulted in a decrease in magnitude of the odds of perpetrating violence (odds ratio [OR] = 2.4, 95% CI = 1.3-4.5). Further adjustment for concentrated poverty of the neighborhood did not change the magnitude of the estimate. Concentrated poverty was independently associated with a higher probability of perpetrating violence (OR = 1.5, 95% CI = 1.1-2.1).

Table 4. 
Multilevel, Hierarchical Nonlinear Models of the Odds of Perpetrating Violence Associated With Prior Violent Victimization and Neighborhood Characteristics PHDCN Waves 1 Through 3; Ages at Enrollment: 9, 12, and 15 Years (n = 637)
Multilevel, Hierarchical Nonlinear Models of the Odds of Perpetrating Violence Associated With Prior Violent Victimization and Neighborhood Characteristics PHDCN Waves 1 Through 3; Ages at Enrollment: 9, 12, and 15 Years (n = 637)

The probability of violent perpetration by girls was not higher in neighborhoods with more reports of violent victimization by residents. However, in neighborhoods with higher homicide rates, there was a trend toward higher probability of violence by girls (OR = 1.1, 95% CI = 0.9-1.02). When we examined cross-level interactions between neighborhood violence and individual victimization (Table 5), we found that the effect of individual-level victimization on subsequent violent behavior was approximately the same whether girls lived in neighborhoods without homicides (OR = 5.5, 95% CI = 2.3-13.0) or in neighborhoods with a 1-unit increase in the homicide rate (OR = 5.4, 95% CI = 4.3-6.8). However, we found a cross-level interaction between neighborhood concentrated poverty and individual victimization. The effect of individual-level victimization (with or without homicides, OR = 3.9, 95% CI = 3.0-4.9 or OR = 4.0, 95% CI = 1.8-9.0, respectively) on violent behavior was greater for girls from neighborhoods which were better off (with or without homicides) while the effect of victimization on violent behavior in neighborhoods with a 1-unit higher level of concentrated poverty was less strong and no longer statistically significant, with or without homicides (OR = 1.9, 95% CI = 0.4-8.7 or OR = 2.0, 95% CI = 0.4-9.0, respectively).

Table 5. 
Multilevel, Hierarchical Nonlinear Models of the Odds of Perpetrating Violence Associated With Prior Violent Victimization and Neighborhood Characteristics, Which Include Cross-Level Interactions Between Individual-Level Victimization and Neighborhood-Level Features: PHDCN Waves 1 Through 3; Ages at Enrollment: 9, 12, and 15 Years (n = 637)
Multilevel, Hierarchical Nonlinear Models of the Odds of Perpetrating Violence Associated With Prior Violent Victimization and Neighborhood Characteristics, Which Include Cross-Level Interactions Between Individual-Level Victimization and Neighborhood-Level Features: PHDCN Waves 1 Through 3; Ages at Enrollment: 9, 12, and 15 Years (n = 637)

In this study, adolescent girls were more likely to act violently if they had previously experienced violent victimization, and if they lived in impoverished and/or severely violent communities. Additionally, a cross-level interaction was found indicating that, in neighborhoods that have less concentrated poverty, the effect of individual-level victimization on violent behavior is even stronger. Further investigation into neighborhood characteristics is needed to explain this differential impact of victimization.

These results highlight the critical role that violent victimization plays in the development of aggression by girls. Even when socioeconomic status, previous violence perpetration, deviant peer behavior, illegal substance use, and other family and individual characteristics were considered, violent victimization remained an important risk factor for subsequent violent behavior by girls.

In the current sample of girls who were aged 9 through 15 years at baseline and followed up for 7 years, violent behavior peaked at the age of 14 years. Other studies of girls’ violence have shown that rates of the most serious self-reported violence by girls peaks about age 14 years and then decline, except for a highly aggressive subgroup.14,30,44 In this study, there was a steep decline in the prevalence of girls’ violent behavior between the baseline interview (38%) and subsequent interviews (28% and 14%, waves 2 and 3, respectively). This decline echos the decrease noted in other assessments of youth violence in the United States during this period.8,13,14,45 The finding that violently victimized girls show higher levels of violence across all age groups indicates a need for early intervention, as well as a possibility that early support for girl survivors of violence would avert later aggression.

In this study, victimization might have happened at home, at school, or in neighborhoods. One could argue that both victimization and perpetration could have been in the context of a single fight between youth. However, we controlled for baseline factors that put youth in such risky contexts, including illegal substance use, deviant peer behaviors, girls’ own previous violent behavior, and neighborhood-level concentrated poverty. With these controlled, the effects of prior victimization on girls’ perpetration of violence persisted over time, independent of being in a peer group with norms of fighting and other risky behaviors.

Studies of the relationship between sexual assault and subsequent violent behavior are scarce in the literature; only 2 previous studies with related findings were identified,46,47 both using cross-sectional, individual-level data. One analyzed 1995 National College Health Risk Behavior Survey data and found that lifetime forced sexual intercourse among girls was associated with over twice the odds of fighting with spouses or boyfriends in the 12 months prior to the survey.46 The second was a nationally representative sample of students in grades 8 to 12 in France; this study found a significant relationship between rape and regular perpetration of violence in the past year for both girls and boys.47 These findings suggest that, from a clinical perspective, violence by girls may serve as an indicator of need for interventions focused on healing from sexual assault. Unfortunately, because of the low prevalence of past-year reported sexual assault among this sample of girls, we had to combine it with other types of victimization, rather than evaluate this relationship separately.

In addition to suggesting the potential importance of early interventions for victimized children and adolescents, results of the current inquiry suggest that a much stronger emphasis on safety for children in their neighborhoods, schools, and families would yield the largest reduction in youth violence. To restrict the focus to individual interventions alone will leave those in unsafe settings with correspondingly poor outcomes.


This study relied primarily on girls’ self-reports. These may have been subject to social desirability bias if girls exaggerated aggressive behaviors or were reluctant to disclose victimization experiences, especially sexual assault. With the longitudinal design, we also had missing data that resulted in the deletion of respondents from the analysis. This may have resulted in the introduction of additional bias since removed subjects were more likely to report deviant peers, less likely to have married caregivers, and more likely to have a higher rate of baseline violence. Additionally, the steep decline in perpetration of violence across the 3 waves of data resulted in a low prevalence of the outcome, which may have affected statistical power to detect associations prospectively.48 For example, we expected to see an association between parent-to-child physical aggression49 and subsequent violent behavior, but this was the case only at baseline. Finally, although the propensity score analysis allowed us to account for many potential confounders, there is a possibility that there were unmeasured confounders that were not included in the analysis.


A major strength of this study is its multilevel, longitudinal design. Prospective designs are particularly important in assessing exposures to or experiences with violence across multiple settings, especially when studying adolescence, a time of so many developmental changes. The focus on perpetration by girls also is significant, given the relatively scant literature on the origin of girls’ aggression using population-based samples. Finally, we used multilevel methods to evaluate the independent contribution of neighborhood effects on girls’ violence, including their interactions with individual exposure to violent victimization.


Knowledge of mechanisms by which victimization and exposure to neighborhood violence affects perpetration by youth is as yet unrefined. Turner and Lloyd,50 discussing their findings on the effects of multiple traumas for a child, suggest that one mechanism may be impairment of the development of coping skills to deal with stressors. Some theorists posit a social learning model in which adolescents learn from others that violent behavior is normative51,52; others offer biosocial explanations for the development of youth violence.3

Current youth violence prevention efforts focus primarily on intervention at the individual level: counseling for children or teens who demonstrate problems, punishment for children as individuals, attributions of adult intentions or characteristics to children, and application of adult remedies. In contrast, successful interventions, to the extent that they have been evaluated, include multilevel interventions integrating family, school, and community-level efforts.53 Our results suggest that a continued emphasis on individual-based remedies or penalties will be unsuccessful unless additional efforts are undertaken to reduce childhood victimization and improve the conditions in which many urban girls live.

Correspondence: Beth E. Molnar, ScD, Department of Society, Human Development, and Health, Harvard School of Public Health, 677 Huntington Ave, Sixth Floor, Boston, MA 02115 (bmolnar@hsph.harvard.edu).

Disclaimer: The manuscript is solely the responsibility of the authors and does not represent official views of the Centers for Disease Control and Prevention.

Back to top
Article Information

Accepted for Publication: February 22, 2005.

Funding/Support: This study was supported by grant R49/CCR118602 from the Centers for Disease Control and Prevention, Atlanta, Ga, to the Harvard Youth Violence Prevention Center. Funding for the PHDCN was provided by the John D. and Catherine T. MacArthur Foundation, Philadelphia, Pa; the National Institute of Mental Health, Bethesda, Md; and the National Institute of Justice, Washington, DC.

Acknowledgment: We extend our gratitude to Harvard Youth Violence Prevention Center director David Hemenway, PhD; to Stephen Raudenbush, EdD; Tobias Kurth, ScD; and John Seeger, DrPh; for their guidance on our statistical analyses; and to the field staff who tirelessly collected these data in Chicago. We also acknowledge the roles of Jeanne Brooks-Gunn, PhD; Felton Earls, MD; Stephen Raudenbush, EdD; and Robert Sampson, PhD, scientific directors of the PHDCN. We are especially grateful to the families participating in the PHDCN, who gave us their time and energy to help us better understand their lives and communities.

Champion  HLODurant  RH Exposure to violence and victimization and the use of violence by adolescents in the United States. Minerva Pediatr 2001;53189- 197PubMed
Menard  S Short- and Long-term Consequences of Adolescent Victimization.  Washington, DC Office of Juvenile Justice and Delinquency Prevention February2002;
Scarpa  A Community violence exposure in young adults. Trauma Violence Abuse 2003;4210- 227PubMedArticle
Grunbaum  JAKann  LKinchen  S  et al.  Youth risk behavior surveillance—United States 2003 [published correction appears in MMWR Surveill Summ. 2004;53(24):536]. MMWR Surveill Summ 2004;53 (2) 1- 96PubMed
Flannery  DJSinger  MIWester  K Violence exposure, psychological trauma, and suicide risk in a community sample of dangerously violent adolescents. J Am Acad Child Adolesc Psychiatry 2001;40435- 442PubMedArticle
Mollen  CJFein  JALocalio  ARDurbin  DR Characterization of interpersonal violence events involving young adolescent girls vs events involving young adolescent boys. Arch Pediatr Adolesc Med 2004;158545- 550PubMedArticle
Schwab-Stone  MChen  CGreenberger  ESilver  DLichtman  JVoyce  C No safe haven, II: the effects of violence exposure on urban youth. J Am Acad Child Adolesc Psychiatry 1999;38359- 367PubMedArticle
Snyder  HN Juvenile Arrests 2002.  Washington, DC Office of Juvenile Justice and Delinquency Prevention2004;
Chesney-Lind  M Girls, Violence, and delinquency: popular myths and persistent problems. White  SOed.Handbook of Youth and Justice New York, NY Kluwer Academic/Plenum Publishers2001;
Chesney-Lind  MSheldon  R Girls, Delinquency, and Juvenile Justice. 2nd ed. Pacific Grove, Calif Brooks/Cole Publishing Co1998;
Belknap  JHolsinger  K An overview of delinquent girls: how theory and practice have failed and the need for innovative changes. Zaplin  RTed.Female Crime and Delinquency: Critical Perspectives and Effective Interventions Gaithersburg, Md Aspen Publishers1998;31- 64
Locy  T Like mother, like daughter: why more young women follow their moms into lives of crime. US News World Rep October4 1999;12718
US Department of Health and Human Services, Youth Violence: A Report of the Surgeon General.  Rockville, Md National Center for Injury Prevention and Control, Centers for Disease Control and Prevention2001;
Kann  LWarren  CWHarris  WA  et al.  Youth Risk Behavior Surveillance—United States, 1993. MMWR CDC Surveill Summ 1995;44 (1) 1- 56PubMed
Song  LYSinger  MIAnglin  TM Violence exposure and emotional trauma as contributors to adolescents' violent behaviors. Arch Pediatr Adolesc Med 1998;152531- 536PubMedArticle
Chang  JJChen  JJBrownson  RC The role of repeat victimization in adolescent delinquent behaviors and recidivism. J Adolesc Health 2003;32272- 280PubMedArticle
Lederman  CSDakof  GALarrea  MALi  H Characteristics of adolescent females in juvenile detention. Int J Law Psychiatry 2004;27321- 337PubMedArticle
Resnick  MDIreland  MBorowsky  I Youth violence perpetration: what protects? what predicts? findings from the National Longitudinal Study of Adolescent Health. J Adolesc Health [serial online] 2004;35424- e1-e10Available at:http://www.jahonline.org/article/PIIS1054139X0400165X/fulltextdoi:10.1016/jadohealth.2004.01.011Accessed February 22, 2005
Herrera  VMMcCloskey  LA Sexual abuse, family violence, and female delinquency: findings from a longitudinal study. Violence Vict 2003;18319- 334PubMedArticle
Rivara  FPShepherd  JPFarrington  DPRichmond  PWCannon  P Victim as offender in youth violence. Ann Emerg Med 1995;26609- 614PubMedArticle
Margolin  GGordis  EB The effects of family and community violence on children. Annu Rev Psychol 2000;51445- 479PubMedArticle
Molidor  CE Female gang members: a profile of aggression and victimization. Soc Work 1996;41251- 257PubMed
Odgers  CLMoretti  MM Aggressive and antisocial girls: research update and challenges. Int J Forensic Mental Health 2002;1103- 119Article
Buka  SLStichick  TLBirdthistle  IEarls  FJ Youth exposure to violence: prevalence, risks, and consequences. Am J Orthopsychiatry 2001;71298- 310PubMedArticle
DuRant  RHCadenhead  CPendergrast  RASlavens  GLinder  CW Factors associated with the use of violence among urban black adolescents. Am J Public Health 1994;84612- 617PubMedArticle
Leventhal  TBrooks-Gunn  J The neighborhoods they live in: the effects of neighborhood residence on child and adolescent outcomes. Psychol Bull 2000;126309- 337PubMedArticle
Bingenheimer  JBRaudenbush  SW Statistical and substantive inferences in public health: issues in the application of multilevel models. Annu Rev Public Health 2004;2553- 77PubMedArticle
Molnar  BEMiller  MAzrael  DBuka  SL Neighborhood predictors of concealed firearm carrying among children and adolescents: results from the project on human development in Chicago neighborhoods. Arch Pediatr Adolesc Med 2004;158657- 664PubMedArticle
Earls  FBuka  SL Project on Human Development in Chicago Neighborhoods: Technical Report.  Rockville, Md National Institute of Justice1997;
Elliott  DSHuizinga  DMorse  BJ Self-reported violent offending: a descriptive analysis of juvenile violent offenders and their offending careers. J Interpersonal Violence 1986;1472- 514Article
Andrich  D Rasch Models for Measurement. 07-068 Thousand Oaks, Calif Sage Publications1988;
Wright  BDMasters  GN Rating Scale Analysis: Rasch Measurement.  Chicago, Ill MESA Press1982;
Raudenbush  SWJohnson  CSampson  RJA Multivariate, multilevel rasch model with application to self-reported criminal behavior. Sociol Method 2003;33169- 211Article
Selner-O'Hagan  MBKindlon  DJBuka  SLRaudenbush  SWEarls  FJ Assessing exposure to violence in urban youth. J Child Psychol Psychiatry 1998;39215- 224PubMedArticle
Hirschi  TGottfredson  M Age and the explanation of crime. Am J Sociol 1983;89552- 584Article
Rubin  DB Estimating causal effects from large data sets using propensity scores. Ann Intern Med 1997;127757- 763PubMedArticle
Cepeda  MSBoston  RFarrar  JTStrom  BL Comparison of logistic regression versus propensity score when the number of events is low and there are multiple confounders. Am J Epidemiol 2003;158280- 287PubMedArticle
Henry  DBTolan  PHGorman-Smith  D Longitudinal family and peer group effects on violence and nonviolent delinquency. J Clin Child Psychol 2001;30172- 186PubMedArticle
Prinstein  MJBoergers  JSpirito  A Adolescents' and their friends' health-risk behavior: factors that alter or add to peer influence. J Pediatr Psychol 2001;26287- 298PubMedArticle
Fergusson  DSwain-Campbell  NHorwood  J How does childhood economic disadvantage lead to crime? J Child Psychol Psychiatry 2004;45956- 966PubMedArticle
Heinze  HJToro  PAUrberg  KA Antisocial behavior and affiliation with deviant peers. J Clin Child Adolesc Psychol 2004;33336- 346PubMedArticle
SAS Institute, SAS User's Guide, Version 8.  Cary NC SAS Institute1999;
Raudenbush  SBryk  ACheong  YFCongdon  R HLM 5: Hierarchical Linear and Nonlinear Modeling.  Lincolnwood, Ill Scientific Software International Inc2000;
Elliott  DS Serious violent offenders: onset, developmental course, and termination: the American Society of Criminology 1993 Presidential Address. Criminology 1994;321- 21Article
Grunbaum  JAKann  LKinchen  S  et al.  Youth risk behavior surveillance–United States, 2001. MMWR Surveill Summ 2002;51 (4) 1- 64PubMed
Brener  NDMcMahon  PMWarren  CWDouglas  KA Forced sexual intercourse and associated health-risk behaviors among female college students in the United States. J Consult Clin Psychol 1999;67252- 259Article
Choquet  MDarves-Bornoz  J-MLedoux  SManfredi  RHassler  C Self-reported health and behavioral problems among adolescent victims of rape in France: results of a cross-sectional survey. Child Abuse Negl 1997;21823- 832PubMedArticle
Sampson  RJMorenoff  JDGannon-Rowley  T Assessing “neighborhood effects”: social processes and new directions in research. Annu Rev Sociol 2002;28443- 478Article
Molnar  BEBuka  SLBrennan  RTHolton  JKEarls  F A multi-level study of parent-to-child physical aggression: results from the Project on Human Development in Chicago Neighborhoods. Child Maltreat 2003;884- 97PubMedArticle
Turner  RJLloyd  DA Lifetime traumas and mental health: the significance of cumulative adversity. J Health Soc Behav 1995;36360- 376PubMedArticle
Bandura  A Aggression: A Social Learning Analysis.  Englewood Cliffs, NJ Prentice-Hall International Inc1973;
Malik  SSorenson  SBAneshensel  CS Community and dating violence among adolescents: perpetration and victimization. J Adolesc Health 1997;21291- 302PubMedArticle
Dahlberg  LL Youth violence in the United States: major trends, risk factors, and prevention approaches. Am J Prev Med 1998;14259- 272PubMedArticle
Sampson  RJMorenoff  JDRaudenbush  S Social anatomy of racial and ethnic disparities in violence. Am J Public Health 2005;95224- 232PubMedArticle