Correlates of physical violence at school: a multilevel analysis of Australian high school students.

Grunseit, Anne C. ; Weatherburn, Don ; Donnelly, Neil 等

Introduction

Most explanations for school violence have their roots in social disorganisation theory (Shaw & McKay 1969; Sampson et al. 1997). According to this theory, variation in rates of crime across areas (or schools) reflects differences in levels of informal social control. In the school environment, the principal sources of informal social control are the norms and rules governing behaviour on school grounds that are established by school authorities. These norms and rules play a key role in shaping the school climate. The manifestations of a breakdown of informal social control include such things as lack of clarity about school rules, inconsistent or haphazard or discriminatory enforcement of school rules, a failure to challenge nascent threats to school order (e.g. failure to discipline students for racist remarks), poor supervision of students by teachers or the absence of explicit rewards for pro-social behaviour. Viewed from a social disorganisation perspective, school violence is attributable, not to any characteristics the students themselves bring into the school environment, but to a failure on the part of school authorities to establish and reinforce norms against violent behaviour.

Social disorganisation is not the only theory within which variations in school violence might be comprehended. Gottfredson and Hirschi (1990) have argued that weak parent-child attachment, poor parental supervision, ineffective discipline, parental criminality, large numbers of children and/or family dissolution lead to lowered self-control. Individuals with low self-control (according to the theory) respond to provocation without thought, have difficulty delaying gratification and are prone to taking risks (Gottfredson & Hirschi 1990: 97-105). Applied to school violence, social control theory predicts that most of the variation between schools in the proclivity of students to engage in violence is associated with variations between individuals in low self-control or the factors that affect it. School factors enter the picture only insofar as the school plays a role in the process of socialisation. Gottfredson and Hirschi (1990: 105-106) maintain that this role is relatively slight.

Social disorganisation theory is supported by a large body of evidence, both inside and outside the school environment (Pratt & Cullen 2005; Gottfredson & Gottfredson 1985; Welsh et al. 1999; Smith & Thomas 2000; Welsh 2000; Jenkins 1997; Welsh Green et al. 1999; Clarke & Lab 2000; Hoffman & Johnson 2000; Payne et al. 2003). But there is also a large body of evidence suggesting that criminal behaviour has its origins in low self-control and the family characteristics that contribute to it (Blumstein et al. 1985; Loeber & Stouthamer-Loeber 1986; Le Blanc & Loeber 1998). Most investigations of school violence make little attempt to control for individual and family factors when assessing the effects of school-related variables on school violence (see, for example, Raudenbush & Bryk 1986; Felson et al. 1994; Costenbader & Markson 1998; Jenkins 1997). This is unfortunate because the profile of students and their families (and therefore distribution of individual and family-related risk factors for violent behaviour) may differ markedly from one school to another. Failure to institute adequate controls for individual and family-related factors could easily result in a misspecification of the relationship between school-related factors and school violence.

Although social disorganisation and social control theory are the dominant theoretical perspectives on school violence, several factors not encompassed within these theories may also play a role in school violence. These include: school size (Gottfredson & Gottfredson 1985; Hoffman & Johnson 2000), school boredom (Costenbader & Markson 1998), level of attachment to school and school performance (Payne et al. 2003); and racial profile (Gottfredson et al. 2005). State education authorities consulted in the course of designing this study also highlighted a number of factors that have not been the subject of previous research but which, in their opinion, are also relevant to an understanding of school violence in Australia. These include having a high proportion of teachers with limited experience, lacking a peer mediation system (1), having a high proportion of students that have low or elementary language ability and being a single sex (boys) school (see: Rigby 1997; Kingery et al. 1998; Zeira et al. 2003).

The purpose of the present study is to report the results of a large-scale survey of secondary school students conducted in NSW, Australia, conducted with a view to determining what contribution school-related variables make to school violence once controls have been introduced for individual and family-related correlates of violent behaviour. Multilevel modelling techniques were employed to examine the joint effect of three classes of variable:

1. Individual background variables (including demographic, parenting style and personal characteristics)

2. School climate variables (including knowledge and perceptions of school rules, classroom culture); and

3. School structure (including size of school, teacher experience, ethnic profile).

The first group emerge from social control theory and were measured and operationalised only at the individual level (i.e. a value was obtained from each student participating in the study). The second and third groups emerge from social disorganisation theory and were operationalised (respectively) at the individual and school level (i.e. all students within a school have the same value on a school-level variable).

Methodology

Sampling

The sampling strategy was geared toward maximising variation on the key dependent variable (self-reported perpetration of assault) rather than toward obtaining a representative sample of NSW Secondary Schools. The aim was to generate a sample of 60 schools from which 60 students each (two classes of 30 students) would be surveyed. This represents a balance between number of clusters and cluster size appropriate for multilevel modelling (Raudenbush & Liu 2000). Therefore, a two-stage cluster sample of schools was generated as follows: First, school districts in NSW (n = 40) were stratified by geographical region into Major Urban, Minor Urban, and Rural (Australian Bureau of Statistics, 2001). Second, school districts within each region were ranked according to the number of suspensions for school violence; and then a random selection of 15 districts was made. Third, schools within each of the 15 selected districts were grouped into low, medium and high violence strata, again on the basis of the number of short suspensions for school violence. The final sample of 60 schools was then randomly chosen within violence strata, within district. Schools (and districts) were chosen across the geographic regions proportionate to share of short suspensions for school violence across the state (e.g., 40% of short suspensions were in rural areas, therefore 40% of schools came from the rural region).

Measures and variables

The key dependent variable in the analysis was whether a student assaulted another student on school grounds in the preceding 12-month period. We conducted a large survey of secondary school students and asked each student participating in the survey:

 "During the past 12 months, how often have you physically attacked
 another student to hurt them at school or on your way to/from
 school?" The response options for this question ranged between
 'never' and 'five times or more'.

Note that, in what follows, the terms 'assault', 'self-reported involvement in assault' and 'attack' all allude to the patterns of responses to this question.

A full list of the independent variables is displayed in Table 1.

As can be seen from Table I, the variables fall into the three groups described previously: Individual background variables, individual perceptions of school climate and rules, and school characteristics (2).

Data Sources

The primary data source for the study was an anonymous self-completion survey administered to 2,616 Year 8 and Year 9 students in 60 high schools throughout the State of New South Wales (NSW) Australia. As can be seen from Table I, the survey sought information on a range of demographic, personal and family characteristics that have in past research been found to be associated with perpetration of violence both in the school (Hellman & Beaton 1986; Loeber & Stouthamer-Loeber 1986; Jenkins 1997; Hope & Bierman 1998; Warner et al. 1999; Welsh et al. 1999) and more generally (Tremblay et al. 1994; Farrington 1997; Gorski & Pilotto cited Warner et al. 1999; Smith & Thomas, 2000; Weatherburn et al. Hua 2003).

In all, four groups of questions were directed at obtaining information on school climate. The first group gauged knowledge and/or awareness of school rules. The second group measured the consistency and perceived fairness with which school rules were enforced. The third covered aspects of the relationship between teachers and students (e.g., whether teachers were perceived of as being well-organised and helpful) and the fourth group sought to measure the presence and control of racism and bullying in the school.

The school climate questions (i.e., groups 2-4) were presented as statements with which students indicated their level of agreement on a four-option scale from strongly disagree to strongly agree or rarely/never to always/almost always. Note that some of the questions used to measure school climate were prompted by consultations with local educational practitioners and a companion qualitative study of students involved in fights on school premises (Grunseit et al. 2005). These variables were operationalised at the individual level and at the school level (as an average of the individual ratings) to examine the effect of both the individual student's attitude and the aggregate "climate" of the school as a whole.

Although student ratings were the principal source of data concerning the influence of school-related factors on violence, consultations with school authorities suggested that a number of school-level factors could also influence such violence. The most obvious of these is a school's discipline policy. The disciplinary environment of the school was measured through four variables: how long ago the school behavioural conduct policy was last reviewed, whether peer mediation was used in the school, whether or not the school employed an explicit grading system to manage antisocial behaviour and the number of suspensions it had imposed in the school year 2001 for violent behaviour. Data on each of these factors was collected via telephone interview with the principals of schools participating in the survey. Data on suspensions for violent behaviour were obtained from the NSW Department of Education and Training (DET) administrative records (see below).

As noted earlier, consultations with school authorities also led to the identification of a number of other potential correlates of school violence. It was suggested by some informants that secondary schools that have a large number of feeder schools may be more prone to violence because they place existing friendship networks under greater strain. Schools without an explicit transition strategy between primary and secondary school were also said to be at greater risk because the move from primary to secondary school can be challenging for some students and may predispose them to antisocial behaviour (Eccles et al. 1993). Finally, since there is reason to believe that schools that are connected with the community have lower levels of antisocial behaviour (Haynes 1996), the study also included an indicator of parent-school interaction (i.e. whether the school canteen was run by parent/community volunteers or as a business). These factors were operationalised at the school level only.

Although most of our measures of school discipline were drawn from the survey of school students and principals, DET administrative records were used to compile the measure of the number of suspensions for violence in each participating school during 2001. DET records were also used to obtain school-level measures of numeracy, reading ability, language ability, writing ability, school size (indexed by the number of student enrolments), socio-economic status (indexed by whether the school gets priority funding (3)), the racial and cultural composition of the school and whether the school was selective and/or a single sex school.

Administration of survey

Once a suitable school had been selected, one class from Year 8 and one from Year 9 were identified. Questionnaires were then administered by the class teacher and were completed in class.

Sixty schools were chosen from 15 school districts throughout NSW. Nine schools refused to take part in the survey. These schools were replaced using the same sampling procedure as described above. Of 3,251 potential respondents, the parents of 75 students (2.3%) withdrew permission for their child to take part and 50 students (1.5%) refused on their own behalf on the day. A further 510 (15.7%) were absent on the day that the survey took place. This left a total of 2,616 completed surveys, giving a response rate of 80.5%.

Analysis

Exploratory factor analysis (using PCA extraction with Varimax rotation of factors) was used to summarise and synthesise questions relating to parenting style drawn from the Parenting Questionnaire, as described by Lempers et al. (1989). Three scales emerged from the data reduction procedure: a scale that represented a nurturing and positive parenting style (Nurture); a scale that represented a punitive or inconsistent parenting style (Punitive); and a scale that indicated the degree to which parents monitored their child's whereabouts and behaviour (Supervision). For each of these scales, raw scores were summed within each scale such that higher scores indicate greater experience of the parenting style in question.

Variables representing school climate were treated as an ordinal scale unless they showed a non-linear relationship with probability of assault, in which case they were transformed into categorical variables. Either "strongly disagree" or "rarely/never" were designated as the reference category in all instances for consistency. For other indicator and categorical variables the reference category will be indicated in brackets in all results tables. As may be seen by Table 1, the school level variables were also a combination of indicator, categorical and continuous variables. At all phases of the analysis these measures were tested in the multilevel format to allow for the regression of an individual-level outcome variable on school-level predictor variables.

Preliminary analysis and the model reduction strategy were based on that recommended by Sribney (2001) and it proceeded in three stages. First, a dummy variable was created which took the value "1" for a student who said that they had attacked another student in the past 12 months, and "0" if the student said they had not. Then a series of bivariate logistic regressions were carried out to determine the effect of the independent variables unadjusted for other factors. Third, all predictors within each class of variables as outlined in the introduction were jointly regressed against the outcome variable, using a manual backward sequential approach in order to determine which factors, within each family, were independently related to the probability of assault.

Finally, variables found to be significant in the multivariate models for each class were combined together in a multilevel model and refined by backwards-manual elimination, following the procedure outlined by Hox (2002). Consistent with the aims set out in the introduction, elimination of non-significant variables started with the school-level school characteristics, followed individual students' perceptions of the school rules and school climate. Individual background variables were tested last.

The major analytic technique used throughout was logistic regression. The two statistical software programs STATA and MLWin were used. These programs make allowance for the fact that those observations within a cluster (i.e., within a school) are not independent and are more likely to be more homogeneous within the cluster than between clusters (i.e., between schools). Further, explicit multilevel modelling also ensured that the relationships between the independent variables and our measure of school violence were tested at the same level as they were to be interpreted.

The parameters in the final multilevel model were estimated with the second order, penalised quasi-likelihood method using restricted iterative generalised least squares (Guo & Zhao 2000). There were 60 level two units (schools) and an average of 43 level one units (students) per school, which is consistent with sampling for adequate power described in the multilevel literature (Raudenbush & Liu 2000).

Results

Descriptive statistics

A summary of the demographic profile of the total sample is displayed in Table 2.

As may be seen by Table 2, the sample was fairly evenly split between males and females, and between Year 8 and Year 9. The majority spoke English at home, and lived with both their parents. Over three quarters were from homes that had three or fewer children living at home and almost half of the respondents had mothers who aged 40 or older. The overall percentage of respondents who reported attacking another student in the past 12 months was 43.7% (55.5% for male students, and 31.6% for female students). Rates of self-reported assault varied across a large number of the independent variables as explained below.

Given the major focus of this paper is on the final multilevel model, the findings from the preliminary bivariate and within-family multivariate analyses will be reported only in brief.

A higher probability of assault was associated with male and ATSI-identifying students, students with mothers younger than 40 years, who lived with 3 or more siblings or were only children, more punitive parenting, higher impulsiveness, more problems with reading and writing, and more frequent problems with their family in the last 6 months. Lower probability of assault was associated with: living with both parents, more nurturing parenting and close supervision by parents.

Among the school climate variables, not knowing whether the school has a discipline policy, not receiving any formal notice of the school rules, not believing that "you get into big trouble" for breaking the rules, not believing that "you still get to tell your side of the story if you break the rules", agreeing with that "you always being told what you shouldn't do rather than what you should do", and not agreeing that "good behaviour is rewarded at school", were all associated with a higher risk of involvement in assault. The risk of perpetration of assault was also higher where a student reported that: he/she spends a lot of time copying from textbooks or the blackboard, his/her teachers often seem unprepared for class, rarely greet students, seem disorganised, seem to spend more time controlling the class than teaching, or if his/her teachers rarely if ever help a student with his or her work. A student was also more likely to have attacked someone where they believe that: students at their school are racist, kids who make racist remarks at their school do not get into trouble, some students bully other students at their school; or teachers at their school fail to stop bullying when they know about it.

Finally, at the school level, assault was also found to be more common in students attending boys' schools, smaller schools, schools where more than 25 per cent of teachers have less than five years experience, schools with no peer mediation system, schools with a high proportion of students with poor reading or poor language ability. The remainder of the variables detailed in Table 1 were not significant.

Multilevel Model

The method described by Hox (2002) was used for building the final multilevel model. In the null model only 6.4% (4) of the variance was at level 2 (Snijders & Bosker 1999). Once the final model had been fitted, the residual level 2 variance based on middle values for all other variables was 4.9% for male students and 3.8% for female students. The final model included fixed effects for 16 variables and is displayed in Table 3.

The multilevel model included correlates of involvement in assault that related to the individual student's background as well as students' feelings about their school. The only random effect to achieve significance was that for the intercept (p = .001). None of the other variables had a significant random slope (i.e., the slope of the regression lines for each predictor variable did not vary significantly across schools).

In terms of significant fixed effects, as with the bivariate and multivariate analyses, male students (p < .001), students living with one (< .001) or neither of their parents (p = .004) and students whose mother was aged younger than 40 years (p = .004) had greater odds of self-reported assault in the past 12 months. Among the three parenting style scales that were tested in the model (Nurture, Punitive, and Supervision) only the scales relating to supervision and punitiveness were significant (Supervision, p = .01; Punitiveness, p = .04). As one might expect, higher levels of supervision were associated with lower odds of self-reported involvement in an assault (OR = .80), while higher levels of punitiveness were associated with higher odds of self-reported involvement in an assault (OR = 1.24). The Nurture scale was eliminated from the multilevel model (p = .09).

Students reporting more frequent problems with their family (p < .001), and/ or with reading and writing (p = .05) also had increased odds of assault. For example, holding all other factors equal, compared with students who said they never have problems with their family, those students who say they have constant problems have almost two and a half times the odds of self-reported involvement in assault. Finally, students who judged themselves to be impulsive very often also had more than triple the odds of self-reported involvement in assault in the past 12 months (p <.001) compared with those students who felt they were never impulsive.

Eight measures relating to the school remained significant in the model after adjusting for the individual background characteristics of the individual, although only one was measured at the school level (greater than 25% of teachers with less than 5 years experience, p = .02). Specifically, students attending a school where more than a quarter of the teachers have less than 5 years experience have almost 60% higher odds of self-reported involvement in assault. The variable measuring whether the school was a single sex boy's school was eliminated (p = .07).

Student assessments of their school's rules and climate emerged as an important group of correlates of assault. Firstly, students were less likely to report involvement in assault if they also reported having first found out about the school rules formally rather than by informal means (OR = .69, p = .002). However, the odds of assault increased for students who felt: that they spent a lot of class time copying from textbooks (p = .001); that their teachers spent more time in behaviour management than educating the class (p = .005) and that their fellow students were racist (p = .02). Smaller odds were associated with increased agreement that teachers would discipline students who make racist remarks (p = .005) and with increased agreement that teachers would stop bullying at school (p = .06).

Another school climate variable that was significant in the multilevel model was that which measured respondent perceptions of whether the school allowed students to tell their side of the story when they got into trouble (p = .01). None of the multiple comparisons between the reference category (strongly disagree) and the other answer options (disagree, agree, strongly agree) were significant, so it is somewhat difficult to interpret the effect of this variable. However, the direction and relative size of the effects of each response category reflected the trend observed in the bivariate analyses (where the multiple comparisons were statistically significant). Generally speaking, the more students felt that they got to tell their side of the story, the lower the odds of self-reported involvement in assault. There was, however, a slight upturn in odds of assault among those who strongly agreed that 'the school allows students to tell their side of the story'. This anomaly may have arisen because some those who strongly agreed that students get to tell their side of the story may have discovered that this was the case as a result of their involvement in an assault.

Although the school climate variable measuring perceived control of bullying by teachers was only marginally significant (p = .06), given its importance in terms of school-related characteristics that may ameliorate the perpetration of physical violence, it was retained in the final model. The results in relation to this variable showed that if a student believes that teachers always or almost always stop bullying they know about, they have .89 the odds of self-reporting involvement in an assault than someone who feels teachers never do this. School climate measures that were removed from the model were the degree to which students felt good behaviour was rewarded at their school (p =.06) (5), whether they felt their teachers helped them with their work, (p = .12) and whether they felt their teachers were prepared for class (p = .09).

Discussion

The present study sought to examine the joint effect of three classes of variable: Individual background variables (including demographic, parenting style and personal characteristics), school climate variables (including knowledge and perceptions of school rules, classroom culture); and school structure (including size of school, teacher experience, ethnic profile etc).

Amongst the variables highlighted by State school authorities as potentially relevant to an understanding of violence in Australian schools, only teacher experience remained significant in the multi-level analysis. All the other factors highlighted by State education authorities as potentially important (single sex schools, schools that lack a peer mediation system and schools with a high proportion of students that have low or elementary language ability) were correlated with self-reported involvement in violence at the bi-variate level but ceased to be significant once controls were introduced for individual background and school-related variables. This is a somewhat unexpected result as the study included a wide range of school level variables tapping a range of features of schools. The finding of only one significant school level variable in the final model has important implications for education policy and anti-violence strategies in school. However, it may be the case that at least some of the effects of school level variables were masked by the individual level measures of school climate of which six remained in the final model. That is, the effect of factors such as the presence (or absence) of a certain disciplinary system (for example) may be somewhat accounted for in the measures of students' perceptions of the school's fairness at the individual level.

From a theoretical perspective, the results of the present study underline the point, originally made by (Agnew 1999) and later highlighted by Brezina et al. (2001) that a full explanation of community differences in crime rates must draw on a range of theories. In the present case it is clear that neither social control nor social disorganisation theory on its own provides an adequate explanation of school violence. A more reasonable approach would be to combine the two theories and argue that social control variables play an important role in shaping an individual's propensity to involvement in violence but that the level of informal social control in the school environment plays a key role in determining whether and in what circumstances that propensity is translated into violent behaviour. Viewed from this perspective, social control theory helps us understand the disposition to violent behaviour, while social disorganisation theory helps us understand the circumstances in which violent behaviour is more or less likely to occur.

Our findings have encouraging implications for the management of school violence. The fact that school climate plays is a significant independent correlate of the risk of school violence reinforces the point that authorities may be able to reduce the level of violence in schools even if many of the distal causes of such violence lie outside their direct control. They may reduce the risk of school violence by ensuring students are properly informed about school rules and by ensuring that they are consistently and fairly enforced. They may also reduce it by providing teaching programs that are highly structured, include positive rewards, are ability-appropriate and are perceived by students as stimulating (Haynes 1996; Ward 1998; McEvoy & Welker 2000). This might seem like commonsense but, as Gottfredson et al. (2005) point out in the context of the US schools, school administrators often seem to ignore or underestimate the importance of these facts. All the schools included in this study, for example, had some sort of school discipline policy. They varied markedly, however, in the forcefulness with which that policy was communicated to the student population.

It is interesting to note that the finding that students are less likely to become involved in violence if they feel that teachers intercede to stop bullying and/ or teasing resonates with the findings of qualitative research conducted in the United States. Devine (1995), in his analysis of field notes collected from observers placed on school campuses in seven New York City schools over nine years found that disruptive and hostile behaviour flourished where there was a belief among students that school authorities had abdicated their role in disciplining students. Consequently, students in such conditions stopped seeking help from the school and attempted to pre-empt being victimised with their own displays of aggression. Thus the finding in the present study may be reflecting the "privatisation" of protection whereby survival is managed through the appearance of "toughness" as manifested by violence and misbehaviour (Devine 1995).

The finding that students from schools with greater than 25% of teachers with less than five years experience are more likely to report self-reported assault is also quite important. Only one other study has investigated the relationship between teacher experience and school violence. Gottfredson and Gottfredson (1985) in their analysis of the National Institute of Health's Safe School Study correlated the average years of teacher experience in a school with student victimisation. Although they also found an inverse relationship (i.e., greater experience less victimisation) the correlation they observed was not statistically significant even at the bivariate level. The fact that this variable remained significant in our study, even after controlling for a wide range of other school factors, suggests more work needs to be done to understand the ways in which experienced teachers reduce the risk of violence in school.

Finally, although the present study adds to a growing body of research confirming the importance of school culture in school disorder, it has some inherent limitations that need to be understood. First, it cannot be assumed that factors eliminated in the multi-level analysis are unimportant. Many of the measures used in the present analysis may be tapping the same underlying construct. When this occurs the most salient measure of that construct will tend to mask the effects of the others. The unexpected lack of significant factors at the school level may indeed be an artefact of such masking. Second, because the study design is cross-sectional we cannot be sure that our significant independent variables all 'cause' an increase in the risk of violence. In fact in some cases the causal relationship may actually run the other way. Antisocial behaviour, for example, may make it harder to teach students and result in less then optimal teaching methods. It may also invite negative teacher-pupil interactions and cause less time to be spent on instructional interactions, thereby further compromising academic success (Mok & Flynn 1997; Wehby et al. 1998). What is clear already, however, is that an individual's background; their perceptions of the school environment; and the school's structure are all implicated in the occurrence of physical violence on school premises.

References

Agnew, R. (1999) 'A General Strain Theory of Community Differences in Crime Rates', Journal of Research in Crime and Delinquency, 36, 681-704.

Anderson, A.L. (2002). 'Individual and contextual influences on delinquency: The role of the single-parent family', Journal of Criminal Justice, 30, 575-587.

Australian Bureau of Statistics. (2001) Australian Standard Geographical Classification (ASGC) 2001. Canberra: Australian Bureau of Statistics.

Brezina, T., Piquero, A.R. & Mazerolle, P. (2001). 'Student Anger and Aggressive Behaviour in School: An Initial Test of Agnew's Macro-level Strain Theory', Journal of Research in Crime and Delinquency 38(4), 362-386.

Blumstein, A., Farrington, D.P. & Moitra, S. (1985) 'Delinquency Careers: Innocents, Desisters and Persisters', Pp. 187-219 in M. Tonry and N. Morris (eds.) Crime and Justice: An Annual Review of Research, vol. 6. Chicago, University of Chicago Press.

Clarke, R.D. & Lab, S.P. (2000) 'Community characteristics and in-school criminal victimization', Journal of Criminal Justice, 28, 33-42.

Costenbader, V. & Markson, S. (1998) 'School suspension: A study with secondary school students', Journal of School Psychology 36, 59-82.

Devine, J. (1995). 'Can metal detectors replace the panopticon?' Cultural Anthropology, 10, 171-195.

Eccles, J S., Midgley, C.. Wigfield, A., Buchanan, C., Reuman, D., Flanagan, C. & MacIver, D. (1993) 'Development during adolescence: The impact of stage-environment fit on adolescents' experiences in schools and in families', American Psychologist, 48, 90-101.

Elliott, D. S., Wilson, W. J., Huizinga, D., Sampson, R. J., Elliott, A., and Rankin, B. (1996). The effects of neighborhood disadvantage on adolescent development. J. Res. Crime Delinq. 33: 389-426.

Farrington, D.P. (1997) 'Human development and criminal careers', Pp. 361-408 in M. Maguire, R. Morgan & R. Reiner (ed.) The Oxford Handbook of Criminology (2nd edition), Oxford, Oxford University Press.

Felson, R.B., Liska, A.E., South, S.J. & McNulty, T.L. (1994) 'The Subculture of Violence and Delinquency: Individual vs. School Effects', Social Forces, 73, 155-173.

Gottfredson, G. D. & Gottfredson, D.C. (1985) Victimization in school. New York and London, Plenum Press.

Gottfredson, G. D., Gottfredson, D.C., Payne, A.A. & Nisha C. Gottfredson. (2005) 'School Climate Predictors of School Disorder: Results from a National Study of Delinquency Prevention in Schools', Journal of Research in Crime and Delinquency, 43(4), 412-444.

Gottfredson, M.R. & Hirschi, T. (1990) A General Theory of Crime. Stanford, California, Stanford University Press.

Grunseit, A.C., Weatherburn, D. & Donnelly, N. (2005) School violence and its antecedents: Interviews with high school students. Sydney, NSW Bureau of Crime Statistics and Research.

Guo, G. & Zhao, H.H. (2000) 'Multilevel modelling for binary data', Annual Review of Sociology, 26, 441-462.

Haynes, N.M. (1996) 'Creating safe and caring school communities: Comer School Development Program schools', Journal of Negro Education, 65, 308-314.

Hellman, D.A. & Beaton, S. (1986) 'The pattern of violence in urban public schools: The influence of school and community', Journal of Research in Crime and Delinquency, 23, 102-127.

Hoffman, J. P. & Ireland, T.O. (2004) 'Strain and opportunity structures'. Journal of Quantitative Criminology, 20, 263-292.

Hoffman, J. P. & Johnson, R.A. (2000) 'Multilevel influences on school disorder: A comment on Welsh, Greene and Jenkins', Criminology, 38(4), 1275-1298.

Hope, T. L. & Bierman, K.L. (1998) 'Patterns of home and school behaviour problems in rural and urban settings', Journal of School Psychology, 36, 45-58.

Hox, J. J. (2002) Multilevel analysis: Techniques and Applications. Mahwah, NJ, Lawrence Erlbaum Associates.

Jenkins, P.H. (1997) 'School delinquency and the school social bond', Journal of Research in Crime and Delinquency, 34, 337-367.

Kingery, P.M., Coggeshall, M.B. & Alford, A.A. (1998) 'Violence at school: Recent evidence from four national surveys', Psychology in the Schools, 35, 247-258.

Le Blanc, M. & Loeber, R. (1998) 'Developmental criminology updated', Pp. 115-198 in M. Tonry (ed.) Crime and Justice: A review of research, Vol. 23, Chicago, The University of Chicago Press.

Lempers, JD., Clark-Lempers, D. & Simons, R.L. 1989 'Economic hardship, parenting and distress in adolescence', Child Development, 60, 25-39.

Loeber, R. & Stouthamer-Loeber, M. 1986 'Family factors as correlates and predictors of juvenile conduct problems and delinquency', Pp. 29-150 in M. Tonry and N. Morris Crime and Justice: An annual review of research, vol. 7, Chicago, Chicago Press.

McEvoy, A. & Welker, R. (2000) 'Antisocial behaviour, academic failure, and school climate: A critical review', Journal of Emotional and Behavioural Disorders, 8, 130-140.

Mok, M. & Flynn, M. (1997) 'Quality of life and students' achievement in the HSC: A multilevel analysis', Australian Journal of Education, 41, 169-188.

Payne, A.A., Gottfredson, D.C. & Gottfredson, G.D. (2003) 'Schools as communities: The relationships among communal organization, student bonding, and school disorder', Criminology, 41, 749-777.

Pratt, T.C. & Cullen, F.T. (2005) 'Assessing Macro-Level Predictors and Theories of Crime: A Meta Analysis', Pp. 373-441 in M. Tonry (ed.) Crime and Justice, A Review of Research, vol .32, Chicago, The University of Chicago Press.

Raudenbush, S.W. & Bryk, A.S.. (1986) 'A Hierarchical Model for Studying School Effects, Sociology of Education', 59, 1-17.

Raudenbush, S.W. & Liu, X. (2000) 'Statistical power and optimal design for multisite randomized trials', Psychological Methods, 5, 199-213.

Rigby, K. (1997) 'What children tell us about bullying in schools', Children Australia, 22(2), 28-34.

Sampson, R.J., Raudenbush, S.W. & Earls, F. (1997) 'Neighbourhoods and Violent Crime: A Multi-level Study of Collective Efficacy', Science, 277/5328, 918-924.

Shaw, C.R. & McKay, H.D. (1969) Juvenile Delinquency and Urban Areas, Chicago, University of Chicago Press.

Smith, H. & Thomas, S.P. (2000) 'Violent and non-violent girls: Contrasting perceptions of anger experiences, school, and relationships', Issues in Mental Health Nursing, 21, 547-575.

Snijders, T.A.B. & Bosker, R.J. (1999) Multilevel Analysis. London, Sage Publications.

Sribney, W. (2001) 'Stepwise regression with the svy commands', College Station, StataCorp. Retrieved 11th June, 2004. (www.stata.com/support/ faqs/stat/stepsvy.html).

Tremblay, R.E., Pihl, R.O., Vitaro, F. & Dobkin, P.L. (1994) 'Predicting early onset of male antisocial behaviour from pre-school behaviour', Archives of General Psychiatry, 51, 732-739.

Ward, C.M. (1998) 'Student discipline and alleviating criminal behaviour in the inner city', Urban Review, 30, 29-48.

Warner, B.S., Weist, M.D. & Krulak, M. (1999) 'Risk factors for school violence', Urban Education, 34, 52-68.

Weatherburn, D., Fitzgerald, J. & Hua, J. (2003) 'Reducing Aboriginal over-representation in prison', Australian Journal of Public Administration, 62, 65-73.

Wehby, J.H., Symons, F.J. & Canale, J.A. (1998) 'Teaching practices in classrooms for students with emotional and behavioural disorders: Discrepancies between recommendations and observations', Behavioural Disorders, 24, 51-56.

Welsh, W.N. (2000) 'The effects of school climate on school disorder', Annals AAPSS, 567, 88-107.

Welsh, WN., Greene, J.R. & Jenkins, P.H. (1999) 'School disorder: the influence of individual, institutional, and community factors', Criminology, 37, 73-115.

Welsh, WN., Stokes, R. & Greene, J.R. (2000) 'A macro-level model of school disorder', Journal of Research in Crime and Delinquency, 37, 243-283.

Zeira, A., Astor, R.A. & Benbenbishty, R. (2003) 'School violence in Israel: Findings of a national survey', Social Work, 48, 471-483.

(1) Peer mediation is a dispute resolution process whereby students from a senior year mediate between younger students in the case of disagreement or conflict.

(2) Surrounding community characteristics were not included in this analysis firstly because there was little evidence in the literature that these factors were important once school, family and/ or individual factors are taken into account (for example, see Gottfredson & Gottfredson 1985; Hellman & Beaton 1986; Welsh et al. 1999; Clark & Lab 2000; Welsh et al. 2000). Further, a test of variance at the district (i.e., surrounding community) level in the present study indicated no between-district variation.

(3) Priority school funding is a NSW Department of Education indicator of disadvantage.

(4) This may seem to be a small degree of between group variation however group level variation of 10 per cent or less is not unusual in the multilevel analysis literature (see, for example, Anderson (2002), Elliot et al. (1996), Hoffman & Ireland (2004) and Welsh et al. 1999).

(5) Although this variable could also be considered marginal it was not retained as none of the multiple comparisons between categories were significant. Nor was there a clear or interpretable trend in the coefficients.

Anne C Grunseit [1], Don Weatherburn [2], Nell Donnelly [3]

[1] Ph.D. (Macquarie University) Senior Research Officer, National Prescribing Service, PO Box 1147 Strawberry Hills NSW 2012, email agrunseit@nps.org.au

[2] Ph.D. (Sydney University) Director, NSW Bureau of Crime Statistics and Research, GPO Box 6, Sydney, Australia, 2000, email: don_j_weatherburn@agd.nsw.gov.au

[3] Ph.D. (Sydney University), Program Evaluation Manager, National Prescribing Service Limited, email: ndonnelly@nps.org.au

Table 1: Independent variables used in analysis of correlates of
student attacking another student in the past 12 months

Variable (reference category) Type Source

Individual level variables

Demographic & personal characteristics
Sex (male) I Q
Year (Year 8) I Q
Aboriginal or Torres Strait Islander
 (non-ATSI identifying) I Q
Language spoken at home (English) I Q
Parents (two parents at home) C Q
Number of siblings at home (one sibling) C Q
Mother's age (Mother aged over 40 years) C Q
Nurturing parenting scale S(1-4) F
Punitive parenting scale S(1-4) F
Supervision by parents scale S(1-4) F
Self-rated family problems in last 6 months O Q
Self-rated reading problems O Q
Self-rated impulsiveness O Q

School rules & climate

Knows school discipline policy exists
 (doesn't know) I Q
Found out about school rules formally (found
 out informally or not at all) C Q
(School climate statements (strongly disagree))
No one at this school has really been told what
 the rules are O Q
You get into big trouble if you break the rules
 at this school C Q
If you break the rules you still get to tell your
 side of the story C Q
You are always being told at this school what you
 shouldn't do, rather than what you should do C Q
Good behaviour is rewarded at this school C Q
My teachers seem unprepared for the class lessons O Q
We spend a lot of time copying from text books O Q
My teachers greet the students when they walk
 into the classroom O Q
My teachers' lessons are very organised O Q
My teachers spend more time controlling the class
 than teaching O Q
My teachers help me with my work C Q
The students at this school are racist O Q
Kids who make racist remarks get into trouble O Q
Some kids bully other kids at this school O Q
The teachers at this school stop bullying if they
 know about it O Q
I see the principal in the playground with the
 students C Q

School-level variables

Selective school (not selective) I DET
Boys school (not boys school) I S
Number of feeder primary schools (2-5 feeder
 schools) (6) C PQ
School size (number of students) Cont DET
Priority school (not priority funded school) I DET
Geographic location (capital city school) C S
Proportion non-English speaking background C DIET
 (NESB) (< 30% NESB)
Proportion ATSI (< 5% ATSI) C DET
Proportion head teachers < 3 years
 experience (<25%) I PQ
Proportion teachers < 5 years experience (< 25%) I PQ
Peer mediation system (no such system) I PQ
Level system of discipline (no such system) I PQ
Time since policy last reviewed in years O PQ
Short suspensions for violence 2001 Cont DET
Canteen operation (operated by parents) I PQ
Parent involvement in discipline policy I PQ
 review (no)
Transition strategy from primary to high I PQ
 school (no)
Proportion low/elementary reading ability Cont DET
Proportion low/elementary language ability Cont DET
Proportion low/elementary numeracy Cont DET
Proportion low/elementary writing ability Cont DET
Aggregate across students of school climate
 statements (n=16) Cont A

KEY

Type: I--Indicator variable (reference category) (coded 1/0);
C--Multiple-category categorical variable (for all statements,
strongly disagree is reference category); O--Ordinal variable;
Cont--Continuous variable; S--Scale score (range)

Source Q--Information taken directly from student questionnaire;
F--Factor analysis of (questions specified); PQ--Principal's
questionnaire; DET--NSW Department of Education records,
S--Determined from sampling frame; A--Aggregate score

Table 2: Basic characteristics of survey respondents (n = 2616)

Variable N %

Year 8 1350 51.6
Year 9 1266 48.4
Male 1321 50.8
Female 1281 49.2
ATSI 176 6.8
Non-ATSI 2409 93.2
English spoken at home (1) 2035 88.1
Language other than English at home 274 11.9
Lives with both parents 1821 70.2
Lives with one parent 711 27.4
Lives with neither parent 62 2.4
Mother is 35 years or younger 348 13.5
Mother is 36-40 years 833 32.3
Mother is over 40 years 1221 47.4
Didn't know mother's age 176 6.8
Respondent only child at home 258 9.9
One sibling at home 958 37.0
Two siblings at home 769 29.7
Three siblings at home 334 12.9
Four or more siblings at home 270 10.4

Notes to table: (1) There were 307 missing values for this variable

Table 3: Results of multivariate analysis of experience off
physically attacking another student on school premises in
the last 12 months and school level variables.

FIXED EFFECTS Variable (reference category) OR1 p-value2

Intercept .34
Individual level variables
Individual background variables
Sex (female students) 2.78 <.001
Parents (lives with both parents) <.001
Lives with neither parent 2.96**
Lives with one parent 1.48**
Mother's age (Mother aged > 40 years) =.002
Mother aged 35 years or younger 1.77**
Mother aged 36-40 years 1.28*
Don't know mother's age .93
Punitive scale4 1.24 =.04
Supervision scale4 .80 =.01
Problems with family 1.35 <.0.001
Self-reported problems reading/writing 1.19 =.05
Impulsiveness 1.48 <.001
School rules & climate
Found out about school rules formally (found
 out informally/not at all) .69 =.002
If you break the rules you still get to tell
 your side of the story (Strongly disagree) .70 =.01
Disagree .75
Agree 1.17
Strongly agree
We spend a lot of time copying out work from
 text books 1.14 =.001
My teachers spend more time keeping control of 1.19 =.005
 the class than teaching
The students at this school are racist 1.17 =.02
Kids who make racist remarks get into trouble
 with the teachers .85 =.005
The teachers at this school stop bullying if
 they know about it .89 =.06
School level variables
Greater than 25% teachers with < 5 years
 experience (< 25%) 1.56 =.02
RANDOM EFFECTS
Intercept (Uoj) .214 =.001

Note: n = 2136 for individual level variables, and n = 60 for school
level variable

Notes to table: * Comparison with reference category significant at 05,
** significant at .01; (1) The odds ratios indicate the change in odds
of physically attacking another student for a one point increase in a
discrete variable, and the category of interest compared with the
(reference category) for categorical variables; (2) P-value for
overall omnibus test of variable adjusted for clustering