Structural Analysis of 1995–2005 School Crime Supplement Datasets: Factors Influencing Students' Fear, Anxiety, and Avoidant Behaviors

Abstract

The 1995–2005 School Crime Supplement datasets were analyzed using structural equation modeling. Converging evidence across multiple analyses suggests that secure school building policies may not be systematically linked to school disorder and may be more a reactive measure in response to other concerns. Most importantly, measures of incivility explain almost double the variance (31–45% vs. 17–25%) in students' fear, anxiety, and avoidant behaviors, compared to theft and attack, when tested using alternate measurement models. The analysis can inform school-based professionals working with at-risk students who may be exhibiting fear, anxiety, or avoidant behaviors related to a hostile school environment. The findings suggest that local school personnel and educational leaders concerned about school violence should facilitate increased resource supports to address incivility in schools.

KEYWORDS:

• school violence
• fear
• anxiety
• school avoidance

School violence has been considered a serious issue in the United States since public and policy-maker attention was drawn to this topic following the release of the 1978 report, Safe Schools–Violent Schools: The Safe School Study Report to Congress (National Institute of Education, 1978). It was perceived to be at epidemic levels by the early- to mid-1990s (Elliott, Hamburg, & Williams, 1998). Several reports show a significant decline in school violence and disruption from the mid-1990s through the early 2000s, including data from the National Crime Victimization Survey (NCVS) (DeVoe, Noonan, Snyder, & Baum, 2005; Dinkes, Cataldi, Kena, & Baum, 2006). For example, from 1992 to 2004, NCVS-based reports of violent crime at school declined from 48 per 1,000, to 22 per 1,000 students. The Centers for Disease Control (CDC) Youth Risk Behavior Survey (YRBS) data show a marked decline from 1993 to 2005 in students fighting in school and bringing weapons to school, with a more recent leveling off of that decline (Dinkes et al., 2006). Whereas 11.8% of respondents reported having brought a weapon to school during the past 30 days in the 1993 YRBS survey, the figure dropped to the 6% range beginning in 1999, with values of 6.9%, 6.4%, 6.1%, and 6.5% for 1999, 2001, 2003, and 2005, respectively. The trend is relatively flat over the entire time period from 1993 to 2005 for YRBS data on the percentage of students reporting being threatened or injured with a weapon at school and missing school due to safety concerns (Dinkes et al., 2006). Overall, school violence for 1993–2005 declined in many respects, but remains a significant issue.

School violence is a multidimensional issue, reflecting the complexities of society. Violence exists at multiple ecological levels relating to a confluence of risk factors and processes at the individual, peer, family, and societal levels (Mayer & Leone, 2007; Osher, Dwyer, & Jimerson, 2006; Osher et al., 2004). Research has identified several factors influencing an orderly school environment, prosocial student behavior, and student investment in school (Aleem & Moles, 1993; Lab & Clark, 1996; Mayer & Leone, 2007). In particular, several aspects of the school environment have been investigated.

Classroom setting variables such as room layout, activity schedules, curriculum supports, and student-teacher interactions can influence student behaviors (Conroy & Fox, 1994; Van Acker, Grant, & Henry, 1996), as can patterns of teacher movement in the class (Shores, Gunter, Denny, & Jack, 1993). School climate, which has been linked to levels of violence and disruption (Gottfredson, 1995; Leone & Mayer, 2004), has been described in terms of: (a) schoolwide goals, (b) rules and procedures, and (c) an “ethic of caring” (Aleem & Moles, 1993). Knowledge and perception of how school rules are enforced can drive students' investment in the system of rules (Aleem & Moles, 1993; Evertson & Emmer, 1982) and students' views of nonlegitimacy of the system of rules may facilitate school violence (Toby, 1993/1994). Several researchers have suggested that schools may promote aggression and disorder through overly controlling and restrictive approaches that make schools seem somewhat jail-like (Colvin, Kameenui, & Sugai, 1993; Noguera, 1999). Research has also shown that using metal detectors in schools does not systematically reduce classroom violence or levels of fighting in schools (Aleem & Moles, 1993; CDC, 1993).

Recent research has addressed so-called low-level aggression in schools that can include general intimidation, bullying, and exposure to hate language, with social rejection as an outcome. Multiple studies have reported on the widespread nature and harmful effects of these behaviors (Arseneault et al., 2006; Bierman, 2004; Boxer, Edwards-Leeper, Goldstein, Musher-Eizenman, & Dubow, 2003; Ladd, 2003; Limber, 2006; Nansel et al., 2001; Skiba et al., 2004). Students who experienced low-level aggression, directly and vicariously, engaged in aggression more, had more negative future expectations, and perceived lower school safety, compared to students without such experiences (Boxer et al., 2003). School climate, students' sense of connectedness, and levels of incivility in school were major factors shaping students' perception of school safety and may be more salient to intervention efforts than high-level aggression and violence (Skiba et al., 2004).

In a comprehensive review of research on child adaptation to school, using a child by environment perspective, Ladd (2003) pointed to research demonstrating the harmful effects of peer victimization and rejection, leading to disinvestment in school and avoidant behaviors. Later research by Buhs, Ladd, and Herald (2006) also demonstrated a strong link between chronic peer maltreatment and school avoidance. Where much of the attention in school violence research has focused on higher profile theft and personal attack, taken together, these studies suggest that low-level incivility in schools can lead to psychosocial harm for the individual student and the larger student group, and may play a more significant role than previously thought with regard to school violence and students' levels of anxiety about school and related avoidant behaviors.

Characteristics of school organization and management, student-adult interactions, and the overall social context of the school experience can exert an influence on violence in schools (Gaddy, 1988; Lab & Clark, 1996; Leone & Mayer, 2004). Hence, the school environment has great importance as a facilitator or inhibitor of school violence. Trade-offs can exist between school efforts to maintain a system of rules through communicative measures versus control and containment via secure building strategies. Different ideas exist on how best to manage schools, and while not mutually exclusive, there must be some balance between tightly controlled and more nurturing school environments. An unanswered question is: What are the relative contributions of security-oriented strategies that focus on control and containment as opposed to more proactive, communicative approaches?

Mayer and Leone (1999) addressed this question, analyzing the 1995 School Crime Supplement (SCS) dataset, a subset of the NCVS. Using a structural equation modeling (SEM) approach, they examined the influence of secure building strategies and students' understanding of the school's system of rules and procedures on school violence and disruption. They used responses to the 1995 SCS questionnaire from public school students in grades 6–12 (N = 7,593) to construct composite measured variables linked to four constructs used in the SEM analysis: (a) secure building, (b) system of law, (c) school disorder, and, (d) self-protect. Secure building reflects efforts of the school to control its premises. System of law encompasses students' awareness of school rules and the degree to which the rules and their consequences are implemented. School disorder represents personal theft and attack, and the degree of gang and drug-related activity at school. Self-protect reflects students' level of fear and anxiety, and protective actions taken in response to perceived threats at school.

They found that with greater efforts to maintain a secure building through personnel (e.g., guards patrolling halls) and equipment-based measures (e.g., metal detectors, locked doors, etc.), more school violence occurred. A key assumption of this model pertains to the directional arrow from the secure building construct to the school disorder construct. An important question is whether the direction can be plausibly reversed. Mayer and Leone (1999) suggested the possibility of a reciprocal process explaining the relationship between the secure building and school disorder constructs. Of greater importance, Mayer and Leone (1999) reported that with increased student understanding of the school's system of rules, less violence existed. They also replicated prior findings showing that where greater school violence existed, students tended to take more self-protective actions such as avoiding places perceived as being dangerous and exhibiting a continued state of fear.

Mayer (2001) extended this line of inquiry, working with the 1995 and 1999 (N = 5,656) SCS datasets. The core research questions were (a) What is suggested by the results of fitting the 1999 SCS data to the Mayer and Leone model?; and (b) Based on analysis of the 1995 and 1999 SCS data, what is known about the direction of causal influence between secure building strategies and school violence? Mayer (2001) reported similar relationships in the 1999 SCS analysis as were found with the 1995 SCS analysis (see Figure 1). The results suggested that the structural relationships in the model reflect a real world process, as evidenced through replication, and that a fairly consistent relationship existed among these constructs throughout the later 1990s. A reversed arrow model was investigated, with a path from the school disorder construct to the secure building construct, but it proved statistically untenable (Mayer, 2001).

FIGURE 1 Structural equation modeling analysis of 1995 and 1999 school crime supplements (Mayer, 2001; Mayer & Leone, 1999).

Mayer (2001) performed structured means modeling analysis (SMM), examining latent means changes among the four latent constructs from 1995 to 1999. The SMM analysis revealed that secure building strategies increased approximately 0.44 standard deviation units from 1995 to 1999, suggesting a general across-the-board increase in schools' procedures such as restricting building entry, patrolling hallways, requiring visitor identification and sign-in, and so forth. The system of law construct increased only 0.10 standard deviation units, suggesting relative stability over time. School violence and disorder decreased about 0.90 standard deviation units, mirroring other data on school violence trends from the same period. The self-protect construct, which reflects students' degree of fear and anxiety, and self-protective actions taken, also declined modestly, with an average standardized effect of just under –0.30.

The data raised questions about a possible contradiction between the consistent positive correlation over 1995–1999 between the secure building and school disorder constructs, and the latent means changes with secure building increasing and school disorder decreasing. Mayer offered two possible interpretations of this change, one pointing to secure building strategies causing reductions in school disorder, and the other pointing to the influence of unmeasured variables external to the model. No conclusive determination of the relationship between secure building and school disorder constructs was offered by Mayer (2001).

The current research builds on prior work by Mayer and Leone (1999) and Mayer (2001). The goal of this study is to address three questions:

1.	How does the structural model fitted to 1995 SCS data by Mayer and Leone (1999) apply to the 1999, 2001, 2003, and 2005 SCS data and what is suggested by the results of using the later SCS data with the prior model?
2.	What is known about the direction of causal influence between secure building strategies and school violence (based on analysis of data from the five administrations of the SCS)?
3.	What is the relationship of incivility, versus theft and personal attack, to students' levels of fear and anxiety, and avoidant behaviors, when considered as part of a structural model?

This research applies SEM and SMM techniques to the Mayer and Leone (1999) model. The current analysis considers so-called low-level behaviors in schools that can include general intimidation, bullying, and hate language, along with an outcome of social rejection, as previously discussed. Changes in the NCVS SCS survey questionnaire included new questions that provide data on these types of behaviors in the more recent 2001, 2003, and 2005 SCS survey versions. That data allow researchers to consider relationships among these low-level behaviors and students' levels of fear and anxiety, and related avoidant behaviors at school. These relationships can be contrasted to students' fear, anxiety, and avoidant behaviors as they relate to more overt high-level problems such as theft and personal attack.

METHOD

Participants

The 1995, 1999, 2001, 2003, and 2005 NCVS SCS datasets used in this analysis employed a nationally representative stratified multi-stage cluster sample design. This research analyzed a subset of the respondents, including only students in grades 6–12 in public school who reported having attended school at least four of last six months of school. Private school students were not included because most analyses and policy efforts concerning school violence have addressed public school settings. Students who attended fewer than four of the previous six months would be atypical participants and could introduce bias. The cutoff of at least four out of six months of school attendance would most likely exclude those students who were out for extended periods due to sickness, suspensions, or other special reasons. Students who were away from school for more than two of the last six months would not be considered credible witnesses for purposes of providing information on day-to-day experiences at school over a half-year period. Since all interviews were conducted during the regular school year, no students were excluded due to summer vacation absences. Proxy interview data were not used in the analysis, since the proxies—parents in most cases—can't be expected to be very knowledgeable about daily life at school.

Constructs and Variables

Although the constructs in this research represent school-level processes, the source data are at the individual student level, drawn from survey responses. Students were questioned about school rules and policies, knowledge of and experience with student violence, ease of accessing drugs in school, presence of gangs in and around school, as well as information on their family, travel to and from school, and neighborhood crime. They were also asked about fear of victimization at school and self-protective actions they had taken while at school. Composite measured variables built from questionnaire items were aligned with four latent constructs: secure building, system of law, school disorder, and self-protect.

Secure building represents the school's efforts to maintain secure premises. The two indicator variables are ordpers and ordphys. Ordpers was constructed from the sum of recoded scores of three questions involving use of security guards, staff watching hallways, and visitor sign-in procedures. These are personnel-based procedures used to secure premises. Ordphys is derived from the sum of recoded scores on three questions addressing metal detectors, locked door policies, and locker checks. These are equipment-related methods for controlling the physical environment.

System of Law represents students' understanding of the school's rules and its systemic approach to applying consequences for rule infractions. The two measured variables are knowlaw and implaw. Knowlaw is derived from recoded scores on two questions reporting students' perceptions of schoolwide student knowledge of school rules and students' awareness of consequences for rule violations. Implaw (implementation) is derived from recoded scores on two questions asking how strictly and consistently school rules are actually enforced.

School disorder represents general levels of school violence and disruption. The three indicators for this construct are gangpres, drugpres, and percrime. Gangpres is derived from three questions on gang activity in school. Drugpres is the composite of responses to nine questions on drug use and availability (one questionnaire item changed in the 2005 SCS, providing data on whether various drugs were available, but not on relative ease of availability, as previous surveys had). Percrime is derived from several questions pertaining to personal attack and theft.

Self-protect reflects students' degree of fear and anxiety and related avoidant behaviors at school. It has two indicator variables: Stay away and fear attack. Stay away is derived from nine questions regarding places around school students may avoid out of fear. Fear attack is a composite of two questions on fear of attack while at school or traveling to or from school. Earlier analysis of the 1995 dataset explored a third indicator, bringwep (bringing a weapon to school), but it was statistically untenable (Mayer & Leone, 1999). More of the self-protect construct would be particularly important to school officials and other stakeholders, because students' apprehension about their safety at school can have deleterious effects on their school performance (Buhs et al., 2006; Schwartz, Gorman, Nakamoto, & Toblin, 2005). Furthermore, a relatively high degree of this construct would suggest the existence of an unacceptable school climate.

A subsequent part of the analysis (discussed later) involved testing alternate measurement models for the school disorder construct, examining explained variance in students' fear, anxiety, and avoidant behaviors, linked to those alternate models. Thus, a new measured variable, incivil, seen in alternate measurement models numbers two and three (see Figure 2), represents the influence of low-level school disorder (incivility) and was constructed as a composite of four questionnaire items that included being bullied, called a derogatory name, exposed to hate language, and feeling rejected. While there is no universally accepted research-based standard for the construct of incivility in schools, there is a substantial empirical basis for conceptualizing a more global measure of school incivility as a function of these four types of measures (Arseneault et al., 2006; Boxer et al., 2003; Buhs et al., 2006; Ladd, 2003; Limber, 2006; Nansel et al., 2001; Skiba et al., 2004). Where the first three of the four incivil variable items reflect forms of victimization, the fourth, feeling rejected, is more properly considered as an outcome of an incivil and unwelcoming environment. The 2005 SCS questionnaire changed the question on rejection to being excluded (linked to bullying). This outcome measure should be included in this indicator variable for two reasons: (a) the notion of an incivil school environment is a more global measure of the degree of overall intimidation, hostile, and exclusionary climate, and feelings of rejection/exclusion are a natural result of experiencing such a climate; and (b) feelings of rejection and/or being excluded, while not the same as loneliness, can be considered as a proxy for loneliness. This is highly relevant based on research by Nansel et al. (2001) which reported the highest risk of being bullied at school among all measured attributes in the study were for feelings of loneliness, with an odds ratio of over 33:1 for students who felt most lonely, compared to those who felt least lonely.

FIGURE 2 Alternate measurement models for school disorder construct in 2001–2005 analyses.

The observed variables described above are composite (or derived) variables, built through combining the responses on multiple questionnaire items. The original dataset codebooks were checked and recoded to ensure that there was consistency in the direction of numerical codes (e.g., 1 represents a lesser value than 5) and that on a substantive level, the order of the codes were consistent with one another. Some composite variable scores were recoded by a multiplicative factor to allow them to fall within the same general range as other composite variables, as recommended in the SEM methodological literature (Bentler, 1995). Other observed variables in the analysis have scores with a range of 0–3, representing ordered categorical variables. Such variables may not have characteristics based on continuous distribution theory. While SEM software typically includes features to address this scenario (West, Finch, & Curran, 1995, p. 68), it assumes that the categorized variables are tied to an underlying continuous distribution. In datasets where coarsely categorized variables have four or more levels (as in this research), they can be appropriately analyzed as if they have an underlying continuous distribution using robust estimation methods (Bentler & Chou, 1987, p. 88; West et al., 1995, p. 64). In turn, a maximum likelihood (ML) robust analysis option in the EQS software was used.

Model Testing and Analysis

SEM and SMM analysis of 1995–2005 SCS data was conducted with the EQS for Windows program. All data recoding and preparation was done using SPSS for Windows. The raw data matrix was analyzed using the method = ML, robust command due to evidence of considerable nonnormality in two of the measured variables (Newcomb & Bentler, 1988; West et al., 1995). West and colleagues (1995, p. 74) reported that skewness values of 2 and kurtosis values of 7 are indicative of nonnormality.

All nine measured variables used in the prior SEM analyses were virtually the same, except for percrime (personal crime experienced—theft and attacks), which could not be developed directly from the 2001, 2003, and 2005 SCS data, and gangpres (presence of gangs around school), which changed slightly in the same SCS questionnaires. Percrime, along with drugpres and gangpres represent the construct of school violence and disorder, based on this dataset and prior research with the Mayer and Leone model. As recommended to the researcher by a statistician working in collaboration with the National Center for Education Statistics on the SCS (M. Noonan, personal communication, April 12, 2005), data from the 2001, 2003, and 2005 NCVS surveys were used to closely approximate the percrime measured variable that was used in prior analyses of the 1995 and 1999 SCS questionnaires.

As previously suggested, this research sought to compare the relative value of alternate measurement models of school violence and disorder (tied to lower- and higher-level problems) in explaining variance in measures of students' fear, anxiety, and avoidant behaviors. Based on changes in the SCS survey questionnaires from 1995 and 1999, to 2001, 2003, and 2005, the analysis examined alternate models using the 2001–2005 datasets. Figure 2 illustrates three alternate measurement models for the school disorder construct.

Additional statistical concerns in the analysis included missing data, ¹ a Heywood case result for one measured variable error term, ² less than three indicator variables for some constructs, ³ design effect, ⁴ and chi-square tests. ⁵

RESULTS

The results of the 1995 and 1999 SEM analyses were previously reported (see Figure 1). Although the respondent counts varied over the 1995–2005 SCS subsamples analyzed, each involved approximately 50–51% males and 49–50% females. Race and ethnicity percentages varied over the years, with approximately 64–78% White, non-Hispanic, 14–16% Black non-Hispanic, 12–19% Hispanic, and 4-5% other, non-Hispanic. Percentages for the subsamples used in this analysis were consistently within 1% of the corresponding total survey sample race/ethnicity percentages.

While reliability analysis of measurement scales for psychometric instruments is typically performed using Cronbach's alpha, it is less appropriate in analyses such as this, where particular indicator variables do not necessarily represent a unified measurement and are derived from questions that may represent distinct and differing facets of an umbrella indicator. For example, the measured variable, percrime, is a composite variable reflecting responses to several questions about threat of personal harm, theft and attack. The indicator percrime is multifaceted and yields higher and lower values for schools through many different combinations of responses. That is to say, a school could appropriately score higher on percrime as a function of threat, theft, or attack, yet the intercorrelations among the questions might be relatively low for particular schools. This just means that schools may have different types of personal crime configurations, based on these questionnaire items. Generally speaking, Cronbach alpha values for indicator variables were fairly stable across the years and fell into low (under 0.40), moderate (0.40–0.70), and high (above 0.70) ranges, with percrime, ordpers, ordphys, and incivil falling in the low range, implaw, knowlaw, and fear attack in the moderate range, and gangpres, drugpres, and stay away in the high range.

The results of the 2001–2005 analysis are found in Table 1. The 2001–2005 model-1 (percrime) had excellent fit. The other runs had moderately good fit. The 2001 analysis was somewhat problematic due to Heywood results on one measured variable error term. The secure building to school disorder path values were 0.16–0.18 in the 2001–2005 model-1 analyses using percrime, compared to 0.52 to 0.43 in similar 1995 and 1999 models. The much lower association in the 2001–2005 analyses is noteworthy when considered along with additional SMM analysis of latent mean changes from 1999–2001, 2001–2003, and 2003–2005, as discussed below. The R ² value for the school disorder construct ranged from 0.06–0.12 in the 2001–2005 analyses, compared to 0.35 and 0.25 in 1995 and 1999, respectively. The model does a poor job of explaining the construct relative to the 2001–2005 data. This is likely the result of many fewer students reporting indications of school disorder combined with the much lower association between secure building and school disorder. The path value from system of law to school disorder was fairly similar across all five administrations of the SCS survey, suggesting a consistent relationship. The path values from school disorder to self-protect were similar across all five years of the SCS survey, also suggesting a consistent relationship.

TABLE 1 Results of Structural Equation Modeling (SEM) Analyses of 2001–2005 School Crime Supplement (SCS) Datasets With Alternate Models

Model	Measured variables for SD beyond gangpres/drugpres	Robust CFI	Robust RMSEA	Path ^a SB→ SD	Path ^a SL→ SD	Path ^a SD→SP	R ^2 (SD)	R ^2 (SP)
SCS 2001 analyses^a (N = 5,750)
1	Percrime	.959	.028	.17	–.25	.50	.09	.25
2	Incivil	.924	.042	.14	–.29	.66	.10	.44
3	Percrime & incivil	.924	.036	.13	–.29	.67	.10	.45
SCS 2003 analyses^a (N  = 6,225)
1	Percrime	.981	.024	.18	–.22	.42	.08	.17
2	Incivil	.917	.052	.18	–.30	.56	.12	.31
3	Percrime & incivil	.933	.043	.17	–.30	.58	.12	.34
SCS 2005 analyses^a (N  = 4,278)
1	Percrime	.977	.027	.16	–.19	.45	.06	.20
2	Incivil	.954	.038	.17	–.28	.57	.11	.33
3	Percrime & incivil	.958	.034	.17	–.28	.58	.11	.33
Note. CFI = Comparative Fit Index; RMSEA = root mean square error of approximation; SB = secure building; SD = school disorder; SL = system of law; SP = self-protect. All main path parameters significant at p < .01. One error correlation path n.s. Forty-eight to 60 cases in 2001–2003 analysis dropped by the SEM software in alternate model runs due to missing data.
^aStandardized path values.

The R ² values for the self-protect construct under the three alternative measurement models associated with the school disorder construct show a dramatic increase from Model 1, with R ² values ranging from 0.17–0.25 for the 2001–2005 SCS datasets, respectively, to Models 2 and 3, with R ² values of 0.31–0.45. There was negligible increase in variance going from Model 2 to Model 3, where the percrime measured variable was used in addition to the incivil measured variable, above and beyond the drugpres and gangpres measured variables. This key finding is discussed further below.

The SMM analysis of changes in means of latent variables (measured in standard deviation units) across the five years of the SCS dataset is seen in Table 2. The secure building construct increased substantially from 1995–1999 and from 1999–2001, with a minor increase from 2001–2003 and a negligible change from 2003–2005. System of law demonstrated very minor changes over the years. School disorder showed a large decline (−0.89) from 1995 to 1999, but little change from 1999 to 2005. Likewise, self-protect had a moderate decline (−0.29) from 1995 to 1999, a minor decrease (−0.10) from 1999-2001, no change from 2001–2003, and a minor increase (0.07) from 2003–2005.

TABLE 2 Latent Means Changes ^a for 1995–2005 School Crime Supplement (SCS)

Latent construct	1995–1999	1999–2001	2001–2003	2003–005
Secure building	.44	.50	.16	–.03
System of law	.10	.06	.06	–.06
School disorder	–.89	.13	–.08	.23
Self-protect	–.29	–.10	.00	.07
MFI/RMSEA Fit Indices (based on covariancematrix and means)	Median robust CFI = .94 (not based on means)	MFI  = .970RMSEA = .051	MFI  = .984RMSEA = .036	MFI = .946RMSEA = .067
Note. These are median values from 154 structured means modeling analysis runs with the 1995 and 1999 SCS data using an ad hoc procedure to address problems with Lagrange tests of constraint validity. See Mayer (2001) for methodological details. Forty-eight cases were dropped in the 1991–2001 and 2001–2003 analyses by the structural equation modeling software due to missing data. MFI = McDonald Fit Index; RMSEA = root mean square error of approximation; CFI = Comparative Fit Index.
^aStandard deviation units.

Discussion

These analyses present several important findings for educators, school-based clinicians, researchers, and policy makers that warrant serious consideration. First, the analyses demonstrated that the prior Mayer and Leone Model, based on the 1995 SCS, replicated with the 1999 SCS, exhibiting a similar solution with excellent fit, supporting the existence of a consistent relationship among these constructs. SEM analysis of the 2001–2005 SCS datasets demonstrated similar structural path coefficients for paths from system of law to school disorder, and from school disorder to self-protect. This suggests that general relationships among these constructs as found in this model further existed across the years from 1995 to 2005. This is important for two reasons. First, the consistent negative relationship of the system of law construct to school disorder across all five survey administrations from 1995 to 2005 provides evidence of the importance of working with students to make sure that they understand the school's system of rules and have knowledge of the degree to which the school implements the rules and their consequences. Second, the consistent strong positive relationship between school disorder and self-protect suggests that in order to lessen students' levels of fear and anxiety, and related avoidant behaviors, school officials must reduce school disorder.

However, there were changes in other model parameters over the years. The structural path values between secure building and school disorder dropped from a range of about 0.4–0.5 for the 1995 and 1999 datasets, to about 0.13–0.18 for the 2001–2005 datasets. The alternating decreasing and increasing latent mean changes in school disorder relative to changes in secure building over the years, as seen in the SMM data (see Table 2), considered along with (a) the much lower path values (association) from secure building to school disorder for the 2001 and 2003 SEM models, and (b) the leveling off of school violence during those years, suggests that in the context of day-to-day violence and disorder in schools, secure building strategies may not be well linked to changes in school violence, but are more of a reactive measure to political or other concerns. This finding signals a need for further research into this question. The lack of clear linkage of security-oriented building strategies to changes in school violence suggests that school-based teams should carefully review their safety and security plans and ensure that the procedures they employ are well thought out and supported by empirical research. Every security procedure should be part of a larger, well-integrated programmatic approach that as a whole meshes well together to promote a safe and positive school environment.

Of greater importance, the pattern of R ² values for the three alternative measurement models suggests that incivil (intimidation, bullying, hate language, feelings of rejection) offers better explanatory value than percrime (theft and attack) of students' fear/anxiety and avoidant behaviors, in the context of this model, based on 2001–2005 SCS data. This result provides moderately strong empirical corroboration to earlier reports by Boxer et al. (2003), Skiba et al., (2004), Ladd (2003), and others, and represent a red flag that necessitates more focused research on the role of incivility as a cause of students' level of fear/anxiety and avoidant behaviors in the context of school violence and disruption. Much of the research on school violence has focused on higher profile measures, but this is the first nationally representative school violence survey data structural analysis suggesting retargeting priorities toward addressing so-called low-level incivility. It is particularly noteworthy because the findings are based on data from the principal federal survey on school victimization, designed and executed according to established principles of complex survey methodology.

This convergence of evidence on lower level forms of violence and disruption should alert the education and allied clinical community at the local, state, and national level, and policy makers at the state and national levels to the need to (a) provide financial and other infrastructure supports to local school districts to help them respond more effectively to incivility in schools, (b) conduct more focused prevention research in this area, (c) invest in professional development for preservice teachers and allied educational staff, and (d) support colleges of education and allied discipline professional preparation programs to include more training in this area.

CONCLUSION

This research has demonstrated the importance of helping students understand the school's system of law, more thoroughly addressing incivility in schools, and careful planning of building security approaches as part of a larger school violence prevention approach. The research has several limitations. First, the structural modeling was dependent on using existing datasets that were not designed for this specific research. As such, the model used two measured variables linked to several constructs instead of generally recommended minimum of three. While not leading to any problems with the 1995 and 1999 SCS dataset analyses, this may have contributed to Heywood case problems with the 2001–2005 SCS datasets. Categorical ordered data were used and this can limit the utility of some parameter estimates in the model, where the researcher is better advised to think of the model path values more broadly as low, moderate, and high (Mayer, 2004; Mayer & Leone, 1999). As with much survey research, there were problems with missing data necessitating imputation procedures. Mayer and Leone (1999) used multiple levels of cross-validation in the earlier analyses, demonstrating that the imputation was not problematic. Problems with Heywood case results in the current analyses were partially remedied. While the nature and degree of the Heywood problems in this research were minor, they still limit the findings.

Changes in the survey questionnaires necessitated using data from the NCVS (from the identical subjects) for the percrime measured variable in the 2001–2005 analyses, pursuant to expert statistician recommendations, as previously mentioned. The normalized weights that were used in these analyses were drawn from the SCS dataset, not the NCVS, as there was no practical method for mixing these weights in this analysis. However, this cannot reasonably be expected to have created any substantive differences in model solutions. Prior unpublished SEM research by Mayer on earlier SCS datasets demonstrated almost identical model solutions between weighted, normalized weighted, and unweighted case data. Low R ² values for the school disorder construct in the 2001–2005 analyses—likely a statistical artifact as discussed above—limit the utility of the overall model, yet do not invalidate the consistent finding regarding the role of incivility based on comparison of alternative measurement Models 1, 2, and 3.

This analysis focused on a limited model of school violence and disorder, based on availability of national level data. Research tells us that school violence and disorder is a far more complex situation. There are general assumptions of changes over time, yet measurements were taken via discrete points in time with no case linkage over time, as the respondents changed with each administration of the SCS survey. However, as indicated by Firebaugh (1997), repeated surveys using the same questions, but different samples, may be used for analysis of trends over time, given comparable sample design (as was the case in this research), with specific advantages over panel surveys in representing change processes for the total group represented through the sample design. Finally, the central question of path directionality and influence between the secure building and school disorder constructs was untestable and could only be argued in terms of most favored interpretations, based on a convergence of data.

Future research needs to focus on several issues. The Mayer and Leone model should be extended by comparing schools that have implemented so-called zero-tolerance policies with schools using a more balanced disciplinary approach. Similar studies should be conducted, applying the model to data representing urban and nonurban schools. Several reports have noted different patterns of violence and disruption among poorer inner city schools. If possible, a larger series of models that examine a reciprocal path between secure building and school disorder should be researched, using a nested approach. Additionally, efforts should be made to study changes over time, linking specific schools over time within the respective datasets.

This study has established the critical need to craft new survey instrumentation that will facilitate and support a more comprehensive set of research studies at the national level to further investigate the connections between incivility and related low-level behaviors in schools, and students' resultant fear, anxiety, and avoidant behaviors. While the current SCS questionnaire does an admirable job of tapping into many key aspects of school violence and disruption, it is too limited to support the next generation of analysis of incivility and its sequelae. The Departments of Education and Justice have demonstrated a long history of providing vital support for similar lines of research and hopefully will recognize and respond to the needs in this area.

As more is learned about processes associated with school violence and disorder, research returns to a basic issue. Does the nation have the resolve and wherewithal to tackle these difficult problems? With substantially increased commitment to provide supports for schools to implement evidence-based solutions, increased funding for targeted research as discussed above, professional development opportunities for teachers and members of allied clinical professions, and enhancement of professional training programs that address school violence prevention, communities can look forward to safer and more productive schools. Schools should see visible signs of improvements as a result of these efforts and more students will be able to feel safe and secure, focusing more on understanding the lesson than worrying about bad things that might happen.

Notes

1. Prior research with the 1995 SCS data also uncovered problems with missing and/or indeterminate data with regard to the drugpres and gangpres indicator variables (see Mayer, 2001, for a detailed discussion). Similar problems existed with the other three datasets. Data were imputed with a saturated regression model (using all other measured variables in the model) that included adding a random noise term.

2. The SEM analysis of the 2001, 2003, and 2005 data using the three alternate measurement models ran into problems with negative variance—Heywood Case—for the error term associated with one of the measured variable, ordpers. The SEM software accordingly set the variance at the lower permissible bound—zero. Causes of a Heywood Case can include small samples, only two indicators for a latent variable, outlier values, empirical under-identification, misspecification of model, and bad start values (Dunn, Everitt, & Pickles, 1993; Kline, 2005). In such cases, the methodological literature suggests setting the indicator error term to have a small positive value or using a forced value drawn from prior empirical research of comparable data (Chen, Bollen, Paxton, Curran, & Kirby, 2001). The problem was not responsive to such approaches in the 2001 SCS dataset. The Heywood situation was resolved in the 2003 and 2005 SEM analysis, and also in the SMM analysis of the 2001, 2003, and 2005 SCS datasets, by setting an error term value based on prior research with the SCS dataset.

3. SEM analysis generally requires three or more indicator variables for each latent variable to avoid specification and related solution convergence problems (Bentler & Chou, 1987). This was not possible in this analysis, due to limitations of the SCS dataset. Prior analysis of the 1995 SCS dataset with the same structural model showed no estimation problems. However, the existence of competing plausible models is greatly reduced with more than two measured variables per construct and strong fit.

4. The SCS survey used a stratified multistage cluster sample design, which can introduce problems due to the design effect (Deff), “the ratio of the variance of a statistic under a particular sample design, to the variance of that statistic under simple random sampling for a sample of equal size” (Mayer, 2004, p. 140). Design effects can lead to incorrect inferences. Formal and ad hoc procedures exist to address design effect challenges (Kish, 1965; Kish & Frankel, 1974; Muthen & Satorra, 1995; Stapleton, 2006). However, if significance tests (z-scores) of path parameters in SEM analyses fall in a high range (e.g., above 4 or so) design effect problems would be rendered moot, as was the case in this analysis.

5. Chi-square statistics are used with SEM models in a hypothesis testing framework, whereas fit indices such as the Comparative Fit Index (CFI) and root mean square error of approximation (RMSEA) are used to determine goodness of fit. Chi-square tests can be problematic because the chi-square distribution is sensitive to large sample size and trivial misspecification of a model with a large N (as is the case in this analysis) will result in rejecting the model, because the p-values will be less than .05 in most cases (Hu & Bentler, 1995, p. 96; Loehlin, 1992, p. 71; Raykov & Widaman, 1995, p. 294). Thus, this analysis only addresses p-values associated with model path parameters and not omnibus chi-square tests of the model. As is customary in SEM research, greater attention is given to fit indices when considering model viability.