School-Based Violent Victimization in Turkey: An Examination of the Cross-National Generality of Lifestyle-Routine Activities and Self-Control Theories

ABSTRACT

The authors examined victimization among Turkish school students as a function of individual lifestyles and routine activities, perceived school guardianship/control, and low self-control. In doing so, they aimed to provide a much-needed explanatory test of school victimization in Turkey while also offering an important test of the cross-cultural generalizability of self-control and opportunity-based theories of victimization. Logistic regression models of violent victimization were estimated using a subsample of over 900 Turkish school students. Regression coefficients were estimated for 20 datasets generated through a multivariate sequential imputation technique, with results then pooled. Lifestyle measures associated with school-based victimization included in-school delinquency, delinquent self-cutting, gang membership, and number of gang friends. Perceived school guardianship/control was also related to victimization, as was low self-control. The authors found little evidence that the effects of low self-control were mediated or moderated by lifestyle characteristics or perceived school security. Findings suggest that the propositions of lifestyle-routine activities and self-control theories regarding victimization risk can largely be generalized to Turkish high school students. Findings imply that school-based victimization prevention in Turkey should target individual-level criminogenic traits and lifestyles as well as risky environmental school characteristics.

KEYWORDS:

• delinquency in Turkey
• lifestyle-routine activities
• low self-control
• school-based victimization

Notes

1. Students indicated that 45.1% of bullying incidents occurred in the classroom or school corridors, followed by on the way to school and outside school (24.1%), in the playground (14.4%), and other places including school canteen and sport centers (16.4%).

2. Inadequate funding prohibited using all 81 provinces for sampling purposes.

3. Smoking, drinking, and drug use were originally measured by three separate items in the survey. Due to the infrequent nature of any of these behaviors, they were summed and then dichotomized into any/none substance use for analysis purposes.

4. Two of the three authors of this study are former officers in the Turkish National Police.

5. Another survey question asked, “How easy it is to bring in weapon/gun to school?” However, this item was highly correlated (more than .7) with “How easy it is to bring in knife to school?” Further, there were many more missing cases association with the question pertaining to gun carrying. Thus, we used the question about knife carrying, and excluded the other one from the analysis in order to retain more cases.

6. Before MI analysis, we first examined the pattern of missing values in the data to determine whether data were missing completely at random (MCAR). Little’s MCAR test was used for such purposes (Little, 1988). The test computes maximum likelihood estimates of the means of study variables (we used expectation maximization), and evaluates the null hypothesis that the means are the same across missing data patterns (Howell, 2007). A significant result indicates that the missing data are not MCAR. In our analysis, Little’s test was nonsignificant, indicating that the data are MCAR and there is no significant difference between those cases with missing data and those without (p < .05). However, because Little’s test was very close to significance, we felt comfortable using the more conservative approach to handling missing data—multiple imputation—as opposed to listwise deletion. We also wanted to retain more cases in our sample, which has less than 950 cases before listwise deletion. SPSS 22 is used to run the Little’s MCAR test. More detailed test results are available from the first author upon request.

7. During this procedure, missing values of a continuous variable with a restricted range are filled in using a truncated regression imputation method. The upper and lower limits of metric variables are also defined before the imputation so that imputed values fall within the range of each variable. Briefly, we illustrate the sequential imputation technique (using chained equations and truncated regression) with 3 variables: v1 (binary), v2 (limited metric), and v3 (limited metric). In this scenario, we impute missing values for v1 using a logistic regression (v1 regressed on v2 and v3), impute missing values for v2 using a truncated regression of v2 on v1 and v3, and impute missing values for v3 using a truncated regression of v3 on v1 and v2. Such a chain of equations is created for all variables with missing values.

8. The imputation procedure assumes that the missing data are missing at random, which means that the probability that a value is missing depends only on observed values and not on unobserved values (Schafer & Graham, 2002).

9. We estimated an unconditional Bernoulli model in HLM software, and found that the p value for the cross-school variation was found to be .07. The small sample size at level 2 (school level), in conjunction with the fact that all schools were drawn from one area of Turkey, undoubtedly affects the ability to observe significant cross-school differences.

10. We did not report a model examining the potential indirect effects of LSC through perceived school guardianship since there is no theoretical reason to expect LSC to influence school environment.

11. Note that Akaike information criterion (AIC) and −2 log likelihood values are provided for all models in addition to pseudo R² (Nagelkerke) values. While AIC or −2 log likelihood values are not meaningful themselves, they are meaningful when compared across models, as smaller values for these statistics are preferred. In other words, better models will have smaller AIC or −2 log likelihood values. In Table 2, we present the average value of each fit statistics obtained from the logistic regression analysis of 20 imputed datasets. Comparison of AIC and −2 log likelihood values across the models in Table 2 presents a similar story as does the comparison of R² values in terms of model fit.

12. Appendix Table 2 indicates that moderation effect for gang membership was significant nine times at p < .05 and 15 times at p < .10 across the 20 datasets created through chained imputation. Interaction with weapon control was significant five times at p < .05 and 12 times at p < .10 across the 20 datasets.