Self-protection against crime: what do schools do?

ABSTRACT

Most economic research about self-protection against crime concentrates on self-protection by individuals, households, and stores – private economic agents. We know much less about self-protection in public economic settings, such as by schools; this article takes a step towards learning more. An economic agent who self-protects benefits by reducing vulnerability to crime but incurs self-protection costs whether a victimization occurs or not; should a crime occur, the agent further bears the cost of the victimization itself. The agent, a school administrator in this application, must determine the optimal level of self-protection within this environment. Empirical results obtained using data from the 2004 and 2006 School Survey on Crime and Safety (SSOCS) show that schools self-protect (use professional security personnel) much in line with theoretical predictions. Among other findings, schools located in larger cities and that have a larger and older student body self-protect more prevalently than other schools, while schools with more academically able students self-protect less than schools with less productive student inputs.

KEYWORDS:

• Crime
• self-protection
• schools
• guards

JEL CLASSIFICATION:

• K42
• I20

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

¹ See, for example, seminal research by Komesar (1973), Bartel (1975), Clotfelter (1977), and McDonald and Balkin (1983), and more recent work by Vollaard and Koning (2009), Allen (2013), and Zimmerman (2014).

² The White House Council on Women and Girls (2014) documents these measures.

³ The implied production function e(s, x) in this setting serves the same analytical purpose as educational production functions featured implicitly or explicitly in earlier economic research on the educational process. For example, Unnever, Kerckhoff, and Robinson (2000) and Barnett et al. (2002) emphasize how schools transform inputs of faculty time and other resources into measurable student outcomes. Foreman-Peck and Foreman-Peck (2006) highlight inputs of technology and that ‘pupils of a certain educational standard,’ i.e., student quality, constitute inputs into the educational process. Andrews, Duncombe, and Yinger (2002) develop an explicit model that also allows for variation in educational input quality.

⁴ The approach taken here effectively assumes that both the hypothetical school administrator and governance body share an interest in the production of safe education, which seems reasonable given that both ultimately have accountability to students and parents. Even allowing for disparate career concerns (e.g. long-run employment as a school principal or elected school official), each has an incentive to facilitate safe education in the short run given the presumed importance of reputation along these lines to future employers or electorates.

⁵ Noonan and Vavra (2007) observe that most illegalities occurring in U.S. schools stem from ‘everyday school interactions.’ From 2000 to 2004, 38% of school offenders were 13–15 years of age and 30.7% were 16–18 years of age; where offender age was known, offenders aged 18 years or younger were 4.5 times more likely to be involved in school crime than older individuals. These facts and the finding that the vast majority of school crimes involved offenders and victims who knew each other point to students as the most common school offenders.

⁶ If we were modelling self-protection by students, as in Allen (2013), we would likely make the same assumption, because the most immediate threat of crime emanates from the same source: fellow students. The assumption bears similarity to the common assumption within the economic model of crime that enhanced licit human capital skills encourage allocation of time more to licit than to illicit activities.

⁷ More explicitly, Z_ss = ∂Z_s/∂s = -(∂²p/∂s²)k + ∂²e/∂s^{2 –} ∂²c/∂s². The first term is negative because ∂²p/∂s² > 0: second-order variation in self-protection reinforces the negative first-order effect of self-protection on the probability of victimization. The second term is negative under the assumption of diminishing marginal productivity of self-protective inputs. The final term is positive because second-order variation in self-protection reinforces the positive marginal cost of self-protection (∂c/∂s). The sign pattern therefore renders Z_ss <0.

⁸ Within this model one can also formally establish that exogenously higher costs of self-protection (independent of the victimization loss) discourage the activity, ceteris paribus. However, because the SSOCS data set does not allow observation of the costs incurred by schools, thus preventing an empirical analysis of this factor, I do not develop this connection in this section.

⁹ As with any data set based on survey responses, I assume random measurement error associated with self-reporting by subjects. The NCES nevertheless took several documented steps aimed to minimizing such error. As a general matter, the NCES instructed each school to have the survey completed by the ‘person most knowledgeable about school crime and policies to provide a safe environment,’ making clear that the SSOCS represents the Department of Education’s primary source of school-level data on crime and safety and an essential resource used by educational leaders and policy-makers. The NCES furthermore communicated with sample schools in advance of sending the survey to enhance their readiness to complete it, engaged in follow-up communication when responses appeared incorrect or incomplete, and informed state- and district-level school officials in writing when schools within their oversight were being surveyed – actions intended to enhance accurate participation by schools.

¹⁰ SSOCS data capture comparable information about the level of criminality at the school’s location, allowing the creation of a similar set of indicator variables. Empirical results obtained using these variables yielded qualitatively similar results as those estimated using the student-neighbourhood variables, thus only the latter are formally summarized here.

¹¹ Education economists have frequently used these and similar variables as measures of educational inputs, as in the present study, and occasionally as measures of educational output. Unnever, Kerckhoff, and Robinson (2000) incorporate student ability as an input in modelling educational production empirically. Rubenstein et al. (2007) document increased public funding for schools with higher LEP percentages, as limited language skills impact the ability of schools to produce educational output. Jones, Toma, and Zimmer (2008) observe that better attendance empirically improves learning and quantifiable academic preparation, i.e. educational output.

¹² To aid interpretation in multivariate statistical settings, these variables will be scaled by a factor of 10; one-unit variation will thus imply a ten-percentage-point change in each variable.

¹³ The crime type variables are summarized in unscaled form in Table 1 but will be scaled by 10 in regression settings; unitary variation in each variable will thus imply a ten-unit change in the quantity of each type of illegality sustained.

¹⁴ The NCES purposely administered the SSOCS at the end of each associated academic year (March–May), allowing reasonable observation of schools’ established crime and security data. Collection of the data at the beginning or in the middle of a school year would have clouded the direction of causality: increased guard use might plausibly have lowered school-level crime. However, the school crime and guard use variables exhibit positive, not negative, correlations for both school years represented in the two-period panel, inconsistent with an inverse relationship.

¹⁵ In their study of urban-suburban student performance gaps, Sandy and Duncan (2010) observed that urban students face greater risks to their overall safety and participate in more risk-taking behaviours, factors associated with lesser academic achievement. In the present context, such factors may make urban students relatively more likely to commit illegal activity or become victims of illegal activity, potentially motivating more self-protection by urban schools.

¹⁶ Along these lines, Jones, Toma, and Zimmer (2008) observe that the older age of high school students makes them more autonomous and their behaviour more difficult for school officials to manage.

¹⁷ Count-data panel models (Poisson and negative binomial regressions) estimated for this study did not universally achieve convergence. To aid comparisons across models, I therefore report findings from GLS models.

¹⁸ Although the SSOCS data set does not indicate the specific grade levels that coexist in the sample combined schools, this pattern of results and others that follow suggest that these schools primarily combine middle and secondary school students.

¹⁹ Personal Crimes averages 7 for the full sample, 1 for schools with Size <300, 2 for Size 300–499, 5 for Size 500–999, and 14 for Size ≥1,000, rounding each mean to the nearest whole number.

²⁰ To aid computation and interpretation of the predictions in relation to the continuous educational input variables, these variables are separated into categorical variables indicating whether they lie above or below their respective means, as detailed in Tables 5 and 6.