Unintentional injuries are the leading cause of death and hospitalization among adolescents in industrialized countries.1 Since children and adolescents spend a great deal of time in school, the school environment is a highly relevant setting for unintentional injuries.2 In Germany, according to the German Social Accident Insurance (DGUV), more than 1.2 million school injuries were recorded in 2019. This translates into 66 injuries per 1000 students.3 Although only a very small proportion leads to death or permanent disability, this figure clearly underlines that unintentional injuries in the school environment pose a serious public health problem. Therefore, it is worthwhile to determine factors that are causally related to them.

Previous research on school injuries almost exclusively relies on official or process data from insurances or hospitals. This data suggests, for instance, that the rate of school injuries is higher in urban and larger schools while on the individual level, boys were consistently found to be more prone to school injuries than girls.4-12 Since process data only includes cases for which the event of interest (the injury) has occurred, its informative value is, however, limited.

Data which includes injured as well as noninjured students usually stems from surveys. Yet, there are only a few empirical studies with most of them not explicitly focusing on school injuries but on unintentional injuries of adolescents in general.13-19 Moreover, the vast majority of these surveys only gather data at one single point in time. Thus, the relative risk of, for instance, physical activity for school injuries can be estimated only by comparing the injury-rates of physical active and inactive students while controlling for other observed variables. Physical active and inactive students may, however, also differ with regard to other, unobserved variables that may affect both, the independent variable as well as the probability to suffer school injuries. Consequently, coefficients that were estimated based on cross-sectional data are prone to bias due to unobserved heterogeneity.

The most promising approach to approximate causal effects on school injuries is to collect panel data and to calculate coefficients based on within-estimators like in fixed-effects regression models.20, 21 This is the main objective of our contribution. More precisely, we aim to approximate the causal effects of variables for which past research found correlations with unintentional (school) injuries. These variables are risk-seeking behavior, mental health problems, physical activity, and exposure to bullying.13, 14, 18, 19, 22-29

METHODS Participants

In our survey Health Behavior and Injuries in School Age, we collected data from students visiting public secondary schools in 14 German federal states. Initially, we surveyed 10,621 students in fifth grade (ages 10-12) and followed the same students by surveying them repeatedly in 5 consecutive, annual panel waves.

Before the first survey in the school year 2014/2015, a stratified random sample was drawn among public secondary schools in Germany. In the first recruitment stage in 2014, we contacted 854 schools in 11 federal states via mail and asked the principals to take part in the study. One hundred fifty-six schools agreed to participate (participation rate: 18%). Since students in 3 German federal states do not enter secondary schools before seventh grade, we carried out a second recruitment stage in 2016 (before the third panel wave). This time, we contacted 120 schools with 17 taking part in the study (participation rate: 14%). In each participating school, we aimed to survey all students in the respective grade.

For our analyses, we relied on panel waves 2 to 5. Thus, we excluded data from the first panel wave because (1) some reported injuries had already occurred in primary school and (2) some of the independent variables of interest were surveyed only from the second wave onwards. We also excluded data from the last (sixth) panel wave, since we had to prematurely stop data collection due to school closures caused by the coronavirus. Students from 2 German federal states (Bavaria and Hamburg) were not included in our sample as we did not receive the permission to carry out the study in the respective schools by the Ministries of Education.

Instruments: Dependent Variable

In our study, we focused on self-reported injuries that happened in the school environment. Before the students were asked about their school injuries, interviewers read out a standardized instruction text dealing with the key criteria of school injuries. First, school injuries must have occurred in a school-related setting, like on the schoolyard, during physical education, in the school building or on students' way to school. Interviewers particularly stressed that injuries students suffered in leisure time or at home do not count as school injuries. Second, injuries must have occurred after the last survey with interviewers explicitly stating the relevant recall period by naming month and year. Third, the instruction text made clear that we were only interested in severe injuries that required medical treatment. Furthermore, it was also stated that diseases (like the flu, or tonsillitis) do not count as injuries even if they required medical treatment.

After these instructions, students were asked how many school injuries they suffered since the last survey or—if they participated for the first time—during the last 12 months. The recall period for both variants was the same since our interviewers visited the participating schools the same month in each panel wave. If students answered to have suffered one or more injuries, they were asked additional questions about this injury or (in the case of several injuries) about the most severe injury. These questions aimed at differentiating school injuries with regard to the setting in which they happened. Moreover, they were used to check whether the reported injury fitted our criteria of school injuries regarding location (occurrence in the school environment and not at home or in leisure time), time (during the last 12 months or since the last survey), and degree of severity (injury must have required medical treatment).

For our analyses, we created a dummy variable coded 0 for students who did not suffer any school injuries since the last survey and 1 for students who reported at least one school injury during that time. Since the effects of our independent variables may differ for the various types of school injuries, we distinguished between (1) school injuries that occurred during physical education and (2) school injuries that occurred on the schoolyard or in the school building. Thus, we excluded certain settings in which school injuries can occur (eg, on students' way to school) mainly due to the small number of cases. However, the 2 types of school injuries accounted for 81% of all reported school injuries in the 4 panel waves.

Instruments: Independent Variables

In our analyses, we included the following characteristics as independent variables measured in the same way in all 4 panel waves:

Risk-seeking behavior: For assessing risk-seeking behavior, we relied on students' evaluation of the statements “I like doing dangerous things” and “I enjoy new and exciting experiences even if they are sometimes dangerous or threatening.” Students were asked to rate each statement on a 5-point scale ranging from “does not apply at all” to “fully applies.” The values for Cronbach's Alpha for these items ranged from .79 (wave 2) to .85 (wave 5), thus both items were transformed into a summative index ranging from 0 to 4, with higher values indicating a higher risk-seeking behavior.

Mental health problems: Information on students' mental health problems is based on 8 items. Students were asked on how many days during the last week they (1) were irritated and in a bad mood, (2) felt fit and comfortable (coded reversely), (3) were full of energy (coded reversely), (4) felt sad, (5) felt lonely, (6) slept badly, (7) had problems to concentrate, and (8) felt unhappy and depressed. To record students' answers, a 5-point scale was used, ranging from “not at all” to “every day.” Alpha values were between .80 (wave 2) and .85 (wave 5). Again, a summative index from 0 to 4 was constructed with higher values indicating more mental health problems.

Physical activity: For assessing students' physical activity, they were asked how many hours they spend each week for (a) club sports and (b) leisure sports. We integrated both variables in our statistical models in their original forms (number of hours per week).

Exposure to bullying: Students were asked to assess how often during the last 12 months they were exposed to bullying from other students. The 5 answering options were “not at all,” “once or twice,” “twice or three times per month,” “about once per week,” and “several times per week.” We recoded this variable to range from 0 (not at all) to 4 (several times per week). Although the distances between the scale points vary considerably, we integrated this variable in the original ordinal-scaled version in our statistical models. However, we also tested the effect by relying on other recoding procedures. The question was introduced with a short explanation of the term bullying, following the criteria proposed by Olweus.30

In the first step of our analyses, we estimated the effect of the following time-invariant characteristics:

Sex: A dummy variable coded 1 for girls and 0 for boys.

Type of school: A dummy variable coded 1 for students visiting the most advanced school type in Germany (Gymnasium) and 0 for students from all other types of school.

Urbanity: A dummy variable coded 1 for students visiting urban schools (the school is located in a city with at least 100,000 inhabitants) and 0 for students visiting schools in smaller villages.

Region: A dummy variable coded 1 if the school was located in former East Germany and 0 if the school was located in former West Germany.

School size: A dummy variable indicating if 100 or more students were taught in fifth grade (1) or less than 100 (0).

Procedure

The study relied on a computer-assisted self-administered classroom survey. Each student that was willing to participate and had a parental informed consent form received a tablet device to answer the questionnaire by him- or herself during a (regular) school period (of 45 minutes). However, trained interviewers were also present in the classroom to introduce the survey, read out the instruction text, explain the handling of the devices and answer students' questions.

Data Analysis

In a first step, we analyzed the effects of our independent variables as if data came from 4 cross-sectional surveys. More precisely, we used logistic multilevel regression models to estimate the effects of our independent variables on school injuries separately for the 4 panel waves. In these models, coefficients are calculated based on comparisons between the surveyed students and may therefore be biased due to unobserved heterogeneity. To alleviate this problem, we included sex, type of school, urbanity, region, and school size as controls. Moreover, since the surveyed students are nested in schools we relied on multilevel models to account for the hierarchical structure of the data.

To examine the extent of bias and to approach the causal effects of our independent variables, we ran fixed-effects logistic regression models in a second step. These models take full advantage of the panel structure and only take into account the variation within individuals. Thus, coefficients of the fixed-effects models represent the effect of a 1 unit change within students on the probability to suffer school injuries.

Since fixed-effects regression models require changes in the variables of interest over time within individuals, it is not possible to estimate the effects of characteristics that do not vary within people over time, like sex or type of school. However, what appears to be a downside of this approach is its real strength. Since all (observed and unobserved) person-specific characteristics are included in the constant (the fixed part), regression coefficients of fixed-effects models do not suffer from bias due to unobserved heterogeneity. To further exclude biased coefficients due to age effects, we also integrated time in our model by including dummy variables for the different panel waves.

Data analysis was performed with Stata/MP 15 using the xtmelogit and the xtlogit command with the fe (fixed effects) option.

RESULTS

We start this section by briefly describing our sample (see Table 1). With regard to the time-varying characteristics, mental health problems in students increased over time while physical activities in leisure time and exposure to bullying reduced. Moreover, the distribution of the time-invariant characteristics reflects that students visiting the highest school type in Germany were overrepresented in our sample since official data suggests that only 34% of all students in secondary schools in Germany are taught in Gymnasien.31 Finally, the proportion of students who reported having experienced injuries on the schoolyard or in the school building reduced remarkably while the proportion of students who suffered injuries during physical education remained rather stable.

Table 1. Students' Characteristics %/Mean (SD) Wave 2 Wave 3 Wave 4 Wave 5 Time-varying characteristics Risk-seeking behavior 1.71 (1.13) 1.81 (1.11) 1.88 (1.09) 1.89 (1.06) Mental health problems 0.91 (0.65) 1.02 (0.72) 1.15 (0.78) 1.22 (0.78) Hours per week: club sports 3.16 (3.11) 3.31 (3.42) 3.24 (3.52) 3.10 (3.59) Hours per week: leisure sports 4.03 (3.83) 3.95 (3.91) 3.76 (3.83) 3.64 (3.74) Exposure to bullying 0.53 (1.01) 0.47 (0.97) 0.43 (0.93) 0.37 (0.86) Age 12.47 (0.58) 13.47 (0.60) 14.49 (0.63) 15.51 (0.66) Time-invariant characteristics % Female students 51.6 52.5 53.0 53.0 % Students in Gymnasien 56.1 57.7 56.6 57.3 % Students in urban schools 23.1 25.6 22.6 25.0 % Students in East German schools 8.3 12.6 12.2 13.0 % Students in large schools 77.0 77.0 77.3 75.9 Dependent variables % Students: injury during phys. educ. 10.2 11.5 11.6 10.2 % Students: injury on yard/in building 10.0 6.9 4.8 3.0 Number of cases 10,085 10,018 9120 8426

Tables 2 and 3 display the results of the multilevel logistic regression models separately for each panel wave and for school injuries occurred during physical education and those that took place on the schoolyard or in the school building. These models also reflect the distribution of the variance of school injuries on the 2 analytical levels. Here, we can observe that only a small portion of the variance was attributed to the school level. Thus, differences between the surveyed students rather than between the schools were responsible for differences in the susceptibility to school injuries.

Table 2. Effects on School Injuries Occurred During Physical Education in Multilevel Logistic Regression Models Wave 2 Wave 3 Wave 4 Wave 5 b SE p OR b SE p OR b SE p OR b SE p OR Time-invariant characteristics Female students .18 .07 * 1.19 .24 .07 ** 1.27 .11 .07 1.11 .02 .08 1.02 Students in Gymnasien −.01 .09 .99 .04 .09 1.05 .01 .09 1.01 .04 .10 1.04 Students in urban schools −.26 .11 * .77 −.02 .10 .98 −.17 .11 .84 −.04 .11 .96 Students in East German schools .36 .15 * 1.44 .17 .13 1.19 .30 .12 * 1.36 .23 .13 1.26 Students in large schools −.11 .10 .89 .08 .10 1.08 −.04 .10 .96 −.14 .11 .87 Time-varying characteristics Risk-seeking behavior .12 .03 *** 1.13 .17 .03 *** 1.18 .16 .03 *** 1.17 .07 .04 * 1.08 Mental health problems .32 .05 *** 1.38 .16 .05 ** 1.18 .25 .05 *** 1.28 .22 .05 *** 1.25 Hours per week: club sports .07 .01 *** 1.07 .05 .01 *** 1.05 .05 .01 *** 1.05 .03 .01 ** 1.03 Hours per week: leisure sports .04 .01 *** 1.04 .02 .01 ** 1.02 .03 .01 ** 1.03 .03 .01 ** 1.03 Exposure to bullying −.00 .04 1.00 .08 .03 * 1.08 .07 .04 * 1.08 .03 .04 1.03 Constant −3.13 −3.09 −3.00 −2.83 Proportion of variance (school level) 2.4 2.5 1.8 2.1 N (students) 9211 9302 8575 8055 N (schools) 138 145 133 124 Table 3. Effects on School Injuries Occurred on the Schoolyard or in the School Building in Multilevel Logistic Regression Models Wave 2 Wave 3 Wave 4 Wave 5 b SE p OR b SE p OR b SE p OR b SE p OR Time-invariant characteristics Female students −.39 .08 *** .68 −.43 .09 *** .65 −.55 .11 *** .58 −.20 .14 .82 Students in Gymnasien −.33 .09 ** .72 −.36 .11 ** .70 −.43 .13 ** .65 −.46 .18 * .63 Students in urban schools .13 .11 1.14 .14 .12 1.16 .20 .14 1.23 .14 .20 1.15 Students in East German schools .02 .02 1.02 −.11 .17 .90 .01 .19 1.02 −.33 .28 .72 Students in large schools .07 .11 1.07 −.04 .12 .97 .15 .14 1.16 .09 .20 1.10 Time-varying characteristics Risk-seeking behavior .14 .03 *** 1.16 .22 .04 *** 1.25 .23 .05 *** 1.26 .16 .06 * 1.17 Mental health problems .24 .05 *** 1.27 .32 .06 *** 1.38 .32 .07 *** 1.38 .30 .09 ** 1.35 Hours per week: club sports .02 .01 * 1.02 .03 .01 *