Abstract
Bullying is a widespread social phenomenon that is thought to have detrimental effects on life outcomes. This paper investigates the link between bullying and later school performance. We rely on rich survey and register-based data for children born in a region of Denmark during 1990–92, which allows us to carefully consider possible confounders including psychological factors. We implement an IV strategy inspired by Carrell and Hoekstra (2010) where we instrument victim status with the proportion of peers from troubled homes in one’s classroom. We show that bullied children suffer in terms of GPA and effects tend to increase with severity.
I. Introduction
A student is characterized as being bullied or victimized when he or she is exposed, repeatedly and over time, to negative actions on the part of one or more other students (Olweus 1993). This paper investigates the determinants and potential effects of bullying in elementary school on academic achievement.
Bullying is a serious and widespread phenomenon: 27 percent of the Danish children that we analyze are reported by their parents and/or teacher to be victims of bullying (similar numbers are reported by, for instance, Brown and Taylor 2008 for Britain, Nordhagen et al. 2005 for Denmark, and Saufler and Gagne 2000 and Centers for Disease Control 2010 for the United States). From an economic point of view, such common negative actions may be extremely costly, not only in terms of immediate individual welfare but also in terms of longer run consequences. Despite this, very little research documents the impact of bullying on economic outcomes. An exception is the paper by Brown and Taylor (2008) that uses regression-based techniques to show that bullying is associated with reduced educational attainment and wages. We know of no other papers studying the link between bullying and long-term economic outcomes.
Our paper contributes to this very small literature by using survey-and register-based data on children born in a region of Denmark during 1990–92 to investigate the determinants and potential effects of bullying at ages 10–12 on grade point average (GPA) after Grade 9.
Our data include exceptionally rich register-and survey-based information on physical and mental health as well as socio-emotional and psychological issues measured prior to exposure to bullying. The survey data also present a unique opportunity to define bullying status as both the teacher and parents answered whether the child was a bully or a victim of bullying. Because we are interested in school bullying, the teacher’s perception is crucial in order to obtain a truthful picture of the interactions among peers. At the same time, it would not be sufficient to restrict ourselves to the teachers’ responses as they do not observe the child for the entire school day, and they do not have the confidence of the child to the same extent as the child’s parents.
We implement an instrumental variables strategy inspired by Carrell and Hoekstra (2010) that finds that domestic violence affects not only children in the family but also their peers in the classroom. Here we exploit administrative data on parents’ criminal history including convictions for violent crime, property crime, or any other non-traffic-related crime. We document that criminal behavior of the parents of one child increases the likelihood that other children in the classroom are bullied. As such, we shed light on a channel through which the results of Carrell and Hoekstra (2010) may operate. We show that our findings are robust to using more standard sibling comparisons. We acknowledge that the problem of nonrandom selection of victims is particularly difficult to control and stress that caution should be made when interpreting our results.
We see that the quality of the family environment, as well as individual child characteristics, is predictive of bullying status. We find that individual characteristics such as poor early mental health, indicators of hyperactive behavior, and physical appearance are important drivers of victimization.
Our results suggest that being bullied significantly lowers ninth grade GPA. The effects tend to increase with the severity of bullying. Robustness analyses suggest that signs are robust though prevalence and magnitudes are sensitive to which informant is used to report on victimization. Our results suggest that teacher-reported victimization is, on average, more severe than parent-reported victimization.
The remainder of the paper unfolds as follows: Section II surveys the literature on bullying and its determinants and consequences. Section III discusses the institutional context and the available data, while Section IV presents baseline ordinary least squares (OLS) regressions and Section V our main empirical strategy and associated results. Section VI shows robustness analyses and investigates heterogeneity while Section VII concludes.
II. Background
As discussed above, bullying is the exposure to repeated negative actions over time on the part of one or more students; Olweus (1993, 1997). Negative actions are intentional attempts to injure or cause discomfort in others. Examples are physical contact, verbal insults, rumors, and intentional exclusion. For the actions to qualify as bullying, an asymmetric power relationship between the bully and the victim should also exist such that the bullied child has difficulties defending him or herself against the perpetrator. The seminal works by Olweus (1993, 1997) describe two victim types: passive and provocative. The typical passive victim is cautious, sensitive, and quiet, and reacts by crying. Boys in this category are generally physically weaker than other boys. The provocative victim, on the other hand, has problems with concentration, causes irritation and tension, and is often hyperactive.
A. Why Would Bullying Affect Future Outcomes?
Psychological explanations of why bullying affects future outcomes distinguish between the effects of being a victim and being a perpetrator of bullying. Victimization is closely related to harassment and violence (Patchin and Hinduja 2011), which are known to have unfortunate long-run consequences, although causal relationships are inherently difficult to establish (Currie and Tekin 2012). The negative long-run consequences may be interpreted in the framework of general strain theory (Agnew 1992), which argues that individuals who experience a strain (such as bullying) may produce negative emotions such as anger, frustration, depression, or anxiety, which may lead to a corrective action in terms of wrongdoing, self-harm, suicide, etc. Ouellet-Morin et al. (2011) shows that bullied children had lower and longer-lasting cortisol response to stress than the comparison group, suggesting that bullying invokes biological changes in victims with potential long-lasting impacts.
Some theories would predict that perpetrators also may be affected by bullying. However, we make no attempt to identify the potential effect of being a perpetrator because the case for this analysis is plausibly weaker.
The mentioned theories may be reconciled with the economic theories of life-cycle skill formation (Heckman 2008). In economics, it has been shown that early investments not only have a large potential payoff, they are also efficient in the sense that an equity-efficiency tradeoff does not exist, which is the case for later investments. The reasons are that skills acquired in one period persist into future periods and that skills produced at one stage raise the productivity of investment at subsequent stages. Importantly, skills are multidimensional and are likely to complement each other. In this context, coping with victimization of bullying early in life directs resources away from investment in other skills. In addition, to the extent that bullying exerts a direct negative impact on self-esteem and other noncognitive skills as suggested above, educational and labor market success are also affected through this channel (Heckman 2008; Waddell 2006). The loss in terms of education, health, and lifetime earnings potential may be enormous if bullying is interpreted in this framework.
B. Prior Evidence About Childhood Bullying
In this section, we review the literature on predictors of being bullied in order to obtain a guideline for defining the conditioning set in our study of the potential effect of victimization on scholastic achievement.
Brown and Taylor (2008) is one of the few existing studies that actually investigates the link between bullying and educational attainment. It finds that strong predictors of being bullied at age 11 are: being a boy, having disabilities, unattractive physical appearance, personality traits, and number of schools attended.
Henningsen (2009) identifies the two main determinants of victimization as low family income and not feeling safe with one’s parents. However, parental education and divorce as well as more rare instances such as serious illness in the family, accidents, foster care, drug abuse, and sexual assault also correlate with victimization. Wolke et al. (2001) confirms that low socioeconomic status correlates positively with victimization and moreover finds that ethnic background/skin color is an important predictor.
A plausible hypothesis is that not only individual characteristics but also the institutional framework matters for the prevalence of bullying. However, Persson and Svensson (2010) finds no effects of class size on victimization. Obviously, school-based anti-bullying programs might also influence the prevalence of bullying. Farrington and Ttofi (2009) systematically reviews evaluations of such programs and finds that long, high-intensity interventions that emphasize teacher and parent training, among other things, effectively reduce bullying and victimization.
Based on the literature reviewed, the conditioning set in our study of the potential effect of victimization on educational achievement should preferably include socioeconomic variables such as gender, age, ethnic origin, family resources and strains, as well as individual characteristics such as personality traits, psychological factors, disabilities, physical appearance, and physical weakness/strength. Among institutional characteristics, the previous literature indicates that class size is of less importance while school and teacher characteristics or fixed effects should be included to account for antibullying prevention and related policies.
III. Institutional Context and Data
This section presents the institutional context within which we perform our analyses and gives a detailed discussion of data sources along with measures of bullying, the outcome, and the conditioning set.
A. Elementary School in Denmark
The vast majority of Danish children attend public elementary school (87 percent)1 and subsequently publicly subsidized after-school care (83 percent).2 After-school care most often takes place at an after-school club set up at schools with the idea that children have an integrated day (93 percent). The personnel may to a minor extent overlap with the personnel during the school day. However, after-school care may also take place at a recreation center detached from schools (7 percent).3 Thus school and afterschool care is by far the most important scene for social interactions between children.
In Grade 0, pupils are taught by a form teacher who is a trained pedagogue. From Grade 1 to Grade 9, pupils are taught by subject-specific teachers rather than form teachers, among which one or two teachers take on the responsibility as a class teacher. Concern for the social climate in class is the responsibility of the class teacher(s) while introduction of antibullying programs are most often school-based policies.
B. Data
The main data used in the analyses below stem from the Aarhus Birth Cohort (ABC). The data consist of initially 10,907 children born by 10,375 mothers in Aarhus, Denmark, during 1990–92. Of these, 525 women gave birth to more than one child during the period of observation, which we exploit in our robustness analysis. All pregnant women were eligible to participate in the survey and were recruited via tax-paid antenatal health services in their 14th gestational week,4 and 98 percent chose to participate. In 2001 (when the children were 9–11 years old) and again in 2002, the parents of the children were surveyed, and in 2002 also the teachers of the children were interviewed and asked to evaluate the children’s behavior. What is crucial for our purposes is that information about teacher and parent assessed incidents of bullying were provided. In addition, measures of socio-emotional and psychological issues are available, which is of major importance in our analysis of bullying.
We drop observations where we observe no information about victimization from either the parents or the teacher as well as observations with no classmates in our data set. This results in 4,490 observations. Finally, we do not observe ninth grade GPA for another 235 children who drop out or skip the exam. Our final sample thus consists of 4,255 children. Appendix 1 contains more information about attrition.
The survey data are augmented with a rich set of register-based information on 1) parents’ socioeconomic background, crime, and health status (level of education, labor market history, settlement patterns, income, prescription drug usage, somatic and psychiatric diagnoses from general hospitals, crime records)5 and 2) children’s early health outcomes including information about circumstances pertaining to the birth of the child, daily information on prescription drug usage, yearly information about hospital use and related diagnoses, type of childcare, the classroom they attended, and ninth grade test scores and yearly marks. We use the register data to strengthen our conditioning set and to construct the outcome measure as detailed below.
1. Bullying
In identifying bullying, we exploit the parent and teacher questionnaires conducted in 2001 (only parents) and 2002. Each supplies a rating of the extent to which the child is a victim of bullying and whether the child bullies other children.6
According to Olweus (1997), negative acts only qualify as bullying if they take place repeatedly, over time, and if the negative acts are intentional and the victim cannot defend him or herself (asymmetric power relationship). In the past decade bullying has received increasing attention in Danish society. Bullying policies have been introduced in school, the media has drawn attention to the problem at several occasions, and politicians have also increased focus on the matter. We therefore assume that the respondents have an appropriate understanding of the concept.7 Of course, we cannot be absolutely certain that the respondents employ the exact same definition as suggested by Olweus.
In our main analysis, we identify a child as a victim of bullying if either the teacher or the parents replied that the child is being bullied “to a small extent,” “to some extent,” or “to a large extent” in the 2001 questionnaire or “somewhat true” or “certainly true” in the 2002 questionnaires. Table 1 displays the bullying status of the children in our sample. Among the 4,255 children, 1,151 (27 percent) are identified as victims of bullying. This largely resembles the prevalence rates obtained in other studies based on self-reporting or parental reporting (see the Introduction). We also see that 20 percent of the victimization is reported to be severe (at least one of the informants state that the child is “bullied to a large extent/certainly true”) while 80 percent is reported as minor victimization (child is “bullied to a small/to some extent/somewhat true”).
Means of Selected Child Characteristics by Bullying Statusa
Of course, one might worry about measurement errors in this context, and Bertrand and Mullainathan (2001) discusses possible pitfalls associated with the use of subjective measures such as bullying.8 Individuals may, for example, answer on different scales; they may misreport due to social desirability; and they may report having attitudes, which are consistent with past behavior. One might also worry that an informant who has observed a change in the child’s behavior during a period of time that indicates that something troubling is going on in the child’s life (truancy, low scholastic performance or nightmares) may be more likely to report victimization. In a similar line of reasoning “victim mentality” may vary across children—what may be considered bullying by one child may be blown off by another child. All these mechanisms may create a spurious correlation between victimization and our outcome.
To address such measurement concerns we exploit that we have bullying information from two sources (teachers and parents) and explore to what extent our results are robust toward changing the definition of bullying to rely on one or the other source or both. We also investigate the consequences of distinguishing between the severities of the bullying experience.
Presumably, teachers and parents possess different sets of information about the child and the child’s behavior. Thus, we expect that exploiting both reporting sources will provide a more truthful picture of the extent of bullying. Although Oliver and Candappa (2003) finds that the majority of pupils would tell their mothers about the bullying episodes, we cannot rule out that some pupils would choose not to inform their parents because they are afraid that this would lead the parents to take action, which might increase victimization. If victimized children are negatively selected, we expect misclassification due to underreporting to cause a downward bias in our formal analysis of consequences of being bullied and this will likely be reduced when we rely also on a teacher’s report.
The correlation between the teacher’s and parent’s responses to whether the child is being bullied is 0.29. The parents in our sample are more likely to report their child a victim of bullying compared to the teachers (23 percent and 12 percent are reported to be bullied according to parents and teachers, respectively). These numbers emphasize the importance of having two informants as well as the importance of careful robustness checks.
The peer relations and the social interactions leading to victimization experiences may vary across gender as well as victim mentality (Espelage et al. 2000). Therefore, we study boys and girls separately in part of the robustness analyses.
2. Characteristics of Children and Parents
Means of selected characteristics of children and their parents by bullying status are shown in Tables 1–3. These variables also enter into our conditioning set in the formal analyses below. Except for psychosocial well-being, height, and minor physical handicaps at ages 9–11, all the child and parental characteristics shown here stem from administrative registers and are measured before the child starts school. Classroom fixed effects are based on the earliest possible classroom identifier (most often Grade 1) in order to avoid potential class and school mobility induced by early initiated victimization.
Means of Selected Characteristics of the Mother by Bullying Statusa
Means of Selected Characteristics of the Father by Bullying Statusa
As suggested by the literature, measures of the quality of family environment such as number of older siblings and parental divorce are predictive of victimization as is immigrant status. Similarly, poor early mental health (as indicated by prescription of antidepressants and a mental or behavioral diagnosis established before the age of seven) predicts bullying status at ages 10–129 as does a higher than average number of early emergency ward visits that may be indicative of hyperactive behavior; see Dalsgaard, Nielsen, and Simonsen (2013). Physical appearance has also been suggested as a driver of victimization. In line with this hypothesis, we see that minor physical handicaps such as impaired hearing, the wearing of glasses, and cross-eyedness are associated with victimization. The type of childcare before starting school also correlates with exposure to bullying.
Among the conditioning variables, we include four variables computed from a factor analysis based on items reflecting socio-emotional and psychological well-being. We obtain four factors (zero mean and unit standard deviation); hyperactive, absentminded, empathic, and anxious, using explorative principal component analysis. Appendix 2 presents a detailed description of the items used and the factor analyses including validity measures. From Table 1, it is clear that the psychosocial factors vary tremendously across bullying status. The gap in means between victims and controls ranges from 36 percent (for anxious) to 52 percent (for empathic) of a standard deviation. We expect these psychosocial factors to be strongly associated with victimization because they are closely related to the two prototypical victims: the passive and the provocative victim. Psychosocial factors might also very well influence school achievement and thus our outcome measure.10 Although it is clear that these variables stand out as potentially very important for our analysis, we also exert some caution because they are measured in 2001, the same year as the first parent survey on victimization, and therefore they may be affected by long-lasting victimization or common source bias. Partly for this reason we include control variables step by step below.
As is evident in Tables 2 and 3, parents of victimized children are negatively selected in terms of observable characteristics: They are younger when they give birth; they have lower levels of education, lower income; are more likely to be unemployed, more likely to be part-time employed; and are less likely to be higher-level employees. Similarly, they are more likely to be treated for cardiovascular diseases, receive antidepressants, and have a mental health diagnosis. Finally, they are significantly more likely to have a criminal history, which is especially true for fathers.
The characteristics of children and parents described above are employed in a rich conditioning set used in the subsequent empirical analysis.
3. Outcome
Our empirical analysis is concerned with potential consequences of victimization on ninth grade GPA obtained from Danish register data. A particular advantage of this study compared to other studies using surveys is that we obtain our outcome from a different data source than our treatment variable, removing concern about common variance. Furthermore, because the register information is available for the population of children born in Denmark, we do not face the problem of missing values in our outcome variables due to nonresponse.
The outcome measure is ninth grade GPA based on the marks at the end of ninth grade in the subjects written and oral Danish and written Mathematics. The average is taken over the preliminary mark (given by the teacher based on the pupil’s effort and achievement throughout the school year) and the mark at the national school exit exam (written and oral exams that are comparable across schools).11 To be able to compare grades across cohorts, we standardize grades to zero mean and unit standard deviation within each cohort.
Table 1 shows mean outcome by bullying status and indicates that being a victim of bullying correlates negatively with school performance measured by the ninth grade GPA. We stress that these observations do not represent causal pathways.
IV. Baseline OLS Results
We begin by estimating the relationship between bullying and GPA using OLS. Our baseline estimating equation is
(1)
where GPA indicates the outcome of interest, bul is an indicator for being a victim of bullying at age 10–12, and X is a rich conditioning set that includes the child and parental variables informative both about exposure to bullying and about GPA. β is our parameter of interest. Remember that we measure bullying status in 2001 and 2002. This implies that we must interpret our parameter of interest as the effect of victimization in elementary school, generally speaking. Victimization could have started earlier on and it may very well continue afterwards.
We gradually expand the conditioning set: We first include classroom fixed effects to address the importance of teacher and classroom characteristics. Next, we add the set of register-based individual (first block of Table 1) and parent specific variables (shown in Tables 2–3). We think of these as representing socioeconomic background (as most often available from administrative data or surveys) plus detailed information about health and criminal records. Third, we consider the consequences of adding the four psychosocial factors (hyperactive, absentminded, empathic, and anxious) described above and in Appendix 2. Finally, we add information about minor disabilities and height as measured in 2001 in an attempt to isolate the effect of bullying from effects of physical appearance.
The corresponding OLS results are shown in Table 4. We see that victims perform significantly worse than others in terms of ninth grade GPA. The size of the estimate is reduced somewhat by adjusting for background variables. When we include classroom fixed effects the estimate is reduced by about 15 percentage points. This means that the association between victimization and GPA is not driven by potentially troublesome variation between well-functioning classrooms and badly functioning classrooms. When we include the sets of register-based variables and psychosocial factors the estimate is substantially reduced, but once these are added the estimate is robust to the inclusion of additional variables measuring physical appearance. Remember that all register-based variables are measured before school start, and therefore the potential for reverse causality up to this point is most likely minor even if bullying has taken place from school start onwards. In Appendix 3 we show the full set of estimates revealing that many of the variables that predicted victimization (Tables A1-A3) are also important conditioning variables: The quality of the family environment, behavioral diagnoses, psychosocial factors, and crime in parents are significant determinants.
OLS Results: Bullying and Standardized Ninth Grade GPA
Estimates are significant in an economic sense: Our richest models suggest that exposure to bullying is associated with a reduced ninth grade GPA of more than 10 percent of a standard deviation (comparable to the effect of adding four extra pupils to the classroom; see Heinesen 2010).
V. Identifying Relationships Between Bullying and GPA: Exploiting Troubled Children in the Classroom
The key problem facing us is that it is not random who is bullied. In fact, as indicated by the literature review and our descriptive statistics above, victims are negatively selected in terms of observable characteristics. Moreover, children involved in conflict are also likely to be negatively selected in terms of unobservable characteristics. Although our conditioning set described above is incredibly rich, we cannot rule out that such unobserved characteristics will lead us to overstate the effects of bullying. An additional complication relates to the measurement of bullying as discussed above. For these reasons, it would be convenient to be able to rely on a more objective measure of classroom conflict.
In an attempt to solve these issues, we implement an instrumental variables strategy inspired by Carrell and Hoekstra (2010): Here we instrument victim status with the proportion of the child’s classroom peers whose parents have a criminal conviction (violent crime, property crime, and other non-traffic-related crime) or have served time in prison.12 For this to constitute a valid instrument, it must affect victim status (and the effect must go only in one direction) yet cannot directly affect academic outcomes for the other children. This means, for example, that we assume that teachers do not redirect resources away from the other children because of the presence of a troubled child, which is not an innocuous assumption.13 We also plausibly assume that the behavior of peers cannot cause a child’s parents to engage into crime.
Let Trouble–i be the proportion of peers whose parents have a criminal conviction (violent crime, property crime, or other non-traffic-related crime) or have served time in prison.14 We can then model victim status as:
(2)
where θi is the error term. We model the relationship between GPA and victim status as detailed in Equation 1 above. In practice we estimate the consequences of bullying using two stages least squares (2SLS).
We first investigate the variation of the instrument. Table 1 shows that 25 percent of victimized children and 22 percent of nonvictimized children have a classmate whose parent has been convicted of a crime. Figure 1 depicts the distribution of the instrument and illustrates that 13 percent of the children attend classes with no such classmates. Around 80 percent of the children are concentrated in the span from 5 percent to 50 percent. We observe a few individuals where the instrument equals one, which may be explained by some of the classes being incompletely observed. In the empirical analyses, we investigate robustness of the results to excluding extreme values of the instrument and excluding implausibly small classes.
Distribution of the instrument
We then investigate the correlation between the instrument and observable characteristics. In practice we regress the instrument on the full set of observable characteristics except for victim status (available on request). Coefficient estimates are small and most are insignificant although own parents’ criminal record as well as education of the mother are significant correlates and thus important control variables in the IV analyses to come.
Table 5 shows the results from the instrumental variables analysis. As in the simple OLS analysis above, we gradually expand the conditioning set.15 We find that the presence of criminal peer parents significantly increases the likelihood that a given child is bullied. The size of the first stage is slightly reduced with the inclusion of additional controls but even the model with the most extensive conditioning set suggests a positive effect of one percentage point increase in victimization when the proportion of troubled kids increases from 0 to 10 percent. The estimated effect on ninth grade GPA is large and negative but unfortunately also somewhat noisy. The second stage estimates are reduced when the conditioning set is enriched.
IV Results: Bullying and Standardized Ninth Grade GPA
Before we proceed, we investigate which parts of the distribution of the instrument drives the results. The results are robust to excluding the 13 percent of individuals with no classmate parents convicted of crimes but not robust to excluding the 7 percent of the distribution with more than half of classmate parents convicted of crimes. This indicates that the variation coming from classes with a high fraction of troubled children is important for the strategy to work. Above we raised a concern that the high fractions were explained by incomplete classes. However, the IV results are literally unchanged when we exclude implausibly small classes (<5 or <10). In the next section we investigate heterogeneity of the results and perform further robustness checks including sibling comparisons.
VI. Robustness Analyses and Heterogeneity
In this section we present a long range of robustness analyses. We show results from including mother FE, heterogeneity of results by gender, and by intensity of victimization. Finally, we explore how our estimates vary with different measurements of bullying.
A. Mother Fixed Effects
As mentioned above, bullying is related to standard socioeconomic measures such as family resources and ethnic origin as well as personal characteristics such as personality traits, psychological factors, disabilities, physical appearance, and physical weakness/strength. To the extent that these characteristics are not already captured by our extensive conditioning set and assuming they are fixed within a family, a mother fixed effects estimator will account for them.
Our data allow us to account for mother fixed effects for siblings who are born within the 1990–92 time period. That is, we consider closely spaced siblings. We exploit sibling pairs where one sibling is the victim of bullying and the other is not. The outcome of the nonvictim sibling can then be used as the counterfactual outcome.
The fixed effects strategy assumes that comparing siblings, perhaps conditional on attributes, eliminates selective differences between victims and controls. A common concern is exactly that although siblings are born into the same family and share this environment, they still may differ along a wide range of characteristics. If less able siblings are more likely to be exposed to bullying, the sibling comparison estimator will tend to bias the effect of bullying upward, just as the simple OLS is expected to do. To accommodate this criticism, our estimations include a wide range of variables descriptive of the child himself and his abilities; see above.
A second concern with within-family estimators is that the identifying population is potentially very small: 141 mothers in our final sample gave birth to more than one child during 1990–92 (43 gave birth to twins, two gave birth to triplets). Of these, we observe 33 sibling pairs where one is a victim of bullying and the other is not; these pairs identify our parameter of interest in the sibling analysis.16 Table 6 shows the percentages of sibling pairs in the different combinations of bullying status.17 Bullying status of the oldest sibling (Sibling 1) is on the vertical axis and bullying status of the younger sibling (Sibling 2) is on the horizontal axis. The table illustrates that the younger sibling is more likely to be reported as a victim of bullying if the older sibling is a victim of bullying and vice versa. Furthermore, we see that slightly more of the older siblings experience bullying.
Sibling Variation in Bullying Statusa
A final concern with this estimator is that we need to assume that one sibling is not affected if the other sibling is exposed or unexposed to bullying. Negative spillovers from exposed siblings will cause a bias toward zero in the fixed effects estimations while positive spillovers stemming from protective effects from unexposed siblings would cause a bias in the opposite direction.
We report the coefficients for the mother fixed effects specification in Table 7. As the sample size decreases significantly when we run the mother fixed effects model we also report the unadjusted OLS estimates for this reduced sample.
OLS Results: Bullying and Standardized Ninth Grade GPA
The important message from the mother fixed effects analysis is that the conclusions from the simple OLS seem largely robust. However, when we include psychosocial factors, the coefficient estimate is no longer statistically significant due to the small sample size.18
B. Gender Heterogeneity
In Table 1, we see that bullying status does not vary significantly by gender. In Table 8, we investigate if the relationship between GPA and victimization varies by gender. The estimates suggest that although the first stage is stronger for girls than for boys, the statistically significant effects of victimization tend to be driven by boys rather than girls.19 However, the samples are too small to draw firm statistical inference about the difference.
OLS and IV Results: Bullying and Standardized Ninth Grade GPA, Gender Differences
C. Intensity of Bullying
In Table 1, we see that 20 percent of the victims report to experience bullying “to a large extent” over the last six months, which we consider severe bullying. In Table 9, we redefine the endogenous variable to take the values zero (no bullying), one (minor bullying) and two (severe bullying). Also, when we use this linear measure of victimization, we find detrimental effects which increase with the intensity of bullying. This robustness check supports the validity of the measurement of bullying.
OLS and IV Results: Bullying and Standardized Ninth Grade GPA, Intensity of Bullying
D. Measurement of Bullying
In Table 10, we explore how the association between GPA and victimization varies with the exact definition of victimization. In the main analysis, we define an individual to be victimized if either the parent or the teacher indicates that the child was victimized. In Table 10, we compare the main results to the results from using the teacher’s report only, the parents’ report only, and from requiring that both informants agree that the individual is being bullied.
OLS and IV Results: Bullying and Standardized Ninth Grade GPA, Different Definitions of Victimization
One may view Table 10 as adding on to the results on the importance of intensity of victimization (see Table 9). The four measures of bullying identify effects at four different margins of severity: From the top to the bottom, the bullying measures identify 27 percent, 23 percent, 12 percent, and 7 percent as victims. Interpreting the point estimates at face value suggests that effects tend to increase with severity.
When we apply measures that require the teacher to agree on the child being bullied, the estimates of the effect of victimization become larger. One explanation might be that teachers apply a threshold other than parents and identify more severe cases. Another explanation might be that the teacher’s report is more strongly associated with the outcome, and one may be concerned that teachers misclassify some victims if their academic achievement is fine. For the measures that require the teacher to agree with the parents, the instrument becomes weak when controls are added, and as a consequence the estimated effect of victimization becomes noisier. This might indicate that the teacher applies a relative standard and potentially misclassifies children as victims due to other conditions, or it might indicate that these other conditions are the true reasons for weak educational performance rather than victimization as such.
VII. Conclusion
This paper investigates the determinants and potential effects of bullying in elementary school on educational performance measured by ninth grade GPA. We employ a number of strategies in order to come closer to identifying impacts of such experiences than previous research. We exploit a rich conditioning set that includes classroom information, parents’ socioeconomic background plus detailed information about health and criminal records as well as detailed accounts of children’s early physical and mental health outcomes and psychosocial factors measured just prior to exposure to bullying. In our main analysis, we implement an IV strategy inspired by Carrell and Hoekstra (2010) where we instrument victim status with the proportion of peers from troubled homes in one’s classroom. We show that bullied children suffer in terms of ninth grade GPA and that the effects of victimization tend to increase with severity. We emphasize that effects of exposure to bullying are particularly difficult to identify and caution that our estimates should be interpreted with this is mind. Of course, Denmark is a very homogenous society, which may limit the potential for conflict, and it is therefore possible that the nature of bullying is less severe than in other places. As such, we think of our estimates as lower bounds.
We show that the quality of the family environment as well as individual child characteristics such as poor early mental health, indicators of hyperactive behavior, and physical appearance are important drivers of victimization.
Given that bullying is likely so costly, can it be limited? Farrington and Ttofi (2009) systematically reviews evaluations of 44 school-based antibullying programs. It finds that the reviewed interventions on average reduce the prevalence of bullying and victimization by roughly 20 percent. Program effectiveness increases with inclusion of more elements, longer duration, and higher intensity. Some of the single elements that are significantly related to successful intervention are teacher and parent training as well as use of disciplinary methods and video and virtual reality video games. Furthermore, programs inspired by the pioneer Olweus are found to be more effective than others.
The details of the Olweus bullying prevention program are described in Olweus (1997). The idea is to combine warmth and positive involvement from adults with firm limits to unacceptable behavior. Violation of the limits and rules should be followed by nonhostile, nonphysical sanctions. The program implicitly requires some monitoring of behavior as well as adults acting as authorities at least in some respects. This relatively simple skeleton underlies bullying prevention programs implemented all over the world. Yet bullying prevails. Of course, such intensive programs are likely expensive and rely at least partly on very specific—and possibly limited—human resources. However, our results indicate that such programs may have longer-run aggregate effects in improving education and subsequently income of the population.
Appendix 1
Attrition
10,907 children were initially included in the ABC survey. Unfortunately, not all parents and teachers reported in the subsequent survey rounds. Those residing outside the region of Aarhus at the time of the surveys were not even asked to complete the survey.
For this reason, we drop 3,231 observations. Out of the remaining 7,676 children, we can identify classmates at school entry for 4,490. Finally we exclude 235 children, who drop out before the ninth grade exit exam or who skip the exam. We thus include 4,255 children in the empirical analyses.
Among the 10,907 individuals initially included, 70 percent of the parents and 52 percent of the teachers respond to the bullying question in the 2001 and 2002 round of the questionnaires. This gives rise to concern about possible bias due to attrition, especially because the subject being surveyed is of sensitive nature. Because the survey is linked to register-based information, we are able to test possible differences in the populations of parents who responded and who did not respond. We find that nonrespondents are more likely to have worse socioeconomic background, were on average younger when the child was born, were more likely to be of ethnic minority origin, and have more psychiatric diagnoses.
Appendix 2
Additional Information Regarding Psychosocial Factors
This appendix presents details behind the factor analyses conducted to arrive at the psychosocial factors. The factor analyses extract the common variance in responses to a set of questions about socio-emotional and psychological issues in the questionnaire conducted in 2001.
The questionnaire contains a range of items from the Child Behavior Check List (CBCL), see Achenbach (1993). However, not all items are included in the questionnaire, and therefore, we are not able to obtain the entire scales. Instead we conduct an explorative factor analysis using principal components where items are chosen based on Children’s Behavior Questionnaire (CBQ), see Rutter (1967). Whereas the CBQ focuses on undesirable traits, the CBCL incorporates other aspects such as prosocial behavior. We therefore include additional items not obtained in the CBQ that describe prosocial behavior. Our analysis is based on 14 items. The KMO for all items is 0.875, which validate the use of factor analysis. The explorative factor analysis suggests four factors. Appendix Table A1 presents the four factors with their respective loadings and Cronbach’s Alpha. Each variable loads highly on one factor and not much on the remaining factors, giving us a clear factor structure. The items loading high on each factor clearly suggests the labels: anxious, hyperactive, empathic, and absentminded. Furthermore, Cronbach’s Alphas are high, which indicates good internal validity of the factors.20 We obtained factor scores on each of the factors using the regression method. These scores were then incorporated in the regression analysis and denoted psychosocial factors.
Questions, Loadings, and Cronbach’s α
Appendix 3
Additional Results
OLS Results: Bullying and Standardized Ninth Grade GPA, Full Set of Estimates
Footnotes
Tine Louise Mundbjerg Eriksen is a PhD student in economics at Aarhus University.
Helena Skyt Nielsen is a professor of economics at Aarhus University.
Marianne Simonsen is a professor of economics at Aarhus University. The data used in this article can be obtained beginning May 2015 through April 2018 from Helena Skyt Nielsen, Fuglesangs Allé 4, DK8210 Aarhus V, hnielsen{at}econ.au.dk. Data security policy means that access can be obtained from Aarhus only. The authors thank the Danish Psychiatric Central Register for access to data, and Carsten Obel for valuable discussions and for kindly allowing us access to the Aarhus Birth Cohort. Financial support from the Strategic Research Council (Grant 09-070295) is gratefully acknowledged. The authors appreciate valuable comments from three anonymous referees, Kristina Risom Jespersen, and Erik Lindqvist and from participants at the conferences: Human Capital Formation in Childhood and Adolescence (IFAU), ESPE 2011, the 2nd International Workshop on Applied Economics of Education, as well as seminar participants at the University of Wisconsin and Aarhus University. The usual disclaimer applies.
↵1. This number includes the pupils attending the voluntary tenth grade. For details, see Ministry of Education (2009).
↵2. The figures for after-school care apply for 6–9-year-olds. See Statistics Denmark (2010).
↵3. The reported figures apply for 6–9-year-olds. Among 10–13-year-olds, 32 percent attend after-school care, and for this age group it most often takes place in a recreation center or in a youth club. See Statistics Denmark (2010).
↵4. 99.8 percent of all pregnant women received this type of care. See Delvaux et al. (2001).
↵5. The psychiatric diagnoses are obtained from the Danish Psychiatric Central Register; see Munk-Jorgensen and Mortensen (1997) for details.
↵6. In 2001, parents are asked whether their child engages in bullying and whether the child is being bullied (No, To a small extent, To some extent, To a large extent). In 2002, parents and teachers are asked to what extent during the past six months are the following statements descriptive of the child: is being bullied or teased by other children in school; often gets into fights or bullies other children (Not true, Somewhat true, Certainly true).
↵7. See the discussion by Wolke et al. (2001) about the problems of defining an internationally comparable measure of the prevalence of bullying when the languages differ.
↵8. These issues turn out to be particularly severe when the subjective measure is used as an outcome.
↵9. See Currie and Stabile (2006) and Fletcher and Wolfe (2008), which argue that children with ADHD suffer in terms of academic outcomes.
↵10. We denote these variables psychosocial factors but they are closely related to personality traits and socio-emotional capabilities, and the underlying questions are widely used in child psychology/psychiatry to describe children’s well-being and to screen for and diagnose mental and behavioral disorders. In addition, they may also be related to victim mentality as discussed earlier.
↵11. The written exams are identical across the country and all exams, whether written or oral, are graded by the teacher and an external examiner, where the opinion of the external examiner dominates the opinion of the teacher. The teachers involved in the ninth grade exam (taking place at age 15–16) are unlikely to be the same teacher who informed about bullying at age 9–11. However, as a robustness check we also compute a GPA based on written Danish and Math, which are the most centralized and objective exams. The results are robust to this alternative measure of the outcome variable and available on request. Only 92 percent of the children sit the ninth grade exam. We ignore the selection into taking the exam.
↵12. Previous research by Espelage et al. (2000) suggests that children who are slapped or hit when they break the rules at home or who lack positive adult role models for conflict management engage in more negative actions towards other students. We cannot check this directly in our data set.
↵13. Figlio (2007) finds that boys who are disruptive because of the stigma associated with their feminine names create disruptive ramification for peer learning.
↵14. A more direct approach would be to use the number of reported bullies in class as an instrument for victimization. When we do that the first stage is incredibly weak, and we suspect that this has to do with under-and misreporting of perpetrators.
↵15. Note that class fixed effects are not identifiable.
↵16. Of these sibling pairs 12 are twin pairs.
↵17. Families that give birth to three children in the period constitute two sibling pairs; sibling two and three are each paired with sibling one.
↵18. In an additional specification we have included an indicator variable for being the oldest of the sibling pair. In this specification the estimated effect of victimization increases, which reflects that the older sibling is more often bullied and does better in school on average.
↵19. When we compare the importance of background characteristics across genders, some differences show up: Type of childcare and a high score on the hyperactivity factor are important for boys but not for girls and the association with mother’s labor market activities differ.
↵20. We tested the Cronbach’s Alpha by deleting and adding items with higher cross loadings. In no case could Cronbach’s Alpha be increased.
- Received September 2012.
- Accepted September 2013.
References
In this issue
Jump to section
- Article
- Abstract
- I. Introduction
- II. Background
- III. Institutional Context and Data
- IV. Baseline OLS Results
- V. Identifying Relationships Between Bullying and GPA: Exploiting Troubled Children in the Classroom
- VI. Robustness Analyses and Heterogeneity
- VII. Conclusion
- Appendix 1
- Appendix 2
- Appendix 3
- Footnotes
- References
- Figures & Data
- Info & Metrics
- References