Data sources
Children’s Internet risk data were derived from the dataset released by the Economic and Social Data Service’s EU Kids Online survey (Study No. 6885, EUKOS) [
17]. The EUKOS provides the data on Internet-related behaviours of children and parents in 25 European countries, including Austria, Belgium, Bulgaria, Cyprus, the Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Lithuania, The Netherlands, Norway, Poland, Portugal, Romania, Slovenia, Spain, Sweden, Turkey, and the United Kingdom. The data were collected in a face-to-face survey in homes with Internet users aged 9 to 16 from the 25 countries; in all, 25,142 children were interviewed in 2010. The survey methodology and findings have been reported elsewhere [
2,
16]. Since another major data source in this study – the EDMD mortality database – only provides aggregated data in age groups 10–14 and 15–19 years, data samples of children aged 10–14, i.e. shared age range in both EUKOS and EDMD datasets, were selected for the analysis.
Four Internet risks were chosen and defined as follows:
1)
Exposure to online information on self-harm or suicide: answering “Yes” to “Have you seen websites where people discuss ways of physically harming or hurting themselves or ways of committing suicide (either one)?”;
2)
Experience of online and traditional bullying: answering “Yes” to “Has someone acted in this kind of hurtful or nasty way [this can include teasing someone in a way that the person did not like (online), hitting, kicking or pushing someone around, or leaving someone out of things (traditional)] to you on the Internet?”;
3)
Experience of exclusively online bullying (excluding traditional bullying): respondent confirmed having had “the experience of online bullying” but answered “no” when asked if the experience had occurred “in person, face to face” and “by mobile phone calls, texts or image/video texts”;
4)
Internet addiction (measured by a composite score): choosing an option among “very often”, “fairly often”, “not very often”, and “never/almost never” for a five-item scale: How often have the following things happened because of the Internet? 1) “Going without eating or sleeping”, 2) “feeling bothered when I cannot be on the Internet”, 3) “catching myself surfing when I'm not really interested”, 4) “spending less time than I should with either family, friends or doing schoolwork”, and 5) “trying unsuccessfully to spend less time on the Internet”.
The reference period for the above questions was twelve months. Missing responses were excluded from the analysis.
Based on each country’s data samples, prevalence rates of exposure to online information on self-harm or suicide, experiences of online and traditional bullying, experiences of exclusively online bullying, and the mean score of “Internet addiction” were calculated.
The European Detailed Mortality Database (EDMD) comprises the number of deaths for each European country, which are stratified by year, cause of death, age group, and sex [
18]. The cause of death in EDMD was coded using the International Classification of Diseases (ICD-10), tenth revision [
19], except for Greece. The sex and age stratified population figures of each country are also provided in the EDMD. Annual unnatural child death is defined as the collection of death causes including accidents (ICD-10: V01-X59), self-harm (ICD-10: X60-X84), assaults (ICD-10: X85-Y09), and undetermined cause of death (ICD-10: Y1-Y34) for children who died at age 10 to 14. For data from Greece, the cause of death is coded using ICD-9, the ninth revision [
20]. The equivalent definition of unnatural death is devised as those who died from causes of death categorised with ICD-9 codes 800–999. For each country, the annual mortality incidence rate of unnatural deaths of children aged 10–14 was calculated (per 100,000).
Most countries’ mortality data were based on year 2010 figures, i.e. same year as the EUKOS study. When unavailable, the latest figures were used. Turkey’s data were not analysed because its death data were not available in the EDMD. Finally, data of twenty-four countries were included and analysed.
In the collected dataset, there were 1,013 unnatural child deaths (ages 10–14) among the twenty-four European countries. Of that total, 789 (77.9%) were accidental deaths, 138 (13.6%) were suicide deaths, 38 (3.8%) were caused by various forms of assault, and 48 (4.7%) were classified as undetermined causes of death. The unnatural death mortality incidence (ages 10–14) was ordered from the lowest 0 (Cyprus) to the highest 12.6 per 100,000 (Romania).
We also tested whether spurious associations of the national psychiatric problems might exist between Internet risk and unnatural causes of death, i.e. testing mediating and confounding effects using the same statistical procedure [
21]. Proxy measures of a country’s prevalent rates of psychiatric problems were derived from the 2006 and 2012 European Social Survey respectively (ESS3 and ESS6) [
22,
23]. As the two datasets do not provide age breakdowns, respondents who were 18 years old or younger were selected as a proxy for each country. The depression symptom score was calculated using an eight-item version of the Center for Epidemiologic Studies Depression Scale Revised (8-item CES-D) [
24]. It ranges from 0 to 24 and the clinical cut-off score is chosen as 7 [
25]. It is interpreted as a higher score indicating a greater severity of depressive symptoms. Each country’s prevalence rates of depressive symptoms in 2006 and 2012 were computed to test for a possible spurious association respectively. Since some countries’ CES-D scores were not available in the two datasets, Czech Republic, Greece, Italy, Lithuania, and Romania were excluded in the 2006 spuriousness test. Austria, Greece, and Romania were excluded in the 2012 test.
The protocol of this secondary data analysis was approved by the Human Research Ethics Committee for Non-Clinical Faculties, The University of Hong Kong.
Statistical analysis
The incidence of unnatural deaths was plotted against the prevalence rates of the four Internet risks and was analysed using quasi-Poisson regression [
26]. The reason for using quasi-Poisson rather than conventional Poisson regression is to adjust for the spurious standard error due to the problem of dispersion. Negative binomial regression, which is also a common choice of over-dispersion adjustment, was not used because that technique gives higher weights to observations with a lower mean (that is to say, the countries with lower incidence of unnatural deaths), which may not provide a good estimate of the overall incidence of unnatural deaths for the current study [
26].
The dependent variable was the raw count of unnatural deaths while the independent variable was the four Internet risks. The population difference among counties was adjusted by the introduction of an offset term of taking the logarithm of the age 10–14 population in the equation. Separate quasi-Poisson regression models were constructed for each Internet risks.
The Pearson’s correlations between the prevalence rates of depressive symptoms in 2006 and 2012 and the prevalence of exclusively online bullying were calculated. A quasi-Poisson regression model was also constructed such that unnatural cause of death was treated as a dependent variable, and the years 2006 and 2012 prevalence rates of depressive symptoms were treated as independent variables respectively. Statistical test of significance was set at the 5% level. We further deploy Cook's distance to quantify the effect of each data point on the model outcome [
27]. An outlier is determined to be influential if its Cook’s distance is substantially larger than the rest [
28].