Study design and population
Children and adolescents aged through 18 years who presented to hospital emergency departments with self-harm and other mental health emergencies during March–April 2020 or March–April 2019 were included in this study. The study uses a retrospective cohort study design to compare the 2020 and 2019 cohorts in terms of characteristics of emergency presentations.
Electronic patient records of 23 hospital emergency departments in 10 countries (England, Scotland, Ireland, Austria, Italy, Hungary, Serbia, Turkey, Oman and the United Arab Emirates) subdivided into 14 relatively geographically homogeneous areas: London, Dublin, Edinburgh, Livingstone, Manchester, the Home Counties (suburban areas around London), Turin, Cagliari, Vienna, Budapest, Belgrade, Istanbul, Muscat and Dubai. The hospital emergency departments serve the total population of approximately 31.2 million with 6.5 million under-18 s. There is a total of approximately 200,000 paediatric emergency presentations per year. The hospitals are representative of health care systems in developed high-income (England, Scotland, Ireland, Austria, Italy, Hungary), developing middle-income (Serbia and Turkey) and developing high-income (Oman and the United Arab Emirates) countries.
We measured the following sociodemographic characteristics for the entire sample of presentations: sex, age, dominant ethnic group (yes/no), decile of deprivation index, young person in education, employment or training (yes/no), young person looked after by the local authority and the young person’s biological parents living together out of those who are looked after by their parents (yes/no). We then measured the following clinical characteristics for all emergency presentations: presentation for self-harm (yes/no), children and adolescents with previous hospital presentations for self-harm (yes/no), or with previous self-harm in the community (yes/no).
Sub-sample with self-harm presentations
For emergency presentations for self-harm we measured sociodemographic characteristics; clinical characteristics, including proximal risk factors for self-harm, clinical diagnosis and distal risk factors for self-harm; clinical management variables. The following clinical variables were measures: severe self-harm (yes/no), presence (yes/no) of each of emotional disorders, behavioural disorders, psychotic disorders, eating disorders, neurodevelopmental disorders, substance misuse disorder, somatoform disorders or personality disorders, use of a violent method of self-harm (yes/no) and suicide intent (yes/no). We then assessed whether or not a row with a family member or social isolation were precipitants of self-harm, whether the children and adolescents used social media to communicate about self-harm, used alcohol at the time of self-harm, had previous hospital emergency presentation for self-harm, had previous hospital emergency presentation for reasons other than self-harm, had previous self-harm in the community, previous psychiatric inpatient treatment and had a family history of self-harm. Finally, we measured the following variables reflecting patterns of clinical management of self-harm: young person detained under a section of Mental Health Act (yes/no), length of stay in the emergency department, dichotomised to more or less than 24 h; admission to observation wards in emergency departments (yes/no), acute wards (yes/no), Intensive Treatment Units (yes/no), psychiatric inpatient wards (yes/no); offer of community follow-up within seven days of the presentation with self-harm (yes/no) and attending at least one follow-appointment within seven days of the original hospital presentation if community follow-up was offered (yes/no). We recorded phone, remote or face-to-face follow-up appointments.
Self-harm was defined using the UK National Institute for Health and Care Excellence clinical guidelines as “any act of self-poisoning or self-injury, irrespective of the underlying intent” (NICE), thus incorporating both non-suicidal self-injury, suicide attempts non-suicidal self-poisoning and self-harm with unclear or mixed intent. Severe self-harm was defined as meeting at least one of the following criteria: (1) A high-lethality method used (drowning, hanging, jumping from heights, shooting, potentially lethal self-poisoning in the absence of medical care and self-injury involving major vessels) (2) Any self-harm resulting in an Intensive Care Unit admission 3. Any self-harm resulting in an admission to an acute ward for 72 h or more.
Sample size calculation
The two main outcomes of interest in this study were (a) frequency of hospital presentation and (b) proportion of presentations with severe self-harm out of all hospital presentations for self-harm. To address (a) we simply collected presentations from as many different areas as possible within the available data collection period. To address (b) we used evidence from one London emergency department obtained in March 2020 that suggested that while the number of presentations might be reduced to a quarter the number in 2019, within those with presentations for self-harm the proportion of severe self-harm might have increased from around 8% in 2019 to around 40% in 2020. Based on a two-sided Fisher’s exact test with significance level 5% and a power requirement of 90% we calculated that a minimum of 84 self-harm presentations in 2019 and 21 self-harm presentations in 2020 would be required to detect such an effect within a hospital area.
Demographic and clinical characteristics of the sample of hospital emergency presentations were described using relevant summary statistics within site and year, and overall. The formal statistical analyses assessed the effect of the pandemic/interventions by comparing 2020 outcomes with 2019 outcomes for each site and across areas.
To compare the number of hospital emergency presentations between 2020 and 2019 a negative binomial model was fitted to the 28 counts of hospital emergency presentations from 14 areas in two years. The negative binomial model was chosen to allow for a positively skewed distribution of the counts as well as over-dispersion, e.g., because hospital emergency presentations were by the same young person. The model contained the hospital emergency presentations count as the dependent variable and year and hospital area as explanatory variables.
To compare (1) proportions of children and adolescents presenting with self-harm, having previously presented with self-harm to a hospital or engaged in self-harm in the community, and (2) the characteristics of those presenting at hospital for self-harm between 2020 and 2019, individual participant data (IPD) meta-analyses will be used. The modelling proceeded in two stages: In the first stage relevant linear or logistic regression models were fitted for each area separately. These models contained the respective outcome as the dependent variable and year (2020 vs 2019) as the explanatory variable. In the second stage, relevant site-specific effects were combined into an estimate of an overall year effect using meta-analysis. We opted for a random effects meta-analysis approach to combine effect size estimates. To account for the correlation between repeated outcome measures from the same young person cluster bootstrapping was used to generate 95% confidence interval with the resampling cluster defined by the person (1000 bootstrap samples). When modelling binary outcomes areas with cell counts of less than five in the cross-tabulation of the outcome by year were excluded from the IPD meta-analysis to ensure that bootstrapping inferences were reliable. We report the overall year effect estimate unless there is evidence for effect modification by area according to Cochran’s Q test. A measure of the amount of between-area year effect variability is provided by the I2 statistic, and which is shown on the resulting forest plots. The analyses were implemented in Stata using the command ipdmetan .
Where (1) self-harm variables or (2) other outcome variables in those who presented with self-harm were missing presentations were excluded from the area-level and overall year comparison. In other words, complete case analyses were used which assumed that within years observations were missing completely at random.