We created the study cohort for all individuals born in 1950 or before in South Korea at baseline in 2010 (n=7,889,086), using the Korean National Health Information Database (NHID) from January 2010 to December 2019. The NHID contains demographic information and medical treatment data for all Korean citizens. Korea’s universal healthcare system consists of National Health Insurance and Medical Aid enrollees (excluding temporary residents). National Health Insurance is administered in 2 ways: 1) workplace health insurance (i.e., workers and employers registered at any workplace) and their dependents, and 2) community insurance (i.e., including those who are not registered as insured employees, including the unemployed, the self-employed, and retirees not supported by employed family members)[20]. Both groups pay a mandatory monthly premium (based on their income). Certain low-income individuals (with no ownership of assets such as real estate or children who can support them) are insured through Medical Aid and receive NBLSS benefits. However, a significant number of low-income individuals do not receive Medical Aid/NBLSS, due to the stringent asset criteria and rules about family support obligations, e.g., elderly parents must demonstrate that they are not receiving support from their children.
NHID contains information of both National Health Insurance and Medical Aid enrollees, and we extracted information on insurance eligibility, sociodemographic data, outpatient and inpatient visits data, and date of death (if applicable). The NHID data was also linked to vital statistics (for cause of death data) created by Statistics Korea.
Outcome: Suicide-Related Behaviour
This study has two outcomes, which includes 1) death from self-harm with suicidal or undetermined intent and 2) SRB-related hospitalisation. For our first outcome, we used the International Classification of Diseases version 10 (ICD-10) codes to identify deaths from self-harm (X60-X84) and deaths that are of undetermined intent (Y10-Y34). It is now common practice to include death by undetermined intent (i.e. no evidence for intent to die) into studies of SRB [21], this is particularly important given the context of this classification in Korea. Where legal authorities find no recorded evidence of intent to die (e.g. suicide note), at the time of the person initiated self-killing (especially by self-poisoning), the death is coded as being due to an “event of undetermined intent”. The classification is applied if the intention to die “cannot be established beyond a reasonable doubt”, which is a legal construct (rather than evaluating the balance of probability for suicide) that leads to an underestimation of true suicides [22]. In a prior study, using psychological autopsy methods (which involve interviews with friends, family, and a chart review of clinical notes), a Finnish study uncovered suicide intent in 87% of 139 deaths that were classified to be undetermined intent [23]. A Korean study, investigating the trends in suicides by pesticide, and undertermined deaths by pesticide poisoning, estimated the misclassification of suicides as undetermined deaths has led to suicides (by pesticides) to be underestimated by 15-31% between 1991-2012 [24]. Therefore, we include deaths of undetermined intent as part of the primary outcome.
For our second outcome, SRB hospitalsation, we used the ICD-10-codes X60-X84, Y87.0, U03, and Y10-Y34 to identify hospitalisation for suicide attempts. In a prior study of Korean adolescent suicide-related emergency department visits from 2016-19 [25], it was found that 3,006 out of 11,462 (26.23%) suicide-related emergency department visits was followed by hospitalisation and the rest were discharged. Of note, those who were hospitalised had significantly higher levels of acuity based on the Korean Triage and Acuity Scale (79.9% of hospitalised patients were considered urgent or greater at triage vs. 47.1% of discharged patients, p<0.001). We also included probable suicide attempts (see supplementary table 1 for all ICD-10-codes), which included wrist lacerations (S61.9) and poisoning from psychotropic drugs (T43). Previous studies have included these codes for probable suicide attempts in Korea [26]. Since intentional self-harm was not covered by Korean National Health Insurance until 2014, using only intentional self-harm codes has been found to underestimate attempted suicides by 62-81% [27].
Measure of levels of income
Levels of income were inferred from National Health Insurance premiums. Premiums are calculated based on monthly wages for employees, and a combination of assets and all income are considered for those who are unemployed, retired, and self-employed. Insurance premiums were provided as categorical data, and divided into quartiles based on the general population distribution. To note, each 10,000 KRW is approximately $10 USD. Based on 2010 numbers, insurance premiums for the employee insurance enrollees range from 0-31,980 KRW in Q1, 31,981-50,740 KRW in Q2, 50,741-105,120 KRW in Q3, and 105,121-1,753,300 KRW in Q4. Given these premiums, the maximum income for Q1 is 600,000 KRW, 1,551,989 KRW for Q2, and 2,924,221 for Q3. For the majority of workers in Q1, their incomes are below the national poverty threshold of 550,000 as of 2010, and for the purpose of this study, we will consider the Q1 group to be living in poverty. The NBLSS group, as mentioned earlier, are low-income individuals who meet additional criteria (e.g. no assets and no direct children supporting them), and are also considered to be in the lowest income. The ‘others’ groups include individuals who do not pay national health insurance premiums, which includes institutionalised populations, soldiers, and those who are on long-term leave (e.g. sick leave). Since the last group do not pay premiums, we are unable to infer their socio-economic status; however, they are included as a separate group in the models.
Covariates
The covariates included in this study are 1) birth year, 2) urban/rural status, 3) employment status, 4) disability status, and 5) comorbidity index. Birth year was included as a continuous variable to adjust for age and cohort effects. Urban/rural status was determined by place of residence. An individual’s place of residence was classified as rural if their place of residence was in a local jurisdiction with population less than 100,000 (“gun”), and urban if in a jurisdiction with more population than 100,000 (“si”). Employment status was recorded as a binary variable (waged workers vs self-employed, out of the labour force, or unemployed). Information on all individuals with disability are registered in the eligibility database, and thus disability status was recorded as a binary variable (any disability vs no disability). Comorbidity was measured using the Charlson comorbidity index (CCI), which was calculated through the ICD-10 codes from 2010 to 2019 and recorded as 0,1,2,3, and 4+ comorbidities [28].
Statistical analysis
Descriptive statistics were calculated for incidence rates of 1) death from self-harm and undetermined intent, 2) deaths from self-harm only, and 3) SRB-related hospitalisation across levels of the main exposure and each covariate at baseline. Royston-Parmar flexible parametric survival analysis was conducted, since the results of Schoenfeld residuals test indicated that the proportional hazards assumption was not met [29]. This survival model is an alternative to Cox regression by allowing greater flexibility in model fitting, and providing both baseline hazard and the effect of time-varying covariates on the baseline hazard [29]. Adjusted Hazards ratios (HRs) and 95% CIs were estimated by gender for each outcome using multivariate flexible parametric survival analysis. Lastly, to investigate the difference of SRB risk between the two groups living in poverty (i.e. Q4 vs NBLSS), we used 1000 non-parametric bootstrap samples (sampled with replacement) to obtain 95% confidence intervals for the difference in risk between Q4 and NBLSS.
We conducted several sensitivity analyses. First, the Schoenfeld residuals indicated that the proportional hazard assumption was not met, potentially driven by the large number of participants, which includes the entire 65+ population [30]. Therefore, we repeated the analyses by fitting Cox proportional hazards regression models to ensure that the results of the alternative survival model were robust. Second, we produced model results using deaths from self-harm only (excluding deaths from undetermined intents) and hospitalisations from suicide attempts only (excluding hospitalisations from probable suicide attempts) to verify the model with hospitalisations from suicide attempts only shows similar patterns to the model containing suicide attempts and probable suicide attempts. Statistical significance was defined as a 2-sided P-value of <0.05.