Study population

We obtained information on all children born in Sweden from 1987 to 2020 from the Total Population Register (n=3,704,422). The children were linked to their biological parents and siblings (full and half) using the Multi-Generation Register. Additional information on the children and their relatives was extracted from the Total Population Register, the Longitudinal Integrated Database for Health Insurance and Labour Market Studies, the National Patient Register (NPR), the Cause of Death Register and the Swedish Prescribed Drug Register. Linkages across registers were enabled through personal identification numbers, a unique 10-digit number that is assigned to all residents in Sweden at birth or immigration [19].

To enable the establishment of attachment relationships [20] and to avoid including events of neonatal diabetes, follow-up started at the age of 1 year. We excluded children who had died or emigrated before that age or who had incomplete migration data (Fig. 1). We also excluded children with missing information on the identity of their parents, and children with one or two parents who had died or emigrated before the start of follow-up. We further excluded children with any diabetes diagnosis (including neonatal diabetes or any other type of diabetes) recorded in the NPR before the start of follow-up. Any diabetes diagnosis before the start of follow-up was defined as having had an inpatient visit with the main diagnosis coded as 250 according to ICD-9 (http://www.icd9data.com/2007/Volume1/default.htm) or E10 or E11 according to ICD-10 (https://icd.who.int/browse10/2019/en). Our final study population comprised 3,598,159 children.

Fig. 1

Flow chart of the study population

Exposure

The main exposure, childhood bereavement, was defined as the death of a biological mother, father or sibling. The date of death of the parent or sibling was ascertained through the Cause of Death Register. We categorised age at loss as preschool (1–6 years), school age (7–12 years) or teenage (13–17 years). We categorised the main cause of death as illness (e.g. cardiovascular disease or cancer) or external causes of morbidity and mortality (including suicide, accidents, environmental exposures and homicides). External causes of morbidity and mortality were defined by ICD-9 codes E800–E999 and ICD-10 codes V01–Y98, while illness was defined as death by all other ICD-9 or ICD-10 codes. The familial relationship with the deceased was also investigated, and categorised as mother, father or sibling.

Outcome

The main outcome was type 1 diabetes in childhood (1–17 years), defined as an inpatient main diagnosis (ICD-9 code 250 or ICD-10 code E10) in the NPR. The date of diagnosis was defined as the date of discharge.

Although ICD-9 does not have separate codes for the various types of diabetes, the risk of misclassification is low, because >98% of Swedish children aged 0–18 years who are diagnosed with diabetes have type 1 diabetes [21]. However, a previous study showed that prescription of insulin in the Swedish Prescribed Drug Register, which holds information on all dispensed medications in Sweden from July 2005 onwards, could be used to reliably assess the occurrence of type 1 diabetes in individuals aged 0–34 years [8]. To explore the validity of our outcome, we assessed the proportion of children defined as having type 1 diabetes who had at least one dispensed prescription of insulin (anatomical therapeutic chemical code A10A) at <18 years, and the proportion of children with one or more dispensed prescription of insulin at <18 years who did not have a type 1 diabetes diagnosis.

Covariates

We obtained information on baseline parental covariates, including age, country of birth (categorised as Sweden or other), region of residence (categorised as Götaland, Svealand, Southern Norrland and Northern Norrland) and the population density of home municipality (calculated as the number of inhabitants per km2) from the Total Population Register. Baseline was defined as the year of birth of the child. Information on parental highest achieved education level (categorised as compulsory, secondary or university), disposable income (presented in quintiles) and marital status (categorised as married, not married but cohabiting with children, or single) was obtained from the Longitudinal Integrated Database for Health Insurance and Labour Market Studies, which collects sociodemographic information on all Swedish residents from 1990 onwards. Parental type 1 diabetes status was based on a diagnosis of type 1 diabetes (ICD-9 code 250 and/or ICD-10 code E10; main inpatient or outpatient diagnosis) in the NPR. Race as a concept is not used in Sweden, and information on race or ethnicity is not available in Swedish national registers and is therefore not included in our analyses [22]. All children in our cohort were born in Sweden.

Prior to the analysis phase, we created a directed acyclic graph using the DAGitty tool (available at http://www.dagitty.net) [23]. The directed acyclic graph (electronic supplementary material [ESM] Fig. 1) is a graphical presentation of the theoretical framework of the study, describes our prior assumptions on how bereavement may have a causal effect on child type 1 diabetes development, and further helps to identify potential confounders.

Statistical methods

We used Cox proportional hazards models, with attained age as the timescale, to assess the association between the death of a family member and the risk of type 1 diabetes. Death of a family member was included in the models as a time-varying variable. An individual was classified as unexposed until the date of death of a family member, and was thereafter classified as exposed until the end of follow-up. If a child lost more than one family member during the study period, the first death was used to classify the exposure status. We did not evaluate the effect of multiple losses. The study individuals were censored at emigration (their own or parental), death, when the individual turned 18 years old, or at the end of follow-up (31 December 2021), whichever came first. To explore how the potential effect of losing a family member may vary by time since exposure to loss, we used restricted cubic splines (four knots at the 5th, 35th, 65th and 95th percentiles, respectively). A robust sandwich estimator of variance was used to account for the within-family correlations [24].

The models were adjusted for the potential confounders identified in our directed acyclic graph, i.e. the baseline variables year of birth of child, maternal age at delivery, parental country of birth, parental type 1 diabetes, region of residence and population density of the home municipality. However, information about parental type 1 diabetes was updated during follow-up and treated as a time-varying covariate. Less than 2% of our study population had missing information for the confounders, and complete-case analysis was performed. To examine the proportional hazards assumption, we used Schoenfeld residuals. These indicated that the assumption was violated for parental type 1 diabetes, and therefore a stratified Cox model allowing for different baseline hazards was used.

We further performed subgroup analyses by the sex of the child.

Analyses were performed using SAS version 9.4 (SAS Institute, USA) and R version 4.3 (The R Foundation, Austria) [25]. The study was approved by the Swedish Ethical Review Authority (DNR 2018/1697-31/1, with amendment 2021-03277).