Medication-only cohort

We identified individuals aged 15–54 years who had collected at least one glucocorticoid prescription (see Measures, below, for details on Anatomical Therapeutic Chemical (ATC) codes) in oral form between 1 January 2006 and 31 December 2020. We focused on individuals aged <55 years to estimate effects in adults. We excluded individuals who had ever been dispensed a glucocorticoid where the prescription text indicated they had Addison’s disease, given that steroid treatment in these individuals is intended to replace insufficient endogenous production. Follow-up started on either 1 January 2006 or the date of the individual’s 15th birthday, whichever came last; it ended on the date of either first emigration, death, reaching age 55 years or 31 December 2020, whichever occurred first.

Indication cohorts – target trial emulation

As an alternative to the medication-only cohort, we defined indication cohorts in order to emulate two target trials. This allowed us to define clear clinical questions: what is the effect of glucocorticoid treatment initiation following any autoimmune or a gastrointestinal autoimmune disorder diagnosis on psychiatric and suicidal behaviour risks? See Supplementary Table 1 for a description of the target trials and how we emulated them.

We identified two cohorts based on a diagnosis of disorders indicating glucocorticoid treatment – any autoimmune disorders and gastrointestinal autoimmune disorders. See Supplementary Table 2 for the diagnoses considered, defined using ICD-10 codes. ¹⁵ We selected individuals who had received a diagnosis of either of these disorders between 1 January 2006 and 30 November 2019, choosing the first recorded diagnosis as index. We then excluded individuals who had been prescribed a glucocorticoid <180 days before they received their diagnosis, to ensure that all included individuals had a washout period from glucocorticoid treatment of at least 180 days. Anyone initiating a glucocorticoid within 28 days of their diagnosis was assigned as an initiator, and anyone who did not as a control. The start of follow-up was the date of glucocorticoid dispensation among initiators. Among controls, the time between diagnosis date and dispensation date in initiators was randomly assigned to define the follow-up start. ^{Reference Lagerberg, Matthews, Zhu, Fazel, Carrero and Chang16} We carried out intention-to-treat and per-protocol analyses (see Analyses, below). For intention-to-treat analyses, follow-up ended at 365 days after follow-start. For per-protocol analyses, follow-up ended at 365 days after follow-start unless individuals stopped adhering to their baseline treatment strategy, at which point they were censored.

Exposure

We considered any glucocorticoid medication licensed for sale in Sweden during the study period. Because we did not have information on individual ATC codes below the second level (H02) for the main cohort, we could not directly select glucocorticoid-only dispensations (H02AB). However, the only medication in the H02 category that is not a glucocorticoid and licensed for sale in Sweden is fludrocortisone (H02AA02), a mineralocorticoid used to treat Addison’s disease and congenital adrenal hyperplasia, both rare conditions. An estimated 1300 Swedish residents have Addison’s disease, ¹⁷ and we excluded any individual where their prescription text indicated this disease, although we did not have structured diagnostic information on it. Meanwhile, only 606 individuals born between 1915 and 2011 had congenital adrenal hyperplasia. ^{Reference Gidlöf, Falhammar, Thilén, von Döbeln, Ritzén and Wedell18} For completeness, we carried out sensitivity analyses in a subset of the follow-up (2006–2013), where we were able to restrict to glucocorticoid-only prescriptions (H02AB; see ‘Analyses’ below). We created continuous treatment periods on the assumption that any prescriptions falling within 120 days of each other within an individual belonged to the same treatment period. ^{Reference Lagerberg, Fazel, Molero, Franko, Chen and Hellner19} The treatment period started on the date of the first prescription in the period. For the last or single prescription in a treatment period, 14 days were added to the end to define the date of treatment period end. ^{Reference Lagerberg, Fazel, Molero, Franko, Chen and Hellner19} Continuous treatment periods were considered in the analyses of the medication-only cohort, and in per-protocol analyses in the indication cohorts.

Outcomes

We considered any unplanned specialist psychiatric healthcare contact where diagnosis was made for anxiety (F4), bipolar (F25.0, F30, F31, F34.0), depressive (F32, F33, F34, excluding F34.0, F38, F39) or schizophrenia-spectrum disorders (ICD-10 codes: F2, excluding F25.0). For suicidal behaviour, we considered any unplanned healthcare contact where the diagnosis was self-harm or death by suicide, of either known (X60–84) or unknown (Y10–34) intent. The date of patient admission to care was assumed to be the date of the event.

Covariates

In the medication-only cohort, all analyses were adjusted for age. Between-individual analyses were additionally adjusted for gender. In the indication cohorts, all analyses were adjusted for gender; age; year of follow-up start; highest level of attained education of individual and their parents; highest income category between individual and parents; and diagnoses at baseline (anxiety, bipolar disorder, depression, schizophrenia-spectrum disorders, substance use disorder and a history of self-harm). The only variable with missing information was highest attained education between individuals and their parents, where the prevalence of missingness was between 1 and 2% (Supplementary Tables 3 and 4). Missing values for education were treated as a separate category in the education variable.

Medication-only cohort

We used stratified Cox proportional hazards models to estimate within-individual hazard ratios. We compared the hazards of outcomes during treated and untreated periods within the same individual, ensuring that all time-invariant confounding was controlled for. ^{Reference Lichtenstein, Halldner, Zetterqvist, Sjölander, Serlachius and Fazel20,Reference Lu, Sjölander, Cederlöf, D’Onofrio, Almqvist and Larsson21} Analyses were carried out overall and stratified by: gender, age category (15–24, 25–34, 35–44 and 45–54 years) and a history of any of the outcomes at start of follow-up (past psychiatric diagnosis). We also conducted a sensitivity analysis where we excluded any individual ever prescribed an inhaled glucocorticoid (ATC code R03BA). All analyses were adjusted for time-varying age.

To investigate whether there were periods of high or low risk during glucocorticoid treatment, we carried out analyses looking at the outcome in each of the following periods since treatment start: <14, 14–119, 120–364 and >364 days, using any off-treatment period as the reference. We also considered average daily dose of glucocorticoid medication collected during treatment. We calculated this by taking the sum of the amount of medication – expressed as defined daily dose (DDD) ²² – that was dispensed over a treatment period. We then divided cumulative DDD by the length of treatment in days. We classified <0.5 DDD per day as low dose, 0.5–1.5 DDD per day as medium dose and >1.5 DDD per day as high dose. ^{Reference Einarsdottir, Ekman, Molin, Trimpou, Olsson and Johannsson23}

We further stratified by diagnosis of an autoimmune or gastrointestinal autoimmune disorder. These diagnoses were defined as being present after the first recorded diagnosis from 1 January 2001 to 31 December 2020. To investigate the impact of between-individual confounding in the medication-only cohort, we carried out between-individual analyses using Cox poportional hazards models, which were adjusted for age and gender. We used robust sandwich covariance estimation when calculating confidence intervals to account for correlation of person–time within individuals. ^{Reference Lee, Wei, Amato, Leurgans, Klein and Goel24}

Finally, to assess whether our results were impacted by the potential inclusion of non-glucocorticoid corticosteroids, we restricted analyses to the follow-up period 1 January 2006 to 31 December 2013. In these data, we had access to information up to the 5th ATC level (e.g. H02AB01), and were able to ensure that the cohort contained only dispensations of glucocorticoids.

Indication cohorts

We used pooled logistic regression models with product terms between treatment and time ^{Reference Hernán25} to estimate cumulative incidence and hazard ratios for each of the outcomes at 14, 120 and 365 days of follow-up. We accounted for baseline confounders by applying inverse probability weighting (IPW), ^{Reference Hernán, Lanoy, Costagliola and Robins26} and calculated the standardised mean difference (SMD) between confounder distributions in initiators and controls before and after weighting. We conducted both intention-to-treat and per-protocol analyses (Supplementary Table 1). In intention-to-treat analyses, individuals were assumed to have adhered to their assigned baseline treatment throughout follow-up. In per-protocol analyses, we censored individuals when they stopped adhering to the treatment strategy they were assigned at baseline. For initiators, this occurred if the individual terminated their treatment during follow-up; for controls, this occurred if the individual initiated glucocorticoid treatment during follow-up. We estimated time-varying treatment adherence weights using baseline confounders and specialist out- or in-patient healthcare contacts over follow-up. We weighted each 2-week period of follow-up by the product of time-varying and baseline IPW. Weights were stabilised and truncated at the 99th percentile. We used non-parametric bootstraps over 500 samples to estimate 95% confidence intervals. All analyses were run overall, and intention-to-treat analyses were stratified by a history of outcomes at the start of follow-up (past psychiatric diagnosis).