The data for this nationwide cohort study were collected from various registers maintained by Statistics Sweden and the National Board of Health and Welfare (Table 1). The Total Population Register (https://www.scb.se/vara-tjanster/bestall-data-och-statistik/register/registret-over-totalbefolkningen-rtb) was used to identify Swedish residents, as well as their age and sex. The Longitudinal Integrated Database for Health Insurance and Labour Market Studies (LISA) (https://www.scb.se/en/services/ordering-data-and-statistics/register/longitudinal-integrated-database-for-health-insurance-and-labour-market-studies-lisa/) provided information on education level and cohabitation status. The National Patient Register (https://www.socialstyrelsen.se/statistik-och-data/register/patientregistret/) was used to obtain diagnostic information from inpatient and outpatient care. Prescription information came from the National Prescription Register (https://www.socialstyrelsen.se/statistik-och-data/register/lakemedelsregistret/). The National Cause of Death Register (https://www.socialstyrelsen.se/statistik-och-data/register/dodsorsaksregistret/) was used to determine the date of death for deceased individuals. Register data of individuals from the different sources were linked using the personal identification number (PIN) that is unique to each Swedish resident. The study was approved by the Swedish Ethical Review Authority (decision id 2019 − 00516). No informed consent was required since no registered person was contacted, and only anonymized data was used in this register-based study.

According to international guidelines [11, 14, 22], people with an alcohol dependence according to the International Statistical Classification of Diseases and Related Health Problems, 10th Edition (ICD-10), corresponding to a severe AUD, are the main target group for pharmacological relapse prevention. The study population therefore consisted of individuals who received an inpatient diagnosis of an alcohol dependence, corresponding to a severe AUD, between 2009 and 2020. Severe AUD, i.e. alcohol dependence, was defined based on the ICD-10 code F10.2. The date of the individual’s first hospitalization with diagnosis code F10.2 (i.e. alcohol dependence) as the main diagnosis during the observational period was used as the individual’s index date, that is, the date on which a one-year follow-up period started. Additionally, individuals were only included if they were between 18 and 64 years old and resided in Sweden at the time of inclusion. To avoid potential bias, we excluded individuals with a history of prescriptions for relapse-preventive medications for AUD, antidepressants, or diagnoses of depressive disorders (ICD-10 codes F32, F33, F34.1-F34.9, F39) or anxiety disorders (ICD-10 codes F40-F41) one year prior to inclusion. The cohort included a total of 14,026 individuals. The follow-up period continued until 365 days after the index date or until the individual was hospitalized with a main diagnosis of a mental and behavioural disorder due to use of alcohol (ICD-10 code F10.0-F10.9), turned 65 years old, emigrated from Sweden, or died, whichever occurred first. We chose the length of the follow-up period because of high relapse rates after initial inpatient treatment. Previous research shows that most individuals return to substance use within the first three months following treatment, and fewer than 30% are abstinent one year after treatment [24].

Exposure to medication was divided into four categories: exposure to antidepressants, exposure to relapse-preventive medication for AUD, overlapping exposure to both, and exposure to neither. Exposure to antidepressant and relapse-preventive medications were defined using data on prescription and dispensing of each medication from the National Prescribed Drug Register. These data included the prescription date, the dispensing date, the anatomical therapeutic chemical (ATC) code, and information on the amount of dispensed antidepressants and relapse-preventive medication for AUD. We included AUD medication (e.g. disulfiram, acamprosate, naltrexone, nalmefene) (ATC code N07BB) and antidepressants from the following groups: SSRIs (ATC N06AB), serotonin and norepinephrine reuptake inhibitors or SNRIs (e.g. venlafaxine, milnacipran, duloxetine)/ serotonin–norepinephrine–dopamine reuptake inhibitors or SNDRIs (e.g. nefazodone)/ tetracyclic antidepressants (e.g. mirtazapine)/ antidepressants with other modes of action (e.g. esketamine, tianeptine, bupropion, agomelatine, trazodone) (ATC N06AX), tricyclic antidepressants or TCAs (ATC N06AA), and monoamine oxidase inhibitors or MAOIs (ATC N06AF and N06AG). Exposure periods were calculated separately for antidepressants and AUD medication. The start of an exposure was set to the date the medication was dispensed to the individual, i.e., the date the prescription was filled. To define the end of an exposure, we first identified the total amount of the specific drug that was dispensed. We then used the defined daily dose (DDD) for AUD medication and the recommended therapeutic starting dose for antidepressants (see Table S1 in the supplement) to define expected daily medication intake and determine the duration and end date of an exposure. For AUD medication, the start and end of the exposure were determined as follows:

$$\:Start\:date\:=Dispense\:date$$

$$\eqalign \right)\>} \over }\> \cr} $$

For antidepressants, the duration of the exposure was calculated by dividing the total amount of medication dispensed in mg by the recommended starting dosage for each medication (see Table S1 in the supplement). We also incorporated a 15-day delay in the start date of the exposure to antidepressants, because previous studies show that the therapeutic effects of antidepressants often have a delayed onset and that the maximum improvement occurs during the first 2 weeks after treatment initiation [19, 25]. For antidepressants, the start date of the exposure period was hence determined as follows:

$$\:Start\:date\:=Dispense\:date+15\:days$$

$$\eqalign \right)} \over \over }} \right)}} \cr} $$

When two or more exposure periods to medication with the same ATC code overlapped, we assumed stockpiling, that is, that the individual obtained additional medication before the previous supply was exhausted. The exposure period was extended by the number of days that the overlapping period lasted. When remaining doses from previous prescriptions could bridge the period to the next time the medication with the same ATC code was dispensed, we considered exposure to be continuous. Exposures to the same ATC code with gaps of up to five days in between were also considered to be continuous. Overlapping exposure to antidepressants and medication for AUD was defined as simultaneous exposure to both medication groups.

The outcome measure was the first inpatient hospitalization during the follow-up period with the main diagnosis of an alcohol-related disorder, including ICD-10 codes F10.0 alcohol intoxication, F10.1 alcohol abuse, F10.2 alcohol dependence, F10.3 alcohol withdrawal, F10.4 alcohol withdrawal syndrome with delirium, F10.5 alcohol-induced psychotic disorder, F10.6 alcohol-related amnesic syndrome, F10.7 alcohol-associated residual and delayed-onset psychotic disorder, F10.8 other mental and behavioral disorders due to alcohol, and F10.9 unspecified mental and behavioral disorder due to alcohol.

We used a Cox regression model as our primary statistical model to calculate hazard ratios (HRs) and 95% confidence intervals (CIs). The outcome was the first alcohol-related hospitalization during the follow-up period. In the model, the time-varying covariate was medication exposure (exposure to antidepressants, exposure to AUD medication, overlapping exposure to both, or exposure to neither) and follow-up time was used as the underlying time scale. To control for potential bias by differences in clinical and sociodemographic baseline characteristics, the Cox regression model was adjusted for sex, age, number of previous alcohol-related hospitalizations, cohabitation status, and education level at baseline (time-invariant covariates) (see Table 1 for a detailed description of covariates). To ensure the suitability of the Cox regression model, the proportional hazards (PH) assumption was checked by visual inspection of the scaled Schoenfeld residuals versus time with a smoothed curve superimposed to assess any trends.

To validate the robustness of the primary Cox regression model, we constructed a logistic regression model to assess the association between medication exposure and hospitalization due to AUD, independent of the specific time point of hospitalization. The logistic regression model tested the association between medication exposure and risk of hospitalization during the one-year follow-up period as a dichotomous outcome. In the logistic regression model, exposure on the index date was used to define exposure groups. Here we considered a three-day grace period. That is, individuals not dispensed medication on the index date or the following three days were counted as unexposed. A three-day grace period was chosen to accommodate a delayed dispensing, e.g. due to closure of pharmacies on weekends and public holidays. The three day period was also selected based on the data indicating that most of the dispenses occurred within this timeframe, while only a small increase of about 15% was observed for the following 4 days (i.e. first week). A three day period was also chosen to reduce the likelihood of inter-current events, which could bias the observed associations between medication exposure and hospitalization risk. Individuals who were hospitalized again within three days after index date (n = 349) or died (n = 484) during the follow-up period of one year were excluded from the analysis, yielding a sample size of 13,193 individuals. Based on their exposure on the index date, 303 individuals were assigned to the group exposed to antidepressants, 1,516 to the group exposed to AUD medication, 188 to the group exposed to both, and 11,186 individuals to the group exposed to neither antidepressants nor AUD medication. The logistic regression model was adjusted for the same covariates as the Cox regression model. To confirm the robustness of the results against different durations of the grace period, models considering a 0 and 7 day grace period were set-up and tested.