The ESTRID study Study population and design

The Epidemiological Study of Risk Factors for LADA and Type 2 Diabetes (ESTRID) is a Swedish, population-based, case–control study with incident cases of LADA and type 2 diabetes, nested within the All New Diabetics in Scania (ANDIS) register. Details on ANDIS and the study design are outlined in the Electronic Supplementary Material (ESM) Fig. 1 and ESM Methods. Sex was identified through each individual’s personal identity number. The proportions of men and women in ESTRID are fairly representative of the Swedish population, whereas people born abroad are under-represented. Participants provide health and lifestyle information through a detailed questionnaire and clinical information is available for cases through ANDIS. The analytical dataset included all incident cases of LADA (n=597) and type 2 diabetes (n=2065) along with control participants (n=2386) enrolled in ESTRID in 2010–2019. All participants provided written informed consent and the study was approved by the regional ethical review board in Stockholm (2018-1036-31, 2018-1036-32, 2023-06428-02).

Diabetes classification

Incident cases in ANDIS were classified as LADA or type 2 diabetes based on age at diagnosis (≥35 years) and analysis of GADA, measured by ELISA (RSR, UK, with sensitivity/specificity 0.84/0.98 [14]), and C-peptide levels (IMMULITE 2000, Siemens Healthcare Diagnostics Products, UK, or Cobas e601, Roche Diagnostics, Germany). Individuals with LADA were GADA positive (≥10 U/ml) and had C-peptide levels of ≥0.2 nmol/l or ≥0.3 nmol/l (IMMULITE/Cobas). Individuals with type 2 diabetes were GADA negative and had C-peptide levels of >0.60 or >0.72 nmol/l (IMMULITE/Cobas). HOMA-IR and HOMA-B were calculated based on fasting C-peptide and glucose levels [15].

The HUNT study Study population and design

In the Nord-Trøndelag region of Norway, the entire population aged ≥20 years has been invited to participate in the Trøndelag Health Study (HUNT) on four occasions between 1984 and 2019. More women than men participated in the HUNT surveys, and the highest participation rates were among those aged 50–79 years [16]. Sex was determined based on Norwegian national identification numbers. Information on ethnicity was not collected in the HUNT surveys. Participants respond to questionnaires and undergo clinical examination and blood sampling. We conducted a nested case–control study among cases and matched control participants in HUNT4, as well as HUNT3 if the year of diagnosis of cases was ≥2006 (prescription data available in Norway from 2004). The analytical sample included incident cases of LADA (n=82) and type 2 diabetes (n=1279) and matched control participants (n=2050). Details of the study design are outlined in ESM Methods. The Norwegian Data Protection Authority and the Regional Committee for Medical and Health Research Ethics approved the study (REK 140824) and all participants provided informed consent.

Diabetes classification

Cases were identified through a questionnaire including the question ‘Have you had or do you have diabetes (yes/no)?’. GADA was assessed at a median of 5 years after diagnosis and used with self-reported age at diagnosis to determine diabetes type. GADA analysis was carried out at the Hormone Laboratory, Oslo University Hospital, Norway, either by immunoprecipitation radioligand assay (Novo Nordisk, Denmark; HUNT3) or by ELISA (RSR; HUNT4). The sensitivity and specificity of the immunoprecipitation assay were 0.64/1.00 (Islet Autoantibody Standardization Program 2003) [17] and of the ELISA assay were 0.84/0.98 (Islet Autoantibody Standardization Program 2020; P. M. Thorsby, Hormone Laboratory, Oslo University Hospital, Oslo, Norway, personal communication). Cases (≥35 years) were classified as LADA if positive for GADA (≥0.08 antibody index) in HUNT3; ≥10 U/ml in HUNT4); otherwise, they were classified as type 2 diabetes. C-peptide data from the time of diagnosis were not available in HUNT.

Register linkages

In ESTRID, we retrieved information on antibiotic dispensations (Anatomical Therapeutic Chemical [ATC] codes J01, A07AA09, J04AB02, P01AB01) from the National Prescribed Drug Register (NPDR) [18]. HUNT participants were linked to the Norwegian Prescription Database (NorPD) from which we retrieved information on antibiotic dispensations according to ATC code J01. These registers record all prescriptions dispensed at pharmacies since 2005 and 2004, respectively. Diagnoses for comorbidity adjustment (ESTRID only) were identified from the Swedish National Patient Register (NPR) [19] and the Scania Healthcare Register (SHR) [20].

Antibiotic use

Antibiotics were categorised into broad or narrow spectrum or any type (broad and/or narrow spectrum) [21] (ESM Table 1). Phenoxymethylpenicillin (ATC code J01CE02) was the most common antibiotic dispensed in both ESTRID and HUNT. Dispensations (number and % of total dispensations) of the most dispensed antibiotics 1–5 years prior to the index date are listed in ESM Table 2. Exposure windows were defined as 0 to <1, 1–5, 6–10 and 0–10 years prior to the index date. In ESTRID, the index date was set to the date of diagnosis (cases) and date of participation (control participants). In HUNT, the index date for cases and matched control participants was set to the year of diabetes diagnosis of the cases. Within the exposure windows, the different types of antibiotics were analysed categorically and continuously as number of dispensed prescriptions (0, 1–2, 3–4, ≥5) or as exposure duration, that is, as consecutive days per longest course (0, 1–14, ≥15) or as cumulative exposure (0, 1–19, ≥20 days [20–49, ≥50 days in the 0–10 years exposure window]), corresponding to the total number of exposed days over the period of interest. Exposed days were calculated as dispensed quantity × defined daily dosage (DDD), which is ‘the assumed average maintenance dose per day for a drug used for its main indication in adults’ [22]. Exposed days were calculated only in ESTRID, as information to calculate this was lacking in HUNT.

Covariates

The covariate selection is illustrated in a directed acyclic graph (ESM Fig. 2). Age at participation and sex were based on or derived from national population registers. BMI (kg/m2) was based on anthropometric measurements in HUNT and self-reported weight and height in ESTRID. Smoking status (never/former/current), physical activity level (four categories ranging from sedentary to high), education level (primary school/upper secondary school/university) and family history of diabetes (yes/no) were self-reported. The percentage of missing values of covariates was <2% in both ESTRID and HUNT. We calculated high-dimensional propensity scores (hd-PS) [23] in ESTRID based on ICD-10-coded diagnoses (https://icd.who.int/browse10/2019/en) and ATC-coded dispensations from the registers, to adjust for potential confounding from comorbidity (see ESM Methods).

Statistical analysis

Two-sided p values, calculated using an unpaired Student’s t test for means (±SDs) of normally distributed variables, Kruskal–Wallis test for medians (IQRs) of non-normally distributed variables, and χ2 test for proportions, were used to assess differences in baseline characteristics between the study groups.

Within the exposure windows 0 to <1, 1–5, 6–10 and 0–10 years prior to the index date, the associations between antibiotic dispensations (one or more vs none; one to two, three to four or five or more vs none; and per dispensation) and LADA/type 2 diabetes were analysed by logistic regression conditioned on participation time (ESTRID) and index year and sex (HUNT), estimating ORs with 95% CIs. Model 1 was adjusted for age and sex (matching variable in HUNT), whereas model 2 was additionally adjusted for BMI, smoking status, physical activity level, education level and family history of diabetes. Imputation was performed on missing values of the covariates based on the median values, and an indicator variable for missingness was included in the models. Study-specific estimates were meta-analysed using fixed-effects inverse-variance weighting (ORpooled). All analyses were performed in SAS 9.4 (SAS Institute, USA). Results from the fully adjusted model (model 2) are presented.

Sensitivity analyses

To account for comorbidity and residual confounding, we adjusted the analyses in ESTRID for age, sex, ethnicity (born in Sweden: yes/no) and a continuous hd-PS considering all other covariates in model 2 in addition to comorbidity and elevated blood glucose. In ESTRID, we could also analyse the association between consecutive and cumulative antibiotic exposure and risk of LADA or type 2 diabetes. We further distinguished between LADA cases with high (GADAhigh) and low (GADAlow) GADA levels based on the median value due to the highly skewed distribution (ESM Figs 3 and 4). Finally, we performed the same analysis with dermatological drugs (ATC codes D01, D04, D05, D10) as a negative control exposure.