This prospective, nested case−control study included twins from the nationwide Swedish Twin Registry (STR), which was started in the 1960s . In 1998–2002, all living twins (≥40 years of age, i.e. born in 1958 or earlier) in the registry were invited to participate in the Screening Across the Lifespan Twin study (SALT), a full-scale screening that gathered data on an extended set of variables via computer-assisted telephone interview. Out of 44,919 twin individuals eligible for the interview, 5009 who had not reached the age of 60 by the end of available follow-up on 31 December 2014 (born after 1954), 417 who died by 31 December 2014, 773 who had CBD before the age of 60 years, 4210 who had type 2 diabetes with onset age <40 years or ≥60 years, 310 who had type 1 diabetes, and 1114 who had transient ischaemic attack were excluded. Finally, 33,086 individuals remained for the current analyses (Fig. 1).
Information on demographics (age, sex and educational attainment), lifestyle (smoking, alcohol consumption), anthropometric measures (weight and height), zygosity and medication use (including treatment for type 2 diabetes) was obtained from the SALT survey . Education was defined by the maximum years of formal schooling attained, and dichotomised into <8 vs ≥8 years. BMI was calculated as weight (kg) divided by height squared (m2), and categorised as <20 kg/m2 (underweight), 20–24.9 kg/m2 (normal weight), 25–29.9 kg/m2 (overweight), and ≥30 kg/m2 (obese). Smoking status was categorised as never smoked, former smoker and current smoker. Alcohol consumption was dichotomised as never consumed alcohol vs past/current consumer of alcohol. Marital status was defined as married/cohabitating vs single (including divorced).
Information on history of type 2 diabetes, CBD, hypertension and heart disease (including coronary heart disease, cardiac arrhythmias and heart failure) was derived from the National Patient Registry (NPR), which covers all inpatient care in Sweden since the 1960s up to the end of 2014. Each medical record in the NPR included up to eight discharge diagnoses recorded as International Classification of Disease (ICD) codes. The seventh revision (ICD-7) was used through to 1968, the eighth revision (ICD-8) from 1969 through to 1986, the ninth revision (ICD-9 [www.icd9data.com/2007/Volume1]) from 1987 through to 1996 and the tenth revision (ICD-10 [http://apps.who.int/classifications/icd10/browse/2016/en]) since 1997 through to the end of available follow-up in 2014.
Informed consent was required from all participants. The data collection procedures were approved by the Regional Ethics Committee at Karolinska Institutet, Stockholm, Sweden, and the Institutional Review Board of the University of Southern California, USA.
Ascertainment of diabetes
Both type 1 and type 2 diabetes were ascertained based on self- and informant-reported history of type 2 diabetes, glucose-lowering medication use, or NPR (ICD-7 code 260, ICD-8 and -9 code 250, and ICD-10 codes E10–E14). The age at type 2 diabetes onset was estimated according to the earliest recorded date of type 2 diabetes in the NPR or the self-reported date of type 2 diabetes onset available in SALT. Midlife type 2 diabetes was defined as type 2 diabetes onset at age 40–59 years. Information on treatment of type 2 diabetes was collected from the SALT survey and the Swedish Drug Registry (The Anatomical Therapeutic Chemical code A10) since 2006. Treatment of type 2 diabetes was categorised as diet, oral glucose-lowering drugs, insulin and combined oral glucose-lowering drugs and insulin.
Assessment of CBD
Information on CBD diagnosis (ICD-7 codes 330–334, ICD-8 codes 430–438, ICD-9 codes 430–437, ICD-10 codes I60–I68) was derived from NPR records between 1960 and 2014. The CBD subtypes in the current study included: (1) cerebral infarction; (2) occlusion and stenosis of precerebral or cerebral arteries not resulting in cerebral infarction (termed as occlusion of cerebral arteries below); (3) subarachnoid haemorrhage; (4) intracerebral haemorrhage; and (5) unspecified CBD . The age of CBD onset was estimated according to the earliest recorded date of the CBD diagnosis in NPR.
The characteristics of the study population by midlife type 2 diabetes or CBD were compared using χ2 tests for categorical variables, unpaired Student’s t test for continuous variables with normal distribution, and Mann–Whitney U test for continuous variables with non-normal distribution. Generalised estimating equation (GEE) models were used in unmatched case−control analyses, controlling for the clustering of twins within a pair. In unmatched case−control analysis, the basic-adjusted models were adjusted for age, sex and education. The multi-adjusted models were further adjusted for BMI, smoking, alcohol consumption, marital status, hypertension and heart disease. Data for the co-twin matched case−control study were analysed using conditional logistic regression, in which twin pairs who were discordant for outcome. In the co-twin control design, the disease-free co-twin (in both monozygotic and dizygotic twin pairs) is used as a control for the diseased twin. Using discordant twin pairs is more informative than using unrelated case−control samples, as cases and controls matched for genetic background and familial environmental factors such as fetal environment, maternal smoking or childhood socioeconomic status [28, 29]. In the co-twin control analysis, the basic-adjusted models were adjusted for sex and education. The multi-adjusted models were further adjusted for BMI, smoking, alcohol consumption, marital status, hypertension and heart disease.
Logistic regression was used to test the difference in ORs from the GEE model and conditional logistic regression by examining the difference in midlife type 2 diabetes between unmatched and co-twin controls . If an OR for the observed association in the unmatched case−control analysis becomes strengthened or attenuated (or even disappears) in co-twin control analyses, and the difference in ORs from the GEE and conditional logistic regression was significant, genetic and/or shared familial environmental factors are likely to play a role in the association [29, 31]. In contrast, if the OR in conditional logistic regression remains similar to that from the GEE, and the difference in ORs was not significant, then the confounding by genetic or shared familial environmental factors in the observed association is small or null [24, 30, 32].
Age, sex, education, BMI, smoking, alcohol consumption and marital status were considered as potential confounders. Missing values on education (n = 1222), smoking (n = 1182), alcohol consumption (n = 1253), BMI (n = 1801) and marital status (n = 757) were imputed using Rubin’s rule for pooling estimates to obtain valid statistical inferences . All statistical analyses were performed using SAS statistical software version 9.4 (SAS Institute, Cary, NC, USA) and IBM SPSS Statistics 20.0 (IBM, New York, NY, USA).