ADDITION–Cambridge (ISRCTN86769081) is a pragmatic cluster-randomised trial comparing multifactorial intervention vs routine care among people with screen-detected diabetes from 49 general practices (GPs) in eastern England [19]. The present study is an observational analysis of the trial cohort. A validated risk score was used to identify eligible individuals aged 40–69 years who were at high risk of diabetes, using the electronic records of participating GPs [20]. From 2002 to 2006, 33,539 high-risk individuals were invited to attend a stepwise screening programme; 24,654 (74%) took part [21]. During screening, 867 participants were diagnosed with type 2 diabetes using the 1999 WHO criteria [22] and all consented to enrol in the study. The GPs were cluster-randomised to intensive treatment (n = 26) or routine care (n = 23). In the routine care group, GPs were advised to follow current UK guidelines for diabetes management [23,24,25]. Intensive treatment included more frequent consultations, provision of educational materials and GP-based academic-detailing sessions encouraging earlier use of medication to improve control of risk factors. The intervention did not include behavioural treatment. Written informed consent was obtained from all participants and ethics approval was obtained from the research ethics committees in Cambridge (ref. 01/063), Huntingdonshire (ref. 00/609), Peterborough and Fenland (ref. P01/95), West Essex (ref. 1511–0103), North and Mid Essex (ref. MH395 MREC02/5/54), West Suffolk (ref. 03/002), and Hertfordshire and Bedfordshire (ref. EC03623) Local Research Ethics Committees and the Eastern Multi-Centre Research Ethics Committee (ref. 02/5/54).
Measurements
Anthropometric, biochemical, clinical and questionnaire-based measures were taken at the time of diabetes diagnosis (baseline) and after 1 and 5 years [19]. Sociodemographic information (age, sex, occupation and ethnicity), smoking status and prescribed pharmacological treatment were self-reported via standardised questionnaires. Information on pharmacological treatment at 5 years was supplemented using GP electronic records. Socioeconomic status (SES) was defined according to the Registrar General’s occupation-based classification: ‘professional, managerial and technical’, ‘skilled-manual and non-manual’ and ‘partly skilled or unskilled’ [26]. Clinical and anthropometric measures were performed by trained staff, according to standard procedures [19].
CVD and mortality outcomes
The outcomes of interest were CVD events and all-cause mortality. The composite CVD outcome included cardiovascular mortality, non-fatal myocardial infarction, non-fatal stroke, non-traumatic amputation, and revascularisation (both invasive cardiovascular and peripheral vascular procedures). Incidence of CVD events and mortality was ascertained from the date of diabetes diagnosis until 31 December 2014. Participants were flagged using National Health Service (NHS) patient numbers for mortality surveillance by the Office for National Statistics. Possible CVD outcomes were identified via searches of GP notes, hospital discharge summaries, hospital notes, electrocardiograms, laboratory results, death certificates, autopsies and the Myocardial Ischaemia National Audit Project (MINAP) [27]. All events were independently adjudicated using standardised case report forms.
Statistical analysis
Of the 867 adults enrolled in the ADDITION–Cambridge study, we were unable to include 137 for whom we could not determine weight change, as they were missing data on weight measurements at baseline or 1 year. We also excluded individuals who had a CVD event in the first year in the study (n = 5), as this was when weight change was measured. Therefore, this study included 725 participants with 99 CVD events. Mortality analyses also excluded deaths occurring within 1 year after weight change was assessed (n = 2), as weight loss might have been due to underlying disease; this left a total of 95 all-cause mortalities during the study period.
We assessed predictors of missing weight information by comparing distributions of factors measured at baseline between individuals who were and who were not missing weight measurements.
The proportion of weight change was determined by subtracting weight at baseline from weight at 1 year, then dividing by weight at baseline. We defined weight change categorically as >2% gain, maintained weight (≤2% gain or <2% loss), ≥2% to <5% loss, ≥5% to <10% loss, and ≥10% loss. These categories were chosen to examine the recommended weight loss target of 5–10% [28, 29] and compare it against greater or lesser weight loss, while separating those who gained weight from those who maintained their baseline weight.
Weight change, CVD events and all-cause mortality
We used Cox proportional hazards regression to estimate HRs for category of weight change, 10 year incidence of CVD events and 10 year all-cause mortality. Due to the small numbers of CVD events among participants who lost the most weight, we combined the ≥5% to <10% loss and ≥10% loss categories.
The timescale was the number of days since diabetes diagnosis. Person-time at risk began 1 year after diabetes diagnosis and ended at the time of the event, death or 31 December 2014. We assessed the proportional hazards assumption by modelling an interaction term between the natural log of time and each covariate, which indicated no departures from proportional hazards. All models adjusted for confounders identified using a directed acyclic graph [30]: age at baseline (continuous), sex (female, male), trial group (intensive treatment, routine care), baseline occupational SES (‘professional, managerial and technical’, ‘skilled-manual and non-manual’ and ‘partly skilled or unskilled’), BMI at baseline (continuous, coded as a quadratic term), cigarette-smoking status at 1 year (current, former, never), and use of antihypertensive (yes, no), glucose-lowering (yes, no) and lipid-lowering (yes, no) medication at 1 year. Clustering of individuals within GPs was accounted for using a robust cluster variance estimator.
To assess heterogeneity of associations by age, we modelled an interaction term between weight change and age at diagnosis. As past research has shown negative impacts of weight loss on mortality among older adults [31,32,33], we performed separate analyses to estimate the associations of interest among those aged ≥65 years at the time of diabetes diagnosis. We were unable to stratify by age at diagnosis <65 years owing to the small number of cases in this age group.
In a sensitivity analysis, we used multiple imputation by chained equations [34] to estimate associations among the full cohort including the 137 individuals with missing weight information. The imputation models included covariates for weight change category; sex; occupational SES; baseline BMI; smoking; treatment group; antihypertensive, glucose-lowering and lipid-lowering medication use at 1 year; outcome status; and the Nelson–Aalen estimate of cumulative hazard. We generated 20 imputed datasets and used these to estimate HRs in the full cohort. Separate imputation models were fit where the outcome was all-cause mortality. We also assessed the robustness of our results to changes in weight after the first year in the study by adjusting for weight change from year 1 to year 5. Additional sensitivity analyses excluded individuals with a history of myocardial infarction or stroke prior to diagnosis of diabetes (n = 77); separate analyses excluded revascularisation and amputations from the composite CVD outcome (n = 37) in order to estimate HRs for a composite of stroke, myocardial infarction or cardiovascular death. To evaluate the association of weight change and CVD occurring in the presence of the competing risk of non-CVD mortality, we estimated subdistribution HRs [35].
Weight loss and CVD risk factors
We used linear regression models to estimate the associations between percentage weight loss in the year following diabetes diagnosis and CVD risk factors (HbA1c, systolic and diastolic BP, triacylglycerols, total cholesterol and LDL-cholesterol) at 1 year and 5 years. Percentage weight change categories were defined as above.
The adjustment set was the same as in the proportional hazards models, except that each model was adjusted for relevant medication use in the year the outcome was measured (e.g. glucose-lowering agents for HbA1c, antihypertensive medication for BP, and lipid-lowering medication for lipid levels), rather than adjusting for each type of medication use. Separate analyses were stratified by medication use at 1 year and 5 years to assess possible heterogeneity of the associations. Sensitivity analyses were conducted: (1) using multiple imputation by chained equations [34] to estimate the associations of percentage weight change and the CVD risk factors among the full cohort; [2] by adjusting for baseline values of the risk factors.