Participants and comparisons
A total of 5102 newly diagnosed patients with type 2 diabetes, defined as fasting plasma glucose above 6 mmol/l on two occasions, aged 25-65 years (mean age 53) were recruited in 23 centres. After initial dietary treatment, 4209 patients had fasting plasma glucose concentrations of 6.1-15 mmol/l without symptoms of hyperglycaemia. Of these, 342 overweight patients were randomised to metformin, leaving 3867 patients who entered the main randomisation and were allocated either to conventional management (mainly through diet, 1138 patients), or to intensive management with insulin (1156) or sulphonylureas (1573). The aim of the conventional policy was to maintain patients free of diabetic symptoms and with a fasting plasma glucose concentration below 15 mmol/l, whereas the intensive policy was aimed at a fasting plasma glucose concentration below 6 mmol/l. All patients in the main randomisation were included in this economic evaluation. The median follow up period to death, the last known date at which survival was known, or to the end of the trial was 10 years. The main clinical end points analysed were death or the development of diabetic complications, including coronary heart disease, cerebrovascular disease, amputation, laser treatment for retinopathy, cataract extraction, and renal failure. All analyses and comparisons were performed on an intention to treat basis.
Type of evaluation and perspective
We performed an incremental cost effectiveness analysis in which the net costs and net effectiveness of intensive compared with conventional management were calculated and expressed as a ratio. The main perspective of the economic evaluation was that of healthcare purchasers. Only direct health service costs were analysed. These costs covered conventional and intensive treatments, visits to diabetic clinics and tests, and treatment of diabetic complications, including inpatient stays and outpatient health care. We also compared the costs of conventional policy with the insulin and sulphonylurea intensive policies separately.
For each patient, data were collected at three monthly clinic visits on the doses of all drugs used for treating diabetes (insulin, sulphonylureas, metformin); the number of home blood glucose tests; the dose of the three main drugs for hypertension (captopril, atenolol, nifedipine); whether the patient was taking diuretics, methyldopa, calcium channel blockers, vasodilators, or other antihypertensive drugs; and whether the patient was taking aspirin, antidepressant drugs, hormone replacement therapy, anxiolytics, or any other drugs. When treatment doses were not recorded, missing values were replaced by extrapolation from adjacent values for that patient. Last observation carried forward was used to impute missing data when necessary.
Data on the date and duration of each hospital admission were collected at every clinic visit. These were coded by using ICD-9 and ICD-10 classifications for prime cause of admission and Office of Population Censuses and Surveys (OPCS-4) codes for all procedures undertaken. In addition, a separate record was maintained of all angiograms, angioplasties, and bypass grafts for coronary or peripheral vascular disease. All hospital admissions were also allocated by two clinicians to one of 40 national standard specialty codes. Missing values for hospital lengths of stay were replaced with the mean value for all patients in that specialty.
Data on non-hospital and outpatient resource use were collected from all patients in the trial through a questionnaire distributed at routine clinic visits between January 1996 and September 1997 and by post to those who did not attend a clinic during this period. This questionnaire collected information on all home, clinic, and telephone contacts with general practitioners, nurses, chiropodists, opticians, dieticians, and eye and other specialists over the previous four months. Retrospective data capture from patients may underestimate resource use, but this is unlikely to introduce systematic bias when analysed by allocation. These cross sectional data were analysed by multiple regression to estimate for each patient the annual non-hospital resource use adjusted for significant variables including age, sex, body mass index, duration of diabetes, and time from a non-fatal diabetes related end point.
Unit costs for all resources used by trial patients were obtained from national statistics and from centres participating in the trial (table ). These unit costs were combined with the resource volumes to obtain a net cost per patient over their time in the trial. Mean net costs and associated 95% confidence intervals were calculated for each arm of the study. Costs are reported both undiscounted and in net present values using the UK Treasury approved 6% annual discount rate.8
All costs are reported in 1997 values (£s).
Main unit costs and sources of conventional and intensive management of type 2 diabetes
All participants in the study attended clinics every three months; the interval was increased to four months in the later years of the study. However, outside a trial it is likely that the frequency and type of visits would be different, particularly for conventional policy patients. To illustrate this, we conducted a complementary analysis in which visits for conventional or intensive treatment are costed to reflect likely standard clinical practice rather than that dictated by the trial protocol. This assumes that the observed differences in complications between trial arms would be maintained with the different pattern of visits. Table outlines the likely pattern of standard practice for conventional and intensive management based on the opinion of primary care and specialist clinic staff in the participating centres. Each patient's actual annual trial visit costs were replaced by the estimated standard practice annual visit cost depending on allocation and the associated probability of the patient receiving insulin or diet and tablet treatment. The costs of other patterns are also considered in sensitivity analyses.
Assumed annual real life visit and blood glucose test schedules for conventional and intensive management of type 2 diabetes
Diabetes related end points were defined as in the clinical trial.6
The trial showed that intensive blood glucose control significantly reduced (P=0.029) the risk of any diabetes related end point by 12% but did not significantly reduce diabetes related deaths or all cause mortality. Consequently, the current analysis measures outcomes in terms of time to first event (myocardial infarction, congestive heart failure, stroke, renal replacement therapy, amputation, cataract extraction, vitreous haemorrhage, or death from any cause).
We used a conservative estimate of time to first event—that is, we assumed that no treatment effects continue beyond the end of the trial. For patients with no event observed during the trial, we used simulation from a parametric model to estimate the time from study closure (or withdrawal from study) to first event. Basic bootstrap confidence intervals were calculated for all simulation results based on 5000 non-parametric bootstrap iterations, each iteration being averaged over 20 runs of the simulation for stability.
For comparison we also estimated time to first event assuming that treatment effects continue beyond the end of the trial. We refer to this as an unbiased estimate, because it attempts to follow the size of the treatment effects that were observed in the trial, whereas the conservative estimate described above forces all treatment effects to zero at study closure. All cost effectiveness results are based on the conservative estimate of effectiveness. Since cost data are available for the duration of the trial only, the appropriate effectiveness measure is the one that allows treatment effects during the trial only.
The model used in the simulation described above was a twofold competing risk model. In the first component, risk of a diabetes related event increases with age at diagnosis of diabetes and with duration of diabetes. In the second component, risk of other death (any death except myocardial infarction, sudden cardiac death, stroke) increases with the age of the patient. For simplicity, the last category includes several causes of death that may be considered diabetes related (hyperglycaemic or hypoglycaemic episodes, renal death, and death from peripheral vascular disease) but occurred too infrequently to be modelled individually (total 26 deaths). The methods used to test the model's validity are given on the BMJ 's website.
All comparisons were carried out on an intention to treat basis. All results are reported as mean values with standard deviations; mean differences are reported with 95% confidence intervals. When descriptive statistics suggested the possible presence of skewness, 1000 bootstrap replications of the original data were performed and the resulting means, mean differences, and intervals were compared. For all reported costs, parametric confidence intervals for the cost differences were compared with the bootstrap confidence intervals and were found to be robust; parametric confidence intervals are therefore reported. Confidence intervals for the mean cost effectiveness ratios were calculated by Fieller's method.9,10
The effect of assumptions on our main results was examined by sensitivity analyses. All data were analysed with SPSS 8.0 and Microsoft Excel 97; the modelling work was carried out in C language.