Our data suggest that, in routine practice, people with type 2 diabetes and initially poor glycaemic control have a similar risk of major macrovascular events (MI, acute coronary syndrome, stroke or death) in the medium term regardless of what type of insulin is started. The mean follow-up was 3.6 years although events in the first year were not included. Fewer patients started insulin after three oral agents resulting in a small number of outcomes and upper 95% confidence intervals close to two. The sensitivity analysis comparing macrovascular disease plus angina in premix insulin versus basal analogue after two OGLDs had a hazard ratio of 1.49 and a lower 95% confidence interval of 0.99 although the same comparison after three oral agents did not show a difference in risk between the insulins with a hazard ratio of 0.76 (95% confidence interval 0.39, 1.51).
Rates of incident microvascular conditions were similar in the treatment groups except in the comparison between basal insulin and NPH after a regimen of three oral agents. The increased rate after NPH is unlikely to be due to better short-term glycaemic control as there was no difference in baseline HbA1c
levels but a statistically significant greater improvement in glycaemic control from 6 to 11 months after starting NPH rather than basal insulin. A sharp increase in both testing for diabetes complications 
and diagnosis of some conditions has been reported after the introduction of the 2004 UK Quality and Outcomes Framework which provided incentives for routine surveillance of people with diabetes. An increased microvascular risk could be due to NPH being started before 2004 more frequently than was the new basal insulin. However, inclusion of year of first insulin in the adjusted analysis should have accounted for this difference. While no difference in the rate of microvascular disease was found after escalation from the larger number of patients who started insulin after two oral agents, there were anticipated differences between those who started insulin from two rather than three oral agents, such as similar HbA1c
after a shorter duration of diabetes, which might impact on the development of these complications. This finding therefore warrants further investigation. Differences in weight gain and glycaemic control were identified between insulin groups in the first year of treatment. Starting basal insulin was associated with less reduction in HbA1c
than other insulins after three oral agents but less weight gain than pre-mix after two oral agents.
No observational studies or clinical trials were found which specifically aimed to compare the rate of macrovascular disease between first insulins. The Treating to Target in Type 2 diabetes (4-T) clinical trial found no significant difference in death rate between basal, biphasic and prandial insulin regimens over 3 years, but a higher rate of death from cardiovascular disease in the prandial than the biphasic and basal groups 
. A five-year trial of insulin glargine versus NPH insulin treatment in patients with type 2 diabetes mellitus showed no evidence of a greater risk of the development or progression of diabetic retinopathy between treatments 
. A smaller decrease in HbA1c
in basal insulin versus NPH treated patients was reported in that trial and in the 4-T study when basal insulin (detemir) was compared to both biphasic and prandial treatment 
levels later converged in the 4-T study after usual addition of other insulins, 
and in the Treat to Target study more glargine users eventually achieved a target of HbA1c
≤7.0% (53 mmol/mol) without nocturnal hypoglycaemia 
Less weight gain in the basal group in routine care was consistent with several clinical trials of basal compared to other insulins 
. Lower weight gain with basal insulin is believed to be a consequence of lesser hyperinsulinaemia in the post-prandial period, with consequent less hunger as a result of lower plasma glucose concentration, and perhaps concomitantly less frequent need for correction of hypoglycaemia with carbohydrates. We were not able to compare the rates of hypoglycaemic events between groups or in relation to HbA1 c
as primary care records only capture episodes known to the physician rather than events treated by family, paramedics or the patient themselves.
Insulin was often started when glycaemic control was very poor, with mean HbA1c
9.9% (85 mmol/mol) (after exclusion of 8% with baseline HbA1c
<7.5% (58 mmol/mol)). This mean is higher than many clinical trials, which often have an upper as well as lower inclusion limit 
, demonstrating that trial populations may not always be representative of routine practice. It is consistent with the ad-hoc observational CREDIT study conducted in Europe, Canada and Japan 
. Overall glucose lowering was good (means 1.0–1.7%) after 12 months and, unlike interventional studies, could not be biased by selection of patients or physicians. Nevertheless, because starting HbA1c
levels were high, control to target was not achieved in the majority of people even after one year.
This study has a number of limitations. While we were able to adjust for many potential confounding variables, severity of prevalent vascular disease (such as angina), prior long term glycaemic control, socioeconomic status and ethnicity were not routinely available on all patients in this observational dataset in the study years and recorded duration of diabetes has not been validated. While these variables are likely to be non-differential between the insulin types, we cannot exclude the presence of some residual confounding. Confounding by indication should be decreased as all subjects initiated insulin when they had poor glycaemic control, although different insulins may be started in different clinical circumstances such as ambulatory care versus in-patient treatment. We chose to include all-cause mortality rather than cardiovascular mortality because the recording of cause of death has been shown to be incomplete in general practice databases 
so non-cardiovascular events could not be reliably excluded. HbA1c
may have been measured more frequently in those whose glycaemic control was poor resulting in higher mean levels during follow-up than were actually the case. As all patients should have levels measured after insulin introduction there should be no difference between insulin types. The finding that the number of patients who have a value measured remains relatively consistent over time, across treatment groups, is also reassuring. The first year of follow-up was excluded from all groups in the analysis of vascular events as the distribution of insulin use in terminally ill patients may not be even, causing bias, and any effect of the choice of therapy on our outcomes is likely to be in the medium- to long-term. The design was an intention-to-treat type analysis so we did not censor follow-up at the next change of glucose lowering treatment. As a result, the study did not investigate the association of individual insulins and cardiovascular outcomes but rather the mid-term consequences of one therapeutic decision when starting insulin. Later changes in treatment could be expected to relate to the efficacy and problems of the first insulin regimen.
Overall, while the number of patients lost due to inclusion/exclusion criteria appears high, the criteria for the initial data cut were untypically wide due to the two stage identification process (OGLD therapy and then insulin). For example, the requirement for a year of data recording pre- and post-insulin resulted in the removal of 29% of the total potential population but allowed time for recording of information on prevalent events as well as laboratory measurements. Many of those who had no baseline HbA1c also had less than one year of data pre-insulin. Many of the remaining exclusions occurred because the HbA1c could not be interpreted rather than because it was not measured. The study covered a time of HbA1c unit change in the UK.
In conclusion, in routine clinical practice no significant difference in the rate of incident macrovascular disease was found between first insulin type started during poor glycaemic control over a period of 1 to 3.6 years. The increased risk of microvascular disease between users of NPH and basal insulin was found in those who escalated from three but not two agents and so warrants further investigation. First time use of basal insulin was associated with a smaller increase in weight gain and a smaller decrease in HbA1c in the first year compared to other insulins in some treatment groups. Glycaemic control was poor at baseline and, while mean HbA1c decreased during insulin therapy, control remained poor across all groups.