Variation in Event Rates in Trials of Patients With Type 2 Diabetes

Early in the course of type 2 diabetes, mild elevations of glucose levels provoke so few symptoms that the date of disease onset may be difficult to establish, and little or no drug treatment may be necessary to minimize symptoms such as polyuria. Diabetes is nonetheless associated with the onset and progression of disabling microvascular and macrovascular disease complications. The strength of these associations with vascular disease risk is pronounced and depends on the duration of diabetes and the blood glucose level. The prevention of complications rather than the minimization of symptoms has usually served as the rationale for tight control of glucose or glycated hemoglobin.

Drug effects on putative surrogate end points such as glycated hemoglobin, however, may not track well with their effects on the incidence of cardiovascular events, and such end points rarely capture off-target effects of drug treatments.¹ The typical inability to establish a reliable and valid risk-benefit profile on the basis of surrogate end points has encouraged the conduct of large clinical trials. The principal purpose of the large prevention trials in patients with type 2 diabetes is to assess directly the risks and benefits of various drug treatments on cardiovascular outcomes.

In the last several years, 5 major trials of treatments for patients with type 2 diabetes have been published.^{2 - 6} Two trials evaluated the addition of thiazolidinedione and 3 trials evaluated intensive glucose control vs standard therapy. The occurrence of myocardial infarction (MI), a well-standardized outcome, was part of the composite outcome in all 5 trials. For unexplained reasons, however, the MI event rate in the control groups of these 5 trials, which enrolled substantially similar patients, varied by a factor of 4 (Table).

The Table summarizes the MI event rates and the baseline characteristics of the patients in the 5 trials as well as in a general, population-based sample from the Atherosclerosis Risk in Communities Study (ARIC).^{7 - 8} The event rates were highest in the Prospective Pioglitazone Clinical Trial in Macrovascular Events (PROACTIVE) trial,⁶ which recruited only patients with macrovascular disease. The event rates were intermediate and similar in the 2 US federally funded studies: Action to Control Cardiovascular Risk in Diabetes³ (ACCORD) and Veterans Affairs Diabetes Trial⁴ (VADT). The event rates were low in the industry-sponsored Action in Diabetes and Vascular Disease: Preterax and Diamicron Modified Release Controlled Evaluation⁵ (ADVANCE) and Rosiglitazone Evaluated for Cardiovascular Outcomes and Regulation of Glycemia in Diabetes² (RECORD) trials, which were each conducted at hundreds of centers in more than 20 countries. Many of the RECORD centers were outside of the United States and western Europe.

Variation in MI event rates among trials does not appear to be explained by differences in baseline risk factors (Table). For instance, the control groups of ACCORD and RECORD have similar mean ages, mean levels of fasting glucose, proportions of smokers, and patients with cardiovascular disease. Patients in the ACCORD trial have a longer duration of diabetes and a higher level of glycated hemoglobin, whereas patients in the RECORD trial have higher levels of diastolic blood pressure and low-density lipoprotein cholesterol. For these 2 studies, the event rates in the control groups differed by a factor of 3, a difference greater than can be accounted for by the variations in distributions of known risk factors.

The baseline data from the white patients in the ARIC cohort study together with the incidence of MI in those aged 55 to 65 years at baseline provide an interesting anchor. Indeed, the event rate in the RECORD control group (4.6 per 1000 person-years) is almost the same as the event rate in the ARIC population-based sample (4.3 per 1000 person-years), only 8% of whom have diabetes. Moreover, the between-trial variation in event rates tends to be much larger than the within-trial differences between treatments.

The low event rate in RECORD was not isolated to the MI end point. The primary outcome in RECORD was cardiovascular hospitalization or cardiovascular death. With an anticipated annual event rate of 11% in the control group, the investigators had expected about 1000 events in the control group over 5.5 years of follow-up; but they observed only 323 events.² In other words, about 677 of expected events in the control group were missing—a number that far exceeds the 272 control patients who were lost to follow-up or had missing end point information. Davis and Whelton⁹ have noted a similarly extreme variation in event rates in several hypertension trials, not only for the outcome of MI but also for a variety of composite end points.¹⁰

Even with common end-point definitions across studies, many factors can contribute to variations in event rates. Risk factors other than those shown in the Table may differ among the geographic areas and study populations, and the magnitude of healthy volunteer effects may vary among study populations. Perhaps the most plausible contributors are variations among trials in adherence to outcome ascertainment protocols, variations among study populations in detection and diagnostic practices, and variations in medical care practices more generally among the diverse study populations.

Even in the presence of substantial MI underascertainment, treatment comparisons remain valid, provided ascertainment rates are precisely the same among randomized groups. An assumption that ascertainment has been nondifferential may be justified if ascertainment is nearly complete, but may be difficult to justify if a large fraction of events are not ascertained, especially if the reasons for underascertainment are unknown.

Moreover, study interpretation may be sensitive to even modest ascertainment differences among randomized groups. Suppose, for illustration, that cardiovascular outcomes are ascertained with probabilities of 40% in the rosiglitazone group and 50% in the control group throughout the follow-up period in the open-label RECORD trial. Hazard ratios (HRs) of 1.2 for the cardiovascular hospitalization or cardiovascular death primary end point, 1.4 for MI, and 1.1 for cardiovascular death would yield corresponding HRs for the ascertained outcomes of 0.96 for the primary end point, 1.12 for MI, and 0.88 for cardiovascular death, which is rather similar to the HRs estimated from the trial data. The investigators' claim that “rosiglitazone does not increase the risk of overall cardiovascular morbidity or mortality”² is highly sensitive to the implicit assumption that any underascertainment of cardiovascular events was identical in the randomized groups.

The complete and accurate ascertainment of events in these trials seems key to the interpretation of their results and provides confidence about efforts to translate related new interventions into practice. The low event rates in some of these trials, which may be particularly problematic in noninferiority trials (such as RECORD), raises some questions about the quality of their conduct. Without an explanation for the deviation from anticipated event rates, it may be unclear whether their findings are free of bias or whether their interventions merit widespread dissemination. In addition to ensuring that outcome ascertainment protocol and procedures are free of differential elements between randomized groups, trial investigators need to ensure that end point ascertainment is as complete as possible.

Although emphasis on disease rate comparisons among randomized groups is central to the interpretation of trials, journals and regulatory bodies could also usefully place emphasis on absolute disease rates in study evaluation as an additional quality-control effort. The globalization of the conduct of clinical trials perhaps requires some new activities to protect the clinical trial information chain. The large trials such as those for type 2 diabetes^{2 - 6} are important for establishing standards of practice and for public health so every effort should be made to produce studies of the highest reliability and clearest interpretation.