Summary

Investigations of the association between diabetes, diabetes treatments, and cancer risk have raised several epidemiological challenges. In particular, a patient’s exposure to glucose-lowering drugs needs to be represented accurately to allow unbiased assessment of the link between the treatments and cancer risk. Many studies have used a simple binary contrast (exposure to a specific drug vs no exposure), which has potentially serious drawbacks. In addition, methods used to determine the duration and cumulative dose of drug exposure differ widely between studies. In this Review, we discuss representation of drug exposure in pharmacoepidemiological investigations of the connection between diabetes drugs and cancer risk. We identify principles that might improve future research (particularly in observational studies), and consider issues related to reverse causation and detection bias.

Diabetes mellitus is a major public health challenge in many countries. WHO estimates that 350 million people have diabetes worldwide.1 Investigators of epidemiological studies have identified associations between diabetes and an increased risk of some cancers. Evidence for these associations has been summarised in reviews and meta-analyses of cancer of the breast,2endometrium,3 prostate,4, 5 liver,6, 7 pancreas,8, 9 bladder,10 colon and rectum,11—14 and kidney,15 and of non-Hodgkin lymphoma.16 In all these types of cancer, except cancer of the prostate, diabetes is linked to increased risk. Many possible explanations for these associations have been reviewed.17—20

Three mechanisms exist that might (alone or in combination) explain the link between diabetes and cancer: risk factors common to both disorders (eg, obesity or physical inactivity), direct causal effects of metabolic derangements in diabetes on cancer development, and the effect of diabetes treatments. We acknowledge the importance of lifestyle interventions (smoking cessation, increased physical activity, a healthy diet, and maintaining a healthy weight) in reducing cancer risk; however, this Review is limited to methodological challenges in studying the effects of diabetes drugs.

Investigation of the association between treatment of diabetes and risk of cancer is complex, and substantial scope exists for false-positive and false-negative conclusions to be made about causality. Accurate representation of exposure is one key dimension of high-quality studies in this specialty, and is arguably as important as other features of study design such as avoidance of other forms of bias, adequate sample size, appropriate adjustment for confounders, and the use of validated data sources. The use of validated data sources is arguably of particular importance in this article; we discussed how drug exposure can be represented in analysis, but the source data used to construct that representation should be of the highest possible quality. One example of a data source that has been subjected to extensive validation is the General Practice Research Database.67 This resource has an international reputation for the study of drug safety,67 and has been used in several studies of diabetes drugs and cancer risk.52—54 Two reports68, 69 have described the validity of cancer outcomes recorded in the General Practice Research Database and contributors suggest that validation could improve data quality.

Our examination of a small and opportunistic sample of studies about diabetes treatments and cancer risk suggests that a wide range of approaches have been used to represent exposure to glucose-lowering drugs. Although we recognise that all investigations are inevitably subject to resource constraints, we suggest that future studies observe three principles for the representation of exposure. First, investigators should carefully consider the restricted validity and value of simple binary representations (ever-or-never). Where such representations are unavoidable, investigators should apply of some form of sensitivity analysis (as Monami and colleagues32 used), if feasible. Second, construction of measures that represent total exposure (ie, dose—duration products), analogous to the pack year used in tobacco studies, offers potential benefits and could be explored. Third, the concept of continuity of exposure (ie, continuous versus interrupted treatment) has received relatively little attention. Representation of this concept in an analysis will be challenging, both to obtain the necessary detailed prescription data and in developing appropriate analysis techniques (eg, multistate models). However, continuity of treatment is, arguably, a meaningful component of the total exposure experience, and representation should be considered if possible. Finally, adherence to drug regimens is not likely to be complete, is difficult to assess, and could result in differential misclassification bias and underestimation of the real effect of a drug.

Source: Lancet