In this prospective cohort study of almost 6000 middle-aged men and women, we found antidepressant use at four clinical examinations over an 18-year period to be associated with physician-diagnosed diabetes. However, we observed no association between antidepressant use and study screen-detected diabetes; that is, diabetes detected for the first time by routine blood testing as part of the Whitehall II study. Furthermore, there was no association between antidepressant use and glucose levels at any of the four clinical examinations, and continued antidepressant use was not associated with progressively increasing levels of fasting or 2-hour postload glucose over time. These data suggest the observed associations between antidepressant therapy and increased risk of diabetes are not causal.
Our findings on physician-diagnosed diabetes are in agreement with the majority of register-based investigations showing a link between long-term antidepressant use and a clinical record of diagnosed diabetes (3–8,27,28
). However, as register-based studies do not capture people with undiagnosed diabetes, such evidence is potentially affected by detection/ascertainment bias (11
). First, the observed higher proportion of physician-diagnosed diabetes cases among antidepressant users may relate to the indications for this drug treatment; that is, diabetes may be detected when ruling out endocrinologic diseases as a cause of depression (29
). Second, antidepressant use requires contact with medical care, which may increase the likelihood of the diagnosis of hidden health problems such as diabetes (8
). These explanations are consistent with the observed differential association with undiagnosed (study screen-detected) and diagnosed (physician-diagnosed) diabetes and the failure to observe a relationship between antidepressant use and undiagnosed diabetes. Our findings are also in agreement with trials on antidepressant medication that do not indicate excess short-term risk of type 2 diabetes (30,31
Unlike clinical record studies, the Diabetes Prevention Program trial targeted people who were at high risk of diabetes because of overweight and elevated blood glucose levels (10
). In that study, participants were randomly assigned to groups of lifestyle changes, glucose-lowering medication (metformin), or placebo (10
). The authors found that in the lifestyle and placebo groups, participants consistently on antidepressants during the study period were about twice as likely as nonusers to develop diabetes, although no such pattern was seen in the metformin group (10
). However, the study did not report stratified analyses for physician-diagnosed versus study screen-detected diabetes or comparisons of glucose trajectories between antidepressant users and nonusers. Furthermore, despite being based on a clinical trial, the analyses of antidepressant use in the Diabetes Prevention Program utilized observational data because the exposure of interest, antidepressant use, was not randomized in that study (10
). Thus, the observed association between antidepressant use and diabetes might have been due to unmeasured differences between the two groups of antidepressant use rather than a causal effect of antidepressant use.
Several potential confounding factors for the association between antidepressant use and diabetes have been hypothesized. For example, antidepressant use could be a proxy of more severe depression or a history of chronic or recurrent depression, which are robust predictors of type 2 diabetes, independent of antidepressant therapy (14,29,32
). In the present study, antidepressant users were also more sedentary at baseline and had a higher prevalence of smoking as compared with nonusers. In addition, antidepressant users were more likely to come from low socioeconomic groups, a predictor of both depression and diabetes. However, adjustment for these factors did not change the association between antidepressant use and physician-diagnosed diabetes. Our study, in combination with the evidence that physician-diagnosed diabetes also predicts future antidepressant use (33
), is consistent with a view that being treated for one condition increases the likelihood of being diagnosed with the other condition, irrespective of other characteristics of the patient.
Confounding factors may inflate, but could also suppress, the magnitude of an unadjusted association, contributing to false null findings. In the present study, antidepressant use was unrelated to screen-detected diabetes and plasma glucose both before and after adjustments for potential confounding factors, suggesting that the absence of associations with undiagnosed diabetes and glucose is not an artifact resulting from a suppression effect of the confounders.
It is important to consider potential limitations to the present study that could contribute to false-negative findings. First, the participants of the Whitehall II study are from an occupational cohort that is likely to cover a more restricted range of health status compared with the general population. However, a large bias due to restricted variance seems unlikely because the magnitude of the association between antidepressant use and diagnosed diabetes was comparable with that observed in other cohorts (3–10). Second, despite a high response to the successive data collection phases, loss to follow-up accumulated over the extended time period, as is inevitable in all long-term prospective studies. However, differences between the included participants and the total baseline population were generally small. Third, despite the large sample size, the number of diabetes cases among antidepressant users was relatively small; thus, the findings should be interpreted with caution. We did not have precise information on prescriptions (e.g., dosage) and the sample size was not large enough for analyses of specific classes of antidepressants. Given that side effects may vary depending on a drug's chemical substance, antidepressant-specific analyses should be undertaken in future studies (34
). Fourth, error in the measurement of glucose and diabetes status is a potential source of false-negative findings. This seems, however, an unlikely explanation of our findings because we used the World Health Organization diabetes definition, based on standard oral glucose tolerance testing, considered to be a gold standard measure (26
). Indeed, few previous studies have data based on glucose test available across repeated examinations. Fifth, we identified persons with depressive symptoms using a standard, validated questionnaire measure: the Center for Epidemiologic Studies Depression Scale (21,22
). This instrument has been shown to be a sensitive measure of mental health problems in the general population and in diabetic patients but was not designed to make a psychiatric diagnosis of first or recurrent major depression. Thus, we cannot exclude the possibility of confounding by unmeasured depression. However, as depression is known to underlie antidepressant use and increased risk of diabetes, unadjusted associations between antidepressant use and diabetes will represent, if anything, overestimates rather than underestimates of the true association. It therefore seems unlikely that unmeasured depression removed the associations of antidepressant use with screen-detected diabetes and glycemia in our study, as indication bias by depression should have inflated these associations.
We have demonstrated that detection/ascertainment bias may have compromised evidence in this field of research. Our longitudinal study of British men and women suggests that the adverse effect of antidepressant use on type 2 diabetes risk is biologically implausible and might have been overestimated in previous epidemiologic studies. This evidence suggests that concerns about important diabetogenic side effects of antidepressants might have been unfounded.