Population-Based Cohort Analyses of the Bidirectional Relationship Between Type 2 Diabetes and Depression

The data analyzed in this study were claims of 1 million beneficiaries randomly selected from all beneficiaries insured in 2000, with age and sex distributions nearly identical to the entire insured population of Taiwan (19). The claims were retrieved from the National Health Insurance Research Database (NHIRD) provided by the Bureau of National Health Insurance (BNHI). The NHIRD provides all inpatient and ambulatory medical claims for ∼96% of Taiwanese people (20,21). To ensure the accuracy of claim files, the BNHI performs quarterly expert reviews on a random sample for every 50–100 ambulatory and inpatient claims. False reports of diagnosis would yield severe penalties from the BNHI (22). By the end of 1996, BNHI had contracted with 97% of the island-wide hospitals and clinics, with 99% of the total Taiwanese population enrolled in the program (21). Therefore, information obtained from the NHIRD is believed to be complete and accurate. We used several NHIRD datasets in this study, including ambulatory care visit claims (ACVC), Inpatient Expenditures by Admissions (IEA), and Registry for Beneficiaries (RB). Access to research data has been approved by the Review Committee of the National Health Research Institutes.

An individual was classified as a diabetic patient if she or he had a diagnosis of type 2 diabetes (ICD-9-CM: 250 × 0 or 250 × 2) at any time in ACVC of 2000 and then experienced another one or more diagnoses within the subsequent 12-month follow-up periods. The first and last outpatient visits within 1 year had to be >30 days apart to avoid accidental inclusion of miscoded patients (23). The eligible diabetic patients must have no prior history of depression (ICD-9-CM: 296, 309, or 311) (3) diagnosis since 1 January 1997. In total, 16,957 prevalent diabetic patients were included in the diabetic group. The control subjects were 16,957 insurers randomly selected, sex and age matched to the diabetic group, from all beneficiaries free from both diabetes and depression in 1997–2000.

We linked the diabetic and control subjects to ACVC in 2000–2006 for possible episodes of diagnosis for depression. The index date for each diabetic patient was the date of his or her first diabetes diagnosis. The index date for subjects in the control group was the first date of enrollment in NHI. If their first date of enrollment was before 1 January 2000, the index date was set as 1 January 2000. The 7-year follow-up period began as early as 1 January 2000 and ended 31 December 2006. The age of each study subject was calculated by the difference in time between the index date and the date of birth. We grouped the area of each member’s insurance unit, either the beneficiaries’ residential area or location of their employment, into four geographic areas (north, central, south, and east) or urbanization status (urban and rural) according to the National Statistics of Regional Standard Classification (24), and such information was obtained from the RB.

The age- and sex-specific hazard rates were determined with person-years (PY) as the denominator under the Poisson assumption. To assess the independent associations of diabetes with the risks of depression, we conducted Cox proportional hazards regression models with age, sex, geographic area, urbanization statuses, and various comorbidities adjusted simultaneously in the model. We adjusted geographic variables for the presence of an urban-rural difference in the accessibility to medical care in Taiwan (25). The comorbidities considered in our analysis included a number of medical diagnoses considered to pose a long-term risk for depressive symptoms (12) and several macrovascular complications that could substantially affect diabetic patients’ quality of life and psychological well-being (14). Information of comorbidities was retrieved from the IEA from the first day of 1997 to the date of encountering a depression diagnosis, or to the date of censoring, which was either the date of withdraw from the insurance or date of the end of follow-up, i.e., 31 December 2006. All statistical analyses were performed with SAS (version 9.2; SAS Institute, Cary, NC). A P value <0.05 was considered statistically significant.

The second cohort analysis managed to assess whether depression predicts the onset of diabetes. The depression subjects were also identified from the above-mentioned 1 million beneficiaries. An individual was classified as a depression patient if she or he had a depression diagnosis (ICD-9-CM: 296, 309, or 311) at any time in 2000, and she or he must have no prior history of diabetes (ICD-9-CM: 250 × 0 or 250 × 2). In total, 5,847 subjects with depression were identified to form the depression group. The control subjects (n = 5,847) were randomly selected and sex and age matched to the depressive patients from all nondepression insurers. The control subjects must have been free from both depression and diabetes in 1997–2000.

We linked the subjects from both groups to ACVC in 2000–2006 for possible episodes of diagnosis of diabetes. The index date for each depression subject was the date of his or her first depression diagnosis in 2000. The index date for subjects in the control group was the first date of enrollment to NHI. If their first date of enrollment was before 1 January 2000, the index date was set as 1 January 2000. The 7-year follow-up period began as early as 1 January 2000 and ended 31 December 2006. The methods of calculation of age and grouping one’s beneficiaries’ residential area or location of employment were the same as those described in the first cohort analysis. The statistical analysis used in the second cohort analysis was also similar to that used in the first cohort analysis. The only difference was that we did not take into account the risk factors for diabetes, since many of the risk factors for diabetes were related to one’s diet, lifestyle, and compliance to medical advice, but such information was not included in the NHI claims.

Male subjects dominated the sample in both diabetic and control groups (53.5%), and diabetic patients were of the same age as their matched controls (mean age ± SD, 60.1 ± 13.2 vs. 60.3 ± 13.1 years). Distributions of urbanization level and geographic location of residential areas were essentially the same between the two groups. On the other hand, the respective prevalence of prior admission due to cancer, hypertension, rheumatoid arthritis, lung disease, and peripheral artery disease was higher in diabetic patients than in control subjects. Moreover, diabetic patients were also more likely than control subjects to suffer from commonly seen diabetes complications and disease consequences, including stroke (7.3 vs. 5.2%), cardiovascular disease (6.1 vs. 4.8%), nontraumatic hip fracture (1.0 vs. 0.3%), and lower-extremity amputations (1.0 vs. 0.2%) (Table 1).

Table 2 shows the overall and age- and sex-specific incidence densities (IDs) of depression for the two study groups. Over the 7-year study period, diabetic patients had an ID of 7.03 (95% CI 6.52–7.55) per 1,000 PY. The control subjects had a lower ID at 5.04 (4.62–5.46) per 1,000 PY. Diabetic patients were observed to suffer from a higher ID than control subjects irrespective of age and sex. Concerning the age-specific ID, it tended to increase with age regardless of sex among control subjects. However, for the diabetic patients, the lowest age-specific ID of depression was noted in male subjects 45–54 years of age (4.93 [3.76–6.10] per 1,000 PY) and in female subjects 55–64 years of age (7.66 [6.20–9.12] per 1,000 PY). Table 2 also shows covariate-adjusted hazard ratios (HRs) of depression in association with diabetes. Compared with control subjects, diabetic patients had a significantly increased hazard of depression (HR 1.43 [95% CI 1.16–1.77]). The corresponding figures for male and female subjects were 1.39 [1.08–1.79] and 1.48 [1.14–1.92]. However, these two sex-specific HRs were not significantly different (P value for diabetes-sex interaction is 0.0749). On the other hand, the diabetes and age interaction was significant for both male and female subjects. The highest age-specific HR was noted in subjects <35 years of age in both male and female subjects with an HR of 1.81 (1.13–2.91) and 2.16 (1.39–3.37), respectively. For both sexes, the age-specific HR gradually decreased as age increased.

Female subjects (59.6%) were dominant in the depression group. The mean ± SD age for the depression group was 44.5 ± 17.6 years, with a majority of study subjects <45 years of age (54.9%). The distributions of urbanization and geographic locations of residential areas were similar between the two groups (Table 3). Table 4 shows overall and age- and sex-specific ID and HR of diabetes onset for the depression and nondepression groups. The overall ID of diabetes for the depression and nondepression groups was estimated at 27.59 per 1,000 PY and 16.45 per 1,000 PY, respectively. The ID of diabetes increased with age for both men and women, and depressive subjects consistently had a higher ID of diabetes than the nondepressive subjects for all sex and age stratifications. The depressive subjects were associated with a significantly increased HR (2.02 [95% CI 1.80–2.27]), with a greater RR estimate for women (2.23 [1.89–2.62]) than for men (1.83 [1.56–2.16), but this sex difference was not statistically significant (P = 0.0721). The age-stratified analyses revealed that age significantly interacted with depression on the risk of diabetes for both men and women (both P < 0.0001), in which younger depressive patients were associated with higher HRs than older patients. The depressive women <35 years of age were significantly associated with the most increased HR (3.50 [2.47–4.97]), followed by depressive men <35 years of age (3.46 [2.37–5.04]). The depressive men and women 34–44 years of age had an HR of 2.13 (1.51–3.00) and 2.21 (1.54–3.18), respectively. The elevated HRs for depressive men >54 years of age and depressive women >64 years of age were no longer statistically significant.

The current study revealed a bidirectional association between diabetes and depression. Similar to a meta-analysis that included 13 publications of depression (9), our observations showed that the relationship between depression and incidence of diabetes was stronger than between diabetes and incidence of depression. The bidirectional relationships between diabetes and depression appreciably weakened with increased age in both sexes.

Several prospective studies have assessed the association between diabetes and risk of developing depression and showed contradictory findings (1,6,11–13,17,26). Mezuk et al. (9) have reported in their meta-analysis of these studies that the pooled RR was 1.15 (95% CI 1.02–1.30). A later investigation indicated that the elevated risk was found in treated diabetic patients but not in those untreated (17). In our analyses, the diabetic group was identified from claims records, likely consisting of a relatively large proportion of patients with severe diabetes who were more likely to seek health care and be treated.

In a meta-analysis report (9), investigators suggested that the relationship between diabetes and incidence of depression may differ by age and sex. Indeed, we found that diabetes was mostly associated with subsequent depression in female subjects <45 years of age and in male patients <35 years of age, and the association weakened in those that were older. A cross-sectional study supported our observation, indicating that the odds ratios for depression associated with diabetes were much higher in individuals 20–39 years of age than in those 40–64 years of age (3). The age difference may partly explain the modest association observed in previous studies, most of which did not include young adults.

Several potential explanations for the age differences have been proposed. First, the association between diabetes and depression among the elderly may be masked because of a greater burden of competing risks for depression in this age-group. Second, evidence has shown that older adults were less likely to endorse the symptoms of depression (27). The nondifferential misclassification of depression could bias the association toward null. Third, previous studies reported that type of treatment, severity of the disease, and its impact on physical and psychological conditions may vary between the diabetic patients of various ages (28). Younger diabetic patients tended to respond negatively to their disease and have more trouble in undergoing complex treatments and intensive health monitoring, which could lead to greater psychological burden (3).

A greater risk of depression in women has been well documented previously (14,29). The sex difference was also observed in patients with diabetes (30). In line with these studies, our observations revealed greater IDs of depression in women than in men in both the diabetic and control groups. However, the test for an interaction effect between sex and diabetes was not statistically significant, suggesting that the relationship between diabetes and risk depression did not differ between men and women. This result was consistent with the finding of Engum (1), which indicated that sex did not modify the diabetes effect on onset of depression and anxiety.

The results of studies linking depression to the risk of developing diabetes were consistent (5–9). Potential mechanisms have been elaborated, involving factors on a psychological, socioeconomic, and biological level (7,17). Different from the analyses for diabetes predicting depression, an increased risk of diabetes in patients with depression was observed in all age-groups, with the exception of men ≥55 years of age and women >64 years of age in this study. Our analyses showed that the adjusted HR reduced as age increased, consistent with the results of the subgroup analysis of the meta-analysis (9). The RR derived from studies including subjects with a mean (or median) age of <50 years was 1.96, higher than that from studies consisting of individuals with a mean (or median) age of ≥50 years (RR 1.50). In older patients with depression, a competing risk of death from other causes may increase, precluding the observation of diabetes. In addition, a misclassification of depression to no depression due to the difficulty in detecting the condition in the elderly may result in an underestimation of HR in this group.

Studies reporting stratified analyses on sex for depression predicting diabetes showed controversial findings. Some studies found greater risk in men (9,31,32), but other reports, similar to our analysis, revealed that sex did not significantly modify the relation between depression and diabetes (1,5). We observed a slightly greater HR in women than in men. A possible explanation is our use of claims data to identify depression and diabetes. It has been suggested that women were more likely to seek medical consultations (33). Women with depression may seek health care even more frequently, and thus their diabetes is more likely to be detected. Further studies are needed to clarify this issue.

A strength of this study is that it included a large, population-based sample that is highly representative, leaving little room for selection bias. The large sample size also made it possible to perform age- and sex-specific analyses without comprising the required sample size. There are also limitations to this study. First, the exposure and outcome in this study were both diagnosed depression and diabetes, as we relied solely on claims data that do not cover those who went undiagnosed. The misclassification for both conditions could lead to biased estimation on the association. Second, the claims data do not contain BMI, diet, and lifestyle information, which are important risk factors of diabetes. However, previous studies have suggested that these factors explained partially, but not all, of the association between depression and diabetes (17,18). Third, surveillance bias might occur in this study, as both patients with diabetes and depression are more likely than their control counterparts to seek medical care. To address this issue, we calculated the frequency of medical visits for each study subject and adjusted this variable in the multivariate regression model. The annual average number of outpatient visits was approximately twofold higher in the patient groups than in the control groups at baseline (i.e., the year 2000). After adjusting for the number of outpatient visits, the overall HR of depression onset reduced from 1.43 (95% CI 1.16–1.77) to 1.25 (1.21–1.30), and HR of diabetes incidence decreased from 2.02 (1.80–2.27) to 1.72 (1.53–1.94), suggesting a small magnitude of surveillance bias in our analysis. Fourth, because the NHIRD dataset covers the claims only beginning from 1997, we were unable to obtain each study subject’s lifetime information on the diagnoses of diabetes and depression.

Given that diabetes is a serious, chronic disease that is usually accompanied by various life-threatening complications and that depression is a complex condition characterized by distraction in all facets of life, including physiological, behavioral, and psychosocial impacts, it is crucial to both public and medical professionals to be aware of the bidirectional relationship between the two conditions. This study, focused on identifying groups at high risk, could help target prevention efforts and priorities. In this study, the association between diabetes and depression was particularly strong in adults <45 years of age, who should be the objects of further attention.

To address the public health impact, we further calculated the overall population-attributable risk percentage associated with diabetes and depression, respectively. According to the NHI statistics (34), the prevalence rate of treated diabetes and depression in 2010 was 6,541 and 4,685 per 10⁵, respectively. Together with the RR estimated from this current study, we estimated that 2.74% of depression was attributable to diabetes. On the other hand, ~4.65% of diabetes was attributable to depression. Given the increasing incidences of both diabetes and depression, efforts on preventing both conditions may help decrease the population-attributable risk percentage associated with these two conditions.

This study was supported by a grant from Taiwan’s National Scientific Council (NSC98-2314-B-227-001-MY2).

No potential conflicts of interest relevant to this article were reported.

P.-C.C. designed the study, researched data, and drafted the manuscript. Y.-T.C. contributed to statistical analyses and drafted the manuscript. H.-F.C. designed the study, research data, and contributed to the discussion. M.-C.K. designed the study and reviewed and revised the manuscript. C.-Y.L. obtained the research grant, designed the study, and reviewed and edited the manuscript. C.-Y.L. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.