Cost-Effectiveness of a Collaborative Care Model Among Patients With Type 2 Diabetes and Depression in India

Physical and mental conditions share multiple physiological and behavioral pathways (1,2), leading to an increased risk for mental health comorbidities, such as depression and anxiety, among patients with chronic conditions (3,4). Among people with type 2 diabetes, depression is twice more common than among people without type 2 diabetes. Furthermore, comorbid depression is associated with an elevated risk for micro- and macrovascular complications (5,6), higher mortality (7), worse self-management (8), and higher costs (9).

The syndemic of type 2 diabetes and depression is particularly pervasive in lower- and middle-income countries (LMICs), where in India, for example, the prevalence is estimated to be >30% (10). In LMICs, a large proportion of mental health disorders remain undiagnosed and untreated, producing a large economic burden for individuals, families, and health systems (11–14). Integrated, collaborative care can address these challenges in caring for people with comorbid mental and physical chronic conditions (15–17).

Especially in resource-limited settings, disease management needs are often unmet because of lack of universal health care coverage and barriers to care through large physical distances, fragmented services by specialists, and an overall shortage of health care professionals. Hence, there is a need for accessible, cost-effective, integrated chronic disease care (18–21). For LMICs, some evidence on the cost-effectiveness of integrated mental health care exists (22), yet economic evaluations of interventions for patients with physical chronic and concurrent mental conditions are lacking to guide health policy and clinical practice in these countries (23).

Previously, we showed that compared with usual care, a 12-month culturally adapted collaborative care intervention, which integrated care for patients with type 2 diabetes and depression in four diverse urban diabetes clinics in India, was associated with a 30% higher likelihood of achieving sustained, clinically significant improvements in composite depressive symptoms and cardiometabolic indices at 24 months (15,24). Here, we evaluated the within-trial cost-effectiveness of this culturally adapted collaborative care intervention.

The Integrating Depression and Diabetes Treatment (INDEPENDENT) study was a multicenter, open-label, pragmatic, patient-randomized controlled trial conducted at four socioeconomically diverse urban diabetes clinics in India. The ethics committees of all involved institutions approved the study, and eligible patients gave written informed consent before they were enrolled. Detailed methods have been published elsewhere (15).

To be eligible for inclusion, patients had to be at least 35 years of age and have a confirmed diagnosis of type 2 diabetes, moderate to severe depressive symptoms (9-item Patient Health Questionnaire [PHQ-9] score ≥10 [range 0–27] where higher scores indicate more severe depressive symptoms), and at least one poorly controlled cardiometabolic marker (hemoglobin A_1c [HbA_1c] ≥8%, systolic blood pressure ≥140 mmHg, or LDL cholesterol ≥130 mg/dL).

Clinic staff reviewed medical records at the four clinics (private clinics in Chennai, Bangalore, and Visakhapatnam and a large public hospital in Delhi) from 9 March 2015 to 31 May 2016 to identify eligible patients on the basis of cardiometabolic parameters and screened them using the PHQ-9. Blinded study staff randomized eligible, consenting patients in blocks of 4, 6, 8, or 10, stratified by site, using an electronic data management system. Patients in the intervention group received collaborative care for 12 months and were followed for 12 additional months without intervention. The collaborative care intervention consisted of three evidence-based components at the patient (e.g., self-management support), clinician (e.g., electronic decision support), and system (e.g., patient case reviews) level. A care team at each clinic implemented these components to supplement usual care by diabetes physicians. Each team consisted of a full-time nonphysician care coordinator who had no previous training in mental health management and specialists (psychiatrist and diabetologist) who served as consultants and provided advice during case review meetings. All teams received a 3-day in-person training on the intervention, and care coordinators received additional training on management of patients with depression, including reinforcement through monthly coaching calls with the study team. Extensive information on the intervention is available elsewhere (15,25).

The control group received usual care over 24 months. Treating physicians were notified of their patients’ depressive symptoms and received basic training in suicide risk assessment and management.

All patients were followed for 24 months. Patients in both groups attended baseline and 6 monthly assessments, which were paid for by the study. All costs of clinical care between study assessments were borne by the respective payers (patient out of pocket, health insurance, or government-funded clinics).

Information on depressive symptoms was collected during all study visits using the PHQ-9 and the 20-item Symptoms Checklist Depression Scale (SCL-20) (range 0–4, with higher scores indicating worse symptoms). To measure health status from a patient perspective (i.e., health utility), data on health-related quality of life was collected at the baseline and 24-month study visit using the Health Utilities Index Mark 3 (HUI-3) (range −0.36 to 1.00, with higher values indicating higher utility) with its standard Canadian value set. The HUI-3 measures eight different attributes of health-related quality of life (e.g., vision, ambulation, cognition), is robust across populations, and has been validated in East Asian populations (26,27). Information on the various cost components was collected at baseline, 12 months, and 24 months using pilot-tested questionnaires (28).

We performed within-trial cost-effectiveness analyses comparing collaborative care with usual care over 24 months from multipayer and societal perspectives. The multipayer perspective included costs occurring directly in the formal health sector (e.g., clinic visits), independent of the actual payer, and out-of-pocket expenses in the informal health sector (e.g., food during inpatient visits). It comprised costs borne by health care providers (i.e., public hospitals) and private health insurance (i.e., reimbursement of patient-reported expenses) as well as out-of-pocket expenditures by patients. The multipayer approach is especially useful in LMIC settings with a high heterogeneity of payers (29). The societal perspective further included all other costs in the informal health sector, such as patient time costs for care and lost productivity due to outpatient care (Supplementary Material 1).

All costs were measured in Indian rupees (INR), with 2015 as the base year, and converted to purchasing power parity adjusted for international dollars (Int’l-$) with an exchange rate of INR20.18 per Int’l-$ based on the average exchange rates over 2015–2018 (30). Costs and health effects were discounted at 5% per year, which is recommended for LMICs because of higher rates of economic growth (31).

We followed a hybrid costing approach to calculate cumulative costs over the 24-month study period. Data sources included 1) self-report on health care utilization and expenditure for both groups; 2) health care utilization, care coordinator interactions, and case review meetings recorded in the decision-support electronic health record system during the active intervention period for the intervention group only; and 3) costs, resource use, and time spent on intervention-related activities from a survey among clinic staff (Supplementary Material 2). Self-reported costs over the previous 6 months were doubled to reflect annual costs. We added 10% overhead and fringe benefit costs to intervention setup and delivery costs, except commodities (e.g., laptops). An extensive description is available in Supplementary Material 3. This study was conducted and reported according to current guidelines for cost-effectiveness research in health and medicine (32,33) (Supplementary Material 4–5).

Total costs related to screening were calculated by multiplying the number of potentially eligible individuals (n = 1,905) identified through clinical records by the time and associated labor costs (10 min × wage) a care coordinator needs per individual to screen for depressive symptoms using the PHQ-9. Per-patient screening costs were calculated by dividing total costs related to screening by the number of randomly assigned patients. To reflect that screening would be required to identify eligible patients in a real-world setting (i.e., screen and treat), screening costs were only applied to intervention group participants.

Costs related to intervention setup and delivery consisted of two components: 1) fixed costs for care coordinator training (onsite and online), one laptop per site, educational materials, and care coordinator salaries for 12 months and 2) care team labor costs for case review meetings, including time spent by psychiatrists and diabetologists. Corresponding time use and unit costs were based on the clinic staff survey (Supplementary Table 1A and B).

We also, separately, investigated the cost containment potential of integrating the intervention into existing clinical workflows by costing care coordinator labor on the basis of actual time spent delivering the intervention (including time for administrative tasks and coordination with specialists) and the inclusion of the cost for development and implementation of the decision-support electronic health record system (Supplementary Material 6).

Costs related to health care utilization comprised those for outpatient care, inpatient care, and medications. Outpatient care was defined as the number of clinic visits, tests, and examinations. We calculated unit costs for all sites as the median value of self-reported cost per item (e.g., consultation fee, HbA_1c test) across the three private sites and cross checked these estimates with the clinic staff survey to ensure the reliability of our approach. To capture outpatient resource use in the intervention group during the active intervention period, we leveraged data on received care recorded in the decision-support electronic health record system. For all other periods and the usual care group, we had to rely on self-reported data. We analyzed the influence of this differential costing approach on the results in a sensitivity analysis (Supplementary Material 5). For inpatient costs only, self-reported costs for hospital stays due to type 2 diabetes–related complications could be calculated, as no other information on inpatient health care utilization was available. Medication costs were calculated using self-reported expenses.

Patient and escort expenses for food and transportation during hospital stays (food is usually not provided during hospital stays in India) were based on self-report. We estimated patient time cost with questions capturing time spent for diabetes care and patient wages. For both outpatient and inpatient care, time costs of escorts were estimated based on the self-reported number of visits the patient was accompanied and the respective escorts’ wages. We assumed a full workday of escort commitment for each escort visit as a conservative estimate to account for round-trip travel time to the clinic and waiting times. Following a human capital approach, patient lost productivity was estimated on the basis of self-reported workdays missed because of outpatient care, assuming an 8-h workday (Supplementary Material 3).

We used quality-adjusted life years (QALYs) and depression-free days (DFDs) to measure health utility and effects. Health status as measured by the HUI-3 was transformed to health utilities and used to estimate QALYs, a measure that incorporates quality and length of life in a single metric. Since 1 QALY indicates 1 year in perfect health (health utility = 1), we calculated accumulated QALYs per patient as the area under the health utility curve over the 24-month study period. Because the HUI-3 was only administered at baseline and 24 months, we linearly interpolated health utilities between these two time points (Supplementary Material 7).

We calculated DFDs on the basis of PHQ-9 score and, alternatively, the SCL-20 depression symptoms score as suggested by Vannoy et al. (34). To calculate DFDs, depression severity at baseline and all 6 monthly study visits was defined using two instrument-specific threshold scores. Patients were assumed to be free of depression (DFD = 1) if below these thresholds (PHQ-9 score <5, SCL-20 score <0.5) and fully depressed (DFD = 0) if above (PHQ-9 score >14, SCL-20 score >1.34). Thresholds are defined according to the literature except for the upper limit of the SCL-20 score, which was defined using the mean SCL-20 score for patients at baseline (34). If the respective score was between thresholds, linear interpolation was used to convert scores into proportionate DFDs between 0 and 1. The final estimate of DFDs accumulated by patients is the number of days between assessment time points multiplied by the respective level of depression (between 0 and 1), which is equivalent to the area under the depression severity curve over time. To achieve this, we linearly interpolated depression severity between study visits (Supplementary Material 7).

We conducted the primary analysis from a multipayer perspective. Secondary analyses were conducted separately from a societal perspective and stratified by clinic. We estimated between-group differences in cumulative costs over 24 months based on average marginal effects using generalized linear models with a γ-distribution and a logarithmic link function (35). Because the logarithm of 0 is undefined, arbitrary costs of INR10.0 were added to all participants with zero costs. Between-group differences in QALYs and DFDs after 24 months were estimated using linear regression models. For all outcomes, we estimated unadjusted differences in costs between the intervention and usual care groups. In the adjusted sensitivity analysis, we controlled for baseline values of outcomes. We calculated incremental cost-effectiveness ratios (ICERs) and constructed cost-effectiveness acceptability curves (CEACs) for each health outcome and costing perspective. Uncertainty in ICERs was computed using a bootstrap procedure with 1,000 replications.

Self-reported data on outpatient health care utilization was missing for up to 20.30% of observations at baseline, 22.52% at 12 months, and 25.74% at 24 months. To account for these missing data and facilitate intention-to-treat analyses, we merged available self-reported economic data with 10 imputation data sets that were previously constructed for the primary effectiveness analysis (15). We next imputed each merged data set fivefold, for a total of 50 imputation data sets (Supplementary Material 8). All statistical procedures were based on these 50 imputation data sets, and results were combined following convention (i.e., Rubin’s rules) (36).

As no official willingness-to-pay threshold for QALYs exists in India, we used guidance from the World Health Organization Choosing Interventions That Are Cost-Effective initiative, which recommends a threshold of one to three times the country-specific gross domestic product (GDP) per capita (37). This resulted in thresholds between INR112,000 (Int’l-$5,552) and INR336,000 (Int’l-$16,654) per QALY based on India’s average per-capita GDP of Int’l-$6,041 from 2015 to 2018 (30,38) (Supplementary Material 9). Similarly, no established willingness-to-pay threshold exists for DFDs. To enable interpretation of the related CEACs, we calculated the theoretical willingness to pay per DFD that would be needed to achieve a prespecified probability of cost-effectiveness that was ≥95%.

All analyses were conducted in R version 4.0.3 (39). Statistical tests were two-sided and used a significance threshold of 5%. Where appropriate, 95% CIs are reported.

We performed sensitivity analyses to explore the influences of our costing decisions and to check the robustness of our results. These included a cost containment analysis in which the intervention is assumed to be integrated in clinical workflows, adjustment for baseline values of outcomes, adoption of a health system perspective, use of self-reported data on health care utilization for both groups, inclusion of costs for development and implementation of the decision-support electronic health record system, and a two-way missing-not-at-random analysis in which costs are varied by ±15% and QALYs by 15% to account for potential over- or underestimation. Each analysis is described in detail in Supplementary Material 6.

In total, 1,905 patients were screened for eligibility. Of 404 randomly assigned patients, 196 were allocated to receive the collaborative care intervention and 208 to receive usual care (Supplementary Fig. 1). Patient characteristics in both groups were similar at baseline. Mean (SD) health utility was 0.66 (0.26) in the collaborative care group and 0.62 (0.25) in the usual care group. Baseline costs across categories were also similar between groups (Supplementary Table 2).

Considering all cost components, patients in the collaborative care group accumulated mean (SD) total costs of INR48,287 (15,799) (Int’l-$2,393 [783.1]) compared with INR20,920 (12,933) (Int’l-$1,037 [641.1]) in the control group over the 24-month study period. A descriptive overview of cost components by group is presented in Table 1. Unadjusted between-group differences in costs are presented in Table 2.

Screening administered by care coordinators to identify eligible patients costed INR181 (Int’l-$8.97) per patient. On average, the intervention costed INR21,836 (Int’l-$1,082) per patient in the collaborative care group (Table 1). Of these, INR3,850 (Int’l-$190.8) were upfront costs associated with intervention setup (e.g., training workshops, laptops). The largest share of intervention costs accrued during intervention delivery, of which salaries for care coordinators constituted the most expensive component (INR14,661 [Int’l-$726.7]) (Table 1).

Health care utilization was the most expensive nonintervention cost component (Table 1). Patients in the collaborative care group had INR3,926 (95% CI 1,881–5,971) (Int’l-$194.6 [93.3–296.0]) higher health care utilization costs compared with the usual care group, particularly because of outpatient care. Medication costs and costs due to inpatient care were similar between groups (Tables 1 and 2).

Costs related to food, transportation, time, and lost productivity were largely similar between groups (Table 1). Although we observed slightly higher time costs (INR808.6 [95% CI −367.5 to 1,985]) (Int’l-$40.1 [−18.2 to 98.4]) and lost productivity (INR468.1 [−537.4 to 1,474]) (Int’l-$23.2 [−26.7 to 73.0]) in the collaborative care group, which were both driven by outpatient care, the respective estimates were subject to considerable uncertainty (Table 2).

Over 24 months, patients in the collaborative care group gained more QALYs and DFDs compared with the usual care group (Table 1). The unadjusted between-group difference in QALYs was 0.084 (95% CI 0.015–0.152). The between-group difference in DFDs was 89.5 (62.8–116.2) using PHQ-9 and 113.6 (81.2–146.0) using SCL-20 (Table 2, Fig. 1, Supplementary Table 3, and Supplementary Fig. 2).

From a multipayer perspective, mean costs in the collaborative care group were INR25,975 (95% CI 23,916–28,035) (Int’l-$1,288 [1,549–1,753]) higher than in the control group. The intervention costed INR309,558 (Int’l-$15,344) per QALY and between INR290.2 (Int’l-$14.4) and INR228.7 (Int’l-$11.3) per DFD gained based on the PHQ-9 and SCL-20, respectively (Table 3 and Supplementary Table 3).

For QALYs, the probability that the intervention is cost-effective was 0.80–56.4% using willingness-to-pay thresholds from INR112,000 (Int’l-$5,552) to INR336,000 (Int’l-$16,654). The required willingness-to-pay per DFD to achieve a probability of cost-effectiveness >95% was INR401.6 (Int’l-$19.9) (Table 3 and Fig. 1). From a societal perspective, cost-effectiveness was only marginally lower for all outcomes (Table 3 and Fig. 1).

Cost-effectiveness decreased when adjusting for baseline values, particularly because of a reduction in the estimated between-group difference in QALYs (Table 3, Supplementary Tables 3–5, and Supplementary Figs. 2 and 3). Stratified analyses showed moderate differences among clinics, albeit patients at one clinic gained no QALYs in the adjusted sensitivity analysis (statistically nonsignificant) (Supplementary Table 6A).

Our findings were sensitive to some costing decisions. Particularly in the cost containment analysis, which assumed that care coordinators were integrated into clinical workflows so that their labor costs were based on actual time spent delivering the intervention instead of annual dedicated full-time salaries, the per-patient intervention costs reduced to INR9,510 (Int’l-$471.4) (Supplementary Table 7). In this scenario, incremental costs were reduced by ∼47% (Supplementary Table 8), and the probability of cost-effectiveness increased to 15.1–89.4% depending on the respective willingness-to-pay threshold (results not shown). However, including costs for the development of the decision-support electronic health record system led to a decrease in cost-effectiveness. When relying solely on self-reported cost data for both groups, cost-effectiveness of the intervention increased, rendering our costing decision in the main analysis to the more conservative approach (Supplementary Table 5A and B). The two-way missing-not-at-random analysis indicated a large cost-effectiveness potential of the intervention if QALYs are truly underestimated and costs overestimated but further highlights that cost-effectiveness might be substantially lower when baseline health utility is adjusted for (Fig. 1 and Supplementary Fig. 2).

In this within-trial economic evaluation of a culturally adapted 12-month collaborative care intervention targeting patients with type 2 diabetes and depression at four diverse urban diabetes clinics in India, we found that collaborative care can be cost-effective over 24 months, especially when integrated into existing clinical workflows. Integrating mental health services in primary care in LMICs is effective, yet there is a paucity of economic evaluations of integrated care models for patients with chronic conditions and mental comorbidities (22). The original effectiveness analysis of the INDEPENDENT trial demonstrated improvements in depressive symptoms and cardiometabolic indices in patients seeking care in diabetes clinics (15). In contrast, the Integrated mHealth System for the Prevention and Care of Chronic Disease (mWELLCARE) trial aimed to assess the effectiveness of an integrated care model for the management of hypertension, type 2 diabetes, and depression in government primary clinics in India but found no benefit (40).

Studies from the U.S. and U.K. that investigated collaborative care in comparable patient populations reported similar, albeit larger, improvements in health-related quality of life and smaller incremental costs. The Team Approach to Improve the Quality of Diabetes Care for Patients With Schizophrenia (TEAMcare) study in Washington State investigated the effectiveness of a patient-centered, team-based collaborative care management intervention in patients with depressive disorder and poorly controlled type 2 diabetes or coronary heart disease over 24 months (41). Patients in the collaborative care group achieved substantial gains in (regression-based) QALYs (0.335 [95% CI −0.18 to 0.85]) and DFDs (114 [79–149]). The intervention in TEAMcare was cost-saving over 24 months (16). A comparable intervention and evaluation was applied in the 24-month cost-effectiveness analysis of the U.K. Collaborative Interventions for Circulation and Depression (COINCIDE) trial, including patients with depression and type 2 diabetes or coronary heart disease (42). Here, collaborative care resulted in 0.136 (0.061–0.212) additional QALYs and incremental costs of £1,777 (95% CI −320 to 3,875) (ICER £13,069 per QALY) (42). The reduction of the incremental intervention cost by ∼50% in our cost containment analysis, which, similar to TEAMcare, replicates a costing approach based on actual care coordinator-patient interactions, supports that collaborative care is very likely to be cost-effective if integrated in existing clinical workflows.

Setting up patient-centered collaborative care is resource intensive and presupposes an essential infrastructure, such as patient registries and electronic records (43). It is therefore crucial to distinguish between upfront implementation and long-run operating costs. Keeping this in mind, we explored the impact of our costing and analysis decisions to provide an overall picture of the cost-effectiveness of collaborative care for patients with type 2 diabetes and depression in India, which may be informative for future extensions in LMICs. On a sidenote, our findings and implications did not change when applying alternative discount rates (results not shown).

We show that a 12-month collaborative care intervention can have sustained effects and is cost-effective at 24 months, yet the current follow-up prohibits an extrapolation of the long-term health effects, costs, and thus, cost-effectiveness because of the uncertainty in the clinical trajectory of both groups. Short-term (re)intervention after deterioration of symptoms and corresponding longer depression-free states and improved health-related quality of life may lead to substantial gains in QALYs. Additionally, reductions in the long-term per-patient costs of the intervention are likely due to the amortization of fixed costs, such as training, increased efficiency in clinical workflows, and declining technology costs.

Because of high integration and resource requirements, the scalability and adaptability of collaborative care across diverse local settings remain important challenges. Implementation in clinics of different types, sizes, and locations, with heterogeneous patient populations exhibiting specific patterns of multimorbidity (including infectious diseases), may presume the modification of specific (expensive) intervention components, such as in-person case review meetings. This is aggravated by a severe global shortage of mental health care professionals, especially in LMICs and even more so in rural areas (44).

Our study has several strengths. It is the first to provide evidence of which circumstances collaborative care for patients with chronic physical conditions and mental comorbidities could be cost-effective in an LMIC setting. We captured context-specific economic impacts of the intervention, such as escort costs, and used Int’l-$ to make our estimates comparable across countries. To measure patient benefits of the intervention, we used QALYs and two DFD measures as clinically relevant outcomes. Finally, we explored the sensitivity of our results to a variety of costing approaches, analysis decisions, and implementation scenarios, exploring their influence on cost-effectiveness.

Our study also has several limitations. First, because of the pragmatic nature of the trial and the heterogeneity in the Indian health system, detailed assessments of health care utilization or validation of self-reported information using patient or insurance records was infeasible. We therefore relied on consistently collected self-reported data for many cost components, including medications; notably, these were no different between groups at baseline. Second, we did not account for costs of restructuring diabetes care to integrate collaborative care in existing clinical workflows in the cost containment analysis. These results, therefore, constitute a lower cost-effectiveness bound. Third, albeit an HUI-3 value set to calculate health utilities among high-income Asian populations exists, no corresponding value set for LMIC settings is available. Fourth, health utility was not collected at the 12-month visit, where the largest between-group difference was expected according to the clinical effectiveness analysis and the association between depressive symptoms and health-related quality of life (15,45). Therefore, underestimation of the incremental QALYs is likely. As a potential consequence of this, the intervention was unlikely to be cost-effective in the adjusted sensitivity analysis because of a smaller incremental difference in QALYs. However, considering the potential underestimation of QALYs and the results from the cost containment analysis, this does not affect the main conclusion of our study. Fifth, a substantial amount of economic data was missing. We conducted missing-not-at-random analyses to account for this further uncertainty in costs in conjunction with the missing health utility measurement at 12 months. Sixth, data on depression-related inpatient costs were not available. However, this is unlikely to affect our results in a meaningful way because patients with already prevalent severe depression and other psychiatric disorders were not eligible for the trial, and depression-related serious adverse events were very rare (15). Finally, we were only able to include lost productivity due to outpatient care but not to illness.

Collaborative care, as applied in the INDEPENDENT trial, could be cost-effective to treat patients with type 2 diabetes and depression in urban India, particularly if integrated into routine clinical workflows. This analysis can serve as guidance for the integration of mental health and chronic disease care in other LMIC settings. From a long-term perspective, cost-effectiveness of the intervention is likely to be more favorable but might be affected by heterogeneity in implementation feasibility because of contextual factors across diverse local settings.

Funding. This study was funded by the National Institute of Mental Health (grant R01MH100390).

The National Institute of Mental Health had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and the decision to submit the manuscript for publication.

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

Author Contributions. K.M.F.E.-F., M.L., S.A.P., K.S., S.P., V.M., L.C., and M.K.A. conceptualized and revised the design of the study. K.M.F.E.-F., M.L., K.S., L.C., and M.K.A. were responsible for the analysis and interpretation of the data. S.A.P., K.S., V.M., R.S., A.R.S., R.M.A., G.R.S., and N.T. critically revised and provided feedback on the interpretation of the data. K.M.F.E.-F., M.L., and M.K.A. prepared the initial manuscript draft. All authors reviewed, edited, and approved the final manuscript. V.M., L.C., K.M.V.N., and M.K.A. obtained funding for this work. K.M.F.E.-F. 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.

Prior Presentation. Parts of this study were presented in abstract form at the 16th Symposium of the International Diabetes Epidemiology Group, 29 November–1 December 2019, Seoul, South Korea.