Health Care Costs in People With Diabetes and Their Association With Glycemic Control and Kidney Function Open Access

We used population-level data from the Alberta Kidney Disease Network (AKDN; www.akdn.info). The AKDN is a province-wide network that captures laboratory measurements, including serum creatinine, lipid profile, HbA_1c, and measures of urine protein (18). These data are linked to Alberta Health administrative data, which captures resource utilization for all provincial residents with public health insurance. All residents of Alberta are eligible for public health insurance, and >99% of residents participate in the government-sponsored insurance plan. Public health insurance covers the cost of all medically necessary physician visits, hospitalizations, investigations, and procedures. In addition, drug insurance is provided for all residents >65 years of age. Alberta Health data capture all health care utilization paid for through the provincial insurance plan. Vital statistics and health insurance registry data were also obtained from Alberta Health. Since public health insurance does not provide universal drug coverage for residents <65 years of age, drug costs are only available for patients >65 years of age. Physician visit, hospitalization, and ambulatory care costs are available for the entire cohort.

A cohort of patients 18 years of age and older with prevalent diabetes as of 1 May 2004 was identified from Alberta Health administrative data (18). Cases of diabetes were defined based on health care encounters incurred between 1 April 1995 and the date of cohort entry, 1 May 2004. We used the validated National Diabetes Surveillance System definition: two or more physician claims for diabetes (ICD-9 code 250.x) within 2 years, or one or more hospitalizations with an ICD-9 code of 250.x, selected from all available diagnostic codes on the Hospital Discharge Abstract prior to 31 March 2002 or equivalent ICD-10 codes (E10–14) after 31 March 2002 (19,20). Due to the use of a single diagnostic code for diabetes, it is often not possible to reliably distinguish between type 1 and type 2 diabetes using administrative data. Approximately 90% of prevalent diabetes cases are type 2 diabetes (1); therefore, all cost estimates based on this cohort are heavily weighted toward type 2 diabetes. The cohort was followed for 5 years, from 1 May 2004 to 30 April 2009.

Age, sex, Aboriginal status, and measures of socioeconomic status were determined from the registry file. These factors were included because they are known modifiers of health care utilization (21,22). Aboriginal race/ethnicity was defined by First Nations status. Socioeconomic status was categorized as high income (annual adjusted taxable family income ≥CDN $39,250), low income (annual adjusted taxable family income <CDN $39,250), and income support (provided to people and families with disabilities or with incomes below specified thresholds, e.g., CDN $14,880 for a two-parent family with three children) according to the Alberta health insurance registry (23). The duration of diabetes was calculated as time from diagnosis to 1 May 2004. Comorbidities were defined using administrative data for health care encounters during the 3 years prior to cohort entry. We calculated the Charlson comorbidity index (24,25), a weighted score of 17 comorbid conditions that has been shown to predict mortality (24,25). We also determined the proportion of patients having a history of cardiovascular disease, hypertension, coronary revascularization, cancer, or end-stage renal disease (ESRD).

Outpatient measures for HbA_1c, eGFR, and proteinuria (urine microalbumin-to-creatinine ratio [ACR] and urine dipstick) were included from 2 years prior to 6 months past the index date. For those patients not on dialysis at baseline, the eGFR was estimated from serum creatinine using the validated CKD-EPI equation (26). The mean of the two outpatient eGFR measurements made closest to the index date (May 1, 2004) was used to categorize patients into standard eGFR categories (eGFR >90, 60–90, 45–60, 30–45, 15–30, and <15 mL/min/m² not requiring dialysis) (27). Patients on dialysis were classified separately, and the eGFR was not considered for this subset. Proteinuria was assessed using the median measurement for urine protein. The ACR was used as the primary measure of proteinuria, supplemented by urine dipstick measurement when ACR was not available. Proteinuria was categorized into three levels: normal (ACR <30 mg/g or dipstick negative), mild (ACR 30–300 mg/g or dipstick 1+ or trace), and heavy (ACR >300 mg/g or dipstick ≥2+). We used the mean of the two HbA_1c measurements made closest to the index date to classify patients according to glycemic control: good (HbA_1c ≤7%), fair (7.1–7.9%), poor (8–9%), or inadequate (>9%).

The primary outcome was 5-year cumulative health care costs for the entire cohort. Drug costs were excluded for all patients in this primary outcome. As a secondary outcome, we studied 5-year cumulative costs for the subset of patients >65 years of age, including drug costs. We adopted the perspective of the health care payer; therefore, nonmedical costs (i.e., patient time and travel costs, as well as costs related to lost productivity) were not included. All costs are reported in 2010 CDN dollars. To inform the generalizability of our results, we determined the incidence of clinical outcomes (myocardial infarction, stroke, congestive heart failure, coronary revascularization, ESRD, and death) over the 5-year follow-up period, enabling a qualitative comparison with rates observed in other diabetes cohorts.

All analyses were performed for the cohort as a whole and for the subgroup >65 years of age (in whom drug costs were included). The mean 5-year direct medical costs of diabetes were determined in both cases. Since <3% of patients were lost to follow-up due to outmigration, imputation for missing costs was not required. Costs were further categorized according to baseline demographic and clinical characteristics, including comorbid illness and laboratory measurements.

The association between measures of baseline glycemic control, kidney function (including a separate category for people with ESRD on dialysis at baseline), proteinuria, and 5-year cost was determined using multivariate linear regression, controlling for age, sex, Aboriginal status, socioeconomic status, duration of diabetes, and Charlson index score. Given its ease of interpretation and to facilitate communication, we used a linear regression model using ordinary least squares estimation to assess factors associated with cost, and to estimate the adjusted mean 5-year cost for each category. We compared the fit of the linear regression model against that of four other candidate models, linear regression on log total costs with smearing retransformation (28) and three generalized linear models using the negative binomial (gamma) and inverse Gaussian distributions, and found that the linear regression model performed well based on mean absolute error, root mean squared error, Lin concordance, pseudo R squared, probability plots, quantile-quantile plots, and predicted versus observed mean costs. Ethics approval for the study was obtained from the conjoint health ethics review board at the University of Calgary. All analyses were undertaken using STATA, version 11.2 (College Station, TX).

Overall 138,662 patients with prevalent diabetes as of 1 May 2004 were included in the cohort. The cohort was 47.8% female, 41.7% of patients were >65 years of age, and the mean duration of diabetes was 5.3 years (Table 1). The most common comorbid condition was hypertension (60.5%), and all comorbidities increased with age. In patients >65 years of age, two-thirds had filled a prescription for one or more oral antidiabetic medication at baseline (62.2%), and approximately one-fifth (17.4%) of patients were on insulin. Over three-quarters (78.4%) filled a prescription for an antihypertensive agent, and approximately one-half had filled a prescription for a statin (52.8%). Compared with patients <65 years of age, patients >65 years of age were less likely to be Aboriginal (6.5 vs. 2.2%), less likely to be on income support (16.7 vs. 4.8%), had a longer average duration of diabetes (5.0 vs. 5.9 years), and had a higher burden of disease measured by a Charlson score of >3 (6.6 vs. 21.1%) (Table 1).

Laboratory measurements were available for the majority of the cohort: 73% for HbA_1c, 66% for proteinuria, and 84% for eGFR. These measurements revealed that 2.6% of the cohort had an eGFR <30, indicating severe renal impairment, 4.1% had heavy proteinuria, and 9.8% had inadequate glycemic control (>9% HbA_1c) (Table 1). Patients >65 years of age had a higher proportion of people with low eGFR and proteinuria but had better overall glycemic control. In all, 16% of the cohort did not have any of the three measurements. Patients without laboratory measurements tended to be younger (mean age 56.4 vs. 61.6 years) and had a lower burden of disease, measured by a Charlson score of >3 (4.6 vs. 14.2%) (Table 1).

Patient characteristics varied across strata of glycemic control and whether or not they had an HbA_1c measurement (Table 2). Those with inadequate glycemic control were younger, had a higher proportion with Aboriginal status and low socioeconomic status, and had less comorbid disease at baseline. To explore the notion that patients without HbA_1c measurements may have been misclassified, we compared patients under and over 65 years of age separately (Table 2). Among patients <65 years of age, those with unmeasured HbA_1c were less likely to have comorbid illness, possibly suggesting that some of the unmeasured patients were misclassified. Among those >65 years of age with unmeasured HbA_1c, some were taking diabetes medications. Those not taking diabetes medications were older and had more comorbid illness, indicating that rather than being misclassified, patients in this category may in fact be at a stage of illness where glycemic control has become less important. Although speculative, taken together, these findings suggest that patients without HbA_1c measurements represent a heterogeneous group comprised of misclassified patients, those with very mild disease, as well as frail older patients not being actively managed for their diabetes.

Unadjusted 5-year costs are presented in Table 3. The mean cumulative 5-year cost of caring for patients with diabetes in Alberta, excluding drug costs was CDN $26,978 per patient (IQR $3,401–30,141). Costs increased with age, Aboriginal status, lower socioeconomic status, longer duration of diabetes, and comorbidity. Medications accounted for $10,000 or approximately one-quarter of the 5-year medical costs for people >65 years of age; the mean cumulative 5-year cost for this group, including drug costs, was CDN $44,511 (IQR $13,758–56,333) per patient. Excluding drug costs, patients >65 years of age had consistently higher costs.

After stratification by kidney function, the adjusted cost of caring for patients with diabetes varied from $25,316 (for patients with eGFR >90 mL/min) to $115,348 (for patients not on dialysis with eGFR <15 mL/min) (Fig. 1). Patients who had no proteinuria had an adjusted mean cost of $24,531 per patient compared with $46,836 for patients with heavy proteinuria. Patients with good glycemic control had an adjusted mean cost of $27,064 per patient compared with $32,629 for patients with inadequate control. Similar trends were noted in the subgroup of patients >65 years of age when drug costs were included. Adjusted costs demonstrated a consistent trend of increasing cost with increasing severity of disease, as assessed by laboratory measures (Fig. 1).

The mean unadjusted costs for patients without an eGFR or proteinuria measurement were slightly higher compared with the normal categories. In contrast, patients without an HbA_1c measurement had lower mean costs than those with HbA_1c ≤7%. When unadjusted costs were examined by level of glycemic control, we noted a similar pattern across all categories of cost (Table 4); those with good control cost less across all categories of health care spending.

The coefficients and P values for the linear regression model for the overall cohort are presented in Table 5. In addition to the laboratory parameters described above, worsening socioeconomic status, Aboriginal status, and increasing Charlson index score were associated with increased 5-year costs. When kidney function was considered, people with progressively lower levels of kidney function had significantly higher costs. The model estimates that patients with an eGFR of <15 mL/min have average 5-year costs $91,419 higher compared with a patient with no renal impairment (eGFR >90 mL/min), patients with heavy proteinuria have costs $22,305 higher per patient compared with those with no proteinuria, and patients with inadequate glycemic control had costs $5,565 higher per patient compared with those with good glycemic control.

The mean 5-year cost of diabetes in Alberta was CDN $26,978 per patient, excluding drug costs, and CDN $44,511 per patient for patients >65 years of age, including drug costs. Our analysis demonstrates that after adjusting for sex, age, duration of diabetes, Aboriginal status, socioeconomic status, and comorbid illness, costs increased with worsening kidney function, higher levels of proteinuria, and worsening glycemic control. Adjusted costs increased fivefold for people with eGFR <15 mL/min/m² compared with eGFR >90 mL/min/m² ($115,348 vs. $25,316) and were twice as high in patients with heavy proteinuria compared with those with no proteinuria ($46,836 vs. $24,531). Costs increased less dramatically as glycemic control worsened; patients with inadequate glycemic control ($32,629 for patients with HbA_1c >9%) had 20% higher costs compared with patients with good control ($27,064 for HbA_1c <7%). Costs were also positively associated with age, Aboriginal status, lower socioeconomic status, duration of diabetes, and Charlson comorbidity index.

It is estimated that 2.8 million Canadians will have diabetes in 2012, and our analysis suggests that health care funders will spend approximately CDN $25 billion per year on the care of people with diabetes. This represents ∼12.5% of total health care spending in Canada, which was estimated at $200 billion annually in 2011 (29). This may be an underestimation of the costs of diabetes, given that we have not accounted for incident cases of diabetes in our 5-year projections nor have we included the cost of people with undiagnosed diabetes.

Other studies have noted an association between poor glycemic control (measured by HbA_1c) and cost (12–15); however, all were based on U.S. HMO populations and therefore may not have reflected patients at all socioeconomic levels. By differentiating HbA_1c levels into four distinct categories, we were able to show that costs do not appear to rise until HbA_1c increases beyond 7.9%. Similarly, Gilmer et al. (14) found that in HMO patients, higher HbA_1c was predictive of costs in patients with HbA_1c >7.5% but not in patients with HbA_1c of <7.5%. Our analysis further demonstrates that costs are associated with two other laboratory measures of direct relevance to patients with diabetes, namely eGFR, a marker of kidney function and proteinuria.

Although this study does not provide direct evidence that improvements in diabetes management would lead to cost reductions, our findings demonstrate a clear association between increased cost and suboptimal glycemic control and markers of kidney disease. It is plausible that better glycemic control in patients with HbA_1c >8% might delay or moderate the increasing costs associated with duration of diabetes through a reduction in diabetes complications (30–32). Wagner et al. (33) studied the association between improvement in glycemic control and cost in a retrospective cohort analysis and found that in patients with high baseline HbA_1c (≥10%) whose glycemic control improved, statistically significant cost savings were achieved. In addition, the optimized use of ACE inhibitors and angiotensin II receptor blockers or improved management of hypertension, through delaying the decline in kidney function (34–41), may lead not only to improvements in health but also to moderation of medical costs.

Mean 5-year costs were lower in patients who were not on oral antidiabetic medications or insulin at baseline (in those >65 years of age) and in patients who did not have laboratory testing during the 2 years prior to and 6 months past the index date. We are unable to determine the reasons why patients in these groups did not fill a prescription for diabetes medication or have laboratory testing within the measured time frame, and there are likely many factors involved. Our regression model included a “not measured” category for each laboratory marker, which is reflected in the adjusted cost estimates. Our data demonstrated that the “not measured” category was comprised of a heterogeneous group of patients with respect to demographics and comorbid illness, both of which were accounted for in the adjusted analyses. We did not adjust for medication use since this information was not available for the entire cohort, nor were we able to adjust for unknown factors, such as mild diabetes, misclassification, and more specific socioeconomic characteristics.

Our study was an observational cohort study and was therefore limited by potential confounding and by the data available. Although we controlled for all available confounders, including age, sex, socioeconomic status, and comorbid illness, we were not able to control for other potentially important variables, including ethnicity and education. In addition, although we found a strong association between glycemic control, proteinuria, and kidney function and costs, it is unknown whether improved management would in fact lead to a decrease in health care costs. Economic evaluations alongside controlled intervention studies are needed to draw definitive conclusions. In addition, we used an administrative definition of diabetes to define our cohort and, although this definition has been shown to perform well, some cases may have been misclassified. Finally, the cohort was representative of the population of Alberta. Although this may not be generalizable to all other settings, we expect that the relative differences in costs that we observed across different categories of laboratory measures would hold in other jurisdictions.

Our study also has many strengths. The large dataset makes it unlikely that any of the associations found between patient factors and cost were due to chance alone. Furthermore, our unique dataset included access to linked laboratory, clinical, and costing data, enabling us to study the association between disease severity markers and cost.

In summary, we have generated updated values for the 5-year cost of caring for patients with diabetes in a universal health care system, which will aid decision makers in planning future resource allocation. After controlling for clinical and demographic factors, we found that the cost of caring for people with diabetes increased with suboptimal glycemic control, proteinuria, and worsening kidney function. Future studies should assess whether strategies to improve these laboratory measures lead to reduced costs.

B.J.M. and B.H. were supported by New Investigator awards from the Canadian Institutes of Health Research. B.J.M., S.W.K., D.R., and B.H. were supported by Population Health Investigator awards from the Alberta Heritage Foundation for Medical Research. K.A.M. is supported by a Clinical Research Fellowship Award from Alberta Innovates Health Solutions.

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

K.A.M., B.J.M., and F.C. were involved in the concept and design, performed statistical analysis, interpreted results, and drafted the manuscript. B.C. and N.W. performed statistical analysis. S.W.K. and D.R. interpreted results. P.R. performed statistical analysis and interpreted results. B.H. acquired data and interpreted results. F.A. acquired data and performed statistical analysis. All authors reviewed the final manuscript. F.C. 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.

Preliminary results of this study were presented at the 2012 Clinician Investigator Trainee Association of Canada (CITAC) Young Investigator Forum.