Journal of Medical Economics
ISSN: 1369-6998 (Print) 1941-837X (Online) Journal homepage: http://www.tandfonline.com/loi/ijme20
Factors associated with increased healthcare costs in Medicare Advantage patients with type 2 diabetes enrolled in a large representative health insurance plan in the US S. Lane Slabaugh, Bradley H. Curtis, Gosia Clore, Haoda Fu & Dara P. Schuster To cite this article: S. Lane Slabaugh, Bradley H. Curtis, Gosia Clore, Haoda Fu & Dara P. Schuster (2015) Factors associated with increased healthcare costs in Medicare Advantage patients with type 2 diabetes enrolled in a large representative health insurance plan in the US, Journal of Medical Economics, 18:2, 106-112 To link to this article: http://dx.doi.org/10.3111/13696998.2014.979292
Accepted author version posted online: 21 Oct 2014. Published online: 06 Nov 2014. Submit your article to this journal
Article views: 94
View related articles
View Crossmark data
Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=ijme20 Download by: [University of Florida]
Date: 05 November 2015, At: 16:40
Journal of Medical Economics 1369-6998 doi:10.3111/13696998.2014.979292
Vol. 18, No. 2, 2015, 106–112
Article 0101.R1/979292 All rights reserved: reproduction in whole or part not permitted
Downloaded by [University of Florida] at 16:40 05 November 2015
Original article Factors associated with increased healthcare costs in Medicare Advantage patients with type 2 diabetes enrolled in a large representative health insurance plan in the US
S. Lane Slabaugh Comprehensive Health Insights, Inc., Louisville, KY, USA
Bradley H. Curtis Eli Lilly and Company, Indianapolis, IN, USA
Gosia Clore Comprehensive Health Insights, Inc., Louisville, KY, USA
Haoda Fu Dara P. Schuster Eli Lilly and Company, Indianapolis, IN, USA Address for correspondence: Dr Bradley H. Curtis, Eli Lilly and Company, Indianapolis IN 46205, USA. Tel.: 1-317-431-3203; [email protected] Keywords: Type 2 diabetes – Healthcare costs – Cost distribution – Quantile regression – Medicare Advantage Accepted: 17 October 2014; published online: 6 November 2014 Citation: J Med Econ 2015; 18:106–12
Abstract Aim: The objective of this study was to apply quantile regression (QR) methodology to a population from a large representative health insurance plan with known skewed healthcare utilization attributes, co-morbidities, and costs in order to identify predictors of increased healthcare costs. Further, this study provides comparison of the results to those obtained using ordinary least squares (OLS) regression methodology. Methods: Members diagnosed with Type 2 Diabetes and with 24 months of continuous enrollment were included. Baseline patient demographic, clinical, consumer/behavioural, and cost characteristics were quantified. Quantile regression was used to model the relationship between the baseline characteristics and total healthcare costs during the follow-up 12 month period. Results: The sample included 83,705 patients (mean age ¼ 70.6 years, 48% male) residing primarily in the southern US (78.1%); 81.2% of subjects were on oral-only anti-diabetic therapy. Co-morbid conditions included nephropathy (43.5%), peripheral artery disease (26.4%), and retinopathy (18.0%). Variables with the strongest relationship with costs during the follow-up period included outpatient visits, ER visits, inpatient visits, and Diabetes Complications Severity Index score during the baseline period. In the top cost quantiles, each additional glycohemoglobin (HbA1c) test was associated with cost savings ($1400 in the 98th percentile). Stage 4 and Stage 5 chronic kidney disease were associated with an incremental cost increase of $33,131 and $106,975 relative to Stage 1 or no CKD in the 98th percentile ($US). Conclusions: These results demonstrate that QR provides additional insight compared to traditional OLS regression modeling, and may be more useful for informing resource allocation to patients most likely to benefit from interventions. This study highlights that the impact of clinical and demographic characteristics on the economic burden of the disease vary across the continuum of healthcare costs. Understanding factors that drive costs on an individual patient level provide important insights that will help in ameliorating the clinical, humanistic, and economic burden of diabetes.
Introduction Type 2 diabetes mellitus (T2DM) affects millions of Americans and is associated with a myriad of co-morbid conditions and complications1. As a result, T2DM places a large and growing financial burden on the healthcare system, and 106
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
www.informahealthcare.com/jme ! 2015 Informa UK Ltd
Downloaded by [University of Florida] at 16:40 05 November 2015
Journal of Medical Economics
several studies conducted over recent years have documented the trajectory of this concerning trend2–4. For example, the total estimated cost of diagnosed diabetes in 2012 in the US was $245 billion, including $176 billion in direct medical costs and $69 billion in reduced productivity; people with diagnosed diabetes on average have medical expenditures 2.3-times higher than estimated medical expenditures in the absence of diabetes4. While the overall numbers seem overwhelming, it is increasingly clear for those seeking to manage these costs that an understanding of the distribution of the burden is essential. With this information interventions may be targeted at those patients most likely to benefit and resources be directed in the most appropriate manner. A recent paper reported that the top 10% of newly diagnosed patients with T2DM account for 68% of healthcare expenditure of that group5, with most medical costs related to management of complications and co-morbidities6. In addition, the costs associated with T2DM are expected to increase in the coming decades due to the growth of the elderly population and the anticipated trends in their health. While the epidemic of T2DM is clearly linked to increasing rates of overweight and obesity in the US population, projections by the Centers for Disease Control and Prevention (CDC) suggest that, even if diabetes incidence rates level off, the prevalence of diabetes will double in the next 20 years due to the aging of the population1. Currently more than 25% of the US population aged over 65 years has diabetes1, and the aging of the overall population is then in itself a significant driver of the diabetes epidemic7. Understanding this data requires careful analysis and interpretation. Ordinary least-squares (OLS) regression models the relationship between one or more covariates and the conditional mean of a response variable8. In other words, it models the average of the response variable across the entire study population. In contrast, quantile regression analysis models the relationship between covariates and the conditional quantiles of the response. Therefore, quantile regression models the impact of each variable within each quantile of the response variable. For that reason it is especially useful in applications where extremes of the response variable are important, such as cost studies where both upper and lower quantiles of cost may be critical in appropriately allocating interventions or resources9,10. Accordingly, the objective of this study was to apply quantile regression methodology to a population from a large representative health insurance plan with known skewed healthcare utilization attributes, co-morbidities, and costs (older patients with T2DM) in order to identify predictors of increased healthcare costs. Further, we provide a comparison of the results to those obtained using OLS regression methodology. ! 2015 Informa UK Ltd www.informahealthcare.com/jme
Volume 18, Number 2
February 2015
Methods This retrospective cohort study was conducted using the Humana medical, pharmacy, and laboratory claims databases. The study population consisted of Humana Medicare Advantage with prescription drug coverage (MAPD) members aged 19–89 years as of July 1, 2010 with a diagnosis of T2DM. Humana Inc. (Louisville, KY) is a large national health insurance company in the US. The Medicare program is a federal health insurance program available to all US citizens 65 years of age or older, young people with certain disabilities, and anyone with end-stage renal disease. Those eligible for Medicare can choose to enroll in a MAPD plan. These plans are offered by private companies, which are paid by the Medicare administration to cover patient benefits. The research protocol was reviewed and approved prior to study initiation by an independent institutional review board, Schulman Associates, Inc. (Cincinnati, OH), and received a waiver of Health Insurance Portability and Accountability Act authorization and waiver of informed consent. Patients were required to be continuously enrolled in a Humana MAPD plan between July 1, 2010 and June 30, 2012, unless terminated by death during the final 12 months. The first 12 months of the study period (July 1, 2010 to June 30, 2011) was used as the baseline period and final 12 months (July 1, 2011 to June 30, 2012) were used as the follow-up period. Patients with T2DM were defined as patients meeting any of the following criteria: (1) At least one medical claim with a T2DM diagnosis with ICD-9-CM codes (250.x0 or 250.x2), no medical claims with a type I diabetes (T1DM) diagnosis (250.x1 or 250.x3), and at least one pharmacy claim for a non-insulin antidiabetic medication; or (2) Claims for both T1DM and T2DM with the ratio of T2DM:T1DM claims 0.5 and either of the following: at least one pharmacy claim for non-insulin antidiabetic medication, or a gap of at least 6 months at any time during the study period with no pharmacy claims for insulin. Additionally, all patients were required to have valid HbA1C values in both the baseline and follow-up periods that were at least 6 months apart, and a serum creatinine laboratory value during the baseline period. A large number of baseline descriptive variables were captured or calculated. The variables included, but were not limited to, demographic information [age, gender, geographic location, race/ethnicity (white, black, Hispanic, or other, as provided by the U.S. Social Security Administration), insurance plan type] and clinical descriptors, including point of service, anti-diabetic prescription treatment type, provider specialty (specialist vs primary), co-morbidities, Consumer Circle, Proportion of days covered (PDC), estimated glomerular filtration rate (eGFR),
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
107
Downloaded by [University of Florida] at 16:40 05 November 2015
Journal of Medical Economics
Volume 18, Number 2
February 2015
CHADs2 score, and Diabetes Complications Severity Index (DCSI), among others. Baseline demographic variables were calculated as of July 1, 2010 (the beginning of the baseline period). Baseline clinical characteristics and baseline cost variables were calculated using the 1-year baseline period beginning July 1, 2010. The CHADs2 score, a clinical prediction rule for estimating the risk of stroke in patients with non-rheumatic atrial fibrillation, was calculated by adding 1 point each for congestive heart failure [C], hypertension [H], age 475 years [A], and diabetes [D], and 2 points for prior stroke or TIA [s2]11. CHADs2 was calculated based on a count of complications captured in the medical claims via ICD-9 codes. Consumer Circles are a proprietary Humana population segmentation, which clusters the members into nine ‘circles’ based on factors such as socio-economic status, age, life stage, spending behavior, hobbies, and several hundred other characteristics. The DCSI was developed to predict adverse outcomes including hospitalization and mortality based on the number and severity of complications associated with diabetes, and is based on a summary score derived from diagnostic and laboratory data12. A modified version of the DCSI, which omits laboratory data in order to allow for the use of calculation DCSI using a claims database, has been validated to explain concurrent and future medical costs13. eGFR was calculated using the following formula: 186 ðCreatinine=88:4Þ1:154 ðAgeÞ0:203
burden on the healthcare system. Confidence intervals (95%) are provided around the estimate of effect for each quantile. All data extraction and statistical analyses were conducted using SAS v.9.3.1, SAS Enterprise Guide v.5.1, and SAS Enterprise Miner v.12.1 (SAS Institute, Cary, NC).
Results A total of 83,705 patients with T2DM met the inclusion criteria for this study. Demographic characteristics are presented in Table 1. The mean age was 70.57 years and males made up 48% of the population. Patients resided primarily in the southern and midwestern US (78.1% and 11.6%, respectively), and lived primarily in urban and suburban areas (76.4% and 17.4%, respectively). The largest race/ ethnicity category was white (74.5%), followed by black (18.1%). These geographic and race/ethnicity trends are consistent with the overall Humana MAPD population. The proportion of the T2DM study population that died during the follow-up year was 4.8%. Baseline clinical characteristics are presented in Table 2. The mean DCSI score for the population was 2.81 (SD ¼ 2.24). The most common anti-diabetic treatment categories were ‘Oral Antidiabetics Only’ (81.2%) Table 1. Demographic characteristics. Characteristic (n ¼ 83,705)
Units
T2DM
ð0:742 if femaleÞ ð1:210 if blackÞ where age was based on the beginning of the study period (July 1, 2010), and female gender and black race were determined as binary variables. A gradient boosting tree (GBT) method was used to narrow down the vast number of candidate variables. The gradient boosting tree method has been shown to have high prediction accuracy and to be useful in variable selection based on their importance14. Baseline cost variables were found to be highly correlated with follow-up costs and masked the relative importance of other variables. Therefore, baseline cost variables were removed from the GBT for the final selection. Only variables with relative importance of 0.10 or greater were included in the final linear least square model and quantile regression model. Quantile regression was used to test for associations between the dependent variable (healthcare costs) and the independent variables which were previously described. In particular, covariate effects related to 10%, 25%, 50%, 75%, 80%, 85%, 90%, 95%, and 98% quantiles of the costs were reported. These quantiles were chosen to focus on the patients at the higher end of the cost spectrum. These patients are of particular interest because of the anticipated complexity of their disease and financial 108
Age, years Gender Male Female Geographic region Northeast Midwest South West RUCA Urban Suburban Rural Unknown Race/ethnicity White Black Hispanic Other/unknown Plan characteristics No special characteristics LIS status only Dual eligibility only LIS status and Dual eligibility Plan type HMO PPO Other Died in Year 2
mean (SD)
70.57 (8.08)
n (%) n (%)
40,279 (48.10%) 43,426 (51.90%)
n (%) n (%) n (%) n (%)
907 (1.10%) 9,719 (11.60%) 65,403 (78.10%) 7,676 (9.20%)
n (%) n (%) n (%) n (%)
63,970 (76.40%) 14,557 (17.40%) 4,320 (5.20%) 858 (1.00%)
n (%) n (%) n (%) n (%)
62,319 (74.50%) 15,125 (18.10%) 2,917 (3.50%) 3,344 (4.00%)
n (%) n (%) n (%) n (%)
63,355 (75.70%) 7,968 (9.50%) 68 (0.10%) 12,314 (14.70%)
n (%) n (%) n (%) n (%)
52,835 (63.10%) 7,736 (9.20%) 23,134 (27.60%) 3,980 (4.80%)
SD, standard deviation; RUCA, Rural-Urban Commuting Area; LIS, Low income subsidy; HMO, Health Maintenance Organization; PPO, Preferred provider organization.
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
www.informahealthcare.com/jme ! 2015 Informa UK Ltd
Journal of Medical Economics
Downloaded by [University of Florida] at 16:40 05 November 2015
DCSI score, Mean (SD) Number of ED visits, Mean (SD) Number of inpatient hospitalizations, Mean (SD) Number of outpatient visits, Mean (SD) Anti-diabetic treatment categories, n (%) No anti-diabetic therapy Oral anti-diabetic only Non-insulin injectable anti-diabetic only Insulin only Oral plus non-insulin injectable anti-diabetics Oral anti-diabetic plus insulin Non-insulin injectable anti-diabetic plus insulin Oral anti-diabetic, non-insulin anti-diabetic, and insulin Adherence to anti-diabetic therapy (PDC), Mean (SD) Persistence on anti-diabetic therapy, Mean (SD) Pill burden (non-insulin anti-diabetic), Mean (SD) Mail-order pharmacy use, n (%) Count of the number of unique anti-diabetic classes, Mean (SD) Duplicate therapy, n (%) Number of claims for glucose test strips, Mean (SD) Number of emergent hypoglycemic events, Mean (SD) Number of non-emergent hypoglycemic events, Mean (SD) Comorbidities, n (%) Depression Low testosterone Erectile dysfunction Retinopathy Nephropathy Chronic kidney disease Stage I Stage II Stage III Stage IV Stage V End stage renal disease Neuropathy Cardiovascular disease
T2DM 2.81 (2.24) 0.58 (1.4) 1.85 (5.89) 25.91 (23.33) 0 (0.00%) 67,946 (81.20%) 15 (0.00%) 677 (0.80%) 506 (0.60%) 14,284 (17.10%) 28 (0.00%) 249 (0.30%) 0.86 (0.16) 250.93 (137.57) 1.23 (0.52) 45,699 (54.60%) 1.7 (0.8) 2,997 (3.60%) 0.72 (1.38)
Characteristic (n ¼ 83,705) 7.0% Primary healthcare professional, n (%) Endocrinologist Nephrologist Primary care/Other Renal function lab results, n (%) Microalbumin (mg/day) Normal (530) Elevated (30) Missing Serum Creatinine Stage 1 (eGFR 90þ) Stage 2 (eGFR 60–89) Stage 3 (eGFR 30–59) Stage 4 (eGFR 15–29) Stage 5 (eGFR 515) Passed diabetes-related HEDIS measures, n (%) Cholesterol Screening (Yes) Cholesterol Controlled (Yes) Eye Exam (Yes) Kidney Disease Monitoring (Yes)
T2DM 50,829 (60.70%) 2,167 (2.60%) 3,202 (3.80%) 78,336 (93.60%) 20,552 (24.60%) 8,101 (9.70%) 55,052 (65.80%) 16,103 (19.20%) 42,659 (51.00%) 23,315 (27.90%) 1,392 (1.70%) 236 (0.30%) 47,751 (57.00%) 35,085 (41.90%) 35,635 (42.60%) 46,452 (55.50%)
DCSI, diabetes complication severity index; ED, emergency department; PDC, proportion of days covered; UAP, unstable angina pectoris; NSTEMI, non-ST segment elevation myocardial infarction; STEMI, ST segment elevation myocardial infarction; CHADS2, congestive heart failure, hypertension, age ¼ 75 years, diabetes mellitus, stroke; eGFR, estimated glomerular filtration rate; HEDIS, healthcare effectiveness data and information set.
0.09 (1.69) 0.32 (3.17) 12,775 (15.30%) 1,852 (2.20%) 6,027 (7.20%) 15,100 (18.00%) 36,408 (43.50%) 30,420 (36.30%) 3,181 (3.80%) 9,944 (11.90%) 19,732 (23.60%) 2,217 (2.60%) 801 (1.00%) 801 (1.00%) 34,018 (40.60%) 40,954 (48.90%)
Major adverse cardiac events (MACE) plus including, n (%) UAP 2,520 (3.00%) NSTEMI or STEMI 1,628 (1.90%) Stroke 9,831 (11.70%) Heart failure 12,277 (14.70%) Peripheral artery disease 22,135 (26.40%) Atherosclerosis 13,756 (16.40%) Aortic aneurysm 1,982 (2.40%) Osteoporosis 8,757 (10.50%) CHADs2 stroke risk score, Mean (SD) 2.56 (0.94) HbA1c values Number of HbA1c values during the 2.45 (1.19) baseline period, Mean (SD) Median (range) 2 (1–19) Categorical HbA1c values, n (%) 49.0% 6,205 (7.40%) 48.0% and 9.0% 7,169 (8.60%) 47.0% and 8.0% 19,502 (23.30%)
(continued )
! 2015 Informa UK Ltd www.informahealthcare.com/jme
February 2015
Table 2.Continued.
Table 2. Baseline clinical characteristics. Characteristic (n ¼ 83,705)
Volume 18, Number 2
and ‘Oral Antidiabetic plus Insulin’ (17.1%). The most common oral anti-diabetics used were metformin and sulfonylureas (used by 70.7% and 52.0% of patients, respectively) followed by insulin (18.2%). The mean number of HbA1c values available during the baseline period was 2.45, (median ¼ 2, range ¼ 1–19). Nearly two-thirds (60.7%) of the T2DM population had a baseline HbA1c value 7.0%. Common co-morbid conditions included nephropathy (43.5%), neuropathy (40.6%), and cardiovascular disease (48.9%). The proportion of patients with stage 3 Chronic Kidney Disease (CKD) or worse was very similar using either of the two employed sources. However, the proportion with Stage 2 CKD was much larger based on eGFR values than identified from diagnosis codes (51.0% vs 11.9%, respectively). Table 3 shows the results of the gradient boosting tree that was used to select variables for inclusion in the QR analysis. After excluding variables with a relative importance 50.10 there were 17 variables remaining. Two baseline cost variables (total pharmacy costs and diabetesrelated medical costs) were excluded due to co-linearity with other cost variables.
Regression models Figure 1 shows the results of the OLS and QR regression analyses. The quantile regression analysis showed that the association between many of the study co-variates and
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
109
Journal of Medical Economics
Volume 18, Number 2
February 2015
Table 3. Gradient boosting tree results.
Downloaded by [University of Florida] at 16:40 05 November 2015
Variable
Outpatient visits ED visits Inpatient visits Diabetes Complications Severity Index Cardiovascular disease Atherosclerosis Chronic kidney disease (yes/no) Nephropathy Number of HbA1c values during the baseline period Age at Baseline Pill Burden Chronic kidney disease stage (by eGFR) Count of the number of unique anti-diabetic classes Chronic kidney disease stage (by ICD-9 code) HbA1c value during the baseline period HEDIS measures - eye exam CHADs2 Stroke Risk score
Discussion Relative importance 1.00000 0.79173 0.78540 0.40434 0.32044 0.23325 0.20989 0.18613 0.16103 0.15940 0.15790 0.14769 0.14491 0.14083 0.13922 0.13258 0.10811
ED, emergency department; eGFR, estimated glomerular filtration rate; HEDIS, healthcare effectiveness data and information set; CHADs2, congestive heart failure, hypertension, age ¼ 75 years, diabetes mellitus, stroke.
follow-up costs were found to change across the cost quantiles. The variables with the strongest relationships with costs during the follow-up period included the four baseline cost variables, baseline measures of healthcare utilization (outpatient visits, ER visits, inpatient visits), and DCSI score during the baseline period. These variables had a statistically significant association with costs at nearly every cost quantile, and had the largest impact at the highest cost quantiles. Among them, baseline DCSI score was associated with the greatest increase in the top quantiles ($1490 and $1788 at the 95th and 98th percentiles, respectively). ER visits were associated with cost increases of 4$500 at the 50th percentile and above. Several other variables also exhibited interesting QR curves. The number of HbA1c tests performed during the baseline period was associated with a statistically significant but relatively small decrease in the total healthcare costs for patients in the lower quantiles (5$260 for patients in the 75th percentile and lower). However, in the top cost quantiles, each additional HbA1c test was associated with noteworthy cost savings (as much as $1400 in the 98th percentile). A similar finding is observed for baseline pill burden. The baseline variables associated with the largest cost increases during the follow-up period was late-stage CKD (as identified by eGFR). Stage 4 and Stage 5 CKD were associated with an incremental cost increase of $33,131 and $106,975 relative to Stage 1 or no CKD in the 98th percentile. Late-stage CKD (as identified by ICD-9-CM code) was found to have obscure patterns and did not have a consistently higher impact on costs relative to having Stage 1 or no CKD. 110
To our knowledge this is the first study that has used quantile regression in an examination of economic/cost drivers among patients with diabetes. While several previous studies have quantified the impact of patient demographic and clinical characteristics on the healthcare cost of patients with diabetes, all of these analyses utilized the OLS regression model approach6,15,16. The method of least squares results in estimates that approximate the conditional mean of the response variable given certain values of the predictor variables, whereas quantile regression aims at estimating different quantiles of the response variable and is particularly useful when modeling cost data which are not normally distributed. High cost patients (i.e., in a high quantile) are intuitively of particular interest to researchers, healthcare providers, and payers both from the perspective of improving health and health outcomes among high risk patients, as well as reducing the costs associated with their care10. Our data indicate that there are a few key co-morbid conditions that have only a marginal impact on future healthcare costs among patients in the lower quantiles of costs, but tremendous impact on costs in higher quantiles. For example the DCSI score or diagnosis of cardiovascular disease were found to have very little impact on cost in the lower cost quantiles but a sharply increased impact among patients in the higher quantiles. As such our findings appear to validate the existing literature that higher DCSI scores are a valid predictor of increased future healthcare costs13,17. Other co-variates, such as ER visits, showed a more steady increase in their impact on cost as the quantile increased. The largest drivers of costs in this study population were stages 4 and 5 CKD. The OLS model shows that the average increase in costs among patients with latestage CKD are substantial, but fails to highlight the enormous increase seen for patients in the highest cost quantile that is unveiled by the quantile regression analysis. Interestingly, the QR cost curves for late-stage CKD identified using claims data do not show patterns that are consistent with late-stage CKD identified using eGFR values. Instead, the QR curves for late-stage CKD identified using claims data showed few distinguishable trends. This finding suggests that documentation of stages of CKD in claims data may be highly inaccurate and caution should be used by those wanting to determine patients to target with interventions focused on patients with chronic kidney disease. While such information is clearly important in tailoring interventions, the practicalities of this approach warrant further investigation. All models suffer from limitations inherent to their very nature, as they rely on the quality of the data to which they are applied. Limitations common to research using administrative claims data apply in
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
www.informahealthcare.com/jme ! 2015 Informa UK Ltd
Downloaded by [University of Florida] at 16:40 05 November 2015
Journal of Medical Economics
Volume 18, Number 2
February 2015
Figure 1. Quantile regression and ordinary least squares estimates for total healthcare costs. $ ¼ U.S. dollars; ED ¼ emergency department; CKD ¼ chronic kidney disease; eGFR ¼ estimated glomerular filtration rate; CHAD2S ¼ congestive heart failure, hypertension, age, diabetes mellitus, prior stroke or transient ischemic attack or thromboembolism.
this study. These include lack of certain information in the database (e.g., health behavior information) and error in claims coding. Causal inferences cannot be ascertained from the current study, as it is a descriptive observational study using retrospective claims data. The use of retrospective secondary coded data does not provide as accurate an assessment as an empirical evaluation by a clinician and is neither randomized nor controlled. It will, therefore, be important to replicate the results of this study in a prospective manner, perhaps including outcome measures other than cost that may be helpful to patients, healthcare providers, and payers. The current study uses data only from Humana MAPD members; therefore, the results may not be generalizable to the general Medicare population. However, Humana is a ! 2015 Informa UK Ltd www.informahealthcare.com/jme
large national health plan with membership in every region of the US. This sample likely represents a slightly healthier population than the general Medicare population of patients with T2DM, as the requirement for having HbA1c values during both the baseline and follow-up periods may bias inclusion towards patients that are being more closely monitored by a provider. This may be the reason that the average HbA1c value was low relative to other large population studies18.
Conclusion Diabetes is a highly prevalent disease that places substantial financial burden on patients, their families, and the healthcare system. In order to help patients manage this
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
111
Downloaded by [University of Florida] at 16:40 05 November 2015
Journal of Medical Economics
Volume 18, Number 2
February 2015
chronic disease, communities, providers, and payers offer a wide range of interventions and disease management programs to patients. However, resources available to support such programs are limited. Quantile regression is a useful econometric tool that may aid in shaping individual patient management based on characteristics most likely to have a significant impact, as opposed to those based on population mean or median estimates of effect. This study demonstrates that the impact of clinical and demographic characteristics on the economic burden of the disease vary across the continuum of healthcare costs. A better understanding of the factors that drive costs on an individual patient level will provide important insights that will help in ameliorating the clinical, humanistic, and economic burden of diabetes.
2. 3. 4. 5.
6.
7.
8. 9.
Transparency
10.
Declaration of funding This research was funded by Eli Lilly and Company. Declaration of financial/other relationships This study was funded by Eli Lilly and company and was conducted under the oversight of an ongoing research collaboration agreement between Eli Lilly and Company and Humana, Inc. SLS is an employee of Humana, Inc. GC is an employee of Comprehensive Health Insights, the business entity within Humana that provides research for external clients. BHC, HF, and DPS, are employees and shareholders of Eli Lilly and Company, a pharmaceutical company that is involved in the manufacture and sales of a variety of products, including those used in the management of Type 2 diabetes. JME peer reviewers on this manuscript have no relevant financial or other relationships to disclose.
11.
12.
13.
14. 15. 16. 17.
References 1. US Centers for Disease Control and Prevention. Diabetes Report Card 2012. Atlanta, GA: Centers for Disease Control and Prevention, US Department of
112
18.
Health and Human Services, 2012. http://www.cdc.gov/diabetes/pubs/pdf/ DiabetesReportCard.pdf. Accessed June 25, 2014 American Diabetes Association. Economic costs of diabetes in the US in 2002. Diabetes Care 2003;26:917-32 American Diabetes Association. Economic costs of diabetes in the US in 2007. Diabetes Care 2008;31:596-615 American Diabetes Association. Economic costs of diabetes in the US in 2012. Diabetes Care 2013;36:1033-46 Maguire J, Dhar V. Comparative effectiveness for oral anti-diabetic treatments among newly diagnosed type 2 diabetics: data-driven predictive analytics in healthcare. Heal Syst 2012;2:73-92 Li R, Bilik D, Brown MB, et al. Medical costs associated with type 2 diabetes complications and comorbidities. Am J Manag Care 2013;19: 421-430 Boyle JP, Thompson TJ, Gregg EW, et al. Projection of the year 2050 burden of diabetes in the US adult population: dynamic modeling of incidence, mortality, and pre-diabetes prevalence. Popul Health Metr 2010;8:29 Montgomery DC, Peck EA, Vining GG. Introduction to Linear Regression Analysis. 5th edn. Hoboken, NJ: Wiley, 2012 Koenker R. Quantile regression. New York, NY: Cambridge University Press, 2005 Chen SL. An introduction to Quantile Regression and the QUANTREG Procedure. Paper 213-30, Statistics and Data Analysis. Cary, NC: SAS Institute, http://www2.sas.com/proceedings/sugi30/213-30.pdf. Accessed June 25, 2014 Gage BF, Waterman AD, Shannon W, et al. Validation of clinical classification schemes for predicting stroke: results from the National Registry of Atrial Fibrillation. JAMA 2001;285:2864-70 Young BA, Lin E, Von Korff M, et al. Diabetes complications severity index and risk of mortality, hospitalization, and healthcare utilization. Am J Manag Care 2008;14:15-23 Chang H-Y, Weiner JP, Richards TM, et al. Validating the adapted Diabetes Complications Severity Index in claims data. Am J Manag Care 2012;18:721-6 Hastie T, Tibshirani R, Friedman J. The elements of statistical learning. New York: Springer, 2009 Bhattacharyya SK, Else BA. Medical costs of managed care in patients with type 2 diabetes mellitus. Clin Therapeut 1999;21:2131-42 Brandle M, Zhou H, Smith BR, et al. The direct medical cost of type 2 diabetes. Diabetes Care 2003;26:2300-4 Chen HL, Hsiao FY. Risk of hospitalization and healthcare cost associated with diabetic complication severity index in Taiwan’s National Health Insurance Research Database. J Diabetes Complications 2014;28:612-6 Dodd AH, Colby MS, Boye KS, et al. Treatment approach and HbA1c control among US adults with type 2 diabetes: NHANES 1999-2004. Curr Med Res Opin 2009;27:1605-13
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
www.informahealthcare.com/jme ! 2015 Informa UK Ltd
ISSN: 1369-6998 (Print) 1941-837X (Online) Journal homepage: http://www.tandfonline.com/loi/ijme20
Factors associated with increased healthcare costs in Medicare Advantage patients with type 2 diabetes enrolled in a large representative health insurance plan in the US S. Lane Slabaugh, Bradley H. Curtis, Gosia Clore, Haoda Fu & Dara P. Schuster To cite this article: S. Lane Slabaugh, Bradley H. Curtis, Gosia Clore, Haoda Fu & Dara P. Schuster (2015) Factors associated with increased healthcare costs in Medicare Advantage patients with type 2 diabetes enrolled in a large representative health insurance plan in the US, Journal of Medical Economics, 18:2, 106-112 To link to this article: http://dx.doi.org/10.3111/13696998.2014.979292
Accepted author version posted online: 21 Oct 2014. Published online: 06 Nov 2014. Submit your article to this journal
Article views: 94
View related articles
View Crossmark data
Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=ijme20 Download by: [University of Florida]
Date: 05 November 2015, At: 16:40
Journal of Medical Economics 1369-6998 doi:10.3111/13696998.2014.979292
Vol. 18, No. 2, 2015, 106–112
Article 0101.R1/979292 All rights reserved: reproduction in whole or part not permitted
Downloaded by [University of Florida] at 16:40 05 November 2015
Original article Factors associated with increased healthcare costs in Medicare Advantage patients with type 2 diabetes enrolled in a large representative health insurance plan in the US
S. Lane Slabaugh Comprehensive Health Insights, Inc., Louisville, KY, USA
Bradley H. Curtis Eli Lilly and Company, Indianapolis, IN, USA
Gosia Clore Comprehensive Health Insights, Inc., Louisville, KY, USA
Haoda Fu Dara P. Schuster Eli Lilly and Company, Indianapolis, IN, USA Address for correspondence: Dr Bradley H. Curtis, Eli Lilly and Company, Indianapolis IN 46205, USA. Tel.: 1-317-431-3203; [email protected] Keywords: Type 2 diabetes – Healthcare costs – Cost distribution – Quantile regression – Medicare Advantage Accepted: 17 October 2014; published online: 6 November 2014 Citation: J Med Econ 2015; 18:106–12
Abstract Aim: The objective of this study was to apply quantile regression (QR) methodology to a population from a large representative health insurance plan with known skewed healthcare utilization attributes, co-morbidities, and costs in order to identify predictors of increased healthcare costs. Further, this study provides comparison of the results to those obtained using ordinary least squares (OLS) regression methodology. Methods: Members diagnosed with Type 2 Diabetes and with 24 months of continuous enrollment were included. Baseline patient demographic, clinical, consumer/behavioural, and cost characteristics were quantified. Quantile regression was used to model the relationship between the baseline characteristics and total healthcare costs during the follow-up 12 month period. Results: The sample included 83,705 patients (mean age ¼ 70.6 years, 48% male) residing primarily in the southern US (78.1%); 81.2% of subjects were on oral-only anti-diabetic therapy. Co-morbid conditions included nephropathy (43.5%), peripheral artery disease (26.4%), and retinopathy (18.0%). Variables with the strongest relationship with costs during the follow-up period included outpatient visits, ER visits, inpatient visits, and Diabetes Complications Severity Index score during the baseline period. In the top cost quantiles, each additional glycohemoglobin (HbA1c) test was associated with cost savings ($1400 in the 98th percentile). Stage 4 and Stage 5 chronic kidney disease were associated with an incremental cost increase of $33,131 and $106,975 relative to Stage 1 or no CKD in the 98th percentile ($US). Conclusions: These results demonstrate that QR provides additional insight compared to traditional OLS regression modeling, and may be more useful for informing resource allocation to patients most likely to benefit from interventions. This study highlights that the impact of clinical and demographic characteristics on the economic burden of the disease vary across the continuum of healthcare costs. Understanding factors that drive costs on an individual patient level provide important insights that will help in ameliorating the clinical, humanistic, and economic burden of diabetes.
Introduction Type 2 diabetes mellitus (T2DM) affects millions of Americans and is associated with a myriad of co-morbid conditions and complications1. As a result, T2DM places a large and growing financial burden on the healthcare system, and 106
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
www.informahealthcare.com/jme ! 2015 Informa UK Ltd
Downloaded by [University of Florida] at 16:40 05 November 2015
Journal of Medical Economics
several studies conducted over recent years have documented the trajectory of this concerning trend2–4. For example, the total estimated cost of diagnosed diabetes in 2012 in the US was $245 billion, including $176 billion in direct medical costs and $69 billion in reduced productivity; people with diagnosed diabetes on average have medical expenditures 2.3-times higher than estimated medical expenditures in the absence of diabetes4. While the overall numbers seem overwhelming, it is increasingly clear for those seeking to manage these costs that an understanding of the distribution of the burden is essential. With this information interventions may be targeted at those patients most likely to benefit and resources be directed in the most appropriate manner. A recent paper reported that the top 10% of newly diagnosed patients with T2DM account for 68% of healthcare expenditure of that group5, with most medical costs related to management of complications and co-morbidities6. In addition, the costs associated with T2DM are expected to increase in the coming decades due to the growth of the elderly population and the anticipated trends in their health. While the epidemic of T2DM is clearly linked to increasing rates of overweight and obesity in the US population, projections by the Centers for Disease Control and Prevention (CDC) suggest that, even if diabetes incidence rates level off, the prevalence of diabetes will double in the next 20 years due to the aging of the population1. Currently more than 25% of the US population aged over 65 years has diabetes1, and the aging of the overall population is then in itself a significant driver of the diabetes epidemic7. Understanding this data requires careful analysis and interpretation. Ordinary least-squares (OLS) regression models the relationship between one or more covariates and the conditional mean of a response variable8. In other words, it models the average of the response variable across the entire study population. In contrast, quantile regression analysis models the relationship between covariates and the conditional quantiles of the response. Therefore, quantile regression models the impact of each variable within each quantile of the response variable. For that reason it is especially useful in applications where extremes of the response variable are important, such as cost studies where both upper and lower quantiles of cost may be critical in appropriately allocating interventions or resources9,10. Accordingly, the objective of this study was to apply quantile regression methodology to a population from a large representative health insurance plan with known skewed healthcare utilization attributes, co-morbidities, and costs (older patients with T2DM) in order to identify predictors of increased healthcare costs. Further, we provide a comparison of the results to those obtained using OLS regression methodology. ! 2015 Informa UK Ltd www.informahealthcare.com/jme
Volume 18, Number 2
February 2015
Methods This retrospective cohort study was conducted using the Humana medical, pharmacy, and laboratory claims databases. The study population consisted of Humana Medicare Advantage with prescription drug coverage (MAPD) members aged 19–89 years as of July 1, 2010 with a diagnosis of T2DM. Humana Inc. (Louisville, KY) is a large national health insurance company in the US. The Medicare program is a federal health insurance program available to all US citizens 65 years of age or older, young people with certain disabilities, and anyone with end-stage renal disease. Those eligible for Medicare can choose to enroll in a MAPD plan. These plans are offered by private companies, which are paid by the Medicare administration to cover patient benefits. The research protocol was reviewed and approved prior to study initiation by an independent institutional review board, Schulman Associates, Inc. (Cincinnati, OH), and received a waiver of Health Insurance Portability and Accountability Act authorization and waiver of informed consent. Patients were required to be continuously enrolled in a Humana MAPD plan between July 1, 2010 and June 30, 2012, unless terminated by death during the final 12 months. The first 12 months of the study period (July 1, 2010 to June 30, 2011) was used as the baseline period and final 12 months (July 1, 2011 to June 30, 2012) were used as the follow-up period. Patients with T2DM were defined as patients meeting any of the following criteria: (1) At least one medical claim with a T2DM diagnosis with ICD-9-CM codes (250.x0 or 250.x2), no medical claims with a type I diabetes (T1DM) diagnosis (250.x1 or 250.x3), and at least one pharmacy claim for a non-insulin antidiabetic medication; or (2) Claims for both T1DM and T2DM with the ratio of T2DM:T1DM claims 0.5 and either of the following: at least one pharmacy claim for non-insulin antidiabetic medication, or a gap of at least 6 months at any time during the study period with no pharmacy claims for insulin. Additionally, all patients were required to have valid HbA1C values in both the baseline and follow-up periods that were at least 6 months apart, and a serum creatinine laboratory value during the baseline period. A large number of baseline descriptive variables were captured or calculated. The variables included, but were not limited to, demographic information [age, gender, geographic location, race/ethnicity (white, black, Hispanic, or other, as provided by the U.S. Social Security Administration), insurance plan type] and clinical descriptors, including point of service, anti-diabetic prescription treatment type, provider specialty (specialist vs primary), co-morbidities, Consumer Circle, Proportion of days covered (PDC), estimated glomerular filtration rate (eGFR),
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
107
Downloaded by [University of Florida] at 16:40 05 November 2015
Journal of Medical Economics
Volume 18, Number 2
February 2015
CHADs2 score, and Diabetes Complications Severity Index (DCSI), among others. Baseline demographic variables were calculated as of July 1, 2010 (the beginning of the baseline period). Baseline clinical characteristics and baseline cost variables were calculated using the 1-year baseline period beginning July 1, 2010. The CHADs2 score, a clinical prediction rule for estimating the risk of stroke in patients with non-rheumatic atrial fibrillation, was calculated by adding 1 point each for congestive heart failure [C], hypertension [H], age 475 years [A], and diabetes [D], and 2 points for prior stroke or TIA [s2]11. CHADs2 was calculated based on a count of complications captured in the medical claims via ICD-9 codes. Consumer Circles are a proprietary Humana population segmentation, which clusters the members into nine ‘circles’ based on factors such as socio-economic status, age, life stage, spending behavior, hobbies, and several hundred other characteristics. The DCSI was developed to predict adverse outcomes including hospitalization and mortality based on the number and severity of complications associated with diabetes, and is based on a summary score derived from diagnostic and laboratory data12. A modified version of the DCSI, which omits laboratory data in order to allow for the use of calculation DCSI using a claims database, has been validated to explain concurrent and future medical costs13. eGFR was calculated using the following formula: 186 ðCreatinine=88:4Þ1:154 ðAgeÞ0:203
burden on the healthcare system. Confidence intervals (95%) are provided around the estimate of effect for each quantile. All data extraction and statistical analyses were conducted using SAS v.9.3.1, SAS Enterprise Guide v.5.1, and SAS Enterprise Miner v.12.1 (SAS Institute, Cary, NC).
Results A total of 83,705 patients with T2DM met the inclusion criteria for this study. Demographic characteristics are presented in Table 1. The mean age was 70.57 years and males made up 48% of the population. Patients resided primarily in the southern and midwestern US (78.1% and 11.6%, respectively), and lived primarily in urban and suburban areas (76.4% and 17.4%, respectively). The largest race/ ethnicity category was white (74.5%), followed by black (18.1%). These geographic and race/ethnicity trends are consistent with the overall Humana MAPD population. The proportion of the T2DM study population that died during the follow-up year was 4.8%. Baseline clinical characteristics are presented in Table 2. The mean DCSI score for the population was 2.81 (SD ¼ 2.24). The most common anti-diabetic treatment categories were ‘Oral Antidiabetics Only’ (81.2%) Table 1. Demographic characteristics. Characteristic (n ¼ 83,705)
Units
T2DM
ð0:742 if femaleÞ ð1:210 if blackÞ where age was based on the beginning of the study period (July 1, 2010), and female gender and black race were determined as binary variables. A gradient boosting tree (GBT) method was used to narrow down the vast number of candidate variables. The gradient boosting tree method has been shown to have high prediction accuracy and to be useful in variable selection based on their importance14. Baseline cost variables were found to be highly correlated with follow-up costs and masked the relative importance of other variables. Therefore, baseline cost variables were removed from the GBT for the final selection. Only variables with relative importance of 0.10 or greater were included in the final linear least square model and quantile regression model. Quantile regression was used to test for associations between the dependent variable (healthcare costs) and the independent variables which were previously described. In particular, covariate effects related to 10%, 25%, 50%, 75%, 80%, 85%, 90%, 95%, and 98% quantiles of the costs were reported. These quantiles were chosen to focus on the patients at the higher end of the cost spectrum. These patients are of particular interest because of the anticipated complexity of their disease and financial 108
Age, years Gender Male Female Geographic region Northeast Midwest South West RUCA Urban Suburban Rural Unknown Race/ethnicity White Black Hispanic Other/unknown Plan characteristics No special characteristics LIS status only Dual eligibility only LIS status and Dual eligibility Plan type HMO PPO Other Died in Year 2
mean (SD)
70.57 (8.08)
n (%) n (%)
40,279 (48.10%) 43,426 (51.90%)
n (%) n (%) n (%) n (%)
907 (1.10%) 9,719 (11.60%) 65,403 (78.10%) 7,676 (9.20%)
n (%) n (%) n (%) n (%)
63,970 (76.40%) 14,557 (17.40%) 4,320 (5.20%) 858 (1.00%)
n (%) n (%) n (%) n (%)
62,319 (74.50%) 15,125 (18.10%) 2,917 (3.50%) 3,344 (4.00%)
n (%) n (%) n (%) n (%)
63,355 (75.70%) 7,968 (9.50%) 68 (0.10%) 12,314 (14.70%)
n (%) n (%) n (%) n (%)
52,835 (63.10%) 7,736 (9.20%) 23,134 (27.60%) 3,980 (4.80%)
SD, standard deviation; RUCA, Rural-Urban Commuting Area; LIS, Low income subsidy; HMO, Health Maintenance Organization; PPO, Preferred provider organization.
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
www.informahealthcare.com/jme ! 2015 Informa UK Ltd
Journal of Medical Economics
Downloaded by [University of Florida] at 16:40 05 November 2015
DCSI score, Mean (SD) Number of ED visits, Mean (SD) Number of inpatient hospitalizations, Mean (SD) Number of outpatient visits, Mean (SD) Anti-diabetic treatment categories, n (%) No anti-diabetic therapy Oral anti-diabetic only Non-insulin injectable anti-diabetic only Insulin only Oral plus non-insulin injectable anti-diabetics Oral anti-diabetic plus insulin Non-insulin injectable anti-diabetic plus insulin Oral anti-diabetic, non-insulin anti-diabetic, and insulin Adherence to anti-diabetic therapy (PDC), Mean (SD) Persistence on anti-diabetic therapy, Mean (SD) Pill burden (non-insulin anti-diabetic), Mean (SD) Mail-order pharmacy use, n (%) Count of the number of unique anti-diabetic classes, Mean (SD) Duplicate therapy, n (%) Number of claims for glucose test strips, Mean (SD) Number of emergent hypoglycemic events, Mean (SD) Number of non-emergent hypoglycemic events, Mean (SD) Comorbidities, n (%) Depression Low testosterone Erectile dysfunction Retinopathy Nephropathy Chronic kidney disease Stage I Stage II Stage III Stage IV Stage V End stage renal disease Neuropathy Cardiovascular disease
T2DM 2.81 (2.24) 0.58 (1.4) 1.85 (5.89) 25.91 (23.33) 0 (0.00%) 67,946 (81.20%) 15 (0.00%) 677 (0.80%) 506 (0.60%) 14,284 (17.10%) 28 (0.00%) 249 (0.30%) 0.86 (0.16) 250.93 (137.57) 1.23 (0.52) 45,699 (54.60%) 1.7 (0.8) 2,997 (3.60%) 0.72 (1.38)
Characteristic (n ¼ 83,705) 7.0% Primary healthcare professional, n (%) Endocrinologist Nephrologist Primary care/Other Renal function lab results, n (%) Microalbumin (mg/day) Normal (530) Elevated (30) Missing Serum Creatinine Stage 1 (eGFR 90þ) Stage 2 (eGFR 60–89) Stage 3 (eGFR 30–59) Stage 4 (eGFR 15–29) Stage 5 (eGFR 515) Passed diabetes-related HEDIS measures, n (%) Cholesterol Screening (Yes) Cholesterol Controlled (Yes) Eye Exam (Yes) Kidney Disease Monitoring (Yes)
T2DM 50,829 (60.70%) 2,167 (2.60%) 3,202 (3.80%) 78,336 (93.60%) 20,552 (24.60%) 8,101 (9.70%) 55,052 (65.80%) 16,103 (19.20%) 42,659 (51.00%) 23,315 (27.90%) 1,392 (1.70%) 236 (0.30%) 47,751 (57.00%) 35,085 (41.90%) 35,635 (42.60%) 46,452 (55.50%)
DCSI, diabetes complication severity index; ED, emergency department; PDC, proportion of days covered; UAP, unstable angina pectoris; NSTEMI, non-ST segment elevation myocardial infarction; STEMI, ST segment elevation myocardial infarction; CHADS2, congestive heart failure, hypertension, age ¼ 75 years, diabetes mellitus, stroke; eGFR, estimated glomerular filtration rate; HEDIS, healthcare effectiveness data and information set.
0.09 (1.69) 0.32 (3.17) 12,775 (15.30%) 1,852 (2.20%) 6,027 (7.20%) 15,100 (18.00%) 36,408 (43.50%) 30,420 (36.30%) 3,181 (3.80%) 9,944 (11.90%) 19,732 (23.60%) 2,217 (2.60%) 801 (1.00%) 801 (1.00%) 34,018 (40.60%) 40,954 (48.90%)
Major adverse cardiac events (MACE) plus including, n (%) UAP 2,520 (3.00%) NSTEMI or STEMI 1,628 (1.90%) Stroke 9,831 (11.70%) Heart failure 12,277 (14.70%) Peripheral artery disease 22,135 (26.40%) Atherosclerosis 13,756 (16.40%) Aortic aneurysm 1,982 (2.40%) Osteoporosis 8,757 (10.50%) CHADs2 stroke risk score, Mean (SD) 2.56 (0.94) HbA1c values Number of HbA1c values during the 2.45 (1.19) baseline period, Mean (SD) Median (range) 2 (1–19) Categorical HbA1c values, n (%) 49.0% 6,205 (7.40%) 48.0% and 9.0% 7,169 (8.60%) 47.0% and 8.0% 19,502 (23.30%)
(continued )
! 2015 Informa UK Ltd www.informahealthcare.com/jme
February 2015
Table 2.Continued.
Table 2. Baseline clinical characteristics. Characteristic (n ¼ 83,705)
Volume 18, Number 2
and ‘Oral Antidiabetic plus Insulin’ (17.1%). The most common oral anti-diabetics used were metformin and sulfonylureas (used by 70.7% and 52.0% of patients, respectively) followed by insulin (18.2%). The mean number of HbA1c values available during the baseline period was 2.45, (median ¼ 2, range ¼ 1–19). Nearly two-thirds (60.7%) of the T2DM population had a baseline HbA1c value 7.0%. Common co-morbid conditions included nephropathy (43.5%), neuropathy (40.6%), and cardiovascular disease (48.9%). The proportion of patients with stage 3 Chronic Kidney Disease (CKD) or worse was very similar using either of the two employed sources. However, the proportion with Stage 2 CKD was much larger based on eGFR values than identified from diagnosis codes (51.0% vs 11.9%, respectively). Table 3 shows the results of the gradient boosting tree that was used to select variables for inclusion in the QR analysis. After excluding variables with a relative importance 50.10 there were 17 variables remaining. Two baseline cost variables (total pharmacy costs and diabetesrelated medical costs) were excluded due to co-linearity with other cost variables.
Regression models Figure 1 shows the results of the OLS and QR regression analyses. The quantile regression analysis showed that the association between many of the study co-variates and
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
109
Journal of Medical Economics
Volume 18, Number 2
February 2015
Table 3. Gradient boosting tree results.
Downloaded by [University of Florida] at 16:40 05 November 2015
Variable
Outpatient visits ED visits Inpatient visits Diabetes Complications Severity Index Cardiovascular disease Atherosclerosis Chronic kidney disease (yes/no) Nephropathy Number of HbA1c values during the baseline period Age at Baseline Pill Burden Chronic kidney disease stage (by eGFR) Count of the number of unique anti-diabetic classes Chronic kidney disease stage (by ICD-9 code) HbA1c value during the baseline period HEDIS measures - eye exam CHADs2 Stroke Risk score
Discussion Relative importance 1.00000 0.79173 0.78540 0.40434 0.32044 0.23325 0.20989 0.18613 0.16103 0.15940 0.15790 0.14769 0.14491 0.14083 0.13922 0.13258 0.10811
ED, emergency department; eGFR, estimated glomerular filtration rate; HEDIS, healthcare effectiveness data and information set; CHADs2, congestive heart failure, hypertension, age ¼ 75 years, diabetes mellitus, stroke.
follow-up costs were found to change across the cost quantiles. The variables with the strongest relationships with costs during the follow-up period included the four baseline cost variables, baseline measures of healthcare utilization (outpatient visits, ER visits, inpatient visits), and DCSI score during the baseline period. These variables had a statistically significant association with costs at nearly every cost quantile, and had the largest impact at the highest cost quantiles. Among them, baseline DCSI score was associated with the greatest increase in the top quantiles ($1490 and $1788 at the 95th and 98th percentiles, respectively). ER visits were associated with cost increases of 4$500 at the 50th percentile and above. Several other variables also exhibited interesting QR curves. The number of HbA1c tests performed during the baseline period was associated with a statistically significant but relatively small decrease in the total healthcare costs for patients in the lower quantiles (5$260 for patients in the 75th percentile and lower). However, in the top cost quantiles, each additional HbA1c test was associated with noteworthy cost savings (as much as $1400 in the 98th percentile). A similar finding is observed for baseline pill burden. The baseline variables associated with the largest cost increases during the follow-up period was late-stage CKD (as identified by eGFR). Stage 4 and Stage 5 CKD were associated with an incremental cost increase of $33,131 and $106,975 relative to Stage 1 or no CKD in the 98th percentile. Late-stage CKD (as identified by ICD-9-CM code) was found to have obscure patterns and did not have a consistently higher impact on costs relative to having Stage 1 or no CKD. 110
To our knowledge this is the first study that has used quantile regression in an examination of economic/cost drivers among patients with diabetes. While several previous studies have quantified the impact of patient demographic and clinical characteristics on the healthcare cost of patients with diabetes, all of these analyses utilized the OLS regression model approach6,15,16. The method of least squares results in estimates that approximate the conditional mean of the response variable given certain values of the predictor variables, whereas quantile regression aims at estimating different quantiles of the response variable and is particularly useful when modeling cost data which are not normally distributed. High cost patients (i.e., in a high quantile) are intuitively of particular interest to researchers, healthcare providers, and payers both from the perspective of improving health and health outcomes among high risk patients, as well as reducing the costs associated with their care10. Our data indicate that there are a few key co-morbid conditions that have only a marginal impact on future healthcare costs among patients in the lower quantiles of costs, but tremendous impact on costs in higher quantiles. For example the DCSI score or diagnosis of cardiovascular disease were found to have very little impact on cost in the lower cost quantiles but a sharply increased impact among patients in the higher quantiles. As such our findings appear to validate the existing literature that higher DCSI scores are a valid predictor of increased future healthcare costs13,17. Other co-variates, such as ER visits, showed a more steady increase in their impact on cost as the quantile increased. The largest drivers of costs in this study population were stages 4 and 5 CKD. The OLS model shows that the average increase in costs among patients with latestage CKD are substantial, but fails to highlight the enormous increase seen for patients in the highest cost quantile that is unveiled by the quantile regression analysis. Interestingly, the QR cost curves for late-stage CKD identified using claims data do not show patterns that are consistent with late-stage CKD identified using eGFR values. Instead, the QR curves for late-stage CKD identified using claims data showed few distinguishable trends. This finding suggests that documentation of stages of CKD in claims data may be highly inaccurate and caution should be used by those wanting to determine patients to target with interventions focused on patients with chronic kidney disease. While such information is clearly important in tailoring interventions, the practicalities of this approach warrant further investigation. All models suffer from limitations inherent to their very nature, as they rely on the quality of the data to which they are applied. Limitations common to research using administrative claims data apply in
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
www.informahealthcare.com/jme ! 2015 Informa UK Ltd
Downloaded by [University of Florida] at 16:40 05 November 2015
Journal of Medical Economics
Volume 18, Number 2
February 2015
Figure 1. Quantile regression and ordinary least squares estimates for total healthcare costs. $ ¼ U.S. dollars; ED ¼ emergency department; CKD ¼ chronic kidney disease; eGFR ¼ estimated glomerular filtration rate; CHAD2S ¼ congestive heart failure, hypertension, age, diabetes mellitus, prior stroke or transient ischemic attack or thromboembolism.
this study. These include lack of certain information in the database (e.g., health behavior information) and error in claims coding. Causal inferences cannot be ascertained from the current study, as it is a descriptive observational study using retrospective claims data. The use of retrospective secondary coded data does not provide as accurate an assessment as an empirical evaluation by a clinician and is neither randomized nor controlled. It will, therefore, be important to replicate the results of this study in a prospective manner, perhaps including outcome measures other than cost that may be helpful to patients, healthcare providers, and payers. The current study uses data only from Humana MAPD members; therefore, the results may not be generalizable to the general Medicare population. However, Humana is a ! 2015 Informa UK Ltd www.informahealthcare.com/jme
large national health plan with membership in every region of the US. This sample likely represents a slightly healthier population than the general Medicare population of patients with T2DM, as the requirement for having HbA1c values during both the baseline and follow-up periods may bias inclusion towards patients that are being more closely monitored by a provider. This may be the reason that the average HbA1c value was low relative to other large population studies18.
Conclusion Diabetes is a highly prevalent disease that places substantial financial burden on patients, their families, and the healthcare system. In order to help patients manage this
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
111
Downloaded by [University of Florida] at 16:40 05 November 2015
Journal of Medical Economics
Volume 18, Number 2
February 2015
chronic disease, communities, providers, and payers offer a wide range of interventions and disease management programs to patients. However, resources available to support such programs are limited. Quantile regression is a useful econometric tool that may aid in shaping individual patient management based on characteristics most likely to have a significant impact, as opposed to those based on population mean or median estimates of effect. This study demonstrates that the impact of clinical and demographic characteristics on the economic burden of the disease vary across the continuum of healthcare costs. A better understanding of the factors that drive costs on an individual patient level will provide important insights that will help in ameliorating the clinical, humanistic, and economic burden of diabetes.
2. 3. 4. 5.
6.
7.
8. 9.
Transparency
10.
Declaration of funding This research was funded by Eli Lilly and Company. Declaration of financial/other relationships This study was funded by Eli Lilly and company and was conducted under the oversight of an ongoing research collaboration agreement between Eli Lilly and Company and Humana, Inc. SLS is an employee of Humana, Inc. GC is an employee of Comprehensive Health Insights, the business entity within Humana that provides research for external clients. BHC, HF, and DPS, are employees and shareholders of Eli Lilly and Company, a pharmaceutical company that is involved in the manufacture and sales of a variety of products, including those used in the management of Type 2 diabetes. JME peer reviewers on this manuscript have no relevant financial or other relationships to disclose.
11.
12.
13.
14. 15. 16. 17.
References 1. US Centers for Disease Control and Prevention. Diabetes Report Card 2012. Atlanta, GA: Centers for Disease Control and Prevention, US Department of
112
18.
Health and Human Services, 2012. http://www.cdc.gov/diabetes/pubs/pdf/ DiabetesReportCard.pdf. Accessed June 25, 2014 American Diabetes Association. Economic costs of diabetes in the US in 2002. Diabetes Care 2003;26:917-32 American Diabetes Association. Economic costs of diabetes in the US in 2007. Diabetes Care 2008;31:596-615 American Diabetes Association. Economic costs of diabetes in the US in 2012. Diabetes Care 2013;36:1033-46 Maguire J, Dhar V. Comparative effectiveness for oral anti-diabetic treatments among newly diagnosed type 2 diabetics: data-driven predictive analytics in healthcare. Heal Syst 2012;2:73-92 Li R, Bilik D, Brown MB, et al. Medical costs associated with type 2 diabetes complications and comorbidities. Am J Manag Care 2013;19: 421-430 Boyle JP, Thompson TJ, Gregg EW, et al. Projection of the year 2050 burden of diabetes in the US adult population: dynamic modeling of incidence, mortality, and pre-diabetes prevalence. Popul Health Metr 2010;8:29 Montgomery DC, Peck EA, Vining GG. Introduction to Linear Regression Analysis. 5th edn. Hoboken, NJ: Wiley, 2012 Koenker R. Quantile regression. New York, NY: Cambridge University Press, 2005 Chen SL. An introduction to Quantile Regression and the QUANTREG Procedure. Paper 213-30, Statistics and Data Analysis. Cary, NC: SAS Institute, http://www2.sas.com/proceedings/sugi30/213-30.pdf. Accessed June 25, 2014 Gage BF, Waterman AD, Shannon W, et al. Validation of clinical classification schemes for predicting stroke: results from the National Registry of Atrial Fibrillation. JAMA 2001;285:2864-70 Young BA, Lin E, Von Korff M, et al. Diabetes complications severity index and risk of mortality, hospitalization, and healthcare utilization. Am J Manag Care 2008;14:15-23 Chang H-Y, Weiner JP, Richards TM, et al. Validating the adapted Diabetes Complications Severity Index in claims data. Am J Manag Care 2012;18:721-6 Hastie T, Tibshirani R, Friedman J. The elements of statistical learning. New York: Springer, 2009 Bhattacharyya SK, Else BA. Medical costs of managed care in patients with type 2 diabetes mellitus. Clin Therapeut 1999;21:2131-42 Brandle M, Zhou H, Smith BR, et al. The direct medical cost of type 2 diabetes. Diabetes Care 2003;26:2300-4 Chen HL, Hsiao FY. Risk of hospitalization and healthcare cost associated with diabetic complication severity index in Taiwan’s National Health Insurance Research Database. J Diabetes Complications 2014;28:612-6 Dodd AH, Colby MS, Boye KS, et al. Treatment approach and HbA1c control among US adults with type 2 diabetes: NHANES 1999-2004. Curr Med Res Opin 2009;27:1605-13
Factors associated with increased healthcare costs in patients with type 2 diabetes in the US Slabaugh et al.
www.informahealthcare.com/jme ! 2015 Informa UK Ltd
