Patrick J. O'Connor, MD, MPH

Despite the significant prevalence of elevated blood pressure (BP) and body mass index (BMI) in children, few studies have assessed their combined impact on healthcare costs. This study estimates healthcare costs related to BP and BMI in children and adolescents.

Prospective dynamic cohort study of 71,617 children aged 3 to 17 years with 208,800 child years of enrollment in integrated health systems in Colorado or Minnesota between January 1, 2007, and December 31, 2011.

Generalized linear models were used to calculate standardized annual estimates of total, inpatient, outpatient, and pharmacy costs, outpatient utilization, and receipt of diagnostic and evaluation tests associated with BP status and BMI status.

Total annual costs were significantly lower in children with normal BP ($736, SE = $15) and prehypertension ($945, SE = $10) than children with hypertension ($1972, SE = $74) (

 <.001, each comparison), adjusting for BMI. Total annual cost for children below the 85th percentile of BMI ($822, SE = $8) was significantly lower than for children between the 85th and 95th percentiles ($954, SE = $45) and for children at or above the 95th percentile ($937, SE = $13) (

This study shows strong associations of prehypertension and hypertension, independent of BMI, with healthcare costs in children. Although BMI status was also statistically significantly associated with costs, the major influence on cost in this large cohort of children and adolescents was BP status. Costs related to elevated BMI may be systematically overestimated in studies that do not adjust for BP status.

Prehypertension and hypertension have strong effects on healthcare costs in children, independent of body mass index (BMI).

This study examines the combined impact of blood pressure (BP) and BMI on healthcare costs.

Although BMI status was associated with costs, the major influence on cost in this large cohort of children and adolescents was BP status.

Costs related to elevated BMI may be systematically overestimated in studies that do not adjust for BP status.

The potential long-term adverse outcomes associated with prehypertension (preHT) and hypertension (HT) in children and adolescents are well recognized.

 Blood pressure (BP) in youth correlates with BP in early adulthood and may be associated with elevated cardiovascular (CV) risk later in life.

 Moreover, elevated BP is associated with elevated body mass index (BMI) in children and adolescents,

Because of the acknowledged increased CV risk associated with childhood preHT or HT, it seems reasonable to suggest that healthcare costs would also increase with preHT or HT as the result of increased testing and evaluation, increased follow-up, or increased CV-related comorbidities. However, information on the cost related to elevated BP in children is scarce.

 Studies of obese children suggest modestly elevated costs of care, but none of these prior studies consider the fact that elevated BMI is significantly related to elevated BP.

Studies assessing the relation between BMI and healthcare costs, but failing to account for BP levels or other clinical conditions associated with elevated BMI, may overestimate the independent impact of BMI.

 In this project, we model the association of elevated BP and costs, while taking into account detailed data on BMI and other comorbid conditions (including depression and other mental health conditions). Using this broader approach permits a more accurate assessment of healthcare costs in children and adolescents with elevated BP than data currently available in the literature.

  Electronic health records (EHRs) were used to identify a dynamic, longitudinal cohort of children and adolescents who were enrolled in 2 large integrated health maintenance organizations in Colorado and Minnesota. Children and adolescents aged 3 to 17 years were included in the study if they were continuously enrolled in the health plans for 1 or more years from January 1, 2007, to December 31, 2010; had measured height, weight, and BP; and had comprehensive insurance coverage including coverage for pharmaceuticals. Only BP measurements obtained in primary care outpatient clinics were included in this study; BP measurements obtained in the emergency department (ED) or urgent care centers were excluded.

Both health plans used an EpiCare (Verona, Wisconsin) EHR. Demographics, including age, gender, and race/ethnicity, were derived from administrative eligibility data. BMI and presence of preHT and HT were ascertained from EHR vital signs data. Age and gender-specific BMI percentiles were calculated for each visit in which both height and weight were measured, and children and adolescents were assigned their maximum BMI percentile during the study period. Age, gender, and height-specific BP percentiles were calculated for both systolic BP and diastolic BP measures obtained in outpatient settings (including emergency departments and urgent care clinics)

Children who had 3 or more BP measurements taken during the study period were included. Normal BP was defined as having all systolic BP and diastolic BP measurements < 90th percentile; preHT was defined as having 1 or more BP measurements ≥90th percentile or ≥120/80 mm Hg; HT was defined as either 3 consecutive BP measurements ≥95th percentile or 1 or more BP measurements ≥99th percentile. Children were also identified as having HT if they received a diagnosis of essential hypertension from their healthcare provider during the study period (

International Classification of Diseases, Ninth Revision, Clinical 

 [ICD-9-CM] codes 401.xx). To focus this study on essential hypertension, children diagnosed with secondary hypertension (ICD-9-CM codes 405.xx) or other cardiovascular conditions were excluded. Normal BMI was defined as <85th percentile, overweight as ≥85th to <95th percentile, and obesity as ≥95th percentile.

EHRs accurately measure utilization, but do not directly measure resource use or costs. Therefore, we applied standard pricing methodologies to estimate annual inpatient, outpatient, and pharmacy costs.

 Inpatient admissions were priced using nationally representative 2010 payment rates based on Diagnostic Related Groups. Costs for outpatient services including physician, outpatient hospital, and clinic services; costs for all other outpatient services such as nursing services, laboratory, and radiology were priced using the nationally representative 2010 payment rates per relative value unit. Costs for phar- maceuticals, including prescriptions for antihypertensive and other medications, were priced at 68% of their average wholesale prices in 2010.

We examined utilization of outpatient services to determine whether preHT and HT might be driving utilization and costs through increased diagnostic testing and evaluation. We used procedure codes to estimate the annual number of preventive, office, or outpatient medical visits and the number of outpatient psychiatry and psychology visits. We also identified the likelihood of ED services or specialist consultations. In addition, we identified the likelihood of receiving diagnostic procedures including urinalysis, laboratory testing, chest x-ray, electrocardiogram, renal ultrasound, and echocardiogram.

We measured medical comorbidities using the Chronic Illness and Disability Payment System (CDPS).

 CDPS is a combined diagnostic and pharmacy-based risk adjustment model commonly used to measure illness burden and to adjust capitated payments to health plans. The CDPS algorithm assigns ICD-9-CM diagnosis codes to 58 diagnostic categories including selected mental health conditions, and national drug codes to 15 pharmaceutical categories, and applies severity weights developed from a national database of medical claims. The resulting CDPS score provides a measure of illness burden based on expected future healthcare expenditure conditional on each person’s diagnostic and pharmaceutical profile.

We analyzed inpatient, outpatient, pharmacy, and total costs using generalized linear models. Generalized linear models are commonly used to estimate healthcare costs when the dependent variable is non-negative and when its distribution is noticeably skewed and kurtotic (with a heavy right-hand tail).

 Inpatient and pharmacy costs were estimated using 2-part models: logistic regression was used to estimate the probability of any use of a service, and a generalized linear model based on a gamma family with a log link function was used to estimate costs conditional on receiving at least 1 service. Total and outpatient costs was estimated using a single generalized linear model. The number of preventive, office, or outpatient visits and the number of outpatient psychiatry and psychology visits were estimated using negative binomial regression, and the probability of receipt of each of the remaining measures of outpatient utilization were estimating using logistic regression. We selected these specific distributions based on standard tests for assessing alternative generalized linear and transformed models.

 We assessed goodness of fit using a modified Hosmer-Lemeshow test and a Pregibon’s link test.

The primary predictor variables of interest were indicator variables for BMI ≥85th to <95th percentile, BMI ≥ 95th percentile, preHT, and HT. Analyses of BMI were adjusted for HT, and analyses of HT were adjusted for BMI. Additional control covariates included age (specified as a set of indicator variables with 1 for each year), gender, race/ethnicity, site, and year. We conducted analyses with and without controlling for comorbidity using the CDPS score. Controlling for comorbidity may reduce confounding of the cost estimates if increased illness burden is correlated with BP or BMI status. However, it may be inappropriate to adjust for comorbidity if an increased illness burden is the result of BP or BMI status. Therefore, we conducted 2 sets of analysis to provide a range of estimates and to demonstrate the sensitivity of the estimates to medical comorbidities.

We calculated cost estimates associated with BMI, preHT, and HT that were adjusted for age, sex, race/ ethnicity, CDPS score, and either BMI or BP status. For example, the standardized costs of children with pediatric HT were calculated as the mean predicted values across all children assuming they all had pediatric HT. Standard errors were calculated using the nonparametric bootstrap, and P values were computed using the percentile method from the empirical distributions of the results from 1000 replicates.

All analyses were conducted in STATA, version 12.

This project was reviewed in advance, approved, and monitored by the HealthPartners Institute for Education and Research Institutional Review Board.

The study sample included 71,617 children aged 3 to 17 years followed for 35.0 (SD = 13.7) months and providing 208,800 child years of data. Study sample characteristics are shown in 

. The mean age was 10.3 (SD = 4.4) years, and 51% were female. The majority (57%) were non-Latino white, 14% were Latino, 10% were African American, and 4% were Asian. Native American, Hawaiian / Pacific Islander, or multiple races were 1% or less, and 14% were of unknown race/ethnicity. BMI percentile was normal in 64%, 18% were overweight, and 18% were obese. BP percentile was normal in 55%, 42% were prehypertensive, and 3% had hypertension. Among children identified with hypertension, 83% were identified using only BP measures. Among the 17% with an ICD-9-CM code for essential hypertension, 35% had BP measurements consistent with the diagnosis, while 65% did not have 3 consecutive BP measurements ≥95th percentile, and 31% filled prescriptions for antihypertensive medications. Among the 83% identified by BP only, 2% filled prescriptions for antihypertensive medications.

 shows the mean standardized annual costs among children by BMI and BP status without controlling for additional morbidity. Estimates of costs related to BMI were adjusted for HT, and estimates of costs related to HT were adjusted for BMI. There were no statistically significant interactions between BP and BMI status. As noted in 

, costs for children with preHT ($945, SE = $10) were significantly higher than for children with normal BP ($736, SE = $15), and annual costs for children with HT ($1972, SE = $74) were significantly greater than for children with preHT or normal BP (P <.001 each). Among children with hypertension, there was no difference in total costs between children who received a diagnosis of hypertension compared to children who were identified only by BP measures in the EMR review (P = .814). The largest cost component for all 3 BP groups was outpatient care, followed by pharmacy and inpatient care.

There were smaller cost differentials related to BMI. Total annual cost for children below the 85th percentile of BMI ($822, SE = $8) was significantly lower than for children between the 85th and 95th percentiles ($954, SE = $454) and for children at or above the 95th percentile ($937, SE = $13;) (P <.001 each). Total annual cost for children between the 85th and 95th percentiles of BMI was not statistically significantly different from the total cost for children above the 95th percentile (P = .754).

 shows that the standardized cost differentials between categories narrowed when controlling for additional comorbidity using the CDPS score. The differential in total costs between HT and normotensive children declined by $383 (SE = $90) to $852 (SE = $53), and the cost differential between preHT and normotensive children declined by $67 (SE = $18) to $141 (SE = $9) (P <.001 each) when adjusting for additional comorbidity using CDPS. The differentials in total cost for children above the 95th percentile of BMI and children between the 85th and 95th percentiles compared to children below the 85th percentile declined by $67 (SE = $19, P <.001) and $64 (SE = $48, P = .192), respectively, to $48 (SE = $11, P <.001) and $68 (SE = $15, P <.001) when adjusting for additional comorbidity using CDPS.

 shows estimates of outpatient utilization associated with preHT and HT. There were statistically significant increases in all measures of utilization except for laboratory testing. Utilization was at least 50% higher in HT than normal BP children in all categories, except for laboratory testing (P <.001 each), with the large percentage increases observed for renal ultrasound, echocardiogram, and chest x-ray. However, these procedures were performed at a relatively low rate, and all diagnostic and evaluation procedures with the exception of urinalysis were conducted in fewer than 10% of children with HT. We reviewed the types of specialist consultations, but could not identify any specific pattern to differentiate visits among the BP categories.

The present study shows the strong incremental effects of both preHT and HT, independent of BMI and comorbidity, on healthcare costs in children. Although BMI status also was significantly associated with cost, the major influence on cost in this large cohort of children was BP status. It should be noted that the percent of children with preHT was higher than previously reported.

Because the definition for preHT requires only a single BP measurement, the requirement of 3 separate clinic BP measurements for inclusion in this study increased the number with preHT.

An important factor in this analysis is that hypertension was identified primarily from review of EHRs, with only 17% recognized by the clinic physicians. The lack of recognition of HT by providers using computerized medical record systems has been reported previously,

as has the extent of the underrecognition.

However, there was no significant difference in cost associated with the EHR method, as opposed to physician recognition and diagnosis. Thus, there may be something inherent in the presence of hypertension or prehypertension that leads to increased outpatient and emergency department visits, specialist consultations, increased diagnostic and evaluation procedures, and their associated costs; or medical issues resulting in more frequent clinic visits with greater number of BP measurements resulting in a larger number of diagnoses of preHT or HT. This may account for the significantly greater use of cardiac studies and chest x-rays in the subjects with hypertension.

Elevated BP in adults has been shown to be associated with substantial increased costs in 3 national data sets.

 Studies of adults have further shown that both obesity and hypertension are independent determinants of costs.

In contrast, little attention has been devoted to the costs or to strategies to address elevated BP in children or adolescents. Given the strong relationship between elevated BP and healthcare costs, a case could be made to health plans and payers to consider implementing strategies to manage healthcare utilization of children with elevated BP. These strategies might include improved primary care, case management, health behavior interventions, family education, or other interventions.

In contrast to prior studies, costs of healthcare services in this study were not significantly different between overweight (BMI ≥85th percentile to < 95th percentile) and obese (BMI ≥95th percentile) children controlling for BP status. Studies assessing healthcare utilization and cost related to overweight or obesity previously have not been able to adjust for BP status due to lack of BP data. By including both BMI and BP status in the analyses, this study has been able to extend previous findings by showing that the effect of BMI is lower than previously reported and is confounded by the associated impact of preHT and HT.

These findings are relevant to ongoing clinical and public policy discussions. First, they suggest that greater attention should be paid to elevated BP in children and adolescents as a driver of healthcare costs. Second, elevated BP has a significant effect on cost independent from elevated BMI and other comorbid conditions. Third, studies that do not account for the impact of BP level on utilization may overestimate the effect of BMI on utilization and costs.

There are limitations that should be considered in the interpretation of these data. First, despite the large, racially and ethnically diverse study population, these results may not be generalizable to costs of care in non—insurance-driven healthcare systems. However, it is currently challenging to conduct this type of study and capture detailed clinical and cost data outside the highly integrated healthcare systems included in this study. Second, it is possible that some allopathic care was obtained outside the insurance system, although prior studies suggest that this is well under 5% of total allopathic care. Third, although we adjusted for comorbidity using CDPS, there may be some residual differences in comorbidities by BP status. Finally, the results from this study do not imply causality between hypertension and the increased cost of patient care. However, the data clearly show an association between BP and healthcare costs in youth.

Department of Family and Preventive Medicine, University of California, San Diego (TPG); HealthPartners Institute for Education and Research, Minneapolis, MN (PJO, EOK, NES, KFA, KLM); Department of Pediatrics, University of Minnesota, Minneapolis (ARS); Institute for Health Research, Kaiser Permanente Colorado, Denver (DJM).

Funding was provided by the National Institutes of Health grant R01HL093345 from the National Heart, Lung, and Blood Institute.

 The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Concept and design (TPG, PJO, ARS, EOK, KLM); acquisition of data (PJO, KLM); analysis and interpretation of data (TPG, PJO, ARS, EOK, DJM, NES, KFA, EDP, KLM); drafting of the manuscript (TPG, PJO, ARS); critical revision of the manuscript for important intellectual content (TPG, PJO, ARS, EOK, DJM, NES, KFA, EDP, KLM); statistical analysis (TPG); provision of study materials or patients (PJO); obtaining funding (PJO, ARS, NES); administrative, technical, or logistic support (TPG); supervision (TPG, POC).

Todd Gilmer, PhD, Department of Family and Preventive Medicine, University of California, San Diego, 9500 Gilman Dr, La Jolla, CA 92092-0622. E-mail: tgilmer@ucsd.edu.

1. Daniels SR. Cardiovascular sequelae of childhood hypertension. 

2. Mitchell P, Cheung N, de Haseth K, et al. Blood pressure and retinal arteriolar narrowing in children. 

3. Raitakari OT, Juonala M, Kähönen M, et al. Cardiovascular risk factors in childhood and carotid artery intima-media thickness in adulthood: the Cardiovascular Risk in Young Finns Study. 

4. Donahue RP, Prineas RJ, Gomez O, Hong CP. Tracking of elevated systolic blood pressure among lean and overweight adolescents: the Minneapolis Children’s Blood Pressure Study. 

5. Klumbiene J, Sileikiene L, Milasauskiene Z, Zaborskis A, Shatchkute A. The relationship of childhood to adult blood pressure: longitudinal study of juvenile hypertension in Lithuania. 

6. Elkasabany AM, Urbina EM, Daniels SR, Berenson GS. Prediction of adult hypertension by K4 and K5 diastolic blood pressure in children: the Bogalusa Heart Study. 

7. Lauer RM, Burns TL, Clarke WR, Mahoney LT. Childhood predictors of future blood pressure. 

8. Yong LC, Kuller LH. Tracking of blood pressure from adolescence to middle age: the Dormont High School Study. 

9. Sinaiko AR, Donahue RP, Jacobs DR Jr, Prineas RJ. Relation of weight and rate of increase in weight during childhood and adolescence to body size, blood pressure, fasting insulin, and lipids in young adults: the Minneapolis Children’s Blood Pressure Study. 

10. Chiolero A, Bovet P, Paradis G, Paccaud F. Has blood pressure increased in children in response to the obesity epidemic? 

11. Muntner P, He J, Cutler JA, Wildman RP, Whelton PK. Trends in blood pressure among children and adolescents. 

12. National High Blood Pressure Education Program Working Group on High Blood Pressure in Children and Adolescents. The fourth report on the diagnosis, evaluation, and treatment of high blood pressure in children and adolescents. 

13. McNiece KL, Poffenbarger TS, Turner JL, Franco KD, Sorof JM, Portman RJ. Prevalence of hypertension and pre-hypertension among adolescents. 

14. Fuentes RM, Notkola IL, Shemeikka S, Tuomilehto J, Nissinen A. Tracking of systolic blood pressure during childhood: a 15-year followup population-based family study in eastern Finland. 

15. Vos LE, Oren A, Uiterwaal C, Gorissen WH, Grobbee DE, Bots ML. Adolescent blood pressure and blood pressure tracking into young adulthood are related to subclinical atherosclerosis: the Atherosclerosis Risk in Young Adults (ARYA) study. 

16. Wang YC, Cheung AM, Bibbins-Domingo K, et al. Effectiveness and cost-effectiveness of blood pressure screening in adolescents in the United States. 

17. Tran CL, Ehrmann BJ, Messer KL, et al. Recent trends in healthcare utilization among children and adolescents with hypertension in the United States. 

18. Samuels J. The increasing burden of pediatric hypertension. 

19. Elliott DJ, Weintraub WS. Population-based health requires population-based change. 

20. Hampl SE, Carroll CA, Simon SD, Sharma V. Resource utilization and expenditures for overweight and obese children. 

21. Finkelstein EA, Fiebelkorn IC, Wang G. State-level estimates of annual medical expenditures attributable to obesity. 

Articles

This study demonstrates a major influence of prehypertension and hypertension on healthcare costs in a large cohort of children, independent of body mass index.

Impact of Hypertension on Healthcare Costs Among Children

To study the relative contributions of insurance product (commercial, Medicaid, Medicare), health plan, and physician group to quality of diabetes care. 

Cross-sectional observational study using data reported by Minnesota Community Measurement (MNCM) on care provided in 2005.

Individual performance rates for glycosylated hemoglobin (A1C) level <7%, lowdensity lipoprotein cholesterol (LDL-C) level <100 mg/dL, blood pressure <130/80 mm Hg, documented tobacco-free status, and aspirin use, as well as an all-or-none composite measure (Optimal Diabetes Care), were obtained from MNCM for 57 physician groups. Results included 7169 patients connected with 1 of 8 health plans and 1 of 3 types of insurance product.

All factors studied had a relationship to quality results. Of the factors, insurance product has the strongest relationship to A1C, LDL-C, and Optimal Diabetes Care. Health plan had the strongest relationship to tobacco-free status and daily aspirin, and physician group had the strongest relationship to blood pressure control. Physician group results varied more than those for health plan or insurance product on most measures.

All factors studied contribute to diabetes quality results. Reporting both physician group performance and health plan performance may offer a greater opportunity to improve care than reporting only health plan or only physician group results.

Physician group, health plan, and insurance product all affect diabetes quality of care. The relative contributions of these elements were examined in this cross-sectional observational study.

Reporting physician group performance in addition to health plan performance may stimulate greater improvement in patient care. 

Collaboration enhances a region’s ability to report physician group results. 

Incentives to focus on improving the quality of care community-wide might encourage collaborative efforts to report physician group performance.

A wide variety of factors are believed to contribute to the level of measured care quality for patients in the US healthcare system.

 Among these factors are the characteristics of the insurance product, health plan, physician group, individual physician, and individual patients.

Patient ethnicity, sex, and socioeconomic status have been shown to impact quality, as have provider group organizational structure, provider group culture, and the use of patient registries, electronic health records, and other practice supports.

 Health plan results vary by geography, for-profit versus not-for-profit status, patient ethnicity, enrollee socioeconomic status, and other factors.12-16 The Medicaid insurance product is associated with generally lower-quality results, whereas Medicare analysis demonstrates that for many measures, quality varies inversely with overall medical spending.

Publicly reporting healthcare quality results is becoming more common.

 The Centers for Medicare & Medicaid Services provides comparative quality-of-care results about hospitals, nursing homes, home care agencies, and health plans online.

 The National Committee on Quality Assurance reports quality-of-care results for health plans and by type of insurance product (commercial, Medicaid, Medicare).

 Regional organizations such as Minnesota Community Measurement (MNCM), Massachusetts Health Quality Partners, the Wisconsin Collaborative for Healthcare Quality, and the California Office of the Patient Advocate report quality-of-care results for physician group practices.

 Other programs report results by individual physician.

Although insurance product, health plan, and physician group have been demonstrated to affect measured quality of care, few studies have attempted to determine the importance of these elements relative to each other.

Studies comparing the relative importance of different levels of the healthcare system generally have found that smaller units within the healthcare system contribute more to quality-of-care results. For example, clinic site seems to have a stronger effect on patient experience of care than physician group and health plan; patient and physician factors appear to have a greater effect on diabetes results than clinic site; and the care delivery system has been found to be more related to clinical performance than the health plan.

 We decided to take advantage of a unique data set in Minnesota to compare the relative contribution of health plan, insurance product, and physician group to quality of care for diabetes.

We hypothesized a priori that quality-of-care results would vary across all 3 levels. We predicted that the variation would be the greatest among physician groups and that physician group would contribute more to quality-of-care results than either health plan or insurance product. We also predicted that the results of a composite quality measure using all-or-none scoring and individual quality measures would evidence similar relative contributions to quality. We expected that physician group practice management decisions, practice culture, and office systems would have a stronger effect on patient quality of care than any influences at a higher level in the healthcare system. Physician groups establish practice policies, make investment decisions about electronic health records, implement and maintain patient registries and other office systems, set performance expectations for staff, and interact directly with patients during encounters for medical care.

Health plans focus on improving quality results for their enrollees, often with strategies that supplement care provided by physician practices like telephonic disease management. Physician groups typically care for a mix of patients from most or all health plans within their region, so health plan quality results reflect the average quality of care provided to enrollees by many different physician groups.

Minnesota Community Measurement is an organization that provides standardized publicly reported measures of care quality across many different physician groups, 8 different health plans, and 3 different insurance product types in Minnesota. Data collected by MNCM provide a unique opportunity to address the relative impact of health plan, insurance product, and physician group on quality of diabetes care.

Seven Minnesota health plans and the Minnesota Medical Association initiated MNCM in 2002. Minnesota Community Measurement combines health plan data to measure and publicly report quality of care across multiple clinical domains in more than 70 primary care and multispecialty physician group practices in Minnesota.

The study was designed to assess how much of observed variation in care quality is attributable to health plans compared with insurance products and physician groups. The study also assessed whether using a single all-or-none composite measure of quality would yield results regarding variation and factor contribution similar to those obtained when assessing individual quality measures.

To simplify the assessment, we decided to limit our analysis to a single condition (diabetes) that has several well-recognized performance measures that have good evidence of a relationship to patient outcomes and for which an all-or-none composite measure has been reported.

 We analyzed comparative performance results for glucose control, blood pressure (BP) control, lipid control, tobacco-free status, aspirin use, and Optimal Diabetes Care (a composite of all 5 measures) as publicly reported by MNCM at www.MNHealthcare.org to address the study questions.

Minnesota Community Measurement uses health plan administrative billing data to identify patients of each physician group with various conditions (eg, diabetes, asthma, hypertension). The data set reflects patients enrolled in managed care plans, including not only patients enrolled in commercial health maintenance organization/point-of-service/preferred provider organization products, but also Medicare (cost and risk products) and Minnesota public healthcare programs (Prepaid Medical Assistance, MinnesotaCare, and General Medical Assistance products) from the 8 health plans that provide nearly all of the non-VA insurance coverage in the state. The data do not include the uninsured, patients who pay out-of-pocket, or patients covered by Medicaid/Medicare fee-for-service products. Data were collected by health plans in a standard way using a combination of Healthcare Effectiveness Data and Information Set (HEDIS) and MNCM specifications. Some measures were reported exclusively from administrative (health plan) data, whereas others (eg, diabetes) required on-site review of random samples of individual clinical records. Record reviews of a random sample of at least 60 adults with diabetes receiving care at each physician group were conducted by trained nurse reviewers following a standardized protocol.

Physician groups are defined by MNCM convention and are mutually exclusive. Groups do not share providers. One group, organized as a staff model group, has patients primarily from 1 health plan, providing care for approximately onethird of that plan’s enrollees. Other groups serve enrollees from multiple plans.

Physician group samples are identified through a series of steps. Individual health plans identify patients meeting denominator inclusion criteria per HEDIS specifications. Patients then are assigned to physician groups by each health plan using attribution logic based on evaluation and management and preventive care billing codes. These codes identify the frequency of patient visits to a physician group. Patients are attributed to the physician group they visited most frequently during the measurement year. If patients visited 2 or more groups with the same frequency, the patient is attributed to the physician group visited most recently. Validation processes are part of data collection, aggregation, and reporting. Detailed information about MNCM procedures is available at www.mnhealthcare.org.

Data available from MNCM consisted of deidentified patient- level diabetes performance data for care in year 2005 along with information to allow assignment of each patient to a specific health plan of enrollment, physician group providing care, and insurance product category of commercial, Medicaid, or Medicare. Patients dually enrolled in both Medicaid and Medicare were excluded from this analysis to simplify the comparisons.

Person-level diabetes performance measures consisted of glycosylated hemoglobin (A1C) level, systolic and diastolic BP, low-density lipoprotein cholesterol (LDL-C) values, documented tobacco-free status, and regular aspirin use for those over age 40 years. In addition to these individual measures, MNCM also reports an all-or-none composite measure, Optimal Diabetes Care, which is coded as positive for each patient meeting all 5 criteria: A1C ≤7%, BP <130/80 mm Hg, LDL-C <100 mg/dL, documented tobacco-free status, and regular aspirin use for those over age 40 years.32 The composite measure was coded as a failure for those not meeting all 5 measures or having missing data on any individual measure. However, those ineligible or contraindicated for aspirin use were coded as compliant. Patients with no test value within a year were classified as noncompliant. Raw data files consisted of the universe of diabetes patients sampled for chart review (n = 8401), reduced for analytic purposes by excluding patients dually enrolled in Medicaid and Medicare (n = 683), those lacking a physician group code (n = 524), and those with values outside of a clinically tenable range (eg, A1C <2%, LDL-C <10 mg/dL, diastolic BP <30 mm Hg) (n = 25). The resulting analytic file consisted of 7169 deidentified patient records.

All measures were analyzed as dichotomies to be in line with clinical goals desired in 2005 (A1C ≤7%, BP <130/80 mm Hg, LDL-C <100 mg/dL, documented tobacco-free status, and regular aspirin use for those over age 40 years) and were expressed as rates meeting the goal. Performance results were summarized by reporting the standard deviation of rates, minimum and maximum rates, and 25th/50th/75th percentiles for rates computed at the product, plan, and physician group level. Unadjusted variation by product, plan, and physician group is illustrated graphically in a bubble chart of raw rates (

Logistic regression assessed the contributions of insurance product, health plan, and physician group to the Optimal Diabetes Care composite and its 5 components for care in 2005. Product, plan, and group were treated as fixed effects and entered into models as sets of dummy coded contrasts. The P value from likelihood ratio tests comparing a full maineffects model (effects for product, plan, group) and a model with 1 set of contrasts deleted was used to assess whether the contribution of a particular set of contrasts (eg, for health plan) was significant. Model fit was assessed with changes in Akaike’s information criterion comparing a main-effects model with 2 classification variables to a full model. Because Akaike’s information criterion penalizes fit for the number of terms added to the model, it was used in this article to suggest which term, when added to the model, improved fit the most. The area under the receiver operating characteristic curve (C statistic) was reported for a model with all 3 classification variables in the equation (product, plan, physician group) to assess the overall ability of the model to predict care quality cross the different quality measures. The unique contribution to the C statistic of each classification variable when the other 2 variables were in the equation also was reported as a descriptive assessment of the unique predictive value of that variable above and beyond the other 2 variables. Because the C statistic was not penalized for model complexity and because each classification variable involved a different number of dummy codes (product had 2, plan had 7, group had 56), the magnitudes of these increments to the C statistic could be interpreted as relative contribution to prediction.

Although performance rates aggregated across product, plan, and physician group for individual measures ranged from 0.41 to 0.67, the rate for the Optimal Diabetes Care measure was only 0.08 because of the requirement of satisfying the threshold on all 5 measures for each individual in the denominator (

Variation in performance results is shown in 

 and the Figure by insurance product, health plan, and physician group. The bubble area in the Figure is proportional to sample size. A plot of model-adjusted rates was not notably different from the raw rates presented in the Figure and is not presented. In Table 2, for example, rates for the 57 physician groups for A1C ≤7% ranged from 0.19 to 0.67 (SD = 0.093) with a median rate of 0.47. The Figure illustrates that the low minimum rate of 0.19 was due to an outlier group with a rate considerably lower than other groups. Insurance product rates for A1C ≤7% ranged from 0.44 to 0.59 (SD = 0.081), and health plan rates ranged from 0.35 to 0.54 (SD = 0.065). Physician group rates ranged from 0.22 to 0.62 for BP <130/80 mm Hg, from 0.24 to 0.66 for LDL-C <100 mg/dL, from 0.32 to 0.86 for documented tobacco-free status, and from 0.40 to 0.90 for daily aspirin use. The plan with the largest number of patient-level observations had the highest rate among the 8 plans for 5 of 6 measures (Figure). In addition, the physician group with the largest number of patient-level observations among the 57 groups had among the 5 highest rates for 4 of 6 measures.

Variation as measured by the standard deviation was highest for the physician group for 5 of the 6 measures. The exception was tobacco-free status, for which health plan showed the highest variation. For all measures except documented aspirin use, denominators (number of patient charts) used to compute each rate in Table 2 for insurance product level summaries ranged from 1949 to 2666 with a median of 2554. For health plan, the denominators ranged from 220 to 2038 (median = 776), and for physician group they ranged from 52 to 965 (median = 63). The Figure illustrates that Medicaid patients had poorer quality on most but not all measures. This finding continued to hold in analyses that also considered the effects of plan and physician group.

Likelihood ratio tests all were significant (

), indicating that any variable when added last to the equation significantly improved the prediction of the measure examined. For all 6 measures, the A1C value was lowest when product, plan, and group were all in the model (data not shown) compared with any other main-effects model. The A1C increment column indicates that the addition of product to the model improved the model fit the most when examining A1C ≤7%, LDL-C <100 mg/dL, and the Optimal Diabetes Care measure. Plan improved model fit the most for tobacco-free status and daily aspirin use. Group improved model fit the most for BP <130/80 mm Hg. The total C statistic column indicates that tobacco status, aspirin use, and Optimal Diabetes Care were more accurately predicted by product, plan, and group (C ≥0.69) than the other measures (C ≤0.65). Product, plan, and physician group all showed statistically significant contributions to classification of goal status for all measures.

When we analyzed our data with the large high-performing physician group removed from the analysis, our findings did not change materially. Across the 6 measures reported in Table 2, standard deviations were reduced on average by 5% at the group level, 12% at the plan level, and 9% at the product level. The relative ranks of variation as summarized by the standard deviations in Table 2 remained the same, as did the relative ranks of A1C values in Table 3. Our conclusions remain the same.

All 3 variables are important. The conclusion one draws about relative importance varies depending on what is being examined. Our prediction that physician group would have the strongest relationship to quality-of-care results was true for only 1 of 6 measures: BP control.

For all measures, there was significant variation in the results by each grouping variable, even after considering the other 2 grouping variables (eg, by medical group after considering product and plan). All-or-none scoring lowered performance scores and dampened variation, but maintained the pattern of relative variation seen with A1C, LDL-C, and aspirin use.

Medicaid performance rates were lower than those of other insurance products for 4 of 6 measures: aspirin use, documented tobacco-free status, LDL-C, and Optimal Diabetes Care. Medicaid patients had the better BP control than patients enrolled in commercial or Medicare products, and Medicaid A1C control was intermediate. Medicare performance rates were the highest for all measures with the exception of BP; differences between insurance products were smaller.

Measuring and reporting quality are necessary to improve quality, but the resources for doing so are limited. Understanding the relative contributions of insurance product, health plan, and physician group can help determine where measurement and reporting resources will have the greatest impact. Our data suggest that measurement and reporting have the potential to stimulate quality gains at each of these levels, so it is important to report at each level if possible.

Reporting quality results by health plan and by insurance product is common. Reporting by provider group is not. Insufficient observations and the potential for atypical patients to distort results limit the ability and accuracy of individual health plans to report physician group quality-of-care results. Combining data from multiple health plans has enabled MNCM to report quality-of-care results for most physician groups in the state. More than 80% of primary care physicians practice in these groups.

Two major factors drove the development of these community- wide measures in Minnesota: the need for quality-of- care measures at the physician group level and costly inefficiencies related to regulatory reporting requirements for health plans. In the year prior to pilot-testing MNCM, Minnesota health plans sampled in excess of 12,000 clinical records to fulfill their diabetes quality reporting requirements. 

Despite this large number of clinical record reviews, results were unavailable for many physician groups. It was estimated that a sample sufficient to report performance of physician groups would require many fewer clinical reviews. Despite the apparent enhanced reporting capability and potential to improve quality, the MNCM model faces barriers to adoption in other regions of the country. Health plans commonly publicly report quality results, and competition for business is based partly on those results. Health plans less frequently collaborate to report physician group performance. Preference for improvement strategies capable of differentially impacting a health plan’s members and avoidance of strategies that “lift all boats” might be predicted in this situation. Indeed, many health plans sponsor telephone-based disease management programs to help differentiate their chronic care quality results, although some have suggested that physician group disease management services might be more effective and efficient.

 In other studies, Medicare results have demonstrated little evidence of effect from mandatory health plan reporting.

 A sole focus on health plan quality-of-care results may create a barrier to the development of physician group reporting or improvement efforts.

The data we present have several limitations. First, they are limited to diabetes. However, further investigation is possible as quality results are available for these physician groups for other conditions as well, including immunizations; breast, cervical, and colorectal cancer screening; an all-or-none cancer screening composite measure; and a range of other acute nd chronic condition measures. All are community-wide and well accepted by physician groups, who share ideas for improving care and compete for patients. Second, individual health plans were not identifiable in our data so we could not directly investigate possible explanations for the fact that health plan had a stronger effect than physician group on documented tobacco-free status and daily aspirin use. One health plan in our data had previously reported on a pay-for-performance program that was associated with significant increases in compliance with tobacco treatment guidelines.39 Some plans offered a subset of products, leading to unavoidable collinearity of product and plan in the analysis. This made it more difficult to disentangle the effects of product and plan. Third, because of uneven and sometimes thin distributions of data across combinations of product, plan, and group, the statistical models we used did not account for the possibility that the effect of physician group on performance varied by health plan influence. Plan influence on group characteristics is likely to vary with the proportion of the group’s patient covered by that plan.

One advantage of reporting community-wide quality-ofcare results is that it documents the variation in care quality across an entire community. Levels of diabetes quality of care in this community were better than levels reported nationally from 1988 to 2002. We hypothesize that if physician group reporting became as widespread as health plan reporting, it would be a positive force for community-wide healthcare improvement in other regions, but we see barriers to achieving that objective. We recommend consideration be given to creating incentives for health plans to collaborate to report physician group performance. Our results suggest such reporting could provide an extra stimulant to improve care.

Reporting physician group performance in addition to health plan performance may stimulate greater improvement in diabetes care.

Diabetes Care Quality: Insurance, Health Plan, and Physician Group Contributions

To ascertain the impact of advanced access and productivity-based pay on physician productivity and compensation, patient satisfaction, and medical group cost of delivering care. 

Study subjects were 105 primary care physicians (PCPs) continuously employed at a Minnesota medical group that implemented major changes in access to primary care appointments and to PCP compensation arrangements in 2000. We tracked physician productivity, physician compensation, patient satisfaction, and cost to the medical group per relative value unit (RVU) of patient care from 1998 through 2002. 

In 1998, 105 PCPs (99.6 full-time equivalents [FTEs]) generated 275 000 work RVUs (WRVUs), and PCP pay averaged $123 500 per FTE. In 2002, the same 105 PCPs (now 98.1 FTEs) generated 374 000 WRVUs and $148 000 pay per FTE. From 1998 through 2002, WRVUs per FTE rose 38%, PCP compensation increased 20%, cost of PCP compensation per WRVU produced fell by 13%, overall direct cost of running the clinics per total RVU fell by 20%, and patient satisfaction remained constant. Improvement in operating costs was due to increased physician productivity, lower physician compensation per RVU, and a decline in support staff per 10 000 RVUs from 6.80 to 4.50. 

It is possible to substantially and simultaneously improve costs of care, physician compensation, and patient access without harming patient satisfaction. Advanced access and productivity-based pay may be effective ways to address the challenges of timeliness, efficiency, and patient centeredness identified by the Institute of Medicine. 

, the Institute of Medicine challenged American medicine to transform its approach to care to solve the simultaneous large gaps in various dimensions of quality, including efficiency.

Although most healthcare policy makers have focused on reduction of overuse, underuse, and misuse or on alternative financing mechanisms to lower costs, an approach often emphasized in other industries is to increase productivity.

The productivity of physicians is commonly measured by relative value units (RVUs) of care produced per full-time equivalent (FTE) physician. Improving this measure without increasing support staff costs or increasing charges could reduce healthcare costs and increase capacity to meet the growing care needs of an aging population. Although a change in compensation from salary to piece rate has been linked to increases in productivity of 15% to 30% in other industries,

only a paucity of published data has evaluated the impact of such changes on physician productivity. Because physician productivity may be affected in intended or unintended ways by the many changes that have reshaped medical care organizations and modified healthcare delivery in recent years, it is important to improve our understanding of trends in physician productivity and the factors that influence it. Medical groups that lower the costs per RVU will have a financial advantage in the marketplace. 

We evaluated a natural experiment that occurred in 2000 through 2001 when a large, multispecialty medical group in Minnesota implemented major organizational changes intended to reduce the cost of delivering care. Changes included: (1) a change in physician compensation from straight salary to a fixed amount of pay per work RVU (WRVU). A guaranteed floor of 70% of 1999 compensation was retained in year 2000, but pay became solely dependent on production starting January 1, 2001; (2) a change in scheduling of primary care appointments to facilitate patient access to "same-day care" from their regular primary care physician (PCP) starting in April 2000

; and (3) improved access to appointments, which enabled reduced size and scope of registered nurse medical triage systems at each clinic and led to a reduction in registered nurse staffing at most clinics. We have previously described the change process and analyzed the impact of these changes on utilization and quality of care.

This study was conducted at a Minnesota multispecialty medical group with approximately 120 PCPs who provided care to 225 000 adults insured by an affiliated health plan and to 60 000 otherwise insured or uninsured patients at 18 primary care clinics. In 1998, the average FTE PCP at this group was paid based on straight salary and produced an average 2764 WRVUs per year, substantially less than the national average of 3737.

Medical group leaders recognized low productivity as a serious threat to the survival of the medical group, so they developed a program of sweeping changes in primary care for their group. 

Study subjects were all primary care generalist physicians in family practice, internal medicine, or pediatrics employed at least 0.5 FTE with the medical group on both January 1, 1999, and on December 31, 2002. Approximately 10% of PCPs left the medical group in these years, slightly more than in preceding years; approximately one third of these physicians left due to stated dissatisfaction with changes in the work environment. The study inclusion criteria reduced confounding effects related to the departure of physicians with low RVU production. 

The RVU is a nationally used measure of units of work produced by physicians. A short (99212), average (99213), or long (99214) office visit yields 1.03, 1.42, and 2.20 total RVUs (TRVUs), and 0.45, 0.67, and 1.09, respectively, of these TRVUs are designated as WRVUs. Within primary care, TRVU revenue is allocated roughly: 47% to WRVU, 3% to malpractice, and 50% to practice expenses including overhead. Physician pay is based only on the WRVU component of revenue. WRVUs in this study did not include any RVUs assigned to physicians in exchange for the performance of administrative duties or the ordering of laboratory tests or other ancillary services, but only RVUs generated by the physician in direct patient care. A few physician leaders had FTE subdivided into a portion for their administrative role and a portion for their clinical role. 

We measured all units of cost and productivity based on FTE PCPs. Physician compensation was expressed as gross pay in dollars, but was not adjusted for inflation and was reported as gross pay (before taxes or withholding) per primary physician FTE. FTE-adjusted gross pay excludes any additional pay for work in urgent care. 

was defined in this analysis as the number of WRVUs generated in a year by a given physician, divided by the physician's FTE. Mean physician productivity is calculated by summing the WRVUs and the FTEs of a defined group of physicians and dividing total WRVUs by total FTE physicians. The higher this number, the more clinical work the physicians have performed on average over a given period of time. Physician production efficiency is estimated for subsets of physicians by summing the compensation of all physicians in the defined subset by the number of WRVUs produced by the same physicians. 

Patient satisfaction was based on routinely conducted quarterly patient surveys of randomly selected new and returning patients seen recently by each PCP. Satisfaction scores were based on 5-point Likert scale responses and scaled so that 100 represented the highest possible satisfaction and 20 represented the lowest possible satisfaction. 

This project was reviewed in advance, approved, and monitored by the HealthPartners Institutional Review Board. 

shows that physician production increased from 2930 WRVUs per FTE in 1998 to 3980 WRVUs per FTE in 2002, with the largest increase (15.8%) occurring from 1999 to 2000. Patient satisfaction scores remained stable in the years before and after the organizational changes in 2000, with no statistically significant increase or decline. 

shows that physician production efficiency (defined as physician compensation paid per physician WRVU) decreased from $44.70 under the salary-based compensation system in place until 2000 to $38.85 in 2002 under the productivity-based compensation system. During the same period, average compensation rose from $123 581 per FTE in 1998 to $148 200 in 2002. 

This increase in WRVU was unlikely due to changes in evaluation and management (E/M) coding practices, because these were closely monitored and physicians provided review of random cases every 2 to 3 months to assure accurate coding. The average E/M code level for office visits for family physicians was 3.24 in 2000 and 3.18 in 2002. General internists had average E/M codes of 3.63 in 2000 and 3.44 in 2002. Pediatricians had average E/M codes of 3.14 in 2000 and 3.20 in 2002. The number of clinic days worked increased from approximately 180 days in 1998 to approximately 201 days in 2002 as physicians attempted to generate more income from more WRVUs by seeing more patients in the clinic as patient access increased and shunting of office patients to after-hours urgent care diminished. 

Total compensation continued to increase because productivity increased faster than the rate paid per WRVU decreased. Between 2000 and 2002, total costs/TRVU at this medical group decreased from $65.00 to $52.08, while staff-to-clinician ratio fell from 4.20 to 3.48 and staff/10 000 TRVUs of care provided fell from 6.80 to 4.50. 

In this medical group, physician productivity (WRVU per FTE), physician compensation per FTE, and physician production efficiency (physician compensation divided by WRVU produced) increased, while costs and staff for delivering 1 RVU of care both substantially decreased during the 5-year study. These changes had no measurable deleterious effect on patient satisfaction, which continued at high levels during this 5-year period. Physician compensation changes represented a significant cultural change from equal compensation for all, to compensation for production and performance. 

We have previously reported that improved continuity of care related to these changes in appointment scheduling, in conjunction with other factors, was the driving force behind significant improvements in care quality for patients with diabetes, heart disease, and depression during the study period.

We have also previously described associated changes in the frequency of primary care, urgent care and emergency department visits, hospital admission rates, and costs of care.

The observed changes in physician work activity in the present study are congruent with previous studies that compared productivity of fee-for-service and salaried workers.

Similarly, physicians paid fee-for-service are more productive,

although selection effects related to employment could account for some of the observed differences in productivity and income in comparisons between the 2 groups of physicians.

It is rare to have the opportunity to longitudinally measure changes in physician productivity and production efficiency among a group of physicians during a time of rapid change in how physicians are compensated, and even rarer to do so during simultaneous changes in patient access. Medical group leaders realized that pay for production was needed to support improved access to PCP appointments, and the market rapidly was evolving from capitation to discounted fee-for-service.

With production-based compensation, physician costs vary as a function of RVUs generated and are no longer fixed. This strategy puts the medical group in a much more competitive position by decreasing the cost of providing services. 

The increase in RVUs produced by the physicians in this study could be attributable to at least 4 factors: (1) a one-time increase in services billed from working down the appointment wait time backlog; (2) an ongoing increase in WRVU production per PCP because fewer PCPs were employed after 2000; (3) decreased patient use of urgent care visits and of some subspecialty providers as primary care access improved; and (4) an ongoing increase in services billed due to better capture of charges and more accurate coding of services. Physicians worked more days per year and more hours per day after these changes, and physician reactions ranged from high satisfaction to enough dissatisfaction to leave the medical group.

It is notable that, despite stresses on physicians and staff and dramatic changes in how care was delivered during this period, patient satisfaction did not deteriorate. 

A number of factors limit the interpretation of these data. First, the joint effects of the multiple interventions cannot be separated.

Second, the granularity of the data was not sufficient to analyze intervals of time shorter than a year, or to support time series analysis. Third, generalization of our findings to other medical groups should be done with great caution because of contextual differences. Similar interventions in other organizations with different cultures and levels of productivity or access may not yield the same results.

Despite these limitations, we believe that the results are interesting and important. First, the data show that a coordinated set of planned changes increased physician production efficiency in a short period of time, without compromising patient satisfaction. Second, compensation changes created concordance between the economic realities facing the medical group and the economic incentives facing individual physicians.

 Third, the experience we have reported suggests that a coherent set of complementary changes may benefit most parties involved–patients obtained timely access and better continuity of care; physicians worked harder but received increased income; and the medical group reduced costs of delivering care, thereby improving its position in a competitive market. The anticipated benefits to multiple stakeholders helped to create the momentum needed to initiate change and helped sustain changes once made. 

Our results support the hypothesis that when a high priority for change is combined with an effective change process and with the right types of changes, medical groups can transform themselves, and collectively can address many of the challenges posed by the Institute of Medicine. 

From HealthPartners Research Foundation, Minneapolis, Minn. This work was supported by Robert Wood Johnson Foundation, grant #041868, from the Improving Chronic Illness Care Initiative. 

Address correspondence to: Patrick J. O'Connor, MD, MPH, Senior Clinical Investigator, HealthPartners Research Foundation, 8100 34th Ave S, MC 21111R, Minneapolis, MN 55440-1524. E-mail: patrick.j.oconnor@healthpartners.com. 

Increasing Primary Care Physician Productivity: A Case Study

Professional Advice and Readiness to Change Behavioral Risk Factors Among Members of a Managed Care Organization

Identifying Diabetes Mellitus or Heart Disease Among Health Maintenance Organization Members:  Sensitivity, Specificity, Predictive Value, and Cost of Survey and Database Methods