Louise H. Anderson, PhD

To better understand the direct costs of insomnia. Our study aimed to compare healthcare costs and utilization of patients diagnosed with insomnia who received care in a managed care organization with a set of matched controls.

Our observational, retrospective cohort study compared 7647 adults with an insomnia diagnosis with an equally sized matched cohort of health plan members without an insomnia diagnosis between 2003 and 2006. We also compared a subset of patients diagnosed with and treated for insomnia with those diagnosed with insomnia but not treated.

A large Midwestern health plan with more than 600,000 members.

Multivariate analysis was used to estimate the association between insomnia diagnosis and costs, controlling for covariates, in the baseline and follow-up periods. Although we cannot conclude a causal relationship between insomnia and healthcare costs, our analysis found that insomnia diagnosis was associated with 26% higher costs in the baseline and 46% in the 12 months after diagnosis. When comorbidities were recognized, the insomnia cohort had 80% higher costs, on average, than the matched control cohort.

These outcomes suggest the need to look beyond the direct cost of insomnia to how its interaction with comorbid conditions drives healthcare cost and utilization.

Increasing costs for treated and untreated insomnia suggest the need to look beyond the cost of insomnia to how its interaction with comorbid conditions drives healthcare cost and utilization: 

		At baseline prior to diagnosis, insomnia was associated with 26% higher healthcare costs than matched controls.

		In the 12 months following diagnosis, insomnia was associated with 46% higher costs.

		Predicted follow-up costs for the insomnia cohort were 80% higher than the controls, reflecting the relatively greater increase in morbidity in the insomnia cohort.

		Of those diagnosed with insomnia, 75% received pharmaceutical treatment.

		Those treated for insomnia had higher utilization and costs than those not treated. n Insomnia and its treatment may be indications of more serious underlying conditions.

Chronic insomnia, difficulty falling or staying asleep or experiencing poor-quality sleep, is a growing health problem with significant consequences for both individuals and the healthcare system.

 Chronic insomnia is associated with a host of physical, psychosocial, and emotional problems, including premature mortality, depression, anxiety, and poor quality of life.

 Chronic insomnia can also exacerbate comorbidities.

 In addition to the personal toll of chronic insomnia on the individual, major economic consequences for the healthcare system include increased direct medical costs and healthcare utilization.

 Research examining the economic consequences associated with chronic insomnia has highlighted the burgeoning indirect societal costs of insomnia, such as workplace absenteeism, lost productivity, and increased workplace errors and accidents.

 Less attention has been focused on direct healthcare costs and increased utilization associated with chronic insomnia.

Direct costs to the healthcare system associated with chronic insomnia are difficult to assess due to complexities and gaps in available data.

 Estimates of the aggregate costs of insomnia vary widely depending on the costs considered. 

 interviewed patients in primary care clinics to measure insomnia prevalence, associated functional impairment, lost productivity, and comorbidities. They found that 10% of patients reported insomnia, which was associated with functional impairment, disability, and increased use of health services. A more recent study estimated the direct and indirect costs of untreated insomnia and found that direct and indirect costs for adults younger than 65 years with insomnia were $1253 greater than for subjects without insomnia for the 6-month study period.

The most common treatment for insomnia is pharmacotherapy, with 2.5% of Americans taking prescription drugs to treat insomnia each year, and about 1 in 4 of them continuing treatment for 4 months or longer.

 Commonly prescribed US Food and Drug Administration (FDA)—approved hypnotics include benzodiazepines such as temazepam and triazolam, and drugs that act as agonists to benzodiazepine receptors such as zolpidem.

 These agents are efficacious in the short-term management of insomnia, but adverse effects include residual daytime sedation, cognitive impairment, motor incoordination, dependence, and rebound insomnia.

A variety of cognitive and behavioral therapy (CBT) programs are as efficacious as approved sleep medications in the short term, side effects are nil, and benefits are more durable.

 There is, however, limited access to these programs due to a shortage of trained professionals.

 A 2002 national survey estimated that 2.2% of Americans use complementary or alternative therapies for insomnia.

 However, herbals, dietary supplements, over the counter medications such as diphenhydramine, and alcoholic beverages are not recommended for insomnia treatment due to lack of efficacy data, potential for adverse effects, lack of standardization, or a combination of these concerns.

The purpose of this study was to estimate the healthcare costs associated with insomnia diagnosis by comparing costs and utilization of patients who have been diagnosed with insomnia to a set of matched controls. We compared costs in the baseline period before an insomnia diagnosis, in the 12-month follow-up period postdiagnosis, and the change in cost from baseline to follow-up. We also compared a subset of patients who received a diagnosis of insomnia and associated treatment with patients who received a diagnosis of insomnia but no treatment. This research will contribute to the scarce literature currently available on the direct costs and health services used by patients with chronic insomnia.

We conducted a retrospective observational study using data from a large Midwestern health plan with more than 600,000 members. A study cohort of 7647 adults with an insomnia diagnosis during 2003 to 2006 was identified and compared with an equally sized matched cohort of health plan members without an insomnia diagnosis. All health plan members were eligible for the study if they met the following criteria: (1) at least 18 months of continuous enrollment from January 1, 2003, to December 31, 2006; (2) aged 18 years or older; (3) continuous pharmacy coverage; and (4) Medicaid or commercial insurance coverage.

Members were considered to have insomnia if they had a qualifying insomnia diagnosis code from January 1, 2004, to December 31, 2005, preceded by a 6-month period free of an insomnia diagnosis (baseline period) and if they remained in the plan for at least 12 months after the diagnosis (follow-up period). Qualifying insomnia cohort

 International Classification of Diseases, Ninth Revisio

n codes were: 307.41 (transient disorder of initiating or maintaining sleep), 307.42 (persistent disorder of initiating or maintaining sleep), and 780.52 (insomnia unspecified). This research was reviewed and approved by the local Institutional Review Board.

A propensity score—matching method was used to create a matched control group that had similar covariate values as the insomnia cohort and therefore reduce confounding by covariate differences.

 First, a 10:1 age- and gender-matched sample of members without an insomnia diagnosis was created. Members were then matched with controls during the 6-month baseline period and 12-month follow-up period. Second, a logistic model was used to produce propensity scores for all members with an insomnia diagnosis and the 10:1 matched cohort. The dependent variable in the model was an indicator of a qualifying insomnia diagnosis, and the predictor variables were age, sex, insurance type, and specific individual baseline physical comorbidities (diabetes, cardiovascular disease, chronic obstructive pulmonary disorder, chronic heart failure, myocardial infarction, peripheral vascular disease, connective tissue disorder, ulcer disease, mild liver disease, moderate to severe renal disease, acquired immunodeficiency disorder, tumor, hemiplegia, leukemia or lymphoma, metastatic solid tumor) and specific mental health comorbidities (alcohol use disorder, opioid and substance use disorder, major depression, anxiety, bipolar disorder, posttraumatic stress disorder, dementia, schizophrenia, organic mental disorders, other psychotic disorders, affective disorders, personality disorders, or adjustment disorders). Finally, propensity scores from the model were used to identify a 1:1 cohort, matched on propensity score. This 1:1 propensity score—matched cohort was the comparison group to evaluate differences in healthcare costs and utilization. Multivariate regression methods were used to adjust for differences that remained between the cohorts after matching.

To compare members with an insomnia diagnosis with and without pharmacological treatment, pharmacological treatment was defined as a prescription fill for benzodiazepines or nonbenzodiazepines, categories of medications that have been approved by the FDA for the treatment of insomnia,

or prescriptions for medications frequently used off label for treatment of insomnia, including: trazodone, ziprasidone, amitriptyline, mirtazapine (with no depression diagnosis), doxepin, quetiapine fumarate (Seroquel), and diphenhydramine (Benadryl).

Healthcare cost and utilization data were obtained from the health plan’s administrative databases. They included type of insurance coverage, diagnostic codes for insomnia and other chronic conditions, plan reimbursed amounts, and member-paid amounts. These data were used to identify major medical comorbidities and total direct medical costs, including inpatient, outpatient, pharmacy, urgent care and emergency department services, for all members. Cost was defined as plan reimbursed amounts plus member-paid.

Multivariate analysis was used to estimate the association between study cohort and costs, controlling for covariates. The outcome was total healthcare costs, which captures the effect of insomnia on the use of healthcare services for all reasons. Estimates of the percentage of cost associated with an insomnia diagnosis were made at 3 times: the baseline period, the follow-up period, and the change from baseline to follow-up. For the baseline and follow-up analyses, a 2-part Heckman estimator was used because of the number of members in the control cohort that had no healthcare costs. First, the association between cohort and positive cost was determined by logistic regression. Second, the association between cohort and costs was estimated for those with positive costs. Because the outcomes were heavily skewed, a log transformation with Duan’s smearing estimator was used.

 For the analysis of change from baseline to follow-up, multivariate linear regression analysis was used and the dependent variable was untransformed dollars.

The covariates were age, sex, insurance type, Charlson Comorbidity Index (CCI) score, psychiatric medication use, and mental health diagnoses. We included interaction and polynomial terms that improved model performance. Comorbidities were identified using clinical classification software maintained by the Agency for Healthcare Research and Quality or defined by the health plan’s algorithms, which were summarized as a CCI score.

 The CCI includes 19 medical conditions, each with a weight of 1, 2, 3, or 6. The CCI score is the sum of the weights for the conditions identified as comorbidities. Total scores can range from 0 (no conditions) to 37 (all 19 conditions).

The parameter for insomnia diagnosis in the regression models for baseline and follow-up costs was used to estimate the percentage of costs that were associated with insomnia. To estimate the impact of insomnia and comorbidities on total costs, we used regression results to estimate predicted costs for each cohort in the follow-up period, with covariate values set at the mean value for each cohort.

We identified 7647 members with a diagnosis of insomnia who met the enrollment criteria during the identification period. They were matched with 7647 controls. In the insomnia cohort, the average age was 48 years, and 64.2% were women, compared with an average age of 49 years and 66.9% women in the control cohort (

). The average CCI score for the insomnia cohort in the baseline period was 0.26, while it was 0.28 for the control cohort. In the follow-up period, the average CCI score increased for both the insomnia and control cohorts and the percentage of members with a mental health diagnosis or psychiatric medication use increased for the insomnia cohort. In the follow-up period, all measures were greater for the insomnia cohort than for the control cohort (Table 1). Baseline and follow-up utilization and costs for each cohort are presented in 

. The insomnia cohort had higher utilization and costs in all categories than the control cohort in both baseline and follow-up periods. The average cost for the insomnia cohort during the 6-month baseline period was $5484, compared with $3937 for the control cohort. In the year-long follow-up period, the average cost for the insomnia cohort was $11,206, compared with $6939 for the control cohort.

Estimated model coefficients for the 2-part models of baseline costs are reported in Table 3. The probability of positive cost in the baseline period was higher for members in the insomnia cohort, women, and those with Medicaid insurance. Among 14,406 members with positive cost in the baseline period, members in the insomnia cohort, women, and those with Medicaid insurance, a mental health diagnosis, and use of psychiatric medication had higher costs. Costs were higher with increased age and CCI score. The parameter estimate for insomnia diagnosis was 0.23, which translates to 26% higher costs associated with insomnia in the baseline period. (The percentage change in US dollars is equal to the exponent of the parameter for log dollars, minus 1.)

Estimated model coefficients for the 2-part models of follow-up costs are reported in 

. The probability of positive cost in the follow-up period was higher for members in the insomnia cohort, females, and those with Medicaid insurance. Among 14,975 members with positive cost in the follow-up period, members in the insomnia cohort, women, and those with a mental health diagnosis and use of psychiatric medication had higher costs. Costs were higher with increased age and CCI score. Presence of Medicaid insurance was not a significant predictor of follow-up cost. The parameter estimate for insomnia diagnosis was 0.38, which translates to 46% higher costs associated with insomnia in the follow-up period. Using mean covariate values for each cohort, the predicted probability of positive costs was 100% for the insomnia cohort and 97% for the control cohort. The predicted follow-up cost for the insomnia cohort was $4395, 80% higher than predicted cost for the control cohort. The analysis of change in costs from the baseline to the follow-up period showed that insomnia was associated with an increase of $1553 from baseline to follow-up (Table 5).

Treated Versus Untreated Subsets of the Insomnia Cohort

In the insomnia cohort, 5773 (75%) were treated with a prescription medication for insomnia during the 12-month follow-up period. Compared with the insomnia cohort subset without treatment, the treated subset was slightly older (48.85 years vs 47.31 years), more likely to be women (65.7% vs 59.6%), more likely to have Medicaid insurance (8.6% vs 4.2%), and had a slightly higher CCI score (0.27 vs 0.21). The treated subset was more likely to have a mental health diagnosis in the baseline period (27.3% vs 16.0%) and to have a pharmacy claim for an antidepressant (47% vs 27.5%) and a mood stabilizer (9.0% vs 3.3%). All measures were statistically different at 

 <.005. Those differences persisted and grew into the follow-up period.

In the follow-up period, the treated subset had a higher CCI score (0.46 vs 0.29) and higher percentage of patients with a mental health diagnosis (50% vs 43%), an antidepressant prescription (77% vs 34%), and a mood stabilizer prescription (15% vs 5%) than the untreated subset. The treated subset also had significantly greater utilization and costs in all categories in both the baseline and follow-up periods. The analysis of cost change from the baseline period to follow-up period showed that the treated subset had a significantly greater increase in total costs than the untreated subset ($5527 vs $3145), and a significantly greater increase in medical cost (excluding pharmacy) than the untreated subset ($4276 vs $2309).

Our results confirm previous findings that untreated insomnia is associated with increased healthcare utilization and costs. Ozminkowski and colleagues

 found that direct costs for patients with untreated insomnia were 25% greater than a matched comparison group, $4755 versus $3831. We estimated that the insomnia diagnosis was associated with 26% higher costs in the baseline period, when members who eventually received an insomnia diagnosis were as yet untreated (Table 3). In addition to estimating the association of healthcare costs with untreated insomnia (during the baseline period), we estimated the association of healthcare costs with diagnosed insomnia during the follow-up period. We found that, in the 12 months after diagnosis, the insomnia diagnosis was associated with 46% higher costs (Table 4). Using regression results, we predicted the follow-up costs for the insomnia and control cohorts, using their respective covariate values. That analysis illustrated the impact of insomnia and comorbidities on total costs. We found that, in the 12 months after insomnia diagnosis, the insomnia cohort had 80% higher costs, on average, than the control cohort. The amount of the change from baseline to follow-up that was associated with insomnia was estimated to be $1553 (Table 5). In general, healthcare costs tend to increase from one period to another, but we found that the insomnia cohort had a greater relative increase in costs from baseline to follow-up than the control cohort. Part of that greater relative increase appeared to be due to the relative decrease in health of the insomnia cohort, as evidenced by the CCI score and mental health measures.

Articles

Patients with an insomnia diagnosis have higher healthcare utilization and costs than a matched control group, both before and after the diagnosis.

Healthcare Utilization and Costs in Persons With Insomnia in a Managed Care Population

To examine the long-term relationships between costs, utilization, and patient-centered medical home (PCMH) clinical practice systems.

Clinical practice systems were evaluated at baseline by the Physician Practice Connections-Research Survey (PPC-RS). Annual costs and utilization of a retrospectively constructed cohort of 58,391 persons receiving primary care at 1 of 22 medical groups over a 5-year period (2005-2009) were compared.

Multivariate regressions adjusting for patient demographics, health status, and autoregressive errors compared PPC-RS scores and study outcomes for the entire cohort and 3 subcohorts defined by medical complexity (medication count 0-2 [n = 29,657], 2-6 [n = 19,505], 

7 [n = 9229]). Outcomes (adjusted to 2005 dollars) were total costs, outpatient costs, inpatient costs, inpatient days, and emergency department (ED) use.

For the entire cohort, a 10% increase in PPC-RS scores was associated with 3.9 (medication count: 0-2), 6 (3-6), and 11.6 (

7) fewer ED visits per 1000 in 2005; and 5.1, 7.6, and 13.6 fewer ED visits in 2009. That 10% increase was not associated with the 0-2 medication subcohort’s total (−$22/person in 2005; $184/person in 2009), outpatient (−$11/person in 2005; $42/person in 2009), or inpatient ($26/person in 2005; $29/person in 2009) costs. However, it was associated with significantly decreased total (−$446/person in 2005; −$184/person in 2009) and outpatient (−$241/person in 2005; −$54/person in 2009) costs for the most medically complex subcohort (

Association of PCMH clinical practice systems with reduced costs appears limited to the most medically complex patients.

Patient-centered medical home (PCMH) clinical practice systems have been associated with quality improvement and short-term (12-24 months) reductions in patient medical costs and utilization. Using a retrospectively constructed longitudinal cohort, we examined the association between PCMH clinical practice systems and annual medical costs and utilization over a 5-year period.

		Improved PCMH clinical practice systems were associated with reduced emergency department utilization for all patients.

		Higher-functioning baseline PCMH clinical practice systems were associated with lower total and outpatient costs for medically complex patients (ie, 7 or more active prescriptions) over 5 years.

The patient-centered medical home (PCMH) is a topic of interest.

 A high-functioning medical home requires coordinated care by a consistent team.

 Although few clinics are PCMHs, adults reporting a usual source of primary care are 25% more likely to report positive clinician attributes22 and reduced disparities.

 Observational studies suggest the PCMH approach results in improved satisfaction

 Pilot studies have found reduced emergency department (ED) use

Clinical practice systems are an important component of a PCMH.

They give access to relevant information, coordinate management of complex conditions, and facilitate delivery of preventive care services. The Physician Practice Connections—Patient-Centered Medical Home (PPC-PCMH), a tool endorsed by the National Committee for Quality Assurance (NCQA), measures practice systems and has been used in PCMH programs.

 have looked at short time frames of 12 to 24 months.

 In theory, the PCMH would reduce medical costs over time by avoiding complications leading to ED visits and inpatient stays, especially among patients with complex illness, as suggested by the chronic care model of Bodenheimer et al.

 This study uses a 2005 measure of the PPC-PCMH and a retrospectively constructed cohort from a large Midwestern health plan to evaluate whether clinical practice systems evaluated at baseline are associated with reduced utilization and costs over a subsequent 5-year period. We present key findings in terms of predicted annual per person amounts to illustrate (1) how predicted costs and utilization change in response to clinical systems and (2) how baseline clinic systems related to different patient groups.

Data Sources, Study Population, and Primary Care Medical Groups

Utilization, billing, provider, medical group, and patient demographic data came from the administrative databases of a large, not-for-profit Midwestern health plan. Practice system measures came from a 2005 survey of medical group directors using the Physician Practice Connections-Research Survey (PPC-RS).

 This instrument is similar to the PPC-PCMH except for fewer questions about the electronic medical record, a focus on 4 chronic conditions (diabetes, cardiovascular disease, asthma, and depression), and graded response categories. A full description of the tool is available on request.

A retrospective cohort over 2005 to 2009 was constructed to compare a baseline measure of clinic systems with subsequent utilization patterns. Subjects needed to meet the following inclusion criteria: (1) have 10 or more months of continuous enrollment in each year; (2) be alive on December 31, 2009; (3) be 19 years or older as of January 1, 2005; and (4) be attributable to the same primary care medical group for 2005 to 2009. Persons were attributed to the medical group with which they had the greatest percentage of primary care visits. Primary care visits were defined as visits with providers in the following specialties: family medicine, internal medicine, general practice, geriatric medicine, and obstetrics and gynecology (Ob-Gyn). Nurse practitioner and physician assistant visits were included. Our decision to include visits with Ob-Gyn providers was made because such visits are a regular source of care for many women of childbearing age. However, such an inclusion does not strictly conform to the personfocused primary care concept of Starfield.

 The implications of this decision are discussed in the Limitations section.

Those with no primary care visits were unattributed and excluded. Those attributed to more than 1 medical group (ie, those who had an equal number of visits to 2 or more groups) were also excluded. We further limited the sample to medical groups with a minimum of 200 attributed members. Per year, of the 318,857 adults attributed in 2005 approximately 5% were excluded due to death or disenrollment, 25% due to a change in attributed medical group, and 5% due to no primary care utilization. This resulted in 58,391 persons across 22 medical groups. Most (n = 48,292) had commercial insurance, 7077 were Medicare enrollees, and 3022 were enrolled in Medicaid. For dually eligible Medicare beneficiaries aged 65 to 75 years, all claims including pharmacy were processed by the health plan in order to track benefits, deductibles, and payer liability. All claims from Medicaid beneficiaries were processed for similar reasons.

Five annualized outcomes were constructed: total cost, total outpatient cost, total inpatient cost, inpatient days, and ED visits. The health plan’s administrative databases contain information concerning insurance product, medical diagnosis, care specialty, costs, and limited demographics (age, address, and sex). These were organized using Evaluation and Management, 

International Classification of Diseases, Ninth Revision

, and Current Procedural Terminology (CPT) codes. Total cost included all reimbursed medical costs, including copays, coinsurance, and deductibles. Outpatient cost included professional services, prescriptions, lab and x-ray tests, and outpatient surgical procedures. Inpatient cost included professional and facility fees for hospital-based services including emergency care. Inpatient days were days with an overnight hospital stay. ED visits included all visits to an ED with reimbursed service. If a subject was not enrolled for the entire year, their cost and utilization was annualized using their monthly average.

To avoid variation in outcomes due to benefit design or provider contract, total, outpatient, and inpatient costs were based on a standardized measure, the relative resource value unit. Relative resource value units are based on Centers for Medicare & Medicaid Services relative value units, inpatient diagnosis-related groups (DRGs), and Ambulatory Payment Classification weights. The logic is to apply a standardized fee schedule across all providers by developing standardized costs for each CPT code, hospital DRG, and National Drug Classification (NDC) code that is dependent upon the type of procedure/service/prescription provided but independent of the place of service, type of insurance coverage, or year. This fee schedule was developed by constructing a weighted average of billed amounts across all contracted providers for each CPT code, hospital DRG, and NDC code. Our measures of costs were developed by adjusting these averages by the ratio of billed to paid amounts across service category and scaled to the base year of 2005.

 asks 53 questions related to delivering preventive services, depression, diabetes, cardiovascular disease, and asthma. Of these, 43 are grouped into domains corresponding to the Chronic Care Model: Health Care Organization (n = 3), Delivery System Redesign (n = 8), Clinical Information System (n = 10), Decision Support (n = 9), and Self-Management Support (n = 23). Items are coded as present and work well (1 point), present but need improvement (1/2 point), or absent (0 points). Domain scores represent the proportion of possible items present and utilized. The PPCRS score is a summation of all items with high scores associated with higher-functioning clinical systems.

Multiple regression models were estimated using generalized estimating equations for continuous outcomes and generalized linear models for discrete outcomes. All models were fit with a subject-level autoregressive error structure (AR1 process). Cost outcomes were log transformed

 and Duan’s smearing estimator was used.

 A majority had neither inpatient nor ED visits within a given year. Thus, 2 outcomes were considered. First, the likelihood of any utilization was modeled using a logistic regression with a subject-level AR1 process. Among those with inpatient costs, a log-transformation was used. For ED utilization and inpatient days, a zero-inflated Poisson model was used.

Our models compared a baseline measure of clinical systems with costs and utilization over a 5-year period. They adjusted for demographics (age and sex), complexity/comorbidity (number of medications and comorbidities), insurance type, and primary care visits. Outpatient prescription medications were our measure of complexity because this information was reliably available from claims data, and it is a validated, easily reproducible measure.

 Certain results categorized subjects by their baseline number of prescriptions, but models allowed that number to vary by year. All of the models were developed in the following manner. First, candidate covariance structures were considered. Second, demographic models were constructed. Covariates significant at the 10% level in univariate models were screened for confounding, multicolinearity, and consistent linear relationships. Appropriate adjustments (transformations, interactions, and polynomial terms) were made. Study year was considered as both a continuous and fixed effect with a fixed-effect specification preferable (likelihood ratio test; 

 = .0002). Finally, PPC-RS scores were added, and the possibility of interactions with both study year and patient demographics was considered.

To control for variation in outcomes due to differences in patient mix when presenting our results, we used a 2-part strategy. First, as discussed, we incorporated multiple patientlevel factors adjusting for patient demographics (age and sex), ability to pay (insurance type), and medical complexity (prescription drug use and comorbidities) into our final multivariate models. Second, we estimated the impact of changes in PCMH-related systems (PPC-RS scores) holding fixed patient demographics.

 contains demographic information and average per patient costs. Total costs averaged $10,347 (standard deviation [SD] $22,384) in 2005 and trended upward to $13,637 (SD $32,905) in 2009. The average 2005 age was 52.3 years with 64% being female, 13% using Medicare, and 1.8% using Medicaid. Over the study, an additional 6% enrolled in Medicare; however, Medicaid enrollment was stable.

Depression was the most prevalent chronic condition, impacting 24% in 2005 and 27% in 2009. The next most prevalent was asthma (10% in 2005, 9% in 2009), then diabetes (10% in 2005, 14% in 2009) and coronary artery disease (5% in 2005, 5.5% in 2009). Patients averaged 3.5 prescriptions in 2005 and 4.2 prescriptions in 2009. A brief discussion of demographics and specific coefficient estimates is in the 

contain predicted amounts from our multivariate models. These predict per person costs (Tables 3 and 5) or utilization (Table 4), adjusting for patient demographics and comorbidities. They illustrate (1) how predicted costs and utilization change according to changes in baseline clinical systems (Table 3) and (2) how a hypothetical 10% improvement in baseline clinical systems potentially affects different groups differently.

Table 2 lists attributed patients and average predicted cost by medical group. Estimated average healthcare costs by medical group ranged from $8717 to $12,155. The medical group closely affiliated with the health plan had the largest attribution (n = 27,331) with an estimated cost of $10,474. Attribution to the remaining groups ranged from 8436 to 227 persons. Table 2 also lists 2005 PPC-RS scores that ranged from 25 to 85.

Relationship Between Costs and PPC-RS Scores

 <.001), no samplewide associations with the PPC-RS were found. However, significant relationships with subgroups existed. Among complex patients (

7 medications), higher scores were associated with lower total costs, outpatient costs, inpatient days, and ED use.

Table 3 presents estimated amounts by year and quartile of baseline PPC-RS scores. In 2005, predicted per patient costs in groups whose PPC-RS score was at the sample’s 3rd quartile (PPC-RS = 70.3) were approximately $100 less than those whose medical group was at the sample median (

 = .025) and $240 less than those whose PPC-RS was at the lowest quartile (

<.01). This difference decreased over time, becoming negative in 2009 (

 = .16). Outpatient utilization differed by $45 (

 <.001) between the 3rd and lower quartiles in 2005, with the difference decreasing over time until it was associated with increases in 2008 and 2009 (

 = .023, respectively). Higher PPC-RS quartiles were associated with increased inpatient utilization across all study years but less ED use. In 2005, the 3rd quartile of PPC-RS scores was associated with 8.5 and 18.4 fewer ED visits per 1000 persons compared with the median and 1st quartile (

 <.01, respectively), and increased to 10.6 and 23.2, respectively, in 2009 (

Tables 4 and 5 estimate the impact of a hypothetical 10% increase in 2005 PPC-RS scores for 3 groups defined by their number of 2005 outpatient prescriptions: 0 to 2 medications (n = 29,657), 3 to 6 medications (n = 19,505), and 7 or more medications (n = 9229). Among those with 0 to 2 medications, the 10% increase was associated with a slight increase in total, outpatient, and inpatient costs not significant at the .05 level. Among those with 3 to 6 medications, the increase was associated with a slight decrease in average total and outpatient costs not significant at the .05 level, but with a slight increase in average inpatient costs not significant at the .05 level. Among those with 7 or more medications, there was a small decrease in average inpatient costs not significant at the .05 level, whereas outpatient costs decreased significantly by 1.12% (

 = .015). The number of ED visits decreased by 3.7% for all 5 years (

 <.01), as did the number of impatient days (

From 2005 to 2009, improved clinical practice systems were associated with cost reductions only for medically complex patients.

Patient-Centered Medical Home Cost Reductions Limited to Complex Patients

 To measure continuity among medical groups of insured patients over a 5-year period and to test whether group continuity of care is associated with healthcare utilization and costs.

 We studied natural patient behavior by using insurance claims data in the absence of any medical group or health plan incentives for continuity. We conducted the study through a retrospective analysis of administrative data of 121,780 patients enrolled from 2005 to 2009 in HealthPartners, a large nonprofit Minnesota health plan. Each year, patients were attributed to the medical group where they received the greatest number of primary care visits. Multilevel multiple regression models were used to estimate the association of annualized medical cost and utilization with attribution and continuity categories.

Although patients with high medical group continuity were older and had more comorbidities than patients with medium or low continuity of care, they had a consistently lower probability of any inpatient expenditure or any emergency department (ED) utilization and lower total medical costs.

Although a small proportion, health plan members who visited a primary care provider but had low or medium continuity among medical groups had higher inpatient and ED use than those with high continuity. Improved coordination and integration has potential to lower utilization and costs in this group.

Patient continuity with a medical group and its relationship to medical expenditures and healthcare utilization are of interest to accountable care organizations.

		Among study patients with at least 1 annual visit to a primary care provider, the majority did not move among medical groups for primary care.

		Patients who moved among medical groups the most often had the highest total costs and the greatest use of inpatient and emergency department care.

		 Although a small proportion, patients who moved among medical groups the most often had the highest total costs and represent an opportunity for better management.

Fisher et al proposed the accountable care organization (ACO) as a way to increase coordination of care and financial responsibility while lowering costs and improving quality.

 The ACO, an organization of physicians and institutional providers such as hospitals, home health agencies, and nursing homes, would be accountable for all of the care for a defined population.

 The Patient Protection and Affordable Care Act authorized Medicare to contract with ACOs to provide care for fee-for-service beneficiaries beginning in January 2012.

 The Medicare contracts include a shared savings program to give financial incentive to ACOs to improve quality and contain costs.4 As the methods to define patient populations, quality measures, and cost containment goals are being developed for the Medicare Shared Savings Program, private payers have also been testing ACO models for their insured populations.

The ACO model is not limited to a single organizational structure. Five types of arrangements have been suggested: integrated delivery systems, multispecialty group practices, physician-hospital organizations, independent practice associations, and virtual physician organizations.

Central to the ACO concept is the organization’s accountability for a defined population. Defining that population requires assigning each patient to a clinician and, by extension, to the ACO. The assignment, or attribution, method can take various forms. It can be retrospective or prospective, can involve single or multiple clinicians, or can be based on patient choice, patient visits, or dollars.

Regardless of the method, attribution is assignment of a patient to an ACO for a single year. As patients seek care from different clinicians from year to year, they may be attributed to different ACOs. Because the ACO assumes financial risk for its attributed patients, the ACO will be concerned not only about the annual attribution but also about the stability of the population and associated financial impacts.

Large medical groups are organizations of healthcare providers that contract with health plans to provide medical care to health plan members. They are potential ACOs because they have the ability to implement clinical, managerial, and financial systems to coordinate the entire spectrum of patient care.

 In this study we explored the importance of patient mobility among medical groups as a proxy for understanding patient mobility among ACOs.

Our objectives were to measure continuity of insured patients among medical groups over a 5-year period and to test whether group continuity of care was associated with healthcare utilization and costs. Thus, it was a retrospective observational study of natural patient behavior tracked through insurance claims data in the absence of any medical group or health plan incentives for continuity. We addressed 2 questions relevant to ACO success. (1) What patterns of patient continuity could be identified? (2) What were the associations among those continuity patterns and total medical costs, inpatient utilization and costs, and emergency department (ED) utilization? Our results will be helpful to policy makers as they evaluate the potential for ACOs to deliver cost savings and as ACOs evaluate the risk they are assuming for their patient populations.

We compared annual medical utilization and costs in patient groups, defined by their pattern of medical group attribution and continuity over a 5-year period. Each year, patients were attributed to the medical group where they received the greatest number of primary care visits, without regard to numbers of providers seen. In case of ties, patients were attributed to the medical group where the most recent visit occurred. Primary care visits were defined by location and specialty of the billing physician and included the following specialties: family medicine, internal medicine, pediatrics, geriatrics, and obstetrics and gynecology. Nurse practitioner and physician assistant visits were also included. Patients without primary care visits in a year were not attributed in that year.

Patients were categorized into 3 groups on the basis of their attribution: not attributed (NotAtt), or patients who were not attributable to a medical group in any of the 5 study years; infrequently attributed (InfAtt), or patients who were attributed only 1 or 2 years during the study period; and frequently attributed (FreqAtt), or patients who were attributable for 3 or more years.

The FreqAtt group was further segmented into 3 continuity categories according to the number of moves the patients made among medical groups: high continuity (HiCont), or patients who were always attributable to the same medical group; medium continuity (MedCont), or patients who made 1 move between medical groups; and low continuity (Low- Cont), or patients who made 2 or more moves among medical groups. A move was defined as a change in attributed medical group. For example, a member who was attributed to medical group A in 2005, who was not attributable in 2006, and who was attributed to medical group A in 2007, did not move. The same member who was not attributable in 2008 and who was attributed to medical group B in 2009 moved once. Patient moves among clinics in the same medical group were not considered moves for this study.

This was a retrospective analysis of administrative data from 2005 to 2009 from HealthPartners, a large nonprofit Minnesota health plan with more than 600,000 members. All HealthPartners members during the study period were eligible if they met the following criteria: (1) at least 10 months of health plan enrollment in each study year; (2) aged 19 years or older on January 1, 2005; (3) pharmacy coverage for each study year; and (4) commercial, Medicaid, or Medicare insurance coverage.

The medical groups included in the study all provided primary care and were equally divided between those having 7 to 40 physicians, 41 to 110 physicians, and more than 110 physicians. Thirty-seven percent were single-specialty primary care groups, and 54% were located outside the metropolitan area of Minneapolis-St. Paul.

Utilization, costs, and patient demographic data came from HealthPartners administrative databases and included information describing type of insurance coverage, diagnostic codes, procedure codes, billed amounts, plan reimbursed amounts, and member liability (ie, patient paid). These data were used to identify major medical comorbidities for all members (asthma, cardiovascular disease, congestive heart failure, chronic obstructive pulmonary disease, depression, and diabetes).

Yearly total cost of care was defined as the total amount paid for medical services to medical providers during the year. To avoid variation from differences in contracted reimbursement rates across medical groups, all costs were based on a standardized measure, the HealthPartners relative resource value unit. These units are based on Centers for Medicare & Medicaid Services relative value units, inpatient diagnosisrelated groups, and ambulatory payment classification weights. Use of standardized costs for each Current Procedural Terminology code, diagnosis-related group, and pharmacy claim made patient costs independent of the provider contract or type of insurance coverage. All costs are expressed as 2005 dollars. Costs and utilization counts were annualized for members with fewer than 12 months of enrollment in a given year.

Multilevel multiple regression models were used to estimate the association of annualized medical cost and utilization with attribution and continuity categories (NotAtt, InfAtt, HiCont, MedCont, LowCont). To account for within-subject correlation across observation years, a generalized estimating equations approach was used. For continuous outcomes of total and inpatient costs, a 2-part Heckman estimator was used.

 First, the association between patient categories and positive expenditure was determined by logistic regression. Second, the association between categories and expenditures was estimated for those with expenditures. Because both continuous outcomes were heavily skewed, a log transformation with Duan’s smearing estimator was used.

 For the count outcome of ED utilization, a zero-inflated Poisson model was used, first with logistic regression to estimate the association between patient categories and positive ED utilization, and second, with Poisson regression to estimate the association between categories and number of ED visits among patients with ED use.

The multivariate models adjusted for patient demographics, complexity, comorbidities, and study year. Number of medications was our measure of patient complexity because it was reliably available from pharmacy claims data, and it is a validated, transparent, and easily reproducible measure.15 We included interaction and polynomial terms that improved model performance.

To more easily interpret model results, we estimated predicted outcomes (probabilities, costs, and utilization) for each patient category, with covariate values set at the mean value within each category. We also estimated the marginal effect of changing category assignment. The predicted outcomes for the average MedCont and LowCont patients were estimated as if they were HiCont patients. These estimates quantified the impact of patient continuity.

We identified 121,780 patients who were covered by HealthPartners insurance and met the inclusion criteria. Among patients enrolled in 2005, 28% were excluded because of age, and 37% were excluded because they were not enrolled with HealthPartners in all years. An additional 16% were excluded because of noncontinuous pharmacy coverage or fewer than 10 months of enrollment each year.

In 2007, the study midpoint, the average patient age was 50.5 years, 52.4% were female, and most patients had commercial insurance (88.2%). Roughly 7% (7.3%) were identified as having asthma, 3.3% coronary artery disease, 0.8% congestive heart failure, 2.6% chronic obstructive pulmonary disease, 19.2% depression, and 7.5% diabetes. Total costs averaged $8547 per patient, and there were an average of 2.2 primary care visits per patient.

Characteristics and utilization of the study population by attribution and continuity categories are presented in 

. Of the 121,780 patients, 5031 (4%) were not attributable (NotAtt) in any of the 5 years, and 14,590 (12%) were categorized as InfAtt. The NotAtt and InfAtt patients were younger, included fewer females, and were less likely to have chronic conditions. They also had the lowest average total annual costs, lowest inpatient hospitalization rate, and lowest rate of ED use.

Most (84%) were categorized as FreqAtt. They had the highest average age (51.5 years), the largest percentage of females (57.1%), and the highest prevalence of chronic conditions, and were more likely than patients in the other categories to have Medicare coverage.

Most of the FreqAtt patients were categorized as Hi-Cont. That group was older (52.1 years) and mostly female (57.0%), and had the highest prevalence of all chronic diseases except depression. Eight percent (7775 of 102,159) of FreqAtt patients were categorized as MedCont, and 3% (3237 of 102,159) were LowCont. The MedCont and LowCont groups were younger than the HiCont group but were similar in the proportion of females and prevalence of asthma and depression.

Patients who often change medical groups have the highest healthcare costs. The improved coordination of an accountable care organization may reduce those costs.

Louise H. Anderson, PhD

Articles

Healthcare Utilization and Costs in Persons With Insomnia in a Managed Care Population

Patient-Centered Medical Home Cost Reductions Limited to Complex Patients

Patient Medical Group Continuity and Healthcare Utilization