: To evaluate variation in outpatient antibiotic utilization among US commercial health plans and the implications of this variation for cost and quality.

: We measured antibiotic utilization rates among 229 US commercial health plans that participated in the 2005 Healthcare Effectiveness Data and Information Set. Rates were adjusted to account for health plan age and sex distribution. To estimate antibiotic costs, we multiplied utilization data for each drug class by national estimates of intraclass distribution of drugs, duration of therapy, and median average wholesale price.

: Antibiotic utilization rates varied markedly among plans, ranging from 0.64 antibiotic fills per member per year (PMPY) at the 5th percentile of plans to 1.08 fills PMPY at the 95th percentile, with a mean of 0.88 (SD ± 0.15) antibiotic fills PMPY. US census region was the strongest predictor of antibiotic utilization. Antibiotic costs averaged $49 PMPY and ranged from $34 to $63 PMPY among plans at the 5th and 95th percentiles of cost, respectively. If a health plan with 250,000 members at the 90th percentile of antibiotic costs reduced its costs to the 25th percentile, annual drug cost savings would be approximately $4.1 million.

: Antibiotic utilization varies substantially among commercial health plans and is not accounted for by differences in the age and sex distribution of plan members. Because reducing rates of antibiotic utilization is likely to lower costs and improve quality, high-utilizing plans may reap considerable rewards from investing in

programs to reduce the overuse of antibiotics.

Outpatient antibiotic utilization varies substantially among commercial health plans in ways not explained by patient case mix. As a result:

Antibiotic utilization is likely to be a valuable marker of prescribing quality for health plans.

Health plans with high rates of antibiotic utilization may benefit from targeted quality improvement programs to reduce unnecessary antibiotic use.

Cost savings from reducing unnecessary antibiotic use are substantial.

Antibiotics are widely overused in ambulatory practice, particularly in the management of acute respiratory tract infections. 

 Although bacterial infections cause a small minority of these illnesses, it is estimated that 40% to 50% of all patients in the United States who seek medical attention because of these conditions receive antibiotics.

 The consequences of this overuse are striking. Every year, millions of people are directly exposed to the side effects of antibiotics, ranging from common, bothersome symptoms to infrequent but devastating complications such as Clostridium difficile colitis and anaphylaxis.

 Moreover, research suggests that community levels of bacterial resistance occur in proportion to the volume of community antibiotic use.

 Thus,  perpetuation of antibiotic overuse promotes the continued evolution of ominous resistance profiles. 

Since early in this decade, the European Surveillance of Antimicrobial Consumption project has been collecting country-level data on  antibiotic utilization and has documented large differences in per capita antibiotic utilization between countries in Europe.

 Because the European Surveillance of Antimicrobial Consumption measures are not based on a particular clinical condition, an optimal rate of overall antibiotic utilization is difficult to establish. Nonetheless, clinical evidence and expert opinion strongly suggest that prescribing is most appropriate in countries at the lower end of the spectrum and that the difference between low- and high-utilizing countries to a considerable extent represents potentially unnecessary and avoidable prescriptions.

Recently, the United States adopted a similar (albeit more limited) approach to track and report antibiotic utilization. In 2005, the National Committee for Quality Assurance (NCQA) developed and implemented a new measure to compare overall rates of antibiotic utilization between US health plans. By evaluating variation in antibiotic use within a country, these data can improve understanding of factors that contribute to variation in overall antibiotic utilization without the confounding effects of different countries’ healthcare systems. In addition, implementation of this measure in an established program to compare and improve healthcare quality can directly facilitate efforts to improve quality of care by identifying high-prescribing plans and stimulating them to investigate the factors that contribute to potentially excessive antibiotic use.

In this study, we analyze antibiotic utilization rates among nonelderly members of commercial health maintenance organization (HMO) and point-of-service (POS) health plans in the United States participating in the NCQA’s Healthcare Effectiveness Data and Information Set (HEDIS) program. We quantified the degree of variation in antibiotic utilization rates across health plans, controlling for member characteristics, and estimated the cost implications of this variation.

Data for this study were collected by commercial HMO and POS health plans in the United States that participate in NCQA’s HEDIS program.16 HEDIS is a voluntary program that collects data from health plans on various domains of effectiveness and utilization. These data are used to benchmark and compare the quality of care across health plans. Participating health plans account for more than 85% of individuals enrolled in US commercial plans. We report data from 2005, the first year plans reported data on overall antibiotic utilization for a new HEDIS measure.

We received a core set of data from 248 commercial health plans, representing 83% of commercial plans in established parts of the HEDIS program. The antibiotic utilization measure was a first-year measure. All first-year measures in HEDIS are not publicly reported in order to evaluate the measure, and, as is common, some plans choose not to report during this first year of data collection. We were able to obtain information on approximately 80% of the plans not contributing core data. Plans that did not report data had fewer members than those which did report (55% of nonreporting plans had fewer than 10,000 members vs 7% of reporting plans; 

Among 248 plans contributing core data, we excluded 4 health plans that did not provide basic data necessary for our analyses, 14 health plans whose rate of overall antibiotic utilization was extreme enough to suggest discrepancies in the way these data were collected or reported, and 1 plan that did not use an HMO or POS model of care. Our final analytic dataset thus comprised 229 health plans.

Our main outcome variable was each plan’s rate of antibiotic utilization per member per year (PMPY), as assessed by pharmacy claims billed. In calculating this rate, we counted each antibiotic prescription fill equally, regardless of drug dose or duration. Our analyses focused on people age 0 to 64 years, as most people age 65 years and older are enrolled in government-sponsored Medicare plans and those remaining in commercial plans may not be representative of the larger population of elders.

Each plan reported antibiotic utilization data for enrollees stratified into age-sex groups. To calculate standardized rates of antibiotic utilization, for each plan we multiplied the PMPY utilization rates for each of these age-sex strata by the proportion of the overall study cohort within each of these strata. Next, we summed each of these weighted rates to create an overall antibiotic utilization rate for all plan members age 0 to 64 years. This approach yielded a PMPY antibiotic utilization rate that adjusted for the age and sex distribution of each plan’s members. This standardized rate yielded figures similar to those of a crude rate that did not adjust for differences in the age and sex distribution of plans (see the 

Next, we evaluated variation in use of specific types of antibiotics. Health plans provided utilization information for each of 15 classes of antibiotics, following a categorization scheme defined by a  multistakeholder expert panel that worked with NCQA to arrive at a consensus definition that was both clinically relevant and feasible to implement. This panel further grouped each of the antibiotic classes into 2 categories: “antibiotics of concern,” agents generally considered to have broad-spectrum activity; and all other agents (“other antibiotics”), generally considered to have a narrow spectrum of antimicrobial activity (see Table 2). For purposes of clarity, we will use the terms “broad spectrum” and “narrow spectrum” to refer to these categories. We identified potential discrepancies in class-level utilization data for 20 plans. These discrepancies were most likely due to rounding errors, as plan-level data on utilization rates were provided to only 2 decimal places (eg, if the reported utilization rate for an antibiotic class was 0.04 antibiotic fills PMPY, the actual

rate could be anywhere from 0.0351 to 0.0449, an error rate of up to ±12%). We excluded these 20 plans from our analyses of antibiotic classes, plus an additional 3 plans that did not report data necessary to complete these analyses, leaving 206 plans with usable data. Because these discrepancies could affect estimates of drug expenditures, we also excluded these plans from our cost analyses, described below.

Next, we evaluated plan characteristics associated with the overall rate of antibiotic utilization. First, we conducted bivariate analyses using linear regression, where the outcome of interest was the plan’s age-sex adjusted rate of antibiotic utilization and the predictor variables were plan characteristics available from HEDIS that we hypothesized might be associated with antibiotic utilization. Next, we entered all variables into a multivariable linear regression model, on which further diagnostic testing confirmed the adequacy of model fit.

To estimate the cost of antibiotic utilization for each plan, we used several steps (see the eAppendix for details). Health plans provided data on utilization of antibiotic classes but not specific drugs within those classes. To compensate, we used data from the 2004 and 2005 National Ambulatory and National Hospital Ambulatory Medical Care Surveys to estimate the frequency of use of specific antibiotics within each antibiotic class. Next, we estimated the typical course of therapy for each antibiotic, including the most commonly used formulation, dose, dosing frequency, and duration of therapy, repeating this process separately for persons age 0 to 3 years, 4 to 9 years, and 10 to 64 years. For each antibiotic, we used data from the 2005 Red Book to assess the median price per pill (or other formulation), averaging across all manufacturers of a given drug and all bottle sizes.17 For multisource (generically available) drugs, we evaluated only the price for the generic versions. Finally, we calculated the estimated cost of a typical course of therapy for each antibiotic and then integrated this cost at the level of the antibiotic class and ultimately at the level of the health plan.

This research was approved by the Committee on Human Research at the University of California, San Francisco and the Research and Development Committee at the San Francisco VA Medical Center.

Characteristics of the 229 health plans are shown in 

. Plans were distributed across the United States and varied widely in the size of their enrolled population. Together, these plans accounted for 42.9 million enrollees age 0 to 64 years. Children age 0 to 17 years comprised 27% of the overall member population.

Across all plans, the mean rate of antibiotic utilization was 0.88 prescription fills PMPY. Rates of antibiotic utilization varied widely among plans (

), with a standard deviation (SD) of 0.15 fills PMPY, and 1.7-fold variation in rates of antibiotic utilization between the plans at the 5th and 95th percentiles of utilization (0.64 and 1.08 prescription fills PMPY, respectively). For a plan with 250,000 members, 89,000 fewer antibiotics would be dispensed each year if plan doctors prescribed antibiotics at the 10th percentile rate compared with the 90th percentile rate. If a similarly sized plan were to reduce its utilization from the 75th to 25th percentile, 47,000 fewer antibiotic prescriptions would be filled.

Variation in rates of antibiotic utilization persisted among age and sex groups (

). Mean rates of antibiotic utilization differed across age-sex groups, and there was considerable variation in utilization rates within each group (SDs for each group ranged from 0.10 fills PMPY for males age 18-34 years to 0.24 fills PMPY for males age 0-9 years).

 shows variation in the distribution of types of antibiotics used among health plans. Across plans, a mean of 47% of antibiotics dispensed were broad-spectrum agents (“antibiotics of concern”). The proportion of broad-spectrum antibiotics varied among plans, ranging from 34% of antibiotic prescription fills for plans at the 5th percentile of broad-spectrum utilization to 59% for plans at the 95th percentile (SD ± 8%). There was little correlation between plans’ overall rate of antibiotic utilization and the proportion of antibiotics that were broad-spectrum agents (Pearson 

We next evaluated factors associated with a plan’s rate of overall antibiotic utilization. On bivariate analyses, plans with “excellent” accreditation status (the highest achievable rank) used more antibiotics than plans with “commendable” status (the second-highest rank), and plans with a smaller percentage of physicians who were board certified used more antibiotics than plans with a larger percentage of physicians who were board certified (

). In addition, national region was strongly associated with rates of antibiotic utilization. Compared with plans in the West, plans in the South used a mean of 0.16 more antibiotic prescription fills PMPY, plans in the Midwest used 0.12 more fills PMPY, and plans in the Northeast used 0.07 more fills PMPY. As the mean rate of antibiotic utilization in the West was 0.78 fills PMPY, these differences correspond to a 21% higher rate of antibiotic utilization in the South, a 15% higher rate in the Midwest, and a 9% higher rate in the Northeast (see Table 3 footnote f for a further explanation of how these percentages were derived). On multivariable analysis, geographic region remained the most notable predictor of antibiotic utilization rates. Thirteen percent of the total variance in antibiotic utilization rates was explained by national region, while multivariable adjustment for all variables shown in Table 3 explained 26% of the total variance.

Finally, we estimated the cost implications of variation in antibiotic utilization. Across 206 plans with 38 million enrollees available for this analysis, total antibiotic expenditures totaled $1.90 billion, corresponding to a mean antibiotic cost of $49 PMPY. The SD of PMPY antibiotic costs was $9, and costs ranged from $34 PMPY for plans at the 5th percentile of costs to $63 PMPY for plans at the 95th percentile of costs. If all plans above the 25th percentile of antibiotic costs PMPY reduced their costs to this level, the mean (SD) drug cost savings would be $10 (±$6) PMPY. Across the 38 million enrollees of the plans we studied, that corresponds to $305 million in savings per year. If plans reduced their utilization to the 10th percentile, the drug cost savings would be greater, with mean (SD) savings of $14 (±$7) PMPY and total savings of $492 million per year.

In this study of 229 commercial health plans in the United States, we found substantial variation in the rate of antibiotic utilization, with high-utilizing plans dispensing nearly 70% more antibiotics per capita than low-utilizing plans. There was similarly large variation in the proportion of broad-spectrum antibiotics used by health plans, ranging from 34% of all antibiotics in plans at the lower end of the spectrum to 59% in plans at the higher end. Geographic region was strongly associated with health plans’ rate of antibiotic utilization: after controlling for other factors, plans in the southern United States used 0.16 more antibiotics PMPY than plans in the western United States, a difference larger than 1 SD of utilization rates nationally.

This variation in antibiotic utilization has substantial implications for health and healthcare costs in the United States. Overprescribing of antibiotics results in unnecessary drug side effects and promotes population-level resistance to antibiotics, which—consistent with our data&mdash;is greater in southern than in western states.

 Although resistance to any antibiotic is promoted by high-volume use, excessive prescribing of broad-spectrum agents is of particular concern because this overprescribing promotes resistance to agents that are commonly used to treat serious or complicated infections for which the consequences of treatment failure can be severe.

 Finally, unnecessary antibiotic use has important cost implications.

 Of an estimated $1.9 billion spent on antibiotics per year by health plans in our study cohort, direct drug costs could be reduced by 16% if health plans reduced their antibiotic utilization to the current 25th percentile of costs and by 26% if plans reduced their utilization to the current 10th percentile of costs. As with European antibiotic studies and other investigations into variation in the delivery of healthcare, it is difficult to determine the “correct” amount of antibiotic utilization. Nonetheless, understanding variation can provide valuable information to health plans and policymakers seeking to improve care quality and reduce unnecessary spending.

 Wennberg and colleagues were among the first to highlight the importance of reporting variation, pointing to large geographic differences within the United States in the use of healthcare interventions such as surgical procedures.

 For interventions for which an optimal rate of utilization is known, these data can stimulate benchmarking, allowing health plans and institutions to set achievable goals for their practice and to monitor progress toward these goals.

 Where an optimal rate of utilization is not known, reports of variation can help health plans and institutions understand their performance relative to that of their peers. This information can prompt plans with relatively high rates of utilization to examine why their delivery of services varies substantially from the norm, to determine whether this  represents a remediable problem in quality, and if so to investigate how to improve their care quality.

Our results suggest that overall rates of antibiotic utilization are well suited to a variation-centered approach. A number of commercial health plans are achieving far lower rates of antibiotic utilization than others. Of course, not all health plans are comparable; for example, plans whose enrollees have greater illness severity and barriers to accessing care may have legitimate reasons for prescribing more antibiotics than others.

 Thus, a high observed rate of antibiotic utilization should not be an end unto itself, but should prompt in-depth analysis to identify nonclinical factors that promote increased antibiotic use. Such analyses can be further guided by evaluation of prescribing rates within age and sex strata to determine whether certain patient subgroups have disproportionately high antibiotic utilization rates relative to a health plan’s peers. Where appropriate, local initiatives can be crafted to address the factors that promote increased antibiotic use, preferably borrowing from previous research to use active forms of clinician education and other methods proven to reduce unnecessary antibiotic prescribing.

 The NCQA took a variation-centered approach in creating this measure for the HEDIS program, in which overall antibiotic utilization has been included in the “use of services” domain, which tracks utilization and can be useful for comparison and identification of achievable goals without any specific performance targets.

Although we cannot definitively establish a “correct” rate of antibiotic utilization in commercial health plans, a variety of data suggest opportunities for improvement in the United States. Numerous studies have documented substantial overuse of antibiotics in the United States, and cross-national comparisons have found that Americans receive approximately 20% more antibiotics per capita than Europeans, with only 3 of 27 European countries having higher rates of antibiotic dispensing than the United States.

 Patients in the Netherlands, the lowest-prescribing European country, receive 60% fewer antibiotics than patients in the United States.10,29 Reducing antibiotic utilization is a complex endeavor and requires attention not only to clinical efficacy but also to patient satisfaction and downstream utilization of health services. Nonetheless, controlled trials to reduce inappropriate antibiotic use in the outpatient setting found no increase in subsequent health services utilization (eg, office visits, telephone calls) and little to no adverse impacts on patient satisfaction.

Our study has several limitations. The data collected were from the first year in which this measure was implemented by HEDIS. Although plans were given detailed instructions for complying with this measure, it is possible that certain plans had not perfected their data-collection and reporting processes. In addition, data are not publicly reported for the first year of any HEDIS measure. As is common, some plans chose not to report their results during this first year, and it is difficult to know whether these plans did not participate because they expected their performance to be poor or because of other factors. (For example, more than half of nonreporting plans had fewer than 10,000 members, compared with 7% of reporting plans, suggesting the possibility that smaller plans might have had fewer resources available to put toward a first-year, non— publicly reported measure.) Nonetheless, we did receive data from 83% of plans, and all HEDIS data are audited, suggesting that our results are representative of the target population.

Other characteristics of our methods merit consideration in interpreting this study’s findings. First, the manner in which data were reported by health plans may result in slight imprecision in our calculation of overall antibiotic utilization rates. However, this imprecision is likely to be small in relation to the large variation in rates among plans. Second, our estimates of drug cost data were based on extrapolations from another national data source (National Ambulatory Medical Care Survey), combined with utilization data from the study plans, and do not precisely reflect the actual drug costs incurred by plans (which may be influenced by negotiated deals with drug suppliers, etc). Thus, our drug cost analyses should be interpreted as reasonable estimates rather than a precise accounting of real drug costs, and they do not account for downstream cost expenditures or savings associated with reduced antibiotic use. Third, we did not have access to clinical data such as comorbid conditions; thus, we could not control for interplan differences in members’ health beyond those that are correlated with patient age and sex. Finally, we collected data only on HMOs and POS plans participating in the HEDIS program. Although the strong majority of eligible commercial health plans participate in the HEDIS program, we cannot know the  generalizability of our results to plans not participating in HEDIS or to persons with public insurance (eg, Medicaid), other forms of commercial insurance, or no insurance at all.

It is difficult to improve healthcare quality unless it can be measured. The substantial unexplained variation in antibiotic utilization across US health plans suggests opportunities to improve the quality and costs of antibiotic prescribing. We believe the NCQA antibiotic utilization measure should stimulate additional efforts to understand and improve antibiotic utilization at the health plan level—particularly for health plans in the higher range of antibiotic use. If successful, these efforts are likely to improve quality of care and to generate meaningful cost savings from reduced antibiotic costs&mdash;a winwin scenario for patients, payers, and the public health.

We thank Min Gayles, formerly at the National Committee for Quality Assurance, for her assistance obtaining the data and facilitating this research, and Saunak Sen, PhD, and John Boscardin, PhD, for their assistance with data analysis.

: From the Division of Geriatrics (MAS), San Francisco Veterans Affairs Medical Center, San Francisco, CA; Department of Clinical Pharmacy (KYY) and Department of Medicine (MAS, JHM, RG), University of California, San Francisco, CA; and the National Committee for Quality Assurance (SCB), Washington, DC.

: This work was supported by Career Development Transition Award 01-013 from the VA Health Services Research and Development Service (Dr Steinman), by grant K23-AG030999 from the National Institute on Aging and the American Federation for Aging Research (Dr Steinman), by the Agency for Healthcare Research and Quality (NCQA) Translating Research into Practice Program (R01 HS13915; Dr Gonzales and Ms Maselli) and by grant KL2 RR024130 from the National Center for Research Resources (Dr Yang). We submitted an advance copy of this paper to NCQA for review and comments, but that institution had no control over the analyses and interpretation of data or over the decision to publish this work. The funding sources had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; or preparation, review, or approval of the manuscript. The content is solely the responsibility of the authors and does not necessarily represent the official view of the Department of Veterans Affairs, the National Center for Research Resources, or the National Institutes of Health.

: The authors (MAS, KYY, SCB, JHM, and RG) 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 (MAS, RG); acquisition of data (KYY, SCB, JHM); analysis and interpretation of data (MAS, KYY, SCB, JHM, RG); drafting of the manuscript (MAS, RG); critical revision of the manuscript for important intellectual content (MAS, KYY, SCB, RG); statistical analysis (MAS, JHM); obtaining funding (MAS); and administrative, technical, or logistic support (MAS).

: Michael A. Steinman, MD, San Francisco VA Medical Ctr, 4150 Clement St, Box 181-G, San Francisco, CA 94121. E-mail: mike.steinman@ucsf.edu.

. Antibiotic treatment of adults with sore throat by community primary care physicians: a national survey, 1989-1999. 

2. Steinman MA, Gonzales R, Linder JA, Landefeld CS

. Changing use of antibiotics in community-based outpatient practice, 1991-1999. 

3. Gonzales R, Malone DC, Maselli JH, Sande MA

. Excessive antibiotic use for acute respiratory infections in the United States. 

4. Shehab N, Patel PR, Srinivasan A, Budnitz DS

. Emergency department visits for antibiotic-associated adverse events. 

5. Seppälä H, Klaukka T, Vuopio-Varkila J, et al

. The effect of changes in the consumption of macrolide antibiotics on erythromycin resistance in group A streptococci in Finland. Finnish Study Group for Antimicrobial Resistance. 

Articles

CLINICAL

Rates of outpatient antibiotic prescribing vary widely between US commercial health plans. High-utilizing health plans may improve quality and lower costs by reducing unnecessary antibiotic use.

Variation in Outpatient Antibiotic Prescribing in the United States

: To evaluate the effect on adherence and medical care expenditures of a pharmacy benefit change that included free generic drugs and higher copayments for brand-name drugs. 

: Quasi-experimental pre—post study of patients with ischemic heart disease (1286 control and 555 intervention) and patients with diabetes mellitus (4089 control and 1846 intervention).

: Medical and pharmacy claims data were analyzed for  continuously enrolled members from January 1, 2005, through December 31, 2008. A generalized linear model was used to predict costs as adherence changed. 

: The rate of switching from brand-name drugs to generic drugs in the intervention group was not statistically different from that in the control group. The net change in adherence was higher only for the intervention group patients taking statins who switched to generic drugs, a 6.2% increase compared with an 8.5% decrease in the control group. The estimate of medical cost savings attributable to this benefit change was significant for only the metformin class of diabetes drugs. Improved adherence independent of this benefit change was estimated to reduce all-cause medical costs for patients taking sulfonylureas, metformin, and thiazolidinediones.

: Altering copayments for pharmaceuticals may affect the rate of conversion to generic drugs but is unlikely in and of itself to result in complete conversion. However, increasing adherence can result in net savings for specific diabetic drug classes, as savings from all-cause medical costs offset the increase in pharmacy costs.

A pharmacy benefit design change increasing copayments on brand-name drugs and simultaneously introducing a zero copayment on generic drugs had little influence on patients’ switching to generic drugs or on improving adherence to medication regimens.

Altering copayments for generic drugs and brand-name drugs may affect the rate of conversion to generic drugs but is unlikely in and of itself to result in complete conversion to generic drugs.

Small increases and decreases in pharmacy copayments have little effect on adherence or medical costs.

Improving adherence lowered all-cause medical costs for 3 diabetes drug classes but raised pharmacy costs for all 5 drug classes studied.

The increasing cost of medical care in the United States is an ongoing concern, consuming more than 16.2% of the nation’s 2006 gross national product.

 Government, payers, business leaders, consumers, and healthcare providers are seeking to control healthcare costs without sacrificing quality of care. One such effort is improving the management of chronic conditions. Caring for chronic conditions in the United States consumes 78% of all healthcare dollars,

 meaning that such an effort has the potential for significant effect.

Payers and providers have implemented several approaches to improve the management of chronic conditions. Disease management programs emphasize one-on-one contact with health coaches and nurses.

 Physician practice models that incorporate aspects of a medical home

 focus on increased involvement of healthcare providers in a team approach. Using insurance benefit design to lower barriers to therapies that effectively manage chronic conditions is the focus of this analysis. Value-based plan design promotes the use of care when the benefits exceed the costs and discourages the use of care when the costs exceed the benefits.

Value-based plan design that focuses on reducing out-of-pocket costs for prescription drugs taken by patients with chronic diseases is premised on a relationship between improved medication adherence and better condition management and between better condition management and lower costs. Most research has focused on the relationship between prescription drug out-of-pocket costs and adherence or on the relationship between adherence and medical cost or utilization, although some studies

 considered all 3 factors. Studies of cost and adherence have demonstrated the following mixed results: increasing pharmacy cost sharing had no effect on adherence,

 increasing pharmacy cost sharing decreased adherence,

 and decreasing pharmacy cost sharing increased adherence. 

 This apparent contradiction may be explained by the conditions studied, the adherence measures used, or the sample size.

Studies of the relationship between adherence and health outcomes or costs and utilization have had more conclusive findings. For example, lower adherence has been associated with worse outcomes, including mortality, in populations with diabetes mellitus or acute myocardial infarction.

 A relationship between increased adherence and fewer emergency department visits and hospitalizations has been observed.6,7,18-20 While some studies

 have shown overall medical cost savings as a result of adherence, other studies

 have not in part because the increase in pharmacy costs exceeded the medical cost savings.

This study relies on a natural experiment created when Blue Cross Blue Shield of Minnesota introduced a free generic drug benefit in July 2006 to groups with fewer than 51 employees. The new benefit eliminated the $5 copayment for generic drugs and increased by $5 the copayment for brandname and nonformulary drugs to $35 and $50, respectively. We hypothesized that adherence would increase as member costs for generic drugs decreased and that adherence would decrease as member costs for brand-name drugs increased. We also hypothesized that improved adherence would lead to better condition management and to lower medical costs. This study extends existing research by quantifying the net effect on medical and pharmacy costs of a net change in adherence due to a simultaneous increase and decrease in pharmacy copayments.

The study used a quasi-experimental pre—post design. The preperiod consisted of 24 months from January 1, 2005, through December 31, 2006. Because the benefit change had a staggered rollout beginning June 1, 2006, the intervention date was defined as January 1, 2007, halfway through the implementation year. The postperiod consisted of 24 months from January 1, 2007, to December 31, 2008.

The intervention group included 44,155 members continuously enrolled in products offered to groups with 2 to 50 employees in the study period. Members enrolled in highdeductible plans paired with federally qualified health savings accounts were not eligible for the free generic drug benefit and were excluded. The control group included 79,853 members continuously enrolled in products purchased by 175 groups with 51 or more employees. The control group experienced no change in pharmacy copayments during the study period. The control and intervention groups had access to the same provider network; consequently, discounts and fee schedules are comparable between the 2 groups.

To evaluate the effect of the copayment changes on adherence and medical costs, the analysis was narrowed to patients who had 2 chronic conditions (ischemic heart disease or diabetes mellitus) easily identified in claims data and for whom the prescribed drugs were taken specifically to treat that condition. Patients were identified in claims data using episode treatment groups (ETGs), which use proprietary algorithms to group medical and pharmacy claims into clinically homogeneous units based on complete episodes of care (Episode Treatment Groups [computer program]. Version 5.0. Symmetry Health Data Systems Inc, Phoenix, AZ). The ischemic heart disease subpopulation (1286 control and 555 intervention) included members with a cardiac episode (ETGs 251-259 and 265) and pharmacy claims indicating that a statin prescription was filled. The diabetes subpopulation (4089 control and 1846 intervention) included members with diabetes-related treatment (ETGs 27-30 and 222-224) and pharmacy claims indicating that a prescription was filled for sulfonylureas, metformin, thiazolidinediones, or insulin. These prescription drugs and their generic equivalents were identified using a proprietary drug database (Master Drug Data Base, Version 2.5. Medi-Span, Indianapolis, IN) that uses national drug codes and proprietary generic product identifiers to indicate drug groups and classes and generic equivalents.

At the time of analysis, no generic equivalents were available for insulin or for thiazolidinediones. These 2 drugs were included to assess the effect of the copayment increase on brand-name drugs. We assumed that once medication therapy commenced, it was continued to manage the condition.

 Using a combination of ETGs and pharmacy claims for case finding minimized the chance of including rule-out diagnoses in the analysis or of including patients taking the specified medications for conditions other than those in the analysis.

We estimated net changes in medication adherence in the preperiod and in the postperiod. This modeling estimated the percentage of patients whose adherence changed after the change in copayments. We then used generalized linear models with a gamma distribution and a log link to estimate the effect of the net change in adherence on the following 3 measures of cost: all-cause medical, condition-specific medical, and condition-specific pharmacy.

. We used the medication possession ratio (MPR) to measure adherence.

 The MPR was computed by dividing the number of drug days supplied by the number of days the patient had the health condition of interest. To accommodate the pre—post design, if the preperiod drug days supplied was greater than the days with the condition, excess days were added to the drug days in the postperiod. In both periods, the MPR was capped at 100%. We then categorized those with an MPR of at least 80% as being adherent and those with an MPR of less than 80% as being nonadherent. The variable of interest identified patients nonadherent in the preperiod but adherent in the postperiod and patients adherent in the preperiod but nonadherent in the postperiod. The mean of this variable indicated the net change in adherence from the preperiod to the postperiod.

The net changes in adherence from the preperiod to the postperiod were compared among patients who only took generics, patients who switched to generics, and patients who only took brand-name drugs. The copayment differentials varied among these groups. Patients who only took generics and patients who switched to generics in the preperiod had a $5 savings for each prescription filled in the postperiod. Patients who switched to generics in the postperiod had savings of $35 for formulary drugs and $50 for nonformulary drugs. Patients who took only brand-name drugs had an increase of $5 for each copayment in the postperiod. Few patients switched from generic to brand or from brand to generic and back to brand; these patients were excluded from the analysis. 

 Tests determined if the differences between the means were statistically significant. A regression model was not used because the value of the available control variables (eg, sex) did not change between the preperiod and the postperiod at the patient level.

. We estimated 3 measures of cost. All-cause medical costs included all medical claims regardless of the underlying condition. Conditionspecific claims for ischemic heart disease were identified using 

International Classification of Diseases, Ninth Revision, Clinical Modification 

) codes 410 through 414 or procedure codes for coronary artery bypass graft or percutaneous transluminal coronary angioplasty. Condition-specific medical costs for diabetes were identified using 

 code 250. Condition-specific pharmacy costs were identified by drug type, including statins for the ischemic heart disease subpopulation and sulfonylureas, metformin, thiazolidinediones and insulin for the diabetes subpopulation. Each of these measures represented the costs paid by the plan and the member in the preperiod and in the postperiod. The medical costs were adjusted to 2008 US dollars using the medical cost portion of the Consumer Price Index, and the pharmacy costs were adjusted to 2008 US dollars using the prescription drug cost portion of the Consumer Price Index.

The cost model estimated how the net change in adherence affected all-cause medical costs, condition-specific medical costs, and condition-specific pharmacy costs in the postperiod. The 2 variables of interest were the net change in adherence and the interaction of that change and membership in the intervention group. Additional explanatory variables included sex, age (in years), the mean annual household income within a census block group (<$25,000, $25,000 to <$50,000, and ≥$50,000), duration of the condition, residence in a metropolitan area, and the probability of having earned a college degree and of being of white or Hispanic race/ethnicity. A log transformation of the condition-specific medical costs in the preperiod was included to be consistent with the log transformation of the dependent cost variables. The values for income, race/ethnicity, and education were imputed using census data and an age-stratified geocoding and surname analysis.

We included several health-related variables specific to the conditions studied that indicated the presence of the following cardiac conditions: acute myocardial infarction, percutaneous transluminal coronary angioplasty, coronary artery bypass graft, chronic ischemic heart disease, other ischemic heart disease, atherosclerosis, and angina. We also included variables indicating the presence of comorbidities, specifically ischemic heart disease (defined as in the ischemic heart disease subpopulation), hypertension (using 

 codes 401-403 and 405), and diabetes (defined as in the diabetes subpopulation). These variables were constructed using claims data from the preperiod.

Characteristics of the study subpopulations are given in 

. Compared with the control group, the intervention group was younger, was less likely to contain female or Hispanic members, was more likely to contain white members, had lower health risks and fewer comorbidities, filled more prescriptions, and was more likely to have low income.

 summarizes the switching patterns from brandname to generic drugs. Generic use was much higher in the diabetes subpopulation than in the ischemic heart disease subpopulation. Although a higher percentage of the diabetes subpopulation switched to generics, most of that switching occurred before the benefit change. However, because generic use was higher in the diabetes subpopulation, the absolute number of members switching was much fewer than in the ischemic heart disease subpopulation. In both subpopulations, there was little difference between the control group and the intervention group. The only statistically significant difference between the control group and the intervention group was the percentage of patients who continued to take brand-name metformin, with 3.0% of the control group continuing compared with 0.4% of the intervention group continuing.

Results of the adherence analysis are given in 

. For each drug type, the net change in adherence was tested between the control group and the intervention group for the total cohort and for patients who only took generic drugs, patients who switched to generic drugs, and patients who only took brandname drugs. Most subgroups had increases in adherence from the preperiod to the postperiod. The only statistically significant difference between the control group and the intervention group occurred in the subgroup taking statins who switched from brand-name drugs to generic drugs. This 14.7% difference reflected a 6.2% net increase in adherence among patients in the intervention group compared with an 8.5% net decrease in adherence among patients in the control group.

The model estimates for the effect of the net change in adherence on costs are given in 

. Summarized are the coefficients from the 3 regression models for the 2 variables of interest (the net change in adherence and a term capturing the interaction of that change and membership in the intervention group). The interaction term captures the effect of the net adherence change for the intervention group, isolating the effect of the benefit change. The change in adherence term captures the effect of the net adherence change unrelated to the intervention. The intervention was the benefit change that introduced a zero copayment on generic drugs and simultaneously increased copayments on brand-name drugs.

The change in adherence coefficients were negative and statistically significant for 3 diabetes drug classes (sulfonylureas, metformin, and thiazolidinediones), indicating that all-cause medical costs decreased as the net change in adherence increased. A similar result was seen for conditionspecific medical costs in the thiazolidinediones subgroup. As expected, the change in adherence coefficients were positive in the condition-specific pharmacy cost model, indicating that pharmacy costs increased as the net change in adherence increased. Although the magnitude of the coefficients in the condition-specific pharmacy cost model was greater than that of the coefficients in the medical cost models, the 3 models suggested overall savings, as pharmacy costs made up only a small portion of total costs.

The interaction term, which isolated the effect of the intervention, showed a decrease in all-cause medical costs and condition-specific medical costs only for the metformin subgroup. The intervention had no statistical effect on condition-specific pharmacy costs. These coefficients represent elasticities, the percentage change in costs due to a 1% change in the net change in adherence. For example, a 10% increase in the net change in adherence for the metformin intervention subgroup would result in an estimated 5.3% decrease in all-cause medical costs, a 4.2% decrease in condition-specific medical costs, and no change in condition-specific pharmacy costs.

The intent of the free generic drug benefit was to encourage switching from brand-name drugs to generic drugs. Additional expectations were that reduced out-of-pocket costs would result in improved adherence,

 in turn resulting in lower medical costs.

 This study found that lowering the cost of generic drugs and raising the price of brand-name drugs did not result in a complete transition to generic drugs for patients with ischemic heart disease or diabetes. The study also found that the $5 differential did not result in more switching to generic drugs compared with a control group that had no change in pharmacy copayment amount.

These results suggest that altering copayments for generics and brands may affect the rate of conversion to generics but is unlikely to result in complete conversion to generics. Additional initiatives to consider include the following: (1) outreach and education to members and physicians outlining alternatives that are available to lower costs without decreasing adherence, (2) a mandatory generics program, or (3) formularies that incorporate step therapy and evidence-based approaches.

While the study showed that the net change in adherence was positive for the intervention group and the control group, the change for the intervention group was statistically different from that for the control group only for patients taking statins. This indicates that there was no difference in the net change in adherence for groups facing small increases, decreases, or no change in copayments. This result is similar to earlier findings of no change in adherence even after an increase in copayment amounts.

With respect to medical and pharmacy costs, the effect of the intervention was limited to the subgroup taking metformin. The finding that increasing the net change in adherence would result in all-cause medical cost savings for 3 of 5 drug classes studied is similar to results of other studies.

 Factors other than a change in copayments that could increase the net change in adherence include disease management programs, provider incentives that reward quality performance for chronic care, and an increase in member income. Unfortunately, data limitations prevented our controlling for these factors in this study. Furthermore, the results suggest that medical cost savings may offset increased pharmacy costs due to that increased adherence. For the other 2 drug classes studied, the finding that adherence did not significantly affect costs may be attributable in part to a postperiod that was too short to capture medical cost savings. In general, these estimates indicate that greater adherence increased condition-specific pharmacy costs. However, for the intervention herein, these estimates indicated no change in condition-specific pharmacy costs, which could be due to the design of increasing some costs and decreasing other costs simultaneously.

There are several limitations to this study. We focused on drug adherence for only 2 chronic conditions, for which the treatment drugs constitute a small subset of those covered by the free generic drug benefit. Therefore, the findings cannot be generalized to all conditions and drugs covered by the free generic drug policy. Although claims data are readily available, their use to study healthcare outcomes, treatment patterns, utilization, and costs is subject to the limitations associated with all observational administrative claims database analyses. In addition, data on race/ethnicity, education, and income are not routinely collected and were imputed. This imputation process does not capture changes in income and education that occurred during the study time period. The absence of additional unobservable characteristics from the claims data such as physician prescribing behavior and other changes in family situations could have affected the estimates.

This study examined how changing members’ out-of-pocket pharmacy costs affected adherence and medical costs. Although a decrease in the generic drug copayment to zero and a simultaneous increase in the brand-name copayment resulted in some switching from brand-name drugs to generic drugs, it was insufficient for complete transition. There was little difference in switching behavior or in the net change in adherence between the intervention group and the control group. Results of regression modeling suggested that improving adherence lowered all-cause medical costs for 3 diabetes drug classes. However, the intervention studied herein resulted in medical cost savings only for the metformin drug class.

: From Blue Cross Blue Shield of Minnesota (HAR, AHH, ARW, NAG, DWP), Eagan, MN. Dr Heaton is now with UCare, Minneapolis, MN. Dr Plocher is now a vice president with Ingenix Consulting in Eden Prairie, MN.

: This study was funded by Blue Cross Blue Shield of Minnesota, and all authors were employees of that organization at the time of the study.

: The authors (HAR, AHH, ARW, NAG, DWP) are or were previously employed by the company whose product is the subject of this analysis.

: Concept and design (HAR, AHH, ARW, NAG); acquisition of data (HAR); analysis and interpretation of data (HAR, AHH, DWP); drafting of the manuscript (HAR, AHH, ARW, NAG); critical revision of the manuscript for important intellectual content (AHH, ARW, NAG, DWP); statistical analysis (HAR); and supervision (DWP).

: Holly A. Rodin, PhD, Blue Cross Blue Shield of Minnesota, 3400 Yankee Dr, R259, Eagan, MN 55122. E-mail: holly_rodin@bluecrossmn.com.

MANAGERIAL

Small increases and decreases in pharmacy copayments have little influence on patients’ switching to generic drugs or on improving adherence to medication regimens.

Plan Designs That Encourage the Use of Generic Drugs Over Brand-Name Drugs: An   Analysis of a Free Generic Benefit 

: During the past decade, experts have devoted substantial efforts to quality improvement for managed care. Although a handful of studies have examined the effect of quality on enrollment, few have systematically investigated the association between managed care quality and plan disenrollment, especially among lowincome populations.

: To examine whether higher-quality measures in managed care plans are associated with lower disenrollment from the State Children’s Health Insurance Program (SCHIP) in New York State.

: Observational study of managed care plan disenrollment for a New York statewide cohort of 2206 new SCHIP enrollees.

: Managed care quality was measured by 7 Consumer Assessment of Health Plans Survey (CAHPS) scores and 3 Healthcare Effectiveness Data and Information Set (HEDIS) scores, obtained from the 2002 New York State Managed Care Plan Performance Report. Disenrollment was defined as being disenrolled from an SCHIP plan for 2 or more consecutive months based on the New York SCHIP universal billing files.

: Nearly 40% of children were disenrolled during the study period. No overall effects of plan quality on disenrollment were detected, but plans with higher scores in “preventive care visits” had a significantly lower disenrollment rate. The disenrollment rate in the eligibility recertification period was 3.2 percentage points higher than that in other time periods.

: Disenrollment was not associated with overall managed care plan quality as measured by CAHPS and HEDIS, suggesting that further study is warranted to determine optimal strategies for enhancing managed care quality in the SCHIP population.

The association between managed care quality and plan disenrollment was examined using data from the New York State Children’s Health Insurance Program (SCHIP).

Plan disenrollment was not significantly associated with overall managed care plan quality, but plans with higher scores in “preventive care visits” had a significantly lower disenrollment rate.

The disenrollment rate in the eligibility recertification period was 3.2 percentage points higher than that in other time periods.

Further study is warranted to determine optimal strategies for enhancing managed care quality in the SCHIP population.

Managed care has been the dominant type of health insurance in the United States during the past 2 decades. As of 2008, 90% of nonelderly people receiving employer-sponsored health benefits and 17% of Medicare beneficiaries were enrolled in managed care

; managed care enrollees accounted for 64% of all Medicaid beneficiaries in 2007

; and most State Children’s Health Insurance Program (SCHIP) enrollees are in managed care.

 During the past decade, enormous efforts have been expended to develop managed care quality measures and reports. Starting in 1997, all Medicare managed care plans have been required to report the Healthcare Effectiveness Data and Information Set (HEDIS).

 The Centers for Medicare & Medicaid Services also issued rules for comprehensive Medicaid managed care quality management and external review of quality improvement programs in 2002 and 2003, respectively.

 As of 2005, pay-for-performance strategies to improve managed care quality in Medicaid programs have been adopted in several states including California, Michigan, New York, and Rhode Island.

 But limited evidence exists on the impact of managed care pay-for-performance programs or other mandatory requirements on quality improvement.

A handful of studies have examined whether consumers’ enrollment choice is associated with plan quality report cards. These studies have produced mixed results, with some studies showing a positive association between quality and enrollment and others suggesting no correlation.

 Additionally, 2 studies directly examined the effect of quality on plan choice, regardless of the sources of information. It has been demonstrated that large employers were more likely to offer their employees plans with higher quality, which was approximated by Consumer Assessment of Health Plans Survey (CAHPS) and HEDIS measures.

 A study on plan choice in an SCHIP population found that parents were more likely to choose a plan with higher CAHPS scores, but plan choice was not associated with HEDIS scores.

 However, few studies have systematically investigated the association between consumers’ disenrollment and managed care quality, especially among low-income populations. In several studies on commercial populations or social/health maintenance organization (HMO) enrollees, satisfaction was found to be negatively correlated with disenrollment or switching behavior.

 Among Medicare HMO enrollees, overall plan rating and some composite ratings of special services strongly predicted a lower disenrollment rate.

 Virtually nothing is known about the relationship between children’s disenrollment from managed care and plan quality. 

To inform the policy debate, we took advantage of a New York SCHIP evaluation to assess the level of disenrollment in relation to quality of care of managed care plans.

 In this program, managed care plans are reimbursed based on prior cost experience, and no risk-adjustment mechanism or payfor-performance program is incorporated into contracting. The premium is determined by family income, with no cost sharing for families below 160% of the federal poverty level (FPL). Families have to recertify every 12 months for continued enrollment of their children. Individuals in this program could disenroll from the current plan and enroll in a private plan (if available), switch to another SCHIP plan or Medicaid plan (if eligible), or become uninsured.

Experts have become increasingly concerned about children’s disenrollment from SCHIP.

 Prior studies have highlighted that administrative requirements during recertification can discourage continued participation of lowincome children.

 Other factors such as cost-sharing also might lead to disenrollment.

 Prior studies also have noted that individual and family characteristics such as being an only child, having better-educated parents, and being healthy were associated with disenrollment, whereas other characteristics such as age, race, and income were not consistently related to disenrollment from SCHIP.

 It also has been shown that compared with those in traditional fee-for-service Medicaid, managed care enrollees were about 70% less likely to disenroll.

Quality of care may be a factor leading to disenrollment if low quality of care decreases the desirability of the program, especially for families paying premiums for SCHIP or families with parents receiving employer-sponsored insurance. Because parents’ decisions hinge on their ability to understand quality information, which is affected by education level and experience with health plan services, we hypothesized that children in health plans with higher quality would be less likely to disenroll and that the effect of quality would be greater for parents with better education and higher family income, and for parents who have children with special healthcare needs or prior health insurance.

This study assessed whether managed care quality, approximated by CAHPS and HEDIS measures, is associated with disenrollment. Determining this relationship is important. If individuals are more likely to disenroll from low-quality plans, this would indicate that market-based policies are worth considering for quality improvement. However, null findings (ie, enrollees do not respond to plan quality) would suggest that further study is warranted to determine optimal strategies for enhancing managed care quality in the SCHIP population.

The study cohort consisted of a stratified random sample of 2644 children who were newly enrolled in New York’s SCHIP between November 2000 and February 2001. Hispanic children were oversampled to ensure the validity of subgroup analysis.

 The program used a 2-month presumptive eligibility period, meaning that every child who applied was enrolled in the program until they were found to be ineligible or 60 days, whichever was shorter. An initial interview was conducted about 4 to 6 months after enrollment to avoid interviewing temporarily enrolled children, and a second interview followed about 12 months later. Individual and family characteristics also were obtained by interview. Enrollment status was derived from the 2000-2002 New York

SCHIP universal billing files. Managed care plan quality measures as well as plan characteristics were obtained from the plan component of the survey and the 2002 New York State Managed Care Plan Performance Report.

Quality measures included 7 CAHPS measures from the performance report and 3 pediatric HEDIS measures, all in percentages. CAHPS measures included 5 composites and 2 overall ratings: provider communication, getting care quickly, problems with getting care needed, problems with services, called or wrote health plan with complaints, overall rating of personal doctor or nurse, and overall rating of health plan. Three HEDIS measures were used: preventive care visits, use of appropriate medications for children with asthma, and childhood immunization. In order to avoid collinearity between variables, the final CAHPS and HEDIS measures were the weighted averages based on the standards and guidelines for managed care organization accreditations from the National Committee on Quality Assurance.

 When CAHPS or HEDIS measures were not available for SCHIP, those of Medicaid or commercial lines of the same plan were used because quality measures were moderately correlated among different product lines within a health plan (correlation coefficient varying from 0.2 to 0.8) or between calendar years (correlation coefficient varying from 0.6 to 0.9).

 disenrollment was defined as being disenrolled from SCHIP for at least 2 consecutive months because New York State has a 30-day grace period after disenrollment and no waiting period for reenrollment.

 Disenrollments due to moving out of a county, turning 19 years old, and becoming Medicaid eligible also were included. Because these disenrollments were due to eligibility changes, we only used the related enrollment records and censored them at the time of disenrollment.

An optimal evaluation of disenrollment would model an individual’s full choice set, including the current health plan, private insurance, and being uninsured. Because information on private insurance after SCHIP disenrollment was not available, a choice set was constructed for each individual based on the Combined Statistical Areas (CSAs) from US Census 2000.

 For the purpose of this study, the state was divided into 8 areas: New York City, Long Island, other suburbs of New York City, the northeast area, the north-central area, the south-central area, the central-west area, and the west or southwest area. It was assumed that the alternative private insurance plans in an area were the same for enrollees in that area. Introducing these reference areas as fixed effects enabled us to model disenrollment decisions as a function of the current health plan quality and other covariates.

Because enrollment status was recorded by month, we used a discrete-time survival model. The logit of the probability of disenrollment was modeled as a function of plan quality and other characteristics, individual/family characteristics, time, and geographic areas. We adopted an order 1 autoregressive correlation structure based on quasi-likelihood information criteria (QIC).

 A variant of QIC was used for selecting the best subset of covariates.

 Independent variables included quality measures, plan outreach activities, child’s age, child’s race, parent education, family income, presence of any special healthcare needs,

 and prior health insurance status before enrollment. Health plan outreach activity was measured by the number of outreach staff in full-time equivalents per 1000 eligible SCHIP or Medicaid population in a county. As outreach information was not available for 8 of 29 plans, an

expectation-maximization algorithm was used to impute the missing outreach activity using other plan characteristics.

 A dummy variable to flag plans without outreach information was added to the model. Dummy variables of time were used to capture the baseline hazard, and geographic areas also were included to allow for fixed effects.

To avoid incorrect inferences for interaction effects of nonlinear models, marginal effects were calculated as derivatives for continuous variables or differences for discrete variables.

 Standard errors were based on bootstrapping with 200 replications. Sampling weights were used in the estimation so that the results could be generalized to the population. Power calculations showed that these data would have the ability to detect a 1.6 percentage point change in the probability of disenrollment for each percentage point change in average quality scores with 80% power.

We performed several secondary analyses to check the robustness of the main results. (1) In addition to the weighted averages of CAHPS and HEDIS measures, we examined individual quality measures as some aspects of quality may have had strong effects compared with others. (2) Nearly 13% of individuals did not complete the second interview, and it is unclear whether the related disenrollment was due to eligibility or not. We did a sensitivity analysis and assumed the disenrollment was not due to eligibility changes. (3) Finally, prior literature has shown that many dropouts from SCHIP were unintentional, especially because of the recertification process.

 Thus, we conducted a secondary analysis by excluding the data after the 11th month when the recertification occurred.

Descriptive Analysis The parents of 2644 children participated in the first interview, and 2290 (87%) individuals completed the second interview. In addition, 84 individuals without enrollment status in the universal billing files were excluded from the analysis. The final sample consisted of 2206 individuals enrolled in 29 managed care plans.

Among 2206 individuals, nearly 40% were younger than age 6 years, and about 44% were Hispanic children (

). The income of more than three quarters of the families was below 160% of the FPL, and only about 5% of the families were above 250% FPL. During the preceding 12 months before joining SCHIP, 40% of the enrollees had some type of health insurance, and about 1 of 6 children had at least 1 special healthcare need. The average CAHPS score was 75.

 with a standard deviation of 4.56, and the average HEDIS score was 69.46 with a standard deviation of 6.92 (

). The average number of full-time equivalents of outreach staff for SCHIP and Medicaid per 1000 eligible population was 0.24, but with tremendous variation. Twenty-two plans were nonprofit, and 13 plans had a commercial insurance business line.

Over the study period, the median observed time was 14 months, with the longest period being 19 months. Nearly 40% of children were disenrolled during this period of time. The Kaplan-Meier survival curve of the first disenrollment and the empirical disenrollment hazard are illustrated in the 

, in which 2 waves of disenrollment are evident. The first wave occurred during the first several months, when a dramatic increase in the number of disenrollees was seen in the fourth month. The second wave of disenrollment occurred around the annual recertification period, with its highest rate being 7.85%.

Association Between Disenrollment and Plan Quality. 

 presents the marginal effects (in percentage points) of various factors on disenrollment. Neither the CAHPS score nor the HEDIS score was significantly associated with disenrollment from SCHIP. We also evaluated quality scores at different cutoff values of CAHPS, and the findings were unchanged (data not shown). The interaction effects between CAHPS and other individual characteristics (except for special healthcare needs) showed negative signs, but none of these interaction effects were significant and thereby did not confirm our hypothesis that quality of care might be more closely related to disenrollment for parents who were better at assessing quality. The results for the HEDIS measures were similar to those for the CAHPS measures: with insufficient evidence to support the hypothesis that children in health plans with higher HEDIS measures were less likely to disenroll. No significant interaction effects were observed between HEDIS and having special healthcare needs, having insurance before SCHIP, income, or education.

Using data from the New York SCHIP program, this study showed that plan disenrollment was not significantly associated with managed care plan quality.

December 2009

Variation in Outpatient Antibiotic Prescribing in the United States

Plan Designs That Encourage the Use of Generic Drugs Over Brand-Name Drugs: An Analysis of a Free Generic Benefit

Managed Care Quality and Disenrollment in New York SCHIP

Can Managed Care Plans Reliably Infer the Quality of Cardiac Surgeons' Outcomes?

Analytic Models to Identify Patients at Risk for Prescription Opioid Abuse

Effect of Antihypertensive Medication Adherence Among Employees With Hypertension

A Physician-Friendly Alternative to Prior Authorization for Prescription Drugs

Pharmaceuticals From Development to Practice: Learnings From an Online Curriculum