To assess the differentiated effects of population-based disease management programs (DMPs) for type 2 diabetes on intermediary clinical outcomes in The Netherlands.

Data covering a period from 20 to 24 months between January 2008 and December 2010 were collected from 18 Dutch care groups (primary care provider networks that have bundled payment contracts for delivery of diabetes DMPs). Meta-analysis and meta-regression methods were used to conduct differentiated analyses of these programs’ effects over time on 4 clinical indicators: glycated hemoglobin, lowdensity lipoprotein, systolic blood pressure, and body mass index. Heterogeneous average results were stratified according to various patient and process characteristics to investigate whether differences in these features could explain variation in outcomes.

Between 56% and 71% of patients (N = 105,056) had valid first- and second-year measurements of the study outcomes. Although average changes in these measures over time were small, stratified analyses demonstrated that clinically relevant improvements were achieved in patients with poor first-year health values. Interactions with age, disease duration, comorbidity, and smoking status were not consistent across outcomes; nonetheless, heterogeneity in results decreased considerably when simultaneously correcting for known patient characteristics. Positive effects tended to diminish with longer length of follow-up, while greater measurement frequency was associated with improved results, especially in patients with poor health.

Our data suggest that tailored disease management, in which not only evidencebased guidelines but also patient characteristics directly determine care processes, including self-management support, has great potential to improve the cost-effectiveness of current chronic care delivery.

Correcting for heterogeneous care processes and patients resulted in differentiated insight into the effects of population-based disease management on intermediary outcomes of diabetes care.

		Patients with poorly controlled diabetes benefit more from standardized disease management than patients with better health, for whom self-management support might be equally effective.

		Positive effects of disease management diminish with longer length of follow-up; greater measurement frequency improves results, especially in patients with poor health.

		Tailored disease management, in which care content and dose are determined by patients’ health, demographics, and social characteristics, has great potential for improving cost-effectiveness in chronic care.

Rising costs, an aging population, and the recognition of severe deficiencies in traditional care have motivated the development of innovative models for chronic illness management. Examples include shared care, case management, and stepped care, but perhaps most well known internationally are disease management and the Chronic Care Model.

 Differences aside, these concepts have in common their basic assumption that better treatment today will result in better health and less expensive care in the future. Specific quality improvement efforts tend to focus on (1) reducing fragmentation between providers and settings, (2) stimulating evidence-based practice, (3) promoting active and planned follow-up, and (4) supporting patients’ self-management.

Developed in the United States, disease management programs (DMPs) have quickly spread to other countries such as Canada, the United Kingdom, and Germany, which now has some of the largest DMPs in the world.

 In the primary care—oriented Dutch health system, chronic disease management initiatives developed from the 1990s onward. Initially, uptake of these initiatives remained limited, mainly due to lack of a structured framework and fragmentary funding.

 The 2006 health insurance reform granted health insurers extended power to negotiate with care providers and, in so doing, facilitated the development of a more integrated funding method, the so-called “bundled payment system.”

 Under this system, insurers pay care groups, which are provider networks based in primary care, a single fee for the full range of outpatient services for a specific chronic condition.

 Bundled payments are seen as a way to stimulate primary care providers, predominantly general practitioners (GPs), to engage in multidisciplinary cooperation and deliver integrated,  evidence-based disease management in an ambulatory setting, therefore limiting the need for specialist care.

Although the evidence base for their impact on the quality and outcomes of care is limited, integrally financed DMPs for type 2 diabetes quickly achieved national coverage in The Netherlands. In March 2010, more than 100 groups representing approximately 80% of GPs had a bundled payment contract for diabetes care.

 In the same year, the payment system was expanded to cover the management of chronic obstructive pulmonary disease and vascular risks. Hence, the support for disease management in The Netherlands appears to be a matter more of faith than fact, as is the case in other countries as well. International studies and reviews of DMPs demonstrate highly heterogeneous results and so far have failed to answer the basic question of what works best for whom, not least because of variation in both methodology and nomenclature.

 It is, however, especially the multicomponent and populationbased nature of disease management that makes it difficult to draw unambiguous conclusions concerning effectiveness. Analyzing complex DMPs necessitates a clear framework that links expected outcomes to the characteristics of both the program and its target population, and measures effects over an adequate period of time.

As part of the DISMEVAL (Developing and Validating Disease Management Evaluation Methods for European Healthcare Systems) project, this study’s objective was to assess the effects of the Dutch diabetes DMPs on a range of intermediary clinical outcomes. To investigate heterogeneity in results across different care processes and patients, we designed the study as a population-based, multilevel meta-analysis and meta-regression. Given that experimental comparisons were not possible because of the nationwide rollout of the DMPs and the unsuitability of using historical controls,

 these methods allow for the most in-depth assessment of effectiveness. Such differentiated insight, which goes beyond the “grand means” that currently inform many health system redesigns, can support professionals and policy makers in their efforts to better meet the complex care needs of the growing and inherently diverse population of chronically ill patients.

The bundled payment system for generic diabetes disease management in The Netherlands obliges care groups to provide insurers with a specific number of performance indicators for both processes and outcomes on an annual basis.

 We retrospectively gathered individual patient data on these indicators from a convenience sample of 18 groups, which were set up between 2006 and 2009 and represent nearly all regions of The Netherlands. Nine groups were part of an experimental pilot concerning bundled payments evaluated by the National Institute for Public Health and the Environment

; to include 9 other, nonexperimental groups, 14 care groups were approached (response rate 64.3%).

Across the included groups, 106,623 patients had at least 1 registered visit during the research period, which—depending on the availability of data&mdash;was either 20 or 24 months between January 2008 and December 2010. We excluded type 1 diabetes patients (n = 1567) because they are treated primarily by specialists. Since patient data were drawn from groups’ clinical information systems, plausibility was verified through range checks. Outliers were removed based on cutoff points determined by Dutch diabetes experts (see

Because patient data were not available for the period before introduction of the bundled payment system, we assessed the effects of the diabetes DMPs by comparing the last measurement of each clinical outcome during the first year of the research period (or the first 8 months for the 2 groups with a 20-month research period) with the last measurement of that outcome in the second year. Per outcome-specific analysis, we excluded patients who (1) lacked registrations of the first- or second-year measurements, or both; (2) missed registrations of 1 or more of the patient and/or process characteristics used for stratification; and/or (3) had an observation period between first- and second-year outcome measurement of fewer than 3 months. The maximum length of follow-up per patient was 23 months. The study flow chart is shown in 

Diabetes Disease Management Programs in The Netherlands

In 2007, 10 Dutch care groups started experimenting with a bundled payment system that allows the different components of care for generic type 2 diabetes to be purchased, delivered, and billed as a single product or service. Care groups are legal entities in primary care, which consist of multiple care providers, are most commonly owned by GPs, and form the principal contracting partners for health insurers involved in bundled payment contracts.

 Groups either deliver the various components of a diabetes DMP themselves or subcontract other providers, including GPs, physical therapists, dieticians, laboratories, and/or specialists, to do so. The price of a specific program (ie, care bundle) as well as the fees for individual subcontractors are freely negotiable between care groups on the one hand and health insurers or care professionals on the other. Patients are automatically enrolled by their GP.

The services to be covered in the diabetes DMPs are codified by a national care standard for generic diabetes care.

This standard is based on existing evidence-based guidelines for GPs and includes general modules (eg, information, education and self-management support, smoking cessation, physical activity, nutrition and diet) as well as disease-specific modules.

The latter comprise a defined frequency of GP visits, regular foot and eye examinations, and laboratory testing. To stimulate task redistribution from GPs to practice nurses, who traditionally play an important role in Dutch chronic care, the standard describes care in terms of functions rather than providers, defining what services must be delivered rather than by whom or where.

 For more information on the contents of the Dutch diabetes DMPs, the diabetes care standard can be accessed online: http://www.diabetesfederatie.nl/zorg/zorgstandaard.html.

Definition of Variables and Data Validation

We defined intermediary clinical outcomes as mean changes in glycated hemoglobin (A1C), low-density lipoprotein (LDL), systolic blood pressure (SBP), and body mass index (BMI) between the first- and second-year measurements. In addition, we assessed and compared the proportions of patients within 3 target range groups for glycemic control (A1C 

75 mmol/mol) at these 2 measurement points.

Care processes were described in terms of measurement frequency and range, and duration of care. We codified measurement frequency as the number of registrations of each clinical outcome during follow-up; to describe measurement range, we assessed the number of different outcomes registered, which across care groups could be a maximum of 8 (ie, A1C, total cholesterol, LDL and high-density lipoprotein, SBP and diastolic blood pressure, BMI, and triglycerides). Duration of care was defined as the number of months between the first- and second-year measurements of each clinical outcome.

To describe patients, we used age in years, disease duration in years, health status, comorbidity, and smoking status. Health status was determined by the first-year values of each clinical outcome. Comorbidity was defined as the presence of 1 or more of the 4 most frequently registered co-occurring conditions across the included care groups (ie, angina pectoris, myocardial infarction, stroke, transient ischemic attack). We dichotomized smoking status as previous smoker or nonsmoker versus current smoker.

We used a 2-step approach to population-based, multilevel meta-analysis, in which we clustered the individual patient data by care group.

 During the first step, we conducted paired-sampled

 tests (2 sided; α = .05) using SPSS version18 (IBM Corporation, Armonk, New York) to calculate the group-specific mean differences in clinical values and associated standard deviations. In the second step, these were synthesized with Review Manager (RevMan version 5.1.1; The Cochrane Collaboration) into pooled mean differences and 95% confidence intervals, for which—owing to significant heterogeneity in results&mdash;we used the random-effects metaanalysis model of DerSimonian and Laird.

 Using this model, we weighted the aggregate effects by the inverse of their variances while assuming random treatment effects across care groups.

 To quantify heterogeneity, we calculated the I2 statistic—which can range from 0% to 100%&mdash;on the basis of the χ2 test.

 >75%) heterogeneity, subgroup analyses were conducted to examine the consistency of results across diabetes care processes and patients.

 For this purpose, most continuous variables were categorized into 2 or 3 groups on the basis of either scientific literature (age

) or median values (measurement frequency and length of followup). We dichotomized measurement range as 8 outcomes (the maximum number of clinical indicators that could be registered) versus fewer than 8 outcomes; first-year health status was categorized according to the target range values for clinical outcomes included in the diabetes care standard.

We used multilevel meta-regression analysis of individual patient data, stratified by care group, to further assess the influence of potential effect modifiers as well as to investigate potential 3-way interactions between care processes, patient characteristics, and clinical outcomes.

 To conduct the meta-regressions, we used the PROC MIXED command in the SAS 9.2 software (SAS Institute Inc, Cary, North Carolina), which uses a random-effects iterative method to provide a maximum likelihood estimate of the regression parameters. The meta-regression models were multivariable, though process and patient characteristics were included separately; continuous covariates were included as such. For each clinical outcome, we calculated the intraclass correlation coefficient to assess the percentage of total heterogeneity in effects occurring between care groups. The larger the intraclass correlation coefficient, the greater the proportion of variance that can be attributed to differences between rather than within groups.

 We examined collinearity with the variance inflation factor: a variance inflation factor value of greater than 10 is generally taken as an indication of serious multi-collinearity.

 Explained heterogeneity was expressed as the percent change in between-group variance (τ

With regard to care processes, patients’ SBP was assessed most frequently (n = 4) during the period between the firstand second-year measurements, followed by BMI (n = 3) and A1C (n = 2). LDL was measured least often (n = 1). Across care groups, the average share of patients with the maximum outcome measurement range varied from 44.4% to 86.7%, with a mean of 62.3%. Median length of follow-up varied from 10 to 12 months between groups.

shows the patients’ main characteristics and firstyear clinical values. Age ranged from 15 to 105 years with a mean of 65.7 (± 11.9) years. Diabetes duration varied from 0 to 76 years with a mean of 4.8 (± 5.6) years. Approximately 16% of patients had a known comorbidity; 18.4% were registered as current smokers. Depending on the clinical outcome of interest, roughly 56% to 71% of patients (n = 105,056) had both a first- and second-year measurement. For most outcomes, the mean first-year values for patients were within normal range. Mean first-year BMI (29.7 ± 5.2 kg/m2), however, signified severe overweight to borderline obesity.

shows the overall results of the random-effects meta-analysis across 18 care groups, presented per clinical outcome. Mean A1C increased by 0.17 mmol/mol (95% confidence interval [CI] —0.60 to 0.93) between the firstand second-year measurements. LDL was reduced significantly by a mean of 0.09 mmol/L (95% CI −0.13 to −0.05); SBP decreased significantly by an average of 0.95 mm Hg (95% CI −1.25 to −0.64). There was a small average decrease in BMI of 0.04 kg/m

 (95% CI −0.10 to 0.02). Except for BMI, the effects of the diabetes DMPs on all intermediary outcomes were moderately to highly heterogeneous, with I

 values ranging from 57% for SBP to 98% for A1C.

 shows the proportions of patients within 3 target range groups for glycemic control at the first- and second-year measurements. Glycemic control improved in the vast majority of patients, especially in the group with highly uncontrolled diabetes (A1C 

75 mmol/mol). Diabetes was brought under control in roughly 68% of these patients; in 19%, the values improved to 53 mmol/mol or lower. Of the patients with first-year A1C concentrations between 53 and 75 mmol/mol, approximately 36% improved their glycemic control, 60% remained within the same range, and 4% deteriorated. Of those within target range at the first measurement (A1C 

 53 mmol/mol), 83% maintained their A1C control, whereas A1C control deteriorated up to levels of <75 mmol/mol in 16.5%. Roughly 0.5% saw their A1C increase to 

75 mmol/mol. In terms of mean age and disease duration, the patients within target range at the first-year measurement were most comparable to the overall patient population. Those in the subgroups with initial A1C values above 53 mmol/mol and 75 mmol/mol were significantly younger but had a longer-thanaverage disease duration.

Articles

MANAGERIAL

The heterogeneous nature of care processes and patients should be taken into account in both the design and evaluation of disease management programs for diabetes.

Toward Tailored Disease Management for Type 2 Diabetes

 To determine if therapeutic gap identification, notification of community pharmacists, and intervention results in increased gap closure, reduced gap closure time, and fewer adherence gaps reopening.

Prospective, controlled, clusterrandomized study.

State of Illinois employees and beneficiaries of State health plans filling prescriptions at independently owned community pharmacies were included. For selected chronic conditions and medications, gaps in medication adherence and omitted essential therapies were identified from prescription claims and sent as alerts for resolution with the patient and/or physician. Adherence and omission gap closure at 90 days were analyzed with Kaplan-Meier (KM) survival curve approach and Cox proportional hazards models including covariates.

 A total of 1433 intervention and 1181 control adherence gaps were identified, while 677 intervention and 534 control omission gaps were generated. Pharmacists intervened on 639 (44.6%) adherence and 506 (74.7%) omission gaps. Gaps were closed more often in intervention than control at 30 days (55.5% in intervention vs 50.6% in control), 45 days (61.1% vs 58.4%, respectively), 60 days (66.1% vs 65.2%, respectively), and 90 days (73.0% vs 72.9%, respectively; adjusted hazard ratio [HR] = 1.242;

 = .022; 95% confidence interval [CI] 1.115-1.385). Adherence gaps reopened less frequently in the intervention group (HR = 0.863; 

 = .012; 95% CI 0.769-0.968). A total of 89 (13.1%) intervention and 29 (5.4%) control omission gaps closed within 90 days (adjusted HR = 1.770;

Independent community pharmacists reduced gaps in care and had fewer reopened adherence gaps, suggesting improvement in adherence. A continuation study will examine the impact of the program on long-term adherence.

This targeted patient outreach program involved identifying patients with poor adherence and omitted therapies (referred to as gaps in care) and notifying independent community pharmacists of these gaps in care.

		Prior to the program, community pharmacists underwent case-based training focused on disease management, motivational interviewing, and communication.

		The program resulted in increased adherence and omission gap closure and fewer adherence gaps reopened. An indicator of adherence was also improved by the program.

		This evaluation of our program can be used to develop more effective, targeted, evidence- based strategies to improve patient adherence and reduce errors of omission.

Underutilization of medications, exemplified by poor medication adherence and errors of omission, is a significant barrier to optimizing patient care. Adherence to chronic medications is reported to be as low as 50% after 1 year of therapy, thereby reducing the clinical effectiveness of drug therapy.

Medication non-adherence is commonly associated with cost, number of medications, duration of treatment, regimen complexity, and patient cognition.3 Likewise, omission gaps, where essential drug therapies are underutilized, are also common. In 1 Veterans Affairs study, omission errors accounted for 20% of all identified medication errors.

 Omissions in therapy have been shown to increase morbidity and mortality with greater potential to cause harm than unjustified prescribing.

In patients with diabetes, hypertension, hypercholesterolemia, and heart failure, poor adherence has been associated with increased hospitalizations and emergency department visits.

Poor adherence to medications has been shown to increase mortality, as exemplified by increased post—myocardial infarction fatality in patients with low adherence to statins and beta-blockers.

 In patients with diabetes and dyslipidemia, poor adherence has also been associated with increased medical and total costs of care.

 Overall, non-adherence to medications is believed to cost the United States healthcare system an estimated $100 billion annually.

While there are no available direct estimates for the implications and cost associated with omitted therapies, these costs could be greater because indicated, evidence-based therapies are not being used at all.

Automated physician notifications of needed omitted therapies have also been tried to increase use of these medicines.

 While many notifications were either ignored or perhaps were not true gaps, automated physician notification did result in increased use of omitted therapies. The involvement of a pharmacist to screen out these unnecessary notifications may be more efficient, but has not been evaluated.

Studies in the community setting aimed at improving patient adherence using mailings10 and telephone reminder systems

 have metwith mixed results. Those programs that focus solely on reminders tend to be less effective, since medication adherence is generally considered to be multifaceted and affected by system and health barriers, patient behaviors and beliefs, and patient self-efficacy.

As such, multifaceted interventions are more likely to be successful in changing patient

behavior and improving adherence. 16 More comprehensive interventions addressing adherence have generally shown better results.

Pharmacists are in an optimal position to address therapeutic gaps in care. Community pharmacists are knowledgeable about potential barriers and solutions to adherence issues. They have frequent contact with their patients, know them well, and when needed, reach out to physicians on their behalf. They are also widely considered a trusted source for providing information to patients and practitioners about medication therapies.

 Therefore, with the right supportive infrastructure, the community pharmacist may be very effective at validating and addressing therapeutic gaps in care related to underutilization of medications. Such an infrastructure includes pharmacists having access to evidence-based gaps in care alerts and comprehensive medication and fill history at the point of care.

The objective of this study was to determine if the identification of therapeutic gaps in care from prescription claims data, notification of community pharmacists of gaps through alerts, and intervention on those alerts results in higher gap closure, reduced time to closure, and reduced proportion of gaps reopening in the case of adherence gaps.

This study was a prospective, cluster-randomized, controlled quality assurance study. The study was approved by the University of Illinois at Chicago institutional review board and conducted in compliance with all regulations of the Health Insurance Portability and Accountability Act of 1996.

All pharmacies were required to 1) service members in the State of Illinois network, 2) be located in Illinois, 3) be independently owned, 4) utilize a commercially available web-based clinical documentation platform capable of relaying real-time gap information from Medco to pharmacists and capturing pharmacists’ documentation of gap causes and status, patient disposition, and comments, and 5) have 10 or more patients with identified care gaps over the 13-week period prior to randomization. In all, 96 independently owned community pharmacies met inclusion criteria and were invited to participate. Pharmacies were randomized to either provide the intervention or not (usual care control).

Population and Patient Inclusion Criteria

State of Illinois employees and beneficiaries (spouses, children, and retired workers) with continuous eligibility during the study period, and all of their pharmacy claims, were included. Patients were assigned to a particular pharmacy based on where the patient was filling the majority of their prescriptions in the 6 months prior to the study. Patient assignment was solely for the purpose of providing the intervention. Specifically, gap alerts were sent to the assigned intervention pharmacy. Control pharmacies did not receive gap in care alerts. At the start of the project, a letter was sent to the intervention pharmacy patients to inform them of the services being offered. Patient participation was optional; patients were not informed about the study. Patients were not required to fill medications at the assigned pharmacy.

Adherence and omission gaps in care were identified daily, using Medco’s prescription claim warehouse and defined clinical criteria. Gaps in care were transmitted as alerts to the pharmacy via the web-based platform. Adherence gaps occurred when a patient on selected chronic medications had a medication possession ratio (MPR) of less than 80% over the last 18 months and was late to fill the medication. Omission gaps occurred when a patient with a selected chronic condition was missing a medication on their profile. Omission gaps were selected based on documentation of clinical evidence and general consensus supporting the medication’s use for a particular condition. The alert definitions and situations for which they were generated are shown in 

Intervention Pharmacy Training and Practice

Control pharmacies and pharmacists continued to provide the services normally provided to their patients and did not receive gap in care alerts. Beyond the usual patient services, intervention pharmacies provided the additional service of addressing care gaps identified by Medco and communicated via a web-based clinical documentation and communication platform. All participating pharmacists at intervention pharmacies were required to undergo a certified continuing education program comprising 7.5 hours of online and 3.5 hours of in-person case-based training conducted by residencytrained clinical pharmacists with experience in medication therapy management and ambulatory care specialty clinics. Training focused on disease management (see Table 1 for conditions), motivational interviewing, and communication. Program sponsors held monthly telephone conferences with the intervention pharmacists to gather feedback on process improvement and best practice sharing and provide updates. Additional education and communication materials were provided when requested, such as the American Diabetes Association treatment guidelines.

The approach to addressing therapeutic gaps depended on the type of gap identified, as well as patient-, pharmacist-, and pharmacy-specific factors. In-person communication was encouraged for the first patient visit, with follow-up visits being in-person or by telephone. Communication with prescribers by telephone or facsimile was encouraged. However, pharmacists were allowed to conduct the intervention as appropriate, according to the pharmacy’s policies, procedures, and established relationships.

An adherence gap was considered “closed” once a prescription was filled for the medication in question. An adherence gap was considered “reopened” if an alert was generated for the gap for a second time during the study period. An omission gap was considered closed when a new prescription was written for an appropriate medication. Typically, adherence gaps were addressed directly with the patient, while omission gaps were addressed with the patient’s primary care or other provider as appropriate.

Pharmacists documented their interventions for the identified gap in care using the web-based clinical platform. All pharmacist activities were within the scope of practice, as defined by Illinois State and national pharmacy organizations and by the Illinois State practice laws. Pharmacies were reimbursed $30 for the initial session and $15 for each of up to 2 follow-up sessions to a maximum of $60 per gap addressed.

A power analysis, conducted prior to the study start, established that a sample of 410 adherence and omission gaps per group were required to detect a difference in absolute gap closure rate of 10%, assuming alpha of 0.05 and power of 0.8. We estimated that approximately 1500 adherence gaps and 1000 omission gaps would be available for analysis if 45 pharmacies were enrolled in each group. We targeted 48 intervention and 48 control pharmacies, in case any pharmacies needed to drop out.

Prescriber, pharmacy, and patient characteristics were analyzed with summary statistics of counts and proportions, as were data regarding number of gap openings, closings, and in the case of adherence gaps, reopening. Gaps that existed at the start of the study as well as new gaps that occurred within 90 days after the study start date were analyzed. Gap closure was analyzed at 90 days after an alert was sent to the pharmacy (or wouldhave been sent in the case of control pharmacies). Where a patient had 2 gaps during the intervention period for the same gap category, only the first gap was used in assessing gap closure. Unadjusted analyses of gap closure at 90 days were initially conducted using the χ

 test statistic. In the case of adherence gaps, an unadjusted χ

 analysis was conducted assessing the proportion of gaps that reopened prior to the end of the study.

Adherence and omission gap closure over time were analyzed using the Kaplan-Meier (KM) survival curve approach

 For the adherence gap model, patient age and gender and total number of Medco prescription claims during the study period and Medco patient count for the pharmacy were included as potential covariates in the model. For the omission gap model, the same patient and pharmacy covariates were included, along with prescriber age, gender, specialty, practice type, practice size (estimated by number of Medco patients), and number of Medco prescription claims. Variable selection methods were used to find the final model following the selection criteria; both the entry and stay

WEB EXCLUSIVE

A program identifying and resolving care gaps involving community pharmacists resulted in increased adherence and omission gaps closure and fewer adherence gap reopenings.

Identification of and Intervention to Address Therapeutic Gaps in Care

To determine whether addition of written-prescription data to existing adherence measures improves identification of nonadherent patients and prediction of changes in low-density lipoprotein (LDL) cholesterol.

Retrospective database analysis of all health plan members prescribed a statin in 2008 and followed through 2010.

We examined statin use in a 4-millionmember health plan with 100% electronic medical  record coverage. A novel type of medication possession ratio (MPR), integrating prescribed with dispensed medication data, was developed. This measure, MPR

, was compared with a standard dispensed-only adherence measure, MPR

. Adherence below 20% was considered nonadherence. The 2 adherence measures were compared regarding (1) the number of patients identified as nonadherent, (2) percent changes in LDL from study enrollment to study termination, and (3) receiver-operator curves assessing the association between adherence and a 24% decrease in LDL.

 A total of 67,517 patients received 1,386,270 written prescriptions over the 3-year period. MPR

 identified 93% more patients as nonadherent than did MPR

 <.001). These newly identified patients exhibited minimal LDL decreases over the course of the study. Adherence by MPR

 was more strongly associated with decreases in LDL than was adherence by MPR

 <.001). During the study period, 18.2% of patients did not fill any prescriptions and were thus unidentifiable by dispensed-only measures.

Addition of written-prescription data to adherence measures identified nearly twice the number of nonadherent patients and markedly improved prediction of changes in LDL.

Commonly used measures of medication adherence use only dispensed medication data. Attempts to integrate written-prescription data into these measures have been limited to “primary adherence” and have not been validated with respect to clinically important outcomes

		Addition of written-prescription data to measures of adherence identified nearly twice as many nonadherent patients and markedly improved prediction of changes in low-density lipoprotein cholesterol.

		Adding such measures to existing health records would identify a large proportion of patients prescribed but not procuring any medication, allowing for improved targeting of nonadherent populations.

Nonadherence to prescribed medication is a pervasive phenomenon, more common in chronic than in acute conditions.

 Poor adherence to medication regimens has been associated with poor medical outcomes over a range of diagnoses,

 and poor adherence to statins has been found to be associated with higher healthcare costs and cardiovascular hospitalizations.

Clinicians and healthcare managers thus have a strong interest in accurately identifying poorly adherent patients. In the past 2 decades, estimating adherence in healthcare systems using administrative databases has become commonplace. Medication possession ratio (MPR) and related measures of medication availability are the most frequently used measures

; however, they are based almost exclusively upon dispensed-medication data. Patients never filling a prescription will be missed, and censoring follow-up at the time of the last refill may overlook periods during which physicians continued to prescribe but patients ceased to adhere. For a comprehensive overview of dispensary-based medication adherence measures, see Steiner and Prochazka.

In recent years, several studies have attempted to integrate writtenprescription data into adherence measures. However, these have focused almost entirely upon primary nonadherence (failure to fill a first prescription for a new medication).

Only a single study using written-prescription data looked past the initial dispensary event to calculate adherence over an extended period.

 Thus, the value of integrating written prescription—based data into medication adherence measures remains unclear.

Healthcare in Israel is delivered primarily through 4 nationwide health plans, which act both as providers and insurers to essentially the entire population and through which patients are guaranteed a legally mandated minimum package of medical services.

 Clalit Health Services (CHS), the largest of these plans, insures approximately 4 million people (53% of the Israeli population) and has comprehensive clinical and administrative electronic medical record coverage. Patients in CHS have a strong incentive to access primary care and obtain prescription medications within the Clalit system because primary care visits are free of charge to members, and medications included in the benefit package are subject to small copayments (approximately US $4-$7). Out-of-pocket quarterly outlays for chronic medications are capped at US $110 to $220, depending on the patient’s eligibility.

Our aim was to ascertain the utility of a novel medication adherence measure incorporating written-prescription data in identifying nonadherent patients and predicting clinically meaningful health outcomes in a large health plan database.

We performed a retrospective database analysis, including all patients 18 years and older enrolled in CHS who had not previously been dispensed a statin, had been

prescribed a statin at least once in 2008, had not since had chronic (mental, rehabilitation, or nursing home) hospitalization, and were alive on December 31, 2010. Patients were monitored for medication adherence between January 1, 2008, and December 31, 2010 (the study period), and were considered enrolled in the study from the date of the first statin prescribed in that period. CHS maintains an electronic medical record with demographic information, clinical information, and records of all medications dispensed since 1998 and of all medications prescribed by any community physician in the system since 2008.

We collected data on age, sex, socioeconomic status, concomitant medications in the month up to and including theday of study enrollment, the occurrence of an acute ischemic event in the year preceding study inclusion, statins prescribed, statins dispensed, the last low-density lipoprotein (LDL)cholesterol level available in the year before study inclusion (prestudy LDL), and the last LDL cholesterol level available at least 90 days after study inclusion but not more than 30 days after the final statin prescription (end-study LDL). All residents of Israel receive a universal identification number at birth or immigration. Each patient interaction with CHS is indexed by this number. Patient data for our study was linked by the universal identification number, but additional personal identifying information was not collected. In-hospital prescribing and dispensing data were not available for this study; however, days of acute hospitalization over the course of the study were tallied. The study was approved by the CHS institutional review board.

Adherence measures used in this study were modeled after the concept of MPR, although with the addition of writtenprescription data.

We created an MPR-type measure integrating written-prescription data with dispensed data (MPR

) and compared it with a commonly used version of dispensedonly MPR (MPR

 was defined as the number of days of supply divided by the number of days between the day the first medication was 

 and the day before the last medication was dispensed in the study period. MPRp was defined as the number of days of supply divided by the number of days between the day the first medication was 

 and the day before the last medication was prescribed in the study period. MPR

 thus reflects adherence over the period for which statins were known to be prescribed.

Study participation was calculated as the time between the first and the last prescribed statin in the study period. Any statin prescribed or dispensed was counted; thus, a change in type of statin was considered ongoing adherence, but a change to another lipid-lowering medication group terminated study participation. Patients with a single written prescription or single dispensed medication were assigned a 29-day study duration. Patients filling more medication than necessary to cover their entire study participation period were considered 100% adherent. MPR is reported as a percent score for each patient. In this study, patients with adherence below 20% were considered nonadherent, following the methodology of Insull et al.

To ascertain whether patients identified as nonadherent by MPR

 indeed had undertreated hyperlipidemia, we calculated the fractional decrease in LDL (the percent change in LDL cholesterol from before study enrollment until the end of study participation) in those identified as nonadherent by MPR

and separately in those identified as nonadherent by MPR

Given the known efficacy of statins in reducing LDL,

 a successful measure of statin adherence should demonstrate a strong association with patient LDL reductions. In order to compare the utility of MPR

 in predicting changes in LDL, we generated receiver operator characteristic (ROC ) curves. We considered a 24% or greater fractional drop in LDL as a clinically important decrease in LDL. This decrease is slightly smaller than that anticipated from full adherence to a variety of statins at low dosages (27%)

 and is double the National Cholesterol Education Program Working Group on Lipoprotein Measurement allowable total error of 12%

; thus, it is unlikely to be due to chance.

Sensitivity analysis was performed in several ways. First, we repeated the ROC analysis using cutoffs of 20% and 30% drops in LDL. To rule out a dominant effect of patients prescribed a statin only once or patients with short durations between the first and last statin prescribed, we repeated our analysis excluding patients whose study participation was 90 days or less, which also excluded those receiving a single written prescription. Additionally, we repeated ROC analysis excluding patients who were not dispensed any statin during the study period. Finally, ROC analysis was repeated including only patients with a prestudy LDL of 100 mg/dL or greater.

Statistical analysis was conducted using SPSS version 18.0.3 (IBM Corporation, Armonk, New York) and WIN-PEPI version 11.15 (freeware; http://www.brixtonhealth.com/ pepi4windows.html). Averages for medication adherence were compared with the

 test. Proportions of nonadherent patients by each measure were compared using χ

 test. Fractional differences in LDL were compared by analysis of variance. ROC curves were generated using standard methodology, and areas under the ROC curve were compared using the Hanley-McNeil method.

A total of 67,517 members of CHS fulfilling all inclusion criteria received a total of 1,386,270 written prescriptions for a statin over the 3-year study period. The average age of the study population was 55.3 ± 12.9 years and 47.2% were male. Approximately half (47.6%) the patients were characterized as of low socioeconomic status. Patients took an average of 3.3 ± 2.9 distinct concomitant medications in the month up to and including the day of index prescribed statin. In the year before study enrollment, 3.7% of patients had sustained an acute coronary event. The average length of study participation was 692.8 ± 332.0 days. A total of 13,667 (20.2%) patients were hospitalized for a median of 4 days (interquartile range 2-8 days). The average number of acute hospitalization days for the whole study population was 1.5 (

Average adherence was 59.0% as measured by MPR

 was 15.9% higher than adherence as measured by MPR

 <.001). A far higher proportion of patients were identified as nonadherent by MPR

 <.001). Whereas our entire study population was available for nonadherence analysis using MPR

, 12,291 patients (18.2% of our sample) were not available for analysis using MPR

because these patients filled no prescriptions during the study period (

A total of 49,325 patients (73.1% of our population) had both prestudy and poststudy LDL tests and were thus available for LDL analysis. The average prestudy LDL

for the entire study population was 151.4 ± 30.3 mg/dL and decreased to 111.4 ± 36.7 mg/dL at the end of the study, indicating a 25.1% fractional decrease in LDL. Patients identified as nonadherent by MPR

 had an 8.2% drop in LDL (153.6 ± 29.4 mg/dL to 139.3 ± 35.1 mg/dL), while those identified as nonadherent by MPR

 had only a 6.8% drop (149.6 ± 31.1 to 136.9 ± 35.1 mg/dL; 

<.001). Thus, patients identified as nonadherent by MPR

 were treated less effectively than those identified as nonadherent only by MPR

The area under the curve for the ROC curve for MPR

 as a predictor of a 24% decrease in LDL was 0.815 (95% confidence interval [CI] 0.812-0.819), whereas that for MPR

 <.001 for the comparison), indicating that MPR

 was a better predictor of LDL change than MPR

under the curve values changed only slightly when the LDL cutoff was set at 20% or at 30% and when only patients with prestudy LDL values higher than 100 mg/dL were included (

). Limiting the analysis to patients who had been in the study for more than 90 days or to patients receiving at least 1 dispensed statin did not alter the primary outcome (data not shown).

Because nonadherence to medications is associated with adverse clinical outcomes, a major concern of clinicians and health plan managers is to identify and

manage nonadherent patients. In this study, we demonstrate that a measure of medication adherence integrating written prescription information with dispensed

medication data better identifies patients who are nonadherent to their medication and more accurately predicts changes in LDL than a commonly used dispensed-only measure. Compared with the integrated measure, the dispensed-only measure overestimated adherence by 16% and failed to identify nearly half of nonadherent patients. In particular, patients filling no prescriptions were identified as nonadherent by the integrated measure, but were invisible to the dispensed-only measure. Comparison of fractional decreases in LDL verified that patients identified as nonadherent by the integrated measure were indeed undertreated and were justly labeled as nonadherent. ROC curve analysis demonstrated that the integrated measure was significantly superior to the dispensed-only measure in predicting changes in LDL over the course of the study. Sensitivity analysis found our results robust when different LDL cutoffs were used, when limited to patients with a pretrial LDL above 100 mg/dL, and even when limited only to those patients who had filled a prescription (ie, including only patients who were “visible” to both measures). Our integrated measure not only uncovered a large population of previously undetected nonadherent patients, but it also was a better predictor of changes in LDL, even among patients visible to both measures.

In our search of the literature, we were able to identify very few studies attempting to integrate written-prescription data into estimates of medication adherence. Four of these studies dealt only with primary (first-fill) nonadherence (ie, failure to fill a first prescription for a new medication) and not with overall measures of adherence as such. Shah andcolleagues identified 30-day nonfillers (ie, patients who failed to fill an initial prescription for the study medication within 30 days, and by inference did not fill the prescription at all) separately for antihypertensive and antidiabetic medications.

 They found 17% and 15% nonfillers, respectively, closely paralleling our rate of patients eluding identification by the dispensed-only measure. In 2 separate settings, Fischer and colleagues in 2010

found 28.2% and 25.2% of patients, respectively, with primary nonadherence to antihyperlipidemic medications in 2 large community cohorts. Those rates are considerably higher than the 18.2% rate of primary nonadherence in our study. Karter et al

 expressed concern about the potential for missed new fills and the resulting differential misclassification bias in the 2010 study by Fischer et al. We lack information on the characteristics of Fischer and colleagues’ study population that would allow an understanding of the marked difference in primary adherence rates. The 2009 study by Karter et al is the work most closely related to our own.

 His group created a novel algorithm to estimate adherence that integrated written-prescription data, although it was based on a methodology measuring “gaps” (periods not covered by supply of medication). In their study population of diabetic patients, they found that 8.5% never filled a prescription for a cholesterol-lowering therapy.

 It is conceivable that patients appearing in a diabetes registry are more adherent than the general population evaluated in our study. Their methodology was not found superior to an existing gap measure in predicting changes in LDL.

Our work has several limitations, some of them common to all pharmaceutical database research. Our data could indicate that a medication was prescribed or dispensed, but we were unable to determine whether that medication had in fact been taken as prescribed. Additionally, our study was not immune to the possibility of patients obtaining medications, and possibly prescriptions, outside the health plan’s system. However, because in the CHS system visits to primary care physicians are free of charge and pharmacy copayments are low and capped, patients have a strong incentive to receive treatment within the system. These considerations may be of lesser consequence to patients of high socioeconomic status; however, poor adherence has repeatedly been found to be positively associated with lower socioeconomic status.

 Our investigations confirm this association (data not shown). A substantial proportion of our patients did not have the 2 LDL tests necessary to enter LDL analysis. Examining this question, we found that poor compliance and short study duration were strongly associated with missing LDL data. This finding suggests that full data would in fact strengthen the association observed between adherence measures and changes in LDL (data not shown).

Integration of written-prescription data into medication adherence measures doubled the number of patients identified as nonadherent and improved prediction of follow-up LDL cholesterol.

EMR-Based Medication Adherence Metric Markedly Enhances Identification of Nonadherent Patients

To examine the impact of financial incentives on physician goal commitment to guideline-recommended hypertension care.

 Clinic-level cluster-randomized trial with 4 arms: individual, group, or combined incentives, and control.

A total of 83 full-time primary care physicians at 12 Veterans Affairs medical centers completed web-based surveys measuring their goal commitment to guideline-recommended hypertension care every 4 months and telephone interviews at months 8 and 16. Intervention arm participants received performance-based incentives every 4 months for 5 periods. All participants received guideline education at baseline and audit and feedback every 4 months.

Physician goal commitment did not vary over time or across arms. Participants reported patient nonadherence was a perceived barrier and consistent follow-up was a perceived facilitator to successful hypertension care, suggesting that providers may perceive hypertension management as more of a patient responsibility (external locus of control).

Financial incentives may constitute an insufficiently strong intervention to influence goal commitment when providers attribute performance to external forces beyond their control.

		Among survey participants, patient nonadherence was perceived to be outside provider control and a barrier to successful hypertension care; consistent follow-up was perceived to be a facilitator.

		Designers of pay-for-performance programs for clinicians should consider rewarding clinicians for behaviors and outcomes under their direct control, as well as for behaviors and outcomes for which individual goal commitment could increase, in order to maximize the likelihood that the incentive will result in a positive clinical outcome.

Use of financial incentives as a tool to improve clinical quality is a growing trend in American healthcare

; insurers and hospitals are using pay-for-performance programs to change clinician behavior and improve quality of care. Importantly, the Affordable Care Act of 2010 will implement a Value-Based Purchasing Program in 2012 that offers financial incentives to hospitals based on the quality of care provided. Consequently, understanding how to incentivize providers’ behavior effectively may significantly improve quality of care.

Some researchers have identified an association between financial incentives and improved care, though these data are limited and additional evidence using rigorously designed methods is needed.

 In the largest randomized controlled trial to our knowledge about the impact of financial incentives on quality of care, preliminary findings by Petersen and colleagues8 were that financial incentives improved blood pressure control or an appropriate clinical response to uncontrolled blood pressure, suggesting that financial incentives 

 impact care if designed correctly. However, we are not aware of any healthcare studies to date that examined the mechanism by which financial incentives may act to change provider behavior. Without this knowledge, it is very difficult to design effective financial incentive programs.

Financial incentives may not be strong enough to influence physician goal commitment to guideline-recommended hypertension care when providers attribute performance to forces beyond their control.

October 2012

Toward Tailored Disease Management for Type 2 Diabetes

Identification of and Intervention to Address Therapeutic Gaps in Care

EMR-Based Medication Adherence Metric Markedly Enhances Identification of Nonadherent Patients

Financial Incentives and Physician Commitment to Guideline-Recommended Hypertension Management

Identifying Frail Older People Using Predictive Modeling

Application of New Method for Evaluating Performance of Fracture Risk Tool

Change to FIT Increased CRC Screening Rates: Evaluation of a US Screening Outreach Program

Implementation of EHR-Based Strategies to Improve Outpatient CAD Care

Pediatric Integrated Delivery System's Experience With Pandemic Influenza A (H1N1)

Medicare Part D Claims Rejections for Nursing Home Residents, 2006 to 2010