- Email: [email protected]
A Pilot Study Identifying Statin Nonadherence With Visit-to-Visit Variability of Low-Density Lipoprotein Cholesterol
A Pilot Study Identifying Statin Nonadherence With Visit-to-Visit Variability of Low-Density Lipoprotein Cholesterol Devin M. Mann, MDa,*, Nicole L. Glazer, PhDa,b, Michael Winter, MSc, Michael K. Paasche-Orlow, MDa, Paul Muntner, PhDd, Daichi Shimbo, MDe, William G. Adams, MDf, Nancy R. Kressin, PhDa,g, Yuqing Zhang, PhDa, Hyon Choi, MDa, and Howard Cabral, PhDc Nonadherence to cardiovascular medications such as statins is a common, important problem. Clinicians currently rely on intuition to identify medication nonadherence. The visit-to-visit variability (VVV) of low-density lipoprotein (LDL) cholesterol might represent an opportunity to identify statin nonadherence with greater accuracy. We examined the clinical and pharmacy data from 782 members of the Boston Medical Center Health Plan, seen at either the Boston Medical Center or its affiliated community health centers, who were taking statins and had ‡3 LDL cholesterol measurements from 2008 to 2011. The LDL cholesterol VVV (defined by the within-patient SD) was categorized into quintiles. Multivariate logistic regression models were generated with statin nonadherence (defined by the standard 80% pharmacy refill-based medication possession ratio threshold) as the dependent variable. The proportion of statin nonadherence increased across the quintiles of LDL cholesterol VVV (64.3%, 71.2%, 89.2%, 92.3%, 91.7%). Higher quintiles of LDL cholesterol VVV had a strong positive association with statin nonadherence, with an adjusted odds ratio of 3.4 (95% confidence interval 1.7 to 7.1) in the highest versus lowest quintile of LDL cholesterol VVV. The age- and gender-adjusted model had poor discrimination (C-statistic 0.62, 95% confidence interval 0.57 to 0.67), but the final adjusted model (age, gender, race, mean LDL cholesterol) demonstrated good discrimination (C-statistic 0.75, 95% confidence interval 0.71 to 0.79) between the adherent and nonadherent patients. In conclusion, the VVV of LDL cholesterol demonstrated a strong association with statin nonadherence in a clinic setting. Furthermore, a VVV of LDL cholesterol-based model had good discrimination characteristics for statin nonadherence. Research is needed to validate and generalize these findings to other populations and biomarkers. Ó 2013 Elsevier Inc. All rights reserved. (Am J Cardiol 2013;111:1437e1442) The visit-to-visit variability (VVV) of cardiovascular risk factors, such as low-density lipoprotein (LDL) cholesterol, in clinical practice has been thought to be due to random variation or measurement error.1e4 Although several physiologic mechanisms have been posited to contribute to VVV, medication nonadherence could be a key contributor. Indirect evidence for an effect of nonadherence on VVV came from a meta-analysis of the effect of different blood pressure medications on VVV. That meta-analysis found that diuretics and calcium channel blockers were associated with lower VVV than were angiotensin-converting enzyme inhibitors and b blockers.5 This observation was thought to possibly result from medication adherence, although few
data have tested this hypothesis.6 These observations created the possibility that the VVV of a biomarker such as LDL cholesterol, which has a strong correlation to a medication effect such as that from statins, might demonstrate an observable phenomenon of VVV according to differences in statin adherence. If established, the VVV of LDL cholesterol could be used to detect and trigger interventions to address statin nonadherence in clinical settings in which pharmacy claims data are not electronically integrated, such as currently the case in the large majority of the United States. To test this hypothesis, we conducted analyses using an integrated pharmacy claims and clinical database from a large urban population of adult medical patients to
a Department of Medicine, Boston University School of Medicine, Boston, Massachusetts; bDepartment of Epidemiology, Boston University School of Public Health, Boston, Massachusetts; cData Coordinating Center and Department of Biostatistics, Boston University, Boston, Massachusetts; d Department of Epidemiology, University of Alabama at Birmingham, Birmingham, Alabama; eDepartment of Medicine, Columbia University Medical Center, New York, New York; fDepartment of Pediatrics, Boston University School of Medicine, Boston, Massachusetts; and gVeterans Affairs Boston Healthcare System, Boston, Massachusetts. Manuscript received October 2, 2012; revised manuscript received and accepted January 21, 2013. This work was supported by grant RC2HL101628-01 (to N. R. Kressin and W. G. Adams, primary investigators). Dr. Kressin was also supported by Research Career Scientist award RCS 02-066-1 from the Health Services
Research and Development Service, Department of Veterans Affairs (Washington, DC). The present study was also supported by PatientOriented Mentored Scientist Award K23DK081665 through the National Institute of Diabetes, Digestive, and Kidney Diseases (Bethesda, Maryland) (to D. M. Mann) and grant UL1-TR000157 from the Boston University’s Clinical and Translational Institute (Boston, Massachusetts). Dr. Mann had full access to all the data in the present study and takes responsibility for the integrity of the data and accuracy of the data analysis. The funders 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. See page 1442 for disclosure information. *Corresponding author: Tel: 617-638-8021; fax: 617-638-8076. E-mail address: [email protected] (D.M. Mann).
0002-9149/13/$ - see front matter Ó 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.amjcard.2013.01.297
www.ajconline.org
1438
The American Journal of Cardiology (www.ajconline.org)
Figure 1. Flow diagram of study inclusion criteria.
determine the independent association of VVV in LDL cholesterol and statin adherence. Methods The study sample was patients enrolled in the Boston Medical Center (BMC) Health Plan from 2008 to 2011 who received care from BMC or any of 8 affiliated community health centers during that time. The patient data were drawn from the Massachusetts Health Disparities Repository (MHDR), which uses the Informatics for Integrating Biology and the Bedside system to aggregate de-identified data for BMC and BMC Health Plan. The MHDR currently contains >650 million electronic health record-based data elements (medications, diagnoses, laboratory test results, visit dates, and clinical observations) and claims data (including filled prescriptions) from the BMC Health Plan for the >1,200,000 subjects who received 1 clinical service at BMC or any of 8 affiliated community health centers during the past 10 years. We used the Informatics for Integrating Biology and the Bedside to access data from the MHDR to examine 74,468 BMC Health Plan members seen at BMC during the sample period. From this group, we limited our analysis to the 2,641 patients taking statin medications from 2008 to 2011. Of those taking statins during this period, 1,886 had 3 prescriptions filled; of those with 3 statin prescriptions filled, 782 had 3 LDL cholesterol measurements between their first and last statin fulfillment dates (Figure 1). The LDL cholesterol measurements that were outside the 0.1th and 99.9th percentiles were top and bottom coded to those values. If multiple LDL cholesterol measures existed for the same date (n ¼ 68 dates), an average of the same day measurements was used as a part of the VVV estimation, and it was counted only
once toward the 3-measure minimum. The primary exposure variable was the VVV of LDL cholesterol between the first and last statin fulfillment dates during the 3-year study period. The VVV of LDL cholesterol was defined as the within-patient SD during the study period. LDL cholesterol VVV was categorized into quintiles. The within-patient mean LDL cholesterol was calculated by averaging the LDL cholesterol measures during the study period. The primary outcome was medication adherence to statins, as determined from the medication possession ratio (MPR). The MPR, also known as the proportion of days covered, is calculated as the sum of the days’ supply of the medication (in this case, a statin) obtained between the first prescription fill and the last, divided by the total number of days in this period.7 This method was used as the main measure of medication adherence. The MPR was calculated using all statin fills during the study period. The statin MPR was dichotomized as nonadherent and adherent according to the standard cutoff of <80% and 80%, respectively.8 Covariate data of previously reported weak correlates of statin adherence were obtained from the MHDR Informatics for Integrating Biology and the Bedside portal. The covariates included age at the first statin fill during the study period, gender, race/ethnicity (white, black, Hispanic, other), total number of outpatient visits during the study period, mean LDL cholesterol level, number of LDL cholesterol measurements during the study period, number of days between the first and last statin prescription fills, and the diagnosis of diabetes mellitus (“International Classification of Diseases, 9th revision,” codes 250.0x to 250.9x), ischemic heart disease (codes 410.0x to 414.9x), hypertension (codes 401.0x to 405.0x), chronic liver disease (codes 571.0x to 571.9x), or cerebrovascular disease (codes 430.0x to 438.9x) at any point during the study period.9 Descriptive data are reported as percentages or the mean SD, as appropriate. All variables were examined for normality and outliers. Bivariate associations between covariates and quintiles of LDL cholesterol VVV, and between covariates and statin adherence, were tested through chi-square statistics for categorical variables and the Wilcoxon rank-sum test for continuous variables. Multivariate logistic regression models were used to examine the association between VVV of LDL cholesterol and statin nonadherence (MPR <80%). We performed an unadjusted model with the VVV of LDL cholesterol alone and an adjusted model that included age at the first statin fill, gender, race/ethnicity (define 2 paragraphs previously), and within-patient mean LDL cholesterol. Additional covariates were examined for inclusion in the models but were not included because they did not materially affect the results (number of LDL cholesterol measurements during the study period and diagnosis of diabetes mellitus, ischemic heart disease, hypertension, chronic liver disease, or cerebrovascular disease during the study period). For all logistic regression models, the odds ratios and 95% confidence intervals (CIs) were calculated. The significance level was set at p <0.05. The performance of each model for predicting the statin MPR of <80% was assessed by plotting the receiver operating characteristic curve and calculating the C-statistic (area under the receiver operating characteristic curve).
Preventive Cardiology/LDL Cholesterol VVV and Statin Nonadherence
1439
Table 1 Descriptive statistics, overall and by within-patient low-density lipoprotein (LDL) cholesterol SD (quintiles), 2008e2011 Variable
Gender Female Male Age at first prescription fill (yrs) Race White Black Hispanic Other Outpatient visits (n) Diagnosis Diabetes mellitus Coronary heart disease Hypertension Chronic liver disease Cerebrovascular disease Framingham risk index Low-density lipoprotein measurements (n) Within-patient mean low-density lipoprotein (mg/dl) Interval between first and last fill dates (days) Statin medication possession ratio <80% 80%
Quintile
p Value
10.4 (n ¼ 157)
>10.4e16.4 (n ¼ 156)
>16.4e23.0 (n ¼ 157)
>23.0e32.3 (n ¼ 155)
>32.3 (n ¼ 157)
77 (49%) 80 (51%) 58
87 (56%) 69 (44%) 53 7
77 (49%) 80 (51%) 53 7
91 (59%) 64 (41%) 52 8
94 (60%) 63 (40%) 52 8
39 (25%) 59 (38%) 18 (12%) 39 (25%) 46 40
44 (28%) 59 (38%) 25 (16%) 27 (17%) 46 39
36 (23%) 66 (42%) 28 (18%) 27 (17%) 45 36
32 (21%) 58 (38%) 37 (24%) 27 (18%) 49 58
34 (22%) 58 (40%) 32 (21%) 29 (19%) 42 31
108 (69%) 38 (24%) 148 (94%) 7 (5%) 17 (11%) 0.17 0.11 4.1 1.6 89 26 983 360
117 (75%) 36 (23%) 145 (93%) 6 (4%) 16 (10%) 0.19 0.13 4.5 1.6 101 32 1,027 351
104 (66%) 31 (20%) 142 (91%) 6 (1%) 14 (9%) 0.20 0.14 4.7 1.6 109 27 979 361
95 (61%) 30 (20%) 134 (87%) 13 (8%) 12 (8%) 0.16 0.13 4.9 2.0 115 27 979 355
88 (56%) 33 (21%) 129 (82%) 11 (7%) 17 (11%) 0.19 0.14 5.0 2.2 133 34 1,003 343
101 (64%) 56 (36%)
111 (71%) 45 (29%)
140 (89%) 17 (11%)
143 (92%) 12 (8%)
144 (92%) 13 (8%)
0.15 <0.01 0.29
0.78 <0.01 0.8 <0.01 0.27 0.86 <0.01 <0.01 <0.01 0.67 <0.01
Data are presented as mean SD or n (%). Table 2 Descriptive statistics, overall and by statin medication possession ratio (MPR), 2008e2011 Variable
Gender Female Male Age at first prescription fill (yrs) Race White Black Hispanic Other Outpatient visits (n) Diagnosis Diabetes mellitus Ischemic heart disease Hypertension Chronic liver disease Cerebrovascular disease Framingham risk index Low-density lipoprotein measurements (n) Within-patient mean low-density lipoprotein (mg/dl) Interval between first and last fill dates (days)
Statin MPR
p Value
<80% (n ¼ 639)
80% (n ¼ 143)
Overall (n ¼ 782)
353 (55%) 286 (45%) 52 8
73 (51%) 70 (49%) 55 7
426 (55%) 356 (46) 53 8
135 (21%) 258 (41%) 122 (19%) 118 (19%) 45 41
50 (36%) 42 (30%) 18 (13%) 31 (22%) 47 44
185 (24%) 300 (39%) 140 (18%) 149 (19%) 46 42
422 (66%) 137 (21%) 568 (89%) 33 (5%) 62 (10%) 0.18 0.13 4.6 1.8 112 33 1,000 338
90 (63%) 31 (22%) 130 (91%) 10 (7%) 14 (10%) 0.19 0.13 4.5 1.8 96 29 968 418
512 (66%) 168 (22%) 698 (89%) 43 (6%) 76 (10%) 0.18 0.13 4.6 1.8 109 33 994 354
0.36
<0.01 <0.01
0.96 0.48 0.95 0.48 0.39 0.97 0.24 0.36 <0.01 0.97
Data are presented as mean SD or n (%).
Sensitivity analyses were conducted by dichotomizing the VVV of LDL cholesterol quintiles into the first and second quintiles versus the third through fifth quintiles. We also substituted quintiles of VVV of total cholesterol for quintiles of VVV of LDL cholesterol and dichotomized
statin MPR as <50% versus 50%. We also conducted analyses restricted to the first 3 LDL cholesterol measures to examine the effect of patients with greater numbers of LDL cholesterol measures. We also conducted an analysis, withdrawing the first LDL cholesterol measure as an indirect
1440
The American Journal of Cardiology (www.ajconline.org)
Table 3 Logistic regression models predicting odds of statin medication possession ratio (MPR) <80%, quintiles of within-patient SD of low-density lipoprotein (LDL) subjects with 3 statin fills and 3 low-density lipoprotein (LDL) measurements within dates of first and last statin fill Unadjusted* OR (95% CI) Visit-to-visit variability (low-density lipoprotein standard deviation quintile) First (referent) Second Third Fourth Fifth Age at first statin fill (1-yr increase) Gender Female (referent) Male Race Hispanic Other White Black (referent) Within-patient mean low-density lipoprotein
1.00 1.37 4.57 6.61 6.14
(—) (0.85e2.20) (2.51e8.32) (3.37e12.95) (3.19e11.82)
p Value <0.0001
Adjusted† OR (95% CI)
p Value <0.0001
1.00 1.12 3.44 4.47 3.43 0.96
(—) (0.68e1.85) (1.83e6.46) (2.19e9.10) (1.65e7.14) (0.94e0.99)
0.01 0.76
1.00 (—) 0.94 (0.63e1.40) 0.01 0.95 0.68 0.46 1.00 1.01
(0.51e1.77) (0.39e1.17) (0.28e0.75) (—) (1.00e1.02)
0.01
OR ¼ odds ratio. * Hosmer and Lemeshow goodness-of-fit test, p ¼ 1.000. † Hosmer and Lemeshow goodness-of-fit test, p ¼ 0.3129.
control for the effect of a “first statin fill,” in which there should be a substantial variation between LDL cholesterol measures; an effect that might dilute the relation between adherence and LDL cholesterol VVV. We also repeated all analyses using the coefficient of variation instead of the SD as the measure of VVV and using a continuous measure of LDL cholesterol VVV. Data analyses were conducted using SAS/STAT software, version 9.2, of the SAS System for Windows (SAS Institute, Cary, North Carolina). Results After restricting the data set to patients with 3 statin pharmacy claims and 3 LDL cholesterol values within the dates of the first and last statin fill, the final analytic data set contained 782 patients. The average within-patient mean LDL cholesterol for our sample was 109.2 32.9 mg/dl. The VVV of LDL cholesterol ranged from 0.6 to 79.6 (mean 21.7 13.1). The distribution by gender, race, number of outpatient visits, and days between the first and last statin fills across VVV quintiles was equal (Table 1). The mean age at the first statin fill declined as the quintile of VVV increased, with a more significant decline with increasing VVV quintile when age was dichotomized at 55 years (data not shown). The proportion of the sample with diabetes also significantly declined as the quintile of VVV increased. The number of LDL cholesterol measurements (mean 4.6 1.8) and the within-patient mean LDL cholesterol level both increased significantly with increasing LDL cholesterol VVV. No significant association was found between statin nonadherence and gender, number of outpatient visits, co-morbidities, number of LDL measurements, or number of days between the first and last statin fills (Table 2). Younger age was associated with statin nonadherence. Race also demonstrated a significant relation with statin nonadherence, with blacks having a greater proportion of statin nonadherence than all other race groups. A greater within-patient mean LDL was significantly associated with statin nonadherence.
Figure 2. Sequential receiver operating characteristic (ROC) curves for identification of statin adherence (defined as statin MPR <80% vs 80%) using age, gender, race, mean LDL cholesterol, and VVV of LDL cholesterol with final model C-statistic of 0.75 (95% CI 0.71 to 0.79).
The prevalence of statin nonadherence increased across greater quintiles of LDL cholesterol VVV (64.3%, 71.2%, 89.2%, 92.3%, and 91.7%). In the unadjusted logistic regression models, a strong positive and significant association was noted between increasing quintiles of VVV and statin nonadherence (Table 3). When adjusted for gender, age at first statin fill, race, and within-patient mean LDL
Preventive Cardiology/LDL Cholesterol VVV and Statin Nonadherence
cholesterol, the association was attenuated but remained statistically significant. The inclusion of the number of LDL cholesterol measurements, number of outpatient visits, number of days between the first and last statin fills, or comorbidities to the model did not appreciably change the association of the VVV of LDL cholesterol and adherence (data not shown). Figure 2 depicts the receiver operating characteristic curves for discrimination between adherence and nonadherence among 4 models: (1) age and gender (C-statistic 0.62, 95% CI 0.57 to 0.67), (2) age, gender, race, and mean LDL cholesterol (C-statistic 0.70, 95% CI 0.65 to 0.74), (3) age, gender, race, and VVV of LDL cholesterol (C-statistic 0.75, 95% CI 0.70 to 0.79), and (4) age, gender, race, mean LDL cholesterol, and VVV of LDL cholesterol (C-statistic 0.75, 95% CI 0.71 to 0.79). The logistic regression analysis results were similar when the VVV of total cholesterol was used instead of LDL cholesterol (data not shown). Dichotomizing the statin MPR at <50% and 50% also did not change the results. Collapsing the VVV quintiles into 2 groups that appeared to cluster (quintiles 1 and 2 vs quintiles 3 to 5) led to no change in the results, with the upper quintiles showing odds about 4 times greater for statin nonadherence in the unadjusted and adjusted models compared to that for the lower quintiles (data not shown). The mean LDL cholesterol divided into quintiles was not a significant correlate of nonadherence (p ¼ 0.16), with odds ratios of 1.3 (95% CI 0.7 to 2.2), 1.4 (95% CI 0.7 to 2.6), 1.2 (95% CI 0.6 to 2.2), and 2.7 (95% CI 1.2 to 5.9) across increasing quintiles. Restricting the analyses to only the first 3 LDL measures to remove any potential confounding from outliers using many LDL cholesterol measures provided similar results (Table E1). Analyses with the first LDL cholesterol measure for each patient not included also found similar, if not stronger, findings (Table E2). No significant interactions were found between the mean LDL cholesterol and VVV and analyses using the coefficient of variation instead of the SD, because the measures of VVV were similar (Table E3 and Figure E1). The analyses using VVV of LDL cholesterol as a continuous measure are presented in (Table E4). Discussion These data have shown a positive association between increasing VVV of LDL cholesterol and statin nonadherence, measured using pharmacy refill data. The relation was maintained in adjusted models that incorporated key covariates that have traditionally been associated with nonadherence, including the mean within-patient LDL cholesterol and number of LDL cholesterol measurements. The magnitude of the relation was substantial, with the adjusted odds of being nonadherent to statins near 4 for the higher quintiles of VVV. The receiver operating characteristic curve that included VVV demonstrated good adherence discrimination characteristics—creating the potential that a VVV-based prediction model might be useful in identifying statin nonadherence. The approaches used to identify medication nonadherence have included self-report, pill counts, pharmacy records, and electronic medication event monitoring
1441
systems.8 Pill counts and electronic monitoring systems commonly fail and are too costly and burdensome for use in routine clinical settings.8 Self-reported questionnaires, although simple, have poor reliability and are difficult to administer in busy clinical settings.8 Databases in which clinical and pharmacy data are linked allow for the generation of objective medication adherence estimates and are frequently used by researchers.10 However, this metric is unavailable in most United States healthcare systems owing to the lack of integration between clinical and pharmacy claims data systems, particularly those serving disadvantaged and minority populations in whom nonadherence is both common and associated with morbidity.11 A great need exists for real-time, point-of-care tools for helping clinicians improve their ability to identify and intervene with nonadherent patients. The observed relation between the VVV of LDL cholesterol and statin nonadherence has the potential to be just such a tool, because it has a relation to nonadherence substantially stronger than that of previously identified correlates. In a previous meta-analysis of 22 cohort studies, the significant markers of statin nonadherence (e.g., age, gender, income, history of cardiovascular disease, diabetes, hypertension) all demonstrated relatively modest relations with peak odds ratios of about 30%.9 Thus, previous clinical and sociodemographic variables are not likely to be useful markers of medication nonadherence.12 The VVV of LDL cholesterol is a novel method to transform clinical data into a useful marker of statin adherence. To our knowledge, this has not been previously exhibited. The substantial discrimination ability of VVV combined with the other significant variables represents a potentially important finding. We began our model building with the variables previously identified as associated with statin adherence in previous data. The covariates included in our final model have previously been identified as weak markers of statin adherence but have never been successfully used to discriminate between adherence and nonadherence.9,13,14 We then added the VVV to the model, and it was the most significant variable included. For the first time, we created a prediction model with a strong association with statin adherence. A compelling clinical rationale exists for why the VVV of LDL cholesterol appears to have a strong relation to statin adherence. Statins, particularly newer generic statins such as simvastatin, have a potent effect on mean LDL cholesterol. As such, nonadherence to these drugs will likely have a relatively dramatic effect on the mean LDL cholesterol. The underlying drivers of nonadherence such as concerns about side effects, doubts about the need for drug therapy, problems with costs, and other psychosocial variables are difficult to detect in a busy clinical setting.13,15 Because VVV likely incorporates the effect of these variables in addition to other more modest epidemiologic variables, this could explain the more substantial relation. Moreover, the VVV of LDL cholesterol could easily be computed at the point-of-care in a modern electronic health record, making it a potentially powerful and scalable nonadherence screening strategy. With the ability to more reliably screen for statin nonadherence, clinicians might be able to avoid unnecessary dose titration in patients who are
1442
The American Journal of Cardiology (www.ajconline.org)
nonadherent and target these patients with specific adherence interventions.16 Our findings were robust to several sensitivity analyses. Using total cholesterol instead of LDL cholesterol and changing the adherence thresholds from 80% to 50% did not alter the results. Altering the categorization of the VVV grouping from quintiles into 2 categories also did not alter the findings; this further strengthened the validity of the observed association. The study findings need to be viewed within the context of several limitations. First, the data set applied to a large, urban medical cohort with a disproportionate number of minorities and a high percentage of Medicaid enrollees. Future studies are needed to replicate these findings in other data sets to ensure their validity and generalizability to other populations. However, Massachusetts health reform, which was enacted before 2008, has helped minimize the effect of cost on nonadherence in the sample. The study sample was also limited to those enrolled in a specific Medicaid health insurance plan, which represents another limitation to the generalizability of the findings but is difficult to predict in which direction this could bias the data. The use of pharmacy claims to assess medication adherence is a standard practice; however, it does suffer from the issues of patient pill dumping or storing medications and does not take into account, nor differentiate between, patient- versus physician-directed discontinuation. Therefore, there is likely some misclassification of the exposure, which would bias our results to the null. Furthermore, because this is a clinical database, the LDL cholesterol sample could not be verified as a fasting sample, which might have reduced the accuracy of the LDL cholesterol estimates. The sample also had a high prevalence of diabetes and hypertension, which likely resulted from the lower threshold for the use of statins in these populations and the need for frequent visits among both groups, giving a greater opportunity for LDL cholesterol measures. We used quintiles of VVV, because no published clinically relevant cutpoints for VVV of LDL cholesterol are available. These limitations were counterbalanced by several study strengths, including the use of a practice-based clinical sample that incorporated a population disproportionately affected by cardiovascular disease and nonadherence.17 The next steps for this research include analyses to identify optimal thresholds of VVV to detect statin nonadherence. However, because the intervention for nonadherence is very low risk (often enhanced counseling), the sensitivity can, in theory, be maximized in favor of specificity. These relations must also be validated in other data sets with different clinical populations and potential interactions with other medications examined. The effect of whether patients are currently at goal for statin therapy or not also needs to be examined in future studies. The relation between the VVV of other cardiovascular biomarkers such as hemoglobin A1c and systolic blood pressure should also be examined to determine whether the observed relation with the VVV of LDL cholesterol, if validated, is a unique phenomenon of statins or is an example of a more robust association between the variability in cardiovascular biomarkers and medication adherence. More work is also needed to identify the number of LDL cholesterol
measurements needed for a reproducible estimate of VVV and to maximize its nonadherence discrimination ability. Disclosures The authors have no conflicts of interest to disclose. Supplementary Data Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.amjcard.2013.01.297. 1. Clarke R, Shipley M, Lewington S, Youngman L, Collins R, Marmot M, Peto R. Underestimation of risk associations due to regression dilution in long-term follow-up of prospective studies. Am J Epidemiol 1999;150:341e353. 2. MacMahon S, Peto R, Cutler J, Collins R, Sorlie P, Neaton J, Abbott R, Godwin J, Dyer A, Stamler J. Blood pressure, stroke, and coronary heart disease. Part 1, prolonged differences in blood pressure: prospective observational studies corrected for the regression dilution bias. Lancet 1990;335:765e774. 3. Gidding SS, Stone NJ, Bookstein LC, Laskarzewski PM, Stein EA. Month-to-month variability of lipids, lipoproteins, and apolipoproteins and the impact of acute infection in adolescents. J Pediatr 1998;133: 242e246. 4. Glasziou PP, Irwig L, Heritier S, Simes RJ, Tonkin A. Monitoring cholesterol levels: measurement error or true change? Ann Intern Med 2008;148:656e661. 5. Webb AJ, Fischer U, Mehta Z, Rothwell PM. Effects of antihypertensive-drug class on interindividual variation in blood pressure and risk of stroke: a systematic review and meta-analysis. Lancet 2010;375: 906e915. 6. Muntner P, Shimbo D, Tonelli M, Reynolds K, Arnett DK, Oparil S. The relationship between visit-to-visit variability in systolic blood pressure and all-cause mortality in the general population: findings from NHANES III, 1988 to 1994. Hypertension 2011;57:160e166. 7. Steiner JF, Prochazka AV. The assessment of refill compliance using pharmacy records: methods, validity, and applications. J Clin Epidemiol 1997;50:105e116. 8. Osterberg L, Blaschke T. Adherence to medication. N Engl J Med 2005;353:487e497. 9. Mann DM, Woodward M, Muntner P, Falzon L, Kronish I. Predictors of nonadherence to statins: a systematic review and meta-analysis. Ann Pharmacother 2010;44:1410e1421. 10. Caetano PA, Lam JMC, Morgan SG. Toward a standard definition and measurement of persistence with drug therapy: examples from research on statin and antihypertensive utilization. Clin Ther 2006;28:1411e1424. 11. Jha AK, Ferris TG, Donelan K, DesRoches C, Shields A, Rosenbaum S, Blumenthal D. How common are electronic health records in the United States? A summary of the evidence. Health Aff (Millwood) 2006;25:w496ew507. 12. Steiner JF, Ho PM, Beaty BL, Dickinson LM, Hanratty R, Zeng C, Tavel HM, Havranek EP, Davidson AJ, Magid DJ, Estacio RO. Sociodemographic and clinical characteristics are not clinically useful predictors of refill adherence in patients with hypertension. Circ Cardiovasc Qual Outcomes 2009;2:451e457. 13. Mann DM, Allegrante JP, Natarajan S, Halm EA, Charlson M. Predictors of adherence to statins for primary prevention. Cardiovasc Drugs Ther 2007;21:311e316. 14. Ellis JJ, Erickson SR, Stevenson JG, Bernstein SJ, Stiles RA, Fendrick AM. Suboptimal statin adherence and discontinuation in primary and secondary prevention populations: should we target patients with the most to gain? J Gen Intern Med 2004;19:638e645. 15. Mann D. Resistant disease or resistant patient: problems with adherence to cardiovascular medications in the elderly? Geriatrics 2009;4:10e15. 16. Lee JK, Grace KA, Taylor AJ. Effect of a pharmacy care program on medication adherence and persistence, blood pressure, and low-density lipoprotein cholesterol: a randomized controlled trial. JAMA 2006;296: 2563e2571. 17. Kanjilal S, Gregg EW, Cheng YJ, Zhang P, Nelson DE, Mensah G, Beckles GLA. Socioeconomic status and trends in disparities in 4 major risk factors for cardiovascular disease among US adults, 1971e2002. Arch Intern Med 2006;166:2348e2355.
data have tested this hypothesis.6 These observations created the possibility that the VVV of a biomarker such as LDL cholesterol, which has a strong correlation to a medication effect such as that from statins, might demonstrate an observable phenomenon of VVV according to differences in statin adherence. If established, the VVV of LDL cholesterol could be used to detect and trigger interventions to address statin nonadherence in clinical settings in which pharmacy claims data are not electronically integrated, such as currently the case in the large majority of the United States. To test this hypothesis, we conducted analyses using an integrated pharmacy claims and clinical database from a large urban population of adult medical patients to
a Department of Medicine, Boston University School of Medicine, Boston, Massachusetts; bDepartment of Epidemiology, Boston University School of Public Health, Boston, Massachusetts; cData Coordinating Center and Department of Biostatistics, Boston University, Boston, Massachusetts; d Department of Epidemiology, University of Alabama at Birmingham, Birmingham, Alabama; eDepartment of Medicine, Columbia University Medical Center, New York, New York; fDepartment of Pediatrics, Boston University School of Medicine, Boston, Massachusetts; and gVeterans Affairs Boston Healthcare System, Boston, Massachusetts. Manuscript received October 2, 2012; revised manuscript received and accepted January 21, 2013. This work was supported by grant RC2HL101628-01 (to N. R. Kressin and W. G. Adams, primary investigators). Dr. Kressin was also supported by Research Career Scientist award RCS 02-066-1 from the Health Services
Research and Development Service, Department of Veterans Affairs (Washington, DC). The present study was also supported by PatientOriented Mentored Scientist Award K23DK081665 through the National Institute of Diabetes, Digestive, and Kidney Diseases (Bethesda, Maryland) (to D. M. Mann) and grant UL1-TR000157 from the Boston University’s Clinical and Translational Institute (Boston, Massachusetts). Dr. Mann had full access to all the data in the present study and takes responsibility for the integrity of the data and accuracy of the data analysis. The funders 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. See page 1442 for disclosure information. *Corresponding author: Tel: 617-638-8021; fax: 617-638-8076. E-mail address: [email protected] (D.M. Mann).
0002-9149/13/$ - see front matter Ó 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.amjcard.2013.01.297
www.ajconline.org
1438
The American Journal of Cardiology (www.ajconline.org)
Figure 1. Flow diagram of study inclusion criteria.
determine the independent association of VVV in LDL cholesterol and statin adherence. Methods The study sample was patients enrolled in the Boston Medical Center (BMC) Health Plan from 2008 to 2011 who received care from BMC or any of 8 affiliated community health centers during that time. The patient data were drawn from the Massachusetts Health Disparities Repository (MHDR), which uses the Informatics for Integrating Biology and the Bedside system to aggregate de-identified data for BMC and BMC Health Plan. The MHDR currently contains >650 million electronic health record-based data elements (medications, diagnoses, laboratory test results, visit dates, and clinical observations) and claims data (including filled prescriptions) from the BMC Health Plan for the >1,200,000 subjects who received 1 clinical service at BMC or any of 8 affiliated community health centers during the past 10 years. We used the Informatics for Integrating Biology and the Bedside to access data from the MHDR to examine 74,468 BMC Health Plan members seen at BMC during the sample period. From this group, we limited our analysis to the 2,641 patients taking statin medications from 2008 to 2011. Of those taking statins during this period, 1,886 had 3 prescriptions filled; of those with 3 statin prescriptions filled, 782 had 3 LDL cholesterol measurements between their first and last statin fulfillment dates (Figure 1). The LDL cholesterol measurements that were outside the 0.1th and 99.9th percentiles were top and bottom coded to those values. If multiple LDL cholesterol measures existed for the same date (n ¼ 68 dates), an average of the same day measurements was used as a part of the VVV estimation, and it was counted only
once toward the 3-measure minimum. The primary exposure variable was the VVV of LDL cholesterol between the first and last statin fulfillment dates during the 3-year study period. The VVV of LDL cholesterol was defined as the within-patient SD during the study period. LDL cholesterol VVV was categorized into quintiles. The within-patient mean LDL cholesterol was calculated by averaging the LDL cholesterol measures during the study period. The primary outcome was medication adherence to statins, as determined from the medication possession ratio (MPR). The MPR, also known as the proportion of days covered, is calculated as the sum of the days’ supply of the medication (in this case, a statin) obtained between the first prescription fill and the last, divided by the total number of days in this period.7 This method was used as the main measure of medication adherence. The MPR was calculated using all statin fills during the study period. The statin MPR was dichotomized as nonadherent and adherent according to the standard cutoff of <80% and 80%, respectively.8 Covariate data of previously reported weak correlates of statin adherence were obtained from the MHDR Informatics for Integrating Biology and the Bedside portal. The covariates included age at the first statin fill during the study period, gender, race/ethnicity (white, black, Hispanic, other), total number of outpatient visits during the study period, mean LDL cholesterol level, number of LDL cholesterol measurements during the study period, number of days between the first and last statin prescription fills, and the diagnosis of diabetes mellitus (“International Classification of Diseases, 9th revision,” codes 250.0x to 250.9x), ischemic heart disease (codes 410.0x to 414.9x), hypertension (codes 401.0x to 405.0x), chronic liver disease (codes 571.0x to 571.9x), or cerebrovascular disease (codes 430.0x to 438.9x) at any point during the study period.9 Descriptive data are reported as percentages or the mean SD, as appropriate. All variables were examined for normality and outliers. Bivariate associations between covariates and quintiles of LDL cholesterol VVV, and between covariates and statin adherence, were tested through chi-square statistics for categorical variables and the Wilcoxon rank-sum test for continuous variables. Multivariate logistic regression models were used to examine the association between VVV of LDL cholesterol and statin nonadherence (MPR <80%). We performed an unadjusted model with the VVV of LDL cholesterol alone and an adjusted model that included age at the first statin fill, gender, race/ethnicity (define 2 paragraphs previously), and within-patient mean LDL cholesterol. Additional covariates were examined for inclusion in the models but were not included because they did not materially affect the results (number of LDL cholesterol measurements during the study period and diagnosis of diabetes mellitus, ischemic heart disease, hypertension, chronic liver disease, or cerebrovascular disease during the study period). For all logistic regression models, the odds ratios and 95% confidence intervals (CIs) were calculated. The significance level was set at p <0.05. The performance of each model for predicting the statin MPR of <80% was assessed by plotting the receiver operating characteristic curve and calculating the C-statistic (area under the receiver operating characteristic curve).
Preventive Cardiology/LDL Cholesterol VVV and Statin Nonadherence
1439
Table 1 Descriptive statistics, overall and by within-patient low-density lipoprotein (LDL) cholesterol SD (quintiles), 2008e2011 Variable
Gender Female Male Age at first prescription fill (yrs) Race White Black Hispanic Other Outpatient visits (n) Diagnosis Diabetes mellitus Coronary heart disease Hypertension Chronic liver disease Cerebrovascular disease Framingham risk index Low-density lipoprotein measurements (n) Within-patient mean low-density lipoprotein (mg/dl) Interval between first and last fill dates (days) Statin medication possession ratio <80% 80%
Quintile
p Value
10.4 (n ¼ 157)
>10.4e16.4 (n ¼ 156)
>16.4e23.0 (n ¼ 157)
>23.0e32.3 (n ¼ 155)
>32.3 (n ¼ 157)
77 (49%) 80 (51%) 58
87 (56%) 69 (44%) 53 7
77 (49%) 80 (51%) 53 7
91 (59%) 64 (41%) 52 8
94 (60%) 63 (40%) 52 8
39 (25%) 59 (38%) 18 (12%) 39 (25%) 46 40
44 (28%) 59 (38%) 25 (16%) 27 (17%) 46 39
36 (23%) 66 (42%) 28 (18%) 27 (17%) 45 36
32 (21%) 58 (38%) 37 (24%) 27 (18%) 49 58
34 (22%) 58 (40%) 32 (21%) 29 (19%) 42 31
108 (69%) 38 (24%) 148 (94%) 7 (5%) 17 (11%) 0.17 0.11 4.1 1.6 89 26 983 360
117 (75%) 36 (23%) 145 (93%) 6 (4%) 16 (10%) 0.19 0.13 4.5 1.6 101 32 1,027 351
104 (66%) 31 (20%) 142 (91%) 6 (1%) 14 (9%) 0.20 0.14 4.7 1.6 109 27 979 361
95 (61%) 30 (20%) 134 (87%) 13 (8%) 12 (8%) 0.16 0.13 4.9 2.0 115 27 979 355
88 (56%) 33 (21%) 129 (82%) 11 (7%) 17 (11%) 0.19 0.14 5.0 2.2 133 34 1,003 343
101 (64%) 56 (36%)
111 (71%) 45 (29%)
140 (89%) 17 (11%)
143 (92%) 12 (8%)
144 (92%) 13 (8%)
0.15 <0.01 0.29
0.78 <0.01 0.8 <0.01 0.27 0.86 <0.01 <0.01 <0.01 0.67 <0.01
Data are presented as mean SD or n (%). Table 2 Descriptive statistics, overall and by statin medication possession ratio (MPR), 2008e2011 Variable
Gender Female Male Age at first prescription fill (yrs) Race White Black Hispanic Other Outpatient visits (n) Diagnosis Diabetes mellitus Ischemic heart disease Hypertension Chronic liver disease Cerebrovascular disease Framingham risk index Low-density lipoprotein measurements (n) Within-patient mean low-density lipoprotein (mg/dl) Interval between first and last fill dates (days)
Statin MPR
p Value
<80% (n ¼ 639)
80% (n ¼ 143)
Overall (n ¼ 782)
353 (55%) 286 (45%) 52 8
73 (51%) 70 (49%) 55 7
426 (55%) 356 (46) 53 8
135 (21%) 258 (41%) 122 (19%) 118 (19%) 45 41
50 (36%) 42 (30%) 18 (13%) 31 (22%) 47 44
185 (24%) 300 (39%) 140 (18%) 149 (19%) 46 42
422 (66%) 137 (21%) 568 (89%) 33 (5%) 62 (10%) 0.18 0.13 4.6 1.8 112 33 1,000 338
90 (63%) 31 (22%) 130 (91%) 10 (7%) 14 (10%) 0.19 0.13 4.5 1.8 96 29 968 418
512 (66%) 168 (22%) 698 (89%) 43 (6%) 76 (10%) 0.18 0.13 4.6 1.8 109 33 994 354
0.36
<0.01 <0.01
0.96 0.48 0.95 0.48 0.39 0.97 0.24 0.36 <0.01 0.97
Data are presented as mean SD or n (%).
Sensitivity analyses were conducted by dichotomizing the VVV of LDL cholesterol quintiles into the first and second quintiles versus the third through fifth quintiles. We also substituted quintiles of VVV of total cholesterol for quintiles of VVV of LDL cholesterol and dichotomized
statin MPR as <50% versus 50%. We also conducted analyses restricted to the first 3 LDL cholesterol measures to examine the effect of patients with greater numbers of LDL cholesterol measures. We also conducted an analysis, withdrawing the first LDL cholesterol measure as an indirect
1440
The American Journal of Cardiology (www.ajconline.org)
Table 3 Logistic regression models predicting odds of statin medication possession ratio (MPR) <80%, quintiles of within-patient SD of low-density lipoprotein (LDL) subjects with 3 statin fills and 3 low-density lipoprotein (LDL) measurements within dates of first and last statin fill Unadjusted* OR (95% CI) Visit-to-visit variability (low-density lipoprotein standard deviation quintile) First (referent) Second Third Fourth Fifth Age at first statin fill (1-yr increase) Gender Female (referent) Male Race Hispanic Other White Black (referent) Within-patient mean low-density lipoprotein
1.00 1.37 4.57 6.61 6.14
(—) (0.85e2.20) (2.51e8.32) (3.37e12.95) (3.19e11.82)
p Value <0.0001
Adjusted† OR (95% CI)
p Value <0.0001
1.00 1.12 3.44 4.47 3.43 0.96
(—) (0.68e1.85) (1.83e6.46) (2.19e9.10) (1.65e7.14) (0.94e0.99)
0.01 0.76
1.00 (—) 0.94 (0.63e1.40) 0.01 0.95 0.68 0.46 1.00 1.01
(0.51e1.77) (0.39e1.17) (0.28e0.75) (—) (1.00e1.02)
0.01
OR ¼ odds ratio. * Hosmer and Lemeshow goodness-of-fit test, p ¼ 1.000. † Hosmer and Lemeshow goodness-of-fit test, p ¼ 0.3129.
control for the effect of a “first statin fill,” in which there should be a substantial variation between LDL cholesterol measures; an effect that might dilute the relation between adherence and LDL cholesterol VVV. We also repeated all analyses using the coefficient of variation instead of the SD as the measure of VVV and using a continuous measure of LDL cholesterol VVV. Data analyses were conducted using SAS/STAT software, version 9.2, of the SAS System for Windows (SAS Institute, Cary, North Carolina). Results After restricting the data set to patients with 3 statin pharmacy claims and 3 LDL cholesterol values within the dates of the first and last statin fill, the final analytic data set contained 782 patients. The average within-patient mean LDL cholesterol for our sample was 109.2 32.9 mg/dl. The VVV of LDL cholesterol ranged from 0.6 to 79.6 (mean 21.7 13.1). The distribution by gender, race, number of outpatient visits, and days between the first and last statin fills across VVV quintiles was equal (Table 1). The mean age at the first statin fill declined as the quintile of VVV increased, with a more significant decline with increasing VVV quintile when age was dichotomized at 55 years (data not shown). The proportion of the sample with diabetes also significantly declined as the quintile of VVV increased. The number of LDL cholesterol measurements (mean 4.6 1.8) and the within-patient mean LDL cholesterol level both increased significantly with increasing LDL cholesterol VVV. No significant association was found between statin nonadherence and gender, number of outpatient visits, co-morbidities, number of LDL measurements, or number of days between the first and last statin fills (Table 2). Younger age was associated with statin nonadherence. Race also demonstrated a significant relation with statin nonadherence, with blacks having a greater proportion of statin nonadherence than all other race groups. A greater within-patient mean LDL was significantly associated with statin nonadherence.
Figure 2. Sequential receiver operating characteristic (ROC) curves for identification of statin adherence (defined as statin MPR <80% vs 80%) using age, gender, race, mean LDL cholesterol, and VVV of LDL cholesterol with final model C-statistic of 0.75 (95% CI 0.71 to 0.79).
The prevalence of statin nonadherence increased across greater quintiles of LDL cholesterol VVV (64.3%, 71.2%, 89.2%, 92.3%, and 91.7%). In the unadjusted logistic regression models, a strong positive and significant association was noted between increasing quintiles of VVV and statin nonadherence (Table 3). When adjusted for gender, age at first statin fill, race, and within-patient mean LDL
Preventive Cardiology/LDL Cholesterol VVV and Statin Nonadherence
cholesterol, the association was attenuated but remained statistically significant. The inclusion of the number of LDL cholesterol measurements, number of outpatient visits, number of days between the first and last statin fills, or comorbidities to the model did not appreciably change the association of the VVV of LDL cholesterol and adherence (data not shown). Figure 2 depicts the receiver operating characteristic curves for discrimination between adherence and nonadherence among 4 models: (1) age and gender (C-statistic 0.62, 95% CI 0.57 to 0.67), (2) age, gender, race, and mean LDL cholesterol (C-statistic 0.70, 95% CI 0.65 to 0.74), (3) age, gender, race, and VVV of LDL cholesterol (C-statistic 0.75, 95% CI 0.70 to 0.79), and (4) age, gender, race, mean LDL cholesterol, and VVV of LDL cholesterol (C-statistic 0.75, 95% CI 0.71 to 0.79). The logistic regression analysis results were similar when the VVV of total cholesterol was used instead of LDL cholesterol (data not shown). Dichotomizing the statin MPR at <50% and 50% also did not change the results. Collapsing the VVV quintiles into 2 groups that appeared to cluster (quintiles 1 and 2 vs quintiles 3 to 5) led to no change in the results, with the upper quintiles showing odds about 4 times greater for statin nonadherence in the unadjusted and adjusted models compared to that for the lower quintiles (data not shown). The mean LDL cholesterol divided into quintiles was not a significant correlate of nonadherence (p ¼ 0.16), with odds ratios of 1.3 (95% CI 0.7 to 2.2), 1.4 (95% CI 0.7 to 2.6), 1.2 (95% CI 0.6 to 2.2), and 2.7 (95% CI 1.2 to 5.9) across increasing quintiles. Restricting the analyses to only the first 3 LDL measures to remove any potential confounding from outliers using many LDL cholesterol measures provided similar results (Table E1). Analyses with the first LDL cholesterol measure for each patient not included also found similar, if not stronger, findings (Table E2). No significant interactions were found between the mean LDL cholesterol and VVV and analyses using the coefficient of variation instead of the SD, because the measures of VVV were similar (Table E3 and Figure E1). The analyses using VVV of LDL cholesterol as a continuous measure are presented in (Table E4). Discussion These data have shown a positive association between increasing VVV of LDL cholesterol and statin nonadherence, measured using pharmacy refill data. The relation was maintained in adjusted models that incorporated key covariates that have traditionally been associated with nonadherence, including the mean within-patient LDL cholesterol and number of LDL cholesterol measurements. The magnitude of the relation was substantial, with the adjusted odds of being nonadherent to statins near 4 for the higher quintiles of VVV. The receiver operating characteristic curve that included VVV demonstrated good adherence discrimination characteristics—creating the potential that a VVV-based prediction model might be useful in identifying statin nonadherence. The approaches used to identify medication nonadherence have included self-report, pill counts, pharmacy records, and electronic medication event monitoring
1441
systems.8 Pill counts and electronic monitoring systems commonly fail and are too costly and burdensome for use in routine clinical settings.8 Self-reported questionnaires, although simple, have poor reliability and are difficult to administer in busy clinical settings.8 Databases in which clinical and pharmacy data are linked allow for the generation of objective medication adherence estimates and are frequently used by researchers.10 However, this metric is unavailable in most United States healthcare systems owing to the lack of integration between clinical and pharmacy claims data systems, particularly those serving disadvantaged and minority populations in whom nonadherence is both common and associated with morbidity.11 A great need exists for real-time, point-of-care tools for helping clinicians improve their ability to identify and intervene with nonadherent patients. The observed relation between the VVV of LDL cholesterol and statin nonadherence has the potential to be just such a tool, because it has a relation to nonadherence substantially stronger than that of previously identified correlates. In a previous meta-analysis of 22 cohort studies, the significant markers of statin nonadherence (e.g., age, gender, income, history of cardiovascular disease, diabetes, hypertension) all demonstrated relatively modest relations with peak odds ratios of about 30%.9 Thus, previous clinical and sociodemographic variables are not likely to be useful markers of medication nonadherence.12 The VVV of LDL cholesterol is a novel method to transform clinical data into a useful marker of statin adherence. To our knowledge, this has not been previously exhibited. The substantial discrimination ability of VVV combined with the other significant variables represents a potentially important finding. We began our model building with the variables previously identified as associated with statin adherence in previous data. The covariates included in our final model have previously been identified as weak markers of statin adherence but have never been successfully used to discriminate between adherence and nonadherence.9,13,14 We then added the VVV to the model, and it was the most significant variable included. For the first time, we created a prediction model with a strong association with statin adherence. A compelling clinical rationale exists for why the VVV of LDL cholesterol appears to have a strong relation to statin adherence. Statins, particularly newer generic statins such as simvastatin, have a potent effect on mean LDL cholesterol. As such, nonadherence to these drugs will likely have a relatively dramatic effect on the mean LDL cholesterol. The underlying drivers of nonadherence such as concerns about side effects, doubts about the need for drug therapy, problems with costs, and other psychosocial variables are difficult to detect in a busy clinical setting.13,15 Because VVV likely incorporates the effect of these variables in addition to other more modest epidemiologic variables, this could explain the more substantial relation. Moreover, the VVV of LDL cholesterol could easily be computed at the point-of-care in a modern electronic health record, making it a potentially powerful and scalable nonadherence screening strategy. With the ability to more reliably screen for statin nonadherence, clinicians might be able to avoid unnecessary dose titration in patients who are
1442
The American Journal of Cardiology (www.ajconline.org)
nonadherent and target these patients with specific adherence interventions.16 Our findings were robust to several sensitivity analyses. Using total cholesterol instead of LDL cholesterol and changing the adherence thresholds from 80% to 50% did not alter the results. Altering the categorization of the VVV grouping from quintiles into 2 categories also did not alter the findings; this further strengthened the validity of the observed association. The study findings need to be viewed within the context of several limitations. First, the data set applied to a large, urban medical cohort with a disproportionate number of minorities and a high percentage of Medicaid enrollees. Future studies are needed to replicate these findings in other data sets to ensure their validity and generalizability to other populations. However, Massachusetts health reform, which was enacted before 2008, has helped minimize the effect of cost on nonadherence in the sample. The study sample was also limited to those enrolled in a specific Medicaid health insurance plan, which represents another limitation to the generalizability of the findings but is difficult to predict in which direction this could bias the data. The use of pharmacy claims to assess medication adherence is a standard practice; however, it does suffer from the issues of patient pill dumping or storing medications and does not take into account, nor differentiate between, patient- versus physician-directed discontinuation. Therefore, there is likely some misclassification of the exposure, which would bias our results to the null. Furthermore, because this is a clinical database, the LDL cholesterol sample could not be verified as a fasting sample, which might have reduced the accuracy of the LDL cholesterol estimates. The sample also had a high prevalence of diabetes and hypertension, which likely resulted from the lower threshold for the use of statins in these populations and the need for frequent visits among both groups, giving a greater opportunity for LDL cholesterol measures. We used quintiles of VVV, because no published clinically relevant cutpoints for VVV of LDL cholesterol are available. These limitations were counterbalanced by several study strengths, including the use of a practice-based clinical sample that incorporated a population disproportionately affected by cardiovascular disease and nonadherence.17 The next steps for this research include analyses to identify optimal thresholds of VVV to detect statin nonadherence. However, because the intervention for nonadherence is very low risk (often enhanced counseling), the sensitivity can, in theory, be maximized in favor of specificity. These relations must also be validated in other data sets with different clinical populations and potential interactions with other medications examined. The effect of whether patients are currently at goal for statin therapy or not also needs to be examined in future studies. The relation between the VVV of other cardiovascular biomarkers such as hemoglobin A1c and systolic blood pressure should also be examined to determine whether the observed relation with the VVV of LDL cholesterol, if validated, is a unique phenomenon of statins or is an example of a more robust association between the variability in cardiovascular biomarkers and medication adherence. More work is also needed to identify the number of LDL cholesterol
measurements needed for a reproducible estimate of VVV and to maximize its nonadherence discrimination ability. Disclosures The authors have no conflicts of interest to disclose. Supplementary Data Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.amjcard.2013.01.297. 1. Clarke R, Shipley M, Lewington S, Youngman L, Collins R, Marmot M, Peto R. Underestimation of risk associations due to regression dilution in long-term follow-up of prospective studies. Am J Epidemiol 1999;150:341e353. 2. MacMahon S, Peto R, Cutler J, Collins R, Sorlie P, Neaton J, Abbott R, Godwin J, Dyer A, Stamler J. Blood pressure, stroke, and coronary heart disease. Part 1, prolonged differences in blood pressure: prospective observational studies corrected for the regression dilution bias. Lancet 1990;335:765e774. 3. Gidding SS, Stone NJ, Bookstein LC, Laskarzewski PM, Stein EA. Month-to-month variability of lipids, lipoproteins, and apolipoproteins and the impact of acute infection in adolescents. J Pediatr 1998;133: 242e246. 4. Glasziou PP, Irwig L, Heritier S, Simes RJ, Tonkin A. Monitoring cholesterol levels: measurement error or true change? Ann Intern Med 2008;148:656e661. 5. Webb AJ, Fischer U, Mehta Z, Rothwell PM. Effects of antihypertensive-drug class on interindividual variation in blood pressure and risk of stroke: a systematic review and meta-analysis. Lancet 2010;375: 906e915. 6. Muntner P, Shimbo D, Tonelli M, Reynolds K, Arnett DK, Oparil S. The relationship between visit-to-visit variability in systolic blood pressure and all-cause mortality in the general population: findings from NHANES III, 1988 to 1994. Hypertension 2011;57:160e166. 7. Steiner JF, Prochazka AV. The assessment of refill compliance using pharmacy records: methods, validity, and applications. J Clin Epidemiol 1997;50:105e116. 8. Osterberg L, Blaschke T. Adherence to medication. N Engl J Med 2005;353:487e497. 9. Mann DM, Woodward M, Muntner P, Falzon L, Kronish I. Predictors of nonadherence to statins: a systematic review and meta-analysis. Ann Pharmacother 2010;44:1410e1421. 10. Caetano PA, Lam JMC, Morgan SG. Toward a standard definition and measurement of persistence with drug therapy: examples from research on statin and antihypertensive utilization. Clin Ther 2006;28:1411e1424. 11. Jha AK, Ferris TG, Donelan K, DesRoches C, Shields A, Rosenbaum S, Blumenthal D. How common are electronic health records in the United States? A summary of the evidence. Health Aff (Millwood) 2006;25:w496ew507. 12. Steiner JF, Ho PM, Beaty BL, Dickinson LM, Hanratty R, Zeng C, Tavel HM, Havranek EP, Davidson AJ, Magid DJ, Estacio RO. Sociodemographic and clinical characteristics are not clinically useful predictors of refill adherence in patients with hypertension. Circ Cardiovasc Qual Outcomes 2009;2:451e457. 13. Mann DM, Allegrante JP, Natarajan S, Halm EA, Charlson M. Predictors of adherence to statins for primary prevention. Cardiovasc Drugs Ther 2007;21:311e316. 14. Ellis JJ, Erickson SR, Stevenson JG, Bernstein SJ, Stiles RA, Fendrick AM. Suboptimal statin adherence and discontinuation in primary and secondary prevention populations: should we target patients with the most to gain? J Gen Intern Med 2004;19:638e645. 15. Mann D. Resistant disease or resistant patient: problems with adherence to cardiovascular medications in the elderly? Geriatrics 2009;4:10e15. 16. Lee JK, Grace KA, Taylor AJ. Effect of a pharmacy care program on medication adherence and persistence, blood pressure, and low-density lipoprotein cholesterol: a randomized controlled trial. JAMA 2006;296: 2563e2571. 17. Kanjilal S, Gregg EW, Cheng YJ, Zhang P, Nelson DE, Mensah G, Beckles GLA. Socioeconomic status and trends in disparities in 4 major risk factors for cardiovascular disease among US adults, 1971e2002. Arch Intern Med 2006;166:2348e2355.