Bone 78 (2015) 174–185

Contents lists available at ScienceDirect

Bone journal homepage: www.elsevier.com/locate/bone

Original Full Length Article

Analysis of osteoporosis treatment patterns with bisphosphonates and outcomes among postmenopausal veterans J. LaFleur a,b,⁎, S.L. DuVall a,b, T. Willson a,b, T. Ginter b, O. Patterson b, Y. Cheng a,b, K. Knippenberg a, C. Haroldsen b,c, R.A. Adler d, J.R. Curtis e, I. Agodoa f, R.E. Nelson b,c a

Pharmacotherapy Outcomes Research Center, University of Utah, 30 South 2000 East, Salt Lake City, UT 84112, USA VA Salt Lake City Heath Care System, 500 Foothill Drive, Salt Lake City, UT 84148, USA Department of Internal Medicine, University of Utah, 30 North 1900 East, Salt Lake City, UT 84132, USA d Hunter Holmes McGuire Veterans Affairs Medical Center, 1201 Broad Rock Boulevard, Richmond, VA 23224, USA e Division of Clinical Immunology and Rheumatology, University of Alabama at Birmingham, 1825 University Boulevard, Birmingham, AL 35294-2182, USA f Amgen, Inc., 1 Amgen Center Drive, Thousand Oaks, CA 91320, USA b c

a r t i c l e

i n f o

Article history: Received 16 July 2014 Revised 24 March 2015 Accepted 14 April 2015 Available online 17 April 2015 Edited by Nuria Guanabens Keywords: Osteoporosis Bisphosphonate Treatment patterns Fracture risk Cost Veterans

a b s t r a c t Purpose: Adherence and persistence with bisphosphonates are frequently poor, and stopping, restarting, or switching bisphosphonates is common. We evaluated bisphosphonate change behaviors (switching, discontinuing, or reinitiating) over time, as well as fractures and costs, among a large, national cohort of postmenopausal veterans. Methods: Female veterans aged 50+ treated with bisphosphonates during 2003–2011 were identified in Veterans Health Administration (VHA) datasets. Bisphosphonate change behaviors were characterized using pharmacy refill records. Patients' baseline disease severity was characterized based on age, T-score, and prior fracture. Cox Proportional Hazard analysis was used to evaluate characteristics associated with discontinuation and the relationship between change behaviors and fracture outcomes. Generalized estimating equations were used to evaluate the relationship between change behaviors and cost outcomes. Results: A total of 35,650 patients met eligibility criteria. Over 6800 patients (19.1%) were non-switchers. The remaining patients were in the change cohort; at least half displayed more than one change behavior over time. A strong, significant predictor of discontinuation was ≥ 5 healthcare visits in the prior year (11–23% more likely to discontinue), and discontinuation risk decreased with increasing age. No change behaviors were associated with increased fracture risk. Total costs were significantly higher in patients with change behaviors (4.7–19.7% higher). Change-behavior patients mostly had significantly lower osteoporosis-related costs than non-switchers (22%–118% lower). Conclusions: Most bisphosphonate patients discontinue treatment at some point, which did not significantly increase the risk of fracture in this majority non-high risk population. Bisphosphonate change behaviors were associated with significantly lower osteoporosis costs, but significantly higher total costs. © 2015 Elsevier Inc. All rights reserved.

Introduction

Abbreviations: BMD, bone mineral density; BMI, body mass index; CDW, corporate data warehouse; COPD, chronic obstructive pulmonary disease; CPT, Current Procedural Terminology; DSS, decision support system; GEE, generalized estimating equations; ICD-9, 9th revision of the International Classification of Diseases; NLP, natural language processing; OLS, ordinary least squares; PCE, Personal Consumption Expenditures; PMO, postmenopausal osteoporosis; SSRIs, selective serotonin reuptake inhibitors; VA, Veterans Affairs; VHA, Veterans Health Administration; VINCI, Veterans INformatics and Computing Infrastructure; WHO FRAX, World Health Organization absolute Fracture Risk Assessment Tool. ⁎ Corresponding author at: Department of Pharmacotherapy, University of Utah, 30 South 2000 East, Room 4765, Salt Lake City, UT 84112, USA. Fax: +1 801 587 7923. E-mail address: joanne.lafl[email protected] (J. LaFleur).

http://dx.doi.org/10.1016/j.bone.2015.04.022 8756-3282/© 2015 Elsevier Inc. All rights reserved.

Poor patient adherence and persistence are common with bisphosphonates, which are considered first-line treatments for postmenopausal osteoporosis (PMO). Reasons for poor adherence and persistence include but are not limited to lack of perceived benefit, lack of understanding, side effects, and inconvenience [1]. In observational studies, the proportions of patients that persisted with therapy for 1 year are low for United States (US) cohorts (21.0–56.7%) and non-US cohorts (21.9–74.8%) [2–11]. Some studies have reported an association between low adherence or persistence and fracture outcomes or healthcare costs (i.e., higher fracture rates or higher costs in patients with lower adherence) [8,12–18]. However, the effect of changing medications on osteoporosis health outcomes has not been well

J. LaFleur et al. / Bone 78 (2015) 174–185

characterized. Several studies demonstrate improved persistence or adherence following a treatment switch [7,19,20], with only one to the contrary [16]. We are aware of one study using a small regional cohort that examined the effect of medication switching on fracture risk. Briesacher and colleagues did not find a significant effect of medication switching on fracture risk in a small sample of patients who switched between once weekly and once monthly bisphosphonates [21]. Given these controversies and inconsistencies in the literature, we sought to examine the effect of switching behaviors on fracture risk and cost outcomes in a large, national cohort of postmenopausal veterans. Few studies have investigated the link between bisphosphonate medication-taking behavior and osteoporosis-related outcomes while controlling for baseline disease severity. One study restricted analysis to patients with a bone mineral density (BMD) T-score ≤ −2.5 and/or a prior vertebral fracture [14]. However, no studies adjusted for baseline BMD in an effort to isolate the independent effect of adherence on outcomes in patients across a range of osteoporosis severity. This is, perhaps, because most studies that use administrative claims datasets typically do not contain clinical data [11,14–17,20]. However, even studies conducted in datasets containing some clinical data [8,13] still lacked the ability to capture BMD in structured data, making adjustment for baseline disease severity challenging. To our knowledge, no research has been done to describe bisphosphonate use and outcomes in the female veteran PMO population while controlling for severity. Studies conducted in the Veterans Health Administration (VHA) are usually overwhelmingly male [22,23], and osteoporosis studies have been no exception [24,25]. Thus, in the bisphosphonate-treated PMO population in the VHA, we sought to characterize bisphosphonate switching patterns; to identify patient and disease characteristics that were associated with switching or discontinuation behaviors; and to investigate the possible relationship between patient medication-taking behaviors and outcomes, including cost and fracture events. As a methodological improvement on past observational analyses and a unique feature of our investigation, we used natural language processing (NLP; a computerized algorithm with which certain data or concepts in an electronic text record are identified) to extract information on BMD and other clinical risk factors for fracture from radiology reports and clinic notes. This information was used to control for baseline disease severity [26–28]. Methods Study design and datasets In this cohort study, we used historical data from several VHA datasets hosted in the VINCI (Veterans INformatics and Computing Infrastructure) environment, linking across datasets with the VHA's unique scrambled social security number system. Datasets included the VHA's Decision Support System (DSS) dataset, from which we identified bisphosphonate fills, other medications, and cost information for pharmacy, inpatient, outpatient, radiology, and laboratory services; the VHA's Medical SAS dataset, from which we identified age, race/ ethnicity, and fracture outcomes; the Corporate Data Warehouse (CDW) dataset, from which we identified vital signs (height and weight for body mass index [BMI] calculation) and narrative records (the clinician progress notes and radiology reports that we used to extract BMD T-scores and other clinical risk factors for fracture). Natural language processing (NLP) We used NLP technology to identify several risk factors from narrative text including radiology reports (for BMD) and clinic notes (for BMD, smoking, alcohol consumption, and family/maternal history of osteoporosis or fracture). Separate, narrow-focused NLP applications were built for each of these four variables. Smoking status was extracted using a system previously developed and validated on the VA clinical notes [27].

175

Applications to extract the other three variables were built following an iterative system development model, in which a system is built incrementally in phases called iterations. Each iteration consists of planning, development, and error analysis phases. For the current system at each iteration, a set of manually developed rules was built or expanded aiming to detect relevant concepts, their context, and relationships. Iterations involve rigorous error checking to compare system output to manual annotations. The error analysis findings inform the next iteration. The measure of the system performance improvement is estimated by relative decrease in the error rate between iterations. When the error rate improvement falls below 1% it is determined to have reached a plateau and the development cycle is determined to be complete [29]. Final performance of these applications was manually validated on a randomly selected set of notes for each involved variable separately. For example, for the T-score application, a random sample of 1000 instances was reviewed to assess the accuracy of the tool for extracting T-score, anatomy (e.g., anatomic sites mentioned in proximity to the T-score such as lumbar spine or femoral neck), and the association between anatomy and T-score (i.e., that a given site was related in the note to a particular anatomic site, such as a femoral neck BMD T-score versus a lumbar spine BMD T-score). The mean accuracy (number of correct extractions divided by the total number of extractions) of the BMD extraction tool was 82.8% for T-score, 92.6% for anatomic site, and 82.8% for the correct BMD being associated with the correct anatomic site. For smoking, alcohol use, and family/maternal history of osteoporosis or fracture, the mean accuracies were 83.4%, 75.9%, and 76.3%, respectively [26,27]. Previous approaches to acquiring variables for family/maternal history of osteoporosis and BMD scores have used administrative data, personal interviews or questionnaires, manual chart review, [30–33] and prospective measurement [34]. As such, they have been limited to smaller cohorts of patients. Several studies have used NLP to identify family history of other medical conditions or distinguish between family history and personal history, with accuracies ranging from 81.3% to 93.8% [35–37]. Although this range is slightly higher than reported here, each of these studies used only 1 to 2 document types from 1 to 2 hospitals, in which document sections could be specifically identified. Data for our study came from thousands of document types from more than 1400 points of care (medical centers, clinics, nursing homes, and long-care facilities) all across the US. Alcohol consumption is also usually recorded through interviews, questionnaires, or manual chart reviews [38–40]. Extraction of alcohol consumption status has been attempted using NLP before, with accuracy of 89.4% [41]. The authors admit that such a high level of accuracy is partially attributable to the high level of consistency in the clinical notes in the sample — the same physician dictated them all. Smoking status has been classified using NLP in several studies, ranging from 23% to 81% [42]. Our system performance is comparable to these previous systems. Patients We identified a national cohort of female veterans aged 50 years and older who had an outpatient encounter and who received an osteoporosis bisphosphonate prescription (oral alendronate, oral or injectable ibandronate, oral risedronate, or injectable zoledronic acid) during the study period (January 1, 2003–December 31, 2011). All patients with an outpatient encounter during the study period and then a bisphosphonate prescription at least 6 months later but still within the study period (i.e., the index prescription) were eligible for analysis and were classified as incident or prevalent bisphosphonate users. The index date was defined, not as the first bisphosphonate prescription filled during the study period, but as the first bisphosphonate prescription filled at least 6 months after the first VHA outpatient encounter in the study period. Prevalent bisphosphonate users were those with any bisphosphonate prescription between the beginning of the study period (January 1,

176

J. LaFleur et al. / Bone 78 (2015) 174–185

2003) and the index bisphosphonate prescription. Incident bisphosphonate users were those with no bisphosphonate use of any kind from the beginning of the study period (January 1, 2003) up to the index bisphosphonate prescription (the first prescription at least 6 months after the first outpatient encounter). For example, a patient may already be taking a bisphosphonate when the study period begins, and may continue on that bisphosphonate, but her index date/prescription would not occur until 6 months after her first VHA outpatient encounter during the study period. This patient would be a “prevalent” user. In the fracture analyses, patients were censored upon the first fracture after the index bisphosphonate or, for those who did not fracture, on their last encounter in the VA system. Patients were excluded if they had one or more of the exclusionary diseases, defined as having at least 2 codes from the 9th revision of the International Classification of Diseases (ICD-9) for a condition on two separate occasions at any time prior to the index date. Exclusionary diseases included Paget's disease, osteogenesis imperfecta, hypercalcemia, malignant cancer, or human immunodeficiency virus infection (HIV; see Supplemental Table 1 for specific ICD-9 codes). To avoid selection bias (because of conditioning study inclusion on an event that occurs after the index date), patients who had a second exclusionary diagnosis after the index date contributed person-time only until the date of the second exclusionary diagnosis and then were censored. Patients were also excluded if their index bisphosphonate was for a dose or dosage form that is primarily used for Paget's disease (e.g., alendronate 40 mg daily or risedronate 30 mg daily). This Health Insurance Portability and Accountability Act (HIPAA)compliant study was approved by the University of Utah Institutional Review Board (IRB) and the Salt Lake City VHA Research and Development office. Definitions Baseline patient characteristics Baseline patient characteristics in the 6 month period before the index prescription included demographics (age, race/ethnicity, BMI, marital status, smoking status, alcohol history, and healthcare visit frequency for prior year); osteoporosis disease characteristics (10-year hip fracture probability and 10-year major osteoporosis-related fracture probability based on the US-adapted World Health Organization [WHO] absolute Fracture Risk Assessment Tool [FRAX], an actionable score being N3% for hip fracture or N20% for any major fracture), such as BMD T-score, history of prior fracture, family/maternal history of osteoporosis or fracture, and index bisphosphonate use; comorbidities (diabetes, dementia, chronic obstructive pulmonary disease [COPD], and depression); and drug exposures (calcitonin, calcium, heparin, hormone replacement therapy, lithium, proton pump inhibitors, raloxifene, seizure medications, benzodiazepines, opioids, selective serotonin reuptake inhibitors [SSRIs], statins, teriparatide, thiazide diuretics, thyroid medications, and vitamin D). In the case of multiple readings of the same characteristic, we always used the one closest to the index date as “baseline”. Patients were classified as high risk or low risk as described below (see the Baseline disease severity section). Baseline disease severity Patients' baseline disease status was classified as high risk if they met any 2 of the following 3 conditions: (1) femoral neck BMD Tscore ≤ − 2.5 [43,44], (2) age ≥ 70 and BMD T-score of the total hip, spine, or one-third radius ≤ −2.5; [43–45] or (3) prior fracture at the hip, spine, forearm, humerus, pelvis, or tibia/fibula at any time [46]. These particular fracture sites were chosen not only because the WHO FRAX uses them, but also because they are documented in the literature as the most common sites for osteoporotic fractures [47]. Some baseline patient characteristics were used as inputs for the WHO FRAX absolute fracture risk calculator [48], including race/ethnicity, age, weight, height, prior major fracture, family/maternal history of fracture, current

smoking, corticosteroid use, rheumatoid arthritis diagnosis, a diagnosis related to secondary osteoporosis, and alcohol exposure. These fields were extracted from structured data or via NLP where structured data were not available (i.e., BMD, family/maternal history of osteoporosis or fracture, current smoking status as of the index date, and alcohol consumption). Bisphosphonate change behavior definitions Patient bisphosphonate change behaviors over time were characterized as non-switching, switching, discontinuing, and reinitiating. Nonswitching was defined as continuing on the index bisphosphonate for the duration of follow-up (until censoring), switching was defined as switching from the index bisphosphonate to a different bisphosphonate (e.g., alendronate to risedronate), discontinuing was defined as having stopped treatment for a gap length of at least 90 days after the end of the prior days' supply, and reinitiating was defined as restarting the index bisphosphonate after a prior discontinuation or switch. Zoledronic acid, administered once per calendar year, was also subject to the 90-day rule. We selected a 90-day gap length for classifying a discontinuation event to be conservative in classification; however, we evaluated alternate gap lengths of 30 and 60 days in sensitivity analyses. All patients were considered to be non-switchers on the index date and their change behavior status was updated every 90 days depending on their change behavior in the prior 90-day quarter. For patients with multiple change behaviors in the same quarter, the patient's exposure was updated to correspond to the most recent change. Fracture outcomes We conducted analyses for two fracture outcomes (hip fracture and any major fracture, [i.e., hip, spine, forearm, or humerus]) and a composite outcome (fracture of the hip, forearm, clinical spine, or humerus) to occur during 2003–2011. Fractures were identified using ICD-9 codes (see Supplemental Table 1). Since a recent publication demonstrated a consensus among clinicians that most fractures in patients over age 50 are osteoporosis-related [47], we included all fractures of the above skeletal sites, regardless of whether they occurred in the context of a trauma code. Cost outcomes We calculated total costs and osteoporosis-related costs during 2003–2011 by summing costs associated with inpatient and outpatient encounters from the DSS dataset. DSS contains records for office visits, pharmacy services, laboratory testing, and radiology. The definitions of “inpatient” and “outpatient” were based entirely on structured data. Inpatient costs were found in the Inpatient table of the DSS dataset, with separate variables for nursing, lab, radiology, surgery, prescription, and other costs. The Outpatient table of the DSS dataset contained outpatient costs. Osteoporosis costs were calculated by including only those cost observations that were associated with a primary ICD-9 code for osteoporosis or fracture, Current Procedural Terminology (CPT) codes for bone density scans, or a pharmacy record for an osteoporosis-related treatment (oral alendronate, oral risedronate, oral or injectable ibandronate, or injectable zoledronic acid; see Supplemental Table 1). All costs were converted to 2011 US dollars using the Personal Consumption Expenditures (PCE) Index from the Bureau of Economic Analysis. Statistical analysis Patient characteristics, overall and by change behavior Descriptive statistics, including means and standard deviations (SD) for continuous variables and frequencies and proportions for categorical variables, were calculated. We calculated descriptive statistics for baseline characteristics for all patients, patients with bisphosphonate change behaviors, and patients without bisphosphonate change behaviors.

J. LaFleur et al. / Bone 78 (2015) 174–185

Descriptive statistics were also used to characterize the proportions with change behaviors in the follow-up period. Predicting the effect of patient characteristics on bisphosphonate discontinuation To identify predictors of bisphosphonate discontinuation (defined as the occurrence of a 90-day gap following the end of a prior days' supply), a multivariable Cox Proportional Hazard regression model was constructed to examine the relationship between baseline patient characteristics and discontinuation. To identify independent predictors of discontinuation, we used a backward stepwise selection procedure, eliminating variables one at a time if p ≥ 0.1, which is the common pvalue used in backwards stepwise selection. Patients were censored at the last encounter with the VHA system in the study period. Sensitivity analyses were conducted for alternate gap lengths of 30 and 60 days using the same methods. Predicting the effect of bisphosphonate change behaviors on fracture, with sensitivity analyses A series of multivariable Cox Proportional Hazard regression models was constructed to examine the relationship between bisphosphonate change behaviors and fracture events, one for overall and one each for high-risk and not high-risk patients. To identify the least biased effect measure for the association between each change behavior and fracture events, our regression models included as covariates all observable variables with a known or theoretical relationship to fracture and bone quality: patient characteristics (age, marital status, race), utilization traits (visits in the prior year), index bisphosphonate (alendronate, ibandronate, risedronate, zoledronic acid), and FRAX risk factors (alcohol use, smoking status, BMI, prior fracture by site, rheumatoid arthritis diagnosis, diagnosis of conditions related to secondary osteoporosis, and steroid exposure). We also adjusted for baseline diagnoses of COPD and depression and baseline drug exposures thought to be associated with fracture risk (calcium, vitamin D, calcitonin, heparin, hormone therapy, lithium, proton pump inhibitors, raloxifene, benzodiazepines, opioids, selective serotonin reuptake inhibitors, statins, and thyroid medications). The bisphosphonate change behavior variable was treated as time-varying and was the only time-varying independent variable included in the models. The unit of analysis for these regression models was a 90-day person-quarter. The exposure was lagged by one quarter to avoid the scenario in which a fracture event is attributed to a change that occurred later in the same quarter. However, an unlagged sensitivity analysis was also conducted. In our primary analysis, we defined the medication change behavior variable in each quarter as a current, momentary point in time based on the most recent behavior (non-switching, switching, discontinuing, reinitiating) in the previous 90 days. However, because the effects of bisphosphonates are thought to take 1–2 years [49], we also performed 2 sensitivity analyses that varied how the bisphosphonate change behavior variable was defined. In the first sensitivity analysis, we treated the bisphosphonate exposures as cumulative rather than momentary; discontinuation was defined as the cumulative proportion of persontime during which the patient was classified as a discontinuer and updated each quarter, and switching was defined as the cumulative count of switches the patient had made since the index date, also updated quarterly. In the second sensitivity analysis, we used only patients with at least 1 year of observation and for whom there were no changes in the first year. In this model, follow-up started at the 1-year mark, so as to model the onset of effect of bisphosphonate therapy. In all models, the dependent variable was the time to a dichotomous fracture event. Multivariable Cox Proportional Hazard regression models were also used for sensitivity analyses. Effect of change behaviors on cost Many patients have no healthcare use while a few outliers have very high healthcare costs; therefore, the assumption required for ordinary

177

least squares (OLS) regression to yield unbiased results is violated in the association between change behaviors and cost outcomes [50]. Several alternative statistical models have been proposed to overcome this problem [50], and we used model fit tests to determine the most appropriate distribution for our data [51]. To obtain the least biased estimate of the effect of change behaviors on cost, we adjusted for all the important covariates listed above, plus total cost in the prior 6 months, and ran three models: one for all patients and one each for high-risk and not high-risk patients. Because each patient in the analysis cohort had the potential for multiple observations (one for each quarter of data), we used generalized estimating equations (GEE) to adjust for the clustering of observations within patients [52]. For these GEE models, we assumed a gamma distributed dependent variable with a log link. Similar to the fracture analyses, the dependent variables in the cost regressions were lagged forward one quarter to allow us to estimate the impact of a change behavior on expenditures in the subsequent quarter, all of which will have occurred after the change behavior occurred [53,54]. We ran separate models for each outcome: total costs, osteoporosisrelated costs, and osteoporosis-related pharmacy costs (see Cost outcomes, above) for overall patients, high-risk patients, and not highrisk patients. Results Patients As shown in Fig. 1, out of more than 1.6 million female veterans aged 50+ with an outpatient encounter in 2003–2011, 43,543 (2.6%) had a prescription for a bisphosphonate and 35,650 (2.2%) met all eligibility criteria and were included in the study. The majority of included patients (90.7%) were incident bisphosphonate users; only a small minority (9.3%) comprised prevalent users. Characteristics of the cohort are summarized in Table 1. The mean (SD) age was 65.7 (12.5), BMI was 27.2 (6.1), and BMD T-score (at the hip, spine, femur, or 1/3 radius)

Fig. 1. Attrition summary for study sample.

178

J. LaFleur et al. / Bone 78 (2015) 174–185

Table 1 Frequencies and percentages for selected baseline patient characteristics in N = 35,650 patients who met all inclusion/exclusion criteria. Overall (N = 35,650)

Demographics Age b55 55–59 60–64 65–69 70–74 75–79 80–84 85+ Race/ethnicity Asian Black Caucasian Hispanic Other Unknown BMI (NIH classification) Underweight (b18.5) Normal (18.5–24.9) Overweight (25.0–29.9) Obesity class 1 (30.0–34.9) Obesity class 2 (35.0–39.9) Extreme obesity (40.0+) Unknown Marital status Divorced Married Never married/single Widowed Unknown Smoking status Current Former Never Unknown Alcohol history Current Past Never Unknown Healthcare visit frequency in prior year b5 5–9 10–19 N19 Disease characteristics High riska Yes No FRAX score (hip)b ≥3% b3% Unknown FRAX score (any major fracture)b ≥20% b20% Unknown Lowest T-score (WHO criteria) Normal T-score (above −1.0) Low bone mass (−1.0 to −2.5) Osteoporosis (below −2.5) Missing Prior fracture Forearm Hip Humerus Spine Otherc Major fracture Family/maternal history of osteoporosis Family/maternal history of fracture Index bisphosphonate

Non-switching cohort (N = 6804)

Change cohort (N = 28,846)

N

%

N

%

N

%

6657 6856 5771 3636 2543 1786 5397 3004

18.7 19.2 16.2 10.2 7.1 5.0 15.1 8.4

769 1204 1274 791 458 393 1113 802

11.3 17.7 18.7 11.6 6.7 5.8 16.4 11.8

5888 5652 4497 2845 2085 1393 4284 2202

20.4 19.6 15.6 9.9 7.2 4.8 14.8 7.6

319 3122 26,645 483 459 4622

0. 9 8.8 74.7 1.4 1.3 13.0

79 438 5012 58 89 1128

1.2 6.4 73.7 0.9 1.3 16.6

240 2684 21,633 425 370 3494

0.8 9.3 75.0 1.5 1.3 12.1

1258 12,535 11,123 5862 2398 1245 1229

3.5 35.2 31.2 16.4 6.7 3.5 3.4

282 2511 2047 1037 413 248 266

4.1 36.9 30.1 15.2 6.1 3.6 3.9

976 10,024 9076 4825 1985 997 963

3.4 34.8 31.5 16.7 6.8 3. 5 3.3

9303 14,513 4544 6936 354

26.1 40.7 12.7 19. 5 1.0

1514 2944 786 1486 74

22.3 43.3 11.6 21.8 1.9

7789 11,569 3758 5450 280

27.0 40.1 13.0 18.9 1.0

9888 6253 10,463 9046

27.7 17.5 29.3 25.4

1942 1165 3697 1915

28.5 17.1 54.3 28.1

7946 5088 15,812 7131

27.5 17.6 54.8 24.7

4742 3145 1911 25,852

13.3 8.8 5.4 72.5

932 603 409 4860

13.7 8.9 6.0 71.4

3810 2542 1502 20,992

13.2 8.8 5.2 72.8

10,012 8183 8792 8663

28.1 23.0 24.7 24.3

1901 1459 1701 1743

27.9 21.4 25.0 25.6

8111 6724 7091 6920

28.1 23.3 24.6 24.0

5478 30,172

15.4 84.6

1079 5725

15.9 84.1

4399 24,447

15.2 84.8

19,337 11,232 5801

54.2 31.5 14.3

3265 2322 1217

48.0 34.1 17.9

16,072 8910 3864

55.7 30.9 13.4

24,808 5761 5081

69.6 16.2 14.3

4336 1251 1217

63.7 18.4 17.9

20,472 4150 3864

71.0 15.6 13.4

4632 10,881 7032 13,105

13.0 30.5 19.7 36.8

685 1795 1265 3059

10.1 26.4 18.6 45.0

3947 9086 5767 10,046

13.7 31.5 20.0 34.8

636 335 363 555 2764 2855 4614 2005

1.8 0.9 1.0 1.6 7.8 8.0 12.9 5.6

124 84 82 114 573 361 866 397

1.8 1.2 1.2 1.7 8.4 5.3 12.7 5.8

512 251 281 441 2191 2494 3748 1608

1.8 0.9 1.0 1.5 7.6 8.6 13.0 5.6

J. LaFleur et al. / Bone 78 (2015) 174–185

179

Table 1 (continued) Overall (N = 35,650)

Alendronate Ibandronate Risedronate Zoledronic acid

Non-switching cohort (N = 6804)

Change cohort (N = 28,846)

N

%

N

%

N

%

32,971 78 2261 340

92.5 0.2 6.4 1.0

6252 24 343 185

91.9 0.4 5.0 2.7

26,719 54 1918 155

92.6 0.2 6.6 0.5

Patient switching categories were characterized over the entire observation period in the descriptive analysis. Also descriptively characterized comorbidities and concomitant drug exposures. Key: BMI — body mass index; NIH — National Institutes of Health; WHO — World Health Organization. a Defined as 2 or more of the following: (1) femoral neck T-score ≤ −2.5; (2) age ≥70 and total hip, spine, or one-third radius T-score ≤ −2.5; (3) prior fracture of the hip, spine, onethird radius, humerus, pelvis, or tibia/fibula. b FRAX score calculated using BMD at the femoral neck, total hip, spine, or 1/3 radius if BMD was known and BMI if BMD was not known. c Excluding head, face, hands, and feet.

was − 1.58 (1.75). Only 63.2% of patients had T-scores available; of these, the majority was in the low bone mass range (T-score of −1.0 to − 2.5; 48.3%). Among the 85% of patients in whom FRAX scores could be calculated, the majority exceeded the National Osteoporosis Foundation (NOF)-recommended FRAX risk thresholds of 3% for hip fracture (54.2% of the cohort) or 20% for any major fracture (69.6% of the cohort), for whom pharmacologic treatment is recommended. Overall, mean (SD) observation time until fracture events was 52.8 (27.9) months. The minimum and maximum were 0.03 and 102 months, respectively. Bisphosphonate change behaviors As shown in Fig. 2, only 6804 patients (19.1%) persisted on their index bisphosphonate and thus were counted as non-switchers over a median follow-up period of 4 years and 95 days (i.e., 1556 days). Among the 28,846 patients (80.9%) who exhibited change behaviors, 2610 (7.3%) switched to a different bisphosphonate, 28,622 (80.3%) discontinued, and 14,452 (40.5%) reinitiated at least once during the follow-up period. At least half of all patients in the change behavior subcohort displayed more than one type of change over time. Among discontinuers, the median time to first discontinuation was 294 days and the median duration of the first discontinuation gap was 159 days from the end of the prior days' supply.

Predictors of bisphosphonate discontinuation Table 2 shows the patient characteristics that remained significant predictors of bisphosphonate discontinuation at p b 0.1. The strongest predictors were a higher number of healthcare visits in the prior year, which increased discontinuation risk 11–23% compared to patients with fewer than 5 visits; and Black and Hispanic race/ethnicity, which were associated with 17% and 21% increased risks of discontinuation compared to Caucasians. Having zoledronic acid as the index bisphosphonate was associated with a markedly decreased risk of discontinuation (only 58% of the risk for alendronate), and increasing age was associated with a 9–22% lower discontinuation risk, compared to a patient younger than age 55.

Effect of bisphosphonate change behaviors on fracture outcomes Crude incidence of any major fracture occurred at a rate of 15.9 events per 1000 person-years overall. The non-switching subcohort had an incidence of 16.9 events per 1000 person-years, and the discontinuing subcohort and reinitiating subcohort incidences were 15.7 and 16.7 events per 1000 person-years, respectively. The switching subcohort's incidence was 20.5 events per 1000 person-years. The differences among these groups were not significant.

Fig. 2. Percentages of patients classified as non-switchers, discontinuers, switchers, and reinitiators in each quarter following the index date.

180

J. LaFleur et al. / Bone 78 (2015) 174–185

Table 2 Significant predictors of discontinuation at p b 0.1. Univariate (unadjusted)

Multivariable (adjusted)

HR

95% CI

HR

95% CI

45.4 43.4 16.6 20.8 8.3

Ref 0.87 0.74 0.64 0.81

– 0.85, 0.90 0.72, 0.77 0.62, 0.67 0.77, 0.86

Ref 0.91 0.84 0.79 0.78

– 0.88, 0.95 0.81, 0.88 0.75, 0.83 0.75, 0.82

21,456 2622 469 603 3473

95.3 16.2 2.4 2.9 14.9

– 1.37 1.26 1.04 0.91

– 1.32, 1.43 1.15, 1.38 0.95, 1.12 0.88, 0.94

– 1.17 1.21 0.99 1.00

– 1.12, 1.22 1.10, 1.32 0.91, 1.07 0.97, 1.04

14,513 21,137

11,458 17,165

49.8 80.7

Ref 1.10

– 1.07, 1.12

Ref 1.08

– 1.05, 1.10

10,463 6253 9888 9046

8621 5050 7882 7070

42.7 24.0 34.0 30.3

Ref 0.95 0.87 0.84

– 0.92, 0.98 0.84, 0.89 0.82, 0.87

Ref 0.98 0.93 0.92

– 0.95, 1.02 0.90, 0.96 0.89, 0.95

10,012 8183 8792 8663

8062 6677 7036 6848

26.6 31.2 38.6 41.1

Ref 1.36 1.52 1.61

– 1.32, 1.41 1.47, 1.57 1.55, 1.66

Ref 1.23 1.22 1.11

– 1.19, 1.27 1.18, 1.27 1.06, 1.16

Disease characteristics Index bisphosphonate Alendronate Ibandronate Risedronate Zoledronic acid

32,971 78 2261 340

26,571 52 1845 155

120.3 0.3 8.7 0.9

Ref 1.00 1.05 0.68

– 0.76, 1.31 1.00, 1.10 0.58, 0.79

Ref 0.86 1.07 0.58

– 0.67, 1.14 1.02, 1.12 0.49, 0.68

Comorbidities (versus none) Depression diagnosis

5803

4705

31.3

1.41

1.37, 1.46

1.10

1.06, 1.14

Other drug exposures (versus none) Calcitonin Calcium Heparin Benzodiazepines Opioids Corticosteroids Thiazide diuretics Vitamin D

823 7258 1180 5645 12,014 4244 8414 11,219

677 5677 897 4504 9625 3373 6627 8941

3.7 36.5 6.9 27.7 60.6 98.3 22.3 34.9

1.18 1.33 1.43 1.30 1.43 1.42 1.13 1.33

1.09, 1.27 1.29, 1.37 1.34, 1.53 1.26, 1.34 1.39, 1.47 1.35, 1.41 1.10, 1.16 1.29, 1.36

1.10 1.06 1.13 1.04 1.15 1.13 1.03 1.09

1.02, 1.19 1.03, 1.10 1.05, 1.21 1.01, 1.08 1.11, 1.18 1.05, 1.13 1.00, 1.06 1.06, 1.13

Demographics Age b55 55–64 65–74 75–84 85+ Race/ethnicity Caucasian Black Hispanic Other Unknown Marital status Married Not married Smoking status Current Former Never Unknown Healthcare visit frequency Visits in prior year b5 Visits in prior year 5–9 Visits in prior year 10–19 Visits in prior year N19

N

Events

9858 10,859 5136 7577 2220

8535 8409 4163 5924 1592

26,645 3075 530 778 4622

Ia

Key: HR — Hazard ratio; CI — Confidence interval; Ref — reference; NS — Not significant; BMI — body mass index; NIH — National Institutes of Health. a Incidence of discontinuation per 1000 person-years.

Crude incidence of hip fracture occurred at a rate of 3.0 events per 1000 person-years overall. The change behavior subcohort with the highest incidence of hip fracture was the switching subcohort, at 3.7 events per 1000 person-years. This was followed by the reinitiating and discontinuing subcohorts, at 3.2 and 2.9, respectively. The non-switching subcohort had 3.2 events per 1000 person-years. The differences across change behavior sub-cohorts were not statistically significant. In the multivariable analyses, which adjusted for patient characteristics, utilization visits in the prior year, index bisphosphonate, FRAX risk factors (alcohol use, smoking status, BMI, prior fracture by site, rheumatoid arthritis diagnosis, diagnosis of conditions related to secondary osteoporosis, and steroid exposure), baseline diagnoses of COPD and depression, and baseline drug exposures thought to be associated with fracture risk, no statistically significant differences in hip or any major fracture events among the compared groups were identified for any of the change behavior groups compared to non-switchers in either the primary or sensitivity analyses (Fig. 3). Findings in sensitivity analyses were no different for gap lengths of 30 or 60 days or for the unlagged exposure.

Cost outcomes Total costs for the cohort over the study period were $1.48 billion, with inpatient costs totaling $395 million and outpatient costs accounting for $1.08 billion. Total osteoporosis-related costs over the study period were $44.7 million. Of this, inpatient costs were $11.6 million, outpatient costs were $12.7 million, and pharmacy costs were $20.4 million. Accounting for variable patient follow-up time, mean annual total costs per patient were $10,291 (with outpatient costs being $7070 and inpatient costs being $3221) per patient. Of this, mean annual osteoporosisrelated costs were $320, and of this, mean annual osteoporosis-related inpatient costs were $96 and mean annual osteoporosis-related outpatient costs were $83. The greatest contributor to osteoporosis-related costs was prescription drugs, with a mean annual cost of $140. The adjusted quarterly total, osteoporosis-related, and osteoporosisrelated pharmacy cost differences associated with change behaviors compared to the non-switching subcohort are shown in Table 3. The coefficients presented in this table represent the percentage change in quarterly cost associated with the change behavior relative to being a

J. LaFleur et al. / Bone 78 (2015) 174–185

181

Key: SW = Switching; D = Discontinuing; RE = Reinitiating; NCE = Non-cumulative exposure (Ref: Non-switching); CCSW = Cumulative count of switches (Ref: None); CQNP = Cumulative quarters of discontinuation (Ref: None). *High-risk defined as 2 or more of the following: (1) femoral neck T-score ≤-2.5; (2) age ≥70 and total hip, spine, or onethird radius T-score ≤-2.5; (3) prior fracture of hip, spine, one-third radius, humerus, pelvis, or tibia/fibula. Fig. 3. Unadjusted (a) and adjusted (b) risks of hip and any major fracture associated with time-varying quarterly bisphosphonate exposure and patient baseline risk.

non-switcher; the quarterly differences are also given in dollars. The total costs were significantly higher for all change behavior subcohorts (p ≤ 0.02) except for the high-risk switchers. In terms of magnitude, the greatest differences in total cost can be seen in the reinitiating subcohort, which was 20% higher for high-risk, 16% higher for not high-risk, and 17% higher overall; these differences corresponded to quarterly total cost differences of $637, $343, and $394, respectively. Total cost differences for switchers ranged from 13–14%, a cost difference of $294–494; the total cost differences for discontinuers ranged from only 5–8%, a cost difference of $102–268. In contrast to total costs, patients with change behaviors had significantly lower osteoporosis-related costs compared to non-switchers. In terms of magnitude, the greatest percentage differences were seen in the discontinuing cohort, which was 63% lower for high-risk and 118% lower for not high-risk switchers (106% lower overall). However, although the magnitudes of these differences were large (versus those seen for total costs), in terms of dollars, these differences were quite small: $76 lower for high-risk and $74 lower for not high-risk discontinuers ($77 overall). For switchers and re-initiators, these differences were even smaller: the percentage differences in osteoporosisrelated costs among reinitiators ranged from 17–24%, which corresponded to a $15–21 lower quarterly cost. For switchers, percentage differences ranged from a 14% lower cost to a non-significant 30% higher cost, or a range from a $10 quarterly decrease to a nonsignificant $36 quarterly increase. Like osteoporosis-related costs, the osteoporosis-related pharmacy costs were all significantly lower in the change behavior subcohorts

compared to the non-switching subcohort. The discontinuing subcohort had between 220% and 236% lower osteoporosis-related pharmacy costs than the non-switching subcohort; this corresponded to a $75 lower quarterly osteoporosis pharmacy cost. Costs for switchers were between 61–67% lower compared to non-switchers, corresponding to a $21 lower quarterly osteoporosis pharmacy cost. For re-initiators, the quarterly cost differences ranged from 58–60%, or $19–20 lower cost per quarter.

Discussion In this study of veterans with PMO, we examined bisphosphonate change behaviors — discontinuing, switching, and reinitiating bisphosphonates — and their association with fracture and cost outcomes. Discontinuation was by far the most common of the change behaviors, with more than 80% having at least one discontinuation in the first 2.5 years of treatment. The strongest risk factors for discontinuation were Hispanic race and having a high number of healthcare visits in the prior year. However, these risk increases were modest, less than 23%. One recently published study by Yun et al. [55] reported modestly increased risk (less than 15%) for Hispanic patients as well as patients with 10+ healthcare visits per year, among other risk factors, in a database cohort study of Medicaid patients. These differences in adherence associated with race/ethnicity may be the result of differences in socioeconomic characteristics of patients. Although the VA's coverage of prescription drugs is quite robust and consistent across the country, the

182

J. LaFleur et al. / Bone 78 (2015) 174–185

Table 3 Unadjusted and adjusted percentage differences in quarterly total and osteoporosis-related costs associated with change behaviors compared to non-switchers. Multivariable (adjusted)a

Univariate (unadjusted) Stratum Total cost differences Overall (reference = non-switching) Switching Discontinuing Reinitiating High-risk (reference = non-switching) Switching Discontinuing Reinitiating Not high-risk (reference = non-switching) Switching Discontinuing Reinitiating Osteoporosis-related cost differences Overall (reference = non-switching) Switching Discontinuing Reinitiating High-risk (reference = non-switching) Switching Discontinuing Reinitiating Not high-risk (reference = non-switching) Switching Discontinuing Reinitiating Osteoporosis-related pharmacy cost differences Overall (reference = non-switching) Switching Discontinuing Reinitiating High-risk (reference = non-switching) Switching Discontinuing Reinitiating Not high-risk (reference = non-switching) Switching Discontinuing Reinitiating

Percentage difference

95% CI

Absolute differenceb

95% CI

Percentage difference

95% CI

Absolute differenceb

95% CI

0.14 0.02 0.10

0.06, 0.22 0.00, 0.05 0.07, 0.14

331.40 51.73 242.55

147.87, 514.3 −6.41, 109.87 158.35, 326.75

0.14 0.05 0.17

0.06, 0.21 0.03, 0.08 0.13, 0.20

317.55 123.13 393.62

140.05, 495.04 66.90, 179.37 311.71, 475.53

0.06 0.03 0.09

−0.14, 0.25 −0.04, 0.10 −0.01, 0.19

186.95 97.19 304.27

−476.60, 850.51 −142.19, 336.57 −31.42, 639.97

0.13 0.08 0.20

−0.07, 0.33 0.01, 0.15 0.10, 0.30

424.39 267.72 637.19

−223.60, 1072.38 33.56, 501.98 307.42, 966.95

0.16 0.02 0.10

0.08, 0.23 −0.00, 0.05 0.07, 0.14

343.71 46.79 221.45

170.23, 517.18 −6.72, 100.31 143.65, 299.25

0.14 0.05 0.16

0.06, 0.21 0.02, 0.07 0.12, 0.19

294.07 102.38 342.64

127.40, 460.73 50.96, 153.80 267.45, 417.83

−0.01 −0.92 −0.18

−0.18, 0.15 −1.01, −0.82 −0.32, −0.04

−1.05 −64.19 −12.82

−12.66, 10.56 −70.66, −57.71 −22.39, −3.26

−0.14 −1.06 −0.22

−0.29, 0.00 −1.14, −0.98 −0.32, −0.11

−10.42 −77.12 −15.62

−21.07, 0.22 −82.85, −71.40 −23.42, −7.83

0.30 −0.45 −0.19

−0.07, 0.67 −0.66, −0.24 −0.46, 0.08

35.28 −53.25 −22.38

−8.96, 79.53 −78.22, −28.29 −54.63, 9.87

0.30 −0.63 −0.17

−0.14, 0.73 −0.79, −0.47 −0.39, 0.04

36.07 −76.40 −21.25

−17.57, 89.71 −95.40, −57.40 −46.94, 4.44

−0.19 −1.08 −0.19

−0.34, −0.05 −1.18, −0.98 −0.35, −0.03

−11.91 −66.12 −11.60

−20.95, −2.88 −72.31, −59.94 −21.29, −1.90

−0.27 −1.18 −0.24

−0.40, −0.15 −1.26, −1.09 −0.36, −0.12

−17.22 −73.94 −14.82

−24.95, −9.48 −79.36, −68.52 −22.18, −7.46

−0.76 −2.35 −0.59

−0.88, −0.65 −2.38, −2.32 −0.61, −0.56

−24.48 −75.40 −18.78

−28.27, −20.69 −77.04, −73.76 −19.74, −17.83

−0.66 −2.34 −0.58

−0.78, −0.54 −2.37, −2.31 −0.61, −0.56

−21.35 −75.14 −18.80

−24.22, −17.48 −76.57, −73.71 −19.65, −17.95

−0.61 −2.22 −0.61

−1.05, −0.17 −2.31, −2.13 −0.68, −0.53

−20.75 −75.71 −20.62

−35.61, −5.89 −81.44, −69.98 −23.84, −17.40

−0.61 −2.20 −0.60

−0.98, −0.24 −2.28, −2.11 −0.66, −0.54

−20.73 −75.07 −20.39

−33.19, −8.27 −79.61, −70.53 −22.92, −17.87

−0.81 −2.38 −0.58

−0.89, −0.73 −2.41, −2.34 −0.61, −0.56

−25.59 −75.31 −18.47

−28.14, −23.04 −76.95, −73.68 −19.45, −17.49

−0.67 −2.36 −0.58

−0.79, −0.56 −2.40, −2.33 −0.61, −0.56

−21.46 −75.21 −18.54

−25.04, −17.89 −76.62, −73.79 −19.41, −17.67

a Adjusted for demographics (age, marital status, race, visits in the prior year), bisphosphonate, FRAX risk factors (alcohol exposure, BMI, prior fracture with site, rheumatoid arthritis, secondary osteoporosis, smoking status, and corticosteroid use), other drug exposures (calcitonin, calcium, heparin, hormone therapy, proton pump inhibitors, raloxifene, benzodiazepines, opioids, somnolence-causing drugs, selective serotonin reuptake inhibitors, statins, teriparatide, thiazide diuretics, thyroid medications, vitamin D), comorbid conditions (COPD and depression), and 6 months of cost prior to index date. b In 2011 dollars.

low-cost copay for a 3-month supply might still be a hardship for people with very low incomes. Notably, having zoledronic acid as an index bisphosphonate reduced the risk of discontinuation substantially compared to alendronate, suggesting that less frequent dosing dramatically improves persistence because patients are not at risk for discontinuation for much of the followup time. Curtis et al. estimate real-world zoledronic acid adherence to be 20% greater than for IV ibandronate, which is dosed quarterly, and also higher than other studies reporting adherence to weekly or monthly oral bisphosphonates [56]. The subcohort with the highest incidence of both fracture types was the switching subcohort; however, after multivariable adjustment for potential confounders, the risk differences between the non-switching cohort and the change subcohorts were no longer significant. This suggests that most of the difference in risk is driven by patient disease severity, and that more severe patients are more likely to switch. The subcohort that most closely approached the trend of increased risk with inconsistent bisphosphonate use was patients at high risk for hip fracture. Switchers, discontinuers, and reinitiators with a baseline high-risk for hip fracture showed increased (adjusted) risks of 28%, 26%, and 44%, respectively (see Fig. 3b), though not statistically

significantly so. Indeed, the 44% increased risk for reinitiators in the high-risk group may be a reassuring sign that clinicians are identifying high-risk patients and are pressuring them to restart their bisphosphonates after discontinuing them. For cost outcomes, quarterly total healthcare costs were higher in most change behavior subcohorts compared to those who did not change, increases that ranged from $102 to $637 per patient per quarter. In contrast, and given that we did not find higher fracture-related risks and costs in the change subcohorts, osteoporosis-related total costs were all significantly lower in the change subcohorts; however, in terms of dollars, these “savings” were modest (ranging from $10–77 per patient per quarter). The “savings” in osteoporosis-related costs among change subcohorts appeared to be entirely driven by lower osteoporosis-related pharmacy costs; osteoporosis-related pharmacy costs were $19–75 lower per patient per quarter in the change subcohorts compared to non-switchers. Although the numbers seem small, particularly the mean annual osteoporosis-related inpatient costs, it is important to realize that these averages are over the whole cohort, including patients who did not have any inpatient costs at all. For example, out of 35,650 patients in the cohort, only 420 (1.2%) had osteoporosis-related inpatient costs; the average in just that subset was

J. LaFleur et al. / Bone 78 (2015) 174–185

$8153. It should also be noted that the large number of discontinuers in the change cohort likely drives the lower osteoporosis-related pharmacy costs. However, when you look at the other change subcohorts (i.e., switchers and reinitiators), costs go down significantly only in patients who were not high-risk. This may be explained by the overlap of these two subcohorts with the discontinuing subcohort; in the clinical setting, there may be less urgency to keep patients who do not present as high-risk on bisphosphonates. Our results confirm what many other researchers have found: patient adherence to bisphosphonates is poor. We found that 80% discontinued their bisphosphonate medications long before they had completed 5 years of treatment. Similar estimates of discontinuation after 1 year of therapy from other US studies have ranged from 43.3% to 83.8% [2–4,6]. These results indicate that optimizing osteoporosis treatment adherence should be a high priority for the clinical community, and while it may result in increased pharmacy costs, it should be associated with lower total healthcare costs overall. Chronic conditions necessitating persistent preventive medication against future outcome are some of the most difficult conditions to treat, not least because patient compliance tends to drop off over time [57]. Likely the preventive care nature of bisphosphonates reduces the patient's perception of its importance: if nothing happens (i.e., no fracture), then the bisphosphonate may be working, but the patient is unlikely to perceive such. This effect can be compounded if the patient suffers from one of the more immediate side effects of bisphosphonates, such as gastric distress [5]. Although switching had a higher incidence of subsequent fracture compared to discontinuing in our data, it was not statistically significantly associated. Switching bisphosphonates was much less common than discontinuing altogether or restarting the index bisphosphonate. A finding of increased risk associated with switching relative to discontinuing does not align with published literature. Briesacher and colleagues similarly did not find a significant effect on fracture risk in patients who switched from once weekly to once monthly bisphosphonates [21], and others found that switching bisphosphonates improved persistence overall, which would be expected to reduce fracture risk [7,19,20]. Our data also suggest that medication-taking patterns, even discontinuation, do not significantly impact fracture risk, which does not correspond to what others have found [8,12–18,58]. We can propose reasons for our inability to detect a relationship between bisphosphonate discontinuations and fracture risk. One is the low proportion of patients who were non-switchers: only 19%. Considering that only a small percentage of these non-switchers fractured (4.2%), our study was likely underpowered for detecting differences. Relatedly, a significant number of fractures may have not been recorded because many fractures are treated outside the VA [59,60]. Existing research is not illustrative, even stratifying by gap length in the definition of discontinuation (3 months to 7 days) [8,12–15,18]. Notably, most of these studies also used medication possession ratio (MPR), [8,12,14,15, 18] which Curtis [17] shows may inflate the difference between the adherent and non-adherent. Finally, most (approximately 85%) of women were not at high risk for fracture at baseline, which may also explain why adherence did not seem to reduce fractures. Our study covers a time period when many relatively young women with osteopenia being given bisphosphonates. Study limitations This study is subject to a few limitations. First, like all observational studies, our findings are potentially confounded by unmeasured characteristics, despite our best efforts to control for those that were available in our data. These unmeasured characteristics include subsequent incident events—such as new incident prescriptions, new incident diagnoses, or other medical or surgical interventions during the observation period—that might have affected bisphosphonate persistence. Given the long observation period, all these factors might eventually affect treatment behavior. Second, our inclusion of zoledronic acid, ensuring

183

100% persistence for 1 year per dose compared to daily, weekly, or monthly bisphosphonates, may have skewed annual persistence (though zoledronic acid is not shown to improve persistence over multiple years/doses [61,62]) and reduced the sensitivity of our 90-day discontinuation gap length to detect changes in bisphosphonate use. However, the 30- and 60-day gap length sensitivity analyses both indicated that our results are robust. Third, there may be some misclassification of outcomes or exposures, particularly if patients received care outside the VHA system. We tried to minimize this by restricting the analysis to patients who showed a pattern of getting their routine care from within the VHA. For example, since some emergent fractures are more likely to be treated at the nearest facility, rather than at the patient's primary facility, there is a high likelihood that many patients had fractures that were treated outside the VHA. Nelson et al. documented that only 13% of dual-eligible Medicare/VHA patients who fracture are treated in the VHA, based largely on distance from the VHA facility [63]. Veterans living in rural settings are reported to have lower healthcare utilization and quality of life, [64,65] which may in turn impact the baseline health state under which fracture events occur. However, a study by MacKenzie et al. reported mixed results in elderly veterans, suggesting socioeconomic factors rather than rurality as a greater driver of overall mortality [66]. It is unknown how VHA/ Medicare utilization may or may not mediate fracture occurrence in our cohort. Our fracture incidence rates were also 1/3–2/3 lower than would be expected based on clinical trial data in an at-risk population [67], although the low fracture rate may also be explained by the fact that only 15% of all women in our study were classified as high-risk, and less than one-quarter were N 80 years old. With regards to the exposure, the inexpensive prescriptions and mail-order prescription service in the VHA provide an incentive for veterans to receive their refills from VHA pharmacies. However, recent changes in the retail market with respect to inexpensive generics may promote non-system use. Fourth, we may also have inadequate characterization of patient baseline risk level. We only had BMD T-scores from notes and radiology reports in 63% of the sample. For patients who received their BMD screening at a non-VA facility, then that report may have been kept in the medical record as an image file, which was not available to us for processing. Thus, if the clinician did not subsequently mention DEXA results in his/her clinic note, then that would explain the high percentage of missing BMD T-scores. Future work should examine the impact of Medicare eligibility on male osteoporosis screening and treatment patterns in the VHA. It should attempt to detect nuances among the various bisphosphonate users (e.g., the difference between patients who restart bisphosphonates after a discontinuation and patients who discontinue bisphosphonates altogether). It should also attempt to predict bisphosphonate usage patterns, with disease severity (i.e., high-risk and low-risk) as the main independent variable. Finally, it should also explore the reasons for changing bisphosphonates, which may be due in large part to adverse effects. Conclusions This study suggests that real-world variations in adherence and persistence during the long-term treatment of osteoporosis are common. Our data show that most patients who use bisphosphonates discontinue them within 30 months, many sooner. However, patients who discontinued or switched were not at significantly higher risk for fracture compared to non-switchers when controlling for disease severity, which may be explained by the low proportion of high-risk patients in our cohort. Therefore, results should be interpreted as applying primarily to low-risk patients. Total costs were higher in patients with change behaviors. While osteoporosis costs were lower in patients with change behaviors, the magnitudes of the reductions were small and appeared to be primarily driven by the smaller pharmacy costs. Furthermore, getting a patient on an individually optimal bisphosphonate, even if it results in

184

J. LaFleur et al. / Bone 78 (2015) 174–185

switches and discontinuations, may result in lower osteoporosis-related costs in the long term. Supplementary data to this article can be found online at http://dx. doi.org/10.1016/j.bone.2015.04.022. Acknowledgments This material is the result of work supported with resources and the use of facilities at the George E. Wahlen Department of Veterans Affairs Medical Center, Salt Lake City, Utah. The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of the Department of Veterans Affairs or the United States government. This research was funded by a grant from Amgen Inc. The authors would like to thank the Agency for Healthcare Research and Quality (Grant #K08-HS018582-01) for their support of Dr. LaFleur during the writing of this manuscript. The authors would also like to thank Joice Huang and Brad Stolshek for their contributions to the initial research idea. References [1] International Osteoporosis Foundation. The Adherence Gap: Why Osteoporosis Patients Don't Continue with Treatment. Nyon, Switzerland: International Osteoporosis Foundation; 2005. [2] Cramer JA, Amonkar MM, Hebborn A, Altman R. Compliance and persistence with bisphosphonate dosing regimens among women with postmenopausal osteoporosis*. Curr Med Res Opin 2005;21(9):1453–60. [3] Ettinger MP, Gallagher R, MacCosbe PE. Medication persistence with weekly versus daily doses of orally administered bisphosphonates. Endocr Pract 2006;12(5): 522–8. [4] Downey TW, Foltz SH, Boccuzzi SJ, Omar MA, Kahler KH. Adherence and persistence associated with the pharmacologic treatment of osteoporosis in a managed care setting. South Med J 2006;99(6):570–5. [5] Penning-van Beest FJ, Goettsch WG, Erkens JA, Herings RM. Determinants of persistence with bisphosphonates: a study in women with postmenopausal osteoporosis. Clin Ther 2006;28(2):236–42. [6] Weycker D, Macarios D, Edelsberg J, Oster G. Compliance with drug therapy for postmenopausal osteoporosis. Osteoporos Int 2006;17(11):1645–52. [7] Ideguchi H, Ohno S, Hattori H, Ishigatsubo Y. Persistence with bisphosphonate therapy including treatment courses with multiple sequential bisphosphonates in the real world. Osteoporos Int 2007;18(10):1421–7. [8] Gallagher AM, Rietbrock S, Olson M, van Staa TP. Fracture outcomes related to persistence and compliance with oral bisphosphonates. J Bone Miner Res 2008; 23(10):1569–75. [9] Cotte FE, Fardellone P, Mercier F, Gaudin AF, Roux C. Adherence to monthly and weekly oral bisphosphonates in women with osteoporosis. Osteoporos Int 2010; 21(1):145–55. [10] Burden AM, Paterson JM, Solomon DH, Mamdani M, Juurlink DN, Cadarette SM. Bisphosphonate prescribing, persistence and cumulative exposure in Ontario Canada. Osteoporos Int 2012;23(3):1075–82. [11] Hadji P, Claus V, Ziller V, Intorcia M, Kostev K, Steinle T. GRAND: the German retrospective cohort analysis on compliance and persistence and the associated risk of fractures in osteoporotic women treated with oral bisphosphonates. Osteoporos Int 2012;23(1):223–31. [12] Siris ES, Harris ST, Rosen CJ, Barr CE, Arvesen JN, Abbott TA, et al. Adherence to bisphosphonate therapy and fracture rates in osteoporotic women: relationship to vertebral and nonvertebral fractures from 2 US claims databases. Mayo Clin Proc 2006; 81(8):1013–22. [13] van den Boogaard CH, Breekveldt-Postma NS, Borggreve SE, Goettsch WG, Herings RM. Persistent bisphosphonate use and the risk of osteoporotic fractures in clinical practice: a database analysis study. Curr Med Res Opin 2006;22(9):1757–64. [14] Rabenda V, Mertens R, Fabri V, Vanoverloop J, Sumkay F, Vannecke C, et al. Adherence to bisphosphonates therapy and hip fracture risk in osteoporotic women. Osteoporos Int 2008;19(6):811–8. [15] Curtis JR, Westfall AO, Cheng H, Delzell E, Saag KG. Risk of hip fracture after bisphosphonate discontinuation: implications for a drug holiday. Osteoporos Int 2008; 19(11):1613–20. [16] Martin KE, Yu J, Campbell HE, Abarca J, White TJ. Analysis of the comparative effectiveness of 3 oral bisphosphonates in a large managed care organization: adherence, fracture rates, and all-cause cost. J Manag Care Pharm 2011;17(8):596–609. [17] Curtis JR, Westfall AO, Cheng H, Lyles K, Saag KG, Delzell E. Benefit of adherence with bisphosphonates depends on age and fracture type: results from an analysis of 101,038 new bisphosphonate users. J Bone Miner Res 2008;23(9):1435–41. [18] Sampalis JS, Adachi JD, Rampakakis E, Vaillancourt J, Karellis A, Kindundu C. Longterm impact of adherence to oral bisphosphonates on osteoporotic fracture incidence. J Bone Miner Res 2012;27(1):202–10. [19] Ideguchi H, Ohno S, Takase K, Ueda A, Ishigatsubo Y. Outcomes after switching from one bisphosphonate to another in 146 patients at a single university hospital. Osteoporos Int 2008;19(12):1777–83.

[20] Kertes J, Dushenat M, Vesterman JL, Lemberger J, Bregman J, Friedman N. Factors contributing to compliance with osteoporosis medication. Isr Med Assoc J 2008; 10(3):207–13. [21] Briesacher BA, Andrade SE, Harrold LR, Fouayzi H, Yood RA. Adherence and occurrence of fractures after switching to once-monthly oral bisphophonates. Pharmacoepidemiol Drug Saf 2010;19(12):1233–40. [22] Goldzweig CL, Balekian TM, Rolon C, Yano EM, Shekelle PG. The state of women veterans' health research. Results of a systematic literature review. J Gen Intern Med 2006;21(Suppl. 3):S82–92. [23] Yano EM, Hayes P, Wright S, Schnurr PP, Lipson L, Bean-Mayberry B, et al. Integration of women veterans into VA quality improvement research efforts: what researchers need to know. J Gen Intern Med 2010;25(Suppl. 1):56–61. [24] Shibli-Rahhal A, Vaughan-Sarrazin MS, Richardson K, Cram P. Testing and treatment for osteoporosis following hip fracture in an integrated U.S. healthcare delivery system. Osteoporos Int 2011;22(12):2973–80. [25] Bass E, French DD, Bradham DD, Rubenstein LZ. Risk-adjusted mortality rates of elderly veterans with hip fractures. Ann Epidemiol 2007;17(7):514–9. [26] LaFleur J, Ginter T, Curtis JR, Adler RA, Agodoa I, Stolshek B, et al. A novel method for obtaining bone mineral densities from a dataset of radiology reports and clinic notes. Natural Language Processing in a National Cohort of Postmenopausal Veterans Paper Presented at: Presented at the 2013 Annual Meeting of the American Society for Bone and Mineral Research (ASBMR). Baltimore, Maryland: Baltimore Convention Center; 2013, October 4–7. [27] De Silva L, Ginter T, Forbush TB, Nokes N, Fay B, Mikuls TR, et al. Extraction and quantification of pack-years and classification of smoker information in semistructured medical records: workshop on learning from unstructured clinical text. International Conference on Machine Learning (ICML); 2011 [Bellevue, WA]. [28] Bellows BK, Lafleur J, Kamauu AW, Ginter T, Forbush TB, Agbor S, et al. Automated identification of patients with a diagnosis of binge eating disorder from narrative electronic health records. J Am Med Inform Assoc 2014;21(e1):e163–8. [29] Manning C, Schutze H. Foundations of Statistical Natural Language Processing. Cambridge, MA: MIT Press; 1999. [30] Kong SY, Kim DY, Han EJ, Park SY, Yim CH, Kim SH, et al. Effects of a ‘drug holiday’ on bone mineral density and bone turnover marker during bisphosphonate therapy. J Bone Metab 2013;20(1):31–5. [31] Soroko SB, Barrett-Connor E, Edelstein SL, Kritz-Silverstein D. Family history of osteoporosis and bone mineral density at the axial skeleton: the Rancho Bernardo Study. J Bone Miner Res 1994;9(6):761–9. [32] Robitaille J, Yoon PW, Moore CA, Liu T, Irizarry-Delacruz M, Looker AC, et al. Prevalence, family history, and prevention of reported osteoporosis in U.S. women. Am J Prev Med 2008;35(1):47–54. [33] de Lusignan S, Chan T, Wood O, Hague N, Valentin T, Van Vlymen J. Quality and variability of osteoporosis data in general practice computer records: implications for disease registers. Public Health 2005;119(9):771–80. [34] Cvijetic S, Colic Baric I, Satalic Z. Influence of heredity and environment on peak bone density: a parent–offspring study. J Clin Densitom 2010;13(3):301–6. [35] Wilson RA, Chapman WW, Defries SJ, Becich MJ, Chapman BE. Automated ancillary cancer history classification for mesothelioma patients from free-text clinical reports. J Pathol Inform 2010;1:24. [36] Friedlin J, McDonald CJ. Using a natural language processing system to extract and code family history data from admission reports. AMIA Annu Symp Proc 2006;925. [37] Goryachev S, Kim H, Zeng-Treitler Q. Identification and extraction of family history information from clinical reports. AMIA Annu Symp Proc 2008:247–51. [38] Reggiardo MV, Fay F, Tanno M, Garcia-Camacho G, Bottaso O, Ferretti S, et al. Natural history of hepatitis C virus infection in a cohort of asymptomatic post-transfused subjects. Ann Hepatol 2012;11(5):658–66. [39] Wu IC, Wu CC, Lu CY, Hsu WH, Wu MC, Lee JY. Substance use (alcohol, areca nut and cigarette) is associated with poor prognosis of esophageal squamous cell carcinoma. PLoS One 2013;8(2):e55834. [40] Casey R, Piazzon-Fevre K, Raverdy N, Forzy ML, Tretare B, Carli PM, et al. Casecontrol study of lymphoid neoplasm in three French areas: description, alcohol and tobacco consumption. Eur J Cancer Prev 2007;16(2):142–50. [41] Zhou X, Han H, Chankai I, Prestrud A, Brooks A. Approaches to text mining for clinical medical records. Paper Presented at: 2006 ACM Symposium on Applied Computing. Dijon, France: Bourgogne University; 2006. [42] Uzuner O, Goldstein I, Luo Y, Kohane I. Identifying patient smoking status from medical discharge records. J Am Med Inform Assoc 2008;15(1):14–24. [43] McClung M, Boonen S, Torring O, Roux C, Rizzoli R, Bone H, Benhamou CL, et al. Effect of denosumab treatment on the risk of fractures in subgroups of women with postmenopausal osteoporosis. J Bone Miner Res 2012;27(1):211–8. [44] Boonen S, Adachi JD, Man Z, Cummings SR, Lippuner K, Torring O, et al. Treatment with denosumab reduces the incidence of new vertebral and hip fractures in postmenopausal women at high risk. J Clin Endocrinol Metab 2011;96(6):1727–36. [45] National Osteoporosis Foundation. Clinician's Guide to Prevention and Treatment of Osteoporosis. Washington, D.C.: National Osteoporosis Foundation; 2010 [46] Papaioannou A, Morin S, Cheung AM, Atkinson S, Brown JP, Feldman S, et al. Clinical practice guidelines for the diagnosis and management of osteoporosis in Canada: summary. CMAJ 2010;182(17):1864–73. [47] Warriner AH, Patkar NM, Curtis JR, Delzell E, Gary L, Kilgore M, et al. Which fractures are most attributable to osteoporosis? J Clin Epidemiol 2011;64(1):46–53. [48] Kanis JA, Oden A, Johansson H, Borgström F, Ström O, McCloskey E. FRAX® and its applications to clinical practice. Bone 2009;44(5):734–43. [49] Orr-Walker B, Wattie DJ, Evans MC, Reid IR. Effects of prolonged bisphosphonate therapy and its discontinuation on bone mineral density in post-menopausal osteoporosis. Clin Endocrinol (Oxf) 1997;46(1):87–92.

J. LaFleur et al. / Bone 78 (2015) 174–185 [50] Manning WG, Basu A, Mullahy J. Generalized modeling approaches to risk adjustment of skewed outcomes data. J Health Econ 2005;24(3):465–88. [51] Manning WG, Mullahy J. Estimating log models: to transform or not to transform? J Health Econ 2001;20(4):461–94. [52] Hanley JA, Negassa A, Edwardes MD, Forrester JE. Statistical analysis of correlated data using generalized estimating equations: an orientation. Am J Epidemiol 2003; 157(4):364–75. [53] De Angelis G, Murthy A, Beyersmann J, Harbarth S. Estimating the impact of healthcare-associated infections on length of stay and costs. Clin Microbiol Infect 2010;16(12):1729–35. [54] Mark TL, Gibson TB, McGuigan KA. The effects of antihypertensive step-therapy protocols on pharmaceutical and medical utilization and expenditures. Am J Manag Care 2009;15(2):123–31. [55] Yun H, Curtis JR, Guo L, Kilgore M, Muntner P, Saag K, et al. Patterns and predictors of osteoporosis medication discontinuation and switching among Medicare beneficiaries. BMC Musculoskelet Disord 2014;15(1):112. [56] Curtis JR, Yun H, Matthews R, Saag KG, Delzell E. Adherence with intravenous zoledronate and intravenous ibandronate in the United States Medicare population. Arthritis Care Res (Hoboken) 2012;64(7):1054–60. [57] Yeaw J, Benner JS, Walt JG, Sian S, Smith DB. Comparing adherence and persistence across 6 chronic medication classes. J Manag Care Pharm 2009;15(9):728–40. [58] Cotte FE, Cortet B, Lafuma A, Avouac B, Hasnaoui AE, Fardellone P, et al. A model of the public health impact of improved treatment persistence in post-menopausal osteoporosis in France. Joint Bone Spine 2008;75(2):201–8. [59] Richardson KK, Cram P, Vaughan-Sarrazin M, Kaboli PJ. Fee-based care is important for access to prompt treatment of hip fractures among veterans. Clin Orthop Relat Res 2013;471(3):1047–53.

185

[60] Fleming C, Fisher ES, Chang CH, Bubolz TA, Malenka DJ. Studying outcomes and hospital utilization in the elderly. The advantages of a merged data base for Medicare and Veterans Affairs hospitals. Med Care 1992;30(5):377–91. [61] Lee YK, Nho JH, Ha YC, Koo KH. Persistence with intravenous zoledronate in elderly patients with osteoporosis. Osteoporos Int 2012;23(9):2329–33. [62] Lee S, Glendenning P, Inderjeeth CA. Efficacy, side effects and route of administration are more important than frequency of dosing of anti-osteoporosis treatments in determining patient adherence: a critical review of published articles from 1970 to 2009. Osteoporos Int 2011;22(3):741–53. [63] Nelson S, Del Fiol G, Hanseler H, Crouch B, Cummins M. A dashboard for health information exchange between poison control centers and emergency departments: the poison control center view. Poster Presented at 2014 Annual Meeting of the North American Congress of Clinical Toxicology (NACCT). New Orleans, LA: Sheraton New Orleans Hotel; 2014, October 17–21. [64] Weeks WB, Wallace AE, Wang S, Lee A, Kazis LE. Rural–urban disparities in healthrelated quality of life within disease categories of Veterans. J Rural Health 2006; 22(3):204–11. [65] Weeks WB, Kazis LE, Shen Y, Cong Z, Ren XS, Miller D, et al. Differences in healthrelated quality of life in rural and urban veterans. Am J Public Health 2004;94(10): 1762–7. [66] Mackenzie TA, Wallace AE, Weeks WB. Impact of rural residence on survival of male veterans affairs patients after age 65. J Rural Health 2010;26(4):318–24. [67] Donaldson MG, Palermo L, Ensrud KE, Hochberg MC, Schousboe JT, Cummings SR. Effect of alendronate for reducing fracture by FRAX score and femoral neck bone mineral density: the fracture intervention trial. J Bone Miner Res 2012;27(8): 1804–10.