Design, Setting, and Subjects
We performed a retrospective, longitudinal cohort study of electronic health records (EHR) and linked radiology records for women aged 40 to 85 years who received primary care at University of California, Davis Health System (UCDHS) clinics from 1 January 2006 through 31 December 2012. UCDHS includes an academic medical center in central Sacramento and a large physician group offering community-based primary care in 13 clinics across the Sacramento region. The institutional review board of the University of California, Davis approved the study.
We identified annual samples of women based on these inclusion criteria: 1) age 40–85 years on 1 January 1st of the study year; 2) one or more primary care or obstetrics and gynecology (OB/GYN) visits during the study year; 3) no DXA test in a prior calendar year; and 4) no prior osteoporosis diagnosis or medication prescription for osteoporosis drugs (including bisphosphonates, raloxifene, teriparatide, calcitonin, denosumab, but not including estrogens, calcium, or vitamins). The EHR and radiology database were searched back to 2002 for evidence of prior DXA use and osteoporosis diagnosis and medications. Women were eligible for inclusion in multiple consecutive study years including the year they received DXA screening, but were excluded in years following DXA screening or when they ceased receiving primary care within the health system (e.g., due to transfer of care or death). However, we included women during years without primary care or OB/GYN visits if these years were preceded and followed by years with visits. When women had two or more consecutive years without primary care or OB/GYN visits followed by a subsequent year with primary care or OB/GYN visits, we included women in the first year following the two-year gap in visits, so that study data reflected the most recent period of continuous observation time.
Incident DXA screening was defined as a DXA screening test that was completed and reported in the radiology records during the study period. We used this measure to determine the cumulative incidence of screening by age and risk factor status. DXA screening in UCDHS is completed at the central academic campus and one community-based radiology site. Women at some UCDHS primary care sites also complete ordered DXAs at outside radiology facilities. Primary care clinics routinely file outside radiology reports for scanning into the EHR. When outside DXA results are scanned into the EHR, the EHR signifies that the ordered DXA test was completed.
Osteoporosis Risk Factors
Based on risk factors in the Fracture Risk Assessment Tool (FRAX),12 we identified from the EHR several osteoporosis risk factors. Age was determined as of January 1st of each study year, although baseline age (on January 1st of the first year of eligibility) was used for some stratified analyses. Smoking status was determined for each study year using social history information collected routinely by staff during outpatient encounters. If smoking status was not documented during or prior to the study year, the earliest recorded smoking status was used. Height and weight were used to calculate body mass index (BMI) as of January 1st of each study year. From pharmacy data, we collected glucocorticoid prescription information; we considered women to be glucocorticoid users if they received one or more prescriptions for a glucocorticoid during the study period. In a sensitivity analysis, we used an alternative specification based on average yearly glucocorticoid dosage. Using International Classification of Diseases, Ninth Edition, Clinical Modification (ICD-9-CM) codes given in Appendix A, we categorized women by the following diagnoses: possible secondary osteoporosis, previous high-risk fracture, rheumatoid arthritis, and alcohol abuse. We also created a binary indicator of whether the patient had one or more of the following six risk factors: BMI < 20, glucocorticoid use, possible secondary osteoporosis, previous high-risk fracture, rheumatoid arthritis, or alcohol abuse.
Sociodemographics and Healthcare Utilization
We collected information on race/ethnicity from the EHR, which includes predefined race and ethnicity categories. Since 2010, providers and office staff have been prompted to enter race/ethnicity information during office visits. For women with more than one race/ethnicity documented in EHR, we classified women using the following exclusive hierarchical categories: Hispanic, Asian, Black, other race/ethnicity, and White. If race/ethnicity was not listed, women were classified as unknown.
We categorized women into the following exclusive hierarchical categories based on their primary insurance: non-Medicare or non-Medicaid preferred provider organization/health maintenance organization (PPO/HMO), Medicare, Medicaid, other (e.g., CHAMPUS, county indigent health program, Workers’ Compensation), and unknown.
As proxy measures of patient comorbidity and predisposition to use healthcare, we constructed several healthcare utilization variables, including: 1) counts of primary care (family medicine, internal medicine and OB/GYN) and specialist visits during each year, and 2) binary indicators of whether women had OB/GYN or endocrinologist visits during each year. We included a count of yearly hospitalizations as a proxy measure of comorbidity.13 We also measured whether women obtained a screening mammogram during each study year, as mammography use may reflect underlying attitudes regarding preventive health service use.
Data Quality Assessment
The accuracy of all EMR-derived variables was assessed by serial review of a random sample of approximately 250 medical records by a physician investigator (either ADA or JJF). During review, study physicians compared variables abstracted from the EHR to variables defined by manual chart review. Discrepancies were discussed with EHR programming staff, and abstraction algorithms were modified until automated and manual abstractions achieved 97.5 % concordance.
Analyses were performed in SAS (version 9.3, Cary, NC). We computed descriptive statistics of the overall sample and women who received incident DXA. We determined unadjusted incidence rates (and incidence rate differences) for women categorized into the following groups by baseline age: < 50 years (when screening is uncommonly indicated), 50–59 years (encompassing the average age of menopause), 60–64 years (when screening may often be recommended based on risk), 65–74 years (when screening is recommended), and 75 years and older (when screening is recommended if previously unscreened). We performed Cox proportional hazards regression to estimate hazards ratios (HR) of incident DXA as a function of fixed and time-varying patient-level covariates and calendar year. In Cox models, patient age was considered a time-varying covariate. Because many osteoporosis risk factors may change from year to year, the Cox model was specified with time-varying covariates. In a first model (Model 1), we modeled incident DXA as a function of sociodemographics, smoking status, health insurance, healthcare utilization, and each osteoporosis risk factor included as separate covariates. We then repeated the regression analysis using the binary risk factor covariate along with other covariates (Model 2). We performed a sensitivity analysis to assess whether there was a glucocorticoid dosage effect on the hazard ratio incident DXA screening. In this analysis, we categorized women based on average yearly glucocorticoid dose (in prednisone equivalents): none, 1–199 mg, 200–499 mg, 500–1499 mg, and ≥ 1500 mg. For each model, we examined the proportional hazards assumption graphically and statistically, and found no statistically significant evidence of violations. We used the Kaplan-Meier estimator to estimate the 3-year, 5-year, and 7-year cumulative incidences of DXA screening by baseline age and baseline risk factor status.