Data sources and setting
This study was conducted at Kaiser Permanente Southern California (KPSC), a large managed care system that comprises nearly 15 hospitals and over 200 medical clinics that serve with 4.7 million members. Patients receive virtually all their medical care and prescription drugs within this integrated healthcare delivery system, and information on any outside procedures and diagnoses are available through claims databases. The health plan’s National Cancer Institute-Surveillance, Epidemiology, and End Results (SEER)-affiliated tumor registry was used to identify subjects with breast cancer. The KPSC Institutional Review Board reviewed and approved this study.
Subjects and design
We assembled a cohort of women diagnosed with first primary breast cancer in 2009–2016 followed through December 2017. Eligible women included adults (> 18 years at diagnosis), with American Joint Commission on Cancer TNM Stage 0-III breast cancer, and with at least one year of continuous membership prior to their cancer diagnosis (n = 21,513). We excluded 167 women with a history of fractures in the prior 3 months before breast cancer diagnosis to reduce confounding, leaving 21,346 for analysis.
Incident fractures outcomes
We identified first incident fractures that occurred after breast cancer diagnosis, from the health plan’s electronic health records (EHR). Fractures of the forearm, femur, lower leg, wrist and hand, vertebrae, and pelvis were identified using ICD9 (International Classification of Diseases, Ninth Revision) codes and ICD10 codes (see Supplement Table S1). The event that occurred first served as the outcome.
The pharmacy dispensing database was used to capture use of all sleep aids (name, date of initiation, days supplied) in the health plan’s formulary: lorazepam, trazodone, doxepin, flurazepam, temazepam, triazolam, eszopiclone, zaleplon, zolpidem, and suvorexant. We also considered the effect of “new use” of the sleep medication on risk of fracture. “New use” was defined as the drug dispensing that occurred without drug possession in the preceding 3 months before breast cancer diagnosis.
A comprehensive set of covariates was captured from the EMR. These included race/ethnicity and tumor factors (age and stage at breast cancer diagnosis, adjuvant cancer therapy, diagnosis year, and tumor characteristics). We also captured prior comorbidities from one year before breast cancer diagnosis, and current comorbidities post breast cancer diagnosis through the end of each woman’s follow-up. Current comorbidities ascertained included osteoporosis, bone mineral density, hypertension, depression, dementia, anxiety, and sleep problems. We extracted data on covariate medications such as antidepressant, anti-anxiety, and bisphosphonate use during the follow-up period. Bisphosphonates (alendronate, which was the most commonly used drug in 93% of patients) was prescribed to combat the bone loss typically associated with adjuvant aromatase inhibitor therapy.
Follow-up commenced on the breast cancer diagnosis date and censored on the date of the earliest study endpoint: first fracture diagnosis date, death, termination of health plan membership, or study’s end (December 31, 2017). The definition of continuous health plan enrollment allowed for gaps of up to 3 months in enrollment during study period, as these were likely administrative gaps. In descriptive analyses, we compared the distribution of all variables (demographics, tumor characteristics, comorbidities, and covariate drugs) by fracture status and sleep medication use. The Chi-square or Fisher exact tests were used to compare categorical variables, and the Kruskal–Wallis tests was used for continuous variables. Because women were followed different lengths of time, we computed the person-year rates of fractures. Odds ratios and 95% confidence intervals (CI) were used to compute the association of mental health conditions with use of sleep medications and correlations between sleep medications and psychiatric medications.
Crude and adjusted hazard ratios (HR) and 95% CIs were estimated for fracture risk by Cox proportional hazards models using time-dependent variables for sleep aids and the other covariate medications (antidepressants, anti-anxiety medications, bisphosphonates) used during the follow-up period. We also adjusted for adjuvant endocrine therapy (tamoxifen and aromatase inhibitors treated as time-dependent) in the models. All covariates selected for adjustment in the model were based on clinical importance and descriptive statistics. The proportional hazards assumption was tested via graphic plots and Schoenfeld residuals; no violations were found.
We also conducted additional analyses to assess the robustness of the multivariable results of the association between sleep medication use and fracture risk. First, we evaluated two models stratified by bisphosphonate use status using the full cohort (n = 21,346). Second, we repeated analyses on the subset of women who were “new users” of sleep medications (n = 16,486). Third, we conducted a sensitivity analysis based on bone mass density (BMD) data during study follow-up; this was available on n = 8,498 survivors with spine BMD; if multiple BMD data were available per patient, we use the lowest to be conservative. Except for body mass index (BMI, < 5% was missing), missing data were rare for the study variables. Unknown BMI was entered as a category in the multivariable models (as done for race/ethnicity which had 1% missing in the other/mixed/unknown category). All analyses were performed using SAS Version 9.4 (SAS Institute, Cary NC).