Benzodiazepines and risk of all cause mortality in adults: cohort study

Sources of data

We collected data from Optum Clinformatics Datamart (OptumInsight, Eden Prairie, MN), a large US commercial insurance database covering more than 14 million people annually from all 50 US states. For each participant, the dataset contains demographic information, health plan enrollment status, inpatient and outpatient medical encounters (coded using the international classification of diseases, ninth revision, clinical modification, and current procedural terminology, fourth edition), and drugs filled on an outpatient basis, including the national drug code numbers, quantity dispensed, and days’ supply. Data were available from January 2004 through December 2013.

Study population, medication use, and outcome definition

The exposed study population included patients aged 18 years and older who started a benzodiazepine (see eTable 1 in supplementary appendix for a list of benzodiazepines) between 1 July 2004 and 31 December 2013, and had six months of continuous health plan enrollment before drug initiation—that is, the index date for patients who start benzodiazepines (fig 1⇓). We defined benzodiazepine initiators as those who had no use of any benzodiazepine in the previous six months. Patients were not allowed to enter the study cohort more than once. We opted for a primary non-active comparator group because the entire benzodiazepine class has been previously associated with an increased mortality risk, which precludes the use of a within class comparator, and because benzodiazepines have a wide range of indications and uses, which makes the selection of an appropriate out-of-class comparator group challenging. Non-users can have quite different characteristics from benzodiazepine initiators, such as a lower burden of comorbidities, and thus a lower mortality risk, or conversely higher barriers to treatment and surveillance for comorbidities, and thus a higher mortality risk23; both scenarios will need to be accounted for through proper study design or analysis choices, or both. To select patients who did not start benzodiazepine treatment with similar opportunity to be evaluated and treated by a physician as patients who started benzodiazepines and within a similar time window, for each benzodiazepine user we identified a random patient who had a medical visit within 14 days either way of the treatment start date for the corresponding benzodiazepine user and fulfilled the same inclusion criteria as benzodiazepine new users—that is, six months of continuous health plan enrollment before the selected medical visit and no use of any benzodiazepine in the six months before and including the date of the visit (ie, the index date for patients who did not start benzodiazepines). We also required patients who either did or did not start benzodiazepines to have at least one filled prescription in the 90 days before the index date and at least one filled prescription in the previous 91 through 180 days, to balance surveillance between groups and to make sure that patients in both groups were in contact with the healthcare system and had equal treatment opportunity. Starting from the patients with the earliest date for starting benzodiazepines in the study period, we repeated these steps for each patient who started benzodiazepines throughout the study period until the last such patient was identified. For patients contributing to both cohorts of users and non-users of benzodiazepines, we only considered the first cohort entry (see supplementary appendix eFigure 1 for study diagram).

Fig 1 Flowchart of study cohort

• Download figure
• Open in new tab
• Download powerpoint

Follow-up began on the day after the index date for both cohorts. Patients were followed in an intention-to-treat approach until the occurrence of death, nursing home admission, end of continuous health plan enrollment, end of the study period (31 December 2013), or end of the observation period, whichever came first. Consistent with an intention-to-treat approach, we disregarded treatment variations occurring during the follow-up for both cohorts. We considered a six month observation period (180 days) in the main analysis and observation periods of 12 and 48 months of follow-up in sensitivity analyses. A 180 day observation period was chosen for the primary analysis as we empirically observed that in routine care approximately two thirds of patients who start benzodiazepines stop treatment within the first two months.

The study outcome was all cause mortality identified by linkage of the Optum Clinformatics Datamart database with the Social Security Administration Death Master File.24 The specificity of this master file for detecting mortality status has been estimated to range between 97% and 100%.2526

Characteristics of patients

Patient characteristics were identified during the six months preceding and including the index date and included demographic information (age, sex), calendar year, comorbidities and lifestyle factors, use of medications, and indicators of healthcare utilization. We purposely considered a large number of patient characteristics to capture both overt risk factors for mortality and variables that may be proxies of risk factors that were unmeasured or incompletely measured in our dataset. Supplementary appendix eTable 2 shows the complete list of patient characteristics.

Statistical analysis

We cross tabulated patient baseline characteristics by those who started or did not start benzodiazepines, and in both groups we calculated number of mortality events and incidence rates. Cox proportional hazards analyses were used to estimate hazard ratios with 95% confidence intervals for the association between benzodiazepines and all cause mortality. Results are presented for three levels of adjustment: unadjusted, adjusted using 1:1 propensity score matching, to control for measured confounders27; and adjusted using 1:1 high dimensional propensity score matching, to further control for potential residual confounding.28 Exposure propensity scores were derived from the predicted probability of benzodiazepine treatment estimated in a logistic regression model that contained all patient characteristics (see supplementary eTable 2) without additional variable selection. Groups were 1:1 matched on their propensity score using a nearest neighbour matching algorithm without replacement.29 We calculated absolute standardized differences to assess covariate balance before and after propensity score matching.30 Through the high dimensional propensity score algorithm, an automated technique that identifies and prioritizes covariates that may serve as proxies for unmeasured confounders in large electronic healthcare databases,28 we selected a total of 200 empirically identified confounders and combined these with the investigator identified covariates (see supplementary appendix eTable 2) to estimate an empirically enriched propensity score. Stratification on the propensity score matched and high dimensional propensity score matched pair in Cox proportional hazards models to account for potential informative censoring did not change the point estimates appreciably, so this was omitted from all the adjusted analyses.

To test the robustness of our primary findings we performed subgroup and sensitivity analyses. Firstly, we varied the definition of exposure risk window and extended the observation period to 12 and 48 months of follow-up. Secondly, as greater risks for outcomes associated with the hypnotic effects of benzodiazepines have been found among older patients31 and possibly among users of benzodiazepines with longer half lives,3233 we tested the potential for effect modification of benzodiazepines on mortality according to age categories (<65 years and ≥65 years) and duration of action (short acting and long acting, ie, <24 hours half life and ≥24 hours half life) (see supplementary appendix eTable 3).34 For these analyses, we estimated separate propensity scores for each subgroup. Thirdly, to address potential residual confounding associated with the use of a non-active comparator, we explored the use of an active comparator and performed analyses comparing initiators of benzodiazepines with initiators of selective serotonin reuptake inhibitor (SSRI) antidepressants for risk of all cause mortality. We chose SSRIs as the active comparator group because benzodiazepines and SSRIs are both psychotropic drugs and have some overlap in underlying indications, and because SSRIs have no known association with mortality, except for a possible association with increased suicidality in children and young adults.3536

All analyses were done using SAS 9.3 Statistical Software (SAS Institute, Cary, NC).

Patient involvement

No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the study. No patients were asked to advise on interpretation or writing up of results. There are no plans to disseminate the results of the research to study participants or the relevant patient community.