Mortality and concurrent use of opioids and hypnotics in older patients: A retrospective cohort study

Background Benzodiazepine hypnotics and the related nonbenzodiazepine hypnotics (z-drugs) are among the most frequently prescribed medications for older adults. Both can depress respiration, which could have fatal cardiorespiratory effects, particularly among patients with concurrent opioid use. Trazodone, frequently prescribed in low doses for insomnia, has minimal respiratory effects, and, consequently, may be a safer hypnotic for older patients. Thus, for patients beginning treatment with benzodiazepine hypnotics or z-drugs, we compared deaths during periods of current hypnotic use, without or with concurrent opioids, to those for comparable patients receiving trazodone in doses up to 100 mg. Methods and findings The retrospective cohort study in the United States included 400,924 Medicare beneficiaries 65 years of age or older without severe illness or evidence of substance use disorder initiating study hypnotic therapy from January 2014 through September 2015. Study endpoints were out-of-hospital (primary) and total mortality. Hazard ratios (HRs) were adjusted for demographic characteristics, psychiatric and neurologic disorders, cardiovascular and renal conditions, respiratory diseases, pain-related diagnoses and medications, measures of frailty, and medical care utilization in a time-dependent propensity score–stratified analysis. Patients without concurrent opioids had 32,388 person-years of current use, 260 (8.0/1,000 person-years) out-of-hospital and 418 (12.9/1,000) total deaths for benzodiazepines; 26,497 person-years,150 (5.7/1,000) out-of-hospital and 227 (8.6/1,000) total deaths for z-drugs; and 16,177 person-years,156 (9.6/1,000) out-of-hospital and 256 (15.8/1,000) total deaths for trazodone. Out-of-hospital and total mortality for benzodiazepines (respective HRs: 0.99 [95% confidence interval, 0.81 to 1.22, p = 0.954] and 0.95 [0.82 to 1.14, p = 0.513] and z-drugs (HRs: 0.96 [0.76 to 1.23], p = 0.767 and 0.87 [0.72 to 1.05], p = 0.153) did not differ significantly from that for trazodone. Patients with concurrent opioids had 4,278 person-years of current use, 90 (21.0/1,000) out-of-hospital and 127 (29.7/1,000) total deaths for benzodiazepines; 3,541 person-years, 40 (11.3/1,000) out-of-hospital and 64 (18.1/1,000) total deaths for z-drugs; and 2,347 person-years, 19 (8.1/1,000) out-of-hospital and 36 (15.3/1,000) total deaths for trazodone. Out-of-hospital and total mortality for benzodiazepines (HRs: 3.02 [1.83 to 4.97], p < 0.001 and 2.21 [1.52 to 3.20], p < 0.001) and z-drugs (HRs: 1.98 [1.14 to 3.44], p = 0.015 and 1.65 [1.09 to 2.49], p = 0.018) were significantly increased relative to trazodone; findings were similar with exclusion of overdose deaths or restriction to those with cardiovascular causes. Limitations included composition of the study cohort and potential confounding by unmeasured variables. Conclusions In US Medicare beneficiaries 65 years of age or older without concurrent opioids who initiated treatment with benzodiazepine hypnotics, z-drugs, or low-dose trazodone, study hypnotics were not associated with mortality. With concurrent opioids, benzodiazepines and z-drugs were associated with increased out-of-hospital and total mortality. These findings indicate that the dangers of benzodiazepine–opioid coadministration go beyond the documented association with overdose death and suggest that in combination with opioids, the z-drugs may be more hazardous than previously thought.

Use of non-study benzodiazepines in the past year disqualified patients for cohort entry. These included clorazepate, chlordiazepoxide, chlordiazepoxide-amitriptyline, diazepam, and oxazepam, but did not include the amnestic midazolam.
We defined current hypnotic use as the day of the prescription fill through the end of the days of supply, offset by one day because hypnotics generally are taken in the evening and thus hypnotic-related deaths should occur either during the night or on the next day. If the patient entered the hospital, current use status was frozen for the hospital stay. If the hospital stay was no more than 7 days, we extended the days of supply by the length of stay to account for the patient receiving the medication in the hospital. For longer hospital stays, we assumed there was a greater likelihood that the prescription would be changed in the hospital and thus set the status after discharge to noncurrent use, unless a prescription was filled on discharge.

B. Study Medications: Opioids
The study opioids (Table B) excluded parenteral opioids (infrequently prescribed for outpatients) and preparations specifically formulated for cough or diarrhea. Although methadone is prescribed as opioid replacement therapy, it also is frequently used for pain and thus was included in the study opioids. Although patients with baseline use of buprenorphine were excluded, use during followup was not. Opioids were classified as short-or long-acting and dose-equivalents were calculated in morphine-milligram equivalents (MME) according to guidelines for chronic opioid therapy for non-cancer pain (Table B). 4 Appendix Table B. Study opioid analgesics, with MME conversion factors a . These do not include the antitussive levopropoxyphene, cough preparations, tincture of opium for diarrhea, and hydrocodone in cough preparations, both tablet (e.g., Hycodan) and syrup/liquid. The study Medicare data reside in the Center for Medicare & Medicaid Services (CMS) Chronic Condition Warehouse. The data may be obtained by applying to the Research Data Assistance Center (RESDAC) at the University of Minnesota in the U.S. The process begins with contact to RESDAC, either by visiting the website (https://resdac.org/), phone (1-888-973-7322), or email (resdac@umn.edu). RESDAC has detailed descriptions of the data and procedures for obtaining access, with staff available to assist researchers.

D. Severe Illnesses
Because the primary study endpoint is death out of the hospital, we excluded patients for whom such deaths were likely to be related to life-threatening pre-existing illnesses. This is particularly important for hypnotics because failure to adequately control for the life-threatening illnesses that both lead to hypnotic prescriptions and increased risk of death has led most to discount the strong association of hypnotics with increased risk of death. 1,5 Appendix The cohort excluded patients with the life-threatening illnesses listed in Table D. These diseases were identified from medical care encounters in the past year, including hospitalizations, emergency department (ED) and outpatient visits, and in some cases medication prescriptions. A single hospital discharge or ED visit was sufficient for exclusion, as was a single outpatient encounter on t0-1, the day of the hypnotic prescription fill (hypnotics occasionally started on the day of diagnosis of a severe illness). However, for [t0-365, t0-2], at least two outpatient encounters on separate days were required.

E. Followup person-time
All study use and followup is restricted to current use of study hypnotics. Fig A provides

F. Endpoints
In-hospital deaths were those that occurred during the hospital stay. For patients who were discharged to a skilled nursing facility (SNF), this period included the subsequent SNF stay.
Deaths were classified according to the underlying cause of death, as specified in Table E. Appendix

G. Covariates
Appendix

H. Propensity Score
Covariates. Because patient characteristics and comorbidity were likely to vary according to both specific hypnotic and opioid use, the analysis controlled for numerous covariates. There were two types of covariates, those that described patient characteristics and comorbidity and those related to current opioid use. The covariates. The patient characteristics included 101 covariates plausibly associated with both risk of death and the use of specific hypnotics or opioids (Appendix Table F). They were defined from medical care encounters in the preceding year and included psychiatric and neurologic disorders, cardiovascular and renal conditions, respiratory diseases, pain-related diagnoses and medications, measures of frailty, and medical care utilization. The selection and definition of covariates was primarily based on our previous studies of out-of-hospital death, 6,7 which included variables based on standard measures of comorbidity. 8,9 We also reviewed the literature to add several indicators of frailty. 10 Because changes in comorbidity during followup could be associated with changes in the likelihood of concurrent opioids, the values were updated on the date of each hypnotic prescription fill.
The risk of opioid use may vary according to duration of action, time since initiation of treatment, and opioid dose. Thus, we controlled for four opioid-related covariates.
Long-acting opioid, see Appendix Table B. New use of opioids, defined as initiation within the past 90 days. Intermediate dose, defined as 30-59 morphine-milligram equivalents (Appendix Table B). High dose, defined as ≥60 morphine-milligram equivalents (Appendix Table B).
Because these covariates could change on any given day (initiate opioid use during the period of days of supply for a hypnotic), these were updated on a daily basis.
Propensity score definition and rationale. The propensity score is the probability that a patient receives one of two treatment options, conditional on the study covariates. If the propensity score is properly calculated, controlling for the propensity score in the analysis is equivalent to adjusting for all of the variables that go into its calculation. [11][12][13] The key advantage of propensity scores is the ability to control for much larger numbers of covariates than traditional multivariate methods, which require 7-10 endpoints for every variable in the model. As Haukoos and Lewis note in a review: 13 "Propensity score methods generally allow many more variables to be included in the propensity score model, which increases the ability of these approaches to effectively adjust for confounding, than could be incorporated directly into a multivariable analysis of the study outcome." Assessing balance. A properly formulated propensity score is a balancing score, that is, the distribution of the covariates conditional on the propensity score is the same in both the treated and untreated groups. Balance is measured by examining the distribution of the covariates in each treatment group after inverseodds-of-treatment (IOT) weighting a and calculating the standardized difference, 12 with a difference of less than 10% considered good balance. 13 Propensity score calculation. Because both the use of opioids and patient characteristics could change during followup, the study propensity score was time-dependent, with covariate values updated at the time of each study hypnotic prescription fill. 14 Three propensity scores were calculated for each of the pairwise comparisons of study hypnotics: trazodone vs benzodiazepines, trazodone vs z-drugs, and benzodiazepines vs z-drugs. The propensity score was estimated with logistic regression models with SAS version 9 PROC LOGISTIC. These included all of the covariates in Appendix Table F except for the opioid use characteristics covariates, which were included directly in the proportional hazards regression. Because the factors that predicted individual hypnotic use might differ depending on opioid use, separate regression models were fit for users and nonusers of opioids, which assured covariate balance for each of these groups.

Use in analysis.
There are four standard methods for analysis with propensity scores. Cohort members can be matched according to the propensity score, each observation can be assigned a weight according to the propensity score value, the analysis can be stratified by quantiles of the propensity score, or the propensity score can be included in a regression model for the study endpoint. [11][12][13] The first three methods are the most commonly used because regression modelling requires specifying the functional form of the relation between the propensity score and the endpoint, which depends on additional assumptions.
We considered both matching and weighting, as these methods are the least subject to residual confounding. 12 However, matching is both more complex with time-dependent exposures and reduces study power by discarding observations. Weighted analyses typically use weights inversely proportional to the propensity score, which commonly leads to large weights that inflate variances and consequently reduce power. Thus, the primary analysis used stratification by the deciles of the time-dependent propensity score. To assess the potential for residual confounding, sensitivity analyses were performed with both matching and weighting.
The strata were defined according to the distribution of the propensity score in the trazodone group. This procedure estimates the average treatment effect in the treated, the same estimate that would result from matching each trazodone user with either a benzodiazepine or z-drug user. 12 We did not include the opioid-related covariates in the propensity score because these could change on any day of followup whereas the patient characteristics were updated only at the time of each prescription fill. Our rationale was that the patient comorbidity changes during the span of a prescription (for example, a fall or a diagnosis of dementia) could be precursors of an endpoint. Furthermore, the theory on which use of a timedependent propensity score is based 14 assumes covariates are updated no more frequently than at each prescription fill. Thus, the opioid-related covariates were controlled for directly in the regression model. Table F shows the distribution of the covariates and the standardized differences (relative to trazodone), with standardized differences greater than 10% for many of the covariates. Appendix Table G shows the distribution and standardized differences after IOT weighting; all differences were <2%, indicating good balance. Table G. Study covariates at baseline according to hypnotic class and baseline opioid use, weighted according to inverse odds of treatment. Opioid-related covariates, which varied on each person-day of followup, were not included in the propensity score because they were directly controlled for in the analysis.

Primary Analysis
The primary analysis was stratified according to deciles of a time-dependent propensity score, with the covariates (excluding those characterizing prescribed opioids) updated at the time of each hypnotic prescription fill. The opioid characteristics were updated for each person-day of followup.
The following SAS program template shows the regression model: PROC PHREG; STRATA PS_STRATA; /* note there is a separate variable for each pairwise hypnotic comparison*/ CLASS HYPOPIOID; /* Design variable that specifies the hypnotics and opioid use, reference trazodone, either with or without opioids according to the comparison */ MODEL (ta,tb)* DEATH(0) = LAOPIOID NEWOPIOID DOSE30TO59 DOSE60PLUS HYPOPIOID; /* SAS counting process syntax, where ta is the day prior to the current day of followup, tb is the current followup day, DEATH is the endpoint (either out-of-hospital or all deaths), and HYPOPIOID is the design variable*/ /* the variables LAOPIOID NEWOPIOID DOSE30TO59 DOSE60PLUS characterize the opioid exposure */ RUN;

Sensitivity Analyses
Clustering by region: variance adjustment. In this analysis, the 5 geographic regions of the U.S. b were considered as a clustering factor, that is, the outcomes for cohort members could be correlated within clusters. Thus, we modified the proportional hazards analysis to do robust variance estimation, as described by Donner and Klar. 15 Clustering by region and control for region. In this analysis we controlled for factors not included in the Medicare data that may vary between regions, such as income or education. Thus, the model included a term for each of the regions.
Covariates fixed at baseline. In the primary analysis, the propensity score was updated at the time of each prescription fill. In this sensitivity analysis, the propensity score was fixed at baseline and the analysis was stratified by deciles of baseline propensity score. Opioid use and related variables updated for each day of followup.
Time-dependent propensity score weights. Because the propensity score deciles leave open the possibility of residual confounding, this sensitivity analysis utilized all of the information available in the propensity score. The proportional hazards regression was weighted by time-dependent matching weights, 16 which were updated at the time of each prescription fill during followup, with robust variance estimation to account for the dependencies induced by weighting. Matching weights, bounded by 0 and 1, produce estimates with lower variance than inverse probability/odds of treatment weights and in populations with good propensity score overlap between the treated/control groups the estimates are asymptotically equivalent to the ATT. 16 Opioid use and related variables were updated for each day of followup.