Sedative Hypnotic Use and the Risk of Motor Vehicle Crash
Sedative Hypnotic Use and the Risk of Motor Vehicle Crash
Group Health Cooperative (GH) is an integrated delivery system that covers approximately 600 000 individuals in Washington State. The GH enrollee population closely resembles the underlying community within Washington State with respect to age, race/ethnicity, and gender. Information on health plan enrollment, medical encounters, and pharmacy use are recorded and maintained in electronic medical records and automated databases. These data were linked by a unique consumer number assigned to each enrollee. The prescription drug formulary at GH included 3 sedative hypnotics during the study period: temazepam, trazodone, and zolpidem. Trazodone, an antidepressant medication, was included because of its near universal use at GH for insomnia rather than depression.
We conducted a retrospective population-based cohort study in the GH population. Eligibility criteria for the study were age 21 to 79 years during the study period (2003–2008), primary residence in Washington State with a Washington State driver's license, and at least 18 months of continuous GH enrollment (defined as < 60-day lapse in membership) between January 1, 2003, and December 31, 2008, unless the participant died. We also required that each participant have a drug benefit through GH. Follow-up went through the earliest of end of the study period (December 31, 2008), disenrollment, 80 years of age, or death.
Among the eligible population, we developed a data set linking GH administrative, medical encounter, and pharmacy records with Washington State Department of Licensing driving license records and Department of Transportation motor vehicle crash records. We extracted demographic characteristics, medical encounters, and prescription records for each study participant. We calculated the Charlson Comorbidity Index for each participant by using an adapted algorithm based on International Classification of Diseases, Ninth Revision, Clinical Modification codes from the administrative database. We used the algorithm described by Gallian (using complete name and date of birth) to calculate driver's license numbers for study participants. All police-reported crashes during the study period in which the primary driver was listed as matching a study driver's license number were returned by the Department of Transportation.
We extracted dispensings of sedative hypnotics on the GH formulary (trazodone [Desyrel], temazepam [Restoril], and zolpidem [Ambien, Ambien CR]) during the study period from the GH pharmacy database. Dispensing data included the drug name, date of dispensing, strength, intended days supplied, and National Drug Code.
We defined duration of exposure by collapsing individual dispenses into periods of continuous use (episodes). We defined exposure to sedative hypnotics as a time-varying covariate with the start date defined as the date of dispensing of the first sedative prescription. We assumed an 80% compliance factor (1.25) multiplied by the days supplied for each dispensing to determine the run-out date of the dispensing. We defined the period of continuous use as within the compliance-adjusted run-out date of one dispensing and fill date of the subsequent dispensing. We defined the end date of a continuous episode as the run-out date of the last prescription in the continuous episode. We evaluated only the initial continuous exposure episode for each participant to maintain a new-user study design and avoid healthy user bias. We matched motor vehicle crash records to the periods of exposure and nonexposure.
We implemented a new-user study design by using the first 3 months of eligibility for each participant to identify current sedative use and exclude those users from the study. We then followed the nonuser group to identify new sedative use in the remaining years of the study. We also stratified exposure by the length (days) of continuous sedative prescriptions (1–30, 31–120, 121–240, 241–360, and ≥ 361 concurrent days).
We used Cox proportional hazards regression to estimate the multivariable adjusted hazard ratio (HR) and 95% confidence interval (CI) for sedative medication use with motor vehicle crash as the outcome. This approach applied a nonparametric time-to-event model, which assumed that the instantaneous rate of crash is constantly proportional between the exposed and unexposed groups over time. By employing robust sandwich estimator standard errors in the model, we allowed for nonlinearity of HRs between the exposed and unexposed groups. We tested the proportional hazards assumption with visual assessment of plotted cumulative sums of the Martingale residuals as well as the Kolmogorov-type supremum test using 1000 simulations.
We first defined all 3 sedatives of interest as a composite exposure to estimate the overall association of sedatives with crash risk. We then individually estimated the association for each medication, with linear hypothesis tests for the equality of the estimates. We adjusted all multivariate models for age (quadratic term), gender, calendar year of exposure (categorical), prescription opioid dispensings, and Charlson Comorbidity Index at study entry (categorized into 0, 1, 2, and 3 or greater with 0 serving as the reference group). We considered other prescription medications (antiseizure, antidementia, antidepressant) and medical diagnoses (neurological and cardiovascular disorders) as possible confounders on the basis of a recent National Highway Traffic Safety report but none were found to be associated with both the exposure and outcome and we therefore omitted them from the final models.
We also translated the model-based HR estimates into blood alcohol concentration (BAC) equivalents by using data from Peck et al. To do so, we matched the HRs found in our analyses with the relative risk values in Peck et al. for participants aged 21 to 55 or more years of age and extracted the BAC equivalents that were associated with that relative risk. Although it is not an exact match, this translation provides the relative context necessary for interpreting the risk of crash.
We generated all analyses with SAS software for Windows, version 9.3 (SAS Institute Inc, Cary, NC).
Methods
Group Health Cooperative (GH) is an integrated delivery system that covers approximately 600 000 individuals in Washington State. The GH enrollee population closely resembles the underlying community within Washington State with respect to age, race/ethnicity, and gender. Information on health plan enrollment, medical encounters, and pharmacy use are recorded and maintained in electronic medical records and automated databases. These data were linked by a unique consumer number assigned to each enrollee. The prescription drug formulary at GH included 3 sedative hypnotics during the study period: temazepam, trazodone, and zolpidem. Trazodone, an antidepressant medication, was included because of its near universal use at GH for insomnia rather than depression.
We conducted a retrospective population-based cohort study in the GH population. Eligibility criteria for the study were age 21 to 79 years during the study period (2003–2008), primary residence in Washington State with a Washington State driver's license, and at least 18 months of continuous GH enrollment (defined as < 60-day lapse in membership) between January 1, 2003, and December 31, 2008, unless the participant died. We also required that each participant have a drug benefit through GH. Follow-up went through the earliest of end of the study period (December 31, 2008), disenrollment, 80 years of age, or death.
Data
Among the eligible population, we developed a data set linking GH administrative, medical encounter, and pharmacy records with Washington State Department of Licensing driving license records and Department of Transportation motor vehicle crash records. We extracted demographic characteristics, medical encounters, and prescription records for each study participant. We calculated the Charlson Comorbidity Index for each participant by using an adapted algorithm based on International Classification of Diseases, Ninth Revision, Clinical Modification codes from the administrative database. We used the algorithm described by Gallian (using complete name and date of birth) to calculate driver's license numbers for study participants. All police-reported crashes during the study period in which the primary driver was listed as matching a study driver's license number were returned by the Department of Transportation.
We extracted dispensings of sedative hypnotics on the GH formulary (trazodone [Desyrel], temazepam [Restoril], and zolpidem [Ambien, Ambien CR]) during the study period from the GH pharmacy database. Dispensing data included the drug name, date of dispensing, strength, intended days supplied, and National Drug Code.
We defined duration of exposure by collapsing individual dispenses into periods of continuous use (episodes). We defined exposure to sedative hypnotics as a time-varying covariate with the start date defined as the date of dispensing of the first sedative prescription. We assumed an 80% compliance factor (1.25) multiplied by the days supplied for each dispensing to determine the run-out date of the dispensing. We defined the period of continuous use as within the compliance-adjusted run-out date of one dispensing and fill date of the subsequent dispensing. We defined the end date of a continuous episode as the run-out date of the last prescription in the continuous episode. We evaluated only the initial continuous exposure episode for each participant to maintain a new-user study design and avoid healthy user bias. We matched motor vehicle crash records to the periods of exposure and nonexposure.
We implemented a new-user study design by using the first 3 months of eligibility for each participant to identify current sedative use and exclude those users from the study. We then followed the nonuser group to identify new sedative use in the remaining years of the study. We also stratified exposure by the length (days) of continuous sedative prescriptions (1–30, 31–120, 121–240, 241–360, and ≥ 361 concurrent days).
Statistical Analyses
We used Cox proportional hazards regression to estimate the multivariable adjusted hazard ratio (HR) and 95% confidence interval (CI) for sedative medication use with motor vehicle crash as the outcome. This approach applied a nonparametric time-to-event model, which assumed that the instantaneous rate of crash is constantly proportional between the exposed and unexposed groups over time. By employing robust sandwich estimator standard errors in the model, we allowed for nonlinearity of HRs between the exposed and unexposed groups. We tested the proportional hazards assumption with visual assessment of plotted cumulative sums of the Martingale residuals as well as the Kolmogorov-type supremum test using 1000 simulations.
We first defined all 3 sedatives of interest as a composite exposure to estimate the overall association of sedatives with crash risk. We then individually estimated the association for each medication, with linear hypothesis tests for the equality of the estimates. We adjusted all multivariate models for age (quadratic term), gender, calendar year of exposure (categorical), prescription opioid dispensings, and Charlson Comorbidity Index at study entry (categorized into 0, 1, 2, and 3 or greater with 0 serving as the reference group). We considered other prescription medications (antiseizure, antidementia, antidepressant) and medical diagnoses (neurological and cardiovascular disorders) as possible confounders on the basis of a recent National Highway Traffic Safety report but none were found to be associated with both the exposure and outcome and we therefore omitted them from the final models.
We also translated the model-based HR estimates into blood alcohol concentration (BAC) equivalents by using data from Peck et al. To do so, we matched the HRs found in our analyses with the relative risk values in Peck et al. for participants aged 21 to 55 or more years of age and extracted the BAC equivalents that were associated with that relative risk. Although it is not an exact match, this translation provides the relative context necessary for interpreting the risk of crash.
We generated all analyses with SAS software for Windows, version 9.3 (SAS Institute Inc, Cary, NC).
Source...