Original ArticleModeling cumulative dose and exposure duration provided insights regarding the associations between benzodiazepines and injuries
Introduction
In the last few decades, the exponential increase in the use of powerful pharmacological agents has led to increased concerns about drug-induced adverse health events [1]. Premarketing clinical trials provide conclusive information on drug effectiveness, but typically they are underpowered for detecting adverse effects. Carefully designed and analyzed large-scale epidemiological postmarketing surveillance studies are therefore essential to establish the associations between drugs and adverse reactions [1], [2]. In many observational studies of the effects of medications, exposure is limited to rather crude measures such as baseline or current drug use. For example, Collet et al. [3] report that, of the many studies focusing on the associations between aspirin use and incidence of various cancers, only two [4], [5] relied on repeated exposure measurements to investigate the role of the dose and duration of use. In most studies, the exposure measurement was limited to aspirin use at entry [6], [7], [8], [9].
Similarly, most postmarketing studies of adverse drug effects are limited to estimating relative risks for drug users, compared to nonusers [10], [11], [12], [13], [14]; however, both current exposure status and current dose of a drug may change frequently and substantially during follow-up [3], [15], [16], [17]. Moreover cumulative measures of exposure, such as total duration of drug use or cumulative dose, by definition, do vary over time. Taking into account such changes over time is important because biological models as well as empirical data indicate that both duration of use and medication dose may affect the outcomes [18], [19], [20], [21]. In such cases, an appropriate representation of a time-dependent exposure may be necessary to avoid underestimating its effect and reducing statistical power [22], [23]. For example, it has been suggested that the discrepant findings regarding the impact of oral contraceptives on breast cancer risk might simply reflect between-studies differences in the duration of use [24].
Little is known, however, about the way complex drug exposure patterns should be represented in epidemiological postmarketing studies. Indeed, recent comprehensive reviews of methodological issues in the evaluation of the safety of medications do not discuss how to model the effects of dose and/or duration of use [1], [2]. Nonetheless, efficient modeling of time-varying exposures is especially important in studies of adverse effects of medications where statistical power is a concern because outcomes are rare and associations are weak [1]. Moreover, an accurate estimation of the effects of quantitative exposure measures is important for clinical management even if the use of a given medication, with a proven therapeutic effect, has been already associated with an increased risk of a particular adverse reaction. For example, knowledge about how the risks change with increasing dose and/or duration of use may promote prescribing that optimizes the trade-off between therapeutic benefits and reduced risk of adverse effects.
Here we address some methodological issues related to a comprehensive assessment of the impact of various aspects of exposure to medication. First, we propose new methods for modeling the impact of cumulative dose and duration of drug use with time-dependent covariates. We then illustrate their ability to yield new insights in real-life analyses that reassess the impact of selected benzodiazepines on the risk of injury.
Benzodiazepines are sedative–hypnotics that are used for a variety of disorders, including anxiety and insomnia [25], [26], and are commonly prescribed to the elderly [27], [28], [29]. Elderly patients are more likely to experience adverse effects of benzodiazepines, such as fall-related injuries [25], [30], [31], [32], [33], [34], but findings regarding the strength of the association have not been consistent [35], [36], [37], [38], [39], [40]. In these studies, classification of exposure has become more sophisticated with time. Earlier studies simply looked at any benzodiazepine use; later studies compared products with similar elimination half-lives or clinical indications [35], [36], [40], [41], [42], [43]. The most recent studies assessed individual benzodiazepines, and suggested that their adverse effects may be quite different [37], [39]; however, the few studies of individual benzodiazepines that accounted for current dose found that these differences may have simply reflected systematic differences between the typical doses of different products [38].
We recently reported that patterns of benzodiazepine use in the elderly are complex and vary across products [15]. For individual subjects, the number of distinct periods of use, during 5 years of follow-up, varied from 1 to 32. For 45% of new users, the average duration of uninterrupted use exceeded the recommended maximum of 30 days [25], [44]. Moreover, 26% of the subjects with several periods of benzodiazepine use had increased their doses over time, and for 31% the doses decreased [15]. Most previous analyses, however, ignored these variations in the duration and/or dose of benzodiazepines, and were restricted to estimating the effects of current use and/or current dose [36], [37], [39], [40], [43].
Section snippets
Data source
In a previous study, we used information from the Quebec Health Insurance Program (RAMQ) and four population health administrative databases, linked through encrypted health care numbers, to assemble a large cohort of new elderly benzodiazepine users [38]. All databases were linked through an encrypted, unique health care number for each subject. The validity of these databases was established previously [45], [46], [47]. The patient demographics database provided age, sex, and date of death.
Results
Table 1 compares the baseline characteristics, follow-up duration, and reasons for early termination of follow-up across the users of the three benzodiazepines. There are no marked differences in the distributions of age and sex, nor in the frequency of previous injury. There are, however, some differences in the mean duration of follow-up. These differences are partly due to differences in the proportion of users of a particular benzodiazepine who were censored at the time they switched to
Discussion
We have proposed methods for representing and modeling complex time-varying medication exposures. Based on both theoretical considerations and limited empirical evidence [18], [22], we expected that carefully designed measures of cumulative dose and/or duration of exposure might enhance the investigators' ability to detect even weak associations between medication use and adverse health outcomes, and possibly advance the understanding of the underlying mechanisms. We then used cohorts of new
Acknowledgments
The collection of data for this study was funded by National Health Research Development Program (File 952249, R. Tamblyn) and the analyses were supported by a grant from the Réseau FRSQ sur l'utilisation des médicaments and by an operating grant from the Natural Sciences and Engineering Research Council of Canada (NSERC) awarded to M. Abrahamowicz. G. Bartlett was supported by a Doctoral Award from the Canadian Institutes for Health Research (CIHR). M. Abrahamowicz is a James McGill Professor
References (65)
- et al.
Principles of epidemiological research on adverse and beneficial drug effects
Lancet
(1998) - et al.
Adverse effect of anticonvulsants on efficacy of chemotherapy for acute lymphoblastic leukaemia
Lancet
(2000) - et al.
Hypersensitivity reactions associated with exposure to naproxen and ibuprofen: a cohort study
J Clin Epidemiol
(2001) - et al.
Occurrence and risk factors of oral candidiasis treated with oral antifungals in seniors using inhaled steroids
J Clin Epidemiol
(2000) - et al.
Oral contraceptive use and breast cancer in young women: A joint national case-control study in Sweden and Norway
Lancet
(1986) Problems with benzodiazepines in elderly patients
Mayo Clinic Proc
(1993)- et al.
Potential adverse outcomes of psychotropic and narcotic drug use in Canadian seniors
J Clin Epidemiol
(1997) - et al.
Using medical services claims to assess injuries in the elderly: sensitivity of diagnostic and procedure codes for injury ascertainment
J Clin Epidemiol
(2000) - et al.
The use of prescription claims databases in pharmacoepidemiological research: the accuracy and comprehensiveness of the prescription claims database in Quebec
J Clin Epidemiol
(1995) - et al.
Information theoretic criteria in non-parametric density estimation: bias and variance in the infinite dimensional case
Comput Stat Data Anal
(1991)