Response to Drs. Kripke and Cervoni
We specifically designed our study to address some concerns about previous investigations1,2, and felt it was critical to control for a wide array of confounding variables, to not measure covariate values after treatment initiation, and to not censor subjects differentially between exposure groups. Extensive consideration of confounding was needed because benzodiazepines (BZDs) are commonly used at the end of life. Also, the focus on BZDs and on a middle-aged insured population were deliberate choices to further control potential confounding.
Drs. Kripke and Cervoni question our use of an intention-to-treat approach. In the context of a non-user comparator, a per-protocol approach would have led to a differential opportunity for BZD users and non-users to be censored (i.e., the likelihood of discontinuing BZD treatment among BZD initiators is much higher than the likelihood of starting BZD treatment among nonusers) and, thus, to differential follow-up and potential informative censoring. Consistent with the intention-to-treat approach, we did not classify patient exposure at cohort entry on the basis of treatment use and variations of use (e.g., duration and dose) occurring during the follow-up, as done in previous research by Dr. Kripke1 and others2. An exposure definition based on future information in terms of treatment use and variations of use may lead to spurious findings3. In the setting of time-varying treatments, standard statistical methods, as used in previous literature by Dr. Kripke1 and others2, fail to produce valid estimates. In such a context, marginal structural models should be the analytical strategy of choice4. To address the concerns related to dose and duration of use during follow-up in a valid manner, yet without completely changing our analytical approach, we implemented an exploratory hybrid approach that showed no evidence of a dose/duration effect associated with the use of BZDs (see results on page 10 of the authors’ response to reviewers, http://www.bmj.com/sites/default/files/attachments/bmj-article/pre-pub-h...). As a final remark, a similar intention-to-treat approach has also been used in a recent literature reporting an over three-fold increased risk of all-cause mortality among adult populations exposed to BZDs.2 Thus, the null finding in our study, compared to previous studies, cannot be explained by the choice of an intention-to-treat analytic approach, but it is likely explained by an exposure definition not based on future information, a more careful selection of the non-user comparators, and a better confounding control (our analyses accounted for over 300 potential confounding variables).
We recognize the potential for residual confounding by characteristics unmeasured in administrative databases (e.g., illicit use of medications), but note that residual confounding is unlikely to explain our null finding since such factors are expected to be more prevalent among BZD users than non-users. Concerns regarding residual confounding should be recognized in previous investigations that accounted for a significantly more limited number of potential confounders (less than 20) and showed an over three-fold increased risk of all-cause mortality among adult populations exposed to BZDs1,2.
Finally, in response to Dr. Cervoni concerns we note that our conclusions reflect the findings from this investigation, and specifically refer to all-cause mortality. We recognize that BZDs are a frequently abused class of medications and may be involved in drug poisoning suicides, but the focus of our investigation was on overall death rates in a broad population, and suicides or overdose mortality are a very small percentage of overall deaths.
1. Kripke DF, Langer RD, Kline LE. Hypnotics’ association with mortality or cancer: a matched cohort study. BMJ Open 2012;2:e000850.
2. Weich S, Pearce HL, Croft P, Singh S, Crome I, Bashford J, Frisher M. Effect of anxiolytic and hypnotic drug prescriptions on mortality hazards: retrospective cohort study. BMJ 2014;348:g1996
3. Suissa S. Immortal time bias in pharmaco-epidemiology. Am J Epidemiol 2008;167:492–499
4. Robins JM and Hernan MA. Estimation of the causal effects of time-varying exposures. In Longitudinal Data Analysis. Ed by Fitzmaurice G, Davidian M, Verbeke G, Molenberghs G.
Competing interests: No competing interests