Participants were recruited as part of the Three-City Study, an ongoing multi-site longitudinal study involving three French cities: Bordeaux, Dijon and Montpellier 26 . Briefly, non-institutionalized participants aged 65 years or over were randomly selected from electoral rolls between 1999 and 2001. The acceptance rate was 37%, yielding a sample of 9,294 individuals.

The study protocol was approved by the ethical committee of the University Hospital of Kremlin-Bicêtre and CPP Sud Méditérannée III, and written informed consent was obtained from each participant. The participants were administered standardized questionnaires and underwent clinical examinations at baseline and after 2, 4, 8, 10 and 12 years.

The exact date of death of the participants was obtained from death registries. The causes of death were collected by the local study centers from medical records and interviews with family physicians, clinicians and other non-medical informants (relatives or caregivers) 27 . A validation committee used all information to classify the cause of death using the tenth revision of International Classification of Diseases (ICD-10) 28 as follows: cancer (ICD-10: C00 to C97 and D37 to D48), coronary heart disease and stroke (ICD-10: I00 to I99 and R960 to R961), respiratory (ICD-10: J00 to J99), and ill-defined causes (ICD-10: R00 to R99).

The standardized interview included questions on demographic characteristics, level of education, living alone, and on health behaviors (for example, consumption of alcohol and smoking status). Information on the health of the participants was obtained through detailed medical questionnaires. Case-level depressive symptoms were defined as a score above the 16-point cut-off on the Center for Epidemiological Studies-Depression Scale 29 . Anxiety trait symptoms were measured using the Spielberger’s State-Trait Anxiety Inventory 30 . In the absence of a validated cut-off score in elderly populations, the state score was divided into tertiles with the highest tertile (higher level of anxiety) being compared to the two lowest tertiles. Global cognitive function was assessed by the Mini-Mental State Examination 31 and participants scoring less than 26 were classified as cognitively impaired. Confinement was defined as social restriction (confinement to bed, home or outings restricted to the neighborhood) 32 . Body mass index (BMI) was calculated as weight (kg) divided by height squared (m²). The presence of hypertension was defined by measured systolic blood pressure ≥160 mmHg or diastolic blood pressure ≥95 mmHg or current antihypertensive treatment. Diabetes was defined as fasting glucose level ≥7.0 mmol/l or treatment for diabetes. Hypercholesterolemia was defined as total cholesterol level ≥6.2 mmol/L or treatment with lipid-lowering agents. Detailed medical questionnaires included past history of respiratory and thyroid disorders, and cardio-cerebrovascular disease (angina pectoris, myocardial infarction, cardiovascular surgery, arteritis, and stroke) established according to standardized questions.

Sleep complaints were assessed at baseline as part of the clinical interview, followed by the completion of a specific sleep questionnaire 33 . The participants self-rated as ‘never, rarely, frequently, or often’ occurrence of being excessively sleepy during the day (EDS), having difficulties in initiating sleep (DIS), having several awakenings during the night (difficulties in maintaining sleep; DMS), having early morning awakening (EMA) without being able to go back to sleep, and snoring loudly. In this analysis, EDS was defined as reporting frequently or often being excessively sleepy. Insomnia complaints based on DIS, DMS, and EMA were dichotomized as frequently/often versus never/rarely and summed up to obtain a number of insomnia complaints ranging from 0 to 3. The risk of obstructive sleep apnea syndrome (OSAS) was defined clinically as being obese (BMI ≥30 kg/m²), with frequent/often EDS, and frequent/often loud snoring.

At baseline and at 2, 4, and 8-year follow-up, an inventory of all prescriptions and over-the-counter drugs used during the preceding month was included in a standardized interview. Medical prescriptions and the medications themselves were checked by the interviewer, thus minimizing exposure misclassification. Current use of antidepressants and hypnotics were coded according to the World Health Organization’s Anatomical Therapeutic Chemical Classification 34 . Hypnotics were classified as; benzodiazepines (BZD), BZD-like compounds (zolpidem, zopiclone), and miscellaneous medications (including barbiturates, antihistamines, and other pharmacological categories such as neuroleptics). At baseline, the participants currently taking hypnotics were also requested to report the duration of hypnotic intake.

Logistic regression models were used to compare the characteristics of participants according to the use of hypnotics at baseline after adjustment for study center, age, and gender. To analyze the associations between hypnotic use and risk of mortality, Cox proportional hazard models with delayed entry and age of the participants as the time scale were used to estimate hazard ratios (HR) and their 95% confidence intervals (CI). This method gives better adjustment for age and is therefore preferable for a sample of elderly individuals over the standard model that uses study time as the time scale, because the covariates are strongly associated with age (for example, chronic diseases) 35 36 . Multivariate models included covariates that were associated with mortality at a conservative level of P <0.15. Model 1 was adjusted for study center, age, and gender. Model 2 was further adjusted for education, alcohol intake, smoking status, BMI, confinement, respiratory disorder, cognitive impairment, history of CVD, hypertension, and diabetes. Two other models were adjusted for diseases associated with hypnotic use to take into account possible prescription bias, for example, number of insomnia complaints and EDS (model 3); and anxiety, depressive symptomatology, and antidepressant use (model 4). The multivariate model 5 was adjusted for all possible confounders. All-cause mortality was the principal outcome defined for the analysis. In secondary analyses, cause-specific mortality due to CVD and cancer was analyzed for separate end points. If both CVD and cancer were reported as cause of death, both causes were considered in the analysis. In all final models, significance level was set at P <0.05. Analyses were performed using SAS statistical software (version 9.2; SAS Inc, Cary, NC, USA).

As shown in the study diagram (Figure  1), the study sample included 6,696 participants free of dementia (58.7% women) with a median age of 72.8 years (range, 65.0 to 95.0 years). The 2,382 participants free of dementia excluded from the study were significantly more likely to be older, have a lower education level, were more frequently female and living alone, with confinement, hypertension, diabetes, respiratory disease, hypercholesterolemia, depressive and anxiety symptoms, cognitive impairment, past history of cardio-cerebrovascular disease, and taking more hypnotics (P <0.05 for all comparisons). They were also more likely to have died during the follow-up period (P <0.0001).

At baseline, 21.7% of the participants (n = 1,454) were taking at least one hypnotic, 6.9% (n = 464) had three insomnia complaints, and 3.9% (n = 260) had no insomnia complaints. More than 3% (n = 212) reported taking two or more hypnotics. Regarding the main classes of hypnotics, 16% (n = 1070) took BZD, 4.8% (n = 321) BZD-like compounds, and 3.0% (n = 204) miscellaneous medications (of whom 54.4% took antihistamines, 25.0% non-BZD anxiolytics, 18.6% barbiturates, and 4.4% neuroleptics). With regard to duration, 4.8% (n = 304) had been taking hypnotics for less than 5 years, 3.9% (n = 244) between 5 and 10 years, 2.0% (n = 127) between 10 and 20 years, and 6.0% (n = 378) for more than 20 years.

Baseline sociodemographic and clinical characteristics of the participants according to hypnotic use are described in Table  1. An analysis adjusted for study center, age, and gender showed that participants taking hypnotics had a lower education level; were more likely to be confined to home; had more symptoms of depression, anxiety, and cognitive impairment; more frequently had a past history of chronic disease (CVD, thyroid disease, diabetes, hypercholesterolemia); consumed less caffeine; and reported more insomnia complaints and EDS (P <0.05 for all comparisons).

^aAdjustment for center study, age and gender; ^buniversity level; ^chistory of cardiovascular disease (stroke or coronary heart disease); ^dnumber of insomnia complaints: difficulties in initiating sleep + difficulties in maintaining sleep + early morning awakening. CI, confidence interval.

The median follow-up time for the study was 8.9 years with a range of 0.06 to 11.7 years. During this period, 1,307 (19.5%) deaths were observed. They were particularly related to CVD (26.3%), cancer (36.8%), and co-morbid CVD and cancer (3.1%). A substantial number of deaths were due to ill-defined causes (21.6%) as the result of multiple pathologies associated with frailty, and 10.2% died from respiratory diseases.

Baseline sociodemographic and clinical characteristics in relation to follow-up mortality (all causes) are given in Table  2. Participants who died during follow-up were more frequently confined to home, obese, past or current smoker, consuming less alcohol, had hypertension, diabetes mellitus, a past history of CVD, respiratory disease, poorer cognitive performance, EDS, depressive symptoms, or were taking antidepressants. They also tended to have a lower level of education, and more frequently reported insomnia and anxiety symptoms (P <0.15). Subsequent analyses were adjusted for these factors. A significant association was also found for participants at risk of OSAS (n = 133; HR = 1.76; 95% CI = 1.31, 2.36; P = 0.0002).

^aAdjusted for center study, gender and age; ^buniversity level; ^chistory of cardiovascular disease (stroke or coronary heart disease; ^dnumber of insomnia complaints: difficulties in initiating sleep + difficulties in maintaining sleep + early morning awakening.

Table  3 shows the associations between hypnotic use at baseline and all-cause mortality over the 12-year follow-up. After adjustment for age, gender, and study center, the risk of all-cause mortality increased significantly with the use of any hypnotic, the number of hypnotics, and alone for BZD (P <0.01 for all comparisons, model 1). When potential lifestyle and chronic disorder confounders were entered into the model (model 2), the HR were reduced and failed to be significant except for BZD (P = 0.05) and this was unchanged when further adjusting for sleep complaints (model 3). When adjusting for anxiety and depressive symptomatology (model 4), the associations were not significant even for BZD (P = 0.22) and this was also the case for the complete multivariate model adjusted for all potential confounders (model 5). BZD-like compounds, and miscellaneous medications intake were not associated with all-cause mortality even in the minimally adjusted model 1. No significant interaction was found for mortality between hypnotic use and EDS, number of insomnia complaints, antidepressant use, chronic diseases, or being at risk for OSAS.

^aAdjusted for age, study center, and gender; ^badjusted for age, study center, gender, level of education, confinement, alcohol intake, smoking status, history of cardio and cerebrovascular disease, respiratory disease, Mini Mental State Examination score, body mass index, hypertension, and diabetes mellitus; ^cadjusted for all the covariates in model 2, plus excessive daytime sleepiness and number of insomnia complaints; ^dadjusted for all the covariates in model 2, plus depressive symptoms, antidepressant use and Spielberger trait anxiety score; ^eadjusted for all the covariates in model 3 plus depressive symptoms, antidepressants use and Spielberger trait anxiety score. BZD, benzodiazepines; CI, confidence interval; HR, hazard ratios.

The relationship between hypnotic intake and the risk of mortality remained unchanged after exclusion of the participants who died during the first two years of follow-up (n = 134), the follow-up rate at two years being 88%. With regard to specific causes of death, the use of hypnotics and BZD as well as number of hypnotics were associated with a significantly increased risk of CVD-related death in model 1, but not in the complete multivariate model adjusted for all potential confounders (Table  4). There was no significant association between hypnotics and cancer-related death regardless of covariates, even in the minimally-adjusted model 1 (P = 0.38).

^aAdjusted for age, study center, and gender; ^badjusted for all covariates in model 1 plus high level of education, confinement, alcohol intake, smoking status, history of cardio-cerebrovascular disease, respiratory disease, Mini Mental State Examination score, body mass index, hypertension and diabetes mellitus, depressive symptoms, antidepressants use, Spielberger trait anxiety score, excessive daytime sleepiness, and number of insomnia complaints. BZD, benzodiazepines; CI, confidence interval; HR, hazard ratios.

Sensitivity analyses were performed to examine the relationship between persistent use of hypnotics during the initial 4 years and all-cause mortality. A total of 3,496 participants (65.9%) did not report hypnotic use at baseline or at follow-up examination, 773 (14.5%) reported use both at baseline and at the first two follow-ups (persistent users) and 1,040 (19.6%) were taking hypnotics at one of two time points (intermittent users). The risk of mortality for the next 8 years was not significantly associated with the persistent use of hypnotics (when compared with non-users, HR = 1.03, 95% CI = 0.84, 1.28 for intermittent users; HR = 1.11, 95% CI = 0.88, 1.39 for persistent users; multivariate model 5). Similar results were obtained when the analyses focused on persistent BZD users in comparison to non persistent BZD users or non BZD users.

We also examined the impact of past hypnotic intake duration and compared participants who were not taking sleep medication at baseline with those having previously reported taking sleep medications for less than 5 years, between 5 and 10 years, between 10 and 20 years, and for more than 20 years. No significant association was observed between duration of hypnotic intake and all-cause mortality, the global P-value ranging from 0.18 (model 1) to 0.76 (model 5) (data not shown).