Study design
We performed a series of cohort studies, using data from the Clinical Practice Research Datalink (CPRD) linked to Hospital Episode Statistics (HES), Office for National Statistics (ONS) mortality data, and Index of Multiple Deprivation data in England. CPRD collates all diagnoses, referrals, and prescribing records for over 11.3 million patients broadly representative of the UK population [21]. Diagnosis information is electronically entered as UK Read codes [22]. HES records all diagnoses made during a hospital admission (coded using the International classification of diseases 10th revision [ICD-10]) and demographic information [23], and ONS provides date and cause of death (ICD-10 coded) [24]. The Index of Multiple Deprivation combines a number of indicators of housing, employment, income, education, and environment at the general practice level [25]. CPRD obtained ethical approval from a National Research Ethics Service Committee, allowing researchers to access anonymised data for observational studies upon approval from an Independent Scientific Advisory Committee.
Study population
We defined dementia patients by record of a dementia diagnosis in CPRD (codes in Additional file 1) or HES (ICD-10 F00-F03, G30, G31.0 or G31.1) or prescription of a cognitive enhancer (i.e. memantine, donepezil, rivastigmine, or galantamine), occurring after January 2000 when aged ≥ 55 years. We excluded patients with < 3 months of ‘up-to-standard’ (research quality) data or with severe mental illness or Down syndrome before dementia diagnosis (codes in Additional file 1) [21].
Exposures
Our primary exposure was new prescription of Z-drugs. We considered three comparator groups to reduce confounding [26]. The primary comparator was record of sleep disturbance without sedative-hypnotic prescription. Secondary comparators were (a) non-users at a proximal GP consultation and (b) new benzodiazepine users. To facilitate these comparisons, three main cohorts were constructed, with their index date as the first date after dementia diagnosis of (a) prescription for a Z-drug (World Health Organization’s Anatomic Therapeutic Classification [ATC] system category N05CF), (b) prescription of any benzodiazepine (ATC N05BA or N05CD except midazolam injection), and (c) code for sleep disturbance (codes in Additional file 1) but without a concurrent sedative-hypnotic (ATC N05C or N05BA) prescription.
Two additional cohorts were created by matching the Z-drug and BZD cohorts on age, sex, and antipsychotic use, to non-users (not prescribed sedative-hypnotics) with proximal GP consultation. We matched three non-users to each Z-drug or benzodiazepine user, without replacement, and assigned an index date as the closest GP consultation within 1 month of the corresponding Z-drug or benzodiazepine index date. Patients could be members of different cohorts over time.
Exclusion criteria for all cohorts were (1) < 12 months data history; (2) sedative-hypnotic prescription in the last 12 months; (3) prior diagnosis of sleep apnoea, sleep-related respiratory failure, or alcohol abuse (codes in Additional file 1); (4) prescription of multiple sedative-hypnotics; and (5) newly prescribed antipsychotics or low-dose tricyclic or related antidepressants (≤ 25 mg amitriptyline or ≤ 50 mg trazodone per day). We additionally performed a separate validation study on the accuracy of our patient selection (details in Additional file 2). In summary, GP practices confirmed dementia diagnoses in 96% of cases; however, uncertainty was raised regarding the accuracy of those identified with sleep disturbance.
To test dose-response relationships, we determined the number of daily defined doses (DDDs) of Z-drugs and benzodiazepines at each prescription. The DDD is the assumed average maintenance dose per day for a drug based on its main indication in adults. We used DDD values from the World Health Organization’s Collaborating Centre for Drug Statistics Methodology (www.whocc.no/atc_ddd_index), where the DDDs for zopiclone, zolpidem, and zaleplon are 7.5 mg, 10 mg, and 10 mg per day, respectively. The British National Formulary recommends these daily doses for adults with insomnia, but to halve them for elderly patients. Missing dosing frequency data was assumed once daily, except diazepam where we applied the most common frequency for the product and quantity prescribed among the complete prescription data.
Outcomes
The selected outcomes were identified from previous studies or were priorities identified by our Patient and Public Involvement (PPI) group members or by an advisory group of healthcare professionals established to support this project. The main outcomes were, in order of importance: (1) incident (a) fracture in any location, (b) hip fracture, and (c) forearm/wrist/hand fracture; (2) incident fall; (3) mortality; (4) acute bacterial infection; (5) ischaemic stroke/transient ischaemic attack; and (6) venous thromboembolism. These were identified via first mention of a relevant Read code in CPRD or ICD-10 code in HES or ONS (codes in Additional file 3). We also examined further healthcare utilisation outcomes: (7) number of (a) hospital admissions and (b) GP consultations and (8) new prescription of (a) antipsychotics, (b) antidepressants, and (c) antibiotics.
Covariates
We considered as potential confounders variables suspected to be linked to dementia, sleep disturbance, benzodiazepine or Z-drug use, or the outcomes examined. They were measured on the index date and covered domains of demographics, health behaviours, dementia subtype, proxies for dementia severity, proxies for sleep disturbance severity, comorbidities, recent medical history (e.g. GP consultations, hospital admissions, falls, fractures, infections, immunisations, body mass index [BMI], systolic blood pressure), and concurrent prescriptions (details in Additional file 4).
Statistical analysis
The primary analysis estimated the association between new prescription of sleep disturbance medication and incidence of each outcome, compared to other groups. We followed patients until the earliest of death, leaving the GP practice, last data extraction, new sedative-hypnotic or antipsychotic prescription, 2 years post-index date, or 31 March 2016. Z-drug and benzodiazepine new users were also censored 90 days following their last Z-drug/benzodiazepine prescription. Matched patients were additionally censored at the censoring date of their corresponding case. Specific exclusion criteria applied at the index date to reduce the chance of repeated coding of the same event are described in Additional file 5 table S1.
Cox proportional hazards regression was used for binary outcomes. We used robust standard errors to account for the correlation due to repeat measurements in some patients [27]. The proportional hazards assumption was checked using Schoenfeld residuals [28]. Negative binomial regression was used to model the number of hospital admissions and GP consultations. Estimates were adjusted for age and sex and all covariates in Additional file 4. Age, BMI, systolic blood pressure, duration since dementia diagnosis, index date, and number of prior GP consultations were modelled using restricted cubic splines (with five knots) to allow non-linear effects [29]. We included an interaction between sex and BMI, due to known sex differences in the relationship between BMI and fracture risk [30]. Absolute risk differences of adverse events and numbers needed to harm (NNH) were estimated using standard formulae for time to event analysis [31].
We examined the average daily Z-drug dose over follow-up, but in post hoc secondary analysis, to reflect changes in dose, we examined time-varying daily DDDs. In the sensitivity analysis, we excluded those with record of > 6 h sleep per night from the sleep-disturbance comparator group (see validation study in Additional file 2 for more detail). This was to increase the chance that the sleep-disturbance group had more comparable insomnia to the Z-drug group. Finally, in the comparison of Z-drug to benzodiazepine new users, we restricted to benzodiazepines likely prescribed for sleep disturbance (loprazolam, lormetazepam, nitrazepam, temazepam, or other benzodiazepines with dosing instructions to take only at night or with a concurrent record of sleep disturbance).
Multiple imputation by chained equations was used to impute missing values of BMI, smoking, alcohol use, residence, ethnicity, and blood pressure (see Additional file 4 for details) [32]. To account for multiple outcomes tested, we used the Benjamini-Hochberg procedure to control the false discovery rate at < 5% for each analysis [33]. Stata version 15.1 was used throughout.