Non-benzodiazepine hypnotic use for sleep disturbance in people aged over 55 years living with dementia: a series of cohort studies

Sleep disturbance in people with dementia

Sleep disturbance can include insomnia, fragmented night-time sleep, night-time wandering and excessive daytime sleeping. The incidence of sleep disturbance in the older population is common, with many older adults reporting insomnia. 9

Studies show that chronic insomnia predicts poor health, with better health leading to improved sleep patterns,9 and that disturbed sleep is actually a rarity in older adults who are healthy. 10 Reasons for age-related decline in sleep duration and quality include physical and psychiatric illness. 11,12 Therefore, it is unsurprising that in people living with dementia, sleep disturbance is high, with around 60% of people living with dementia affected. 13,14

Sleep disturbance in Alzheimer’s disease is correlated with further cognitive dysfunction, suggesting that treatment of sleep disturbance could be a strategy to improve cognition in patients. 15 Poor sleep impacts greatly on the quality of life (QoL) of both the patient and their informal carer. 16,17 Furthermore, frequent night-time waking can cause further risks to the individual, including wandering outside (which contributes to significant reductions in carer sleep and increases carer burden),18–20 and increases the rate of care home admissions for people living with dementia. 21 Sleep disturbance is highly prevalent in people living in care homes, and residents with dementia are reported to have more sleep disturbance then those without dementia. 22 In many cases, this leads to pharmacological interventions being commonly used to manage sleep disturbance in community, hospital and care home settings. 23

Pharmacological treatment of sleep disturbance: benzodiazepines and Z-drugs

There are several pharmacological treatments available for the management of sleep disturbance in the older population, including hypnotic benzodiazepines (BZDs) and Z-drugs (e.g. zolpidem, zaleplon and zopiclone), hormones (melatonin), antidepressants [typically low-dose tricyclic antidepressant (TCAs)] and antihistamines. The most widely used medications for sleep disturbances in older people are BZDs, low-dose TCAs and Z-drugs. 24

Low-dose TCAs have sedation effects due to their pharmacological effects of blocking histamine-1 receptors and are used for promoting sleep, but also in pain management and at higher doses for treating depression. 25 BZDs are long acting and act to enhance the effect of gamma-aminobutyric acid (GABA), an inhibitory neurotransmitter, causing sedation and thus promoting sleep. 26 BZDs are not only given for sleep disturbance, but are also prescribed for a wide range of conditions, including anxiety and agitation. 27 BZDs are associated with a range of adverse side effects, including cognitive impairment, daytime sedation, tolerance, dependence and falls. 28–31 Z-drugs are a newer generation of BZD-like drugs, prescribed exclusively for sleep disturbance. Z-drugs were introduced in the late 1980s, also targeting GABA receptors (similarly to BZDs), to enhance GABA transmission, but with shorter half-lives and claiming to have fewer side effects. 32

However, there is increasing emerging evidence of adverse effects. 33 The National Institute for Health and Care Excellence Guidance on the Use of Zalepon, Zolpidem and Zopiclone for the Short-Term Management of Insomnia recommends that non-pharmacological approaches should be considered first, but that short-acting BZD or Z-drugs can be used for up to 4 weeks, if appropriate. 34 There is no explicit recommendation in these guidelines for use in people living with dementia. Updated guidelines35 for hypnotics discuss the growing evidence of increased risk of cognitive impairment associated with hypnotic use, but do not specifically address if they are safe for use in people living with dementia. There are no recommendations for pharmacological interventions for sleep disturbance in National Institute for Health and Care Excellence dementia guidelines. A Cochrane systematic review in 2016 concluded that there was a distinct lack of evidence to help guide drug treatment of sleep problems in people living with dementia. 36

Adverse effects of hypnotic use in the older population

Benzodiazepines and Z-drugs can improve sleep quality, but clinically relevant adverse events are a cause for concern. 37 Risks in people living with dementia have been largely unexplored, but findings from the older population suggest that the consequences of any side effects in people living with dementia could be more severe.

Cognitive functioning

Sleep is crucial for cognitive performance, memory consolidation and mood regulation. 81,82 Sleep loss is associated with reduced cognitive performance. 83,84 Furthermore, observational and experimental studies suggest that poor sleep is a risk factor for cognitive decline and dementia. 85 The underlying mechanisms are unknown, but this suggests that improving sleep quality could be an intervention strategy for dementia, potentially through sedative hypnotics. However, there is evidence that sedative hypnotic use is associated with cognitive decline, with a stronger association observed with long-term exposure,86–89 although results are conflicting. 90,91 A recent systematic review and meta-analysis suggests that BZD use is associated with increased dementia risk; however, observational studies cannot determine whether the association is a causal effect or due to unmeasured confounders. 92,93 Evidence is emerging that Z-drug use is also associated with dementia risk. 94,95 Despite these findings, there is a lack of studies addressing whether or not sleep drugs are safe for individuals already suffering with significant cognitive decline. 96 It is therefore important to understand if increasing sleep in people living with dementia through pharmacological interventions can maintain cognitive performance.

Quality of life

Sleep disturbance can have a detrimental impact on QoL and health-related quality of life (HRQoL). 97,98 Insomnia has been independently associated with worsened HRQoL to a similar extent as some chronic conditions, including congestive heart failure. 99 Insomnia and sleep disturbance are common in the older population and are associated with decline in QoL. 100 The QoL of carers is also severely reduced when there is a sleep disturbance. 19 Use of hypnotics could be a strategy to increase sleep and thus improve QoL. There have been studies, however, to suggest that use of hypnotics can have a negative impact. In older community-dwelling adults requiring pharmacological treatment for insomnia, hypnotics are also associated with reduced HRQoL. 101 A study of 10,430 older women identified that use of sleeping medications was associated with lower QoL scores in vitality, social functioning, general mental health and bodily pain. 102

Improving sleep

In dementia, for which sleep disturbance is common, improving sleep could be a valuable intervention strategy to improve cognitive performance. Hypnotics are widely prescribed in this population in an attempt to increase sleep. However, there is increasing evidence that the use of sedating hypnotics in the general population does not improve sleep, cognitive performance or general health, calling into doubt their clinical effectiveness. 103,104

Primary care study

• To estimate the effects of first prescription of Z-drugs in people living with dementia with sleep disturbance compared with alternative treatments and no treatment. Patient outcomes included:

  • incidence of falls and factures

  • mortality

  • incidence of infection

  • incidence of cerebrovascular events, including ischaemic stroke and venous thromboembolism

  • incidence of behavioural and psychological symptoms

  • additional medication use, including sedatives, antipsychotics, antidepressants and antibiotics

  • health-care utilisation [general practitioner (GP) visits and hospital admissions].

• To validate recorded dementia diagnosis and sleep disturbance codes through a GP questionnaire.

Clinical cohort studies

• To repurpose existing clinical data sets to estimate the impact of concurrent use of Z-drugs on patient- and carer-reported outcomes, including:

  • QoL for patients and carers

  • functional ability

  • cognitive function and sleep disturbance in people with dementia.

Study design

This was an inception cohort study of people with dementia and sleep disturbance, using data from a UK primary care database. The study design is summarised in Figure 1. Briefly, people living with dementia in cohorts prescribed various treatments were followed for the rate of adverse events from their first prescription for sleep disturbance in dementia (index date), until they had a different sleep drug or antipsychotic prescribed, left their general practice, died or at 2 years’ follow-up (censor date).

FIGURE 1.

Schematic of study design for the primary care cohort study of people living with dementia (n = 6809). The four cohorts of patients were followed for at most 2 years or until March 2016. Follow-up for the cohort diagnosed with sleep disturbance but not prescribed sleep medications ended when they received a first sleep medication or antipsychotic prescription. The figure displays theoretical patterns of prescriptions for the cohorts defined by a first Z-drug, BZD or low-dose TCA prescription. Here, follow-up ended if patients had a first sleep medication or antipsychotic prescription, or if it had been > 90 days since their last prescription for a Z-drug, BZD or low-dose TCA, respectively. Rx, prescription.

Setting

This study was undertaken using data extracted from the CPRD. The CPRD includes anonymised diagnosis, referral and prescribing records for > 11.3 million patients from 674 primary care practices across the UK, and is broadly representative of the UK population in terms of age, sex and ethnicity. 114 CPRD data have been widely used for pharmacoepidemiology applications. 115 The diagnosis information recorded in the CPRD is entered as computer-recorded Read codes. 116 Access to ‘free text’ entered by the GP is not available because of patient anonymity issues.

We used linked data from the CPRD to Hospital Episode Statistics (HES) data version 14, the Office for National Statistics (ONS) mortality data and Index of Multiple Deprivation (IMD) data. HES provides records of all diagnoses made during a hospital admission in England, details on the patient’s demographics and place of residence. Diagnoses during the hospital stay are recorded using the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10). ONS data include the date and cause of death in England, with cause of death coded according to ICD-10. The IMD is a weighted sum of a number of government indicators, including housing, employment, income, education, living environment and crime levels for each general practice’s neighbourhood. 117

The HES and ONS data sets were linked to CPRD patients by NHS number, date of birth, sex and postcode. The vast majority (92%) of eligible CPRD patients matched to HES, matched on all four characteristics, whereas a further 8% were matched on all but postcode.

Eligibility criteria

Patients were included if they satisfied all the following criteria:

• Their general practice was in England.

• They were diagnosed with dementia, defined as the first of a code for dementia or prescription of a cognitive enhancer (memantine, donepezil, rivastigmine or galantamine), occurring after 1 January 2000. The year 2000 onwards was chosen to exclude historical primary care records when recording was less reliable,114 and linked mortality and hospital records were only available since 1998.

• They were aged ≥ 55 years when diagnosed with dementia.

• There was evidence of a sleep disorder, defined as the first primary care record for sleep disorder diagnosis, symptom or referral (see Appendix 1, Table 27, for Read codes), or prescription of a Z-drug, low-dose TCA or melatonin, on or after the dementia diagnosis date and before 31 March 2016 (this first sleep disturbance date defined the ‘index date’).

• Their records contained at least 3 months of good-quality data before the dementia diagnosis and at least 12 months of data before the index date.

Record of a dementia diagnosis was identified by Read codes in the CPRD (see Appendix 1, Table 26 for Read codes) or ICD-10 codes F00-F03, G30, G31.0 or G31.1 in HES.

Patients were excluded if there was:

• uncertainty regarding the timing of dementia diagnosis, specifically Read codes for a dementia annual review, history of dementia, assessment of psychotic and behavioural symptoms of dementia, or antipsychotic drug therapy for dementia, prior to meeting our dementia diagnosis definition

• diagnosis of severe mental illness or Down syndrome prior to dementia diagnosis (see Appendix 1, Table 28 for Read codes)

• diagnosis of sleep apnoea, sleep-related respiratory failure or alcohol abuse prior to the index date (see Appendix 1, Table 29 for Read codes)

• diagnosis of neuropathic pain in the 12 months prior to the index date (see Appendix 1, Table 30 for Read codes)

• a prescription of sedatives, TCAs or BZDs in the 12 months prior to the index date

• a prescription of multiple sleep medications on the index date

• a prescription of a new antipsychotic, other sedative or other TCA on the index date

• no linkage possible between CPRD and HES data.

Patient selection validation

To validate the accuracy of the coding on which our patient selection was based, a GP questionnaire-based validation study was conducted. 118 In collaboration with CPRD validation services we developed a questionnaire to send to GPs, requesting confirmation of the dementia diagnosis and sleep disturbance status of their patient. The questionnaire was sent to the GPs of 106 randomly selected patients who were still registered with their GP in 2017. The patients selected represented those identified as having sleep disturbance via various Read codes, who were or were not prescribed sleep medications. The questionnaire (see Report Supplementary Material 1) asked GPs if the patient had been diagnosed with dementia and the date of diagnosis. It also asked whether or not the patient had a record of a sleep disturbance since being diagnosed with dementia and, if so, on what date. We reported the number of questionnaires returned and the proportions of patients, with their dementia and sleep disturbance confirmed by the GP according to their first medication for sleep disturbance. When the records of dementia or sleep disturbance were not confirmed by GPs, we explored possible causes. We also performed a sensitivity analysis restricted to patients with sleep codes with better validity.

Exposures

The primary exposure of interest was treatment with Z-drugs; secondary exposures were treatment with other medications used for sleep disturbance. The CPRD contains detailed information on all medications prescribed in primary care, including the date of each prescription, drug name, dose, quantity and frequency. We extracted the details of all prescriptions for medications for sleep disturbance for patients in our cohort from their index date up until 31 March 2016.

Sleep disturbance medications were defined according to the World Health Organization (WHO)’s Anatomical Therapeutic Chemical (ATC) Classification System as:

• Z-drugs (N05CF)

• BZDs (N05BA, N05CD)

• melatonin (N05CH)

• low-dose TCA or related (N06AA09, amitriptyline at ≤ 25 mg/day; N06AX05 trazodone at ≤ 50 mg/day).

Patients were classified according to their prescription class on the index date, or to a ‘no prescription for sleep disturbance’ group. Patients in the ‘no prescription for sleep disturbance’ group who then went on to be prescribed any of the medications for sleep disturbance and still met the study eligibility criteria, were then assigned a second index date as the date of this first prescription.

To test for possible dose–response relationships, we determined the cumulative number of prescribed defined daily doses (DDDs) of Z-drugs. A DDD is defined as the assumed average maintenance dose per day for a drug based on its main indication in adults, using the DDD values assigned by the WHO Collaborating Centre for Drug Statistics Methodology (URL: www.whocc.no/atc_ddd_index; accessed 28 November 2019). For each prescription, we multiplied the number of tablets by the dose strength in milligrams and converted this into the number of DDDs, using the values assigned by the WHO. We then summed individual prescriptions to determine the cumulative number of prescribed DDDs, regardless of gaps in use. If the quantity of tablets prescribed was missing, we assumed that one per day (28 tablets/prescription) was prescribed. Prescribing instructions for Z-drugs were also categorised according to whether or not instructions stated pro re nata (PRN) (i.e. use as needed).

The Z-drugs historically prescribed in the UK are zopiclone, zolpidem and zaleplon. The WHO considers the DDD of zopiclone to be 7.5 mg per day, and zolpidem and zaleplon to be 10 mg per day. These values are consistent with the British National Formulary (BNF) prescribing guidelines for the recommended daily dose in adults with insomnia. 119,120 For use in the elderly, the BNF recommends using half these daily doses.

Outcome variables

The selected outcome variables were based on adverse effects of medications used to treat sleep disorders identified from previous studies or priorities identified by carers, family members and health-care professionals with direct experience of sleep disturbance in patients.

The 16 outcomes assessed were:

1. incident fracture in any location

2. incident hip fracture

3. incident forearm/wrist/hand fracture

4. incident fall

5. mortality

6. incident acute bacterial infection

7. incident urinary tract infection (UTI) or acute lower respiratory tract infection (LRTI)

8. ischaemic stroke/transient ischaemic attack (TIA)

9. venous thromboembolism

10. incident agitation or psychosis (including symptoms of hallucinations, delusions or aggression)

11. additional use of sedatives and other sleep medications (BNF version 69 subsection 4.1)120

12. additional use of antipsychotics (BNF version 69 subsection 4.2.1)120

13. additional use of antidepressants (BNF version 69 subsection 4.3)120

14. additional use of antibiotics (BNF version 69 subsection 5.1)120

15. health-care utilisation of number of GP visits

16. health-care utilisation of number of hospital admissions.

Patients with each outcome were identified using the first mention of a relevant Read code in the CPRD, or ICD-10 code in HES or as a cause of death [Part 1(a) or 1(b)] on the death certificate. See Appendix 1, Tables 31–38, for a list of the codes to identify the outcomes. Read code lists were drawn, when applicable, from Quality and Outcomes Framework business rules, published studies, keyword searches and UK GP experience within the study team.

Confounding variables

Potentially confounding variables were coded from the CPRD (unless otherwise stated) at the index date. Covariates were selected that are potentially linked to sleep disturbance, dementia or sedative use and at least one of the outcomes, as well as the availability, completeness and reliability of data within CPRD.

Statistical methods

We described the characteristics, comorbidities and concurrent medications of the patients in the cohort, using summary statistics according to their initial treatment for sleep disturbance. Using a backwards stepwise procedure retaining a p-value of < 0.15, we developed a multinomial regression model estimating the association between patient characteristics at index date and treatment choice for sleep disturbance.

The primary statistical analysis comprised a series of survival analyses to assess the association between sleep disturbance medication and adverse outcomes. Patients entered the analysis on their index date and were followed until the earliest of:

• the date they left the general practice

• the last general practice data extraction date

• their date of death

• 90 days after their last prescription for their medication for sleep disturbance

• the date of an additional ‘medication for sleep disturbance’ or other sedative prescription

• the date of a new antipsychotic prescription (for those not prescribed antipsychotics in the year before the index date)

• 2 years after the index date

• 31 March 2016.

Patients were excluded from the analysis of each outcome depending on their history of that outcome. For the fall and fracture outcomes, we excluded patients with a recorded fall or fracture in the 32 days before index date, because of the chance of repeated coding of the same event. 121 For the infections and antibiotics outcomes, we excluded those with an infection or prescription for a medication for infection in the previous 30 days. For the ischaemic stroke/TIA and venous thromboembolism outcomes, we excluded those with a history of these in the previous 30 days. For the analysis of incident agitation/psychosis, we excluded patients who already had a record of agitation or psychosis. For the antidepressant and antipsychotic outcomes, we excluded patients with a prescription for these in the 12 months before the index date.

We used Cox regression models to estimate the HR for the effect of sleep disturbance medication class compared with no prescription on each binary outcome. Sleep medication exposure was modelled as time varying, such that patients at the index date are included for analysis in the ‘no treatment’ group until initiation of their first treatment, and re-enter the study at time 0 as exposed thereafter to avoid immortal time bias and reduce channelling bias. CIs and p-values were calculated using robust standard errors, accounting for the correlation due to some patients appearing twice in the analysis. We checked the proportional hazards assumption using Schoenfeld residuals. 122

As the number of GP visits and hospital admissions was overdispersed (i.e. the variation was greater than the mean), we used negative binomial regression to calculate incidence rate ratios (IRRs) for the effect of sleep disturbance medication class on number of GP visits and hospital admissions in the 2 years after the index date. For the analysis of GP visits, patients were censored only at the first of either death, leaving the general practice, 31 March 2016 or 2-year follow-up. For the analysis of hospital admissions, patients were censored only at the first of either death, 31 March 2016 or 2-year follow-up.

The primary analysis reported associations relative to no sleep disturbance medication use, and secondary analysis reported associations for the effect of BZDs and low-dose TCAs compared with Z-drug use. Estimates are provided, adjusted for age and sex, and adjusted for all potential confounders listed above and age2 to allow non-linear effects with age. A quadratic association between age seemed an appropriate parametrisation for most outcomes by examining the association between age in 10 equal categories and the outcome. We performed a sensitivity analysis to this by instead modelling age using fractional polynomials and restricted cubic splines (with five knots). Owing to a small number of events for forearm fractures we adjusted for other eye conditions (ARMD, retinal disorders and glaucoma combined), region combined into only five areas (South, East, Midlands, North and London) and did not adjust for ethnicity or systolic blood pressure.

For statistically significant associations with Z-drugs, interactions between Z-drug and age and sex were tested. In addition, the absolute risks of adverse events and number needed to harm (NNH) were estimated using a standard formula for NNH in time-to-event analysis. 123

We also carried out analyses of Z-drug exposure according to both time-varying cumulative DDDs (categorised as 0, 1–13, 14–27, 28–41 and ≥ 42 DDDs) and by Z-drug dosing instructions on the index date (PRN or not).

Stata^® version 14 (StataCorp LP, College Station, TX, USA) was used for data management and statistical analysis. Owing to examining 16 outcomes, the chance of finding a statistically significant association by chance alone had increased. To address this, we used a stricter threshold instead of the traditional p-value threshold of 0.05. We used the Benjamini–Hochberg procedure to estimate this critical p-value threshold in order to control rate of the false discovery rate (i.e. the proportion of rejected null hypotheses that are incorrect rejections) at < 5%. 124 This is less conservative than the Bonferroni correction,125 which simply divides the classical p-value threshold of 0.05 by the number of hypotheses tested (so our critical threshold would be 0.05/16 = 0.003). To control the false discovery rate, the Benjamini–Hochberg procedure depends on the distribution of p-values resulting from the tests, as well as the number of tests performed.

Missing data

We expected incomplete recording of variables, such as smoking and BMI. 126 In the primary analyses, patients with missing covariate data were coded in a missing data category. For covariates with at least 10% of patients with missing data, we summarised the characteristics of those with and without missing data and performed sensitivity analyses, including restricting to patients with and without the missing variable. 127 Finally, we estimated a single imputation model predicting the missing variables using the covariates in Table 2 and in Appendix 2, Table 39. For missing BMI data, we additionally used the most recent previously recorded BMI record to impute missing BMI value. We repeated our main analyses using the imputed values.

Sensitivity analyses

We performed various sensitivity analyses of our main analysis. First, based on the findings of the validation study, we repeated our main analysis with an amended definition of sleep disturbance that included diagnoses of sleep disturbance only when not accompanied by a record for a ‘satisfactory’ sleep pattern. Second, to examine the impact of the source of data used to identify outcomes, we restricted our main analysis to only those outcomes recorded in the CPRD. Third, we compared adverse events in those who initiated Z-drugs compared with those who had recently discontinued them. From the Z-drug cohort we created a further ‘discontinued Z-drug’ cohort, who after a 90-day period with no Z-drug prescriptions were still eligible for entry to the study. We compared the rates of adverse events in the ‘discontinued Z-drug’ to the ‘Z-drug’ cohort.

Protocol changes

The following changes to the analysis plan were made since our protocol:

• We additionally censored patients at their first antipsychotic prescription. Patients initiated on Z-drugs or BZDs were more likely to then start antipsychotics. We additionally censored on first antipsychotic prescription to increase the likelihood that effects observed were due to the Z-drug use and not other sedating medications.

• In our protocol, we aimed to examine incident ischaemic stroke/TIA and incident venous thromboembolism and to exclude anyone with a history of these conditions. However, we found more patients had these histories than expected, so instead we analysed recurrent or incident ischaemic stroke/TIA and venous thromboembolism, and only excluded patients with a history of these in the last 30 days. Findings when excluding all patients with a history of these conditions are available from the authors.

• We made minor changes to the list of potential confounders. We omitted the following prespecified covariates in the analysis due to balance across treatment groups and rare occurrence: phobia, motor neuron disease, human immunodeficiency virus (HIV)/acquired immunodeficiency syndrome (AIDS) and multiple sclerosis. We omitted epilepsy and prescription for medications for Parkinson’s disease and diabetes, due to collinearity with already included covariates of prescription for antiepileptic medications, diagnosis of Parkinson’s disease and diabetes, respectively. Finally, we omitted depression severity and depression duration, as depression diagnosis, depression symptoms, anxiety diagnosis and anxiety symptoms were considered sufficient. Subsequent to the protocol, we have also included the potential confounders of recent prescription for calcium or vitamin D and end-of-life care, based on discussions with our health-care professional advisory panel.

• We omitted a sensitivity analysis using instrumental variables as a result of not finding a sufficiently strong instrument, such that using it would bring greater uncertainty and risk of bias than the main analysis (see Report Supplementary Material 2 for further details), but instead completed an additional analysis of those patients who discontinued Z-drugs.

A priori power calculation

We performed this a priori power calculation before we had access to HES data and made assumptions as to what the HES data would contain; hence the estimates do not match the final numbers included. There were 32,961 patients with sleep disturbance post dementia diagnosis in the CPRD (July 2017 version). Of these, 10,554 patients met study eligibility criteria and a further 6117 patients were eligible for HES linkage (HES version 14). Preliminary analysis suggested that on the index date, 2546 patients received a Z-drug, 68 patients received a melatonin, 255 patients received a BZD, 1745 patients received a low-dose TCA antidepressant and 1503 patients received none of the above treatments. We estimated that an additional 50% of dementia patients would be identified via HES (HES set 14 feasibility). Thus, when comparing 3819 Z-drug patients with 2254 patients with no treatment for a common outcome over 2 years (e.g. falls at 36% per year) we can detect a HR of 1.07 with 90% power (two-sample test, p < 0.05). For rare outcomes, such as hip fractures at 1.7% per year, we can detect a HR of 1.54 with 90% power.

Potential studies

We identified several RCT and cohort studies for which data were obtained and descriptive analysis performed to determine their appropriateness for use in this study. Studies were required to have:

• a validated dementia diagnosis for participants

• specific assessment of sleep disturbance for all participants

• at least two assessments per participant, with assessments separated by ≤ 1 year

• at least one key primary outcome (cognition, neuropsychiatric symptoms, sleep disturbance, disability or QoL of the person living with dementia or their carer) included

• systematic assessment of participant medication use, with a documented protocol

• a well-characterised sample, recruited from a defined population with clear inclusion and exclusion criteria

• sufficient sample size (in particular sufficient numbers of participants taking Z-drugs).

Data sources explored, with at least one primary outcome measure, are described in Table 1. Those which were not deemed suitable for use in our study were not explored further from initial descriptive analysis. Three data sources that fulfilled the required criteria underwent analysis of patient-reported outcomes. These data sources were:

1. the Resource Use and Disease Course in Dementia – Nursing Homes (REDIC) Norwegian observational longitudinal study

2. the University of Washington’s National Alzheimer’s Coordinating Centre (NACC) clinical data set

3. the Improving Well-being and Health for People with Dementia (WHELD) RCT based in UK nursing homes.

Setting

The REDIC study is a longitudinal cohort study of people admitted to nursing homes in Norway. The REDIC study is fully described elsewhere,133 but in brief the study began in 2012 and recruited 696 patients at admission to one of 47 nursing homes across four Norwegian counties.

Baseline assessments were made within 1 month of admission to the nursing home and then every 6 months thereafter. Comprehensive REDIC study assessments included questions on sleep disturbance, medication use and measures of cognitive function [recorded by the Mini Mental State Examination (MMSE) and Severe Impairment Battery (SIB-8)], QoL [measured by the Quality of Life in Late-Stage Dementia (QUALID) scale and EuroQol-5 Dimensions (EQ-5D)], neuropsychiatric symptoms [measured using the 12-item Neuropsychiatric Inventory – Nursing Home (NPI-NH)] and depression [measured using the Cornell Scale for Depression in Dementia (CSDD)]. These are described in detail in Outcome variables.

We received data from the REDIC study team for assessments up to October 2016, including up to five assessments per participant. Assessments were conducted and supplied in Norwegian, with translation kindly provided by Professor Sverre Burgh, the REDIC study chief investigator.

Participant inclusion and exclusion criteria

Participants were eligible for inclusion in the REDIC study if they had an expected stay in the nursing home for > 4 weeks and were excluded if their life expectancy was < 6 weeks. Participants were included in the Z-drug Evaluation in Dementia (ZED) study if they had a diagnosis of dementia or mild cognitive impairment (MCI) on admission. These diagnoses were made and recorded by the REDIC study team on the first visit after admission to the care home.

Medication exposures

Regularly used medication was recorded at each wave, coded according to WHO ATC code and dose, for up to 18 medications per participant. We coded the presence or absence of three classes of hypnotics at each visit, based on the following codes: Z-drug (Z) – ATC N05CF; BZD – ATC N05BA, N03AE or N05CD; and antipsychotic – ATC N05A.

Outcome variables

Confounders

Recoding of outcome measures for analysis

Statistical methods

The sample is described with respect to distribution of exposures, outcomes and covariates at baseline, stratified by dementia severity. The number of participants completing each visit is described and the predictors of dropout at subsequent assessment are estimated using logistic regression. The distribution of each outcome with respect to age, visit number, total number of visits completed, CDR and concurrent Z-drug use is then described using data from all pooled together.

Setting

The NACC data set includes standardised longitudinal patient-level data from Alzheimer’s disease centres (ADCs) across the USA. Data include clinical evaluations and records of medication use at the time of assessment. Patients are assessed annually following referral to an ADC, with the first patient entering the study in 2005, and up to 12 assessments recorded for some participants. Full details of the NACC data set and the uniform data set that is collected at each assessment are found in successive publications describing the evolution of the resource over the years and the NACC website (URL: www.alz.washington.edu/; accessed 6 December 2019). The NACC data set is widely used for dementia research and has been previously used for looking at the impact of medication use on different outcomes among people with MCI or dementia. 156 We requested data from every assessment of all participants included in the NACC data set, with an updated data extraction supplied to us in August 2018. This data set included 131,354 visits among 38,249 unique participants.

Participant inclusion and exclusion criteria

For the current analysis, we included only participants with a study diagnosis of dementia at any time and included only their assessments from the point at which they are first recorded as having dementia. Whether or not the participant meets clinical criteria for dementia is recorded at each assessment. This is made according to clinical judgement or by specific criteria, depending on the version of assessment. In analysis we included only those visits when there is a clinical diagnosis of dementia recorded in the study (NACC coded data set variable ‘demented’ takes the value of ‘1’). This reduced data set included 43,286 annual visits among 17,055 unique participants.

Medication exposures

At each annual visit, the medications each participant had used in the 2 weeks before the visit were recorded. Participants were asked to bring medication to the NACC assessment or a detailed list of their medications. When this was not available, it was followed up by a telephone call following the assessment.

Outcome variables

Statistical methods

Our analysis of the NACC data set follows closely the analysis of the REDIC study data set. First, the characteristics of participants are described stratified by the CDR-global score and the distribution of the total number of study visits with dementia for each participant is shown.

Software

Clustered linear models and weighted linear models are estimated using the ‘survey’ package and fixed-effects linear models are estimated using the ‘plm’ package version 1.6-6 (linear models for panel data) for R statistical software version 3.5.0 (The R Foundation for Statistical Computing, Vienna, Austria). The data set was cleaned and coded using R as well as Stata. All other analysis was conducted using R, with the ‘stargazer’, ‘dplyr’ and ‘ggplot2’ packages for data manipulation, managing outputs and plotting, respectively.

Setting

This study was undertaken using anonymised data from the WHELD cluster randomised controlled two-arm trial, covering 69 care homes in England. 128,159 The RCT was carried out between January 2013 and September 2015, with a primary aim to determine whether or not an optimised staff training intervention improved the QoL and mental health of people with dementia living in nursing homes.

Participant inclusion and exclusion criteria

Care homes and their residents were eligible for inclusion in the WHELD RCT if the:

• care home has > 60% of residents with dementia

• care home was not receiving special support from their local authority

• care home met the five CQC care home quality standard checks

• resident had a diagnosis of dementia or had a score of ≥ 1 on the CDR.

Care homes and their residents were excluded from the WHELD RCT if:

• the care home had insufficient staffing resources or anticipated major change during the study period

• the care home was involved in other research or undergoing a systematic programme of service improvement

• consent or advice from a consultee could not be obtained for the resident.

A total of 971 participants met the inclusion criteria, with 504 participants randomised to receive treatment as usual and 467 to receive the WHELD intervention between January 2013 and April 2014. Follow-up assessments were available for 553 participants 9 months later and non-completion was mainly due to mortality.

Participants were not included in the ZED analysis if it was not possible to record their medication usage at baseline or had a reported diagnosis of severe mental illness (schizophrenia or bipolar disorder) at any assessment.

Medication exposures

Antipsychotic and other psychotropic drugs taken at each assessment were classified according to the BNF. 120 Drugs were coded as non-BZD hypnotics, BZDs, antidepressants, carbamazapine, sodium valproate, other anticonvulsants, memantine, cholinesterase inhibitors, barbiturates, clomethiazole, nuspirone and others.

Outcome variables

The following outcomes were assessed prior to randomisation and after 9 months of the intervention.

Confounding variables

The following confounding variables were considered at baseline: age, sex, ethnicity (white or not), marital status (single/widowed, married/long-term partner, divorced/separated), CDR (mild, moderate, severe), sleep disturbance (a NPI-NH sleep score > 0), Abbey Pain Scale score (0–2 = none, 3–7 = mild, ≥ 8 = moderate/severe), comorbidity (depression, anxiety, respiratory illness, gastrointestinal illness, cardiovascular condition, endocrine illness, musculoskeletal disorder and nervous system illness) and co-medication use (BZD, meprobamate/buspirone, clomethiazole, antidepressant, antipsychotic cholinesterase inhibitor and memantine).

Statistical methods

Care home participant characteristics were described for those who were and were not taking Z-drugs at baseline. Negative binomial regression was used to estimate IRRs for the association between Z-drug use at baseline and QUALID, NPI-NH excluding sleep and NPI-NH sleep scores at baseline. Logistic regression was used to estimate ORs for the association between baseline Z-drug use and mortality by 9-month follow-up. Linear regression models were used to estimate the mean decline in QUALID, NPI-NH excluding sleep and NPI-NH sleep scores from baseline to 9-month follow-up for Z-drug use compared with no use at baseline, as well as for continuing, starting or stopping Z-drug use between baseline and 9-month follow-up compared with no use. All associations are presented both unadjusted and adjusted for the covariates listed above. For the analysis comparing changes of Z-drug use across the 9-month follow-up, changes in the use of antidepressants, memantine, antipsychotics and BZDs were also adjusted for.

Ethics approval

The primary care study was approved by the Independent Scientific Advisory Committee for CPRD for CPRD research (reference 16_181). No further ethics approval was required for the analysis of the data. The CPRD group has obtained ethics approval from a Multicentre Research Ethics Committee for all purely observational research using CPRD data.

The Regional Ethics Committee for Medical Research in South-Eastern Norway approved the REDIC study (reference 2011/1738a). The patient’s capacity to consent to participation in the study was considered by the nursing home staff, including the physician. Written consent for participation was obtained from all participants with the capacity to consent. For participants lacking the capacity to consent, their next of kin gave consent on behalf of the patients. The next of kin gave written consent for their own participation in the study, as they provided information about themselves.

All participants or their legally authorised representatives signed informed consent prior to participation in the NACC study. The institutional review board overseeing each ADC approved local procedures.

Ethics approval for the WHELD RCT was obtained from the South-Central Oxford Research Ethics Committee C (reference 11/SC0066). The trial is registered as a clinical trial (ISRCTN40313497). Consent for nursing home involvement was obtained from the management of the homes. If residents lacked capacity, informed consent was obtained through the involvement of a nominated or personal consultee who represented the residents’ interests and wishes in accordance with the Mental Capacity Act. 160

Funding

The NIHR HTA programme (reference 14/221/02) funded the Z-drugs in dementia (ZED) study.

Sponsorship

The University of East Anglia was the study sponsor.

Absolute risks

The absolute annual risks of fractures, hip fractures and mortality in the unexposed group were 7.4%, 3.2% and 15.4%, respectively. These equate to annual risks of 10.2%, 5.0% and 20.1% for fractures, hip fractures and mortality when taking Z-drugs if our estimated risks are causal. This is equivalent to a NNH of 36, 54 and 21, and extra cases per 1000 treated of 28, 18 and 47, respectively.

Dose–response

When examining outcomes by cumulative Z-drug dose, we found that fracture risk increased with cumulative exposure to Z-drugs (Table 8). During periods when cumulative prescriptions totalled < 28 DDDs of Z-drugs, equivalent to 56 days at the recommended half-dose in the elderly, there were no excess risks of fractures. For 42–55 prescribed DDDs the HR for a fracture increased to 2.81 (95% CI 1.47 to 5.37).

The picture was similar for hip fractures, with no excess risk when prescribed < 28 DDDs of Z-drugs, rising to a HR of 3.33 (95% CI 1.41 to 7.91) when cumulatively prescribed 42–55 DDDs. There were few forearm, wrist or hand fracture outcomes, but there was evidence of a similar pattern to that of fractures with no risk before 28 DDDs and greater risks after. There was evidence of increased falls risks after > 28 DDDs of Z-drugs had been prescribed, but this association disappeared after 56 DDDs were prescribed. The risk of mortality was generally fairly similar, regardless of cumulative exposure to Z-drugs. In addition, we found no evidence of differences in risk according to whether or not the first Z-drug prescription was PRN (see Appendix 5, Table 57).

Dose–response

When examining outcomes by cumulative Z-drug dose, we found some suggestion of increased acute bacterial infections with cumulative exposure to Z-drugs (Table 10); however, this was only during periods when cumulative prescriptions totalled 28–41 DDDs and not for > 42 DDDs. There was no evidence of increased risks of ischaemic stroke, TIA, or venous thromboembolism with cumulative Z-drug use. However, the incidence of agitation or psychosis increased with greater cumulative Z-drug prescriptions, such that for 42–55 prescribed DDDs, the HR increased to 2.55 (95% CI 1.31 to 4.98). In addition, we found no evidence of differences in risk according to whether or not the first Z-drug prescription was PRN (see Appendix 5, Table 57).

Participant characteristics

Of the 696 participants recruited to the REDIC study, 17 participants had no dementia or MCI at baseline and one participant did not provide enough information at their baseline assessment to be included in the current analysis. Therefore, these participants were excluded from the current analysis, leaving 95 participants with MCI and 583 participants with dementia at baseline.

The number with usable assessments at each wave, along with the prevalence of use of each hypnotic reported at each wave, is shown in Appendix 3, Table 42. Forty-two per cent of the sample gave assessments at visit 5 (2 years after baseline). The rate of Z-drug and BZD use was high, with around 20% of participants reporting Z-drug use, 20% reporting BZDs and up to 20% reporting antipsychotics at each visit. Antidepressants were also widely used (30–40% of participants at each wave), with lower numbers reporting antihistamines or antiepileptics. The distribution of key outcome measures and dementia severity at baseline stratified by Z-drug, BZD and antipsychotic use is shown in Table 13.

The vast majority of those who left the study between waves had died, but some had refused to continue, were withdrawn by their nursing home, moved home or had moved to a nursing home not participating in the REDIC study.

Dynamics of hypnotic use throughout the study

The ability to estimate the effect of hypnotics is ultimately derived from within-person comparisons of hypnotic use compared with no hypnotic use. Hence, the number of transitions between hypnotic use and no hypnotic use within individuals between visits is crucial. The number of such transitions is shown for Z-drugs, BZDs and antipsychotics in Table 14, between each of the 1540 pairs of consecutive visits in the REDIC study data set. There is a positive correlation between use of pairs of medications (e.g. out of 421 occasions where Z-drugs were reported, BZDs were also reported on 126 occasions) (Table 15).

Predictors of dropout and mortality

A logistic regression model was applied with drop-out at the next wave as the outcome, and hypnotic use, outcomes and demographics covariates as predictors (see Appendix 3, Table 43). The vast majority of dropouts from the study were due to death and so the different forms of outcome were not modelled individually. Clustered standard errors were used to account for multiple data points per individual contributing to this analysis.

After adjusting for covariates, those who used Z-drugs were less likely to complete the next assessment (OR 1.47, 95% CI 1.03 to 2.09), although those using BZDs were more likely. Antipsychotic use did not predict subsequent loss to follow-up independently of disability and cognitive function (as measured with CDR-global). Those who were more severely disabled or older were more likely to be lost, but net of this there was no clear link between dropout and cognitive function. There was little association between neuropsychiatric symptoms and subsequent loss to follow-up.

Predictors of hypnotic use

Tables 14 and 15 show that medication use at a previous wave is an extremely strong predictor of medication use at the current wave. Hence, for each medication of interest two logistic models were estimated, one for stopping and another for starting, conditional on participant characteristics and lagged values of time-varying covariates. Clustered standard errors were used to account for multiple data points per individual contributing to this analysis.

Table 16 shows the independent predictors of starting (current use conditional on no prior use) and Appendix 3, Table 44, shows the factors predicting continuing medication use (current use conditional on prior use) estimated using these models, which are also used to identify confounding variables to develop the marginal structural model analysis below.

Z-drugs and BZDs are more commonly started among those with less severe dementia, more strikingly so for Z-drugs than for BZDs. Conversely, antipsychotics are more commonly started among those with more severe dementia. Prior use of Z-drugs predicts new use of BZDs. Prior levels of neuropsychiatric symptoms independently predict new use of hypnotics, in particular prior sleep disturbance predicts new Z-drug and BZD use, anxiety predicts new use of all three classes. However, a considerable number of participants started Z-drugs without having any prior sleep disturbance, anxiety or agitation recorded in any previous assessment (not shown).

There were fewer clear predictors of continuing hypnotic use (equivalently of stopping use) between visits. In particular, levels of neuropsychiatric symptoms and dementia severity were not significantly linked to use of any hypnotic if there was use at a previous wave.

Dynamics of outcome measures and estimating the effect of hypnotic use on outcome measures

Figures 13–22 in Appendix 3 describe how each outcome measure changes between waves and vary with age, CDR stage and with concurrent Z-drug use. The left-hand panels of Figures 3–12 show how outcome measures change with changing hypnotic use status. The right-hand panels of Figures 3–12 show the association between current hypnotic use and outcome measures, with weights applied to each observation such that levels of previous hypnotic use, cognitive function and neuropsychiatric variables are balanced between those currently using each medication and those not using each medication. Thus, these graphs show the effect of current medication use controlling for key patient characteristics, in particular those characteristics that predict new or continuing medication use at the previous wave. The difference in current mean levels of outcomes between groups can then be interpreted as the effect of the medication, although when weights do not cause prior levels to perfectly coincide, comparing the change over time between groups may better estimate the effect.

Participant characteristics

The distribution of the characteristics of the 17,055 people living with dementia on study entry (defined as their first assessment with a record of dementia) stratified by baseline CDR-global is shown in Table 18. The median year of joining the study is 2009 and the median age is 76 (range 21–110, IQR 67–82) years. Fifty-two per cent of the cohort were female. There is little association between age at study entry and dementia stage at entry. Although most participants join the study at dementia stage ≤ 1, 14% joined at CDR 2 (moderate dementia) and 8% at CDR 3 (severe dementia).

The reported prevalence of Z-drug and BZD use is far lower in NACC data set than in the REDIC study cohort. Only 2–3% of participants report Z-drug use, whereas 7–15% report BZD use, with highest levels of use among those with severe dementia. Antipsychotic use increases markedly with dementia severity, from 3% (CDR < 1) to 34% (CDR = 3). A substantial proportion of participants at all stages (35–45%) report use of an antidepressant. On fewer than 2% of visits the medication use form was not completed (3% among CDR 3 patients).

Neuropsychiatric symptoms are reported relatively frequently at all stages. Some sleep disturbance is reported for 26–41% of participants across groups; anxiety and agitation have similar numbers. These assessments are present for 96% of participants in CDR stages 0.5–2 and 88% at CDR stage 3. The average total NPI symptom score excluding sleep disturbance increases from 3.5 to 7.6 with dementia severity.

Average disability score increases with dementia severity, such that almost all of those with severe dementia score 29 or 30 out of 30 on the disability scale. In contrast, the level of depression is generally low (median GDS 2/10, IQR 1–4), but in severe dementia 77% of participants are not able to complete the GDS.

The number of participants completing each annual visit is shown in Appendix 4, Table 48. Only 9889 (59%) participants complete more than one visit, 4020 (24%) of participants complete visit 4, whereas only 838 (5%) complete visit 7. Prevalence of Z-drugs does not change markedly with visit number, whereas antipsychotics become more common at later visits.

Dynamics of hypnotic use

The numbers of participants starting and stopping Z-drugs, BZDs and antipsychotics between pairs of consecutive visits are shown in Appendix 4, Table 49. Only 1% of visits report new Z-drug use, although this still corresponds to 219 instances of new Z-drug use, which is a reasonable number from which to estimate changes in outcome measures. Of those who use Z-drugs and return medication data at a subsequent visit, 51% still report Z-drug use. For antipsychotics, the rate of new use (6%) and continuing use (79% of antipsychotic users still use antipsychotics at the next wave) is higher.

Predictors of starting or stopping hypnotics between waves

Table 19 and Appendix 4, Table 50, show the independent effects of cognitive function, neuropsychiatric variables and demographic factors at each visit on the use of hypnotics at the subsequent visit, conditional on current hypnotic use.

Among those not using Z-drugs, sleep disturbance did predict new use of Z-drugs at the next wave, but was only slightly linked to new BZD use and was not independently linked to new antipsychotic use. Conversely, anxiety and agitation predicted new use of antipsychotics and BZDs at the subsequent visit. Cognitive function did not predict new use of Z-drugs, but those with more severe dementia were more likely to start BZDs and antipsychotics (see Table 19). Independently of cognitive function and symptoms, older people with dementia were less likely to start BZD or antipsychotics, but there was no age effect with Z-drugs.

Among those using hypnotics there were very few variables that were able to predict which participants continued to report hypnotics at the next visit (see Appendix 4, Table 50). Older people and those with more severe cognitive impairment appeared less likely to continue BZDs, but, crucially, current levels of psychiatric symptoms did not predict whether or not current medications were continued.

Predictors of dropout in the NACC data set

As in the REDIC study, medication use did not strongly predict whether or not each participant completed a subsequent assessment, but older and more severely impaired participants were more likely to drop out. Psychiatric symptoms also did not predict dropout. Note, we do not distinguish dropout through leaving the study, death or censoring and so this should not be interpreted as an assessment of the risk of hypnotics on any outcome.

Distribution and dynamics of outcome measures

Appendix 4, Figures 23–32, describes how each outcome measure changes between waves and varies with age, CDR stage and with concurrent Z-drug use. Delta trail time does not appear to work well among those with severe cognitive impairment. The score is missing for the vast majority of participants with CDR stage 3, and for those who do return a score it is often at the maximum allowed time for both trails; hence the information provided by this measure is likely to be extremely limited.

Association between each outcome and starting or stopping hypnotic medication

Fixed-effects models

Each of the results discussed above using marginal structural models is also reflected in the fixed-effects models (see Appendix 4, Tables 52–54). There is a strong, consistent association seen between worse cognitive function and greater disability with concurrent BZD and antipsychotic use, but no association with Z-drug use. Depression is not linked to the use of any hypnotics, except for a positive association with antipsychotics among those with no hypnotic use on study entry, whereas other neuropsychiatric symptoms are consistently positively associated with use of hypnotics, as above.

Participant characteristics

After excluding 21 participants with missing medication data and 24 participants with a severe mental illness diagnosis, the 926 participants in the WHELD trial with dementia were mainly women (n = 659, 71%) and the mean age at study entry was 85.3 (SD 8.8) years. A total of 123 (13%) participants reported Z-drug use at baseline and a further 27 (3%) participants reported Z-drug use at 9-month follow-up. Those participants taking Z-drugs were more likely to have a partner, have a nervous system illness and be taking concurrent BZDs, memantine or antipsychotics (Table 23).

Cross-sectional analysis

The results of the cross-sectional analyses comparing Z-drug use at baseline and baseline QoL and sleep scores are shown in Table 24.

At baseline, the mean QUALID scores for those taking Z-drugs or not at baseline were 22.9 (SD 8.0) and 21.3 (SD 7.3), respectively. Z-drug use was not associated with any difference in QUALID scores at baseline (RR 0.96, 95% CI 0.92 to 1.01, adjusted for confounders).

At baseline, the mean NPI-NH sleep scores for those taking Z-drugs or not at baseline were 1.6 (SD 3.0) and 1.0 (SD 2.5), respectively. There was no significant difference in sleep disturbance at baseline in Z-drug users (RR 1.42, 95% CI 0.71 to 2.85, adjusted for confounders).

At baseline, the mean NPI-NH score (excluding sleep) for those taking Z-drugs and not taking Z-drugs at baseline were 21.6 (SD 17.6) and 14.1 (SD 15.1). Greater neuropsychiatric symptoms were reported in those taking Z-drugs after adjusting for confounders (RR 1.24, 95% CI 1.00 to 1.54).

Mortality analysis

A total of 20 (16%) and 177 (22%) participants taking and not taking Z-drugs, respectively, at baseline died during follow-up. Baseline Z-drug use was not associated with mortality after adjusting for confounders (OR 0.66, 95% CI 0.38 to 1.15).

Longitudinal analyses

A total of 627 people living with dementia completed the 9-month follow-up interviews. The mean increase in QUALID, NPI-NH sleep score and NPI-NH excluding sleep over the 9 months’ follow-up was 0.25 (SD 7.67), 0.11 (SD 0.60) and 0.17 (SD 16.59), respectively. Z-drug use at baseline was not significantly associated with any greater increase in either score, with adjusted additional average increases in QUALID, NPI-NH sleep, and NPI-NH excluding sleep scores of 1.43 (95% CI –0.33 to 3.19), 0.08 (95% CI –0.07 to 0.22) and 0.60 (95% CI –3.26 to 4.46), respectively (Table 25).

Considering changes in medication use between baseline and 9-month follow-up, 26 (4%) reported Z-drug use at baseline but not at 9 months (‘stopping’), 62 (10%) participants reported Z-drug use at both times (‘continuing’) and 25 (4%) reported Z-drug use at 9 months but not at baseline (‘starting’). Although those starting, stopping and continuing Z-drugs reported, on average, increased NPI sleep scores over the 9-month follow-up, none were statistically different from the increase for those not using Z-drugs (see Appendix 6,Table 65). Although those initiating Z-drugs saw a decline in their NPI score (excluding sleep), this was not statistically different from those not using Z-drugs. Although those stopping Z-drugs reported a worsening in their QoL over the 9-month follow-up, this was not statistically different from those not using Z-drugs.

Primary care study

By examining linked primary care and hospital admission data we found evidence of an increased risk of fractures and, in particular, hip fractures, in people with dementia taking Z-drugs. The risk also increased with cumulative exposure to Z-drugs, suggesting a causal relationship. However, multiple outcomes were examined, increasing the risk of false-positive findings. We observed a greater mortality rate in people living with dementia who were prescribed Z-drugs; however, the association did not vary significantly by cumulative exposure to Z-drugs, suggesting that this finding may be due to reverse causation (i.e. that patients in later stage dementia were more probably prescribed Z-drugs).

We did not detect a significantly increased risk of stroke, infection or venous thromboembolism with Z-drug use. We observed that people living with dementia who were prescribed Z-drugs were more likely to be further prescribed antipsychotics, antidepressants and antibiotics. People living with dementia who were prescribed Z-drugs also visited their GP more frequently and were more often admitted to hospital in the next 2 years.

Our study was not sufficiently powered to examine adverse events of BZDs, but the risk estimates for Z-drugs and fractures and mortality were generally similar to those for BZDs.

We did not find any associations between the use of low-dose TCAs and adverse events; however, these findings should be interpreted with caution, as we were unable to distinguish whether patients had been prescribed low-dose TCAs for sleep disturbance or for chronic pain. Patients with chronic or neuropathic pain may be at a lower risk of fractures than patients with sleep disturbance.

Clinical cohort studies

In cohort studies of people with dementia in Norway (the REDIC study) and the USA (the NACC data set), we observed no impact of Z-drug use on cognition, neuropsychiatric symptoms, disability or QoL. There was some suggestion of an association between recurrent Z-drug use and lesser declines in cognition and disability; however, we believe that this is more likely to reflect the people living with dementia who are stable on their Z-drugs and not progressing onto using alternative medications, such as antipsychotics or BZDs. In a cohort study of UK care home residents with dementia (the WHELD trial), we observed greater neuropsychiatric symptoms in those taking Z-drugs at baseline, but Z-drug use at baseline was not associated with greater improvement in neuropsychiatric symptoms over the following 9 months. We also observed no greater mortality risk in those taking Z-drugs.

Primary care study

The primary care study is the first study to examine a range of adverse events for Z-drug use in people with dementia. We used data from a large number of patients in a population-representative primary care database linked to hospital records, which allowed detailed analysis of the prescriptions of Z-drugs and patient-relevant outcomes. We designed the study to aim to minimise possible sources of bias and consequently this reduced the number of patients and follow-up in the study, reducing our statistical power to detect adverse events. We shall discuss the strengths and limitations in the context of the main sources of bias. 161

Clinical cohort studies

Generally, there were few studies reporting the outcomes of interest with sufficient Z-drug use in people living with dementia. However, the three included cohort studies were of a large size, with frequent follow-up and detailed outcome recording. Our results varied, and this could reflect differences in the prescribing of Z-drugs across the USA, UK and Norway, and the settings of nursing homes or the community. We were limited by not having a record of the outcome measurements just before the Z-drug was first prescribed. As each study measured outcomes at fixed intervals, this did not necessarily coincide with the initiation of Z-drugs or the start of a sleep disturbance. We were limited by which outcome measures each study recorded. With NPI, it is not obvious whether a respondent is reporting on the level of symptoms that are currently experienced given current treatment, or those that they believe would be experienced if the patient was not treated. Hence, these findings should be interpreted with caution.

Although in the REDIC study and the NACC data set our analysis was strengthened by using marginal structural models to balance users and non-users on prior covariates, the extent of balance was not always perfect and could not be improved using different models for estimating weights. However, balance was very good for prior CDR, suggesting that our findings are not driven by underlying differences in dementia severity. Perhaps more importantly, this approach is valid for causal inference only if the assumption of exchangeability is met, that is after balancing over the covariates included in our model there is no longer any indirect path from the exposure in question to the outcome measure. 175 In our case it is difficult to be certain of this and, as the exact timing of medication initiation is not known, it is impossible to know whether the outcome preceded each exposure or vice versa.

Fractures

In the primary care study, we found evidence of a 40% greater risk of fractures with Z-drug use in people living with dementia, which is lower than that reported in studies in the older population. A systematic review, including a subgroup analysis of the findings from five observational studies in the older population, estimated a pooled relative risk for fractures of 1.73 (95% CI 1.31 to 2.27) with zolpidem use. 176 However, none of these studies accounted for sleep disturbance, which is the indication for zolpidem. A further systematic review and meta-analysis investigated Z-drugs in general and risks for falls and fractures from 14 eligible observational studies. 60 They estimated an OR of 1.70 (95% CI 1.36 to 2.12) for Z-drugs and fractures in the older population, but noted substantial heterogeneity. 60

Although several studies have examined the risk of fractures associated with Z-drug use in the older population, studies targeting people living with dementia have been limited. 36 A recent Cochrane systematic review concluded a distinct lack of evidence to help guide drug treatment of sleep problems in dementia,36 despite being widely prescribed to this large patient group. 177 To our knowledge, no studies have previously examined the risk of general fractures specifically in people living with dementia taking Z-drugs.

Current BZD use was not found to be associated with fractures in a study of 8036 people living with dementia in London. 41 We observed similar rates of fractures among new BZD and new Z-drug use. Other studies in the older population that compare BZDs with Z-drugs for fracture risk, report varying results, often dependent on the type of BZD. 59,178

We found evidence of a 59% greater risk of hip fractures with Z-drug use in people living with dementia. A systematic review and meta-analysis pooled the findings from six observational studies in older adults and estimated a 90% (95% CI 68% to 113%) greater risk of hip fracture with Z-drug use. 61 However, studies suggest that the risk is lower in people living with dementia. 57,179 A Finnish study of 67,072 community-dwelling people diagnosed with Alzheimer’s disease reported no association between any history of BZD or Z-drug use in the 5 years before study entry and incident hip fracture (HR 1.03, 95% CI 0.97 to 1.09); however, this exposure definition may not be sufficiently specific and would not necessarily cover medication use immediately prior to most of the incident hip fractures (i.e. was measured only at baseline and not updated during follow-up). 179 A study of 15,528 US nursing home residents estimated an OR for non-BZD hypnotic drug use and hip fracture in residents with moderate to severe cognitive impairment of 1.43 (95% CI 1.15 to 1.77), which was lower than in the residents with no or MCI (OR 1.86, 95% CI 1.56 to 2.21). 57

We observed that hip fracture risk increased with greater cumulative Z-drug exposure, but then tapered after 56 DDDs. This is consistent with an initial expression then down-regulation of GABA receptors,180,181 or a depletion of susceptible patients among long-term users. 182 Other studies in the older population found similar findings, but with risks tapering over shorter periods, than we did in dementia. 59,183 This perhaps represents delays getting the prescription dispensed or many people living with dementia not taking their prescribed Z-drugs during the first month in our sample, or a slower effect of Z-drugs in dementia. We found no effect in the first 14 days, contrary to other studies in the older population. 54,56,57 This could be partly explained if some patients receiving only one prescription did not take the Z-drugs. Excluding those with only one Z-drug prescription had little effect on our cumulative Z-drug prescription findings (results available from the authors).

We observed evidence of a greater risk of hip fractures for those prescribed BZDs than Z-drugs, but numbers of hip fractures were small and this could represent residual confounding by dementia severity. Studies differ on whether they find those prescribed BZDs or Z-drugs at greater risk of hip fracture,55,56,58,184 but generally suggest greater risks with long-acting BZDs than with Z-drugs. 56,185

Evidence is lacking to indicate that Z-drugs increase the risk of osteoporosis,186 in contrast to other psychotropic drugs (e.g. antidepressants and antipsychotics). 187 Although studies suggest that long-term sleep disturbance may increase osteoporosis risk though disrupting the rhythm of bone turnover. 188,189 The mechanism by which Z-drugs increase fracture risk is probably due to their effects on gait and balance, although this is contradictory to our findings on falls outcomes (see Falls). 59,186 A trial randomised 25 adults to 5-mg zolpidem or placebo 10 minutes before scheduled sleep and asked them to perform 10 tandem walks on a beam on night-time awakening. 190 The NNH was estimated as one tandem walk failure for every 1.7 (95% CI 1.4 to 2.0) older adults treated with zolpidem. The trial authors concluded that the ‘use of nonbenzodiazepine hypnotic medications may have greater consequences for health and safety than previously recognized’. 190

In a comprehensive review of the evidence for major adverse outcomes of BZDs and Z-drugs, there was compelling evidence to establish a causal connection between BZD or Z-drug use and only motor vehicle accidents, falls and fractures. 186 Our study suggests that the increased risk of fractures extends to those people living with dementia taking Z-drugs.

Falls

Counterintuitive to the fractures findings in the primary care study, we did not find an increased falls risk with Z-drug use. Similarly, current BZD use was also not found to be associated with falls in a study of 8036 people living with dementia in London, once accounting for differences in ethnicity, living environment and disability. 41 A systematic review published in 2008, reported modest associations (RRs 1.2–1.3) between BZDs and falls in nursing home residents with dementia,191 and BZDs are consistently associated with around a 20% increased falls risk in older people. 48

Z-drugs were originally claimed to cause fewer falls,51 but a recent meta-analysis estimated an OR for Z-drug use and falls of 2.40 (95% CI 0.92 to 6.27) when pooling the findings from three observational studies in adult populations. 60 However, there was considerable heterogeneity (97%) between the studies and reasons for this were cited as difficulty recording falls, particularly in patients with multimorbidity, or recall bias in the studies relying on self-reported falls. Among 4450 US community-dwelling older men enrolled in the Osteoporotic Fractures in Men study, Z-drug use was associated with a 37% (95 CI 9% to 71%) greater risk of a self-reported fall. 192

Falls are thought be the main mechanism whereby increased fractures and injury occur in Z-drug users resulting from the sedation effect. Falls also account for 90% of hip fractures. 193 A study of older people in Taiwan found that high doses of Z-drugs were associated with an increased risk of fall-related injuries requiring hospitalisation (OR 1.24, 95% CI 1.05 to 1.48), so the lack of association between Z-drug use and falls in our study might be considered surprising. 53

To gain a potential clinical explanation we discussed this finding with our health-care advisory panel, who suggested that a fracture would be coded in electronic records preferentially over falls as the cause for their referral or reason for admission to hospital. This does suggest that in such cases, if a fracture is a direct result of a fall, it could be missed in our analysis due to our reliance on the complete recording of all required coding during contact with a GP or hospital. Alternatively, although the rate of falls may have been similar between groups, it is possible that the nature of the falls may have been more severe or of greater physical impact in patients taking Z-drugs.

Mortality

We observed a 34% increased mortality rate in people living with dementia prescribed Z-drugs, using primary care data; however, no association in UK nursing homes (the WHELD trial). Other studies have been conflicting on the effect of hypnotic use on mortality. A study of 31,140 community-dwelling people with Alzheimer’s disease in Finland, found that BZD use was associated with increased risk of mortality (HR 1.59, 95% CI 1.35 to 1.88), but not Z-drug use (HR 1.06, 95% CI 0.83 to 1.35), after accounting for psychotropic drug use, comorbidities, socioeconomic status, hip fractures and stroke. 75 In a US study,74 crude mortality rates were higher in people living with dementia taking hypnotics, but the association was not tested statistically.

A systematic review and meta-analysis examining the effect of hypnotics on mortality estimated a HR of 1.73 (95% CI 0.95 to 3.16) when pooling the findings from five observational studies on Z-drug use and mortality risk in adults. 71 However, there was considerable heterogeneity (I² = 97%), with three studies reporting no effect and two with HRs > 3. Two of these studies used similar UK primary data to our study. One matched 2435 Z-drug initiators to two non-initiators on age, sex and general practice in the UK adult population during 1998–2001, and reported a HR for mortality of 3.32 (95% CI 3.19 to 3.45), adjusted for age, sex, comorbidities and co-medications, with a fairly constant risk across cumulative Z-drug prescriptions, similar to our study. 194 A further study of 4964 UK patients with pneumonia showed that patients were not at a greater mortality risk if taking zopiclone (HR 1.11, 95% CI 0.77 to 1.60). 64 Studies on Z-drug or BZD use and mortality have been conflicting,64,70–72,194 and reported associations may simply stem from confounding. 73 A large database study of US adults observed an increased mortality rate in 1.25 million BZD initiators matched to 1.25 million patients not initiating BZDs, but not once accounting for comorbidities, concomitant medications and health-care utilisation via propensity score matching (HR 0.89, 95% CI 0.85 to 0.94 in those aged > 65 years). 195

We found the same risk of mortality regardless of cumulative Z-drug exposure from our dose–response analysis, suggesting that the relationship may not be causal between Z-drug use and mortality risk in this case. We suspect that our findings are at least partially due to residual confounding by indication, whereby the cohort of patients on Z-drugs are already at increased risk of death. 73 However, it could also be speculated that the increased risk of hip fractures in this group drives an increased mortality rate, given that older adults have a five- to eightfold increased risk for all-cause mortality during the first 3 months after a hip fracture. 196

Infections

In the primary care study we did not find a statistically significant increased risk of infections in Z-drug users (HR 1.10, 95% CI 0.92 to 1.32 for UTI and LRTI). However, there was a suggestion of an increased infection risk after 28 DDDs of Z-drugs were prescribed, and we observed a greater rate of antibiotic prescribing. A Finnish study of community-dwelling adults with Alzheimer’s disease found that BZD use was associated with increased pneumonia rates and also found a very similar Z-drug use estimate to our study (HR 1.10, 95% CI 0.84 to 1.44). 67

A meta-analysis pooled data from published or Food and Drug Administration RCTs on 2432 participants randomised to eszopiclone and on 1626 participants randomised to zolpidem, estimated RRs of 1.48 (95% CI 1.25 to 1.74) and 1.99 (95% CI 1.21 to 3.26) for infections, respectively. 63 Observational studies in the adult population vary according to whether or not they detect an increased risk of infection with Z-drug use. A UK study using similar primary care to our study did not detect a significant association between Z-drug prescription and influenza-like illness-related pneumonia (HR 1.18, 95% CI 0.89 to 1.56). 66 A study in Taiwan found that patients using zolpidem with cumulative DDDs of 1–28, 29–84 and > 84 had HRs of 1.67 (95% CI 1.32 to 2.11), 1.91 (95% CI 1.47 to 2.49) and 1.62 (95% CI 1.32 to 1.98), respectively, compared with patients who did not use zolpidem during a 3-year follow-up. 65 However, there is probably residual confounding due to accounting only for a comorbidity score and few co-medications. A later case–control study in Taiwan in adults with chronic kidney disease found no difference in Z-drug use at the time of pneumonia (OR 1.07, 95% CI 0.80 to 1.44). 197

The mechanism whereby Z-drug use may increase infection is uncertain. It has been speculated that BZDs may suppress immune surveillance, which is supported by animal study findings. 63 Alternatively, Z-drugs could impair the clearing of oral secretions through inhibiting the swallowing of saliva. 68 Another possibility is that Z-drugs may increase gastro-oesophageal reflux events during sleep by relaxing the lower oesophageal sphincter. 198

Our study and the supporting literature suggest that if there is a greater risk of acute infections with Z-drug use in people living with dementia, then it is likely to be small and our study was inadequately powered to detect it.

Cardiovascular outcomes

In the primary care study we did not find a statistically significant increased risk of ischaemic stroke (HR 1.33, 95% CI 0.85 to 2.07) or venous thromboembolism (HR 1.66, 95% CI 0.69 to 3.98) to people living with dementia prescribed Z-drugs. However, there was a suggestion of an increased stroke risk after 28 DDDs of Z-drugs were prescribed. A study of 45,050 community-dwelling persons with Alzheimer’s disease in Finland reported that BZD or Z-drug prescription was associated with a 21% (95% CI 4% to 40%) increased risk of stroke, and this risk did not vary when stratified by BZD or Z-drug use. 77 However, residual confounding is possible as they were unable to account for sleep disturbance or dementia severity. A case–control study of adults in Taiwan found that exposure to zolpidem was associated with increased risks of ischaemic stroke (OR 1.37, 95% CI 1.30 to 1.44), increasing further with greater exposure to zolpidem. 78 They also found that regardless of whether or not a person presented with a sleep disorder, the risk of stroke was increased with zolpidem use (OR 1.37 without a sleep disorder and 1.41 with a sleep disorder), suggesting that the increased risk was due to drug use and not the sleep disturbance itself.

Mechanisms for Z-drugs causing increased stroke risk are uncertain, but speculated to relate to a decrease in local cerebral blood flow. 77 However, there is evidence that sleep disturbances increase stroke risk,199 suggesting that residual confounding by indication could underlie reported associations. We are not aware of any studies examining Z-drug use and risk of venous thromboembolism, although sleep disturbance has been observed to be a risk factor. 200

Prescriptions and health-care utilisation

In the primary care study, we found that people living with dementia who were prescribed Z-drugs were more likely to be initiated on antipsychotics and antidepressants. In line with our findings, a study in Taiwan reported that Z-drug users aged > 65 years were more frequently exposed to antipsychotics (OR 1.31, 95% CI 1.14 to 1.50), antidepressants (OR 1.45, 95% CI 1.28 to 1.65) and opioids (OR 2.38, 95% CI 2.07 to 2.73). 53 We also found that people living with dementia who were prescribed Z-drugs had more GP and hospital visits. This could reflect the increased risk of fractures in these patients, which would drive increased service use. The study in Taiwan also found that older adults who were prescribed Z-drugs were at an increased risk of fall-related injuries necessitating hospitalisation (OR 1.24, 95% CI 1.05 to 1.48). 53

Cognition, quality of life and sleep disturbance

In the clinical cohort studies, we found no evidence that Z-drug use impacts significantly on cognitive function or QoL. This differs from other studies in older adults with insomnia, which reported that hypnotic use is associated with reduced HRQoL. 101,102 Greater cognitive declines have been reported in people living with dementia taking BZDs;96 however, to the best of our knowledge, no other studies have examined the impact of Z-drugs on cognition in people living with dementia.

Estimating the effect of Z-drug use on sleep disturbance was challenging using the observational data sets, given that sleep scores were not measured immediately prior to Z-drug initiation; however, we observed no impact of Z-drug use on sleep scores. This is in line with other studies, suggesting that the use of sedating hypnotics in the general population does not improve sleep, human performance or general health, calling into doubt their effectiveness. 103,104