Understanding the association between antidepressants and the risk of being diagnosed with dementia in older people: a self-controlled case series analysis

BACKGROUND Given concerns about adverse outcomes for older people taking antidepressants in the literature, we investigated whether taking antidepressants elevates the risk of dementia. METHODS We conducted a population-based self-controlled case series analysis of older people with dementia and taking antidepressants, using territory-wide medical records of 194,507 older patients collected by the Hospital Authority of Hong Kong, to investigate the association between antidepressant treatment and the risk of developing dementia in older people. RESULTS There was a signicantly higher risk of being diagnosed with dementia during the pre-drug-exposed period (incidence rate ratio (IRR) 20.42 (95% CI: 18.66-22.34)) compared to the non-drug-exposed baseline period. The IRR remained high during the drug-exposed period (IRR 8.86 (7.80-10.06)) before returning to a baseline level after washout (IRR 1.12 (0.77-1.36)). CONCLUSIONS The higher risk of dementia before antidepressant treatment may be related to emerging psychiatric symptoms co-occurring with dementia, which trigger medical consultations that result in a decision to begin antidepressants. Our ndings do not support a causal relationship between antidepressant treatment and the risk of dementia.


Background
Dementia is a major cause of disability and dependency in older people 1 . It has become increasingly prevalent worldwide as the population ages, and it is associated with signi cant healthcare and economic burden. 1,2,3 However, there is convincing evidence that dementia is not inevitable in the course of ageing. 2 Signi cant efforts are thus being made to identify potentially-modi able risk factors of dementia in older people, especially those that in uence the early stage of the disease progression, when intervention might provide the most therapeutic bene t. 4 Potentially-modi able risk factors for dementia include hypertension, obesity, hearing loss, depression, physical inactivity, social isolation and diabetes. 5 Among the risk factors that have been studied, the relationship between antidepressants and dementia in the elderly remains contentious. For instance, whilst concerns have been raised that taking tricyclic antidepressants (TCA) poses an increased risk for the onset of dementia in older people 5,6 , there is con icting evidence that treatment of depression in the elderly using TCA is associated with a reduced, 7 or a negligible, 8 risk of dementia.
These controversial ndings are likely caused by the complex confounding effects on the association between taking antidepressants and dementia, of pre-existing mental health conditions, coupled with the high prevalence of comorbid mental health issues in patients treated with antidepressants. 9 For example, depression is a potential confounder for dementia, 10 and the disease processes itself may also lead to the onset of dementia. 10 Thus the reason for being prescribed antidepressants may underlie this argument. Moreover, antidepressants can also be prescribed for the management of health concerns other than depression. It has been reported that approximately 50% of antidepressants are prescribed for other conditions (such as off-label indications for prescription e.g. schizophrenia, insomnia or incontinence). 11 These conditions might indicate other issues of cognitive distress or body systems deterioration, for which diseases of the nervous system and cerebral cortex (including dementia) are important complications. [12][13][14] Consequently, research designs applied to test causal associations between antidepressants and dementia must take account of potentially-complex confounding variables. Research to date has used standard epidemiological case-control and cohort study designs to examine associations between antidepressants and dementia. These studies reported the possibility of causal relationships between antidepressant drugs and dementia, but controlled for none or only some, mental health-related disorders. [5][6][7][8] This paper reports the use of a self-controlled study design (Self-Controlled Case Series (SCCS)) developed to investigate associations between an exposure (antidepressant prescriptions) and an outcome of interest (dementia), using only data on cases. The within-individual inference enabled controlling for any xed covariate effect.

Ethics approval
The study protocol was approved by the Hospital Authority of Hong Kong ethics committee KC/KE-19-0001/ER-2.

Aim
This study aimed to characterize the association of taking antidepressants with the risk of being diagnosed with dementia, in people aged 65 years and over.

Data source
Access to comprehensive territory-wide deidenti ed electronic inpatient health records was provided by the Hospital Authority of Hong Kong (HKHA), a statutory body that manages all Hong Kong public hospitals and their ambulatory clinics. The HKHA services are available to all Hong Kong residents (>7.3 million), re ecting approximately 80% of all hospital admissions. 15 The HKHA collects and collates public-sector inpatient, outpatient, and emergency department admissions records. The HKHA electronic health records have been validated, and previously used for investigations of medication safety. [16][17][18][19][20][21] The patient-level data includes diagnosis, prescription, information on hospital admissions and discharges, payment method, and prescription and dispensing information.

Observation periods
For this study, the observation periods began on January 1, 2008, or the 65th birthday of the patient (whichever was later) and ended on December 31, 2010, or date of registered death (whichever was earlier) (a three year study period). We also imposed a ve-year pre-study window as a historical period for index dementia diagnosis.  Table S1 for detailed names).
Because of the concerns that repeat coding of the index dementia diagnosis might be found within the database, we considered only those potentially-eligible people with no dementia diagnosis in their historical electronic medical records ( ve years before study entry). This assumed that the index dementia event was correctly identi ed as occurring during the study period. Subsequent dementia diagnoses (if any) were excluded from the analysis.
Potentially eligible participants also required at least one antidepressant medicine to be prescribed during the study period. Antidepressants were identi ed using British National Formulary (BNF) codes 4.3.x, speci cally, tricyclic and related antidepressant drugs (4.3.1), monoamine-oxidase inhibitors (4.3.2), selective serotonin re-uptake inhibitors (4.3.3), and other antidepressant drugs (4.3.4). Refer to supplementary Table S2 for detailed drug names and their distributions within the database.

Exclusion criteria
Potential subjects were excluded if (a) their antidepressant drug-exposure was shorter than seven days (to eliminate the impact of potentially-ineffective exposures) and/or (b) the washout period and subsequent exposure to antidepressants overlapped (thus violating model assumptions). 25 2.6 Study design A Self-Controlled Case Series (SCCS) design was used. 25 This design relies on within-person comparisons in a population of individuals who have experienced both the outcome and exposure of interest. A major advantage of this design over classic case-control and cohort study designs is that it allows for controlling for potential effects of measured and unmeasured time-invariant confounders that vary between individuals, such as underlying health status, genetic factors, hospital location, underlying frailty, socio-economic status, etc. Using this design also enabled us to adjust for time-varying factors, such as age, which is a known associate of the onset of dementia. 2 2.6.1 Measurement periods and event date The default baseline period, and four measurement periods were de ned. Firstly, the drug-exposed period was de ned as time receiving antidepressants, with the duration between prescription start and end dates recorded within the database for each prescription episode. Drug-exposed period was further classi ed as index drug-exposed period and subsequent drug-exposed period. The index drug-exposed period was de ned as the rst time subjects received antidepressant medication. There were two other measurement periods: pre-exposure period (50 days prior to index antidepressant prescription) and post-exposure period (washout period). The pre-exposure and post-exposure periods enabled comparison of dementia rates prior to, and after, index antidepressant episode. The washout period was applied after index drugexposure ceased, as drugs generally require time to be excreted from body systems. We applied a washout period only after the index antidepressant administration, because subsequent washout periods (if any) might be in uenced by previous drug administrations. All remaining time within the observation period constituted the baseline period, to which the measurement periods were compared. Each participant at baseline was allocated a (default) risk of dementia (risk=1).
It was possible that (presumed) continuous drug-exposed periods might have been interrupted for a range of reasons. 17,26 We designed an algorithm (see pseudo-codes in the supplementary information) to obtain continuous treatment periods. We set the lengths of pre-exposure period and washout period to an arbitrary 50 days, and relevant sensitivity analyses were performed to examine the effects of the settings of the lengths of pre-exposure period and washout period. The study design and data capture periods are outlined in Figure 1.
The date of rst lifetime dementia diagnosis in the HKHA Electronic Health Record (EHR) was de ned as the event date. Only the rst recorded dementia event for each patient during the study period was included in the analysis.

Main analysis
The association between antidepressant treatment and dementia diagnosis was calculated by comparing the rate of dementia diagnosed during the four measurement periods, with that during the baseline period. Speci cally, the crude incidence rate per 1000 patient-days (CIR) (unadjusted by age) was calculated by dividing the number of events by the patient-days for each period. 16 The age-adjusted Incident Rate Ratio (IRR) was calculated in the standard SCCS analysis by modeling dementia diagnoses within individuals as a non-homogeneous, age-dependent Poisson process and contrasting incident rates within the same individual's person-time, 25 using age in 365-day bands.
In addition to standard SCCS analysis, the nonparametric spline-based SCCS approach was applied to investigate risk changes during the observation period. 27 A signi cance level of 5% was applied in all statistical analyses. Python (version 3.6) and R (version 3.3.2) were used for data processing and analysis. Relevant codes are publically available (https://github.com/zhongzhixu/SCCS).

Sensitivity analyses
Sensitivity analyses were conducted for: (a) the length of the washout period; (b) the length of the preexposure period; (c) the exclusion of individuals who died during the study period; (d) the exclusion of individuals whose drug duration was less than 30 days; (e) age-adjustment using 200-day age bands and 500-day age bands; and (f) the association of dementia and sub-categories of antidepressant drugs (i.e. BNF 4.3.1-4.3.4).

Post-hoc analysis
Post-hoc analysis was conducted to investigate whether depression-related disorders may confound the case-control study designs and cohort study designs, using the whole population within the database, to test prevalence among different subsets.

Results
Of the 194,507 people aged 65+ years in the database, 24,646 (12.6%) received a diagnosis of dementia during the period under study, and of these, 18,825 received their rst lifetime diagnoses of dementia.
Moreover, 3,757 (20.0%) had been prescribed antidepressants during the study period. We further excluded 1,300 people whose drug exposure periods were less than seven days or not available, and 71 patients whose washout period and subsequent drug-exposed period overlapped. As a result, 2,386 eligible participants were included for SCCS modeling (837 (35.1%) men; mean sample (SD) age at the point of enrolment was 81.7 (7.26) years). The characteristics of the included participants are described in Table 1.   Table 2.
The standard SCCS analysis indicated some association between the decision to start antidepressant treatment and dementia (Table 3). More speci cally, a higher risk of receiving a diagnosis of dementia was observed during the index drug-exposed period (IRR, 8.86, 95% con dence interval, 7.80 to 10.06) and subsequent drug-exposed period (IRR, 4.83, 95% CI, 3.28-7.13), relative to the baseline period. Tables 4-6 report the CIRs and SCCS analyses for the three major drug subcategories respectively, using the same population as the main analysis. Analysis was not undertaken on subjects who took MAOIs because of the small numbers. Figure 4 and Supplementary Table S2 reports a detailed distribution of the number of patients using different drugs. The analyses for each subcategory (Table 4-6) and the additional sensitivity analyses (supplementary Table S3-S10), did not change the overall picture of results. Post-hoc analysis indicated that during the study period, the prevalence of receiving a rst-ever dementia diagnosis among people experiencing depression-related disorders was 26.6% (648/2432) which was 2.7 times higher than the prevalence of a rst-ever dementia diagnosis among the whole population (9.7%, 18825/194507).
We examined the time intervals between two dementia diagnoses (which could have been the result of repeated coding). The mean (SD) of the time interval between two dementia diagnoses in the whole population was 114.14 (157.27) days. This indicated that our admission criterion of ve years' retrospective was appropriately strict to ensure that we captured people who truly had a rst lifetime development of dementia within the study timeframe.

Discussion
This pharmacoepidemiologic SCCS analysis, conducted in a large-sample, comprehensive, territory-wide electronic health record database, found an elevated risk of being diagnosed with dementia during the index exposure to antidepressants, compared to baseline. This is consistent with ndings from other case-control and cohort studies. 2, [5][6][7][8]11 However, our results do not support earlier conclusions 5-8 that a causal association between antidepressant treatment and the development of dementia may exist (i.e. antidepressant drugs may increase the risk of dementia). We present the following evidence to support our claims: (a) both the main analysis and the sensitivity analyses supported a higher risk of being diagnosed with dementia during the 50-day pre-exposure period, higher even than the index drug-exposed period. This nding was also supported by the nonparametric spline-based SCCS analysis, where the IRR peaked before the index drug initiation date, then decreased. A possible explanation for such a risk pattern is that the observed increased risk of developing dementia is not due to antidepressant treatment but in fact precedes it, perhaps re ecting changes in comorbid depressive complications and/or associated impairment that lead to a medical consultation, which in turn may contribute to the decision to prescribe antidepressant drugs; (b) both the main analysis and the sensitivity analyses indicated that the IRR of developing dementia during the washout period was similar to the default risk allocated to the baseline period. Since there was no obvious washout effect, it was reasonable to deduce that there is no effect of antidepressant drugs on dementia; and (c) the post-hoc analysis indicated that during the study period, the prevalence of a rst-ever diagnosis of dementia among people diagnosed with depression was approximately three times as high as the dementia prevalence among the general population. This indicates that depression-related disorders may be closely associated with dementia and this should be strictly controlled in correlation models for prescription of antidepressant drugs and the development of dementia. Earlier studies 5-8 reporting a potential causal-relationship between antidepressant drugs and the risk of being diagnosed with dementia controlled for none or only part of depression-related disorders.
Furthermore, diagnosed depression is not the only factor that should be controlled for in a case-control study design or a cohort study design that examines an association between antidepressant drugs and dementia. This is because almost half of the antidepressant prescriptions are for other (on or off-label) indications of other mental health-related disorders such as anxiety disorders, schizophrenia and insomnia 11 , with which dementia usually co-occurs 28 . Indeed, mental health-related disorders of antidepressant drug users might indicate cognitive deterioration (co-occurring with diseases of the nervous system and cerebral cortex (including dementia)). Thus, a strong association between dementia and antidepressant drugs under these circumstances again may not indicate a true causal relationship. 9 However, even if all of these factors are considered, misdiagnosis (both over-and under-diagnosis) for mental disorders is not uncommon. 29,30 As a result, information on relevant diagnoses may not be available for patients who are experiencing mental disorders. 29 This may well bias ndings, and lead to concerns about unmeasured confounders.
Consequently, a case-control or a cohort study design may not be the most appropriate for studies aiming to conclude causal associations, because those designs may not have the capacity to comprehensively control for important confounding variables. We propose that an SCCS design may be a better way of examining whether an exposure truly represents a risk factor for an event because the design eliminates all time-invariant confounders among individuals (including mental disorders) and introduces temporal awareness around the event of interest. This approach has been successfully applied in other epidemiological studies which, for example, explore the association between herpes zoster and stroke, 31 or the association between methylphenidate treatment with suicide attempts, 16 psychosis 17 and seizures. 18 Given that this study's results do not support the argument that antidepressant treatment is causally associated with dementia, suggestions to reduce exposure to antidepressant drugs in older people 5 must be prudently considered, because of the potential harm of stopping antidepressant drugs that could worsen symptoms of depression, incontinence, or pain, for which antidepressants have been prescribed.
There were only three patients who were prescribed MAOIs (BNF 4.3.3) during the study period. This is probably because MAOIs are older types of antidepressants with a wide range of side effects. According to the Hong Kong drug o ce, 32 this drug is only suggested as a second-line intervention if other antidepressants are not effective.
The SCCS method has been adopted in many publications previously. However, we have observed several drawbacks: (a) Standard SCCS only accepts non-overlapping segments, whereas some studies may not ensure this. 27 Our study potentially faced the same pitfalls, where there was a possibility that the washout period and period of subsequent antidepressant use may overlap. We attempted to exclude these individuals (71) to ful ll the SCCS model assumption; (b) The spline-based SCCS 27,34 use splines to smooth the exposure effect. As a result, the potential association trend between the timing of drug exposures or other risks, and adverse events, can be continuously depicted. However, we found that the spline-based SCCS method is underutilized by researchers to date.
Study limitations. Our study design had potential limitations. For example, a possible time lag between the time of developing dementia and the time of a formal clinical diagnosis of dementia may exist.
However, to the best of our knowledge, extracting diagnoses from electronic medical records is the most e cient (if not the only feasible) way to access comprehensive and large-scale database information (including prescriptions) among a population with dementia. More importantly, in this study, if there was a time lag between developing dementia and its diagnosis, this will not contradict our conclusions.

Conclusions
The observed higher risk of dementia before the commencement of antidepressant treatment may re ect emerging psychiatric symptoms, which can co-occur with dementia. These symptoms could trigger medical consultations where a decision is made to begin antidepressant treatment. Therefore, the ndings of this pharmaco-epidemiologic study do not support a causal association between antidepressant treatment and the risk of being diagnosed with dementia. All data generated during this study are included in this published article and its supplementary information le. Figure 1 Illustration of different periods and study design for a hypothetical patient. The symbol "]" indicated that the corresponding day was included in the interval to the left.

Figure 2
Distribution of the number of patients with different time intervals. The time interval of a patient is de ned as the gap between the date receiving a diagnosis of dementia and the date initiating drugs.

Figure 3
Association between the timing of drug exposure and the risk of dementia using spline-based Self-  Distribution of the number of patients using different drugs.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.