Hallucinations, antipsychotic use, and mortality in older adults with dementia: retrospective cohort study of two Medicare-linked national health surveys

Abstract

Background:

Hallucinations are associated with earlier death in older adults with dementia, but antipsychotic medications are also associated with mortality, and comparisons of their relative harms are lacking.

Objective:

To determine the individual and combined association between hallucinations, antipsychotic use, and mortality.

Methods:

We performed a retrospective cohort study using Medicare-linked survey data from two nationally representative studies (the National Health and Aging Trends Study and the Health and Retirement Study) containing validated dementia identification algorithms and a screening question for hallucinations. Using Medicare claims, we identified participants with dementia who had no history of antipsychotic use during the year of or prior to entry. We used extended Cox regression with time-varying covariates to analyze the association between hallucinations, antipsychotic use, and mortality adjusting for confounders.

Results:

We identified 1,703 eligible subjects who contributed 4,819 person-years of follow-up. 555 (32.6%) had hallucinations at baseline, 705 (41.4%) reported hallucinations at least once during follow-up, and 284 (16.7%) received antipsychotics. Hallucinations were associated with an increased risk of death in unadjusted models (HR 1.36, 95% CI: 1.18-1.5), but antipsychotic use was not (HR 1.03, 95% CI: 0.85-1.2). After adjusting for age, race, gender, dementia severity and comorbidities, the hazard ratio for hallucinations attenuated and was no longer statistically significant (1.15, 95% CI: 0.98-1.34). There was no significant interaction between hallucinations and antipsychotic use.

Conclusion:

Hallucinations are associated with an increased risk of death that is greater than the risk associated with antipsychotic use, though this is partially confounded by dementia severity comorbidities.

1. INTRODUCTION

Hallucinations affect one in five patients with Alzheimer disease (AD) and over half of patients with Parkinson disease dementia (PDD) and dementia with Lewy bodies (DLB) and have been identified as an independent risk factor for morbidity and mortality in these patients[1-3]. However, epidemiologic studies of dementia-related hallucinations and mortality have several fundamental limitations. Foremost among these is the fact that antipsychotic medications, which are used to treat hallucinations, are themselves associated with stroke and death and carry an FDA Black Box warning for these outcomes[4,5]. It is therefore unclear to what extent adverse outcomes are attributable directly to hallucinations versus medications in these patients. Hallucinations are also frequently under-reported in clinical practice due to perceived stigma and embarrassment[6,7].

Administrative claims data (e.g., Medicare) have been used to study the risk of mortality associated with antipsychotic use[8-12]. Although their large sample size and use of real-world prescription and outcomes data improve generalizability, claims based studies are also limited by the underascertainment of hallucinations and psychosis in claims data, and we have previously shown that Medicare claims data fail to capture 75% of patients with dementia-related hallucinations[13]. Dementia research cohorts, which are typically recruited from within tertiary memory centers, are able to improve the detection of hallucinations through systematic screening[14,15] while simultaneously adjusting for antipsychotic use[16-21], but at the expense of sample size and generalizability. Specifically, racial and ethnic minorities are under-represented in dementia cohorts that are recruited from memory centers compared with population-based studies, and memory centers are also biased towards earlier age of onset and death[22]. In this study, we combine the depth of cognitive and neuropsychiatric assessments usually found in a clinical cohort with the diversity and generalizability of administrative claims to determine the relationship between hallucinations, antipsychotic use, and mortality using Medicare-linked data from two longitudinal, nationally-representative health surveys.

2. METHODS:

2.1. Standard Protocol Approvals, Registrations, and Patient Consents

This study was approved by the University of Pennsylvania Institutional Review Board. Informed consent was previously obtained at the time of study enrollment and was not separately required for this secondary analysis.

2.2. Study Overview and Data

We analyzed patient-reported outcomes and Medicare claims from two longitudinal surveys of the older U.S. adult population: the National Health and Aging Trends Study (NHATS) and the Health and Retirement Study (HRS). We used validated assessments of cognitive status and proxy-informed hallucinations to ascertain dementia and hallucinations, and linked Medicare claims data to identify antipsychotic prescriptions and death dates.

NHATS is a nationally representative sample of Medicare beneficiaries aged 65 or older who have been surveyed annually since 2011, with replenishment of the sample in 2015 [23]. HRS is a large, nationally representative longitudinal survey of U.S. adults over the age of 50 [24]. It is the result of two surveys from 1992-1996, HRS and Asset and Health Dynamics among the Oldest Old (AHEAD), which were merged in 1998. Participants have been re-surveyed biennially, with replenishment in 1998, 2004, 2010, and 2016. For this study, we used data from NHATS 2011-2017 and HRS 2006-2016 (since linked Medicare Part D data became available for HRS beginning in 2006).

The Medicare program is the primary insurer for 97% of the U.S. population ages 65 and above. Medicare linkage is available for all NHATS participants and for the subset of HRS participants who are Medicare-eligible and consent to data linkage (>80%). We used the Master Beneficiary Summary File (MBSF) to obtain Medicare enrollment and eligibility information and death dates, and the Part D Drug Event file to obtain prescription drug coverage and utilization information.

2.3. Inclusion and exclusion criteria

Study inclusion and exclusion criteria are summarized in Figure 1. NHATS and HRS respondents were eligible for cohort entry if they had dementia and data on the presence or absence of hallucinations (as described below). To exclude prevalent antipsychotic users, we required subjects to have complete Part D coverage with no antipsychotic prescriptions or alternative prescription insurance (e.g. retiree drug subsidy) during the year of cohort entry (unless death was the reason for incomplete coverage) and year prior to entry.

Figure 1: Flowchart of inclusion and exclusion criteria for Medicare-linked NHATS and HRS study participants.

CMS, Center for Medicare & Medicaid Services; HRS, Health and Retirement Study; NHATS, National Health and Aging Trends Study

2.4. Identification of dementia and hallucinations

The NHATS and HRS survey questionnaires include a number of screening items that assess cognitive function. These include questions about self-reported cognitive difficulties (e.g., trouble remembering appointments or handling finances), physician diagnoses of AD or related dementia, and in-person cognitive testing (e.g., orientation, delayed recall, clock drawing). NHATS and HRS have developed algorithms that combine individual survey items to predict the presence or absence of dementia, which have been validated against clinically diagnosed dementia in the Aging, Demographics, and Memory Study (positive predictive value = 71%) [25,26]. NHATS classifies dementia status as “probable dementia”, “possible dementia”, or “no dementia”, and the HRS Langa-Weir classification system categorizes cognitive function as “dementia”, “cognitively impaired but not demented”, and “normal”. For this study, we defined dementia as a rating of “probable dementia” in NHATS or “dementia” in HRS, hereafter referred to collectively as dementia.

In both NHATS and HRS, if a study participant is unable to complete the cognitive assessment independently, a proxy (i.e., spouse, child, or caregiver) is allowed to assist, and several additional questions about cognitive function are asked (proxy cognitive questionnaire). One of these questions assesses for the presence of visual or auditory hallucinations: “In the last year, do you/does he or she ever see or hear things that are not there?” A trained rater incorporates both patient and proxy observations in recording a response to this question.

Because NHATS and HRS are both longitudinal studies, subjects with dementia could complete multiple proxy cognitive questionnaires, and hallucinations were treated as a time-varying exposure. NHATS and HRS provide survey year but not the specific date on which a questionnaire was completed, so all NHATS and HRS-derived variables were discretized by year and analyzed as time-varying exposures.

2.5. Identification of antipsychotic prescriptions

Antipsychotic use was also summarized by year for each subject during and after cohort entry. We identified prescriptions for typical and atypical antipsychotics (Online Resource 1) using National Drug Codes from Multum Medisource Lexicon (Denver, Colorado). We did not include prescriptions for pimavanserin since it was not approved by the FDA until 2016. To avoid capturing prescriptions that were not actually filled or taken, we considered a subject to be “exposed” to antipsychotics in a given year if they had at least two prescriptions with a supply of seven days or longer. In addition to Part D eligibility requirements at the time of cohort entry, we required complete Part D coverage during the year of and year prior to a subject’s first antipsychotic prescription. To minimize the potential for gaps in prescription data and missed antipsychotic prescriptions, we censored subjects if they had no antipsychotic prescriptions during a given year and lacked complete Part D coverage during that year.

2.6. Covariates

With the exception of antipsychotic prescriptions and death dates, all other variables (including gender and race/ethnicity) were directly patient- and proxy-reported in NHATS and HRS. In addition to basic demographics, we calculated several composite measures of functional impairment and comorbid illness. Activity of daily living (ADL) limitations were summarized as the number of the following activities (ranging from 0-6) that respondents reported difficulty, inability, or requiring assistance to perform: dressing, eating, bathing, toileting, getting out of bed, and getting around inside. Instrumental activities of daily living (iADL) limitations were similarly quantified (range: 0-5) based on shopping, meal preparation, using a telephone, taking and managing medication, and managing finances. ADL and iADL limitations have previously been used as markers of dementia severity in NHATS[27]. We also calculated a multimorbidity index using a composite of demographics, comorbid conditions (self-reported diabetes, cancer, chronic lung disease, and heart failure), smoking, BMI, and functional disability that has previously been shown to predict 4-year mortality in HRS[28].

2.7. Statistical Methods

Baseline characteristics at the time of cohort entry were summarized using descriptive statistics. We used extended Cox regression to analyze survival as a function of hallucinations and antipsychotic use. Extended Cox regression is similar to a standard Cox regression analysis but allows for time-varying covariates (that is, for the values of variables such as hallucinations, antipsychotic use, ADL/iADL limitations, and multimorbidity to change from one year to the next during an individual’s time in the study). We performed both unadjusted and adjusted analyses and explored the use of a hallucinations x antipsychotic interaction term. Stratified Kaplan-Meier curves were also produced. Subjects were censored at loss of Part D eligibility or the end of the observation period (whichever came first).

In addition to our prespecified analyses, we performed a sensitivity analysis in which models were also adjusted for baseline history of Parkinson disease (PD) or dementia with Lewy bodies (DLB). For this analysis, we restricted the sample to NHATS and HRS participants who had Medicare Part A and B coverage in addition to Part D coverage. We defined a history of PD or DLB as the presence of a corresponding ICD-9 (332.0 or 331.82) or ICD-10 (G20.0 or G31.83) code in the carrier file during or prior to the year of cohort entry. Individuals without PD or DLB were required to have no previous record of diagnosis and complete (12 months) of Part A and B coverage during the year of and the year prior to cohort entry.

Because the sample was restricted to proxy cognitive questionnaire respondents, we did not apply NHATS and HRS sample weights to the analysis. Statistical analysis was performed using STATA/MP 17.0 (College Station, TX), and statistical significance was defined at the p<0.05 level.

3. RESULTS:

We identified 880 NHATS participants and 823 HRS participants who met eligibility criteria (Figure 1), resulting in a total sample size of 1,703 individuals and 4,819 person-years of survival time for analysis. All subjects had no history of antipsychotic use during the year of and year prior to cohort entry, but 284 individuals (16.7%) received an antipsychotic prescription during the study period. 555 (32.6%) had hallucinations at baseline, and 705 reported hallucinations at least once during follow-up (period prevalence 41.4%). Additional baseline characteristics are presented in Table 1. Because the proxy cognitive questionnaire (which is used to identify hallucinations) is completed only when NHATS and HRS respondents require assistance with cognitive assessment, multiple indicators of greater dementia severity are reflected at baseline: 1,259 (73.9%) were 80 years of age or older, and subjects had difficulty or required assistance with a median of 4 out of 6 ADLs and 4 out of 5 iADLs.

Table 1:

Baseline characteristics of the Medicare-linked NHATS and HRS cohorts

Age
≤69 years	49 (2.9%)
60-69 years	395 (23.2%)
≥80 years	1,259 (73.9%)
Race (non-White)	607 (35.6%)
Gender (female)	1,126 (66.1%)
Smoking (current/former)	748 (44.3%)
Multimorbidity index (median, range)	14 (2-23)
ADL difficulties (median, range)	4 (0-6)
iADL difficulties (median, range)	4 (0-5)
Visual impairment (n, %)	723 (43.2%)
Metropolitan (n, %)	1,305 (77.2%)
Married (n, %)	539 (31.7%)
Nursing home (n, %)	405 (23.8%)
Hallucinations (n, %)	555 (32.6%)

Categorical variables are presented as counts and frequencies, and continuous variables are presented as median and range. ADL, activities of daily living; iADL, instrumental activities of daily living

In unadjusted Cox regression models, the presence of hallucinations was associated with an increased risk of death (HR 1.36, 95% CI: 1.18-1.5), but antipsychotic use was not (HR 1.03, 95% CI: 0.85-1.2). This pattern persisted when adjusting for both hallucinations (HR 1.38, 95% CI: 1.20-1.60) and antipsychotic use (HR 0.87 95% CI: 0.70-1.10). After additionally adjusting for baseline age, race, gender, education, smoking, multimorbidity index, ADL and iADL impairment, visual impairment, metropolitan residence, marital status, and nursing home residence, the hazard ratio for hallucinations attenuated and was no longer statistically significant (HR 1.15, 95% CI: 0.98-1.35), and the point estimate for antipsychotics was similar (Table 2). The addition of a hallucinations-by-antipsychotic interaction term was not statistically significant in either unadjusted (p=0.12) or adjusted (p=0.29) models, and additional adjustment for baseline PD or DLB also did not significantly affect our results (Table 2). Kaplan-Meier curves stratified by hallucinations and antipsychotic use are shown in Figure 2.

Table 2:

Extended Cox proportional hazards model of mortality in NHATS and HRS

	HR (95% CI)1	HR (95% CI)2	HR (95% CI)3	HR (95% CI)4	HR (95% CI)5	HR (95% CI)6
Antipsychotic	1.03 (0.85-1.24)	0.87 (0.70-1.10)	0.86 (0.69-1.08)	0.84 (0.67-1.06)	0.81 (0.64-1.02)	0.94 (0.68-1.28)
Hallucination	1.36 (1.18-1.57)	1.38 (1.20-1.60)	1.30 (1.12-1.51)	1.22 (1.05-1.43)	1.15 (0.98-1.35)	1.13 (0.90-1.42)

¹Unadjusted

²Adjusted for antipsychotic use and hallucinations

³Adjusted for antipsychotic use, hallucinations, age, race, gender, metropolitan residence, marital status, and nursing home residence

⁴Adjusted for antipsychotic use, hallucinations, age, race, gender, metropolitan residence, marital status, nursing home residence, smoking, multimorbidity index, and visual impairment

⁵Adjusted for antipsychotic use, hallucinations, age, race, gender, metropolitan residence, marital status, nursing home residence, smoking, multimorbidity index, visual impairment, ADL limitations, and iADL limitations

⁶Adjusted for antipsychotic use, hallucinations, age, race, gender, smoking, metropolitan residence, marital status, nursing home residence, multimorbidity index, visual impairment, ADL limitations, iADL limitations, and baseline PD/DLB

ADL, activities of daily living; DLB, dementia with Lewy bodies; HR, hazard ratio; HRS, Health and Retirement Study; iADL, instrumental activities of daily living; NHATS, National Health and Aging Trends Study; PD, Parkinson disease

Figure 2: Kaplan-Meier curves stratified by hallucinations and antipsychotic use.

AP, antipsychotic

4. DISCUSSION:

In this study, we examined risk factors for mortality in older adults with dementia using validated dementia algorithms, patient- and caregiver-reported hallucinations, and Medicare Part D prescription claims data in two nationally-representative health surveys. We found that hallucinations were associated with an increased risk of death and that this was greater than the risk of death associated with antipsychotic use and at least partially explained by dementia severity and other confounders. These findings highlight the importance of dementia-related hallucinations with respect to health outcomes in older adults.

Our association between hallucinations and mortality replicates findings from tertiary memory center studies[16-21]. However, these studies are limited by the under-representation of racial and ethnic minorities, and memory center cohorts are also characterized by an earlier age of onset and death than the general population[22], which may reflect systematic differences in disease pathology and severity (e.g. higher burden of Lewy body or mixed pathology) that have the potential to bias the association between hallucinations and mortality. With nearly 75% of our cohort aged 80 or older and more than one third from racial and ethnic minority backgrounds, our sample is more representative of the general population and features groups traditionally underrepresented in dementia research.

The association between hallucinations and death remained unchanged after adjusting for antipsychotic use, but after further adjustment for age, race, gender, dementia severity and other comorbidities, the hazard ratio attenuated and was no longer statistically significant. This indicates that at least some of the relationship between hallucinations and mortality is confounded by shared risk factors, though given the relatively modest decrease in hazard ratio and increase in confidence interval width, we suspect that statistical power in our multivariable model was limited and that a larger sample would have yielded a statistically significant result even after covariate adjustment. It remains unclear whether hallucinations and psychosis truly cause death in older adults with dementia – and thus whether the hypothetical elimination of hallucinations could reduce this risk – or if they are merely a marker of greater disease severity. Addressing causal inference in future studies of dementia-related psychosis would have important implications for treatment models, which require comparing the risks and benefits of medication use in this population.

In contrast to previous pharmacoepidemiologic studies[8-12], we did not find an increased risk of death associated with antipsychotic use in this cohort. This discrepancy may be due in part to the fact that previous administrative claims studies have been unable to fully adjust for confounding by indication in non-randomized treatment allocation. Specifically, the association between antipsychotics and mortality is confounded by the fact that the symptom for which antipsychotics are prescribed (psychosis) is itself associated with an increased risk of death. Dementia severity is not directly measured in administrative claims data; time from dementia diagnosis is typically used as a proxy for dementia severity in these studies, though the validity of this approach (especially in datasets where the coding of dementia itself may be variable) is unclear. Hallucinations are also not adequately captured in administrative claims data, as we have previously shown using linked Medicare claims data from NHATS and HRS[13]. In this study, we were able to take advantage of direct measurements of dementia-related functional status and hallucinations, including repeated measurements to account for their potentially waxing and waning nature over time, to better adjust for confounding by indication.

Antipsychotic use has also been associated with mortality in clinical trials of older adults with dementia[4], which are not subject to confounding by indication because of their randomized design, so in addition to confounding by indication, there are other methodologic explanations for the lack of association between antipsychotic use and mortality in our study. In NHATS and HRS, participants complete surveys every year or every other year, and because the exact date of questionnaire completion is unknown, their responses are applied across the entire year for analysis. We used the same approach to summarize antipsychotic use by year, but if antipsychotic use only occurred for part of the year, it could have biased the association between antipsychotic use and mortality towards the null. Therefore, our data should not be interpreted as evidence of antipsychotic safety in this population, and we agree with recommendations to minimize their use as much as possible in older adults with dementia[29].

In addition to the challenge of analyzing antipsychotic-exposed person-time in a binned manner, other limitations include a lack of data on underlying disease pathology (e.g. Alzheimer, Lewy body) and the potential for some hallucinations to have been missed despite population-based screening (for example, if caregivers were unaware of them or mistook them for other symptoms such as delusions, confusion, or agitation). Because hallucinations were classified as present or absent, we lacked information on the type (e.g. visual or auditory), frequency, or severity, which would be interesting to explore in the future in the context of antipsychotic utilization and outcomes. We also lacked data on other neuropsychiatric symptoms such as agitation and delusions, and while we used activity limitations as a marker of dementia severity, it is important to note that these limitations are not specific to dementia and may actually serve as a composite indicator of multiple confounding comorbidities. Because the proxy cognitive questionnaire was only asked of respondents who had difficulty completing cognitive assessments independently, our cohort reflects a target population of patients with moderate to severe dementia and may not be generalizable to those with mild disease. However, in AD (the most common cause of dementia in the general population), psychosis typically occurs as a late manifestation of disease[30], so we likely captured the majority of dementia-related hallucinations in NHATS and HRS. Finally, we were unable to include data on pimavanserin since it was approved by the FDA towards the end of our study period, though pimavanserin is currently only indicated in Parkinson disease-dementia, which comprises a small fraction of our cohort[13].

5. CONCLUSION

In summary, our results show that the mortality risk of hallucinations may be higher than that of antipsychotics when accounting for dementia severity and other confounders. Antipsychotics should still be used sparingly in older adults with dementia, and future research should assess the extent of confounding between hallucinations and mortality and compare outcomes associated with treated versus untreated hallucinations in order to inform treatment guidelines, especially as newer medication safety data becomes available.

KEY POINTS.

• In a population-based study of over 1,700 older adults with dementia, hallucinations were associated with a greater risk of death than antipsychotic medications, but both were confounded by dementia severity and other factors.

• Future research should assess the extent of confounding between hallucinations and mortality and compare outcomes associated with treated versus untreated hallucinations in order to inform treatment guidelines, especially as newer medication safety data becomes available.