This active comparator, new-user cohort study identified patients initiating pimavanserin or a comparator atypical antipsychotic between April 1, 2016, and December 31, 2019, as pimavanserin was launched in the USA in April 2016. The study used Medicare claims and assessment data, including Medicare enrollment information, Part A (hospital insurance), Part B (physician and professional services), Part D (outpatient prescription drug coverage), and MDS (Minimum Data Set 3.0 and Minimum Data Set – Swing Bed 3.0) assessment information about beneficiaries in LTC settings. Medicare data from before the study period (2010–2016) were used to define patient characteristics before treatment initiation.

The study population included adults (aged ≥ 65 years at the time of treatment initiation) with PDP who initiated treatment with an atypical antipsychotic. Patients were identified at the first pharmacy dispensing record for pimavanserin or one of the comparator atypical antipsychotics during the study period (Fig. 1). The date of the antipsychotic initiation was considered the index date for cohort entry. Patients were required to have at least 365 days of continuous, fee-for-service enrollment in Medicare Parts A, B, and D before the index date; have enrolled in Medicare due to age or have reached their 65th birthday before 2010 (the first year of available data); have no use of any antipsychotic (atypical or conventional) during the year before the index date; and have diagnoses of both PD and psychosis at any point on or before the date of atypical antipsychotic treatment initiation. Patients were excluded if they had use of multiple antipsychotics on the index date or another psychiatric condition typically treated with antipsychotics, including bipolar disorder, schizophrenia or schizoaffective disorder, and major depressive disorder with psychotic symptoms. To ensure medical stability at the time of treatment initiation, patients were also excluded if they had claims for hospice care or palliative care in the 365 days before the index date.

Study design schematic illustrating cohort eligibility and inclusion criteria. DC District of Columbia. See Online Resource eTable 2 for covariate details. aInitial Medicare enrollment other than entitlement due to age (65 years) or aged 65 years before 2010; enrollment in a managed care plan; incomplete or intermittent enrollment in Medicare Parts A, B, and D. bBipolar disorder, schizophrenia, schizoaffective disorder, or major depressive disorder with psychotic symptoms. cDeath, end of available data, disenrollment from eligible Medicare plan, end of continuous use period of index medication, switching to or adding a different study medication

To qualify as being diagnosed with PD, patients must have met one of the following criteria: 1) one inpatient or SNF claim with a recorded diagnosis of PD in any coding position; 2) two outpatient claims with recorded diagnoses of PD separated by at least 30 days but within 365 days; or 3) one outpatient claim with a recorded diagnosis of PD with at least two pharmacy dispensing claims on different days for a PD treatment [19]. Diagnosis codes for PD included 332.0 in International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) or G20 in International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM). Psychosis was identified with diagnoses in Medicare claims or from recorded MDS 3.0 assessment information for hallucinations or delusions. The ICD-9-CM and ICD-10-CM codes for identifying a psychosis diagnosis are displayed in Online Resource eTable 1.

Follow-up for each patient started at the index date (date of atypical antipsychotic treatment initiation) and ended on the date of death or censoring at the earliest occurrence of one of the following events: end of available data (December 31, 2019); disenrollment from fee-for-service Medicare Parts A, B, or D; discontinuation of the index study drug; or use of a different study medication (patients in the pimavanserin group receiving a different atypical antipsychotic, or patients in the comparator group receiving a prescription for pimavanserin or a different comparator antipsychotic).

The primary exposure was initiation of antipsychotic treatment with pimavanserin, and the comparator was initiation of oral formulations of clozapine, quetiapine, risperidone, olanzapine, aripiprazole, or brexpiprazole (“comparator atypical antipsychotic”). Treatment initiation and use were assessed by identifying study medications from pharmacy dispensing information.

In the primary analysis, patients were considered exposed only during time on medication. Current use of pimavanserin and comparator atypical antipsychotics was defined as the duration of the days’ supply of medication prescriptions plus a grace period of 30 days, and patients were considered to have discontinued treatment 30 days after the end of the days’ supply of the previous prescription without receipt of an additional prescription.

The primary outcome of all-cause mortality (including the date of death) was evaluated from Medicare enrollment and beneficiary summary data for both treatment groups.

Patient characteristics for both treatment groups were identified from Medicare enrollment, claims, and MDS assessment data for descriptive analyses and as covariates in comparative analyses. Considered covariates included demographic characteristics, psychiatric diagnoses, comorbidities (including components of the Charlson Comorbidity Index, frailty indicators [20], substance use, infections, and other chronic and acute conditions), comedication use, and healthcare utilization (complete list of characteristics shown in Online Resource eTable 2).

Demographic characteristics and prescriber specialty were measured on the index date. Healthcare utilization and most comorbidities and comedications were evaluated in the 365 days before and not including the index date, although some chronic conditions were defined using all available baseline data before the index date. The lack of claims containing evidence of the condition was interpreted as a lack of the condition; no covariate information was explicitly missing (race/ethnicity was occasionally recorded as unknown).

The characteristics of all patients meeting the eligibility criteria were reported by treatment group; the extent of imbalances in the relative distribution of characteristics across treatment groups was estimated with absolute standardized differences (ASDs) [21]. To account for confounding arising from differences in characteristics between treatment groups, we implemented propensity score (PS) matching. Propensity scores, or the predicted probability of initiating pimavanserin treatment based on the observed covariates, were estimated for each patient using multivariable logistic regression, with all covariates included as independent variables. The distributions of the estimated PSs were plotted by treatment group to evaluate the extent of overlap, with greater overlap indicating better comparability between treatment groups. A PS-matched cohort was created in which one comparator patient was matched to each pimavanserin patient using a greedy-matching 5- to 1-digit algorithm with a maximum caliper of 0.2 times the standard deviation (SD) of the PS [22,23,24] (initially, a 2:1 comparator-to-pimavanserin matching ratio was attempted, but less than 90% of pimavanserin patients were retained because of failure to find two comparator matches; thus, a 1:1 match was implemented). The covariate distributions and ASDs were reevaluated in the matched cohort to ensure balance of confounders.

The incidence rates (IRs), 95% CIs, and cumulative incidence of mortality during follow-up were estimated, by treatment group, in the cohorts before and after matching. Time to mortality was compared between treatment groups, with HRs and 95% CIs estimated with Cox-proportional hazards models. The proportionality assumption was tested through visual inspection of the smoothed hazard curves [25, 26]. As a measure of absolute difference between the matched treatment groups, the incidence rate difference (IRD) was estimated by subtracting the treatment group–specific IRs, and 95% CIs were estimated by using the Wald method [27].

Previous studies of pimavanserin use have suggested different risk profiles between nursing home residents and community dwelling populations or increased mortality risk among patients in nursing homes [17, 18]. Thus, eligible patients in the PDP cohort who were residing in LTC or SNFs on the index date were identified and evaluated separately as an LTC/SNF sub-cohort.

Previous studies have suggested an increased risk of mortality among users of antipsychotics during the initial 180 days of use [28] and potential differences in the risk between users of pimavanserin and users of other antipsychotics or untreated patients in the first 30, 90, or 180 days of treatment [17, 18]. Therefore, time period–specific HRs were estimated in the matched cohort for the first 30, 90, 180, and 365 days of follow-up to evaluate potential changes in mortality risk over time.

Additionally, subgroup analyses were performed to evaluate the potential effect measure modification in clinically relevant subgroups, including sex, age groups, and dementia diagnosis. The overall cohort before matching was restricted to each subgroup, and new PSs were estimated and used to match 1:1 within each subgroup; HRs were estimated within each subgroup.

Two separate sensitivity analyses were performed. First, we implemented an “intent-to-treat” analysis, in which treatment discontinuation was removed from the list of censoring criteria to evaluate potentially lagged events after discontinuation and to address potential informative censoring if treatment was discontinued in anticipation of a patients’ imminent death. Second, we removed the requirement for patients to have a recorded psychosis diagnosis, as many psychosis symptoms may be under-recorded in claims data [29,30,31], and other recent studies of pimavanserin have not required psychosis diagnoses [17, 18]. As a post hoc analysis, the E-value was estimated as a measure of the potential impact of unmeasured confounding (Online Resource Supplementary Methods) [32].

All analyses were performed using SAS Version 9.4 (SAS Institute Inc.; Cary, NC). This study was performed using US Medicare data, which are available for use upon approval and licensing under a data use agreement with the US Centers for Medicare and Medicaid Services (CMS). The study protocol was registered in the European Union Post-authorisation Study (PAS) Register (EU PAS register number 46331) [33].