Antipsychotic drugs in first-episode psychosis: a target trial emulation in the FEP-CAUSAL Collaboration

Abstract

Good adherence to antipsychotic therapy helps prevent relapses in first-episode psychosis (FEP). We used data from the FEP-CAUSAL Collaboration, an international consortium of observational cohorts, to emulate a target trial comparing antipsychotics, with treatment discontinuation as the primary outcome. Other outcomes included all-cause hospitalization. We benchmarked our results to estimates from the European First Episode Schizophrenia Trial, a randomized trial conducted in the 2000s. We included 1097 patients with a psychotic disorder and less than 2 years since psychosis onset. Inverse-probability weighting was used to control for confounding. The estimated 12-month risks of discontinuation for aripiprazole, first-generation agents, olanzapine, paliperidone, quetiapine, and risperidone were 61.5% (95% CI, 52.5-70.6), 73.5% (95% CI, 60.5-84.9), 76.8% (95% CI, 67.2-85.3), 58.4% (95% CI, 40.4-77.4), 76.5% (95% CI, 62.1-88.5), and 74.4% (95% CI, 67.0-81.2), respectively. Compared with aripiprazole, the 12-month risk differences were -15.3% (95% CI, -30.0 to 0.0) for olanzapine, -12.8% (95% CI, -25.7 to -1.0) for risperidone, and 3.0% (95% CI, -21.5 to 30.8) for paliperidone. The 12-month risks of hospitalization were similar between agents. Our estimates support use of aripiprazole and paliperidone as first-line therapies for FEP. Benchmarking yielded similar results for discontinuation and absolute risks of hospitalization as in the original trial, suggesting that data from the FEP-CAUSAL Collaboration sufficed to remove confounding for these clinical questions.

This article is part of a Special Collection on Mental Health.

Introduction

Pharmacological therapy with antipsychotic agents is recommended for individuals with a first episode of psychosis (FEP). Because in most cases good adherence to therapy may prevent relapses, the decision to choose a particular antipsychotic needs to be informed by the comparative effectiveness and tolerability of the medication. In the early 2000s, the European First Episode Schizophrenia Trial (EUFEST) compared the 12-month risk of treatment discontinuation in FEP patients randomly assigned to either haloperidol (a first-generation antipsychotic agent) or to one of 4 second-generation agents: quetiapine, olanzapine, ziprasidone, or amisulpride.¹ Second-generation antipsychotics were better tolerated than haloperidol, and no differences were found between second-generation drugs.

Although EUFEST was a milestone in FEP research, it is not fully applicable to current FEP care: Ziprasidone is not widely used, amisulpride is not approved in the United States, and drugs that are commonly used (eg, aripiprazole, risperidone, paliperidone) were not studied. EUFEST recruited nearly 500 FEP patients from more than 14 countries and 49 centers over a 6-year period. Understandably, such an effort cannot take place every few years when clinical practice changes. More recent randomized trials have found it challenging to recruit adequate sample sizes of FEP patients.^2,3 Conversely, observational studies can benefit from a larger sample size. However, the few observational studies of antipsychotic effectiveness in individuals with first-episode psychosis only studied the most widely prescribed medications (olanzapine, risperidone, aripiprazole)⁴ or were not able to adequately adjust for confounding due to disease severity.^4,5

An alternative approach to obtain this information is to use observational data to explicitly emulate a pragmatic (hypothetical) randomized trial that would answer the causal question of interest—the target trial.⁶ An explicit emulation of a target trial eliminates design-induced biases related to selection and immortal time,⁷ although the observational estimates may still be biased if important confounders remain unmeasured. To increase confidence in estimates from observational data, one can emulate an existing randomized trial and then compare the observational estimates with those from the trial, a process referred to as benchmarking.⁸

Here, we used data from the FEP-CAUSAL Collaboration, an international consortium of observational cohorts of FEP patients in North America and Europe, to emulate a target trial of first-line antipsychotic therapy. We then benchmarked our results to estimates from EUFEST.

Methods

Our approach has 2 steps: (1) specify the protocol of the target trial and (2) emulate the target trial using the available observational data.⁶

Specification of the target trial

We designed the target trial emulation to follow closely the protocol from the EUFEST trial. The key components of the protocol are summarized in Table 1.

Table 1.

Specification of the protocol of a target trial and its emulation using data from the multi-country FEP-CAUSAL Collaboration.

Protocol component	Index trial (EUFEST)	Target trial	Emulation using FEP-CAUSAL
Eligibility criteria	(1) Individuals aged 18-40 between 2002-2006; (2) diagnosis of schizophrenia, schizophreniform, or schizoaffective disorder validated using MINI questionnaire; (3) less than 2 years since psychosis onset; (4) antipsychotics used for less than 2 weeks in the previous year or 6 weeks at any time; (5) no known intolerance to study drugs and; (6) no contraindications to study drugs.	Same as EUFEST, except: (1) dates are 2012-2020 and (2) includes patients with psychosis NOS (clinical diagnoses were used)	Same as target trial. The use of a drug is assumed to imply that no contraindications existed for that drug.
Treatment strategies	Initiation and continuation of the following agents: (1) haloperidol 1-4 mg; (2) olanzapine 5-20 mg; (3) quetiapine 200-750 mg; (4) ziprasidone 40-160 mg; (5) amisulpride 200-800 mg. Individuals may initiate any other psychotropic agent (eg, mood stabilizers, antidepressants) but no other antipsychotic agents under either strategy. Patients are not allowed to transition to a depot formulation.	Initiation and continuation of the following agents: 1) First-generation agents: haloperidol 5-10 mg, fluphenazine 2.5-40 mg, perphenazine 4-64 mg; (2) olanzapine 5-30 mg; (3) quetiapine 200-750 mg; (4) risperidone 2-6 mg; (5) aripiprazole 5-30 mg; (6) paliperidone 3-12 mg. Individuals may initiate any other psychotropic agent (eg, mood stabilizers, antidepressants) but no other antipsychotic agents under either strategy. Patients are allowed to transition to a depot formulation of the same agent they initiated orally, but not to depot formulations for other antipsychotics.	Same as target trial
Treatment assignment	Patients were randomly assigned to 1 of the 5 treatment strategies and were aware of the drug they were randomized to.	Same as EUFEST	Randomization was assumed conditional on the following baseline covariates: age, race, sex, housing status, diagnosis, comorbid substance abuse, prior use of antipsychotics, employment and educational status, and baseline symptomatic status (CGI-S score) and GAF score.
Outcomes	Discontinuation of treatment strategy, defined as dose out of range, or prescription of another antipsychotic for 14 consecutive days (including depot formulation). Other outcomes: psychiatric hospitalizations, concomitant prescription of a mood stabilizer or antidepressant, mean CGI-S, GAF at 12 months.	Same as EUFEST, except that a transition to a depot formulation of the same antipsychotic agent initiated orally was not considered a treatment discontinuation.	Same as target trial
Follow-up	Starts at treatment assignment and ends at the first to occur of loss to follow-up, death, 12 months, or outcome occurrence.	Same as EUFEST	Same as target trial
Causal contrast	Intention-to-treat effect	Intention-to-treat effect, per-protocol effect	Observational analogue of the per-protocol effect
Statistical analysis	Intention-to-treat analysis: for failure time outcomes, comparison of 12-month risks (estimated via Kaplan–Meier or pooled logistic models) and hazard ratios (estimated through Cox proportional regression models); for continuous outcomes, comparison of means at 12 months. IP weighting to adjust for losses to follow-up. Per-protocol analysis: Same with standardization for baseline variables.	Same as EUFEST	Same as per-protocol analysis of target trial

Abbreviations: CGI-S, Clinical Global Impression–Severity; EUFEST European First Episode Schizophrenia Trial; FEP, first-episode psychosis; GAF, Global Assessment of Functioning; IP, inverse-probability; MINI, Mini International Neuropsychiatric Interview; NOS, not otherwise specified.

Eligibility criteria were ages of 18-40 years, less than 2 years since psychosis onset, less than 6 weeks of antipsychotic use, and a diagnosis of schizophrenia, schizoaffective or schizophreniform disorder, or psychosis not otherwise specified (NOS).

The treatment strategies were initiation and continuation of monotherapy with: (1) a first-generation antipsychotic agent (haloperidol 5-10 mg, fluphenazine 2.5-40 mg, or perphenazine 4-64 mg), (2) olanzapine 5-30 mg, (3) quetiapine 200-750 mg, (4) risperidone 2-6 mg, (5) aripiprazole 5-30 mg, and (6) paliperidone 3-12 mg. We grouped 3 high- or medium-potency first-generation antipsychotic agents together based on their similar mechanism of action and tolerability profile.⁹ Individuals may initiate any other psychotropic agent (eg, mood stabilizers, antidepressants) but no other antipsychotic agents under either strategy. The use of a long-acting injectable formulation of the initially assigned antipsychotic agent was allowed but not the use of a long-acting injectable formulation of a different antipsychotic.

The primary outcome of interest was discontinuation of the treatment strategy, defined as 14 consecutive days with: (1) a dose outside the specified range assigned antipsychotic drug, or (2) the prescription of another antipsychotic agent. Other outcomes were all-cause psychiatric hospitalization and use of a concomitant mood stabilizer or antidepressant, as well as Clinical Global Impression–Severity (CGI-S) score and Global Assessment of Functioning (GAF) score at 12 months (although changes in clinical scores are difficult to detect over 1 year^10,11).

For each eligible individual, follow-up started at treatment assignment (baseline) and ended upon the outcome of interest, death, loss to follow-up (defined as 12 consecutive weeks without a clinical appointment), 1 year after baseline, or administrative end of follow-up (September 31, 2022), whichever occurred first.

The causal estimands of interest were the intention-to-treat effect and the per-protocol effect. For the intention-to-treat analysis of failure time outcomes (ie, discontinuation, hospitalization, coprescription), we compared the 1-year risk under each strategy via risk differences. The risks can be estimated nonparametrically by using a Kaplan–Meier estimator or parametrically by fitting a pooled logistic regression model with an indicator for assigned strategy, time of follow-up, and a product term between the strategy indicator and time of follow-up. For continuous outcomes, we compared the mean scores between groups at 12 months.

For the per-protocol analysis, the analysis was restricted to those who initiated the assigned treatment, and prognostic predictors of adherence were added as covariates to the same outcome model. Nonparametric bootstrapping with 500 samples can be used to calculate percentile-based 95% CIs for risk estimates.

Emulation of the target trial

Data source

The FEP-CAUSAL Collaboration is an international consortium of observational cohorts of individuals with FEP. Appendix S1 provides details on the information collected by the collaboration from FEP clinics in North America and Europe, and Table S1 provides information on study variables. This report includes data from OnTrack at McLean Hospital (Massachusetts), STEP at Connecticut Mental Health Center/Yale University-Yale (Connecticut), RAISE-ETP (United States), NYU Langone Health (New York), Pittsburgh-STEP at UPMC (Pennsylvania), University of British Columbia (British Columbia, Canada) and 7 hospitals participating in AGES-CM (Madrid, Spain). Baseline data on sociodemographic characteristics (eg, age, sex) and clinical factors (eg, diagnoses, comorbidities) were extracted from the intake interview of the FEP clinics or the first interview with a mental health clinician. Longitudinal data were extracted from electronic health records. For each recorded clinical encounter, trained research assistants extracted data on medications used, substance use disorder, employment and educational statuses, hospitalizations, emergency room visits, and concomitant services received; psychiatrists and psychologists extracted information on symptomatic status and side effects from the clinical notes. Appendix S2 provides intraclass correlation coefficients to evaluate the reliability of the different symptom scales assessed by psychiatrists and psychologists from reading clinical notes.

Data linkages were used to identify antipsychotic therapy from outpatient prescription data and hospitalizations. Most data extraction used clinical databases; one of the cohorts (RAISE-ETP) was a randomized trial launched by the National Institute of Mental Health, and all data were collected from research questionnaires.

Target trial emulation

We used the observational data to emulate each protocol component of the target trial as closely as possible (Table 1). We identified patients who met the eligibility criteria (see Table S2 for details on eligibility criteria implementation) and classified them into one of the 6 treatment groups according to the strategy that their data were compatible with at baseline.

We assumed treatment groups were exchangeable at baseline conditional on the baseline covariates age, race, sex, housing status, diagnosis, comorbid substance abuse, prior use of antipsychotics (including treatment failures), employment and educational status, and baseline symptomatic (CGI-S score) and functional (GAF score) statuses. We also included measures of frequency and intensity of substance use.

The analysis was identical to the per-protocol analysis of the target trial except that we emulated a sequence of target trials.¹² Specifically, we emulated a new target trial each week between August 1, 2010, and September 1, 2022. Each weekly trial included all patients who met the eligibility criteria and started one of the 6 antipsychotics of interest in that week. Each individual could participate in multiple trials.^7,13 Sequential emulation accommodates the fact that individuals may meet the eligibility criteria at several times during the follow-up and results in more precise estimates than choosing just one of those times as time zero. See Appendix S3 for details on the structure of the analytical dataset. To explore the robustness of our findings, we conducted a number of sensitivity analyses. First, we changed the outcome definition from all-cause psychiatric hospitalization to all-cause hospitalization to avoid relying on the valid recording of the main reason for hospitalization. Second, we changed the functional form of the continuous variables in our models (ie, CGI-S, GAF, and intensity of substance use) to categories (Table S1) to explore the sensitivity of the estimates to different model specifications. Finally, we conducted our analysis by using only the first eligible week for each patient to assess the gain in statistical efficiency of sequential emulation.

Benchmarking

We emulated the protocol of the EUFEST trial with 3 differences.

First, because current clinical guidelines often recommend the use of second-generation drugs (largely based on EUFEST findings¹⁴), few patients in our study initiated haloperidol. Therefore, we combined the first-generation drugs haloperidol, fluphenazine, and perphenazine into a single group.

Second, in current clinical practice many psychiatrists may initiate a particular antipsychotic agent (eg, paliperidone) with the a priori intention of switching to a long-acting injectable formulation during follow-up. However, switching to a long-acting injectable formulation of the same antipsychotic agent was not allowed in the protocol of EUFEST. Therefore, to emulate EUFEST, we modified the protocol of our target trial so that switching to a long-acting injectable formulation was not allowed.

Third, we used average hazard ratios estimated from a Cox model as effect measures for comparability with EUFEST. We considered the emulation successful if these estimates would have led to the same clinical decision as those from the index trial.

In sensitivity analyses, we standardized our estimates to the McLean population, which is similar to the EUFEST population in terms of distribution of risk factors (Appendix S4), and we conducted separate analyses by site to explore geographic variability (Appendix S5). Finally, we included psychosis NOS as an eligible diagnosis in our target trial emulation because it is common practice in the United States to record this diagnosis at first encounter and update it to a more specific one when more clinical information is available. Psychosis NOS was not an eligible diagnosis in EUFEST, so in sensitivity analysis we repeated the analysis restricting to those with other diagnoses.

Results

Figure 1 shows a flowchart for the 1123 individuals who initiated one of the eligible antipsychotic drugs during the follow-up. Of 1236 eligible initiations (corresponding to 1097 unique individuals), 11.1% were first-generation drugs, 22.4% were aripiprazole, 18.0% were olanzapine, 8.2% were quetiapine, 9.3% were paliperidone, and 31.0% were risperidone. Table 2 shows the baseline characteristics of initiators. Initiators of first-generation drugs were less likely to be White and had higher mean scores of CGI-S. Aripiprazole initiators were more likely to be female and employed, and less likely to have a schizophrenia diagnosis at baseline as compared with initiators of other agents. Paliperidone initiators were more likely to have a diagnosis of schizophrenia. The median follow-up was 31.0 weeks (interquartile range, 12.0-52.0). The proportion of patients who were lost to follow up was 16.4%; 9 patients died during follow-up. Most patients were eligible for only 1 trial during follow-up (92.3% of the sample), and patients who were eligible more than once had a higher mean CGI-S score at their second or third trial as compared with their first eligible trial. At the first trial, risperidone was the most used agent (33.1%), followed by aripiprazole (21.7%) and olanzapine (18.1%); at the second and third trial, aripiprazole was the most used agent (46.7%), followed by risperidone (20.0%) and paliperidone (13.3%).

Figure 1.

Flowchart for selection of eligible persons from the multi-country FEP-CAUSAL Collaboration when emulating the European First Episode Schizophrenia Trial (EUFEST). Numbers in parentheses represent unique individuals in each group. FEP, first-episode psychosis; FGA, first-generation antipsychotic.

Table 2.

Baseline characteristics (%) of initiators of antipsychotics, FEP-CAUSAL Collaboration, multiple countries, 2012-2022.

	First-generation antipsychotics (n = 137)	Aripiprazole (n = 277)	Olanzapine (n = 223)	Quetiapine (n = 101)	Risperidone (n = 383)	Paliperidone (n = 115)
Sociodemographic Characteristics
Age, yearsa	23.1 (3.5)	24.2 (4.6)	23.9 (4.6)	24.2 (5.0)	23.5 (3.9)	24.1 (3.8)
Female sex	27.7	38.6	19.7	30.7	20.6	20.0
White race	43.6	57.2	57.2	53.6	51.4	48.7
Years of educationa	11.8 (1.6)	12.3 (2.5)	12.1 (2.0)	11.6 (1.3)	11.9 (1.9)	11.8 (1.6)
Living alone	11.3	11.6	6.0	11.2	7.4	10.7
Employed or student	46.7	57.4	52.5	54.5	54.0	40.0
Diagnoses
Schizophreniform	4.4	9.7	11.2	6.9	9.9	12.2
Schizoaffective	16.8	16.8	13.9	15.8	11.0	17.4
Schizophrenia	50.4	37.2	50.7	50.5	46.2	60.0
Psychosis NOS	28.5	26.4	22.9	25.7	29.2	9.6
Clinical Characteristics
Antipsychotic-naive	61.3	67.5	64.6	77.2	79.6	57.4
Substance use	26.3	24.5	28.7	30.7	25.1	28.7
CGI-S scorea	4.1 (0.8)	3.9 (1.0)	3.9 (0.9)	4.0 (0.8)	3.9 (1.0)	3.9 (1.0)
GAF scorea	51.1 (11.5)	57.4 (10.6)	51.6 (11.4)	50.9 (11.9)	53.6 (11.8)	52.6 (12.2)
Overweight per BMIb	27.0	18.4	10.8	18.8	16.2	16.5

Abbreviations: BMI, body mass index; FEP, first-episode psychosis; CGI-S, Clinical Global Impression–Severity; GAF, Global Assessment of Functioning; NOS, not otherwise specified; SD, standard deviation.

^aValues reported as mean (SD).

^bBMI of >25, calculated as weight (kg)/height (m)².

The intraclass correlation coefficient was 0.67 for the Mental Illness Research, Education, and Clinical Center (MIRECC) version of the GAF symptom scale, 0.83 for the CGI-S scale, 0.75 for the Positive and Negative Syndrome Scale, and 0.68 for the Brief Psychiatric Rating Scale (Appendix S2).

Discontinuation of treatment strategy

The nonstabilized inverse-probability weights had a mean of 5.9 (standard deviation, 6.7) and a range from 1.0 to 100.3.

Figure 2 shows the standardized curves for time to discontinuation of treatment strategy. The 12-month risks of discontinuation were 61.5% (95% CI, 52.5, 70.6) for aripiprazole, 73.5% (95% CI, 60.5-84.9) for first-generation drugs, 76.8% (95% CI, 67.2-85.3) for olanzapine, 58.4% (95% CI, 40.4-77.4), for paliperidone, 76.5% (95% CI, 62.1-88.5) for quetiapine, and 74.4% (95% CI, 67.0-81.2) for risperidone. Compared with aripiprazole, the 12-month risk differences (95% CI) were −12.0% (95% CI, −27.9 to 5.0) for first-generation drugs, −15.3% (95% CI, −30.0 to 0.0) for olanzapine, −14.9% (95% CI, −33.7 to 1.4) for quetiapine, 3.0% (−21.5 to 30.8) for paliperidone, and −12.8% (−25.7 to −1.0) for risperidone. Results were similar in sensitivity analyses (Table S3).

Figure 2.

Estimated cumulative incidence of all-cause treatment discontinuation (A) and all-cause psychiatric hospitalization (B) by antipsychotic therapy, FEP-CAUSAL Collaboration (multiple countries, 2012-2022). FEP, first-episode psychosis.

All-cause psychiatric hospitalization

Figure 2 shows the cumulative incidences of all-cause psychiatric hospitalization. The 12-month risks of all-cause hospitalization were 19.1% (95% CI, 12.9-26.1) for aripiprazole, 19.2% (95% CI, 12.4-28.1) for first-generation antipsychotic agents, 19.3% (95% CI, 12.6-26.4) for olanzapine, 23.4% (12.1-35.4) for paliperidone, 21.4% (11.8-32.0) for quetiapine, and 28.1% (95% CI, 22.4-34.2) for risperidone.

Compared with aripiprazole, the 12-month risk differences were −0.9% (95% CI, −12.0 to 12.0) for first-generation agents, −0.2% (95% CI, −8.9 to 8.9) for olanzapine, −2.2% (95% CI, −15.5 to 10.1) for quetiapine, −4.2% (95% CI, −18.2 to 8.9) for paliperidone, and −8.9% (95% CI, −17.0 to 0.0) for risperidone. In sensitivity analysis, the risk differences were similar when considering all-cause hospitalization: 2.5% (95% CI, −9.8 to 11.6), 1.9% (95% CI, −7.5 to 11.7), −8.6% (95% CI, −23.0 to 4.6), −2.4% (95% CI, −18.7 to 11.8), and −9.3% (95% CI, −19.1 to 0.0), respectively, and when considering only the first trial per individual (Table S3).

Other outcomes

Table 3 shows the mean CGI-S scores at 12 months and the percentage of patients on mood stabilizers and antidepressants.

Table 3.

Estimated effectiveness measures (% and 95% CI)^a by antipsychotic treatment in first-episode psychosis, FEP CAUSAL Collaboration, 2012-2022.

	First-generation agents	Olanzapine	Quetiapine	Risperidone	Paliperidone	Aripiprazole
	n = 137	n = 223	n = 101	n = 383	n = 115	n = 277
12-month risk of discontinuation	73.5 (60.5-84.9)	76.8 (67.2-85.3)	76.5 (62.1-88.5)	74.4 (67.0-81.2)	58.4 (40.4-77.4)	61.5 (52.5-70.6)
12-month risk of hospitalization	19.2 (12.4-28.1)	19.3 (12.6-26.4)	21.4 (11.8-32.0)	28.1 (22.4-34.2)	23.4 (12.1-35.4)	19.1 (12.9-26.1)
12-month mean CGI-S scoreb	3.6 (3.3-3.9)	3.3 (3.1-3.5)	3.8 (3.4-4.1)	3.2 (2.9-3.5)	3.5 (3.2-3.8)	3.6 (3.4-3.8)
12-month mean GAF scoreb	54.8 (49.7-60.0)	58.0 (53.5-61.0)	57.5 (50.1-58.4)	61.0 (58.2-63.5)	58.9 (52.2-63.4)	57.2 (54.1-58.5)
12-month coprescription
Antidepressants	43.6 (25.4-61.8)	25.4 (15.1-35.8)	35.2 (24.3-46.2)	44.2 (25.0-63.4)	19.7 (0.0-39.7)	42.2 (29.5-55.0)
Mood stabilizers	7.3 (0.0-14.6)	11.6 (5.0-18.2)	12.7 (6.8-18.6)	8.7 (0.0-17.3)	27.5 (6.3-61.2)	1.5 (0.0-2.6)

Abbreviations: CGI-S, Clinical Global Impression–Severity; FEP, first-episode psychosis; GAF, Global Assessment of Functioning; NOS, not otherwise specified.

^aAdjusted for age, sex, race, housing status, baseline and time-varying CGI-S score, baseline and time-varying GAF score, years of education, antipsychotic use in the past, diagnosis (schizophreniform, schizophrenia, psychosis NOS, schizoaffective), center, diagnosis of substance use disorder, frequency and type of substance use, employment and educational status at baseline and time-varying. Estimated risk differences were standardized to the joint distribution of the baseline covariates.

^bValues reported as mean (standard deviation).

Benchmarking against the EUFEST randomized trial

The absolute risk estimates of discontinuation and hospitalization, but not of coprescription of antidepressants and mood stabilizers, were similar to those in EUFEST (Table S4). The proportion of patients that were lost to follow-up in EUFEST was 29.4%. Compared with first-generation drugs, the hazard ratios of discontinuation were 0.32 (95% CI, 0.24-0.43) for olanzapine and 0.41 (0.28-0.59) for quetiapine, similar to 0.28 (95% CI, 0.18-0.43) and 0.52 (95% CI, 0.35-0.76), respectively, in EUFEST. CGI-S scores were comparable and GAF scores were lower in our emulation than in EUFEST, but the differences between groups were largely compatible. Sensitivity analyses yielded results that were similar to those estimated in the main analyses (Table S3 and Appendix S4).

Discussion

We used observational data to emulate a target trial of antipsychotic treatment in individuals with FEP. We found that adherence to paliperidone and aripiprazole was higher than adherence to other second-generation agents, and that the risk of psychiatric hospitalization was greater for risperidone compared with other agents. In line with previous studies,^3,15 these findings support the consideration of both aripiprazole and paliperidone as first-line options. An advantage of both drugs is the possibility of transitioning to a long-acting injectable formulation.

The design of our target trial is an update of that of the EUFEST randomized trial, which compared therapeutic strategies that were common in the early 2000s. When we used our observational data to emulate a target trial with a design similar to that of EUFEST, we found similar hazard ratios of discontinuation for olanzapine and quetiapine vs first-generation antipsychotics, as well as similar estimates of the 12-month risk of psychiatric hospitalizations and 12-month average CGI-S scores for first generation agents, olanzapine, and quetiapine. The larger sample size of our study resulted in more precise estimates.

We found two notable differences between results from the original EUFEST and estimates from our observational emulation. First, our observational emulation found that the risk of discontinuation in the olanzapine arm was about 40% higher as compared with the risk estimated in EUFEST. Second, the proportion of patients coprescribed antidepressants was around 25% higher for quetiapine in the FEP-CAUSAL Collaboration. These discrepancies persisted even after standardizing our estimates to the patient characteristics of McLean OnTrack, the component of FEP-CAUSAL that most closely resembles the EUFEST population, which was older and had a higher proportion of individuals who were White, female, and employed or students. That is, even though the FEP-CAUSAL Collaboration collects high-quality data on treatments, outcomes, and confounders, a target trial emulation may not exactly replicate the index trial results because of differences in study population and variations in clinical practices. For example, current clinical guidelines recommend discontinuation of olanzapine if weight gain over 7% is detected,¹⁶ but this recommendation was not in place at the time EUFEST was conducted.

Our analysis has several limitations. First, like in any observational analysis, residual confounding may bias the results. However, we adjusted for key prognostic factors associated with treatment initiation (eg, substance abuse, comorbidities, educational and working statuses, symptomatic assessment) and our estimates benchmarked those of a previous randomized trial. Second, misclassification of some outcomes is possible. The reason for psychiatric hospitalization was not always available or may have been recorded imperfectly. However, when we modified our outcome definition to include all-cause hospitalizations, results were largely unchanged, except for a higher risk of hospitalization in those initiating quetiapine. Also, symptoms may be ascertained with error. However, our reliability analysis showed moderate to good reliability of our symptom measurements performed by chart review. Finally, model misspecification may have introduced bias in our effect estimates. However, our estimates were robust to changes in functional form of continuous covariates (ie, from linear and quadratic to categorical).

Our estimates might not be generalizable to other FEP populations with a different distribution of variables such as age, prior treatment failures, or psychiatric comorbidities. However, the inclusion of different international cohorts as part of the FEP-CAUSAL Collaboration consortium allows for a wider representation of patients with FEP, which facilitates transportability analyses.

In summary, our estimates support the selection of aripiprazole and paliperidone as first-line therapies for the treatment of individuals with FEP. The compatibility with the randomized trial estimates suggests that the observational data from the FEP-CAUSAL Collaboration data sufficed to approximately remove confounding for these clinical questions.

Funding

A.G.S. was partly supported by a 2022 NARSAD Young Investigator Grant from the Brain and Behavior Research Foundation.

Conflict of interest

The authors declare no conflicts of interest.

Data availability

The data used for this work is restricted ­use but can be obtained upon reasonable request and required credentialing approval.