Predictors of Hospitalization of Individuals With First-Episode Psychosis: Data From a 2-Year Follow-Up of the RAISE-ETP

RAISE-ETP enrolled English-speaking individuals between ages 15 and 40 years with a DSM-IV diagnosis of schizophrenia, schizoaffective disorder, schizophreniform disorder, brief psychotic disorder, or psychotic disorder not otherwise specified. Individuals were excluded if they had affective psychosis, substance-induced psychotic disorder, psychosis resulting from a general medical condition, clinically significant head trauma, or a serious general medical condition. All participants had experienced only one episode of psychosis (although this episode might have resulted in multiple hospitalizations) and had taken six or fewer months of lifetime antipsychotics.

Written informed consent was obtained from adult participants and from legal guardians of participants younger than age 18, who provided written assent. The institutional review boards of the coordinating center and the participating sites approved the study. The National Institute of Mental Health Data and Safety Monitoring Board provided study oversight.

Thirty-four outpatient community mental health centers in 21 states were selected via a national search. Site eligibility criteria included experience treating individuals with schizophrenia; interest in offering EIS for FEP; sufficient staff to implement the experimental intervention; ability to recruit an adequate number of participants; and institutional assurance that research assessments would be completed. Academic centers or sites with existing first-episode programs were excluded. All participants were outpatients at the time of their baseline assessment.

RAISE-ETP used cluster randomization (i.e., randomization by clinic rather than individual participant; 22). The study statisticians randomly assigned 17 of the clinics to the experimental intervention and 17 to standard care.

NAVIGATE (20), the experimental EIS, is team based and includes four interventions: personalized medication management, family psychoeducation, resilience-focused individual therapy, and supported education and employment (manuals available at www.raiseetp.org). The primary outcome measure and therefore the goal of RAISE-ETP was improved quality of life, not the prevention of hospitalization per se. These goals are not mutually exclusive in that hospitalization impedes progress toward improving quality of life. With respect to factors that might influence hospitalization risk, personalized medication management included assessment of symptoms, side effects, adherence, and substance use at each visit. The psychosocial interventions included illness management strategies and modules on adherence and making decisions about substance use. The control condition, community care, was psychosis treatment determined by individual and clinician choice and service availability.

Enrollment occurred between July 2010 and July 2012. The minimum potential trial duration for each participant was two years (longer for early enrollees); these two years are the focus of this report. Study assessments were suspended during periods of incarceration or hospitalization but resumed after release or discharge. Research assessments continued even if participants discontinued NAVIGATE or community care treatment.

Centralized assessors, masked both to individual treatment assignments and to the overall study design, administered the following measures via live, two-way video conferencing: the Structured Clinical Interview for DSM-IV (SCID-IV; 23) for diagnosis and to obtain the information required to determine the duration of untreated psychosis (DUP), the Positive and Negative Syndrome Scale (PANSS; 24), the Clinical Global Impressions Severity Scale (CGI-severity; 25), the Calgary Depression Scale for Schizophrenia (CDSS; 26), and the Heinrichs-Carpenter Quality of Life Scale (QLS; 27), which was the primary outcome measure. Remote assessment via two-way video conferencing is comparable to face-to-face assessment in patient acceptability and reliability (28). The SCID-IV was completed at baseline; the other measures were completed every six months.

Site research assistants interviewed participants monthly with the Service Use and Resource Form (SURF; 29, 30) to capture psychiatric inpatient and emergency services use and self-reported days of alcohol or drug use. Emergency department visits that lasted more than 24 hours were considered hospitalizations. Participant-reported assessments allowed us to obtain information not only about treatment that participants received at their study site but also about treatment they received outside the site (e.g., inpatient admission at another agency). Participant self-report has proven to be generally accurate (29). The outcome of interest for this article was mental health hospitalization occurring after study entry (all participants were outpatients at this time point). We obtained data on hospitalizations before study entry through individual and family interviews and medical record search; these data were examined as predictors of hospitalization during the study.

At baseline, 3-month follow-up, 6-month follow-up, and every 6 months thereafter, participants completed the Intent to Attend measure (31), the Adherence Estimator (32), Brief Evaluation of Medication Influences and Beliefs (BEMIB) scale (33), seven items from the Stigma Scale (34), a subset of the Perceived Well-Being Scale (35), the six-item Autonomy Support Scale short-form version of the Health Care Climate Questionnaire (36), and an abbreviated version of the Mental Health Recovery Measure (37). They also rated their current state of mental and emotional health on a scale ranging from 1, worst possible, to 100, perfect health, and how they felt about their life as a whole on a scale ranging from 1, terrible, to 7, delighted.

We used time-to-event analysis for hospitalization. The Cox proportional hazard model was used with site included as a frailty term to account for clustering of individuals within site. Clustered randomized trials often have a limited number of clusters, and this can result in an imbalance on baseline measures between the randomized treatment conditions. Such imbalances may confound the relationship between treatment and individual-level outcomes. Therefore, significant baseline differences between the treatment conditions were included as adjustment variables. We assessed whether the two treatment conditions differed in hospitalization and adjusted for the baseline covariates of gender, student status at entry, and total PANSS score, which were found to be significantly associated with treatment condition in previously reported analyses (16).

For analysis of the longitudinal assessments, we constructed time-varying predictor variables that consisted of the results of the assessment concurrent with or, if not concurrent with, the assessment closest in time to an individual’s first hospitalization. The severity and intensity of a factor often change over time. The use of time-varying predictor variables allowed us to examine the effects of a variable of interest assessed at the time closest to a hospitalization, when it might have had the greatest impact on hospitalization. For example, if an individual had a hospitalization at month 18, we used the month 18 assessment; if those results were not available, we used the closest preceding assessment to month 18. For individuals with no hospitalizations, we used the results from the last assessment.

To determine hospitalization predictors, we first performed univariate analyses using a Cox proportional model with frailty of site for each candidate of the baseline and time-varying covariates. In developing the multivariate model for baseline predictors, the variables we screened for entry into the analysis were those with significant or trend-level associations in the univariate analyses of baseline predictors. Because inclusion of correlated variables can result in unstable multivariate correlations, we used several criteria to determine which correlated variables to enter. We gave preference to variables that would provide more clinically meaningful information if associations with hospitalization were found (e.g., we preferred factor scores over total scores because the former describe more circumscribed symptoms than the latter and could provide more precise targets for intervention development). The final set of baseline variables for model entry included DUP longer than 74 weeks, number of prior hospitalizations, PANSS positive and excited factors, CDSS, and the Perceived Well-Being Scale.

The strategy for developing the final multivariate models that integrate both baseline and time-varying variables was to consider for entry first, the baseline variables with significant associations with hospitalization in the multivariate baseline analysis and second, the time-varying variables with significant associations with hospitalization from the univariate analyses of time-varying variables and hospitalization. By examining the correlations among variables and using our strategy of considering clinical meaningfulness in variable selection, we developed two groups of variables for entry into the analyses. Both groups included the baseline predictors DUP and number of prior hospitalizations and the time-varying predictors of days of illegal drug use and being a student or working. In addition, in analysis 1 we added time-varying variables rated by the central assessors—the PANSS positive and excited factors and the CDSS—plus the self-rated Adherence Estimator. In analysis 2, we added time-varying variables rated by the participant, the BEMIB, and the Perceived Well-Being Scale total score. The participant-rated Adherence Estimator and BEMIB measures were highly correlated (r=−.43). Thus, we could include only one of them in analysis 2, which focused on participant-rated assessment. We chose to include the BEMIB in analysis 2 because it taps participants’ beliefs about the value of medication for themselves. The Adherence Estimator taps general attitudes toward medication and was not significantly correlated with any of the central assessor-rated variables; it was included in analysis 1.

For each analysis, we checked proportional hazard assumptions by dividing time into 6-month intervals and assessing whether the coefficients were statistically different across time intervals.

Characteristics of the full RAISE-ETP sample of 404 individuals have been published (16). Some participants did not have any postbaseline assessments. Supplemental Table 1, which presents the characteristics of the 382 participants who had at least one postbaseline assessment and thus have postbaseline hospitalization data, is available as an online supplement to this article. Overall, the 382 participants were young (mean±SD age =23.2±5.1 years), mostly male (73%, N=279), and of diverse racial background. Outpatient community center sites typically receive most of their FEP referrals from inpatient units. Consistent with this pattern, only 84 participants had never had an inpatient psychiatric hospitalization.

Of the participants, 112 had at least one psychiatric hospitalization during the two-year observation period. On the basis of a survival analysis, 34% of NAVIGATE and 37% of community care participants had a hospitalization (this estimate is the same as that previously reported [16] for the sample of 404 individuals as a result of censoring effects with survival analysis of individuals who did not have postbaseline assessments). Hospitalization rates did not differ between participants receiving NAVIGATE and those receiving community care treatment (hazard ratio=0.892, χ²=0.35, df=1, p=.557).