Transdiagnostic Modeling of Clinician-Rated Symptoms in Affective and Non-Affective Psychotic Disorders

Abstract

Prevailing factor models of psychosis are centered on schizophrenia-related disorders defined by the DSM and ICD, restricting generalizability to other clinical presentations featuring psychosis, even though affective psychoses are more common. This study aims to bridge this gap by conducting exploratory and confirmatory factor analyses, utilizing clinical ratings collected from patients with either affective or non-affective psychoses (n = 1042). Drawing from established clinical instruments, such as the Positive and Negative Syndrome Scale, Young Mania Rating Scale, and Montgomery-Åsberg Depression Rating Scale, a broad spectrum of core psychotic symptoms was considered for the model development. Among the candidate models considered, including correlated factors and multifactor models, a model with seven correlated factors encompassing positive symptoms, negative symptoms, depression, mania, disorganization, hostility, and anxiety was most interpretable with acceptable fit. The seven factors exhibited expected associations with external validators, were replicable through cross-validation, and were generalizable across affective and non-affective psychoses.

General Scientific Summary

The aim of this study is to formulate a transdiagnostic dimensional model by integrating well-established clinical ratings that encompass a wide range of core symptoms observed in both affective and non-affective psychoses. We demonstrate that a multidimensional symptom model representing both psychotic (positive symptoms, negative symptoms, disorganization) and mood symptoms (depression, mania, hostility, and anxiety) is most interpretable and applicable across psychotic diagnoses. Considering the complex interrelationships among identified symptom dimensions, a dimensional approach may be more suitable for characterizing the symptom profiles of psychotic patients than a categorical diagnostic approach.

Growing evidence from research has identified a substantial overlap in genetics, endophenotypes, and clinical expressions between psychotic and mood disorders, highlighting that some aspects of the underlying pathophysiology are shared (Baker et al., 2019; Bora et al., 2010; Cohen, 2016; Cohen, Öngür, et al., 2021; Huang et al., 2010; Lichtenstein et al., 2009; Purcell et al., 2009). However, standard psychiatric diagnostic systems, such as the Diagnostic and Statistical Manual of Mental Disorders (DSM-5; American Psychiatric Association, 2013) and the International Classification of Diseases (ICD-11; World Health Organization, 2019), continue to categorize mental disorders into discrete, non-overlapping categories. While these systems have historically been instrumental in shaping clinical practices and developing treatments, they now lag behind contemporary research, failing to adequately represent the complex etiological and pathophysiological structures of mental illnesses (Hyman, 2010).

The DSM-5 and ICD-11 take into account factors such as the severity, frequency, and duration of psychotic symptoms and the co-presence of affective disturbances in their criteria for diagnosing non-affective and affective psychoses. Non-affective psychosis, encompassing schizophrenia-related disorders, is characterized by enduring delusions and hallucinations, often accompanied by disorganized thinking and “negative” symptoms (e.g., disattention and avolition). In contrast, affective psychosis typically involves prolonged mood dysregulation, such as depression or mania, usually preceding the onset of psychotic symptoms. Although these conditions are categorized as distinct disorders, the prevalence of affective symptoms in psychotic disorders with varying degrees of severity poses a challenge in fitting patients neatly into these predefined categories (Buckley et al., 2009; Ravichandran et al., 2021). Furthermore, the co-occurrence of affective and subclinical psychotic symptoms often precedes the disorder’s onset (Addington et al., 2017; Fusar-Poli et al., 2013), indicating a possible etiological link between these conditions rather than being an epiphenomenon (Wigman et al., 2012). Therefore, developing a comprehensive measurement model that represents the multifaceted clinical profiles with varying severity is critical for clinicians to identify and treat. Such an approach will further facilitate advancing precision psychiatry and personalized care (Cohen & Öngür, 2023).

Past research on the factor structure of clinician-rated symptoms in schizophrenia and other primary psychotic disorders has broadly settled on five dimensions, often referred to as the “pentagonal model” (Lehoux et al., 2009; Os & Kapur, 2009; Potuzak et al., 2012; Wallwork et al., 2012a). The specific symptom dimensions often include positive symptoms, negative symptoms, and disorganization representing the non-affective domain and emotional distress/depressive and excited/mania representing the affective domain. Two limitations of these methods limit the generalizability and construct coverage of this research. First, by focusing on schizophrenia samples, most previous studies have excluded or underrepresented participants with affective psychosis and subclinical presentations. This approach has likely restricted the variance of affective symptoms and suboptimally characterized the covariance between affective and non-affective dimensions. This may also contribute to variation in the factor structure that best describes each diagnostic category featuring psychosis (Anderson et al., 2017). Second, clinical ratings covering affective symptoms are often underspecified in the clinical instruments designed to assess psychotic symptoms (e.g., the PANSS; Kay et al., 1987). For example, while a multitude of signs and symptoms need to be observed to capture the scope of depression and mania reliably, PANSS has limited items available that assess these constructs (e.g., General 6-depression, Positive 4-excitement). Therefore, previous factor models solely relying on PANSS items are prone to underspecification, underdetection, and instability in the affective factors.

There have been few factor analytic studies to date that (a) examined a sample with both affective and non-affective psychoses and (b) included adequate items assessing mania and depression. (Ravichandran et al., 2021) developed a dimensional model with four correlated factors (positive symptoms, negative symptoms, depression, and mania) from a transdiagnostic psychosis sample assessed with multiple clinical instruments—the PANSS, Young Mania Rating Scale (YMRS; R. C. Young et al., 1978), and Montgomery-Åsberg Depression Rating Scale (MADRS; (Montgomery & Åsberg, 1979)—derived from the same database as the current study. Their aim was to predict the class of psychotropic medication prescribed at discharge using a data-driven and dimensional approach. Factor scores derived from their model provided better discriminability than the DSM diagnoses (although diagnoses did correlate with treatment choices, as expected). Consistent with these findings, clinicians have been noted to choose treatments based on factors or symptoms more than DSM or ICD diagnoses (Cohen, Ravichandran, et al., 2021; First et al., 2018).

A few previous studies (Anderson et al., 2018; Reininghaus et al., 2013, 2019) have pooled together participants diagnosed with schizophrenia, schizoaffective, and bipolar disorder diagnoses and applied so-called “multifactor” models (Reise et al., 2013). Such models extend the traditional correlated factors structure by replacing the factor intercorrelations with general and domain factors. Specifically, a bifactor model includes a general factor that all items load on (thereby capturing the common liability across dimensions of psychopathology; Caspi et al., 2013) as well as multiple specific factors that subsets of items also load on (similar to the factors in a correlated factors model except that they are estimated in the context of the general factor and are typically assumed to be orthogonal). Similarly, a trifactor model includes a general factor and specific factors like a bifactor model but further adds multiple domain factors that other subsets of items also load on. Domain factors can be used to capture shared method variance (e.g., patient-reported items vs. clinician-rated items) or, in this case, groupings of specific factors based on theoretical constructs (e.g., into affective and nonaffective domains). Depictions of these different factor structures are provided in Figure 1.

Figure 1. Path Diagrams of (a) Correlated Factors, (b) Bifactor, and (c) Trifactor Structures.

Note. Rectangles depict observed items, ovals depict latent factors, single-headed arrows depict factor loadings, and connected arrows depict factor correlations. Other model parameters are omitted for visual clarity.

Anderson et al., 2018 have shown that a bifactor model of the PANSS, comprising five specific factors (positive symptoms, negative symptoms, disorganization, excited, anxiety/depression) exhibited a superior fit over the correlated five-factors model in a large transdiagnostic sample. Reininghaus et al., 2019 proposed a trifactor model that included five specific factors (positive symptoms, negative symptoms, disorganization, mania, and depression), two domain factors (affective symptoms and non-affective symptoms), and a general factor; they found that this model fit the data better than more conventional approaches (e.g., correlated factors models or bifactor models). Multifactor—especially bifactor—models have become increasingly popular in recent years as they tend to fit better than correlated factor models and seem to reflect many patients’ presentations (G. T. Smith et al., 2020). However, these approaches are controversial as their high flexibility makes them susceptible to overfitting (i.e., failing to generalize) and the general factors they estimate are not guaranteed to be true nosological entities (Bornovalova et al., 2020; Watts et al., 2019).

Recent advances in studying the structure of psychopathology more broadly provide clues about the structure and relationship of affective and non-affective psychosis. The Hierarchical Taxonomy of Psychopathology (HiTOP) is a transdiagnostic and dimensional alternative to traditional categorical nosologies that is hierarchical in the sense that maladaptive traits and symptom components are organized within “spectra,” which are in turn organized within “super spectra” (Kotov et al., 2017). Spectra are also sometimes further divided into “subfactors.” HiTOP conceptualizes psychosis as a super spectrum containing two spectra: psychoticism and detachment (Cicero et al., 2022; Kotov et al., 2022). Psychoticism¹ includes maladaptive traits and symptom components related to perceptions, cognitions, and behaviors that are disconnected from reality (e.g., fantasy proneness, unusual experiences, unusual beliefs, peculiarity, reality distortion, and disorganization). Detachment includes maladaptive traits and symptom components related to apathy, disinterest in people, and blunted affect (e.g., emotional detachment, anhedonia, social withdrawal, romantic disinterest, inexpressivity, and avolition). For affective psychosis, the internalizing spectrum is also highly relevant and encompasses several subfactors including fear (e.g., anxiety, panic, phobias, and rituals) and distress (e.g., dysphoria, suicidality, irritability, and dissociation). More research is needed to determine whether maladaptive traits and symptom components related to mania (e.g., expanded mood, decreased sleep, grandiosity, recklessness, and emotional lability) are best conceptualized as a subfactor of the psychoticism spectrum (Kotov et al., 2022) or as a special subfactor that cross-loads on both psychoticism and internalizing (Cicero et al., 2022).

In the present study, we conducted numerous factor analyses to elucidate transdiagnostic symptom dimensions of psychosis by aggregating symptom ratings from three well-established clinical instruments (i.e., PANSS, YMRS, MADRS) that cover a broad array of core symptoms observed in affective and non-affective psychotic disorders. To do so, we accessed a large and heterogeneous dataset of patients hospitalized for psychosis across numerous diagnostic categories, including schizophrenia spectrum disorders, bipolar disorders, depressive disorders, and anxiety disorders (N = 1042). Our main research questions were as follows: (1) What is the best correlated-factors model structure for the PANSS, YMRS, and MADRS items? (2) Do bifactor and trifactor versions of this structure improve model fit above and beyond the correlated factors model? (3) If one of these versions does improve model fit, do they remain to retain interpretability and theoretical coherence? (4) Is our favored measurement model transdiagnostically applicable across patients with affective and non-affective psychoses? (5) Can our results inform research on the broader structure of psychopathology (e.g., HiTOP)?

Transparency and Openness

All R code, results, supplementary materials are available at https://osf.io/arvsf. This study involved analysis of pre-existing data and was not pre-registered.

Methods

Participants

The dataset used in this report was drawn from an ongoing multidisciplinary (nosologic, imaging, genetic, cell biology) study of psychotic and mood disorders. The majority of participants, primarily in the early phase of psychosis, were recruited from the McLean Hospital’s program specializing in psychotic disorders, with a significant portion initially recruited from the inpatient unit. The eligibility criteria and study procedures are detailed in a prior report (S. Young et al., 2013). Patients were not eligible for the study if their psychiatric symptoms were due to a general medical condition, head trauma, or substance use before the clinical onset. The Partners HealthCare Institutional Review Board approved the study and all participants provided informed consent.

From this database, we included a total of 1042 patients who were hospitalized for affective and non-affective psychotic illnesses. Primary diagnoses include schizophrenia (SCZ; n = 241), schizoaffective disorder (SZA; n = 222), bipolar disorder with psychotic features (BD; n = 444), major depressive disorder with psychotic features (MDD; n = 58), and psychosis not otherwise specified (PNOS; n = 33). Other diagnoses included delusional disorder, schizophreniform disorder, obsessive-compulsive disorder, and any anxiety disorders (n = 44). The Structured Clinical Interview for the DSM-IV (SCID-I; First, 1997) was used for diagnosis. Research assistants, licensed psychologists, and attending psychiatrists who conducted the clinical interviews attended training sessions to maintain reliability (Öngür et al., 2009).

All participants in this report had data available for the following three clinical instruments (described in detail below): the PANSS, MADRS, and YMRS. Subsets of the participants also had the Beck Anxiety Inventory, the Beck Depression Inventory, the Multnomah Community Ability Scale, and the North American Adult Reading Test (NAART) data available, which were used as external criterion variables.

Materials

Positive and Negative Symptom Scale (PANSS)

The PANSS is a 30-item rating scale completed by a clinician to assess distinct symptom domains in schizophrenia (Kay et al., 1987, 1989). It comprises three sub-domains: the Positive scale (7 items) for assessing positive symptoms (e.g., delusions, conceptual disorganization, and excitement), the Negative scale (7 items) for assessing negative symptoms (e.g., blunted affect, emotional withdrawal, and poor rapport), and the General Psychopathology scale (16 items) for assessing broader aspects of the psychiatric illness (e.g., somatic concerns, mannerisms and posturing, and poor attention). Each item is rated on a scale from 1 (absent) to 7 (extreme), with item-specific descriptive anchors for all seven response options.

Montgomery & Åsberg Depression Rating Scale (MADRS)

The MADRS is a 10-item rating scale completed by a clinician to assess depressive symptom severity based on patient reports and clinical observation (Montgomery & Åsberg, 1979). It was developed to enhance sensitivity to treatment effects by deriving items relevant to core symptoms of depression from the 65-item comprehensive psychopathology rating scale (CPRS; Åsberg et al., 1978). Each item is rated on a scale from 0 (normal/not present) to 6 (extreme) with item-specific descriptive anchors for the even-numbered response options only.

Young Mania Rating Scale (YMRS)

The YMRS is an 11-item rating scale completed by a clinician to assess mania symptom severity based on patient reports and clinical observation (R. C. Young et al., 1978). Note that depressive symptoms (which are common in bipolar disorder) are not assessed by this scale. Four items are rated on a scale from 0 to 8 (with item-specific descriptive anchors for the even-numbered response options only) and the other seven items are rated on a scale from 0 to 4 (with item-specific descriptive anchors for all response options).

Criterion Variables

The following measures were only available for a subset of participants and were used as external criteria to validate the factors derived from the PANSS, MADRS, and YMRS. The Beck Anxiety Inventory (Beck et al., 1988) is a 21-item self-report questionnaire for measuring the severity of anxiety symptoms; it was available for 313 of our participants (30%). In our total sample, the reliability of the BAI mean score was very good (⍵ = .93). The Beck Depression Inventory (BDI; (Beck et al., 1961) is a 21-item self-report questionnaire for measuring the severity of depression symptoms; it was available for 327 of our participants (31%). In our total sample, the reliability of the BDI mean score was very good (⍵ = .93). The Multnomah Community Ability Scale (MCAS; (Barker et al., 1994)is a 17-item rating scale for measuring the degree of functional ability in the community in adults with psychiatric disorders; it was available for 679 of our participants (65%). In our total sample, the reliability of the MCAS mean score was good (⍵ = .85). The North American Adult Reading Test (NAART; (Blair & Spreen, 1989) is a clinician-administered neuropsychological test to estimate premorbid intelligence based on word pronunciation ability. Estimates of premorbid Full Scale IQ (FSIQ) derived from the NAART were available for 343 of our participants (33%). While we were unable to estimate its reliability because only a single FSIQ estimate is derived per participant, previous study demonstrated that NAART has high test-retest reliability of .98 (Crawford et al., 1989). Furthermore, a prospective longitudinal study reported that premorbid IQ estimates based on pronouncing irregularly spelled words remain stable over time in patients with schizophrenia (D. Smith et al., 1998).

Statistical Analyses

Data Preprocessing

Several preprocessing steps were undertaken to prepare the data for analysis. First, missing values for the PANSS, YMRS, and MADRS were minimal (0.002%) and were considered missing completely at random (e.g., due to rater error). Some of these missing values were imputed if a similar item was available in another clinical instrument. Subjects with missing values that could not be imputed were omitted from the analysis. Second, each item was re-coded to have three or four levels to minimize categorical sparsity (DiStefano et al., 2021). Adjacent levels were collapsed to create four levels of increasing severity; then, if the fourth level did not contain at least 10% of the sample, the third and fourth levels were further collapsed to create three levels. Third, we decided to drop 7 of the 51 total items from analysis due to extremely low variability (i.e., PANSS Mannerisms and Disorientation) or high overlap with other included items (i.e., PANSS Depression; MADRS Sleep; and YMRS Insight, Aggressive Behavior, and Thought Content). Our analyses thus start with 44 items. See Supplementary Figures S1 to S5 for more details on these preprocessing steps (e.g., code and histograms).

Data Partitioning

The previous literature on transdiagnostic symptoms in psychosis was insufficient to provide concrete structural hypotheses. Therefore, instead of pre-registering specific structures to test, we explored many different structures within a rigorous cross-validation procedure. By splitting our data in half, we were able to use one portion to discover potential structures in a data-driven manner and then test the generalizability of those structures in the other unseen portion, thereby mitigating the risk of overfitting (Rooij & Weeda, 2019). Our total sample (N = 1042) was thus partitioned into two independent datasets using the anticlustering method (Papenberg & Klau, 2021). One half (n = 521), which we named the “discovery set,” was used for learning about our data, trying different modeling approaches, and ultimately electing a small number of candidate models. The other half (n = 521), which we named the “validation set,” was used for model evaluation, comparison, and selection. Once a single model was selected as our “favored” model, we conducted further analyses aimed at validating this model using the total dataset in order to maximize statistical power and minimize sampling error.

Correlated Factors Models

In the discovery set, we used parallel analysis (Lim & Jahng, 2019) to inform the number of factors to extract and examine. We then used the lavaan R package (Rosseel, 2012) to estimate EFA models with the indicated number of factors. The EFA models were fit using the ML estimator (which treats the items as continuous) and “geomin” rotation (which is oblique). We then derived several candidate confirmatory factor analysis (CFA) models from the results and estimated them in the discovery set using the WLSMV estimator (which treats the items as ordinal; Muthén, 1993). The candidate models, described in the results section, differed in terms of which items they included (e.g., some models dropped items that had low loadings on all factors) and whether cross-loadings were permitted. Modification indices (MI) were consulted to determine which cross-loadings would most improve model fit, but a cross-loading was only included if it could be justified theoretically (see Supplementary Figure S6 for more details on how the candidate models were derived). Considering both theoretical coherence and empirical performance (Fabrigar et al., 1999), we compared these models using the root mean squared error of approximation (RMSEA), comparative fit index (CFI), and standardized root mean square residual (SRMR). A model’s fit was considered “acceptable” if its RMSEA ≤ 0.08, CFI ≥ 0.90, and SRMR ≤ 0.10 (Kline, 2015), whereas a model’s fit was considered “good” if its RMSEA ≤ 0.06, CFI ≥ 0.95, and SRMR ≤ 0.08 (Hu & Bentler, 1999).

Multifactor Models

In the discovery set, we adapted our best correlated factor model into a candidate bifactor model by replacing all cross-loadings and factor correlations with loadings for all included items onto a general factor. We then further adapted this bifactor model into a candidate trifactor model by also adding loadings for non-affective items onto a new non-affective domain factor and also adding loadings for affective items onto a new affective domain factor. See Supplementary Figure S14 for multifactor models’ path diagrams. Given estimation issues (e.g., model non-identification and singular solutions) when trying to implement these models in lavaan, we followed Reininghaus et al., 2019 in fitting these models with the mirt R package (Chalmers, 2012), which implements multidimensional item response theory analyses using the Metropolis-Hastings Robbins-Monro (MHRM) algorithm for model estimation. All items were treated as ordinal using the graded response model (Samejima, 1969). We compared these models to each other (as well as to our final correlated factors model, re-fit in mirt) using RMSEA, CFI, and SRMR, as well as the Akaike information criterion (AIC), and the Bayesian information criterion (BIC). AIC and BIC do not have absolute thresholds but can be used to compare models fit to the same data in a relative manner; specifically, models with lower AIC and BIC show better performance.

Model Generalizability and Selection

All candidate models were re-fit in the validation set to evaluate their generalizability to unseen data. The correlated factors models were fit using lavaan and the WLSMV estimator and compared using RMSEA, CFI, and SRMR. (Note that AIC and BIC are not available when using the WLSMV estimator in lavaan.) Our multifactor models, as well as our best correlated factors model, were also fit using mirt and the MHRM estimator and compared using RMSEA, CFI, SRMR, AIC, and BIC. Considering both theoretical coherence and empirical performance in both the discovery and validation sets, we settled on a single candidate model as our “favored model” to further validate (as described next) and potentially recommend to others.

Measurement Invariance

We tested whether the structure of our favored model was invariant across groups of participants diagnosed with affective psychosis (i.e., bipolar disorder, major depressive disorder, and any anxiety disorders with psychotic features) and non-affective psychosis (i.e., schizophrenia, schizoaffective, psychosis not otherwise specified). Using lavaan and WLSMV, we fit a series of models to the total sample where each successive model constrained an additional set of parameters to equality between the groups. We began with a “configural” model that simply required the same structure in both groups. We then estimated a model with equal factor loadings, a model with equal factor loadings and intercepts, and finally a model with equal factor loadings, intercepts, and latent means. At each step, we calculated fit indices, the change in each fit index, and the scaled chi-squared difference test for adjacent models. We interpreted the results as consistent with non-invariance if CFI decreased by more than 0.010 while RMSEA also increased by more than 0.015 or SRMR also increased by more than 0.010 (Chen, 2007)

External Validation

To validate the factors from our favored model, we correlated them with several external criterion variables (BAI, BDI, MCAS, and NAART) using the total sample. This was done within a single, larger CFA model with MLR estimation so that the correlations would be based on the latent variables rather than estimated factor scores. (MLR was used in place of WLSMV so that we could account for missing data in the criterion variables using full information maximum likelihood.) Our favored model was replicated with all items and factors, and these factors were correlated with parcel factors representing the BAI, BDI, MCAS, and NAART. A parcel factor (using the “all-item-parcel” approach; (Matsunaga, 2008) is constructed by calculating the mean of all items on a scale and then estimating a single-indicator factor where the loading is constrained to 1.0 and the item’s residual variance is constrained to (1 − ω)s², where ω is the scale’s estimated reliability and s² is the sample variance of the mean indicator. This hybrid approach, combining multiple-indicator factors with parcels, allowed us to reap the benefits of parceling for the external criterion variables (e.g., enhanced normality and reliability, reduced multicollinearity, and a simplified measurement model) without suffering the potential downsides of parceling for our factors of primary interest (e.g., masking of poor item performance or multidimensionality).

Results

Sample Characteristics

The mean age of all participants was 36.2 years (SD = 12.9, range = 18 to 71 years old) and 56% of the sample was male. Representing the local demographics, the sample was 61% White, 29% African American, and 4% Asian, with around 60% of the sample having some college education or more. Patients across categorical diagnoses showed elevated scores in PANSS positive total (M = 19.14, SD = 7.37) and PANSS general total (M = 33.59, SD = 9.70), indicating the presence of psychotic symptoms and significant psychological distress. Descriptions of sociodemographic and clinical variables in the overall sample and its subdivisions are provided in Table 1. Sociodemographic table by DSM diagnosis is available in Supplementary Table S1.

Table 1.

Demographic and Clinical Characteristics in the Discovery Set, Validation Set, and Total Sample

Characteristic	Discovery Set	Validation Set	Total Sample
Count: n	521	521	1042
Age: M (SD)	36.6 (12.9)	36.0 (12.8)	36.3 (12.9)
Sex: n (%) Male	300 (58%)	281 (54%)	581 (56%)
Race
White: n (%)	317 (61%)	322 (72%)	639 (61%)
Black: n (%)	161 (31%)	147 (28%)	308 (30%)
Asian: n (%)	14 (3%)	22 (4%)	36 (3%)
AI/AN: n (%)	9 (2%)	11 (2%)	20 (2%)
Other: n (%)	19 (4%)	18 (3%)	37 (3.56)
Education
No Diploma/GED: n (%)	28 (6%)	34 (7%)	62 (6%)
Diploma/GED: n (%)	79 (16%)	74 (15%)	153 (15%)
Some College: n (%)	207 (41%)	198 (39%)	405 (40%)
2-Year College: n (%)	24 (5%)	26 (5%)	50 (5%)
4-Year College: n (%)	88 (18%)	97 (19%)	185 (18%)
Some Graduate School: n (%)	28 (5%)	31 (6%)	59 (6%)
Graduate Degree: n (%)	47 (9%)	46 (9%)	93 (9%)
Primary Diagnosis
Schizophrenia: n (%)	124 (24%)	117 (22%)	241 (23%)
Schizoaffective Disorder: n (%)	108 (21%)	114 (22%)	222 (21%)
Psychotic Disorder NOS: n (%)	19 (4%)	14 (3%)	33 (3%)
Bipolar Disorder: n (%)	230 (44%)	214 (41%)	444 (43%)
Major Depressive: n (%)	26 (5%)	32 (6%)	58 (6%)
Other Disorder: n (%)	14 (3%)	30 (6%)	44 (4%)

Notes. AI/AN = American Indian or Alaskan Native, GED = General Educational Development Test, NOS = Not Otherwise Specified, Other Disorder = Schizophreniform, delusional, brief psychosis, or any anxiety disorder.

Correlated Factors Models

Parallel analysis suggested that the first four factors explained considerable variance in the observed data and that adding more than seven factors would capture unreplicable noise (see the scree plot in Supplementary Figure S7). Given these results, we extracted and examined EFA models with four, five, six, and seven factors (i.e., models A through D). The seven-factor EFA model (model D) had the best fit of the four considered and had good fit according to RMSEA and SRMR as well as acceptable fit according to CFI (Table 2). This model’s factors were recognizable as capturing positive symptoms, negative symptoms, disorganization, mania, depression, anxiety, and hostility.² The YMRS and MADRS items loaded predominantly, but not exclusively, on the mania and depression factors, respectively. Complete details of all the EFA models (i.e., code and results for 4 to 7 factors model) are presented in Supplementary Figure S7.

Table 2.

Description and Performance of Factor Analysis Models in the Discovery Set

Exploratory (ML)	Observed	Simulated	df	Chi2	RMSEA	CFI	SRMR
A: Four-factor Oblique	.920	.460	776	2446.6 *	.064 a	.839	.041 b
B: Five-factor Oblique	.606	.424	736	1907.2 *	.055 b	.887	.036 b
C: Six-factor Oblique	.486	.387	697	1554.4 *	.049 b	.917 a	.030 b
D: Seven-factor Oblique	.377	.359	659	1343.4 *	.045 b	.934 a	.026 b
Confirmatory (WLSMV)	Items	Cross	df	Chi2	RMSEA	CFI	SRMR
E: Simple Structure	44	0	881	4516.0 *	.089	.811	.136
F: Trimmed Items	33	0	474	2012.6 *	.079 a	.910 a	.101
G: Cross-loadings	33	2	472	1759.5 *	.072 a	.924 a	.091 a
Confirmatory (MHRM)	General	Domain	df	M2	RMSEA	CFI	SRMR
G: Cross-loadings	0	0	437	1597.4 *	.071 a	.942 a	.140
H: Bifactor Model	1	0	421	2123.3 *	.088 a	.915 a	.177
I: Trifactor Model	1	2	388	1354.4 *	.069 a	.952 b	.145

Notes. ML = Maximum likelihood, WLSMV = Weighted least squares means and variance adjusted (lavaan), MHRM = Metropolis-Hastings Robbins-Monro (mirt), Observed = Eigenvalue of the final observed component, Simulated = Eigenvalue of the final simulated component during parallel analysis, df = Degrees of freedom, Chi2 = Scaled chi-squared, RMSEA = Scaled root mean squared error of approximation, CFI = Scaled comparative fit index, SRMR = Standardized root mean square residual, Items = Number of unique items, Cross = Number of items with cross-loadings, General = Number of general factors, Domain = Number of domain (or method) factors, M2 = M2 statistic

^*p < .001

^aAcceptable fit

^bGood fit. In each table section, the best value per performance metric is bolded.

We next adapted the seven-factor EFA solution into three candidate CFA models and compared their fit in the discovery set. The first candidate model (model E) included all 44 items and had a simple structure such that each item loaded on a single factor. The second candidate model (model F) modified the first by dropping 11 items that had low loadings (factor loading < 0.5) and high cross-loading (complexity index >2.5) on all factors in the EFA. Finally, the third model (model G) modified the second by allowing two items to cross-load (i.e., YMRS Language on mania and disorganization, PANSS Motor Retardation on depression and negative symptoms). The performance of these models is provided in Table 2 (with full details in Supplementary Figure S8). Model E did not show acceptable fit on any fit index, Model F showed acceptable fit on RMSEA and CFI only, and Model G showed acceptable fit on all three.

Analyses in the Validation Set

All three candidate CFA models were re-fit in the validation set. The results were similar to those from the discovery set in that Model E did not show acceptable fit on any index and Model G showed acceptable model fit on all three indexes; this time, however, Model F also showed acceptable fit on all three indexes due to a slight improvement on SRMR in the validation set (see the top of Table 3 and Supplementary Figure S8). In addition to showing the best fit in both data partitions, Model G also made theoretical sense and was selected as our final correlated factors model. In order to fairly compare this model with the bifactor and trifactor models described in the next subsection, we also re-fit it in the validation set using mirt instead of lavaan (which uses a slightly different parameterization). This version of the model had good fit according to CFI and acceptable fit according to RMSEA, but was a little high according to SRMR (see the bottom of Table 3 and Supplementary Figure S9).

Table 3.

Performance of Factor Analysis Models in the Validation Set

Confirmatory (WLSMV)	df	Chi2	RMSEA	CFI	SRMR	AIC	BIC
E: Simple Structure	881	4792.6 *	.092	.793	.142	—	—
F: Trimmed Items	474	1864.2 *	.075 a	.918 a	.098 a	—	—
G: Cross-loadings	472	1572.0 *	.067 a	.935 a	.088 a	—	—
Confirmatory (MHRM)	df	M2	RMSEA	CFI	SRMR	AIC	BIC
G: Cross-loadings	437	1370.4 *	.064 a	.952 b	.133	27678.2	28205.9
H: Bifactor Model	421	1864.9 *	.081	.926 a	.165	28041.5	28637.3
I: Trifactor Model	388	1121.8 *	.060 b	.963 b	.136	27688.6	28424.9

Notes. In each table section, the best value per performance metric is bolded. WLSMV = Weighted least squares means and variance adjusted (lavaan), MHRM = Metropolis-Hastings Robbins-Monro (mirt), df = Degrees of freedom, Chi2 = Scaled chi-squared, RMSEA = Scaled root mean squared error of approximation, CFI = Scaled comparative fit index, SRMR = Standardized root mean square residual, AIC = Akaike information criterion, BIC = Bayesian information criterion, M2 = M₂ statistic

^*p < .001

^aAcceptable fit

^bGood fit.

Bifactor and Trifactor Models

The fit of our bifactor model (Model H) in the validation set was acceptable according to CFI but was a little high on RMSEA and SRMR; surprisingly, this model was worse than our final correlated factors model on all included fit indices (see the bottom of Table 3). We inspected the factor loadings of the bifactor model to assess the coherence of the general and specific factors (full results are presented in Supplementary Figure S9). Loadings on the general factor were strongly³ positive for items from the disorganization (M = 0.66), hostility (M = 0.60), and mania (M = 0.53) factors; moderately positive from items from the positive symptoms factor (M = 0.44); and weakly positive or negative for items from the negative symptoms (M = 0.07), anxiety (M = 0.05), and depression (M = −0.13) factors. All items’ loadings on their corresponding specific factor were positive but tended to be smaller in magnitude compared to their corresponding loadings in the final correlated factors model. See Supplementary Figure S10 for bifactor model fit indices.

To fit our trifactor model (Model I), we added the following parameters to our bifactor model: loadings on a new non-affective domain factor for all items in the disorganization, negative symptoms, and positive symptoms specific factors, and loadings on a new affective domain factor for all items in the anxiety, depression, hostility, and mania specific factors. The trifactor model had acceptable fit in the validation set according to the CFI and the RMSEA, outperforming our bifactor model across all fit indices. However, while the trifactor model had better RMSEA and CFI scores compared to our final correlated factors model, its SRMR, AIC, and BIC were worse (Table 3).

We next inspected the factor loadings of the trifactor model to assess the coherence of its various factors (full details in Supplementary Figure S9). Loadings on the general factor were strongly positive for items from the mania (M = 0.61), disorganization (M = 0.59), and hostility (M = 0.57) factors; moderately positive for items from the positive symptoms factor (M = 0.35); weakly negative for the items from the anxiety (M = −0.08) and negative symptoms (M = −0.13) factors; and moderately negative for items from the depression factor (M = −0.33). Loadings on the affective domain factor were strongly positive for items from the anxiety (M = 0.57) factor, moderately positive for items from the depression (M = 0.40) and hostility (M = 0.31) factors, and weakly positive for items from the mania factor (M = 0.14). Finally, loadings on the non-affective domain factor were strongly positive for items from the negative symptoms factor (M = 0.59) and moderately positive for items from the positive symptoms (M = 0.49) and disorganization (M = 0.25) factors. Once again, all items’ loadings on their corresponding specific factor were positive but tended to be smaller in magnitude compared to their corresponding loadings in the final correlated factors model.

Model Selection

With similar performance between the final correlated factors model (Model G) and its trifactor version (Model I), we selected the former as our “favored” model due to its greater parsimony and interpretability. Following the model evaluation and identification of our favored model, we combined the discovery and validation sets and re-fit our favored model to the total sample using lavaan with the WLSMV estimator (Chi2 = 3019.4, RMSEA = .072, CFI = .926, SRMR = .085). The structure and loadings from this analysis are depicted in Figure 2 and the factor correlations are depicted in Figure 3. The network plot (Figure 3b) shows two broad clusters of factors with disorganization, hostility, mania, and positive symptoms on one side and anxiety, depression, and negative symptoms on the other; mania was unique in this structure in that, unlike the others in its cluster, it had strong negative correlations with depression and negative symptoms. Item-level network plot is shown in Supplementary Figure S15. Furthermore, the factor score distributions based on our preferred model showed varying patterns of symptom severity across the factors for participants in each DSM category, as expected (see Supplementary Figure 11).

Figure 2. Path Diagram Showing the Standardized Loadings of the Favored Model in the Total Sample.

Note. Rectangles depict observed items, ovals depict latent factors, single-headed arrows depict factor loadings, and connected arrows depict factor correlations. Item prefixes denote sources (P = PANSS Positive, N = PANSS Negative, G = PANSS General, M = MADRS, and Y = YMRS) and numbers correspond to the original instruments.

Figure 3. Inter-correlations Among the Favored Model’s Factors in the Total Sample Depicted as a (a) Heatmap Matrix and (b) Network Plot.

Note. The proximities of factors in the network plot are determined by multidimensional scaling. All correlations were significant (p < .05) except anxiety-mania, depression-positive symptoms, and hostility-negative symptoms.

Measurement Invariance

We then tested for the measurement invariance of our favored model across groups of participants with affective and non-affective psychosis (in the total sample) to determine the generalizability of factors across these clinical groups. Results are presented in Table 4 and full details are provided in Supplementary Figure S12. The configural model (same structure) had acceptable fit according to all three fit indexes. Constraining the factor loadings to equality between the two groups resulted in a significant chi-squared difference test but had negligible impact on the fit indexes. Similarly, also constraining the intercepts to equality resulted in a significant chi-squared difference test but negligible impact on the fit indexes. However, constraining the latent means to equality between the groups resulted in a significant chi-squared difference test and had a meaningful impact on both the CFI and RMSEA fit indexes (i.e., ΔCFI < −0.010 and ΔRMSEA > 0.015). In summary, the results suggested that these two groups had invariant structure, loadings, and intercepts but differed on their mean scores on one or more of the factors (which is not unexpected and does not preclude comparison). Our model can thus be applied to both affective and non-affective psychoses and yield meaningful comparisons.

Table 4.

Results of Measurement Invariance Testing for the Favored Model (Model G) in the Total Dataset

Constrained Model	df	Chi2	RMSEA	CFI	SRMR
Same Structure	944	3043.1 **	.065 a	.934 a	.088 a
Same Loadings	972	3096.8 **	.065 a	.934 a	.092 a
Same Intercepts	1006	3303.4 **	.066 a	.928 a	.089 a
Same Means	1013	4755.9 **	.084	.883	.089 a
Model Contrast	Δdf	ΔChi2	ΔRMSEA	ΔCFI	ΔSRMR
Loadings − Structure	28	143.2 **	−.001	−.001	.004
Intercepts − Loadings	34	61.9 *	.001	−.005	−.003
Means − Intercepts	7	519.2 **	.018 †	−.045 †	.000

Note. The groups being compared are affective and non-affective psychosis, df = degrees of freedom, Chi2 = Scaled chi-squared, RMSEA = Scaled root mean square error of approximation, CFI = Scaled comparative fit index, SRMR = Standardized root mean square residual, Δ = Change

^*p < .01

^**p < .001

^aAcceptable fit

^†Noninvariance.

External Validation

Finally, we tested the associations of our favored model’s factors with external criterion variables in the total sample (Table 5 and Supplementary Figure S13). The anxiety factor was positively correlated with the BAI and BDI and negatively correlated with the MCAS; in support of our “anxiety” label, its strongest correlation was with the BAI. Similarly, the depression factor was positively correlated with the BAI and BDI and negatively correlated with the MCAS; in support of our “depression” label, its strongest correlation was with the BDI. The disorganization factor was negatively correlated with the BDI, MCAS, and NAART FSIQ estimate. The hostility factor was negatively correlated with the MCAS. The mania factor was negatively correlated with the BDI. The negative symptoms factor was positively correlated with the BDI and negatively correlated with the MCAS. Finally, the positive symptoms factor was negatively correlated with the MCAS.

Table 5.

Correlations Between the Favored Model’s Factors and External Criteria in the Total Dataset

Factor	External Criterion Variables
	BAI	BDI	MCAS	NAART
Anxiety	0.53 **	0.47 **	−0.26 **	0.09
Depression	0.36 **	0.74 **	−0.17 **	0.02
Disorganization	0.03	−0.15 *	−0.50 **	−0.12 *
Hostility	0.03	−0.06	−0.39 **	−0.07
Mania	0.08	−0.24 **	0.02	−0.06
Negative Symptoms	0.03	0.25 **	−0.51 **	0.06
Positive Symptoms	0.04	0.00	−0.48 **	0.01

Notes. The favored model has seven correlated factors and two cross-loading items. Results in each sub-dataset are also reported in the Supplementary Figure S13 . BAI: Beck Anxiety Inventory; BDI: Beck Depression Inventory; MCAS: Multnomah Community Ability Scale; NAART: North American Adult Reading Test Premorbid Full Scale IQ Estimate;

^*p < .01

^**p < .001.

Discussion

Our analyses revealed that a correlated factors model with seven symptom dimensions (and a handful of cross-loadings) best characterized the clinical symptom ratings collected from a transdiagnostic sample of patients with primary and affective psychoses. This model was designated our “favored model,” and we labeled its seven factors (in alphabetical order) anxiety, depression, disorganization, hostility, mania, negative symptoms, and positive symptoms.

The inter-factor correlations in our favored model aligned with our expectations. Positive symptoms showed positive associations with all other factors, although the strength of these associations varied. A pronounced correlation was observed among positive symptoms, mania, disorganization, and hostility. This pattern can be attributed to the high proportion of hospitalized patients diagnosed with bipolar disorder experiencing psychosis in our cohort, suggesting that these symptom dimensions frequently co-occur during severe manic episodes. Moreover, the non-affective dimensions—namely positive symptoms, negative symptoms, and disorganization—were robustly associated with each other as well, highlighting the comorbid set of symptoms commonly observed in schizophrenia-related disorders and suggesting that the factors are not fully independent. Conversely, while depression was strongly and positively correlated with both anxiety and negative symptoms, it was negatively correlated with mania and hostility and was not correlated with disorganization. Although disorganized symptoms can occur in depressive psychosis, the relatively small proportion of patients with major depressive disorder recruited in this study may be a limiting factor for establishing this relationship.

Although negative symptoms and depression are considered to be distinct psychological constructs, the robust correlation (r = .54) between these factors may be attributable to the overlap in behavioral symptomatology or shared underlying determinants (Krynicki et al., 2018). In accordance with the theoretical frameworks, the main attributes defining the negative symptoms are blunted affect, alogia, and social/emotional withdrawal. In contrast, the main features defining depression are items related to low mood and energy, along with suicidal or pessimistic thoughts. Considering the mutual presence of psychomotor retardation in both dimensions (Krynicki et al., 2018), the item PANSS motor retardation was cross-loaded onto both factors.

The language item on the YMRS is designed to measure the severity of thought disorder by examining speech patterns and content. Patients presenting incoherent speech patterns, such as tangentiality and flight of ideas, which often co-occur with pressured speech during mania, are likely to score high on this item. Likewise, individuals with schizophrenia-related disorganization presenting loosely connected ideas and disruption of thought patterns leading to communication difficulties are likely to score high on this item. To account for this overlapping symptomatology, we cross-loaded this item onto both mania and disorganization factors.

Given that mania is often associated with symptoms such as elevated mood and increased activity, while depression and negative symptoms typically present with anhedonia and reduced activity levels, it was expected that the mania factor would show a strong inverse correlation with depression (r = −0.46) and negative symptoms (r = −0.56) in these between-person analyses. This finding underscores the well-observed clinical phenomenon that manic states and broadly depressed states typically manifest in an inversely related or often cyclical pattern. However, the relationships between these factors are complex and characterized by within-person variability. It is not uncommon for individuals experiencing mania to concurrently endorse depressive and anhedonic symptoms, a phenomenon known as a “mixed state.” The complex interplay of these factors highlights that the heterogeneity cannot be readily parsed into distinct categories, emphasizing the critical need for multidimensional and longitudinal approaches to fully capture the diversity in clinical profiles between and within individuals.

Comparisons to Previous Dimensional Models of Schizophrenia

Most previous research on schizophrenia, using a range of clinical instruments and analytic approaches, has consistently identified a dimensional structure of symptoms with four or five dimensions. These usually encompass positive symptoms, negative symptoms, disorganization, and affective symptoms, with the affective dimensions characterized by manic and/or depressive symptoms (Potuzak et al., 2012).

Our correlated factors model echoes previous PANSS-based factor structures in that it includes factors very similar to those from the pentagonal models (i.e., positive symptoms, negative symptoms, disorganization, mania, and depression). Yet, it offers a more nuanced subdivision of affective symptoms with the addition of hostility and anxiety factors. Notably, our model also provides a more comprehensive (and arguably construct-valid) representation of the mania factor, thanks to the added YMRS items. In previous PANSS-based models, the mania/excitement factor inadequately represented core manic symptoms (e.g., elevated mood and increased activity) and more closely captured irritability/hostility (Anderson et al., 2018; Reininghaus et al., 2013; Wallwork et al., 2012a). Our model is able to separate these phenomena into distinct constructs that show different correlations with external criterion variables and therefore have different clinical implications.

Similarly, previous PANSS-based factor models often merged symptoms of depressed mood and anxiety into a single “depression” or “emotional distress” factor. However, incorporating MADRS items enabled our model to separate depression and anxiety into distinct constructs with similar but distinguishable correlations with external criterion variables.

Dimensions capturing anxiety and hostility have been less frequently reported in psychosis research, but they do appear in some studies (e.g., Potuzak et al., 2012). Their sporadic reporting may be attributed to the design of these clinical instruments, which is primarily geared toward assessing the severity of non-affective symptoms. For instance, the Clinician-Rated Dimensions of Psychosis Severity Scale only captures mania and depression for the affective constructs (Barch et al., 2013). In contrast, the Brief Psychiatric Rating Scale (BPRS), which is not exclusively for evaluating psychosis, encompasses a broader spectrum of symptoms, including depression, mania, anxiety and hostility; thus offering a more comprehensive coverage across affective dimensions (Overall & Gorham, 1962). While anxiety and hostility may not be considered core symptoms of psychosis, their high comorbidity with depression and mania (respectively) is well-recognized. Consequently, while collapsing these comorbid symptoms together may offer parsimony, it is valuable to know that they can be further delineated into theoretically coherent constructs for enhanced granularity.

Comparisons to Broader Transdiagnostic Models

Our results have both similarities to and differences from those expected by the HiTOP framework. Our positive symptoms and disorganization factors would fit easily within HiTOP’s psychoticism spectrum and our negative symptoms factor would fit easily within HiTOP’s detachment spectrum. Our anxiety and depression factors are also similar to HiTOP’s fear and distress subfactors (of the internalizing spectrum), respectively. The two-cluster structure of factor correlations in our model (depicted in Figure 3b) also mirrors HiTOP’s division of the psychoticism and detachment spectra, and the positive correlations between our negative symptoms factor and our depression and anxiety factors are also not too surprising, given that moderate positive correlations between the detachment and internalizing spectra are commonly found in HiTOP research (Ringwald et al., 2023).

Most interesting were our mania and hostility factors, which have less certain positions within the HiTOP framework. Our results are more consistent with the conceptualization of mania as a subfactor of the psychoticism spectrum, as our mania factor was highly correlated with our positive symptoms and disorganization factors and negatively correlated with our depression factor. It is important to note that most HiTOP research is based on self-reported symptoms whereas our results are based on clinician ratings. Additionally, our data is cross-sectional and thus our results likely represent a mix of between-person (trait) variability and within-person (state) variability. Some of our findings, such as the negative correlations between our mania factor and our depression and negative symptoms factors, may be a result of participants being interviewed during acute symptom elevations (e.g., mood episodes). Future research is needed using longitudinal samples that can decompose trait and state effects.

Our hostility factor was highly correlated with our mania and disorganization factors, suggesting that it may be capturing uncooperativeness related to the impulsiveness and emotional lability of mania combined with the distractibility of disorganization. Notably, this factor contrasts with the callous and aggressive hostility seen in HiTOP’s externalizing super spectrum and is not currently well represented by any existing maladaptive traits or symptom components in the psychoticism spectrum; it is worth exploring in future HiTOP revisions.

The HiTOP framework also reveals several maladaptive traits and symptom components that seem to be underrepresented by the popular clinical ratings scales used in this study. Whereas delusions, hallucinations, and disorganization are relatively well captured by the PANSS, MADRS, and YMRS items, dissociation, fantasy proneness, and eccentricity are not well represented. Similarly, whereas anhedonia, inexpressivity, intimacy avoidance, and social withdrawal are relatively well captured, avolition and romantic disinterest are not. Additionally, although we estimated a factor for anxiety, its construct coverage is lacking and more comprehensive assessment of anxiety in this population seems warranted.

Measurement Invariance Test Across Psychoses

Our favored model held structural consistency across both affective and non-affective psychoses, with its loadings and intercepts showing measurement invariance across these groups. Significant group differences emerged only in the latent mean scores (i.e., the extent to which each transdiagnostic symptom factor was elevated), but this is expected as several of the factors quantify affective symptoms, and the groups are defined in terms of the presence or absence of such symptoms. In sum, we aimed to develop a multidimensional symptom model that is well-balanced in characterizing both psychotic and mood-related constructs that are commonly observed across psychotic disorders. Our results suggest that the constructs represented by our factor model can be interpreted and compared across affective and non-affective psychoses.

Associations with External Criterion Variables

Using the external variables available in the dataset, we were able to explore and begin validating the nomological network of our transdiagnostic factors (Cronbach & Meehl, 1955). We found strong support for the labeling of our depression and anxiety factors in that the former most strongly correlated with BDI and the latter most strongly correlated with BAI. We also observed significant negative correlations between six of the seven factors with the MCAS, a measure of community functioning. These findings support the clinical relevance of our factors, hinting at their real-world functional consequences. The observed correlation between MCAS and hostility, but not mania, can be attributed to the nature of these symptoms. While hostility, which often co-occurs with mania, affects critical components of community functioning, such as interpersonal interactions and social behavior, mania can manifest in a variety of ways which may not immediately disrupt functioning.

Previous research by Wallwork et al., 2012 reported that general cognitive ability (including FSIQ and ‘g’ from the Wechsler Adult Intelligence Scale) was most strongly correlated with the disorganization factor from the pentagonal model developed with a schizophrenia cohort. In our transdiagnostic cohort, we observed a similar pattern between FSIQ (from the NAART) and our disorganization factor. Although our estimate of FSIQ is based on a single word-reading assessment, as opposed to a more extensive evaluation of verbal and non-verbal cognitive abilities, the convergence of our results to previous research support that the disorganization factor is associated with cognitive impairments (O’Leary et al., 2000).

The Apparent Lack of a General Factor

Given that affective and non-affective psychoses are often grouped together under the common banner of “psychosis” (at least conceptually), we followed the example of recent work (Anderson et al., 2018; Reininghaus et al., 2013, 2019)in exploring multifactor models. These models estimate a general factor that accounts for the variance shared among all items, which might represent an overall psychosis dimension (in this specific context) or a general liability for all forms of psychopathology more broadly (sometimes called the ‘p’ factor; (Caspi et al., 2013).

In our data, we did not find evidence of unidimensionality with a strong general factor. Rather, there was evidence of strong multidimensionality; loadings on the general factor were variable across items both within and between group factors. This observation also fits what is known about the biological mechanisms underlying psychotic disorders: They are numerous and interacting (Cohen, 2016; Cohen & Öngür, 2023). Consistent with prior bifactor models of psychosis, the items with the largest loadings on the general factor were related to delusions and disordered thought processes. Indicators of hostile behavior and poor attention were also highly loaded on the general factor, which may be capturing distress involving symptoms of delusions and disorganized thoughts. Factor loadings of items related to mania, anxiety, and negative symptoms loaded on the general factor at moderate levels, but loadings related to depressive symptoms were collectively small. The weak loadings of depressive items on the general factor may be due to features of our sample, e.g., high proportion of inpatient population and a relative overrepresentation of patients with affective psychosis assessed during manic rather than depressive episodes. Statistically, the two broad clusters of factors observed in Figure 3b (and in particular the negative correlations of mania with depression and negative symptoms) made it challenging for a general factor to capture a large proportion of the overall variance. Overall, there did not seem to be evidence of a unifying general factor underlying the affective and non-affective psychotic symptoms in this sample, which may suggest that these are being driven by correlated but ultimately distinct processes (such that bifactor and trifactor models are not appropriate). Furthermore, the distinction between affective and non-affective symptoms may be more conceptual than empirical, as our data seems to support a different grouping of factors. Rather than affective (anxiety, depression, hostility, and mania) and non-affective (disorganization, negative symptoms, and positive symptoms) clusters, our data suggest a different grouping with anxiety, depression, and negative symptoms on one side and disorganization, hostility, mania, and positive symptoms on the other. It would be interesting to see whether this pattern can be replicated in other samples and populations in future work.

It is important to note that our findings in this study do not necessarily contraindicate the existence of a general factor in psychopathology overall (e.g., the ‘p’ factor). Although this is a contentious topic (e.g., Bork et al., 2017), many epidemiological, genetic, and endophenotypic studies (e.g., Baker et al., 2019; Bora et al., 2010; Huang et al., 2010; Lichtenstein et al., 2009; Purcell et al., 2009) seem to support a common biological basis underlying different forms of psychopathology. Due to our study focusing on mostly hospitalized patients at the height of distress, the lack of a general factor in our data should be interpreted as supporting the multidimensional nature of psychosis specifically rather than refuting the existence of a general factor of symptoms or liability more broadly.

Limitations and Future Directions

Our results should be interpreted with several limitations in mind. First, the clinical instruments we analyzed did not encompass all symptom dimensions typically observed in psychotic disorders. For example, cognitive impairments, which are core deficits in these disorders (Barch & Sheffield, 2014; Lewandowski et al., 2014), may not have been adequately represented. Likewise, utilizing clinical instruments designed for in-depth analysis of each symptom domains may reveal additional subtypes within those dimensions. For instance, previous research employing clinical ratings from the Scale for the Assessment of Negative Symptoms or Clinical Assessment Interview for Negative Symptoms has identified sub-factors, such as experiential and expressive deficits, among others, offering a more precise characterization of the latent structure of negative symptoms (Blanchard & Cohen, 2006; Strauss et al., 2018). Second, our model was developed largely using an inpatient sample and therefore its results may not generalize well to individuals with less severe psychotic illnesses. Epidemiological studies indicate a continuum of psychotic experiences ranging from healthy individuals to psychiatric patients (Os & Reininghaus, 2016). Given that our model primarily reflects the more severe end of the psychotic spectrum, future research should aim to incorporate a wider range of severities for better characterization of the entire spectrum. Third, our cross-sectional data mostly collected from hospitalized patients primarily reflects acutely heightened clinical states and cannot capture the within-person fluctuations in symptoms that are known to occur over time. Future studies should consider longitudinal symptom assessments to better understand the dynamic nature of symptom dimensions and networks over time. Indeed, course is another key factor or component characterizing and differentiating psychotic disorders.

Broadly, these results have important implications for refining the representation of psychosis (and its relationship with affective symptoms) in broader transdiagnostic and dimensional frameworks, such as the thought disorder spectrum in HiTOP and the psychoticism domain in Alternative Model for Personality Disorders (American Psychiatric Association, 2013). Furthermore, our study reveals several areas in which the clinical assessment of psychosis can be further improved, such as the more detailed measurement of anxiety, eccentricity, submissiveness, and emotional lability.

Conclusion

Establishing the phenomenological structure of psychosis through a transdiagnostic and dimensional framework is critical for advancing scientific understanding of the etiology, pathophysiology, and treatment of this debilitating form of mental illness. While the factors identified here are similar to those observed in past studies, the current study extends previous work on this important topic by rigorously examining the structure of a broader range of symptoms in a large transdiagnostic sample including various forms of affective and non-affective psychosis. The correlated factors model was replicated in unseen data, generalized across affective and non-affective psychoses, and had a fit that was either comparable to or better than that of competing bifactor and trifactor models while being far more theoretically coherent and concordant with clinical assessments of patients. Taken together, these results suggest that the symptoms of affective and non-affective psychosis are multidimensional and cannot be easily reduced to a single underlying “psychosis” general factor.

The findings from our study have significant implications for refining the accuracy of diagnostic models currently used in research and treatment. Our results demonstrate that various symptom dimensions are not confined to specific diagnoses, and using dimensional factors could offer a more comprehensive characterization of clinical expression at the individual level. The seven factors identified by the model correspond to well-established symptom constructs for which treatment options are already available. Further research is essential to assess the feasibility and utility of incorporating these dimensional measures of symptom severity into clinical decision-making. This approach will help streamline the process of monitoring illness progression and treatment responses, thereby enhancing personalized care.