Clinical Features of Psychotic Disorders: Comparing Categorical and Dimensional Models

Subjects (N=934) were from McLean Hospital inpatient units specializing in psychotic disorders (16). Information on medications was collected from discharge records. The Partners HealthCare Institutional Review Board approved the study. All subjects provided informed consent.

Clinical interviews were performed by trained research staff who completed regular exercises to maintain consistency and reliability. Interviews included the Young Mania Rating Scale (YMRS) (17), Montgomery‐Asberg Depression Rating Scale (MADRS) (18), and the Positive and Negative Symptom Scale (PANSS) (19). DSM‐IV diagnoses were based on the Structured Clinical Interview for DSM‐IV‐TR (20).

The first stage of analysis explored the dimensional structure of the item ratings under the assumption that the items were indicators of an underlying set of latent continuous variables. Items on the YMRS, MADRS, and PANSS were submitted for exploratory factor analysis using maximum likelihood estimation and oblique geomin rotation. The primary analysis used the ratings as originally coded, treating them as continuous. Sensitivity analyses treated the items as ordinally rated and collapsed categories for items with a skewed rating distribution. Candidate solutions including different numbers of factors were evaluated based on scree plots of the factor eigenvalues, the root mean squared error of approximation (RMSEA), the standardized root mean squared residual value (SRMR), the Bayesian information criterion (BIC), interpretability of the solution, and robustness of the solution to choice of statistical method. After choosing a solution, items with loadings less than 0.40 on all factors and items with loadings greater than 0.40 on more than one factor were omitted. The resulting fit was submitted for confirmatory factor analysis to obtain factor‐based scores. Factor analysis was conducted using MPLUS (version 6.0) statistical software, which accommodates missing item‐level data under the missing‐at‐random assumption (21). To characterize differences in symptom presentation among DSM‐IV diagnoses and variability of symptom presentation within DSM‐IV diagnoses, we plotted the distribution of factor‐based scores by DSM‐IV diagnosis. Malhalanobis distance (22) was used as a summary metric for describing the distance of an individual's set of factor‐based scores from the mean for each diagnosis.

The second stage of analysis explored the categorical structure of the data using the factor‐based scores, agnostic to DSM categories. Multivariate normal mixture modeling, a method closely related to Malhalanobis distance, was used to identify clusters of patients with similar symptom presentations based on their factor scores. Fits of models with more constraints on the variance of scores within clusters were compared to fits of models with fewer constraints using the BIC. Fits of models with different numbers of clusters were compared using the BIC and the bootstrap likelihood ratio test. The approximate correct model probability (CmP) was used to quantify the certainty that our choice of model was correct under the assumption that the correct model was included among our candidate models (23, 24). A sensitivity analysis compared results obtained from multivariate normal mixture modeling to results obtained by categorizing factor‐based scores based on standard deviation units and submitting category indicators to latent‐class analysis (LCA). Fits of LCA models with different numbers of classes were compared using the BIC and the Vuong‐Lo‐Mendell‐Rubin likelihood ratio test. Multivariate normal mixture modeling was conducted using the mclust package (25) for R statistical software (version 3.5.0), and latent‐class models were fit using MPLUS (version 6).

In the third stage of analysis, logistic regression evaluated factor scores, DSM‐IV diagnoses, and cluster membership as predictors of psychotropic medications at discharge. The c‐statistic was used to compare the discrimination of models with alternate clinical predictors. Factor scores were evaluated both with and without categorical predictors in the model to assess if inclusion of both factors and categories improved discrimination over the use of either feature alone. All logistic regression models controlled for age and sex and were fit using SAS software (version 9.3).

A four‐factor solution was chosen among candidates including between three and ten factors. This solution was suggested as most favorable by the scree plot, was the most parsimonious solution with RMSEA and SRMR estimates below our preferred thresholds of 0.08 and 0.05, and had a solution that was clinically meaningful and robust to decisions made in the course of analysis. Factor loadings for items in the final exploratory factor analysis model are in Table S1. The four factors corresponded to clinical presentations that fit the descriptions: manic, depressive, negative symptoms, and positive symptoms. YMRS and MADRS items loaded almost exclusively on the manic and depressive factors, respectively. PANSS items loaded on all four factors. This solution had an RMSEA estimate of 0.07 and SRMR value of 0.04, values generally associated with adequate model fit (26, 27).

Three items were excluded prior to analysis: insight from YMRS, hostility from PANSS‐P, and inner tension from MADRS. The first two items were excluded because of similarity to items from other scales, reflected in high observed correlations, and the third because it was interpreted inconsistently by raters. Seven additional items were excluded in the course of arriving at a final solution: five because they did not meet the 0.40 threshold for loading on any factor (appearance from YMRS and somatic concern, tension, disorientation, and preoccupation from PANSS‐G) and two because of similar loadings on multiple factors (concentration difficulties from MADRS and conceptual disorganization from PANSS‐G). These cognitive symptoms are known features of all psychoses (28). Likelihood‐based fit statistics did not contribute to choice of solution because, as is common for large datasets, model fit consistently improved with inclusion of additional factors, even when solutions had no clear interpretation and only one or two items loaded on one or more factors. Results for the four‐factor solution were qualitatively similar when treating the items as ordinally rather than intervally scaled, when excluding more extreme item values, and when using least‐squares estimation rather than maximum‐likelihood estimation. The six‐factor solution, which included factors appearing to correspond to thought disturbance and hostility‐aggression, in addition to those identified using the four‐factor solution, was considered as an alternative solution because of potential clinical relevance. However, the six‐factor solution only minimally improved fit over the four‐factor solution, and treatment of the items as ordinally rated resulted in a six‐factor solution with a factor loading on sleep items replacing the thought disturbance factor.

Figure 1 shows density plots of factor‐based scores for the four most common DSM diagnoses in the sample. Overall, means on the factor‐based scores differ among diagnoses in expected ways. Patients with schizophrenia and schizoaffective disorder have the highest average factor scores for positive and negative symptoms. Patients with major depressive disorder are distinguished by higher scores on the depressive factor. Patients with bipolar disorder are distinguished by higher scores on the manic factor and lower scores on the negative symptom factor. However, the plots also show the substantial overlap in symptom presentation among patients categorized by DSM diagnoses for all four factors. That is, at the individual level, many symptom presentations are observed in multiple DSM diagnoses.

The overlap in symptom presentations is further reflected in scatterplots of factor‐based scores. The plots along the upper diagonal of Figure 2 show the substantial variability in scores within each diagnostic category and the substantial overlap in symptom profiles between diagnoses for each pair of scores. Differences across all four factor‐based scores by diagnosis are summarized in Figure 3, which displays the Malhalanobis distance, or the multivariate distance of each patient's set of factor scores from the observed means for different DSM diagnoses. As seen, all pairs of diagnoses had many patients with factor‐based scores more consistent with alternate DSM diagnoses than their own DSM diagnosis.

Based on BIC, the best‐fitting parsimonious categorical model allowed seven different distributions (or clusters) and allowed both the variance of factor‐based scores and the correlations between factor‐based scores to differ across clusters. The CmP of the selected model among all 36 candidate models (a minimum of one and a maximum of nine clusters for each of four possible covariance structures) was 1.00, highly favoring our selection. Patients were assigned to clusters for which they had the highest probability of membership. Average estimated probability of assigned membership was 87%, and 99% of patients had an estimated probability greater than 50%. Details of the clusters derived from our sample are provided in Table S2. In brief, the analysis identified seven clusters named according to their profiles of mean factor‐based scores, with capital letters indicating a mean factor‐based score in the high range (>0.75) and lowercase letters indicating scores in the medium range (≥−0.75 and ≤0.75). In order of highest prevalence in the sample, the clusters were Dnp, Mnp, mdnp, DNmp, Dn, m, and dn (M=manic, D=depressive, N=negative, and P=positive).

All clusters contained substantial numbers of patients with all three most common diagnoses: schizophrenias, schizoaffective disorders, and bipolar disorders (see Table S2 and Figure S1). Unlike the results of factor score analysis, results of cluster analysis demonstrated substantial sensitivity to statistical assumptions and modeling approach. Placing additional constraints on variance tended to produce results favoring different numbers of clusters with different mean profiles. For example, assumptions of equal variance of factor‐based scores across clusters and zero correlation between scores within clusters, which are equivalent to the assumptions underlying traditional profile analysis (29), resulted in the selection of a nine cluster solution (using BIC‐based criterion). In contrast, the latent‐class approach using cut‐points of the factor scores to predict class membership favored the choice of three classes, none of which could be characterized as combinations of the clusters.

Physicians prescribe medications based on their overall evaluation of their patients. All three groups of clinical predictors were significantly associated with the use of all four classes of psychotropic medications at discharge (Table 2). Models with only factor‐based scores had discrimination better than or comparable to models with only DSM categories or only clusters as predictors (Table 2). When both factor‐based scores and DSM categories were included, both were significant as predictors for antidepressants and mood stabilizers, but only factor‐based scores remained significant for antipsychotic use. When both factor‐based scores and cluster assignments were included, only factor‐based scores were significant predictors for antipsychotics, antidepressants, and mood stabilizers. Neither factor‐based scores, DSM diagnoses, nor cluster assignments were significant in models with both scores and categories as predictors of anxiolytic use, and models had fair to poor discrimination for this treatment.

Odds ratios for prescription of the four medications at discharge are provided in Table S3. For antipsychotics, the model with only factor‐based scores had good discrimination (c=0.72). Higher positive symptom scores were most strongly associated with use of antipsychotics (OR=1.4, 95% CI=1.2–1.6 for a one standard deviation increase in score). For antidepressant and mood stabilizer use, discrimination was excellent (c=0.81 antidepressants, c=0.82 mood stabilizers, models with both factor‐based scores, and DSM diagnoses). Antidepressant use was more likely for patients with lower scores for the manic factor (OR=0.4, 95% CI=0.3–0.5), higher scores for the depressive factor (OR=1.4, 95% CI=1.3–1.6), and a DSM diagnosis of major depressive disorder (OR vs. schizophrenia=2.0, 95% CI=0.9–4.5). Antidepressant use was less likely for patients with diagnoses of psychosis NOS and bipolar disorder (OR vs. schizophrenia=0.5, 95% CI=0.2–1.6 for psychosis NOS; 0.6, 95% CI=0.3–1.0 for bipolar disorder). Prescription of mood stabilizers was more likely for patients with higher scores for the manic factor (OR=1.5, 95% CI=1.2–1.9), and DSM diagnoses of bipolar disorder (OR vs. schizophrenia=9.7, 95% CI=5.8–16.4), and schizoaffective disorder (OR vs. schizophrenia=3.8, 95% CI=2.4–6.0).