Latent Profile Analysis and Conversion to Psychosis: Characterizing Subgroups to Enhance Risk Prediction

Kristin M. Healey; David L. Penn; Diana Perkins; Scott W. Woods; Richard S. E. Keefe; Jean Addington


Schizophr Bull. 2018;44(2):286-296. 

In This Article


Latent Profile Analysis

LPA Model Selection.Table 1 provides fit indices from the LPA. The AIC, BIC, and ssa BIC values decreased with each class addition and did not readily discriminate a model of best fit. BLRT value remained significant (P < .0001) with each class addition. Entropy values remained high for each class model (k = 2–5), ranging from 0.88 to 0.93. Fit indices and BLRT alone indicated the 5-class model. However, accepting the model associated with the lowest values does not prioritize model interpretability and parsimony.[44]

Supplementary Figure 1 provides scree plots of AIC, BIC, ssa BIC, and log likelihood value. Leveling off point of the curves occurred at 3 classes in each plot, indicating that significant improvements in model fit are not gained through further class additions. Of note, the BIC is considered to be the best of the presently available information criteria,[41] which showed clearest leveling off at 3 classes. The 3-class solution indicated high classification quality, adequate entropy score of 0.88, and mean posterior probabilities ranging from 93.9% to 95.6%. Supplementary Table 3 summarizes latent class membership based on estimated posterior probabilities. Indicators evidenced meaningful separation. The 4- and 5-class models were examined and evidenced poor separation across a majority of indicators and thus did not result in substantively meaningful or interpretable class structures.

Supplementary Figure 1.

Scree Plots. Y-axis represents information criteria value for each plot (e.g., AIC value). X-axis represents k classes. AIC, Akaike's Information Criteria; BIC, Bayesian Information Criteria; ssa BIC, sample size adjusted Bayesian Information Criteria.

Individuals were assigned to classes as indicated by highest posterior probability value as such: class 1 (mild cluster; n = 124), class 2 (paranoid-affective cluster; n = 106), and class 3 (negative-neurocognitive cluster, n = 41).

Classes and Risk Probability. The overall transition rate in the full combined sample of CHRs and HSCs at 2 years was 12.5%. Transition rate significantly differed across groups in the overall model (χ 2(2, N = 271) = 16.08, P < .001). Pairwise comparisons indicated that transition to psychosis was more likely in individuals in class 3 (negative-neurocognitive; transition rate 29.3%, n = 12 converters) than class 1 (mild; transition rate 5.6%, n = 7 converters) at the P < .05 level. There were no significant differences in pairwise comparisons between class 2 (paranoid-affective; transition rate 14.2%, n = 15 converters) and classes 1 or 3. Diagnoses at transition are provided in Supplementary Table 4.

Characteristics of the 3-Class Solution. Table 2 shows results from the LPA and ANOVAs. Figure 1 and Figure 2 show latent profile plots of estimated means. ANOVA results indicated that all indicators were influential in the clustering process, with the exception of SOPS grandiose ideas (P3) and bizarre thinking (D2).

Examinations of the SOPS latent profile plot and pairwise comparisons indicated that class 1 (mild) evidenced the lowest scores across SOPS and CDSS total. Class 1 largely evidenced SOPS estimated means of 1–2, which indicates mild/questionable presence and depression comparable to UC sample norms (normative mean: 2.6, SD: 2.7).[45]

Figure 1.

Latent profile plot of Scale of Prodromal Symptoms (SOPS) and Calgary Depression Scale for Schizophrenia (CDSS) total score.

Figure 2.

Latent profile plot of neurocognitive scores. CAT, category instances; RAVLT, Rey Auditory Verbal Learning Test; WCST PE, Wisconsin Cart Sorting Test Perseverative Errors; Trails A, Trail Making Test A; Trails B, Trail Making Test B; D'3, Continuous Performance Test-Identical Pairs (CPT-IP) D'3.

Class 2 (paranoid-affective) estimated means were significantly more severe for suspiciousness/persecutory ideas than classes 1 and 3. Class 2 evidenced significantly more severe ratings than class 1 on unusual thought content and perceptual abnormalities. Class 2 had significantly higher depression ratings (on SOPS dysphoric mood and CDSS total scores) and significant sleep disturbance compared to other classes. Class 2 had mild negative symptom ratings (≤2), with the exception of occupational functioning, which was near moderate (3).

Class 3 (negative-neurocognitive) membership was associated with the highest ratings (between 2–4) in a majority of negative symptoms, and to a lesser degree, disorganized symptoms. This was confirmed through pairwise comparisons. Class 3 evidenced comparable ratings to class 2 on avolition and decreased experience of emotions.

Regarding neurocognitive performance, classes 1 (mild) and 2 (paranoid-affective) performed comparably across indices. Class 3 (negative-neurocognitive) evidenced significant impairment compared to classes 1 and 2 across neurocognitive indices (P < .05). As neurocognitive test scores were not age corrected in the LPA model, comparisons among classes on neurocognitive indices were also run as ANCOVAs with age as a covariate. All overall models remained significant (P < .001) and pairwise comparisons using Bonferroni correction for multiple comparisons remained significant (P < .05), indicating that classes significantly differed on neurocognitive performance when accounting for age-related variance.

Characterizing the 3-Class Solution With Covariates. Table 3 provides results from ANOVAs and pairwise comparisons between classes regarding demographics and covariates.

Demographic Characteristics. There were significant differences in age and clinic location between classes. Individuals in class 3 (negative-neurocognitive) were significantly younger than class 2 (paranoid-affective). Individuals from Yale were more likely to be classified in class 3 and less likely to be classified in class 2. Conversely, individuals from UNC were more likely to be classified in class 2 and less likely to be classified in class 3.

Given site effects, comparisons among classes on indicators (SOPS, CDSS total score, neurocognitive indices) were conducted as ANCOVAs with site as a covariate. All results were unchanged, indicating that classes significantly differed on indicators when accounting for site-related variance. Classes showed no significant differences in sex or racial/ethnic composition.

Risk Group. CHR individuals were significantly more likely to be categorized in class 2 (paranoid-affective) than class 1 (mild). Conversely, HSC individuals were more likely to be categorized in class 1 than 2. Supplementary Table 5 provides symptom and functional descriptives of risk subgroups within each class.

Premorbid Functioning. Classes had significant overall group differences across PAS subscales. From childhood through early adolescence (age ≤ 15), individuals in class 3 (negative-neurocognitive) showed significant social and academic maladjustment scores compared to classes 1 and 2, whereas classes 1 and 2 had comparable impairment during this time. Regarding late adolescence (age 16–18) social maladjustment ratings, class 3 continued to perform at the most impaired level compared to classes 1 and 2. However, class 2 evidenced significant social maladjustment compared to class 1, suggesting that for class 2, poor functioning begins in late adolescence.

Social Functioning. Classes 2 and 3 had significant impairment on the SFS compared to class 1.

Role Functioning. Class 3 had significant impairment in QLS total score compared to classes 1 and 2. Class 2 evidenced significant impairment in QLS total score compared to class 1.

Social Cognition. Classes had significant overall models measuring group differences on the Eyes Task, FEIT, and AP. The overall model for FEDT approached significance (P = .053). Pairwise comparisons indicated that class 3 (negative-neurocognitive) had significant social cognitive deficits compared to classes 1 (mild) and 2 (paranoid-affective) across measures, indicating class 3 was impaired in ToM and EP.

As social cognitive performance tends to be associated with age and IQ,[46] comparisons among classes on social cognition were repeated as ANCOVAs with age as a covariate. Overall models and pairwise comparisons remained significant, indicating that classes evidenced significant differences in social cognitive performance when accounting for age-related variance.

Age-scaled IQ was added as a covariate and overall models for Eyes Task and AP Task remained significant (P < .05); however, FEDT was no longer significant. Pairwise comparisons for AP remained significant (P < .05). Eyes Task contrast between classes 2 and 3 was no longer significant. Thus, significant group differences in facial EP performance and ToM may be partially accounted for by neurocognitive ability, but not for AP.

Intelligence. Classes were compared across age-scaled IQ. Classes were significantly different, with impairment in class 3 (negative-neurocognitive) compared to classes 1 (mild) and 2 (paranoid-affective).