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


Consistent with prior work, the present study found that classes were best distinguished by separation in negative/general symptoms and classes that exhibited the greatest baseline negative symptoms and behavioral change ratings had the highest risk of transition to psychosis (ie, class 3).[9] This is consistent with the growing literature establishing an association between high baseline negative symptoms and subsequent conversion to psychosis.[10,35,47–56] Class 3 (negative-neurocognitive) was further characterized by significantly impaired neurocognition. Inclusion of neurocognition in the model may have elicited the emergence of class 3, a novel putative subgroup.

The CHR paradigm was recently conceptually revised into a clinical staging model comprised of subgroups associated with increasing clinical severity and risk of transition.[57] The first stage (CHR−) is characterized by moderate negative symptoms, neurocognitive symptoms, and minimal positive symptoms (none ≥ 3).[58] It is possible that class 3's (negative-neurocognitive) symptomatology is consistent with the CHR− stage and thus they represent a discrete subgroup on the prodromal illness trajectory.

Conversely, class 2 (paranoid-affective) was characterized by significantly higher suspiciousness and CDSS total near the cutoff associated with major depression.[45] Class 2 largely evidenced nonspecific distress, with an emphasis in affective symptoms and sleep disturbance compared to other classes. Class 2 was not clearly consistent with any subgroup in Carrión et al's[58] clinical staging model and instead may be at risk for a broad range of psychopathology (eg, affective disorders). Given that the inclusion criteria of this study was 1 follow-up visit (ie, 6 mo), it may be that CHR criteria are sensitive to emergent psychosis for some, but that timing was insufficient to capture emergence of nonpsychotic disorders, which take years to manifest past adolescence/early adulthood (eg, average age was 15.7–19.6 across classes).[59] This is consistent with clinical staging model theory, which posits that nonspecific distress crystallizes over time into discrete categorical syndromes. Identifying subgroups at this time may be difficult due to the ephemeral nature of distress and symptomatology through adulthood.[60]

Rate of Transition to Psychosis

Class 3 (negative-neurocognitive) had the highest conversion rate (29%) and was not characterized by significantly greater positive symptoms as would be expected based on the clinical staging model.[58] The rate of transition is higher in the present sample (29.3% in class 3) than the comparable class in the clinical staging model (5.9% in CHR−).[58]

Class 2 (paranoid-affective) was associated with the highest suspiciousness, greatest depressive symptoms, intact neurocognition, and lower conversion rate (14.9%). Given that clinical depression is both associated with and predictive of persistent paranoia,[60] it is possible that effective treatment of depression in class 2 may reduce severity of positive symptomatology and prevent subsequent transition to psychosis.

Further Characterizing Subgroups With Covariates

Class 3 (negative-neurocognitive) had significantly lower social cognitive performance consistent with the proposed conceptualization of class 3 as an early risk stage of developing schizophrenia. In contrast, classes 1 and 2 performed comparably to UCs on measures of ToM and facial EP according to norms from age-matched UCs. Results from a meta-analysis of social cognitive performance in CHR individuals found medium effect sizes for EP (d = 0.47) and ToM impairment (d = 0.44).[61] Thus, one would expect class 2 (paranoid-affective) to have EP and ToM deficits, given that CHR individuals comprised 74.5% of this class. Further, results comparing CHR and HSC individuals from this sample found no significant differences in EP or ToM performance.[13,14] Thus, it is possible that specific constellations of symptoms (ie, those associated with class 3) account for social cognitive deficits in heterogeneous CHR samples.

Regarding demographics, class 2 (paranoid-affective) was significantly older than class 3 (negative-neurocognitive). Longitudinal findings indicated negative symptom onset predates positive symptom onset[62] and that negative/disorganized symptoms predicted positive symptoms over time.[47] Further, CHR− individuals were the youngest subgroup in the clinical staging model.[58] Thus, it follows that the youngest group may be characterized by predominant negative symptoms.

Classes also had significant differences in clinic of origin. Each of the 3 clinics used standardized inclusion criteria, screening/assessment measures, and recruitment methods, and raters evidenced significant agreement in routine assessment reliability checks. Although such processes were standardized, site differences may be due to selective recruitment processes.

Class 3 (negative-neurocognitive) exhibited the greatest premorbid academic/social and baseline social/role dysfunction, with scores comparable to individuals with established schizophrenia.[29] Class 2 (paranoid-affective) evidenced functional deterioration over time, and was statistically comparable to class 3's dysfunction in late adolescent academic maladjustment score. Class 2 (paranoid-affective) had significant social/role impairment, but to a lesser degree and with later onset than class 3. Taken together, such findings are consistent with the view of class 3 as an early risk stage of developing schizophrenia subgroup.

Limitations and Strengths

As LVMM are influenced by subtle sample differences, the present model must be replicated to ensure validity of the present class structure. Sample size prohibited cross-validation, which would enhance confidence regarding taxon validity. However, the present model is complex with several indicator variables and parameters; thus, use of cross-validation procedures would likely generate results with increased error.[63,64] Further, the present model does not include other predictive indicators such as basic symptoms (ie, subtle, subjective disturbances in one's mental processes) and biological markers (eg, electrophysiological, imaging, metabolic, genetic markers).

The present class structure evidenced significant site differences. We elected not to include site as a covariate in the LPA model, because in the case of employing a single covariate, the log-linear model is identical whether site is treated as an active covariate or an additional indicator variable.[65–67] Given that there were no significant site differences in transition rate, we instead used site as an inactive descriptive covariate. Significant differences between indicators remained when controlling for site, indicating true variance in symptomatology drove the LPA.

Strengths of the present study include ecological validity in application of LPA to the combined sample. Our use of neurocognitive scores as indicators is novel and the first study to utilize such. The current study is further strengthened by inclusion of a range of covariates (functioning, social cognition) to characterize subgroups.


Overall, the results support a subgroup approach to research, assessment, and treatment of help-seeking individuals. Three classes emerged with adequate separation on a majority of indicator variables (SOPS, CDSS, neurocognition). Despite the well-established association between poor outcome, negative symptoms, and neurocognitive deficits, such symptom clusters are insufficiently targeted in CHR individuals. We join other researchers who have advocated for a transdiagnostic, heuristic approach to CHR individuals that has been emphasized in understanding the progression to psychotic and other mental illnesses.[68,69]