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



The sample consisted of 171 CHR participants (98 males, 73 females) with a mean age of 19.8 (SD = 4.5) and 100 HSC participants (56 males, 44 females) with a mean age of 19.4 (SD = 3.9) years. Data were collected as a part of National Institute of Mental Health (NIMH) funded, multisite study "Enhancing the Prospective Prediction of Psychosis" (PREDICT). Procedures are described in greater detail in prior publications (eg,[11–14]). PREDICT was conducted at the Universities of North Carolina at Chapel Hill (62 CHR, 24 HSC), Toronto (69 CHR, 45 HSC), and Yale (40 CHR, 31 HSC). All CHR participants met Criteria of Prodromal Syndromes (COPS) derived from the Structured Interview for Prodromal Syndromes (SIPS[15]). Twenty-nine CHR individuals converted to psychosis (17.0% within CHR; 10.7% within total sample).

The HSC group was comprised of individuals who responded to CHR recruitment, appeared to have prodromal symptoms at phone screen but upon administration of the SIPS did not meet COPS criteria. The HSC group contains the following subgroups: (1) family high risk, no deterioration in Global Assessment of Functioning (n = 16), (2) attenuated symptoms present for more than 1 year (n = 39), (3) current attenuated symptoms but due to another disorder (n = 2), (4) only negative symptoms (n = 4), and (5) attenuated symptoms not meeting severity or frequency criterion (n = 24). HSC individuals were included as a clinically relevant control group, as CHR and HSC individuals are more symptomatically similar to one another than non-psychiatric controls. Inclusion of such self-presenting, help-seeking individuals typically seen at CHR clinics provides greater better representation of clinical realism and diversity. Further, five HSC individuals converted to psychosis (5.0% within HSC; 1.8% within total sample).

Exclusion criteria included presence of an axis I psychotic disorder, age-scaled intelligence quotient (IQ) < 70, history of a clinically significant central nervous system disorder that may confound/contribute to CHR symptoms, or past/current use of antipsychotics.


SIPS and Scale for Assessment of Prodromal Symptoms (SOPS[15]) were used to assess criteria for prodromal syndrome, conversion, and severity of attenuated psychotic symptoms. Structured clinical interview for the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV[16]) was used to assess current/lifetime substance abuse/dependence.

Conversion to psychotic disorder is defined as at least 1 of 5 attenuated SOPS positive symptoms reaching a psychotic level of intensity (rated 6) for a frequency of ≥1 h/d for 4 d/wk in the past month. If the symptom meets intensity but not frequency criteria, it must seriously impact functioning (ie, severely disorganized or dangerous to self/others) to be considered conversion.[15]

Calgary Depression Scale for Schizophrenia (CDSS[17]) was used to measure depression and has been validated in CHR individuals.[18]

Neurocognitive Measures. Neurocognitive measures were selected based on demonstrated reliability, validity, absence of ceiling/floor effects in CHR population, ability to discriminate individuals with schizophrenia from UCs, and appropriateness for administration in individuals as young as 14. Verbal fluency was measured with category instances (CAT[19]), executive functioning with Wisconsin Card Sorting Test, 64-card computerized version (WCST[20]) and Trail Making Test B (TMT B[21]), speed of processing with TMT A,[21] verbal explicit memory with Rey Auditory Verbal Learning Test (RAVLT[22]), and attention with Continuous Performance Test-Identical Pairs (CPT-IP[23]). Neurocognitive tests, indices, ranges, and normative UC data are provided in Supplementary Table 2.

IQ was measured using the Wechsler Adult Intelligence Test or the Wechsler Intelligence Scale for Children-III, depending on the participants' age.[24,25]

Social Cognition. Theory of mind (ToM) was assessed with the "Reading the Mind in the Eyes" Task (Eyes Task[26]), emotion perception (EP) in faces with the Face Emotion Identification Task (FEIT[27]) and the Face Emotion Discrimination Task (FEDT[27]), and EP in voices with the Affective Prosody Task (AP[28]). All social cognitive tests, ranges, and normative data from UC groups are provided in Supplementary Table 2.

Functioning Measures. Premorbid functioning was assessed using the Premorbid Adjustment Scale (PAS[29]) using administration and scoring procedures outlined by van Mastrigt and Addington.[30] Adult PAS ratings were not included in the present analyses due to young age of the sample (44.6% <19 y). Social functioning was measured using Social Functioning Scale (SFS[31]) with the employment item removed (range: 0–213).[32,33] Role functioning was measured using the employment subscale of the Heinrichs-Carpenter Quality of Life Scales (QLS[34]) (range: 0–18).


PREDICT was a longitudinal study of predictors of conversion to psychosis. Study protocols and informed consent documents were reviewed and approved by institutional review boards of the 3 study sites. Formal consent procedures were conducted with participants. Clinical raters were experienced research clinicians who underwent a training program developed at Yale to identify prodromal syndromes with adequate reliability and demonstrated reliability throughout PREDICT.[35] Gold standard post-training agreements were excellent (κ = 0.90).

JA chaired weekly conference calls with all clinical raters to review inclusion criteria for all participants. Research assistants were trained in neurocognitive assessments by R.S.E.K. and social cognitive assessments by D.L.P.

Statistical Analyses. Data analyses were performed using Mplus version 7 with Mixture Add-On[36] and SPSS version 23.

Model Selection

Number of classes were not estimated a priori, but were ascertained from a combination of model fit statistics and interpretability. Model of best fit was determined from examinations of: (1) Akaike's Information Criteria (AIC[37]), Bayesian Information Criteria (BIC[38]), sample-size adjusted BIC (ssa BIC[39]) (lower values indicate the model of best fit), (2) Bootstrapped Likelihood Ratio Tests (BLRT[40]), (3) Mean estimated average posterior probabilities, and (4) Entropy indices.

An alternative interpretation of information criteria (eg, AIC, BIC, ssa BIC) and log likelihood values is to plot indices against the number of latent classes and examine for the "leveling off" point of the curve (eg, scree plot).[41] The model associated with a subsequent decrease in absolute value of slope may provide a model that balances model fit statistic improvement and parsimony.[41] Substantive interpretability and parsimony of models were considered in model selection.

Data Analytic Plan

Transition rates were computed as the percentage of converters within each class and were compared using χ2 tests of significance. Separation of LPA model indicators was assessed using univariate ANOVAs and effect sizes as measured by r2. Indicator profiles were generated depicting estimated sample means. ANOVAs, independent samples t tests, and chi-square tests of significance were conducted to compare classes on covariates. When appropriate, pairwise comparisons were conducted using Bonferroni correction for multiple comparisons. Z-square cell comparison tests with Bonferroni correction were used to probe significant omnibus chi-square tests and determine which groups significantly differed.[42,43]