Longitudinal Cognitive Changes in Young Individuals at Ultrahigh Risk for Psychosis

Key Points Question Do baseline and longitudinal cognitive architecture discriminate healthy controls from subgroups of young individuals at risk for psychosis? Findings This multiple-group design study involving 384 healthy controls and 173 individuals at ultrahigh risk for psychosis found that baseline cognitive architecture differentiated healthy controls from converters and nonremitters. Remitters were found to recover their cognitive deficits over time, but nonremitters did not. Meaning Cognitive deficits appear to identify the individuals most likely to develop psychosis and appear to reflect an underlying deterioration of a person’s clinical condition over time.


eAppendix 1. Statistical Approaches for Testing Principal Component Loadings: Data Analysis to Test Cognitive Structure Differences Across Groups Prospectively
Principal Component Analysis (PCA) has been used as a technique for data reduction and to identify differences in data structure. It is proposed that PCA and its loadings could potentially be utilized as a means of assessing differences in data structure for cognition between groups of subjects and within subjects over time.
There have been various methods to compare PCA loadings e.g. Tucker Congruence Coefficient (TCC; Lorenzo-Seva & Ten Berge, 2006) and Correlation Coefficient Comparison algorithms 2 .
The TCC had been designed to measure equivalence of latent trait via EFA (Tucker, 1951). However, there are limitations to the methodology 4,5 . Some of the issues that were highlighted with the use of the TCC were i) the lack of hypothesis testing -though Lorenzo-seva and Ten Berge (2006) suggested that R TCC = .85-.95 indicates congruent factor structure, it is notable that these were obtained from consensus ratings and not necessary statistical in nature 4 ii) the TCC was developed mainly for inter sample comparisons, and hence the notion of repeated testing within a sample would tend to inflate TCC values iii) TCC is a similarity index hence if the hypothesis is one that suggest that factors are differential the TCC would not have been an optimal test to show difference iv) factor similarity is not synonymous with factor invariance. To highlight this point, two 10item vector s1 = {.872, .736, .622, .502, .213, .123, .021, .014, .003, .002} and s2 = {.523, .441, .373, .301, .128, .074, .013, .008, .001, .001} is likely to give a very high congruent rating close to if not beyond .90. The reason being the loadings in vector s2 is in fact a function of s1 i.e. f(s2) = 0.6(s1) with a coefficient of a constant 0.6. In this circumstance, the factor loadings might be invariant, but not similar.
The second approach of involve using various correlation estimators 6 . While the approach is compelling, and allows between and within subject comparisons, considering sample sizes and repeated sampling, the approach involves singular coefficient comparisons. The challenge with performing singular comparisons results in massive multiple testing, reducing statistical power and introducing noise to the interpretation of coefficient differences. The challenge is escalated when factor loadings of large number of overlapping items from a neuropsychological test battery is compared.

eAppendix 3. Data Analysis Workflow: Discussion of Data-Analytic Strategies Carried Out for the Current Report
The following data analysis workflow chart details the data analysis workflow that was carried out as part of the investigation reported in the current manuscript. Prior to data analysis basic preprocessing of the cognitive data was undertaken. Demographical factors such as age and gender were adjusted for using baseline age and gender via the following linear regression model: * * Where, Test(n) represents the vector of cognitive tests administered to all participants in the study. Linear and non-linear effects of age at baseline were adjusted for to ensure latent differences that are brought about by Age at baseline did not result in extraneous effects during inter-group comparisons downstream. zResiduals represents the adjusted standardized scores after adjustment for baseline demographic factors. Special note should be taken that only baseline age and gender are used for adjustment of cognitive performance for each of the time points. The rationale for doing so is two fold-First, neurodevelopmental trajectories related to cognition may be embedded within the trajectories associated with clinical outcomes. Hence, we adopted this strategy of demographic adjustment rather than using Age within subsequent longitudinal modelling, of which adjusting away the Age*Time effect could remove potentially interesting differential maturational effects across clinical groups. Second, even though there does not appear to be significant differences in terms of age at recruitment of either healthy controls or UHR individuals, one cannot assume that cognition and age correlations do not exist. Hence, the approach that we employed is most likely to balance methodological rigor for the analysis of cognitive data, but at the same time giving enough room for longitudinal effects to vary.
Ordinal logistic regression was conducted on (i) healthy controls, at-risk individuals and converters; and (ii) healthy controls, remitters, and non-remitters. These analyses aimed to establish baseline differences between groups. Each cognitive test was entered as a predictor to group membership in multiple univariate models. Test of parallel lines across all cognitive tests were not significant, which indicate that the method was interpretable for the purpose of the current report. At the preliminary stage of the analysis, a liberal approach allowed significant thresholds of p < 0.05 to select cognitive tests for subsequent analysis. This is however, already considered stricter than recommended p-values of between 0.1-0.15 7 . Ordinal Logistic Regression was implemented via the PLUM module in IBM SPSS 22.0 Prospective modeling of cognitive change was completed via linear mixed models. Time points were coded in 0.5 year increments with baseline as "0". Intercept and Time were modeled as random effects, using unstructured covariance structure. Time, Time2, Group, Group*Time, Group*Time2 were modeled as fixed effects. Test performance was modeled as dependent variables in separate models. Two sets of Linear Mixed Model analysis were conducted for (i) Healthy Controls, Remitters, and Non-Remitters, and (ii) Healthy Controls, Non-Converters, and Converters. The overall model was: A liberal p-value of .05 was also used to identify potential group by time interaction effects. Additionally, to investigate the possibility of differential maturational trajectories in our sample, a median plot was carried out for the Age variable at baseline, resulting in Age <= 21 and Age >= 22 groups based on the baseline demographics. The Linear Mixed Model analysis then repeated for both sets of analysis via the following model: In models (1), (2) and (3), Test(n) represents the vector of all test included in the cognitive assessment battery administered to each participant. Post-hoc tests were then conducted to evaluate potential maturational stage*Time effects and Age dependent trajectory in each of the clinical groups. Linear Mixed Modelling was implemented via the MIXED module in IBM SPSS 22.0.
Distributions of predicted scores from linear mixed models across cognitive tests were compared between baseline and 24-month follow-up for each group (healthy controls, remitters, non-remitters, nonconverters and converters) using the Stuart-Maxwell marginal homogeneity (MH) test implemented in IBM SPSS 22.0.
Principal components analysis of tests that were found significant (P < .05) in the initial ordinal regression analysis was carried out for healthy controls, remitters and non-remitters for both baseline (BL) and 24 month (FU) time points. Factor loading patterns were then compared using the Kolmogorov-Smirnov test via the ks.boot() module in R 3.4.1 8 , that allow bootstrapping of the input vectors. Default bootstrapping parameters of n = 1000 resampling option was selected.   Table 7 Test Factor Loading patterns Kolmogorov-Smirnov Table 2 Factor scores weighted for differential factor structure perturbations were computed based on the following model:

Data Analysis Workflow
.

* *
Where n is the vector of tests included in the PCA. And Subtest n represents the vector of actual task performance of vector n. L represents the group vector of healthy controls, remitters and non-remitters. There are ten cognitive domain scores that reported, Social Cognition, Attention, Verbal, General Cognitive Function (GCF) and Perception for follow-up and baseline. These factor scores were standardized using healthy controls as the reference group. Non-weighted factor scores were computed similarly, however for non-weighted scores factor loadings for items belonging to a particular domain were constrained to 1 while items not belonging to the domain were constrained to 0. Cognitive domain change scores were computed by subtracting follow-up scores with baseline scores. Two sets of change scores were also computed i) weighted scores, representing the change score weighted by factor structure changes over time ii) non-weighted scores Downstream comparisons of the factor scores were carried out using One-Way ANOVA with Bonferroni Correction for between subject differences and Repeated Measures ANOVA for 2 (BL vs FU) x Group (HC v UHRR v UHRNR) for Time based comparisons. Further examination on how changes in component scores between baseline and follow-up might account for social and occupational functioning outcomes. The SOFAS differential represented the social and occupational functioning range within 6 months of the assessment time point. The larger the range the less likely an individual was doing well. Multivariate repeated measures general linear model was carried out on the percentage SOFAS differential at baseline and follow up. Additional a time*group term was included to assess differential changes of functioning across healthy controls, remitters and non-remitters and additional cognitive component change* time elements were included to examine if changes in cognition might explain differential functional changes across groups. All cognitive component change scores were included in the multivariate model. Analyses were carried out using the GLM module in IBM SPSS 22.0.
Healthy Controls, Non-converters, and Converters BACS Verbal Memory