Signal Detection (Psychology)

Validity and reliability analyses were conducted using SPSS version 22.0 [28 ], and factor analyses were conducted using SPSS AMOS version 23.0 [29 ]. Alpha was calculated for the total PCL-5 and its subscales to assess internal consistency. In the French sample, intraclass correlation coefficients were calculated using scores from time 1 and time 2 to determine test-retest reliability. Convergent validity was assessed via correlations between the PCL-5 and the IES-R, and between the PCL-5 subscales and their corresponding IES-R subscales. Using the Fisher r-to-z transformation we compared the magnitude of the correlation between the PCL-5 and the IES-R to that observed between the PCL-5 and the CES-D to assess divergent validity.
Signal-detection analyses were conducted using the DSM-5 diagnostic guidelines applied to the PCL-5 to dichotomize participants into ‘Probable PTSD’ and ‘Non-PTSD’ groups, as suggested by Weathers et al. [2 ,9 ]. Thus participants with scores 2 or above on at least one re-experiencing symptom, one avoidance symptom, two symptoms of negative alterations in cognition and mood, and two arousal symptoms were classified as having probable PTSD. Using the results of a previous study as a starting point [11 (link)], PCL-5 scores were examined to determine which best predicted the prevalence of probable PTSD as per this grouping. The score that yielded a prevalence proportion that most closely reached that determined by the DSM-5 guidelines (without exceeding it), and with the highest specificity, sensitivity and efficiency ratings, was selected.
Three structural models of PTSD were tested using confirmatory factor analysis (CFA). The first tested the DSM-5 four-factor model of PTSD, using the four PCL-5 subscales. The second tested the six-factor anhedonia model [14 (link)], and the third tested the seven-factor hybrid model of PTSD [15 (link)]. In each case, maximum likelihood estimation procedure was applied, and factor variance for each latent variable was set to 1. Because latent variables were theoretically expected to correlate and to ensure the models were properly identified, latent variables were allowed to correlate with one another. Goodness-of-fit indices were interpreted according to guidelines by Hu and Bentler [30 (link)], thus adequate model fit was determined based on cut-offs of ≥ .95 for the comparative fit index (CFI), ≤ .06 for the root mean square error of approximation (RMSEA) and ≤ .08 for the standardized root mean square (SRMR). In order to compare models, chi-square difference tests and the Akaike information criterion (AIC) were examined. Regarding the AIC, the lowest value of those produced by each model indicates better comparative fit. An analysis of measurement invariance was also performed in order to test the potential differences in fit between the English and French versions of the measure. Less than 2% of the PCL-5, IES-R and CES-D values were missing, thus a single imputation was performed.

The PVCA approach (http://www.niehs.nih.gov/research/resources/software/pvca/index.cfm) was used to estimate sources of variability and compare batch effects before and after adjustment in the VAS and SMRI microarray data sets. We first selected the top PCs, enough to explain a proportion of overall variation larger than a pre-defined threshold (60%–90%, 60% in this case), and retained the corresponding eigenvalues. Each factor was treated as random in a mixed linear model and matched to a PC, then weighted by that PC's corresponding eigenvector. After we standardized the variation attributable to each factor, we calculated the proportion of total variance each factor explained. Including residuals, four factors' variations were estimated in the VAS data and ten factors' variations were estimated in SMRI data.
Signal detection slopes were calculated using the spkTools[21] R package. The significances of differences between slopes were assessed with a test for homogeneity of slope[32] , which was done with the NCStats R package[33] . The evaluation of overall performance was performed using the ROCR[34] (link) R package from Bioconductor[35] (link).