The functional neurobiology of negative affective traits across regions, networks, signatures, and a machine learning multiverse

Figures

• 

  Figure 1.. Schematic illustration of the study design. (a) Theory-driven analyses for regions, networks, and signatures. The three ROIs (amygdala, anterior insula, dACC) are extracted from different sagittal slices and projected on the same slice for visual simplicity. ROI = region of interest, PINES = Picture Induced Negative Emotion Signature, VIFS = Visually Induced Fear Signature. (b) Structure of scenes and faces task. (c) Whole-brain machine learning models across three sequential analytic stages (first two preregistered). Each stage adds modeling choices to the previous stages. PLS = Partial Least Squares Regression, PCR = Principal Component Regression, SVR = Support Vector Regression, PCA-RF = Principal Component Analysis with Random Forest Regression.

• 

  Figure 2.. Correlations between neuroticism and brain measures across two fMRI tasks and three neural levels: Regions, networks, and validated multivariate neural signatures. For region sub-models, “pattern” refers to machine learning models limited to these regions. Hollow circles for the best region approach indicate the point estimates from the split-half selection sample.

• 

  Figure 3.. Cross-validated accuracy in the training data as a function of (a) neuroticism facet and task and (b) self-reports vs reports of close others. Includes only preregistered models. (c) Significant regression weights for the stress vulnerability pattern after correction at α = .05 and minimum cluster size of 50 (positive weights in red, negative in blue). (d) Correlations between brain pattern and hold-out data (N=102) after removing single regions and networks. “Sparse mask” represents the performance of the full model after applying a more conservative gray matter mask, which decreases the occurrence of artifactual voxels but also has a higher probability to miss relevant voxels.

• 

  Figure 4.. Correlations between theory-based neural targets and vulnerability to stress with versus without image-wise centering.

• 

  Figure 5.. Cross-validated accuracy of all fitted models, separated by outcomes (a) neuroticism facets and (b) different affective traits. Models are ordered based on their accuracy along the x-axis. “Task ratings” refers to a participants’ average negative affective self-ratings during the scenes task.

• 

  Figure 6.. Design factors and cross-validated correlations. Blue bars show variable importance when predicting correlations from design factors using random forest regression. It reflects the drop in the accuracy when a given predictor is randomly shuffled in the metric of root mean squared error (i.e., average deviation between predicted and actual correlations). Red bars show standard deviations from a linear mixel model variance decomposition (i.e., how much correlations differ between design choices within a given factor).