Statistical analyses

Statistical analyses were carried out using R statistical software (www.r-project.org) (65), unless specified otherwise. Voxel-wise differences in BP_ND between TBI participants and healthy controls, between each of the disabled TBI and good recovery TBI subgroups and healthy controls, and between these subgroups were assessed using permutation tests in the FMRIB Software Library (FSL) (66). The same procedure was used to assess voxel-wise differences between TBI participants and healthy controls in (i) VBM-derived white matter density, (ii) gray matter density, and (iii) diffusion MRI–derived skeletonized FA. Voxels corresponding to structural lesions were excluded from the analyses using participants’ lesion maps as individual voxel-wise repressors. All cerebral gray and white matter voxels were included except those corresponding to the striatum, where there is known off-target binding (12). One thousand permutations were used, and results were cluster-corrected for multiple comparisons using threshold-free cluster enhancement (TFCE) and a family-wise error rate of 0.05. Localization of voxel clusters was reported on the basis of the Harvard-Oxford probabilistic atlas (67) and the Johns Hopkins University (JHU) DTI-based white matter atlas (68).

To investigate whether white matter tau pathology shared common localization with TAI and/or white matter atrophy in TBI, we compared skeletonized FA and, separately, average white matter density (z-scores) within areas of high flortaucipir signal (BP_ND z > 1.645) versus white matter areas where binding was not increased (BP_ND z < 1.645) using (paired) Wilcoxon signed-rank tests. Any voxels corresponding to focal lesions were excluded. Only TBI participants with 15 or more white matter voxels with increased flortaucipir binding within the regions of interest were included.

To examine the localization of white matter microstructural damage predicted by increased flortaucipir cortical signal in TBI participants, standardized BP_ND within the right parietal cortical area of flortaucipir increase in TBI was used as a regressor in a voxel-wise general linear model (GLM) of skeletonized FA and, separately, in a voxel-wise GLM of white matter density. Statistical significance was tested using permutation tests in FSL (66). As previously, voxels corresponding to lesions were excluded for each individual, 1000 permutations were used and results were obtained using TFCE, corrected for multiple comparisons using a family-wise error rate of 0.05. Localization of voxel clusters was reported on the basis of the JHU DTI-based white matter atlas (68). Post hoc correlations were examined using Spearman’s rank correlation between standardized BP_ND within the right parietal cortical area of flortaucipir increase in TBI participants and skeletonized FA in the following tracts: right and left superior longitudinal, inferior longitudinal, inferior fronto-occipital fasciculi and cingulum bundles, corpus callosum (genu, body, and splenium), as well as in the right and left corticospinal tracts as a control.

Correlations between clinical/neuropsychological measures and PET markers in TBI participants were examined using Spearman’s rank correlation with FDR correction for multiple comparisons. Interactions between APOE genotype (presence of at least one ε4 allele) and age, time since injury, CSF biomarkers, white matter MRI measures, and neuropsychological performance on cerebral flortaucipir BP_ND and flortaucipir spatial extent were examined in the TBI group using linear regression in each case to assess the statistical significance of the interaction term and interaction plots to determine the direction of the effect in ε4 carriers versus noncarriers. Assumptions of multiple linear regression analysis were tested as follows: Analysis of standard residuals was carried out to identify outliers; variance inflation factor was used to test for multicollinearity; Durbin-Watson test was conducted to test for independence of errors; and approximate normal distribution of errors, homoscedasticity, and linearity were assessed by plotting regression-standardized residuals.