Fig. 4.

Effects of non-gaussianity for second level slope parameter and hyperparameter estimation accuracy. Generalized normal distributions (type I and II) were used for generation of trajectory data with non-Gaussian first and second level errors. We simulated ensembles of 64 subjects with 5 annual scans per person. These were sampled under balanced/unbalanced designs and linear Bayesian mixed-effects model inversion was performed. Skewness (top row) and kurtosis (bottom row) were independently manipulated from mean and variance. Estimated slope parameter and hyperparameter were compared to ground truth values computing the mean absolute error (MAE) over 200 independent realizations. Darker to brighter shading of MAE in plots depicts increasing first level errors std of 0.01, 0.02, 0.04, 0.08, and 0.16 respectively.