Table 4.

Prediction performance across models and methods.

Statistic	Method	GLM	SVR	DNN
Predicition performance	Without DT	r = 0.91, MAE = 5.07	r = 0.83, MAE = 6.82	r = 0.90, MAE = 5.18
	With DT	r = 0.93, MAE = 4.58	r = 0.86, MAE = 6.11	r = 0.90, MAE = 5.42
Comparison of models	Without DT	–	z = −7.44, p < 0.001	z = 0.23, p = 0.817
	With DT	–	z = −5.95, p < 0.001	z = −4.12, p < 0.001
Comparison of methods	With DT vs. Without DT	z = −4.90, p < 0.001	z = −4.78, p < 0.001	z = 4.04, p < 0.001

The first row lists performance measures for three decoding algorithms (GLM, SVR, DNN) and two prediction methods (without DT, with DT). The second row reports results from Wilcoxon signed-rank tests comparing GLM against SVR and DNN (thus no entries in the GLM column). The third row reports results from Wilcoxon signed-rank tests comparing each decoding algorithm with and without DT. Negative z-values indicate significantly lower absolute errors for GLM (second row) or DT (third row), respectively. GLM, multiple linear regression; SVR, support vector regression; DNN, deep neural network regression; DT, distributional transformation.