Table 4. Moderators.

	Moderator
	Training characteristics							Sample characteristics
Outcome	training paradigm (AS vs other)	imagery instructions (y/n)	word generation (y/n)	mode of presentation (visual/auditory)	feedback administration (yes/no)	Training items/total items (ratio)	# sessions (frequency)	health status (symptoms/ healthy)	mean age	% males (sex)
cognition
difference between positive and negative interpretation bias after benign training	0.90 ***	1.05 a ***	−0.34	0.15	0.06	0.17	n.ab	0.30	−0.00	−0.02 *
change in positive interpretation bias from pre- to post benign training	−0.95 ***	0.50 **	n.a c	−0.50	0.46 *	−2.78 **	0.08 **	0.69 **	−0.02	−0.03 **
mood
change in negative mood frompre- to post benign training	−0.29 d *	0.50 **	Interaction with paradigme	−0.14	0.20	−0.34	0.07 *	0.15	−0.00	−0.01 **
emotional challenge
change in negative mood from pre- to post emotional challenge	0.23	−0.07	0.16	−0.15	0.05	0.46	−0.01	−0.20	0.01	0.02 **

Note. Regression coefficients of the meta-regression models including the respective moderator. If paradigm resulted in a significant moderation of training effect, it was included in the tests of the other training-characteristic variables. *p<.05, ** p<.01, ***p<.001; ^aperfect confounding with paradigm; ^b there were 33 samples with one training session, one sample with four, and one sample with eight training sessions, it could thus not reasonably be tested; ^c there were only two samples asking individuals to generate word fragments and 15 that exposed participants to training stimuli, therefore no reasonable testing was possible; ^d p = .05; ^e there was a significant interaction with paradigm (β = −.75, p = .02) indicating that the effect of generation depends on the paradigms used, for all combined other paradigms (i.e. homograph, WSAT, or other; k = 6) the effectiveness is reduced when generation is employed as compared to when no generation is employed, whereas this is not true for the AS paradigms.