Descriptive statistics of individual differences, control variables, and direct and generalized training effects are reported in Table 1.
Chi square analyses and independent samples t-tests showed there were no significant group differences in gender distribution (10 men and 50 women, and 9 men and 51 women in the optimism and pessimism condition, respectively), age (M = 19.97, SD = 2.30 and M = 20.27, SD = 3.74 in the optimism and pessimism condition, respectively), university course (49 psychology and 10 other, and 55 psychology and 5 other in the optimism and pessimism condition, respectively), or marital status (57 single and 3 other in both conditions), all p > .15. There was a significant difference in the distribution of nationality, with relatively more Belgians (48) than non-Belgians (7 Dutch and 5 other) in the pessimism condition compared to the optimism condition (58 Belgians and 5 Dutch), χ2(2) = 8.72, p = .01.
Baseline group differences in social anxiety (LSAS total score, fear subscale, avoidance subscale), depression (BDI-II, DASS-21 depression scale), anxiety and stress (DASS-21 anxiety and stress subscales), and adult attachment (ECR anxiety and avoidance subscales) were tested with independent samples t-tests. There were no significant group differences, all t < 1.95, all p > .05, except for the ECR scales. Attachment anxiety was marginally significantly higher in the pessimism condition, t(118) = 1.95, p = .05, and attachment avoidance was significantly higher in the pessimism condition, t(118) = 2.04, p = .04. This was due to two outliers in the pessimism condition, with scores higher than three times the standard deviation above the group mean. These cases did not have outlying scores on any of the dependent variables (SST or M-AMT). DASS-21 scores were entered as covariates in the analyses which tested the effectiveness of the training and the effect of training on the interpretation bias and valence of autobiographical memories and future projections. This was done because depression, anxiety and stress may plausibly influence the effect of the training. The DASS-21 can be entered as a covariate because there are no significant group differences at baseline (Miller and Chapman 2001), and controlling for the covariate can therefore increase statistical power.
A 3 (Time: pre-CBM, post-CBM, post-filler) × 2 (Condition: optimism, pessimism) mixed model repeated measures MANOVA was run with the mood items as multiple dependent variables, Time as a within-subject variable and Condition as the between-subjects variable. Multivariate tests showed a main effect of Time, F(10, 109) = 7.72, p < .001, Cohen’s f = 0.85, suggesting that the CBM training affected mood across conditions. Univariate tests showed significant Time effects for all mood items, smallest F > 4.14, largest p < .02, smallest f > 0.18. Pairwise comparisons (significance level of α = 0.05) showed that, across conditions, optimism, anxiety, and happiness decreased significantly from pre-CBM to post-CBM, all p < .01, but did not change from post-CBM to post-filler, all p > .12. Sadness decreased from pre-CBM to post-CBM, and decreased further from post-CBM to post-filler, both p = .03. Arousal decreased from pre-CBM to post-CBM, p = .03, but returned to baseline levels at post-filler, p = .91. There was no significant main effect of Condition, F(5, 114) = 0.72, p = .61. The Time × Condition was not significant, F(10, 109) = 1.78, p = .07. Importantly, independent t-tests showed that post-filler mood was comparable between the two conditions for all mood items, largest t < 1.01, smallest p > .31. Thus, it was unlikely that any mood effects of the training were responsible for training effects.
Independent t-tests showed no significant differences between conditions on the valence of the autobiographical memory at baseline, t(118) = 0.12, p = .91.
In the optimism group, there were three participants who suspected that the CBM training was aimed at making them more positive. In the pessimism group, two participants suspected that the CBM-training was aimed at making them more negative. This difference was not significant, χ2(1) = 0.21, p = .65. In each group, the number of participants that suspected the goal of the study was low. All analyses were rerun without these five participants but the pattern of results was similar. Therefore, it was decided not to exclude their data from the analyses.
Training Effectiveness: RTs and Expectancy Bias Assessment
Reaction times (RTs) to positive and negative target scenarios were analyzed as a measure of training effectiveness. In total, 13.20% of trials were excluded, which were trials with incorrect answers either to the word fragment (55.56% of total excluded trials) and/or the comprehension question (44.44% of excluded trials). No trials needed to be excluded based on extremely short (< 200 ms) or long RTs (M + 3SD), M = 1757.78 ms, SD = 1984.05 ms. A 2 (Target: positive scenario vs negative scenario) × 2 (Condition: optimism vs pessimism) mixed model ANCOVA was run with reaction times (in ms) as the dependent variable, Target as a within-subject variable, Condition as a between-subjects variable, and DASS-21 subscale scores as covariates. There was no significant main effect of Condition, F(1, 115) = 1.78, p = .19, thus, average reaction times during the training were comparable in the two conditions. There was no significant main effect of Target, F(1, 115) = 1.57, p = .21, indicating that across conditions participants responded equally fast to positive and negative target scenarios. Importantly, the Target × Condition interaction was significant, F(1, 115) = 24.62, p < .01, f = 0.46, see Fig. 1. Because the covariates (DASS-21) were not significantly related to reaction times in the ANCOVA (all p > .17), these were not included in the follow-up tests.
Paired samples t-tests showed that participants in the optimism condition were significantly faster in responding to positive targets relative to negative targets, t(59) = 4.53, p < .01, d = 0.86, whereas participants in the pessimism condition were significantly faster in responding to negative targets than positive targets, t(59) = 2.19, p = .03, d = 0.36.
Expectancy Bias Assessment
The effect of the training on the expectancy bias was tested by comparing the change in the expectancy bias index between the two conditions. This was done with a 2 (Time: pre-CBM vs post-CBM) × 2 (Condition: optimism vs pessimism) mixed model ANCOVA with the positive expectancy bias index as the dependent variable, Time as the within-subject variable, Condition as the between-subjects variable, and the DASS-21 subscale scores as covariates. There was no significant main effect of Time, F(1, 115) = 2.10, p = .15. There was a significant main effect of Condition, F(1, 115) = 16.92, p < .01, f = 0.38, showing a more positive expectancy bias across time in the optimism condition compared to the pessimism condition. Importantly, the Time × Condition interaction was significant, F(1, 115) = 18.00, p < .01, f = 0.40 (see Fig. 2, Top). DASS-21 depression and stress were significant in the ANCOVA (p = .003 and p = .043, respectively), therefore, these were included in the follow-up tests. Repeated measures ANCOVA within each condition showed that the positive expectancy bias did not change significantly in the optimism condition, F(1, 57) = 0.03, p = .857. However, positive expectancy significantly decreased in the pessimism condition, F(1, 57) = 8.67, p = .005, f = 0.39, indicating that, as intended, the manipulation resulted in a difference in optimism between the two conditions.
A similar analysis was run for the generalized positive expectancy bias (based on the general foil statements), with similar results. There was no significant main effect of Time, F(1, 115) = 0.90, p = .35. There was a significant main effect of condition, F(1, 115) = 5.14, p = .03, f = 0.21, showing a more positive generalized expectancy bias overall in the optimism condition. The Time × Condition interaction was also significant, F(1, 115) = 13.58, p < .01, f = 0.34 (see Fig. 2, Bottom). The DASS-21 depression scale was a significant covariate (p = .001) and was included in follow-up ANCOVAs. There was a significant increase in the optimism condition, F(1, 58) = 4.12, p = .047, f = 0.27, and a significant decrease in the pessimism condition, F(1, 58) = 4.59, p = .036, f = 0.28.
Autobiographical Memory and Future Projection Bias
Two ANCOVAs were run to test whether expectancy bias affected the valence of autobiographical recall and autobiographical future projections. Condition (optimism, pessimism) was the independent variable, the number of positive memories/future projections on the M-AMT the dependent variable respectively, and DASS-21 subscale scores were included as covariates. There were no significant differences between the two experimental conditions in the number of positive autobiographical memories, F(1, 115) = 0.45, p = .506, or future projections, F(1, 115) = 0.56, p = .456, on the M-AMT. These findings suggest that the CBM training effect did not generalize to biases in autobiographical recall or future projections for social situations.
To test the effect of the CBM-E training on social interpretation bias, a one-way ANCOVA was run with Condition (optimism vs pessimism) as the independent variable, the number of positive solutions on the social items of the SST as the dependent variable, and DASS-21 subscale scores as covariates. The main effect of condition was significant, F(1, 115) = 4.04, p = .047, f = 0.19, and indicated a more positive interpretation bias in the optimism condition. An ANCOVA was run with the mean number of positive solutions for the future items on the SST as the dependent variable. This indicated no significant difference between the two conditions, F(1, 115) = 0.74, p = .39.
Moderation by Social Anxiety, Attachment Anxiety and Attachment Avoidance
It was tested whether social anxiety or attachment moderated the relationship between the training and the memory/future projection bias (Table 2) and interpretation biases (Table 3). The DASS-21 subscales correlated significantly with each of the moderators and were therefore not included in these regression models. The first block always contained the valence rating of the autobiographical memory at baseline. The second block added a dummy variable for experimental condition. The third block added the moderator (LSAS total score, ECR anxiety, or ECR avoidance). Finally, the interaction term between condition (dummy variable) and the standardized scores of the moderator were entered to test the moderation effect.
Interpretation Bias (SST)
There was no significant moderation of the relation between the CBM-E training and interpretation bias for social sentences on the SST. Attachment anxiety was the only significant moderator of the relation between the CBM-E training and the interpretation bias for future-related sentences on the SST. The interaction equation indicated that higher attachment anxiety was related to fewer positive interpretations of future-oriented sentences in the pessimism condition than in the optimism condition (see Fig. 3, Top).
Memory and Future Projection Bias
There was no significant moderation of the relation between the CBM-E training and valence of autobiographical memories on the M-AMT. Social anxiety was the only significant moderator of the relation between the CBM-E training and the valence of future projections on the M-AMT. The interaction equation indicated that higher social anxiety was related to fewer positive future projections in the optimism condition than in the pessimism condition (see Fig. 3, Bottom).