Fig 1.
A. The task comprised three phases, with affect inductions (90-second emotional film clips) presented prior to each of three blocks in the compound-generalization phase. For positive/negative affect groups, the induction was preceded by one additional compound-generalization block to measure generalization in neutral affect. B. Sequence of visual events for phase 1 (simple cue learning). One cue was presented on each trial, to the left or right of a central fixation cross; after a “planet” was selected for “mining”, its outline color changed to green to indicate the participant’s selection. The chosen planet’s outline color changed again to purple after a jittered interval, signifying that the planet could now by mined. At this point, the participant could press any key to reveal the outcome, which was indicated both by a change in planet fill color to blue (win) or orange (non-win) and a distinctive win or non-win sound. C. Sequence of visual events for task phases 2 and 3 (simple cue test and compound-generalization test). The structure of each trial was the same as in phase 1, except that two stimuli were presented and the participant could choose between them, and feedback was uninformative (a neutral purple color and neutral sound). Three example cue configurations are shown, representing (from top to bottom) a simple-versus-simple choice in the simple cue test phase, a simple-versus-compound choice in the compound-generalization phase, and a compound-versus-compound choice in the compound-generalization phase. Cues in compound stimuli were symmetrically offset from the center of the stimulus, ensuring that the distance of all cues from fixation was approximately equal.
Fig 2.
A. The simple cue test evidenced no significant between-groups differences in choice accuracy (i.e., proportion of trials on which the higher-value cue was chosen). B. As expected, there was a significant effect of affect-induction videos on self-reported mood valence (β = 2.63, p < .001; mood reports standardized within subject with reference to the average between-block mood change in the learning blocks, where no videos were presented). C. In the compound-generalization phase, there were no significant between-groups differences in choice accuracy (the proportion of trials in which the compound with the higher mean value was chosen). D. There was, however, a significant between-groups difference in preference for the L/H stimulus as opposed to the M/M stimulus in compound probe trials (β = 0.79, p < .001). E. There was also a significant between-groups difference in preference for the L/H stimulus as opposed to the M cue in simple probe trials (β = 0.56, p = .004). F. Preference for the L/H stimulus in simple probe trials and preference for the L/H stimulus in compound probe trials were positively correlated (Spearman ρ = .81, p < .001). Dot color reflects participant condition as per subplots A-E. For all panels, groups are presented as mean ± 95% CI. Points in scatterplots represent condition means for individual participants, and are jittered to prevent overplotting (horizontal jitter in subplots A-D, 2% horizontal and vertical jitter in subplot F).
Table 1.
First-stage computational models.
Table 2.
Goodness of fit of models to neutral-affect data (model numbers as per Table 1).
Table 3.
Results of second-stage model fit to all compound-generalization data.
Fig 3.
Estimated posterior distributions for group-mean parameters.
Top histograms: group-level means for the parameters αV (A), ϕ (B), and β (C). Bottom histograms: group level means for the effect of affect on αV (D), the effect of affect on ϕ (E), and the block-wise decay in the strength of these effects, λ (F). Horizontal error bars denote the 90% credible interval for each parameter, and vertical dotted lines denote the reference value for each parameter (i.e., the parameter value at which there is no influence of the parameter on behavior: 0 for additive parameters such as αV, 1 for parameters that were multiplicative or exponents such as ϕ and λ). Asterisks denote parameters for which there is credible evidence (i.e., estimated probability in excess of 95%) that the group mean is different from the reference value.
Fig 4.
Illustration of model estimates of attention to valence across time and between groups.
Negative values of αV indicate more attention to low-value cues; positive values indicate more attention to high-value cues. There was a strong affect-congruent modulation of attention to value in the generalization blocks, the strength of which decayed over time. The dynamics of mood as plotted here were calculated using medians of the posterior distributions of the αV, ΔαV, and λ parameters.
Fig 5.
Eye-tracking validation results.
A: Proportion of time spent looking at each cue in simple versus simple trials, as a function of whether a cue was subsequently chosen or not. For all groups, participants tended to look more at the to-be-chosen cue. B: Proportion of time spent looking at each cue in simple versus simple trials, as a function of cue value (25, 50 or 75% probability of reward). For all groups, participants looked longer at more valuable cues. Diamond markers denote the mean of each group and its 95% confidence interval; dots indicate individual participants. C: Relative looking time to the high-valued cue within a L/H compound (normalized to the total time looking at all cues in the trial), plotted against the tendency to choose the L/H compound in probe trials. D: Attention to value estimated from the computational model, plotted against the relative looking time to the H cue within a L/H compound during probe trials. Data points are color-coded by condition (neutral, positive and negative affect). Each point corresponds to one participant in each of the three groups.
Fig 6.
Affective modulation of relative looking time.
Average relative looking time to the affect-congruent cue (i.e., the low-value 25% cue for participants receiving a negative affect induction and the high-value 75% cue for participants receiving a positive affect induction) pre- versus post-affect induction. Data are plotted separately for compound probe trials (A) and simple probe trials (B). There was a significant increase in time looking at affect-congruent cues after the affect induction in compound probe trials (main effect of time; p < .05, denoted by *), but not in simple probe trials. Insets: data separated by affect-induction condition (red: positive affect; blue: negative affect). The affect by timepoint interaction was significant for simple probe trials (p < .05) but not for compound probe trials, though this interaction was driven by group differences in eye-gaze before the affect induction, and as such is of no interest. Error bars denote the standard error of the mean computed based on estimated marginal means from the linear mixed-effects analysis.