Introduction

Memory illusions have been in the spotlight of cognitive and neurocognitive scientists for decades. The discovery that memory does not lead to static representations of events is now a milestone concept; memory scientists have begun to investigate in detail the numerous variables that influence the occurrence of distortions, including the emotional state or mood of the rememberer. The present study is focused on the effect of transient mood (negative, positive, and neutral), induced immediately before retrieval, on the creation of memory distortions for everyday scripted events.

When individuals are exposed to mood-inducing emotional stimuli, the tendency to incur memory errors—or false memories—may depend on the valence of their affective state. Some studies have shown that negative mood reduces and positive mood enhances memory errors [1–3], and that this is especially true when mood is induced before learning. In these studies, the Deese-Roediger-McDermott paradigm (DRM; [4]) was employed. The DRM paradigm consists of the presentation of a series of wordlists including semantically related words (e.g., nurse, sick, medicine). Within each list, words also converge in meaning toward a non-presented lure (e.g., doctor); the acceptance of this word at recognition represents a false memory. According to the affect-as-information hypothesis, negative mood affects encoding as it favors item-specific processing, which limits the incorporation of non-presented lures; on the other hand, positive mood promotes relational processing, which facilitates the endorsement of items that share the semantic properties of the encoded events [5].

Other studies using the same paradigm have found that arousal (i.e., level of activation) induced before encoding, and not valence per se, affects the likelihood to incur memory errors. Nonetheless, evidence is mixed: in some cases, high arousal has been found to elicit higher false recall and false recognition than low arousal [6], but also the opposite has been found, with high arousal leading to fewer memory errors than low arousal [7]. Importantly, arousal exerts its influence on memory distortions not only at encoding but also when induced after learning: indeed, it has been found to reduce acceptance of misleading information in a misinformation paradigm [8]. The authors of this study argue that arousal enhanced memory consolidation via modulation of the hippocampal memory system by the amygdala; indeed, the amygdala is activated by many different emotional conditions [9]. However, in this study only negative-arousing and neutral stimuli were employed to induce mood, making it difficult to draw conclusions about the influence of valence on false memory production.

Furthermore, a number of studies have shown that high levels of stress influence subsequent true and false memory. For example, stress has been found to reduce false memories when induced either before encoding [10] or before retrieval [11]. However, other studies found that stress experienced at retrieval did not affect false memories, but only hindered true memory [12], and that stress experienced prior to learning DRM wordlists enhanced subsequent critical lures endorsement [13]. While deriving different conclusions, all these studies highlight the importance of the biological mechanisms underlying stress/arousal influence on memory; stress hormones (i.e., cortisol and adrenaline/noradrenaline) influence hippocampal activity, and through interaction with the amygdala they can mediate learning and subsequent memory, especially for emotional events. All these studies focused on induced stress leading to negative and disturbing perceived arousal; they were not aimed at studying the role of arousing pleasant events, leaving unsolved what the different impact of both valence and arousal on memory would be.

Whether valence or arousal play a major role and whether retrieval-related processes are important in modulating the influence of mood on memory is still poorly understood. Bless and colleagues [14] did not find evidence of a distinct influence of mood on false memories when induced at encoding or retrieval. In their study, participants had to listen to a script-based story, on which they were later going to be tested; either before or after encoding the story, they were asked to provide a written report of a personal life event which had to be either happy (inducing a happy mood) or sad (inducing a sad mood). The authors found that–regardless of time of mood induction–people in happy moods were more likely than people in sad moods to incorporate non-presented typical information into their memory, given their higher reliance on the general knowledge structure resulting from the script. However, as in Storbeck and Clore’s study [3], the emotional negative condition was low-arousing; thus, whether high arousing stimuli presented before retrieval influence subsequent true and false memories is still to be clarified.

The present study ought to clarify the effect of induced mood on false memories for everyday events, comparing high-arousal mood states of positive and negative valence with low-arousal mood of neutral valence, concurrently analyzing objective and subjective measures of false recognition. A false-memory paradigm for photographs depicting script-like events [15] was employed; this paradigm allows for the investigation of two types of errors which arise when an individual remembers either a script-consistent but non-presented event (i.e., gap-filling error) or the non-presented cause of a studied action effect (i.e., inferential causal error [16]; for further adaptations see [17–20]). For example, in the “family dinner” script, a gap-filling error could result from wrongly recognizing the adult protagonists eating pasta, whereas instead they were drinking wine; and a causal error could be caused by wrongly recognizing the cause of an action such as the child protagonist knocking over a bottle of water, whereas only the effect scene had actually been presented (e.g., pieces of a broken bottle on the floor). Furthermore, subjective measures of recollection and familiarity were collected.

The assessment of the influence of mood on subjective experiences related to memory errors has been neglected thus far. The remember-know paradigm [21] requires introspection about one’s own memory states in order to provide a more detailed recognition response at the time of testing. Remember judgments are to be given not only when the person is able to recognize a certain item as previously seen, among new ones, but also when some qualitative features are available, including personal thoughts or feelings that were linked to the encoding of the item. Know judgments reflect, instead, that the person is able to recognize a certain item, but no other qualitative features are available.

Recently, one study showed that participants in a negative mood associated more remember responses to negative DRM critical lures compared to neutral critical lures, and this happened both at immediate recognition and one week later [22]. However, a positive mood group was not included in this study, thereby preventing the specific assessment of differently valenced emotional states. Another study [23] found mood-congruent subjective recollection for false remembering—that is, individuals in negative moods associated more remember responses to negative DRM wordlists than either positive or neutral mood groups, and individuals in positive moods associated more remember responses to positive wordlists compared to negative and neutral. In the current study, asking participants to provide remember-know judgments has a twofold objective: on the one hand it would provide qualitative information on the nature of memory errors; on the other hand, it would help clarifying the importance of retrieval-related processes (such as monitoring strategies based on recollection).

In the present study, mood was induced immediately before recognition, in order to study retrieval-related processes. Based on extant studies which investigated the influence of valenced mood—induced before retrieval—on false memories [3, 14], one could expect that positive and negative mood states would influence false memory production to a different extent (with negative mood preventing individuals from incurring false memories and positive mood enhancing the probability of committing errors). However, these studies only examined low arousal negative condition (i.e., sadness), making it difficult to evaluate the effect of arousal. Based on research showing that arousal induced after learning diminishes the propensity to incur distortions [8], one could expect that both positive and negative moods would prevent individuals from incurring memory errors compared with neutral mood. Most important, to our knowledge, the effect of arousing mood at retrieval on false memories for everyday events has not been studied yet. In the current study, positive and negative stimuli used to induce moods were equated for arousal levels, allowing for the investigation of the specific role of valence and arousal: if differences in false memory production should emerge between positive and negative moods, then valence plays a major role. However, if differences emerge between neutral mood and both positive and negative moods, then arousal must be considered as being the central factor that affects false memory formation.

Finally, as typically done in this field of research [e.g., 3; 7], the experiment was designed to test participants belonging to different mood groups. Therefore data analyses were initially conducted by comparing groups against each other. However, at a later stage of the analyses we wanted to investigate whether using the participants’ own ratings on their perceived valence and arousal levels before and after mood-induction would better predict memory outcomes compared to analyzing the impact of mood on false memory in a way that categorizes participants as belonging to either one of three mood groups (i.e., positive, negative or neutral). Indeed, while traditionally used in studies with mood induction, comparisons between groups neglect part of the variability, as they overlook the fact that within the same group one participant may be strongly affected by mood induction, while another may be only mildly affected.

Materials and Methods

Results

Mood manipulation check

A mixed 3 (mood group: negative vs. positive vs. neutral) × 2 (pre-mood vs. post-mood induction) ANOVA, with mean valence ratings as the dependent measure, revealed a significant interaction between mood group and pre-post mood induction condition, F(2,72) = 27.12, p < .001, η_p² = .43; post hoc analyses (with Bonferroni correction) showed that valence decreased in the negative group (p < .001) and increased in the positive group (p = .04), while it did not change in the neutral group. In addition, groups did not differ in valence in pre-mood condition (all ps > .87), whereas they did differ in post-SAM reports, with valence being lower in the negative group than both the positive and neutral groups (ps < .001). A similar analysis conducted on the mean arousal ratings as the dependent measure revealed a significant interaction between mood group and pre- and post-mood induction, F(2,72) = 47.79, p < .001, η_p² = .57. Post hoc analyses (with Bonferroni correction) showed that arousal increased in both negative and positive mood groups (ps < .001), whereas it did not change in the neutral group (p = .26). Further, groups did not differ in arousal ratings prior to mood induction (all ps > .90), whereas they did differ in post-SAM reports: arousal was higher in both negative and positive groups than in the neutral group (all ps < .001), and it did not differ between negative and positive groups (p = .61). See Table 1 for the descriptive statistics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Mean valence and arousal ratings before (1) and after (2) mood induction.

https://doi.org/10.1371/journal.pone.0148716.t001

Memory performance

False recognition.

To analyse the effect of mood on false recognition, the responses to causal and script-consistent distractor photographs were considered. As responses were of a binomial type (“yes”: 1, or “no”: 0), a logistic mixed-effects model approach was used [29, 30]. Causal and gap-filling errors were the “yes” responses to causal and new script-consistent photographs, respectively. In both cases, the fixed effect was Group (3 levels: negative, positive, neutral), while Participants and Scripts were treated as random effects. The significance of both fixed and random effects was tested through a series of likelihood ratio tests for nested models based on the chi-square distribution [31]. The Akaike Information Criterion (AIC; [32]) was also reported for each model; lower AIC indicates a better model. Odds ratios were used as a measure of effect size.

For causal errors, both random effects (i.e., participants and scripts) were significant: for Participants, χ²(1) = 4.89, p = .03 (full model: AIC = 543.3; model without Participants: AIC = 546.2); for Scripts, χ²(1) = 87.33, p < .001 (model without Scripts: AIC = 628.6). With regard to the fixed effect of Group, a significant main effect was found, χ²(2) = 9.75, p = .008 (model without Group: AIC = 549.1). The probability of producing causal errors in the three groups is reported in Fig 1A. Planned comparisons showed that the probability of producing causal errors was lower in both negative and positive groups than in the neutral group (ps < .05). See Table 2 for detailed information on the model, including estimated parameters and odds ratios.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

Mean proportions of causal errors (A) and gap-filling errors (B) as a function of mood group; error bars represent standard errors.

https://doi.org/10.1371/journal.pone.0148716.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Fixed effect of Group on Causal Errors, Gap-filling Errors, and Hits using a logistic mixed-effects model.

https://doi.org/10.1371/journal.pone.0148716.t002

For gap-filling errors, both random effects were significant: for Participants, χ²(1) = 102.22, p < .001 (full model: AIC = 1476.3; model without Participants: AIC = 1576.5); for Scripts, χ²(1) = 170.38, p < .001 (model without Scripts: AIC = 1644.7). The fixed effect of Group on gap-filling errors was not significant, χ²(2) = 0.78, p = .68 (model without Group: AIC = 1473.1). The probability of producing gap-filling errors in the three groups is reported in Fig 1B. See Table 2 for detailed information.

Responses to script-inconsistent distractors were not included in the analysis given that false alarms relative to these items approached floor effect (5%). They were included in the paradigm only to avoid primacy and recency effects on the relevant material.

Accuracy.

A similar logistic mixed-effects model was tested to analyse the effect of mood on hits (i.e., the “yes” responses to targets; “yes”: 1, “no”: 0). As above, the fixed effect was Group, and random effects were Participants and Scripts. Both random effects emerged as significant: for Participants, χ²(1) = 101.01, p < .001 (full model: AIC = 2143.4; model without Participants: AIC = 2242.3); for Scripts, χ²(1) = 180.76, p < .001 (model without Scripts: AIC = 2322.1). With regard to the fixed effect of Group, a significant main effect was found, χ²(2) = 7.85, p = .02 (model without Group: AIC = 2147.2). Planned comparisons showed that the probability of correctly recognizing target photographs was higher in both negative and positive groups than in the neutral group (negative: p = .005; positive: p = .05 see Fig 2). See Fig 2 and Table 2 for detailed information.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Mean proportions of hits as a function of mood group; error bars represent standard errors.

https://doi.org/10.1371/journal.pone.0148716.g002

Remember and familiar responses.

Additional analyses were performed to examine whether remember and familiar responses associated with causal errors and hits would change according to mood. Again, Participants and Scripts were treated as random effects, and significance was tested through likelihood ratio tests for nested models.

The analysis on the remember responses associated with causal errors revealed a significant main effect of Group, χ²(2) = 14.46, p < .001 (full model: AIC = 283.11; model without Group: AIC = 293.57). Planned comparisons showed that the probability of associating remember responses to causal errors was lower in both negative and positive groups than in the neutral group (ps < .05). On the contrary, with regard to familiar responses associated with causal errors, no main effect of Group was found, χ²(2) = 0.74, p = .69 (full model: AIC = 463.91; model without Group: AIC = 460.65). See Table 3 for detailed information on the models.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Fixed effect of Group on remember and familiar responses associated with Causal Errors.

https://doi.org/10.1371/journal.pone.0148716.t003

The analysis on the remember responses associated with hits did not show any significant main effect of Group, χ²(2) = 2.08, p = .35 (full model: AIC = 1891.7; model without Group: AIC = 1889.8). Furthermore, the analysis on the familiar responses associated with hits did not show any main effect of Group as well, χ²(2) = 2.24, p = .33 (full model: AIC = 1887.1; model without Group: AIC = 1885.3).

Comparing mood ratings vs. groups in predicting memory performance.

To test whether actual mood ratings predict memory performance better than groups, we computed another series of logistic mixed-effects models on causal errors, gap-filling errors, and hits, but using the actual ratings reported by the participants through SAM as the predictors. In particular, we calculated differential valence and differential arousal as the difference between the valence and arousal ratings reported before and after mood induction. The significance of these predictors was tested through a series of likelihood ratio tests. Participants and Scripts were again included as random effects. Finally, the resulting models were compared with the corresponding models having group as the predictor. As the to-be-compared models are non-nested, comparisons were made through the relative likelihood based on the models’ AICs–this quantifies the strength of evidence in favour of one model over the other [33]; specifically, the relative likelihood was calculated as Exp((AIC₁-AIC₂)/2) [34].

For causal errors, differential valence did not have a significant effect, χ²(1) = 2.31, p = .13 (model with differential valence: AIC = 548.8; model without differential valence: AIC = 549.1), but a significant main effect of differential arousal was found, χ²(1) = 14.57, p < .001 (model with differential arousal: AIC = 536.5; model without differential arousal: AIC = 549.1); as shown by the parameters reported in Table 4, this indicates decreasing probability of causal errors with increasing arousal ratings. No interaction between differential arousal and valence emerged, χ²(1) = .57, p = .45 (model with the interaction: AIC = 539.0; model without interaction: AIC = 537.6). The relative likelihood of the model with differential arousal as the predictor (i.e. the best model) with respect to the corresponding model with group was 30.3, indicating clear evidence in favour of differential arousal (following [33], this implies that the first of the two models is 30.3 times more likely to be the best model than the latter).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Fixed effect of differential valence and differential arousal on Causal Errors, Gap-filling Errors, and Hits using logistic mixed-effects models.

https://doi.org/10.1371/journal.pone.0148716.t004

For gap-filling errors, differential valence did not have a significant effect, χ²(1) = 0.64, p = .42 (model with differential valence: AIC = 1474.5; model without differential valence: AIC = 1473.5). A significant main effect of differential arousal emerged, χ²(1) = 4.68, p = .03 (model with differential arousal: AIC = 1470.4; model without differential arousal: AIC = 1473.5); as shown by the parameters reported in Table 4, this indicates decreasing probability of gap-filling errors with increasing arousal ratings. No interaction between differential arousal and valence emerged, χ²(1) = .62, p = .43 (model with the interaction: AIC = 1473.6; model without interaction: AIC = 1472.2). The relative likelihood of the model with differential arousal as the predictor with respect to the corresponding model with group was 19.1, again indicating clear evidence in favour of differential arousal.

Finally, also for hits, no main effect of differential valence was found, χ²(1) = .01, p = .94 (model with differential valence: AIC = 2149.2; model without differential valence: AIC = 2147.2), but significant main effect of differential arousal was found, χ²(1) = 5.13, p = .02 (model with differential arousal: AIC = 2144.0; model without differential arousal: AIC = 2147.2); as shown by the parameters reported in Table 4, this indicates increasing probability of hits with increasing arousal ratings. No interaction between differential arousal and valence emerged, χ²(1) = .44, p = .51 (model with the interaction: AIC = 2147.3; model without interaction: AIC = 2145.8). The relative likelihood of the model with differential arousal as the predictor with respect to the corresponding model with group was 0.70, this time indicating that while both group and differential arousal are good predictors of hits, it is not possible to establish which of the two is a better predictor.

Discussion

Mood was found to affect true and false memories for events when induced before retrieval. Specifically, arousal—but not valence—predicted the probability of committing memory errors: both negative and positive mood groups (similar for arousal levels) were less likely to produce causal errors compared to the neutral group. This is even more evident when we directly used the difference in the participants’ ratings between pre- and post-mood induction as the predictor of memory errors. Indeed, the increase in arousal after exposure to emotional scenes (compared to the individual’s baseline) predicted the extent to which individuals would be later influenced in their memory performance. Larger increases in perceived arousal predicted lower probability to produce overall false memories (i.e., individuals were less likely to incorporate unseen events into their memories), and higher probability to correctly recognizing target events. On the contrary, differential valence did not predict memory performance. Most important, analyses of the relative likelihood of the models clearly indicated that self assessed mood ratings predict false memory performance better than mood groups. This is important because it shows that assigning participants to different mood groups and subsequently conducting between-group comparisons–while still useful–may not be the most adequate and powerful way to test the effect of emotional states on memory; for example, in our study, a significant effect of mood on gap-filling errors emerged only when we analyzed arousal as a continuum. Analyses on the subjective ratings allow to evaluate the individual’s perception of their changing emotional state, and to take into consideration individual differences in the extent to which participants respond to mood induction (and subsequently how this impacts their memory performance).

The present findings extend previous evidence that arousal, in some circumstances, reduces the susceptibility to false memory (e.g., [8]). In their study, English and Nielson found that arousal exerted its influence on consolidation processes (i.e., when the memory trace has to be maintained in long-term memory). While we cannot rule out that consolidation processes are still at play in our study, we argue that arousal affects retrieval to a higher extent. Participants were not instructed to remember the photographs presented at encoding (i.e., the subsequent memory test was incidental) and, furthermore, they performed filler tasks during the retention interval that distracted them from thinking of the viewed photographs. Finally, participants were exposed to mood-induction immediately before they had to retrieve the episodes. All these aspects suggest that the current findings are due to processes taking place at retrieval. Future research should address whether mood affects false memory to a different extent depending on the memory stage involved (encoding, consolidation or retrieval). The findings relative to subjective experiences of recollection and familiarity–described below–sheds further light on the role of the retrieval processes involved in the current study.

Given that arousal affected both true and false memories, there may be a common mechanism that should explain the current findings. Most of the available research on the effects of arousal and stress on memory highlights the impairment of memory performance when arousal is experienced before retrieval [e.g., 12]. One recent study shows that whether memory retrieval is impaired depends on time, and on the specific physiological processes involved [35]. These authors tested the hypothesis that if individuals are experiencing stress and are simultaneously required to retrieve previously studied information, then an immediate release of adrenaline and noradrenaline from the autonomic nervous system occurs and could facilitate memory retrieval. Indeed, noradrenergic arousal has been linked to successful memory retrieval [36]. However, if retrieval takes place several minutes after the stressful experience, then the hypothalamus-pituitary-adrenal axis triggers a cascade of events which lead to the secretion of glucocorticoids (i.e., cortisol) and would impair memory retrieval. They found that, while a specific effect of stress on memory was not found, individual differences in change in blood pressure (indexing activity of the autonomic nervous system) were positively correlated with memory retrieval when participants were still under stress (but not 25 min after stress). Conversely, individual differences in cortisol response were negatively correlated with memory retrieval 25 min post stress (but not when participants were under stress). Schönfeld and colleague [35] argue that, consistently with previous research, rapidly occurring noradrenergic mediators improve the capacity of the hippocampus and the prefrontal cortex, while cortisol is not yet at its peak. Given that our retrieval task took place immediately after mood-induction (and thus individuals were under the effects of arousal while performing the retrieval task), we speculate that our findings may be consistent with this view that arousal enhances memory when experienced during retrieval. Furthermore, our findings are consistent with the U-inverted function relation between arousal and level of performance, described by the Yerkes-Dodson law [37] (where performance increases with increasing levels of arousal only up to a point, after which performance drops), assuming that the arousal level induced in the current study was moderate-high, and the task complexity was not high.

High arousal may be associated with increased monitoring abilities at retrieval (see [38] for a review on theories on monitoring processes in false memory research). When a person is in a transient negative or positive mood, he or she may be prone to actively control the available memory traces in order to discern events that were actually experienced from those only inferred. For example, viewing the photograph of a child with pieces of a broken bottle on the floor (effect) leads to the inference that the child knocked over the bottle from the table (cause); however, when the causal scene is available at recognition, people in arousing moods tend to reject it, likely recognizing that the cause was only inferred.

The present findings on the subjective experiences accompanying false memories shed more light on the processes involved. Once false memories of causal scenes occur in individuals in negative and positive moods, they are not characterized by subjectively compelling memory traces; negative and positive moods lead to a lower recollection for unstudied events than neutral mood. However, remember responses associated with hits were not different according to the mood state, suggesting that the impact of arousal on memory performance may be stronger for memory distortions than memory accuracy, at least at the subjective level.

One of the qualifying features of recollection is the ability to retrieve qualitative aspects of the studied events’ context, along with thoughts that came to mind when the rememberer first encountered the event [39]. Illusory recollection, or phantom recollection, occurs when individuals erroneously claim to remember non-experienced events, and these events are accompanied by vivid details (e.g., [40]). For example, in the DRM paradigm, when a related lure is presented at retrieval, details originally linked to the presentation of target words are recollected and mistakenly attributed to the critical lure in a process that can be thought of as “content borrowing” [41]. In the current study, it is likely that fewer contextual and qualitative details of unstudied items are available at retrieval for people in arousing mood states, reducing the risk of illusory memory. In turn, this may serve as evidence of recollection-based monitoring processing that limits the overall rate of false memories. From a dual-process perspective, the fuzzy-trace theory [40] states that under certain conditions, the retrieval of verbatim traces (i.e., surface features of studied events) may limit the influence of gist traces (i.e., the general meaning of studied events, which is thought to be at the base of false-memory formation). Thus, when a person is exposed to non-presented causal scenes, he or she may retrieve the corresponding effect scenes in detail and consequently reject the cause because it seems familiar but it is not recollected or felt as “encountered before.” We found evidence of a reduced subjective recollection for false memories when individuals were exposed to mood before retrieving the episodes, suggesting that mood influences retrieval processing via recollection-driven mechanisms.

The current findings may have important applications. First, not only is it important to consider the affective state of the rememberer at the time when he or she experienced some (emotional or neutral) event, it is also vital to assess the transitory affective state that accompanies the person at the time of retrieving those events from memory. People in negative and positive highly arousing moods discriminate to a higher extent between experienced and non-experienced events than people in a neutral mood and are more able to ward off false remembrance, especially that which stems from causal inferential thinking. Finally, the participants’ subjective ratings about their perceived arousal states are valuable predictors of false memory performance. To conclude, high arousal at retrieval seems a valuable predictor of diminished false memory production and enhanced memory performance.

Acknowledgments

We are grateful to Alessandra Palmieri for helping with data collection.