The devil's in the detail: Accessibility of specific personal memories supports rose-tinted self-generalizations in mental health and toxic self-generalizations in clinical depression.

Autobiographical Memory Priming Task

The Autobiographical Memory Priming Task (AMPT) was closely derived from that used by Klein et al. (2001) . Each trial comprised an initial component—processing a trait adjective that is either positively or negatively valenced—followed by a second component—recalling a specific memory (see Figure 1 ). For the initial component, participants completed either (a) a trait self-rating task, in which they were asked to think about how well the presented trait adjective described them or (b) a definition task, in which they were asked to think about the “dictionary definition” for that trait characteristic. The definition task served as a control condition as it required the individual to access information about the characteristic in a manner that was not explicitly self-referential. In the second component, participants were asked to recall a specific autobiographical memory that either: a) involved behavior consistent with the trait characteristic (i.e., a consistent memory), or b) involved behavior inconsistent with the characteristic (i.e., an inconsistent memory). The dependent variable was the time taken for participants to recall these specific memories.

The AMPT therefore comprised eight conditions in a 2 (Initial Component: Self-Rating, Dictionary Definition Control) × 2 (Valence of the Initial Component: Positive, negative) × 2 (Consistency of the specific memory with the initial component cue: Consistent, inconsistent) design. Participants completed 8 trials per condition (totaling 64 trials; see Figure 1 ).

For the trait characteristic stimuli we selected 32 positive and 32 negative person-descriptive words from Dumas, Johnson, and Lynch (2002) . The positivity or negativity of the characteristic was determined by likeableness and pleasantness ratings in the Dumas et al. corpus (see Dumas et al., 2002 for details of how ratings were obtained). Positive cues were significantly more likable ( M = 4.97, SD = 0.56) than negative cues ( M = 2.07, SD = 0.50), t (126) = 31.04, p < .001, d = 5.53. Positive cues ( M = 755.74, SD = 59.29) were also more pleasant, (negative M = 223.42, SD = 63.63), t (78) = 38.73, p < .001, d = 8.77. To determine antonyms for each of these characteristics, we searched the characteristic in Roget’s Thesaurus ( Davidson, 2006 ), and selected an antonym listed for the characteristic that was also included in the Dumas et al. (2002) wordlist to allow us to use the Dumas et al. ratings. The experimental stimuli were comparable between conditions, with no significant interactions between valence (i.e., positive or negative) and the task-component the word was used for (i.e., as a cue for the initial component or as a specific memory cue) in predicting Kuèera-Francis word frequency, F (1, 124) = 2.55, p = .11, η p 2 = .02, meaningfulness, F (1, 88) = 0.59, p = .45, η p 2 = .01, and familiarity, F (1, 124) = 0.03, p = .88, η p 2 < .01, according to the MRC Psycholinguistic Database ( Wilson, 1988 ). Cue words were randomly assigned to experimental condition between participants.

The AMPT was presented on a computer and programmed using E-Prime 2.0 software ( Psychology Software Tools, 2012 ). Prior to beginning the test trials, on-screen instructions were presented for the initial component stating that participants would need to either think about how much the personality characteristic described them or to think about the dictionary definition for the word. For the second component, participants were instructed that they would need to recall a specific incident: “ an event that occurred on one particular time and lasted for less than a day, ” and an example was provided. Participants were informed that it was important that they immediately indicate when they had completed each component by pressing any computer key.

For each test trial, a short reminder of the task instructions and the relevant cue were simultaneously presented and remained on the screen until participants selected a key to show that they had completed that component. The time interval between screen presentation and pressing the key was recorded. The time it took for participants to press a key to indicate recall of a specific memory on the second component of the task was our dependent variable. Following completion of the initial component, a blank screen appeared for 1 second, and then the specific memory instructions and cue were presented. To remind participants of task instructions for the second component, after indicating that they had brought a memory to mind, a screen asked, “Was your memory of one particular time/one particular event?” If the participant indicated that it was not a specific incident, they were reminded that they needed to recall a specific incident. The reason for this was that with consistent prompting, depressed individuals are able to overcome their natural difficulty in accessing specific memories ( Dalgleish et al., 2007 ). There was no significant difference in the proportions of initial declared nonspecific responses between the healthy ( M = .04, SD = .06), depressed ( M = .11, SD = .20), and remitted ( M = .13, SD = .20) groups, F (2, 54) = 1.95, p = .152, η p 2 = .07. After a 2-second intertrial interval the next trial was presented.

Prior to the 64 test trials, participants completed eight practice trials (one for each experimental condition) during which they were required to report their memories aloud to ensure that task instructions were understood. After completing the AMPT, participants were presented with each cue used in the experiment, and asked to rate the cue on a 9 point scale from 1 = ‘ extremely unlike me ’ to 9 ‘ extremely like me ’. These ratings allowed us to consider the effect of cue self-relevance on our experimental findings.

Procedure

Ethics approval was obtained from the East of England committee of the NHS National Research Ethics Service. All participants provided informed consent. Participants were informed that they would be completing a number of tasks involving the processing of personal trait adjectives. Participants first completed the experimental task, followed by a short break and then they completed the cue word ratings. Participants next completed the Beck Depression Inventory- Second Edition (BDI-II), a gold-standard self-report measure of depression symptoms ( Beck, Steer, & Brown, 1996 ). To allow us to assess whether verbal ability and fluency affected performance on the experimental tasks, participants also completed the Mill Hill Vocabulary Scale—Synonyms Test Form B ( Raven & John Hugh Court, 1998 ), a measure of acquired verbal knowledge in which participants are required to select a synonym for 34 English words, and the Verbal Fluency Task, which required participants to retrieve as many words as possible in one minute that matched the given category (words begining with the letter A and types of animals; Tombaugh, Kozak, & Rees, 1999 ). Participants received an honorarium of £12 for their time.

Results

The latencies to retrieve specific memories were transformed using a square root transformation (see supplemental materials for details) to remove the positive skew common to response time data ( Bargh & Chartrand, 2014 ; Tabachnick & Fidell, 2001 ). Transformed data were therefore used in all analyses. Data were excluded from three participants in the never-depressed group, one in the remitted and one in the depressed group, due to noncompliance with the experimental protocol (see supplementary materials for details). The analyzed dataset therefore comprised 24 participants in the never-depressed group, 17 in the remitted group, and 16 in the depressed group.

As anticipated, the depressed group had higher BDI-II scores (descriptive statistics in Table 1 ) than the remitted, p < .001, d = 0.58, and never-depressed, p < .001, d = 2.91, groups, F (2, 54) = 38.74, p < .001, η p 2 = .64. As further anticipated, the remitted group also demonstrated significantly more symptoms on the BDI-II than never-depressed participants, p = .009, d = 1.25. The mean BDI-II score indicated a ‘moderate’ level of symptom severity for the depressed group and a ‘minimal’ level of symptom severity for the never-depressed and remitted groups ( Beck et al., 1996 ). In terms of the number of previously experienced depressive episodes, 12 of the currently depressed participants had experienced too many depressive episodes to count the number of distinct episodes (this represents one of the coding categories on the SCID; First et al., 1996 ). This was also the case for three of the remitted individuals. For those remitted participants able to identify the number of previous episodes, the mean number of episodes was 4.40 ( SD = 3.11), consistent with the experience of recurrent depression. Age did differ between the never-depressed and remitted, p < .001, and depressed conditions, p < .001, F (2, 59) = 21.72, p < .001. Although age was not associated with performance on any of the outcome measures, r s ( n = 62) < −.10, p s > .423, we repeated our key between-groups analyses with age included as a covariate and the results were unchanged (and in fact the effects were stronger). All groups were comparable in terms of scores on the Mill Hill Vocabulary test, F (2, 54) = 1.88, p = .163, η p 2 = .05, and the Verbal Fluency Task F (2, 54) = 1.30, p = .280, η p 2 = .07.

Hypothesis Testing

Figure 2 presents the response times (square root transformed) for recall of an inconsistent specific memory across the different groups and task components. Complete response time data for recall of a specific memory across all of the different AMPT conditions is presented, by group, in supplementary Table 1 . If the speed of access to a specific memory is an index of its capacity to delimit the scope of a preceding trait self-rating for which the memory is inconsistent, then we would expect response time on inconsistent memory trials to be shorter when the initial component involved a trait generalization self-rating (relative to the dictionary definition comparison condition) - the priming effect ( Klein et al., 2001 ). As outlined in the Introduction, we had two key hypotheses. We expected that for the healthy, never-depressed group, this priming effect would be greater following negative trait self-ratings than for positive trait self-ratings. For the depressed group, we expected the opposite with the priming effect being greater following positive trait-self-ratings than following negative trait self-ratings. We had no a priori prediction for the remitted group.

We initially conducted an omnibus 3 (Group: Never-depressed, remitted, depressed) × 2 (Valence of the trait self-rating: Positive, negative) × 2 (Initial component: Self-rating, dictionary definition) × 2 (Consistency of the specific memory: Inconsistent, consistent) mixed-model ANOVA with time taken to recall a specific memory (square root transformed) as the dependent variable. We observed main effects of task, F (1, 54) = 4.17, p = .046, η p 2 = .07, and consistency, F (1, 54) = 3.97, p = .051, η p 2 = .07, and an interaction between task and consistency, F (1, 54) = 5.29, p = .025, η p 2 = .09, all qualified by a significant four-way interaction, F (2, 54) = 5.38, p = .007, η p 2 = .17, consistent with our specific predictions. 2 None of the other main effects or interactions reached significance F s < 2.61, p s > .083.

We examined this four-way interaction between each pairing of the three groups. Performance on the AMPT was significantly different between never-depressed and depressed conditions, F (1, 39) = 9.74, p = .003, η p 2 = .20, as expected. However the remitted group did not differ from either the never-depressed, F (1, 40) = 1.34, p = .253, η p 2 = .03, or the depressed groups, F (1, 32) = 3.76, p = .061, η p 2 = .11.

Having established that there was a significant differential effect across the depressed and never-depressed groups, we next examined each of these two groups separately to explore our specific hypotheses. For the never-depressed group, there was a significant 3-way interaction of Valence × Task × Consistency, F (1, 23) = 5.32, p = .031, η p 2 = .19, consistent with the differential pattern of task performance across positive and negative trials that we predicted (see Figure 2 ). Breaking this effect down by valence, a Significant Task × Consistency interaction was observed for negative trials, F (1, 23) = 4.45, p = .046, η p 2 = .13. In line with faster access to inconsistent specific memories delimiting the scope of the prior negative self-rated trait, response time on inconsistent trials was significantly shorter following a negative trait self-rating, relative to a definition of a negative trait, t (23) = 2.13, p = .044, d = 0.24 (see Figure 2 ). This difference was not evident on consistent trials, t (23) = 0.15, p = .882, d = 0.02. The Task × Consistency interaction was nonsignificant for positive trials, F < 1.

For the depressed group there was a near significant 3-way interaction of Valence × Task × Consistency, F (1, 15) = 4.28, p = .056, η p 2 = .22. Deconstructing this by valence, we found a Significant Task × Consistency interaction for positive trials, F (1, 15) = 6.57, p = .022, η p 2 = .31, but not for negative trials, F < 1. As predicted, response time on inconsistent trials was significantly shorter when following a self-rating of a positive trait than when following a dictionary definition of the trait, t (15) = 4.37, p = .001, d = 0.35. Again, this effect was not significant for consistent trials, t (15) = 1.69, p = .112, d = 0.15.

Although the pattern of performance in the remitted group was not significantly different to that of either of our other groups, we nevertheless examined the profile of task performance within this group. However, unlike the depressed and never-depressed groups, we found no support for a 3-way interaction of Valence × Task × Consistency, F < 1.

Discussion

Klein and colleagues’ Scope Hypothesis ( Klein et al., 2001 , Klein, Cosmides, Tooby, & Chance, 2002 ) argues that when we retrieve personal semantic information ( Tulving, 2002 ) that has context-dependent veridicality, memories of specific autobiographical episodes that represent exceptions to that information are primed such that the boundary conditions or ‘scope’ for when the semantic information is true, versus when it is invalid, can be delimited. For instance, a self-referent trait judgment such as “ Am I friendly ?” would involve retrieval of generic personal semantic information about my overall degree of friendliness, paralleled by priming of memories of specific episodes when my level of friend-liness had deviated significantly from this personal norm. The theoretical rationale is that any increases in speed of cognizing that are leveraged by the rapid retrieval of generic self-referent trait information would be offset by limitations in the appropriateness of that information in any given situation. However, accuracy can be preserved when the boundary conditions on the validity of the generic trait information—it’s scope—are set in place by primed episodic memories of exceptions. These specific memories are slower to retrieve, thus creating an overall speed–accuracy trade-off ( Klein et al., 2002 ).

The question we wanted to examine was why, if the system is thus optimized for accuracy, are self-relevant trait judgments routinely biased? This is the case not only in those with mental health problems, such as depression ( Dalgleish & Werner-Seidler, 2014 ), who are characterized by biased and maladaptive views of the self, but also in healthy individuals who are characterized by biased overly positive self-views ( Taylor & Brown, 1988 ).

We hypothesized that such biases in trait judgments may in part result from asymmetries in the aforementioned episodic memory priming effects, such that for those experiencing depression, there would be greater priming of negative episodic memories that were thus inconsistent with positive self-referent traits, than vice versa. We predicted the opposite pattern in healthy participants. We also examined performance in individuals with a history of depression who were in remission, but for whom we had no a priori predictions.

Our results using the memory priming paradigm established by Klein ( Klein et al., 2001 , 2002 ) provided support for both hypotheses. In individuals with depression, recollection of countertrait specific negative memories was relatively faster following their evaluation of a positive trait (relative to the definition control condition), compared to the other way around. This contrasted with never-depressed controls who exhibited relatively faster recollection of positive countertrait memories following evaluation of a negative self-referent trait, compared to the other way around. Individuals remitted from depression did not appear to exhibit a significant priming effect following evaluation of either positive or negative self-traits, suggesting an attenuation of the bias evident in those currently experiencing a depressive episode, but also a potential lack of reinstatement of the bias associated with stable mental health.

The present findings, as discussed in the Introduction, suggest that the relatively reduced accessibility of specific negative information necessary to set realistic constraints on the “scope” ( Klein et al., 2002 ) of positive self-generalizations may be one potential process supporting the ubiquitous positive self-biases associated with mental health ( Taylor & Armor, 1996 ; Taylor & Brown, 1988 ). In contrast, relatively “unbounded” ( Klein et al., 2002 ) negative self-generalizations in depression may be one process that fosters, maintains, and amplifies globalized negative self-referent thinking patterns (including overgeneralization, catastrophization, negative self-attributions, disqualification of positive information), as well as perpetuating rumination ( Rimes & Watkins, 2005 ) - factors that are known to exacerbate depression severity and persistence and that have been shown to permeate daily cognition in depression sufferers ( Beck, 1967 ; Beck et al., 1979 ). Interventions designed to ameliorate disturbance in the accessibility of autobiographical memories (e.g., by improving accessibility of specific memories of positive events; cf. memory specificity training; Raes, Williams, & Hermans, 2009 ) may thereby help to shift the over-generalized negative thoughts and beliefs that are characteristic of depression. These interventions have shown considerable promise in alleviating depressive symptoms (for review see Hitchcock, Werner-Seidler, Blackwell, & Dalgleish. 2017 ) and the current results suggest that these interventions may work in part through reinstating memory retrieval patterns that help to set ‘boundaries’ on negative self-beliefs. Further exploration of this potential mechanism may help to elucidate whether interventions translated from basic science offer a low-intensity option for shifting maladaptive cognitions targeted in more complex therapies (e.g., cognitive therapy).

Although our results are consistent with the theoretical notions advanced by Klein et al. (2002) , it is important to note that priming of countertrait specific memories following a trait judgment does not itself directly demonstrate that those trait generalizations have been bounded by the retrieved specific event memories. At the very least, one would want to show that trait judgments systematically differ when putative relevant boundaries are in place relative to when they are not. However, to demonstrate this a methodology that permits the manipulation of these mnemonic boundaries would be required, and it is not immediately clear what such a methodology would look like. It is also unclear whether the specific memories actually need to be explicitly retrieved as episodic recollections in order to exert