Self-Serving Bias in Memories.

Protecting one's positive self-image from damage is a fundamental need of human beings. Forgetting is an effective strategy in this respect. Individuals show inferior recall of negative feedback about themselves but unimpaired recognition of self-related negative feedback. This discrepancy may imply that individuals retain negative information but forget that the information is associated with the self. In two experiments, participants judged whether two-character trait adjectives (positive or negative) described themselves or others. Subsequently, they completed old-new judgments (Experiment 2) and attribution tasks (Experiments 1 and 2). Neither old-new recognition nor source guessing bias was influenced by word valence. Participants' source memory was worse in the negative self-referenced word processing condition than in the other conditions. These results suggest there is a self-serving bias in memory for the connection between valence information and the self.

Viewing oneself as positive is a vital human need (Leary, 2007). However, individuals in daily life frequently receive negative comments from friends, relatives, employers, and even strangers. Therefore, one’s positive self-image may often be threatened. Consequentially, how to protect one’s self-image from damage is a critical problem.

In this respect, the first choice is to seek affirmative information and to avoid disapproving information. Some studies have shown that individuals selectively attend to positive information about themselves (Ditto & Lopez, 1992) and report feeling better after receiving positive feedback (Kwang & Swann Jr., 2010). When asked to rate the degree to which they expected to receive four types of feedback (self-enhancing, self-effacing, self-improving, and no-feedback) from others, both Chinese and Americans were uninterested in self-effacing feedback (Gaertner, Sedikides, & Cai, 2012). In contrast, Baumeister and Ilko (1995) asked participants to furnish public or private accounts of their recent experiences of success. Compared with private accounts, public accounts contained more references to receiving aid from others. This suggests that self-presentation prompts individuals to consider possible reactions from other people and to behave modestly in public situations (Baumeister & Ilko, 1995; Gould, Brounstein, & Sigall, 1977). In short, people actively pursue information that enhances the self, but evade information that hurts the self (Gregg, Sedikides, & Gebauer, 2011; Sedikides, 1993). Unfortunately, reality does not always follow the will of human beings. In the long term, everyone will receive self-threatening feedback. In response to this frustrating situation, individuals can choose another coping strategy to maintain a positive self-image: selectively forgetting information that has disadvantageous implications for the self.

Since self-threatening information is usually upsetting, individuals try their utmost to erase such information from memories (Anderson & Green, 2001; Gagnepain, Henson, & Anderson, 2014; Kim & Yi, 2013). One of the most powerful lines of evidence is autobiographical memory research, which indicates that it is more difficult to recall negative life-events than positive life-events. For example, in one study, participants were asked to remember real-life behaviors at Time 1 and recall them at Time 2. If the target was the self, participants recalled a significantly lower proportion of negative than positive behaviors (Ritchie, Sedikides, & Skowronski, 2017).

Researchers have systematically examined the selective forgetting of self-threatening negative feedback – the mnemic neglect effect (MNE). In an initial experiment by Sedikides and Green (2000), participants read either self- or other-referent mixed-valence lists of evaluations that consisted of central and peripheral trait-relevant behaviors. Then, participants were given a surprise free-recall task. Participants in the self-referent condition recalled fewer central-negative behaviors than those in the other-referent condition, and self-referential participants recalled fewer central-negative than central-positive behaviors. Mnemic neglect occurred both in mundane reality settings, wherein evaluations were based on a purportedly reliable and valid personality inventory, and in minimal-feedback settings, wherein feedback was hypothetical (Sedikides & Green, 2000; Sedikides, Green, Saunders, Skowronski, & Zengel, 2016).

As is commonly appreciated, feedback usually contains two aspects: the referent and the content. The referent refers to the person who is evaluated. The content refers to the valence information (positive or negative) implied in the feedback. Although there are many studies exploring the memory of self-threatening negative information (Green, Sedikides, & Gregg, 2007; Newman, Sapolsky, Tang, & Bakina, 2014; Sedikides & Green, 2004), no researchers have explored the memory of connections between the valence information and the referent (self or others) that were developed while processing feedback. How do individuals perceive the connection between valence information and themselves, especially when the information is obviously negative? Are they likely to establish a weaker connection when the information is self-threatening? Do they have a preference for generating a connection with positive information? More generally, what strategy do individuals use to protect the self-image from negative feedback, when addressing the relationship between the self and self-threatening information?

Evidence from several sources suggests that individuals weaken their connection with negative information or strengthen their connection with positive information. According to psychodynamics, to narrow the gap between the actual and ideal self, defense mechanisms (e.g., denial, projection) inhibit awareness of unfavorable events, especially when those events are highly self-threatening (Baumeister, Dale, & Sommer, 2010). In addition, studies indicate there is a self-serving bias in attribution, which is helpful for maintaining one’s positive self-image and even physical well-being (Hu, Zhang, & Yang, 2015; Sanjuán & Magallares, 2014). Moreover, people make assimilation-based judgments about positive self-related information, which leads positive information to be integrated with stored self-knowledge. However, individuals make contrast-based judgments about negative self-related information, which leads threatening information to be separated from stored self-knowledge (Pinter, Green, Sedikides, & Gregg, 2011). All of these results imply that individuals prefer to sever their connection with negative information. Therefore, we expected that participants would poorly memorize the connection between negative information and the self.

The Current Study

We borrowed and modified the classic self-reference paradigm (Conway & Dewhurst, 1995; Rogers, Kuiper, & Kirker, 1977) to address the above hypothesis in two experiments. Both experiments consisted of two phases. During the learning phase, participants were asked to process trait words in both the self-reference and other-reference conditions. During the testing phase, memory performance was measured. In the first experiment, the test asked participants to attribute each word they processed to the word’s referent. If participants forgot the word’s referent, they could choose “not remember” as their response. In the second experiment, the test first asked participants to judge whether the words presented were “new” or “old.” If a word was judged as “old,” the test also required participants to attribute the “old” word to its referent. Collectively, with the two experiments we elucidated the memory of the connections generated while processing positive and negative words in both self- and other-reference conditions.

Experimental Materials

To obtain valence, arousal, familiarity, and meaningfulness measurements of the stimulus words, pretests were conducted with individuals who did not participate in the main experiments. First, 413 two-character trait words were selected from a pool of Chinese personality-trait adjectives (Wang & Cui, 2005) and randomly divided into two groups (the first group contained 210 words, while the second contained 203 words). Second, 66 college students were recruited to rate the words (33 raters for the first-group words, 33 raters for the second-group words) in terms of valence using a 7-point Likert scale (1 = extremely bad to 7 = extremely good); 66 college students rated the words (33 raters for the first-group words, 33 raters for the second group) in terms of arousal (1 = extremely unexciting to 7 = extremely exciting); 64 college students rated the words (32 raters for the first-group words, 32 raters for the second group) in terms familiarity (1 = extremely unfamiliar to 7 = extremely familiar); and 60 college students rated the words (30 raters for the first-group words, 30 raters for the second group) in terms of meaningfulness (1 = extremely meaningless to 7 = extremely meaningful). Each student only rated one dimension of the words. Finally, 40 positive and 40 negative trait words were selected as the experimental materials for the first experiment and 80 positive and 80 negative trait words were selected as the experimental materials for the second experiment. Trait words of both experiments are provided in ESM 5.

Experiment 1

Past research has established that individuals selectively forget negative information about themselves, relative to positive information. However, memory of connections between positive or negative information and the self or other entities has not been examined. The first experiment examined participants’ overall attribution of words to their referents. The attribution is likely to include memory components and judgment bias.

Participants

Thirty-nine Chinese university students (27 females, 12 males; aged 18–24 years, Mage = 20.18 years) were recruited as paid volunteers. All participants were right-handed, native Chinese speakers with normal or corrected-to-normal vision. None of the participants had a history of, or currently reported, neurological or psychiatric conditions.

Design and Materials

The experiment used a two-factorial design that varied both the valence (positive vs. negative) and the referent (self vs. other) within subjects. The number of correct responses each participant produced during word attribution was used as the dependent variable.

The experimental stimuli were 80 two-character trait adjectives, consisting of 40 positive words and 40 negative words. Positive words were significantly more desirable than negative words, t(78) = 43.84, p < .001. These words were divided into two lists. Each list contained 20 positive words and 20 negative words. Word valence was matched between the two lists. In addition, word arousal, word familiarity, and word meaningfulness were matched both within and between lists. One list was assigned to the self-reference condition, and the other list was assigned to the other-reference condition. The two lists were exchanged between conditions across participants, so that trait words that appeared in the self-reference condition for half of the participants appeared in the other-reference condition for the other participants.

Procedure

Upon arrival in the lab, participants were given a brief overview of the study and signed the informed consent form. Then, participants completed the Beck Depression Inventory (M = 8.33, SD = 4.84) (Beck, Ward, Mendelson, Mock, & Erbaugh, 1961), the State-Trait Anxiety Inventory (Mstate = 35.54, SDstate = 6.81; Mtrait = 43.62, SDtrait = 9.90) (Spielberger, Gorsuch, Lushene, Vagg, & Jacobs, 1983) and the Rosenberg Self-esteem Scale (M = 27.97, SD = 4.06) (Rosenberg, 1979). Finally, the primary experimental task was started.

The primary experiment was conducted via E-prime 2.0 (retrieved from https://pstnet.com/welcome-to-e-prime-2-0/). During the learning phase, participants were asked to perform a two-choice judgment task of individually presented trait words paired with referent cues (Figure 1). In the self-reference condition, each trait word was presented with the first-person pronoun “I” (我), and participants were asked to indicate whether the word could be used to describe themselves. In the other-reference condition, each trait word was presented with the third-person pronoun “he” (他), and participants were asked to indicate whether the word could be used to describe others. All stimuli were randomly presented on a computer monitor in white font centered on a gray background. Participants were asked to indicate their choice by pressing the appropriate “d” or “k” key with one of their index fingers. The assignment of keys to response alternatives was balanced across participants. After completion of the study task, participants were asked to count backward from 100 to 1 aloud for 5 min to prevent maintenance rehearsal.

Figure 1

Each trial started with a fixation (200 ms), followed by a blank screen (500 ms). Next, a target adjective (positive/negative) was presented with a personal pronoun (I/he) until a response was given. Then, 500 ms after the response, the next trial started.

During the testing phase, the 80 words were randomly presented, one word at a time, on the computer screen. Participants needed to attribute the words to their referents (“I “or “he”) by pressing the appropriate “d” or “k” key. If they could not remember which personal pronoun (“I “or “he”) the trait word had been paired with, they were to press the space bar. It should be noted that, before the memory test was officially started, participants did not know there was a test.

Results and Discussion

The raw data of both experiments is provided in ESM 1–4. Attribution counts were the numbers of the trait words that were correctly attributed to their referents in each condition by each participant (Table 1). Attribution counts were subjected to a 2 (Valence: positive vs. negative) × 2 (Referent: self vs. other) ANOVA. The interaction between referent and valence was significant, F(1, 38) = 19.25, p < .001, η2 = .336. In the self-reference condition, participants correctly attributed fewer negative words than positive words, t(38) = 5.09, p < .001. In the other-reference condition, participants correctly attributed more negative words than positive words, t(38) = 2.45, p < .05. For positive words, participants correctly attributed more self-referenced words than other-referenced words, t(38) = 7.97, p < .001. However, for negative words, there was no significant difference between the attribution of self-referenced words and that of other-referenced words, t(38) = 0.05, p > .05. These results suggest that (1) participants more accurately attributed negative other-referenced words than positive other-referenced words, but less accurately attributed negative self-referenced words than positive self-referenced words; (2) self-reference processing can improve the attribution of positive information rather than that of negative information.

Table 1

Participants’ behavioral performance in the testing phase in Experiment 1

	Participants’ responses
Actual referent during learning	Attributed to “I”	Attributed to “He”	Do not remember
Note. Rows represent presentation during learning, columns denote the response of participants, and cells contain average numbers (and standard deviations).
Referent “I”
Positive	14.97 (2.87)	3.33 (2.85)	1.69 (1.70)
Negative	10.92 (4.05)	6.05 (3.95)	3.03 (3.41)
Referent “He”
Positive	7.23 (3.60)	9.21 (3.16)	3.56 (2.98)
negative	5.13 (3.32)	10.97 (3.92)	3.90 (3.37)

The experimental outcome is consistent with attribution studies, which suggest that individuals are always ready to separate themselves from negative information or to link themselves to positive information (Mezulis, Abramson, Hyde, & Hankin, 2004; Seidel et al., 2010). When negative information is unavoidable, one can achieve self-protection by severing the connection between negative information and the self.

It is important to note that there are two other possible reasons for this experimental outcome, in addition to the memory of the connections between trait words and their referents. The first reason is that participants might have forgotten more negative self-referenced words than positive self-referenced words, and thus they failed to recognize whether the word had been presented. The second reason is that there might have existed judgment bias in cases where the referent was forgotten, and thus participants attributed more positive words than negative words to the self.

Experiment 2

The second experiment examined the memory effect and judgment bias simultaneously. We added some “new” words in the testing phase, and participants needed to judge whether each word was “old” or “new” and to attribute the words judged as “old” to their referents (“I” or “he”).

Seventy Chinese university students (54 females, 16 males; aged 18–24 years, Mage = 20.63 years) were recruited as paid volunteers. All participants were right-handed, native Chinese speakers with normal or corrected-to-normal vision. None of the participants had a history of, or currently reported, neurological or psychiatric conditions. No participant took part in the first experiment.

The experiment used a two-factor design that varied both the valence (positive vs. negative) and the referent (self vs. other) within subjects.

A set of 160 two-character trait adjectives (80 positive, 80 negative) was selected as the experimental stimulus. These words were divided into two lists, with 80 words (40 positive, 40 negative) in each list. The two lists were equated in terms of valence, arousal, familiarity, and meaningfulness. There was no significant difference between positive and negative words in terms of arousal, familiarity, and meaningfulness in each list. One of the two lists was used as study list. Half of the study list words (20 positive, 20 negative) were assigned to the self-reference condition, and half to the other-reference condition. There was no significant difference between the two halves in terms of valence, arousal, familiarity, and meaningfulness. The two halves were exchanged between conditions across participants, so that trait words that appeared in the self-reference condition for half of the participants appeared in other-reference condition for the other participants. All 160 trait words were used in the test phase.

Procedure

Upon arrival to the lab, participants were given a brief overview of the study and signed the informed consent form. Then, participants completed the Beck Depression Inventory (M = 8.26, SD = 4.64) (Beck, Ward, Mendelson, Mock, & Erbaugh, 1961), the State-Trait Anxiety Inventory (Mstate = 34.99, SDstate = 9.02; Mtrait = 40.54, SDtrait = 8.49) (Spielberger, Gorsuch, Lushene, Vagg, & Jacobs, 1983), and the Rosenberg Self-esteem Scale (M = 27.91, SD = 2.76). Finally, the primary experimental task was started.

The primary experiment was conducted via E-prime 2.0. The learning phase was the same as in Experiment 1. For the test, the full set of 160 trait adjectives was randomly presented, one word at a time, on the computer screen (Figure 2). Participants were instructed to indicate whether the word had been presented by pressing the appropriate “d” or “k” key. When a trait adjective was judged as an “old” word, participants then needed to indicate which personal pronoun (“I” or “he”) had been paired with it by pressing the appropriate “d” or “k” key. In the case of “new” judgments, 500 ms after the response, the next trial started. The assignment of keys to response alternatives was balanced across participants.

Figure 2

Each trial started with a fixation (200 ms), followed by a blank screen (350 ms). Then, an adjective (old/new) was presented, and the participant pressed keys representing “old” or “new.” If the word was judged as “old,” the participant needed to indicate which personal pronoun (I or he) had been paired with it. Otherwise, the word was judged as “new.” The next trial started 500 ms after the response.

To disentangle item memory, source memory, and guessing biases, we applied the two-high threshold source memory (2HTSM) model (Bayen, Murnane, & Erdfelder, 1996; Bröder & Meiser, 2007; Vogt & Bröder, 2007). The 2HTSM model is depicted in Figure 3. There is one processing tree for each of six classes of words (i.e., Source I positive words, Source I negative words, Source He positive words, Source He negative words, new positive words, and new negative words). Source I positive (Source I negative, Source He positive, Source He negative) words will be detected as “old” with probability DIp (DIn, DHp, DHn). New positive (negative) words will be detected as “new” with probability DNp (DNn). If old words are detected as old, their source can either be remembered (dIp, dIn, dHp, dHn) or not (1 − dIp, 1 − dIn,1 − dHp,1 − dHn). In the latter case, there is a guessing probability ap (an) of guessing “Source I” and 1 − ap (1 − an) of guessing “Source He.” If old words are undetected, they can nevertheless be guessed as “old,” depending on bias bp (bn). Since the participant has to judge a source in this case, there is another guessing probability gp (gn) to guess “Source I” and 1 − gp (1 − gn) to guess “Source He.” A “new” response only results for old words if they are neither detected nor guessed as old. If new words are not detected as new, they can be guessed as “old,” and the source guessing probability gp (gn) also applies.

Figure 3

Processing tree depicting the 2HTSM model. The words are either from Source I, Source He, or they are new. DIp/DIn/DHp/DHn = probabilities of detecting Source I/He words as old; dIp/dIn/dHp/dHn = probabilities of remembering a word’s Source; ap/an = probabilities of guessing “Source I” when the source was forgotten; bp/bn = probabilities of guessing “old” if a word is not detected as old; gp/gn = probabilities of guessing “Source I” when a word was guessed as old; DNp/DNn = probabilities of detecting a new word as new.

Considering that self-referential processing may promote participants’ old-new recognition and the promotion may be influenced by words’ valence, we let DIp and DIn vary freely. In addition, a fully restricted model (all D equal) did not fit the data, but leaving DIp and DIn free led to a very good model fit. Therefore, the final restrictions on parameters were as follows: DHp = DHn = DNp = DNn; ap = gp; an = gn.

The model-based analyses were conducted based on the aggregated data presented in Table 2. Parameter estimates and the goodness-of-fit index G2 were obtained via the EM algorithm using the multiTree v046 (Moshagen, 2010). The parameter estimates and confidence intervals are shown in Table 3. The model fit the data, G2(1) = 0.06, p = .81.

Table 2

Participants’ behavioral performance in the testing phase in Experiment 2

	Participants’ responses
Actual source during learning	Attributed to “I”	Attributed to “He”	Do not remember
Note. Rows represent presentation during learning, columns denote the response of the participant, and cells contain raw frequencies.
Source “I”
Positive	922	165	313
Negative	713	310	377
Source “He”
Positive	365	527	508
Negative	274	584	542
New words
Positive	438	309	2,053
Negative	404	290	2,106

Table 3

Parameter estimates and confidence intervals for the baseline model

Parameters	M	SD	95% CI
Note. M = mean; SD = standard deviation; CI = confidence interval.
DIp	0.61	0.02	[.57–.65]
DIn	0.56	0.02	[.52–.60]
DNn = DNp = DHp = DHn	0.37	0.01	[.35–.39]
dIp	0.80	0.04	[.72–.89]
dIn	0.36	0.06	[.24–.48]
dHp	0.52	0.06	[.40–.64]
dHn	0.75	0.06	[.64–.87]
bp	0.42	0.01	[.40–.45]
bn	0.39	0.01	[.37–.41]
ap = gp	0.59	0.02	[.55–.62]
an = gn	0.58	0.02	[.55–.62]

Using this model as a baseline model, model fit significantly decreased when dIp, dIn, dHp, and dHn were restricted to be equal, ΔG2(3) = 37.91, p < .001, suggesting an overall effect of learning condition on source memory. Then, we conducted pair-wise comparisons among parameters dIp, dIn, dHp, and dHn. The parameter dIn was significantly smaller than dIp, ΔG2(1) = 36.68, p < .001, suggesting that in the self-reference condition, participants had worse source memory for negative words than positive words. The parameter dHn is significantly larger than dHp, ΔG2(1) = 7.98, p < .01, suggesting that in the other-reference condition, participants had better source memory for negative words than positive words. The parameter dIn is significantly smaller than dHn, ΔG2(1) = 16.57, p < .001, suggesting that for negative words, participants had worse source memory in the self-reference condition than in the other-reference condition.

For source guessing parameters, the parameter gp was significantly larger than 0.5, ΔG2(1) = 22.39, p < .001; the parameter gn was also significantly larger than 0.5, ΔG2(1) = 18.81, p < .001; and there was no significant difference between parameter gp and gn, ΔG2(1) = 0.03, p = .87, suggesting that there was a general guessing bias to attribute words (although judged as old, participants did not remember their source) to the self, but the guessing bias was not influenced by words’ valence.

For old-new recognition parameters, the parameter DIp was significantly larger than DHp, ΔG2(1) = 82.22, p < .001; the parameter DIn was also significantly larger than DHn, ΔG2(1) = 51.81, p < .001; and there was no significant difference between parameter DIp and DIn, ΔG2(1) = 3.39, p = .07, suggesting that self-reference processing promoted participants’ old-new recognition performance, but this promotion was not influenced by words’ valence.

Although there was a general source guessing bias, this bias was not influenced by words’ valence. Hence, source guessing bias is an unlikely candidate for the result in Experiment 1. Self-reference processing promoted old-new recognition performance, but this promotion was not influenced by words’ valence. Hence, it is also unlikely that the attribution differences in Experiment 1 were caused by old-new recognition. Participants had inferior source memory in the negative self-referenced words learning condition compared to other learning conditions, suggesting that individuals selectively forget the connection between negative information and oneself.

General Discussion

The present study explored individuals’ memory of the connection between positive or negative words and their referents. In the self-reference condition, the memory of the connection between negative words and the self was significantly worse than that between positive words and the self. For negative words, the memory of the connection between negative words and the self was significantly worse than that between negative words and the others. What could explain the inferior memory of the connection between negative information and oneself?

One possibility is the role of self-protection motivation in memory. Self-protection refers to the motivation for defending positive components of the self-image against harm (Sedikides, Skowronski, & Gaertner, 2004). Self-protection motivation not only guides people to avoid self-threatening information, but also influences people’s memory of self-related information by processing self-threatening information in a shallow manner (Alicke & Sedikides, 2009; Sedikides & Green, 2009; Sedikides et al., 2016). As a result, the memory of self-threatening information is weakened. Autobiographical memory research indicates that unpleasant life experiences are recalled more poorly than pleasant life experiences (Ritchie et al., 2017; Skowronski & Walker, 2004; Skowronski, Betz, Thompson, & Shannon, 1991). Moreover, individuals selectively forget evaluations that have unfavorable implications for the self (Green et al., 2007; Newman et al., 2014; Pinter et al., 2011; Sedikides & Green, 2000, 2004; Sedikides et al., 2016). In the present study, to achieve the purpose of self-protection, the individual may avoid building up a connection with negative information by allocating less cognitive resources to the connection. Consequently, negative information becomes weakly connected with the self in memories.

Individuals may process self-related positive and negative information not only to different degrees, but also in different ways. Previous studies indicate that people have poor recall of self-threatening information compared to nonthreatening information (Newman, Nibert, & Winer, 2009; Newman et al., 2014; Sedikides & Green, 2000, 2004, 2009; Sedikides et al., 2004). However, they show equivalent recognition for threatening and nonthreatening information (Green et al., 2007). These findings imply that people process positive and negative self-related evaluations in different ways: they make assimilation-based judgments about positive evaluations (“they are suitable to describe me”), but contrast-based judgments about negative evaluations (“they are not suitable to describe me”). Therefore, self-related positive evaluations would be integrated into stored self-knowledge, but negative evaluations would be separated from stored self-knowledge. This claim is supported by a study in which participants received instructions designed to make them integrate some items of behavioral information, but separate other items. The integration instruction led to better information recall than did the separation instruction. However, the manipulation had no impact on levels of recognition (Pinter et al., 2011). In our experiments, there was no significant difference in recognition performance for positive and negative self-referenced words. This is consistent with the previous study that showed self-protectively neglected information can be retrieved (Green et al., 2007). Nonetheless, participants’ source memory was extremely poor in the negative self-referenced words learning condition. This is probably because people integrate self-related positive information into stored self-knowledge but keep negative information separate.

When processing self-referenced information, participants’ source memory is worse for negative information than positive information. When processing negative information, participants’ source memory is worse for self-referenced information than other-referenced information. These results suggest that there is a self-serving bias in memory for the connection between valence information and the self.

The electronic supplementary material is available with the online version of the article at https://doi.org/10.1027/1618-3169/a000409

Raw data of Experiment 1 (part 1).

Raw data of Experiment 1 (part 2).

Raw data of Experiment 2 (part 1).

Raw data of Experiment 2 (part 2).

Stimulus material and word properties.