(Eye-)tracking the escape from the self: guilt proneness moderates the effect of failure on self-avoidance

ABSTRACT Failure increases the motivation to escape self-awareness. To date, however, the role of self-conscious emotions (shame and guilt) in triggering escape responses after failure has not been sufficiently addressed. In this pre-registered study (N = 156 undergraduates), we adapted a classic paradigm (avoidance of one’s image in a mirror) to a modern eye-tracking technology to test the hypothesis that shame proneness moderates the effect of failure on self-awareness avoidance. Individual differences in guilt and shame proneness were assessed before priming thoughts of failure or success. Then, an eye-tracking paradigm was used to monitor gaze avoidance of one’s screen-reflected face during a neutral, unrelated task. Unexpectedly, results showed that guilt but not shame proneness exacerbated self-avoidance after failure. The present findings challenge the dominant view that shame fosters avoidance more so than guilt.

The mere act of focusing on one's self can be highly aversive to some individuals, such as those suffering from body dysmorphic or eating disorders (Shafran et al., 2004;Toh et al., 2017;Veale & Riley, 2001), low self-esteem (Brockner & Wallnau, 1981) or borderline personality disorder (Winter et al., 2015). However, self-awareness can be aversive to most individuals in certain unpleasant situations, as for example, after a major failure or setback. In effect, failure widens the separation of one's actual self from one's ideal self and triggers escape thoughts and behaviours (Baumeister, 1990;Chatard & Selimbegović, 2011;Duval & Wicklund, 1972;Gibbons et al., 1985;Higgins, 1996;Steenbarger & Aderman, 1979). Given the myriad of negative clinical consequences associated with escapist behaviours (e.g. Baumeister, 1990;Blackburn et al., 2006;Deleuze et al., 2019;Heatherton & Baumeister, 1991), understanding when and why failure induces self-focus avoidance is an important scientific endeavour. In this perspective, the aim of the present study was to better understand the role of individual differences in guilt and shame proneness in the effect of thoughts of failure on escape behaviours.

Guilt and shame
Shame and guilt are both negative self-conscious emotions associated to self-evaluation after a failure or a transgression. Although often confounded, it has been suggested that shame and guilt can be distinguished on the basis of the specific object of evaluation that elicits these emotions (Tracy & Robins, 2004. In this view, while guilt results from negative evaluation of a specific behaviour, evaluated as controllable and unstable (I did a bad thing, but I can control it and it won't necessarily happen again in the future), shame results from negative evaluation of the whole self, evaluated as uncontrollable and stable (I am a bad person, there is nothing I can do about it, and it is not likely to change). Dominant accounts suggest that shame should be associated with escape and withdrawal more than guilt, whereas guilt is believed to elicit repair and approach tendencies (Lewis, 1971;Tangney & Dearing, 2002). This view was rarely questioned until recently.
Arguments suggesting that shame is an avoidanceoriented and guilt an approach-oriented emotion can be found in both clinical and social psychology literature (Cunningham et al., 2019;Keltner & Harker, 1998;Schmader & Lickel, 2006;Sheehy et al., 2019;Sheikh & Janoff-Bulman, 2010, Study 1; Tangney & Tracy, 2012). However, authors mostly rely on correlational studies and/or self-reported measures to support this view, while experimental evidence supporting this stance is relatively scarce. Rebega et al. (2013) noted that, as of 2013, only seven studies in the literature experimentally manipulated guilt, and their analyses did not account for studies orthogonally manipulating both shame and guilt.
To our knowledge, the most extensive experimental tests of this link between shame and guilt and the ensuing action tendencies stem from de Hooge's work (de Hooge et al., 2008(de Hooge et al., , 2011(de Hooge et al., , 2018. However, inconsistent with the dominant view that shame would be associated to avoidance, her results suggest that shame could be associated with social approach if such a behaviour has the potential to restore a damaged self. Baumeister, in yet another account contrasting with the dominant view of shame and guilt's respective associations to avoidance and approach, suggested that guilt is the key self-conscious emotion involved in escape responses, as guilt relates to self-blame (although it must be noted that Baumeister endorsed in his later work the idea that guilt should have prosocial consequences, e.g. Baumeister et al., 1994;Leith & Baumeister, 1998). In an attempt to reconcile the diverging views about reactions to guilt in terms of approach and avoidance, Amodio et al. (2007) suggested in their dynamic model of guilt that avoidance is an early reaction to guilt, in order to limit the damage associated to the transgression or failure, while approach tendencies are set in motion when opportunity arises to repair the harm done.
In sum, the literature on shame and guilt gravitates around the idea that shame is avoidance-oriented and guilt approach-oriented, but this account critically lacks experimental tests. Moreover, alternative accounts have posited that shame might elicit approach and that guilt might elicit avoidance, while empirical studies yield mitigated evidence (de Hooge et al., 2018;Graton & Mailliez, 2019). Given the current state of the literature, it is, therefore, crucial to experimentally investigate behavioural consequences of these two often confounded emotions.
To the extent that situations appear to be poor predictors of whether shame or guilt will prevail in terms of affective state (Tangney, 1998), authors often rely on measures of one's tendency to experience shame and/or guilt instead of manipulating these emotions that are oftenif not alwaysfused (see Rebega et al., 2013). Shame and guilt proneness qualify individuals who are more at risk of experiencing shame and guilt, respectively, when confronted to a failure or a transgression. As such, individuals high in shame proneness would be expected to report more shame following a failure than individuals low in shame proneness scores, and the same reasoning could be made for guilt proneness. These two traits appear to be associated to several distinct personality traits with shame proneness being globally seen as more maladaptative than guilt proneness (Abe, 2004). For example, shame proneness was found to negatively predict self-compassion and self-esteem (Cohen et al., 2011). Moreover, whereas guilt proneness does not appear to be related to neuroticism, shame proneness tends to correlate with measures of neuroticism (Abe, 2004;Cohen et al., 2011;Muris et al., 2018;Paulus et al., 2016).
Shame and guilt proneness have also been found to be associated to behavioural inhibition system (BIS) and behavioural activation system (BAS), respectively (Sheikh & Janoff-Bulman, 2010; but see Cornwell & Higgins, 2015 for mixed results). These two motivational systems are in turn considered to regulate aversive (BIS) and appetitive (BAS) motives and behaviours. Moreover, Carpenter et al. (2016) observed across two samples that guilt proneness was positively related to self-forgiveness, whereas shame proneness was negatively related to self-forgiveness. These elements corroborate the idea that, as it could be expected for the actual affective states of shame and guilt, shame proneness also entails a negative self-evaluation and a greater avoidance motivation, whereas guilt proneness would be associated to approach motivation. Hence, it should be expected that shame-prone individuals would display greater self-focus avoidance after failure than in the absence of failure, whereas guilt-prone individuals should be less affected.
In the present study, we adapted Duval and Wicklund's (1972; Wicklund, 1975) classic paradigm (avoidance of self-reflection in a mirror) to modern eye-tracking technology to directly but unobtrusively measure behavioural self-avoidance. We hypothesised that, consistent with the dominant view, shame proneness should moderate self-focus avoidance following failure priming. We thus conducted an experimental study where individuals' proneness to feel both guilt and shame was measured in order to assess the respective potential of these emotions to moderate self-focus avoidance following failure (Chatard & Selimbegović, 2011;Moskalenko & Heine, 2003;Twenge et al., 2003).

Ethical approval
The study was approved by the institutional ethical committee and conforms with the European regulations on data protection.

Participants
We registered a sample of 150 participants, and recruited 156 students from a medium-sized French university. Eleven participants were removed from the sample due to bad acquisition of gaze positions (because of lens, glasses or other variables that created noise, n = 3), genetic disorders (i.e. nystagmus and important strabismus, n = 2) and experimenters' mistakes (n = 6). Hence, the reported analyses were conducted on a sample of 145 participants (M age = 18.52, 95% CI [18.33,18.71], 133 women). This sample size provided 99%, 99% and 40% power to detect a large (f 2 = 0.35), medium (f 2 = 0.15) and small ( f 2 = 0.02) effect size, respectively (for the R 2 increase relative to a given predictor, in a regression with five tested predictors, as in the main analyses reported hereafter).

Individual differences in proneness to self-Conscious emotions
To measure individual differences in shame and guilt proneness, we used a modified version of the state shame and guilt scale (SSGS; Marschall et al., 1994). Words such as "generally" or "often" were added at the beginning of each item to make the questions less relevant to the immediate context and more trait-like, compared to the original scale. Thus, this measure assessed individual differences in the propensity to feel shame (e.g. Oftentimes, I want to sink into the floor and disappear, Generally, I feel like I am a bad person, α = .70), guilt (I generally feel tension about something I have done, I generally feel like apologising, confessing, α = .75) and pride (Generally, I feel good about myself, I often feel capable, useful], α = .75) generally in life, rather than in a specific situation. This measure was preferred to classical measures of guilt, shame, and pride proneness such as the Test of Self Conscious Affects-3 (TOSCA-3; Tangney et al., 2000), or the Guilt and Shame Proneness (GASP) scale (Cohen et al., 2011), because existing measures tend to conflate shame and guilt proneness with an evaluation of moral standards (Tangney & Dearing, 2002). According to Tangney and Dearing (2002), the set of personal moral standards should be distinguished from one's experience of guilt or shame (which are affective experiences). The use of a modified version of a state shame and state guilt measure thus allows to match as close as possible the actual emotional experience of guilt and shame and evaluate one's proneness to actually feel those emotions. The finding of a pilot study showed that higher state shame (but not state guilt) scores on this scale was associated with increased escape-thought accessibility following a failure (vs. success) priming manipulation similar to the one described in the failure induction section below.

Face attention task
Using a reflective Mac Screen (iMac, 27", 44.5 cm × 65 cm or 17.5" × 25.6" or 1440 × 2560 pixels) equipped with an eye-tracking device (EyeLink® Portable Duo), gaze behaviours were assessed directly but unobtrusively while the participants were engaged in an unrelated cognitive task. The experimental room was organised to have the light source facing the participants, behind the computer screen. The computer screen could thus clearly reflect the participant's face (see Figure 1 for an illustration; additional photographs of the experiment setup are available online, https://osf.io/sp94m/?view_ only = 9338c22020d949c19fd3e91870dbd96b). The use of an eye-tracking device allows a direct and precise behavioural measure of self-focus avoidance in an ecological situation (exposure to one's reflected face). The unrelated cognitive task used to maintain the participant's attention on the screen was a lexical decision task with words and nonwords displayed randomly in one of the four corners of the screen. Participants were instructed to attend to strings of letters appearing randomly in one corner of the screen at random time intervals. Their task was to indicate as fast as possible whether displayed strings of letters were words or non-words by pressing one of the two triggers of a controller remote (Microsoft SideWinder Controller, right trigger associated to the detection of a word, and left trigger associated to the detection of a nonword). A blank black screen separated each pair of trials for a random duration set between 300 and 9000 ms. The intertrial intervals were randomly generated before the beginning of the study, and all participants had the same intertrial intervals to ensure that they all had the same total time available to look at their screen-reflected faces. 1 The unpredictable duration of the intertrial interval also ensured that participants stayed focused on the task.

Failure induction
We manipulated failure using a scenario-based method. Participants were asked to read a vignette describing a student who, in the company of their successful friends, learned that they had failed (vs. passed) an exam. Specifically, the failure condition described a situation in which the protagonist discovered their failure on an important exam in front of their friends who had themselves succeeded. On the other hand, the success condition described a situation of an unimportant validation of an easy task that both the protagonist and all of their friends achieved. They were randomly assigned to one of these two experimental conditions (failure vs. success). The text was written from the perspective of the protagonist in first person singular. They were then instructed to write for 10 min, about how they would feel, behave and what they would be motivated to do in such a situation. This procedure was pre-tested in a study using a sample of 200 participants where it was shown that it significantly decreased pride scores and increased shame (but not guilt) scores reported in the SSGS (for details, see supplementary online material, https://osf.io/ awp54/?view_only = 9338c22020d949c19fd3e91870d bd96b).

Procedure
After being greeted and introduced to the experimental room, participants read an information notice regarding what was expected from them in the study and then completed the informed consent form. They were seated in front of the reflective screen equipped with the eye-tracking device and were told that the experiment focused on how individuals read words depending on their mood. After completing the modified version of shame and guilt proneness, they completed a practice block consisting of 12 trials of the word recognition task used for the face attention task. The practice block had a twofold purpose: training the participant for the main measures and accustoming the participant to his/ her own reflected image with the chin-rest. After completing the practice block, participants completed the failure induction task. Then directly afterward, they proceeded with the 36-trial critical block of the face attention task. Each block of the face attention task was preceded by a 9-point calibration task. Participants were then thanked and debriefed.

Data analysis
Data were analysed using R version 3.6.3 (R Core Team, 2021). We performed linear regression models separately for the two predicted variables (i.e. dwell time and number of gazes entering the AOI). The experimental condition, shame proneness scores, guilt proneness scores, and the interactions of shame proneness on the one hand and guilt proneness on the other hand with the experimental condition were used as predictors (hence the models contained five predictors). Although we initially registered two different models (one for each emotion), computing only one model offers the benefit of a more complete model, reduces the number of analyses, and provides tests of shame and guilt proneness effects independent from each other. That being said, computing separate models for shame and guilt proneness did not change any of the reported inferences.
As registered, participants whose scores on the dependent variables deviated from more than 2.5 Median Absolute Deviations (MAD) from the median (Leys et al., 2013) of the dependent variables were excluded (but see sensitivity analysis section).

Transparency and openness
We report how we determined our sample size, all data exclusions (if any), all manipulations and all measures in the study. Registration, data, analysis codes, data cleaning codes, and supplementary information regarding the experimental setup are available online (https://osf.io/sp94m/?view_only = 9338c22020d949c19fd3e91870dbd96b).

Preliminary analyses
For each model tested, distributions of the dependent variables' residuals failed to conform to normal distributions, as evidenced by Shapiro-Wilks' tests: for dwell-time, W = .87, p < .0001, and for number of gazes entering the AOI, W = 0.96, p < .001. However, we decided to report the registered analyses despite the suboptimal conditions. Additionally, a sensitivity analysis was conducted to account for how correcting for this deviation from a normal distribution influenced the reported results. This analysis revealed that results appear globally insensitive to analytical decisions (p < .05 in 75% of the specifications and p < .10 in 100% of the specifications, average p = .032, see section "Sensitivity Analyses (Multiverse Approach)").
Using the full sample (N = 145), the expected correlation between shame and guilt proneness did not exceed the internal consistencies of these scales (Pearson's product-moment correlation, r = .51, p < .001), and the dependent variables were positively related when using a Spearman's rank correlation, p = .23, p < .01, but not when using a Pearson product-moment correlation, r = .013, p = .87, suggesting a non-linear correlation between those variables and a deviation from normal distribution (see Figure A in online supplementary material, https://osf.io/49ef2/?view_only = 9338c22020d949c1 9fd3e91870dbd96b). Zero-order Pearson's correlations in the full data set are reported in Table 1.

Dwell time
As registered, ten participants whose scores deviated for more than 2.5 MAD from the median were excluded from the model. We performed a global linear regression model predicting dwell time with the experimental condition, shame proneness scores, guilt proneness scores, and the interactions of shame and guilt proneness with the experimental condition as predictors, F(5, 129) = 0.65, p = 0.66, R 2 = .025.
In this model, there was no effect of experimental condition, B =

Number of saccades into the AOI
Six participants whose scores deviated for more than 2.5 MAD from the median were excluded from the model. We once again performed a global linear regression model predicting number of saccades performed from outside the AOI into the AOI (thereafter, saccades into the AOI) with the experimental condition, shame proneness scores, guilt proneness scores, and the interactions of shame and guilt proneness with the experimental condition as predictors, F(5,133) = 2.21, p = 0.057, R 2 = .077.
Given the high correlation between shame and guilt measures, we also conducted analyses without using guilt as a predictor to test whether the effect of shame was spuriously hidden by its shared variance with guilt. In this model including manipulation of failure and shame scores as predictors, F(3, 135) = 1.34, p = .27, R 2 = .03, the experimental condition was not a significant predictor, B = Moreover, a significant effect of the experimental condition emerged, B = -3.63, t(135) = -2.00, p = .48, η 2 = .029, 95%CI[-7.21, -0.039], such that participants made significantly less fixations outiside of the AOI followed by fixations in the AOI in the failure condition (M = 24.64, SD = 11.09) than in the success condition (M = 28.25, SD = 10.54). These results suggest that shame does not seem to predict self-avoidance even when it includes the common variance that it shares with guilt. We next turn to an alternative interpretation of the reported results in terms of task focus, and report analyses of gaze patterns in the task-relevant areas of the screen.

Testing an alternative hypothesis: increased performance motivation
Results tend to indicate that guilt proneness scores interacted with the experimental manipulation of failure to predict self-focus avoidance, as measured through the registered evaluation of fewer gazes from outside the AOI to the AOI. However, to the extent that guilt is associated to approach orientation (e.g. Sheikh & Janoff-Bulman, 2010) and an action tendency to repair for wrongdoings (e.g. Cohen et al., 2011), an alternative account of such results could be that guilt-prone individuals in the failure condition were more motivated to perform the cognitive task. In other words, they could have been approaching the task, rather than avoiding their faces. Accordingly, higher motivation to correctly detect words displayed in the four corners of the screen could result in fewer fixations in the AOI. Thus, we conducted additional analyses in order to examine this alternative hypothesis.
We computed four additional AOI to capture fixations occurring in zones comprising the expected locations of the targets (rectangle AOI of a size of 575 × 193 pixels). We then computed the number of fixations into these AOI that were preceded by a fixation outside of the AOI. None of the predictors had a significant effect on the number of saccades into these task-relevant AOI (all p > .1). Additional details regarding the examination of this "motivation hypothesis" are described in a supplementary document available online (see https://osf.io/sp94m/? view_only = 9338c22020d949c19fd3e91870dbd96b)

Multiverse approach: sensitivity analyses
Results show that guilt but not shame proneness moderated self-avoidance after failure. Because this finding was not anticipated, we conducted a sensitivity analysis to determine the reliability of this result.
Multiverse analyses consist of testing an effect multiple times with different combinations of analytical choices (Steegen et al., 2016). It allows the observation of influential and non-influential decisions in the emergence of a significant effect. The interaction between guilt proneness and the experimental condition on the two dependent variables, the number of gazes entering the AOI and the dwell time on the AOI, was examined. Three parameters were considered: the cut-off for excluding outliers (with five different modalities), the size of the AOI (with two different modalities) and the type of data transformation (with two different modalities). Hence, we report the results of 5 × 2×2 = 20 analyses for each dependent variable.

Outlier exclusion threshold
Five modalities were used to study the influence of the outliers on the results: excluding scores greater and less than 2 MAD from the median, 2.5 MAD from the median and 3 MAD from the median, using the full sample without excluding outliers and using robust MM estimators (using the lmrob command in the "robustbase" R package, see Yohai, 1987) with weights applied to residual scores depending on their distance from the sample mean (a technique less sensitive to outliers).

Size of the AOI
We registered a wide interest area for the current study. However, we observed that the screen zone capturing the face of the participants was substantially smaller than the registered zone (see Figure 1). Arguably, a smaller interest area would avoid capturing unnecessary noise. Hence, we included two modalities of the size of the AOI in the multiverse analysis: a large version of the AOI (as in the main analyses) and a narrower AOI.

Data transformation
Neither dwell time nor the number of gazes entering the AOI were normally distributed. Data for both these variables were in fact following an ex-gaussian distribution. We thus decided to include log-transformation as a relevant data analysis pathway, for this transformation is widely used to handle ex-gaussian distributions. Log-transforming dwell time resulted in a normal distribution (W = .99, p = .38). However, it failed to normalise the number of saccades into the AOI (W = 0.96, p < .001).
To see how the logarithmic transformation would influence the reported results, we used data transformation as a parameter of the multiverse analyses with two parameters: untransformed and log-transformed scores. Table 2 presents a summary of the multiverse analyses, showing how analytical choices influenced the p-values associated with the guilt proneness by experimental condition interaction.

Multiverse analyses results
The multiverse analysis revealed that the guilt proneness by experimental condition interaction on the number of gazes entering the AOI was robust to a number of analytical choices. Evidence for the reported effects were only mitigated by the findings of mixed results when using the large AOI and logtransformation. This inconsistent result might point to the necessity of considering a more precise AOI but also to the importance of using untransformed scores.
Regarding dwell times, results pointed to an absence of interaction when using a large AOI. However, when using a more precise AOI, most results were significant and in the same direction as those observed on the number of saccades into the AOI. Notably, combinations of log-transformation and a narrow AOI yielded significant interactions regardless of the method used for outlier detection (with the exception of the 2 MAD sample, p = .055).
Robust regressions using MM estimators (initial weight of the re-descending function for weighting residuals is an S-estimator; see documentation of the lmrob command in the "robustbase" package for further details) used in all the analytical modalities, indicated effects consistent with our interpretation: a significant interaction for the number of gazes entering into the AOI (with the exception of the log-transformed scores in the large AOI that indicated a trend towards significance) and significant effects for dwell times using the narrow AOI but not the large AOI.
Overall, the multiverse analysis demonstrated that the effect on dwell times was contingent on using a relatively narrow AOI. Regarding the number of saccades into the AOI, results were more robust without log-transformation of the data. 2 To sum up the results of the sensitivity analysis, guilt proneness rather than shame proneness appeared to have a consistent relation to self-focus avoidance upon failure, in almost all the analytical combinations considered, hence lending strong support to the reliability of the previously reported results.

Discussion
Individuals who fall short of personal or social standards of value are often motivated to escape selfawareness (Chatard & Selimbegović, 2011;Morin, 2011;Moskalenko & Heine, 2003;Silvia & Duval, 2001). To the best of our knowledge, the role of proneness to feel negative self-conscious emotions in such situations has never been studied. Drawing on the classic mirror paradigm (Wicklund & Duval, 1971), we developed an experimental paradigm designed to unobtrusively assess self-focus avoidance using an eye-tracker and a reflexive black screen. Using this new paradigm, it was shown that, unexpectedly, guilt but not shame proneness moderated effects of priming failure on self-focus avoidance. Participants relatively high in guilt proneness oriented their gazes less frequently towards the area of their reflected face after imagined failure than after imagined success. This effect was quite robust to a number of alternative analytical decisions. The present study is the first to use this novel paradigm and points to a number of important theoretical, methodological and practical implications.

Theoretical implications
Self-focus avoidance after failure was found for individuals highly prone to guilt. The present findings are generally consistent with objective self-awareness theory (Duval & Wicklund, 1972;Silvia & Duval, 2001), but also with accounts of guilt as an emotion that might elicit avoidance (Amodio et al., 2007;Baumeister, 1990). The current results have important implications for the theoretical distinctions between shame and guilt, inasmuch as they contradict the dominant view of shame as eliciting avoidance and guilt as eliciting approach. Some theoretical perspectives might account for these results. For instance, in an effort to reconcile findings that guilt is associated to approach (Baumeister et al., 1994;Tangney & Dearing, 2002) but also to avoidance (Devine et al., 1991;Monteith, 1993;Monteith et al., 2002), Amodio et al. (2007) advanced a dynamic model of interpersonal guilt. These authors suggested that this emotion could be associated to both approach and avoidance in a two-stage reaction: guilt would act as negative cue motivating one to immediately display avoidant behaviours in order to avoid further damage. However, if and when an opportunity for reparation appears, approach behaviours are elicited, in order to repair for the wrongdoing. Hence, the avoidant reaction displayed by individuals prone to guilt in the current study could be accounted for by an increased motivation to halt further damage until an opportunity for reparation is presented. This prediction could be tested by contrasting conditions with and without opportunity for reparation (see, for instance, Graton et al., 2016).
Interestingly, whereas guilt proneness did not significantly predict self-focus avoidance in the failure condition, it was associated to a larger number of saccades into the AOI in the success condition. This latter effect might indicate that guilt-prone individuals are more sensitive to success-related cues than failurerelated cues. Increased sensitivity to rewards/pride (vs. punishments/guilt) among guilt-prone individuals might facilitate the elicitation of reparation behaviours often described in the guilt literature (Graton & Mailliez, 2019;Tracy & Robins, 2006). However, concluding that the link between the effect of guilt proneness on self-focus upon success indicates that guilt predicts approach might be unwarranted. Indeed, participants in the success condition were unlikely to feel guilt, regardless of their level of guilt proneness. As such, it is difficult to interpret this effect as suggesting that guilt is associated to approach. Nevertheless, future studies directly investigating how increased sensitivity to success (vs. failure) might foster approach and reparation in failure situations among guilty individuals could yield interesting insights into the motivational mechanisms operating to predict pro-social outcomes.
Moreover, a notable divergence of the current study in comparison to prior research is that it investigates avoidance of self-focus, while most prior research focuses on avoidance of social activities. This nuance is of importance, inasmuch as guilt is often described as other-oriented and associated to more perspective-taking than shame, which should be more self-oriented (Leith & Baumeister, 1998;Yang et al., 2010; for a review, see Tangney & Dearing, 2002, chapter 5). It is noteworthy that avoiding others (e.g. de Hooge et al., 2018) and avoiding the self are not equivalent activities, and further research should aim to disentangle how shame and guilt motivate individuals to approach or avoid the self and the others. In fact, engaging in social activities could potentially be considered as a distraction from the self. In this case, the present findings are not in contradiction with, but rather complement and extend prior research linking guilt to prosocial tendencies.
Crucially, the current study focuses on personality traits rather than emotional states. Indeed, the failure manipulation was used based on the initial assumption that it would elicit shame (based on a pilot study, see https://osf.io/awp54/?view_only = 9338c22020d949c19fd3e91870dbd96b). As such, individuals prone to shame should feel shame and hence avoid self-focus. However, the current results indicate that, although the failure manipulation should induce shame, it is guilt-prone individuals that are more likely to react to this manipulation with self-focus avoidance. It is possible that failure might prompt externalisation of blame among shame-prone individuals. Indeed, as stated by Tracy and Robins (2006), although internalising the blame is required to feel shame or guilt, it has been observed that shame proneness was associated to a tendency to blame the others or the environment (Thomaes et al., 2011), and a greater proclivity toward external attributions (Kaplánová & Gregor, 2021;Sales et al., 2007;Woien et al., 2003). As such, shame-prone individuals might avoid feelings of responsibility over their failure, and hence avoid emotional consequences of failure. On the other hand, guilt-prone individuals would feel more responsible for their failures (as reflected in greater self-focus avoidance after failure) and successes (as reflected in less selffocus avoidance after success).

Limitations
As discussed above, the current study investigated the role of proneness to feel specific secondary emotions in self-focus avoidance. While this allows the evaluation of inter-individual differences in coping with selffocused attention, it prevents us from concluding that guilt plays a causal role in interacting with failure to predict self-focus avoidance. Indeed, guilt and shame proneness are personality traits rather than affective states. Future research would thus benefit from studying the influence of the actual emotion of guilt instead of relying on a personality variable that might be strongly correlated, and thus partly confounded with other dispositional factors.
In the current work, we used a derivation of a measure of guilt and shame proneness initially intended to tap into state shame and guilt. However, as noted by Carpenter and Tsang (2020), the SSGS and trait measures of shame and guilt are correlated without any transformation of the SSGS. Crucially, the SSGS is a measure emphasising phenomenological aspects of shame and guilt (feelings of regrets, inferiority, powerlessness, etc.), rather than hypothetical responses to hypothetical situations. Items of the SSGS present two notable advantages in our case: (1) they do not make an explicit mention of shame and guilt (as opposed to shame and guilt proneness measures such as the PFQ-2; Harder & Greenwald, 1999), and (2) they do not directly address approach versus avoidance responses to transgressions and failure, but ratheras transformed in a trait measurefeelings generally experienced in daily life (in contrast to shame and guilt proneness measures such as the GASP emphasising the relation between shame and avoidance and guilt and approach; Cohen et al., 2011). Hence, the modified version of the SSGS provides a subtle measure of proneness to feel guilt and shame that does not conflate the measures of shame and guilt with avoidance-approach behaviours. However, although the use of the SSGS as a trait scale in the current study provided satisfying internal consistency, investigating the external validity of this measure is warranted.
Finally, our participants were predominantly female (92% of female participants in the current sample), and it has been shown that women tend to have a higher proclivity to guilt than men (Bybee, 1998). It is thus possible that the present findings are limited to or stronger among women. 3 Applications A consequential number of psychological disorders have been associated with problematic self-directed attention, such as depression (Pyszczynski & Greenberg, 1987), body dysmorphic disorders (Barnier & Collison, 2019;Toh et al., 2017), dissociative disorders (Lebois et al., 2019) or eating disorders (Tuschen-Caffier et al., 2015; for a review, see Griffen et al., 2018). Hence, an eye-tracking device for "tracking" pathological patterns of self-directed gaze is a novel and interesting tool that could help identify symptoms of such pathologies. Several studies in clinical psychology used participants' photographs to assess gaze patterns on the self (Bauer et al., 2017;Bortolon et al., 2016). However, exposing participants to photographs of themselves involves a risk that they can easily realise what is being measured (after all, they are in front of their photographs and an eye-tracking device). Thus, relying on participants' screen-reflected images to track gaze patterns might provide interesting future directions by enabling researchers to more easily dissimulate the aims of their studies and thus, limit demand characteristics.

Conclusion
Facing the self in a situation of failure could facilitate self-improvement, whereas escaping from the self has been associated with a number of behaviours with potentially dire consequences, ranging from alcohol intake (de Timary et al., 2013;Hull & Young, 1983) to suicide (Baumeister, 1990;Landrault et al., 2020). Thus, a better understanding of what might drive people to engage in such behaviours and how to limit them or transform them into something more constructive is a crucial endeavour in the effort to promote psychological adjustment and well-being. The current findings identified that guilt rather than shame proneness was related to self-focus avoidance after failure. Results point to possible recommendations for clinicians or counsellors when dealing with individuals prone to guilt feelings who are trying to cope with failure. Notes 1. In order to randomise the times of presentation of the stimuli, we drew 24 random values between 300 and 9000 ms and 24 values between 300 and 1500 ms. The reason for these two different pools of random values was to ensure a satisfying quantity of short times (between 300 and 1500 ms) and of short to long times (between 300 and 9000 ms). The whole operation was performed using Excel randomiser. 2. An exhaustive multiverse analysis with 5 (predictors) × 2 (types of AOI) × 2 (types of transformation) × 5 (methods for handling outliers) × 2 (DV) = 200 p-values, revealed that small p-values associated with an effect of the experimental condition (failure vs. success induction) were found regardless of outliers cut-off decision in the analyses of the untransformed numbers of saccades into the AOI when using the large AOI. Moreover, without any exception, shame proneness did not show any relevance in self-focus avoidance in a failure situation (see Supplementary Online Material, https://osf.io/sp94m/? view_only=9338c22020d949c19fd3e91870dbd96b). 3. Gender did not influence the interaction on number of saccades into the AOI observed in the reported confirmatory analyses as shown in a model evaluating the main effects of gender, experimental manipulation, guilt proneness and their interaction using 2.5 MAD as an outlier criterion, t(131) = 0.32, p = .75. But it must be noted that our sample was composed of 92% of female participants and hence this analysis might not be meaningful.