Face your fears: direct and indirect measurement of responses to looming threats

ABSTRACT This study investigated the emotional and behavioural effects of looming threats using both recalled (self-reported valence) and real-time response measurements (facial expressions). The looming bias refers to the tendency to underestimate the time of arrival of rapidly approaching (looming) stimuli, providing additional time for defensive reactions. While previous research has shown negative emotional responses to looming threats based on self-reports after stimulus exposure, facial expressions offer valuable insights into emotional experiences and non-verbal behaviour during stimulus exposure. A face reading experiment examined responses to threats in motion, considering stimulus direction (looming versus receding motion) and threat strength (more versus less threatening stimuli). We also explored the added value of facial expression recognition compared to self-reported valence. Results indicated that looming threats elicit more negative facial expressions than receding threats, supporting previous findings on the looming bias. Further, more (vs. less) threatening stimuli evoked more negative facial expressions, but only when the threats were looming rather than receding. Interestingly, facial expressions of valence and self-reported valence showed opposing results, suggesting the importance of incorporating facial expression recognition to understand defensive responses to looming threats more comprehensively.

Imagine being faced with threatening moving stimuli, such as a spider crawling closer or a stranger approaching with visible anger.How would you react emotionally?And would these threats, from animals or humans, elicit equally intense emotional reactions?Being shaped by our evolutionary history, we have learned to be vigilant toward potential moving threats, enabling us to react appropriately and ensure our survival (Rossini, 2014).
Our sensitivity to motion is particularly pronounced for approaching stimuli.People have learned to protect themselves from rapidly approaching (looming) sounds, animals, or individuals, leading to the emergence of a cognitive bias known as the looming bias.This bias causes people to reliably underestimate the time it takes for looming stimuli to arrive or make an impact.Although this bias can increase overall error rates and thus appear irrational, it minimizes overall fitness costs.Specifically, the perceptual underestimation (rather than accurate estimation) of the arrival time of looming stimuli may provide people with advanced warning of the approaching source.By doing so, they create a temporal safety margin that allows for more time to initiate defensive actions such as fight, flight, or freeze (Neuhoff, 2001(Neuhoff, , 2016;;Vagnoni et al., 2012).Looming motion, characterized by a sudden increase in an object's perceived size, indicates an object rapidly approaching and likely signifies a threat to the observer.For instance, a snake becomes increasingly threatening as it crawls closer (Mühlberger et al., 2008).Looming motion is perceived when a predator launches an attack or when a ball is hurled in the observer's direction (Moher et al., 2015).The evolutionary explanation for the development of the looming bias is that looming stimuli are potentially more harmful and dangerous than receding stimuli, so it is advantageous to underestimate the time it takes for looming stimuli to reach us.Preparing for a rapidly approaching stimulus ahead of time minimizes fitness costs compared to potentially reacting too late (Neuhoff, 2001).
Previous investigations into the looming bias have primarily focused on animal responses, such as mice, locusts, zebrafish larvae, and crabs, rather than humans (e.g.De Franceschi et al., 2016;Huang et al., 2017;Temizer et al., 2015).Many of these studies employed predominantly unnatural or artificial research stimuli such as graphics, geometrical shapes, or auditory tones to assess animal reactions to looming motion.
Studies on human exposure to looming motion have investigated cognitive, affective, and behavioural responses.For example, looming stimuli have been found to capture attention (Franconeri & Simons, 2003;Moher et al., 2015;Rossini, 2014), elicit heightened startle responses (Mühlberger et al., 2008), and trigger defensive freeze-like reactions (De Paepe et al., 2016), similar to observations in animals (Riskind et al., 2016;Sagliano et al., 2014).Other studies (Brendel et al., 2012;Vagnoni et al., 2012) have examined the perceived time-to-collision (TTC) of looming stimuli compared to the actual TTC.The perceived TTC is influenced by the affective content of looming stimuli, with individuals judging threatening (versus neutral) stimuli to have a shorter TTC.Thus, threatening stimuli appear to approach more rapidly than non-threatening stimuli, particularly for individuals already fearful of such stimuli.Visual looming, therefore, serves as a direct perceptual indicator of threat.
Furthermore, research indicates that people have an evolutionarily ingrained aversion to approaching or looming stimuli.Hsee et al. (2014) demonstrate that individuals exhibit more negative feelings toward approaching stimuli than non-approaching stimuli (i.e.receding and static), referring to this phenomenon as the approach aversion bias.This bias has become overgeneralized, extending to ambivalent stimuli, causing individuals to fear stimuli that are approaching, even if they are nonthreatening (i.e.neutral or positive; Hsee et al., 2014).Previous studies have confirmed the presence of the approach aversion bias by showing that individuals experience more negative emotions when perceiving a stimulus as looming rather than receding or static (Davis et al., 2011).Moreover, Mühlberger et al. (2008) show that approaching (versus receding) unpleasant stimuli evoke more intense emotional responses in terms of both valence and arousal.
Most of these studies have relied on self-report measures to assess emotional responses to looming threats after exposure to such stimuli.However, the recall accuracy of emotions is rather weak (Thomas & Diener, 1990), although negative emotions tend to enhance memory accuracy more than positive emotions (Kensinger, 2007).Importantly, facial movements express emotional responses.People use these facial expressions as non-verbal cues for communication, as the interpretation of facial movements contributes to the evaluation of intent in others (Zeinstra et al., 2009).Behavioural data suggest that facial expressions can convey both the emotional state of the individual displaying them and their behavioural intentions or action demands to the observer (Horstmann, 2003).Anticipating behavioural intentions can elicit appropriate approach or avoidance (withdrawal) responses from the observer.So, the observer is primed to approach individuals with positive facial expressions (e.g.joy) and avoid those with negative expressions (e.g.anger).Detecting another person's behavioural intentions through their facial expressions plays a crucial role in social interaction (Zeinstra et al., 2009).Thus, facial expression recognition can capture responses during exposure to emotionally charged stimuli (Noordewier & van Dijk, 2018).Therefore, estimating emotional experiences based on objectively measured facial expressions has become an increasingly important research area (Terzis et al., 2013).
However, the existing research lacks investigation into facial expressions elicited by looming threats.The present study aims to address this research gap by examining emotional and behavioural responses to looming threats using both recalled (i.e.selfreported valence) and real-time response measurements (i.e.facial expression recognition).Our study aims to replicate and extend previous research on the looming bias in three key ways.
First, we investigate facial expressions in response to looming versus receding threats.Facial actions support the two basic reactions to stimuli: approach versus avoidance.Facial expressions are typically interpreted in a dichotomous way as positive or negative affect, with approach behaviour observed in response to pleasant stimuli and avoidance behaviour observed in response to unpleasant stimuli (Mühlberger et al., 2008).Moreover, individuals tend to exhibit stronger emotional reactions to potentially dangerous or disliked stimuli than safe or liked stimuli (Zeinstra et al., 2009).Thus, negative (versus positive) facial expressions are expected to be more intense and quicker to appear as warning signs.This is crucial for alerting people to potential looming threats, such as a rapidly approaching spider (Zeinstra et al., 2009).Following this, we hypothesize that looming threats elicit more negative facial expressions than receding threats.
Secondly, we examine facial expressions in response to threats in motion, considering the strength of the threats.Building upon previous findings on the looming bias, it has been observed that more (versus less) threatening stimuli appear to approach faster (Brendel et al., 2012;Vagnoni et al., 2012).We aim to replicate and extend these findings by examining facial expressions to more versus less threatening looming and receding stimuli.The strength of threat is operationalized through two distinct threat categories known to elicit fear: animal versus human facial threats.Prior research on facial expressions perception has demonstrated that individuals exhibit an attentional bias toward angry faces compared to neutral faces (e.g.Van Honk et al., 2001).Brendel et al. (2012) further show that this attentional bias is also observed when angry faces are approaching the observer.Notably, the utility of processing facial emotions differs from that of processing an immediate threat, particularly from an animal.Brendel et al. (2012) suggest that people would respond more strongly to frontal attack pictures than to angry facial pictures, as human faces present a more ambiguous and less existential threat than a frontal animal attack.Therefore, in line with Brendel et al. (2012) and the evolutionary psychological theory on the looming bias, it is hypothesized that animal and human facial threats will not evoke equally strong facial expressions.Specifically, we expect more threatening looming stimuli (animal threats) to elicit more negative facial expressions compared to less threatening looming stimuli (human facial threats).Spiders and snakes are utilized as animal threats in the study due to their high biological and evolutionary relevance as potential sources of strong aversive reactions (Brendel et al., 2012;Dan-Glauser & Scherer, 2011).
Thirdly, we explore whether measuring facial expressions of valence in response to looming versus receding threats provides additional insights compared to self-reported valence.While selfreports and direct methods offer subjective accounts of the emotional component, they can be influenced by cognitive biases as they rely on individuals' perceptions of their emotional reactions or experiences.In contrast, physiological or indirect methods, such as facial expression recognition, can capture automatic emotional reactions in real-time, free from cognitive biases (Terzis et al., 2013).The benefit of physiological measurements is that they can uncover emotional responses that individuals may be unconscious or unaware of (Terzis et al., 2013).It is worth noting that direct and indirect measurements do not always align, indicating that physiological data can reveal responses that may not be fully captured by self-reports (Kostyra et al., 2016).Therefore, we aim to investigate the potential advantages of facial expression recognition as an objective indicator of emotional responses complementing self-reported valence.By incorporating indirect response measurements, we can better understand the defensive responses to looming threats and explore emotional experiences that may not be fully accessible through self-reports.

Design and participants
This study utilized a 3 (between-subjects: looming vs. receding vs. static threats) × 3 (within-subjects: animal vs. human facial threats vs. neutral images) mixed experimental design.The effect sizes η 2 from 12 previous studies investigating the effects of stimulus direction and picture content on valence were extracted from Mühlberger et al. (2008), Davis et al. (2011), andHsee et al. (2014).The average η 2 across these studies was 0.28, with a range of 0.01 to 0.96.To ensure adequate statistical power of 0.80, an a priori power analysis using Optimal Design Plus (Raudenbush et al., 2011) suggested a minimum sample size of 76 participants for a multilevel analysis with an alpha of 0.05, three levels of the between-subjects factor, one covariate, 24 trials per participant, and an expected effect size η 2 of 0.28.
A total of 78 healthy individuals, mostly undergraduates (45 women, 33 men; M age = 27.36,SD = 9.80), participated in the study and were randomly assigned to the looming (N = 27), receding (N = 24), or static (N = 26) condition.The study was conducted in a university lab and received approval from the local ethics committee.Participants provided informed consent and were compensated with 8 euros.One male participant was excluded from the final dataset due to incomplete demographics and data.All data and supplemental material are available at osf.io/7 × 9m5.

Stimuli
The study employed 16 pictures obtained from validated databases as experimental stimuli (see Figure 1).Eight animal threat pictures (four spiders and snakes respectively) were selected from the Geneva Affective Picture Database (GAPED; Dan-Glauser & Scherer, 2011, p. 640 × 480 pixels), and eight human facial threat pictures (four angry or threatening male and female faces respectively) were obtained from the Karolinska Directed Emotional Faces database (KDEF; Calvo & Lundqvist, 2008;Lundqvist et al., 1998, p. 410 × 307 pixels).A pre-test with 110 participants confirmed that these selected pictures expressed the emotion threat, and that animal threats were perceived as more threatening than human facial threats (see supplemental materials).Additionally, eight neutral pictures (four inanimate objects from the GAPED, and four neutral faces from the KDEF database) were included as fillers in the experiment to distract participants from the study purpose, but were not pre-tested or considered in our analyses.
The stimuli were presented in the form of 4-second videos, each representing looming, receding, or static conditions.Motion was simulated by altering the size of the stimuli on the screen (Hsee et al., 2014;Mühlberger et al., 2008).Looming stimuli were initially small (205 × 153 pixels) and gradually increased in size until they were large (614 × 461 pixels).Receding stimuli changed in the reverse direction (614 × 461 to 205 × 153 pixels).Static control stimuli stayed unchanged, always between the smallest and the largest size in the other two conditions (409 × 307 pixels).

Procedure
Participants were individually filmed during the experimental sessions using a webcam mounted on a computer screen (1600 × 900 pixels).They were unaware of the purpose of the study until the session concluded.Participants looked directly toward the screen (60 cm from the camera) and were randomly presented with the 24 experimental stimuli trials.Each trial began with a 2-second fixation period (centrally placed black cross to draw participants' eyes to the middle of the screen), followed by the presentation of a stimulus.After each trial, participants rated their feelings toward the stimulus using a Self-Assessment Manikin scale for valence (Bradley & Lang, 1994) from 1 (very positive) to 5 (very negative), recoded to a scale from −1 (negative) to 1 (positive).We refer to these values as "selfreported valence".After the 24 trials, participants completed the Fear Questionnaire to assess selfreported fear (Marks & Mathews, 1979;Vagnoni et al., 2012), which consists of 15 items, forming a Total Phobia Score (sum of all items; score range between 0 and 120; M = 34.31,SD = 12.82; α = 0.66).We included this individual trait variable as a covariate.Finally, participants reported their demographics and were debriefed, compensated, and asked for permission to use their video footage.
Facial expressions of each participant were analysed using FaceReader (Noldus, 2012;version 5.1), a robust automated system often applied in psychology research (e.g.Kostyra et al., 2016;Lewinski et al., 2014;Noordewier & van Dijk, 2018).This software is based on the Facial Action Coding System (FACS; Ekman et al., 2002), which is used to describe facial activity and states that basic emotions correspond with facial models (Terzis et al., 2013).FaceReader assessed facial expressions in terms of the six basic emotions (happiness, sadness, anger, surprise, fear, and disgust) and neutrality, with an accuracy of 89% (Lewinski et al., 2014;Terzis et al., 2013).The lack of emotions was recognized as neutrality (indifference), which is a baseline state, and the appearance of any emotion replaces this state (Kostyra et al., 2016).FaceReader analysed every third frame of the recordings (320 × 240 resolution at 25 frames per second), with the face model "General" (Kostyra et al., 2016).
For each frame, FaceReader computed intensity scores for expressions of the six basic emotions (0 = emotion not detected or present, to 1 = emotion fully present).We calculated average values of these intensity scores for each stimulus from onset to offset.We then pooled the negative emotions sadness, anger, fear, and disgust to form a negative emotions measure.We calculated the average values of this measure for each stimulus from onset to offset, which we repeated for each threat strength (animal and human facial threats).We refer to these values as "negative facial expressions".
As people often express emotions as short impulses (Kostyra et al., 2016;Terzis et al., 2013), or express multiple emotions at the same time, especially in the middle of changing displays from one emotion to another (Sacharin et al., 2012), pooling the negative emotions can better capture the mix of emotions that are expressed through facial expressions.Moreover, because participants predominantly express neutral facial expressions in face reading experiments (Kostyra et al., 2016), combining the different negative emotions provides an emotion measure that sufficiently differs from the baseline state of neutrality, lowering the possibility of floor effects.The correlations among the four distinct emotions and the pooled emotion measure (see supplemental materials) indicate that the emotions sadness, anger, and disgust correlate significantly with negative facial expressions (p < .001).The emotion fear does not correlate significantly with negative facial expressions (p > .05),but fear does correlate significantly with sadness and anger (p < .05).Further, disgust does not correlate significantly with the other distinct emotions (p > .05).As prior emotion research generally classifies disgust and fear as negative emotions and confirms an overlap between these two emotions (Davey, 2011), both emotions remained in the pooled measure of negative emotions.
FaceReader also computed general valence (happiness minus negative emotions, excluding surprise; −1 = negative, to 1 = positive).We calculated the average values of this valence score for each stimulus from onset to offset, which we repeated for each threat strength.We refer to these values as "facial expressions of valence" and use this measurement to enable comparisons with self-reported valence.
FaceReader cannot generate data when certain circumstances or participant behaviours make the ).Bottom panel: angry male face (BM29ANS), angry female face (AF26ANS), and neutral male face (BM14NES; filler) from the KDEF database (Calvo & Lundqvist, 2008;Lundqvist et al., 1998).GAPED and KDEF images reprinted here with written permission of the copyright holder.
correct assessment of all the particular emotions impossible (e.g.participants yawning; Noordewier & van Dijk, 2018;Terzis et al., 2013).Therefore, our final dataset contained 21.3% missing FaceReader data on trial level, which could not be sensibly interpreted as face images.These empty records are not desirable, but they make the rest of the records more reliable (Kostyra et al., 2016).
The data were analysed using multilevel modelling, accounting for the repeated measures structure of the dataset, where each participant provided measures for 16 experimental stimuli trials (excluding 8 filler trials).Three linear mixed effects models were used, with (1) negative facial expressions, (2) facial expressions of valence, and (3) self-reported valence as dependent variables respectively.Motion (between-subjects: looming vs. receding vs. static threats), threat strength (within-subjects: animal vs. human facial threats), and the interaction between motion and threat strength were the independent variables, with self-reported fear serving as a covariate.Each analysis included a randomly modelled intercept to account for participant-level effects, the restricted maximum likelihood estimation technique, and variance components as the default covariance structure.All data were analysed using IBM SPSS Statistics 25 (IBM Corp., Armonk, NY, USA).

Discussion
This study examined the emotional and behavioural reactions to looming threats by incorporating facial expressions as real-time indicators.The findings highlight the significance of considering both selfreported measures and facial expression recognition to gain a comprehensive understanding of defensive responses to looming threats.We contribute to the existing literature on the looming bias in three ways.
First, we showed that looming threats elicit more negative facial expressions compared to receding threats, although there was no main effects difference in facial expressions of valence.These results only partially confirm previous findings that relied on selfreports to measure the emotional effects of looming (vs.receding or static) threatening stimuli after exposure to these stimuli, revealing negative responses (e.g.Davis et al., 2011;Hsee et al., 2014;Mühlberger et al., 2008).However, we show that emotional and behavioural responses are also expressed by negative facial expressions during exposure to looming threats.This aligns with research showing similarities between human and animal responses to looming sounds.According to Neuhoff (2016), the additional converging evidence for the adaptive looming bias hypothesis comes from comparative work on rhesus monkeys who show a pattern of response to looming versus receding sounds that mirrors that found in humans (Ghazanfar et al., 2002;Maier & Ghazanfar, 2007).Particularly, studies with humans and other animals have identified neural correlates of looming sounds and have implicated cerebral regions that respond preferentially to looming (as opposed to receding) sounds (e.g.Bach et al., 2009;Ghazanfar et al., 2002;Maier & Ghazanfar, 2007;Seifritz et al., 2002).
Secondly, we found that more threatening stimuli (animal threats) elicit more negative facial expressions and valence than less threatening stimuli (human facial threats), but only when the animal threats are looming (vs.receding).These findings support previous research on the looming bias, suggesting that more threatening stimuli appear to approach faster than less threatening stimuli (Brendel et al., 2012;Vagnoni et al., 2012).While Brendel et al. (2012) suggest that people respond more strongly to animal threats than to human facial threats, because of their higher existential threat potential, our findings revealed that the processing of different threats depends on their motion path.The utility of processing immediate threats from animals is higher than from human faces, but only when approaching, which is in line with the evolutionary explanation of the looming bias (Mühlberger et al., 2008;Rossini, 2014).
Thirdly, our findings revealed opposing results for facial expressions of valence and self-reported valence in response to more threatening (animal) versus less threatening (human facial) looming stimuli.Facial expression recognition indicated more negative facial expressions of valence toward more (vs.less) threatening looming stimuli, confirming the looming bias.However, self-reported valence suggested more negative valence toward less (vs.more) threatening looming stimuli, contradicting the looming bias.These conflicting findings are consistent with previous research showing that emotional expressions and liking ratings can contradict (Kostyra et al., 2016).The use of facial expression recognition data can capture unconscious or unaware responses that may not align with conscious selfreports (Terzis et al., 2013).Human behaviour is driven by both conscious and unconscious emotional and behavioural mechanisms (Lewinski et al., 2014).People might consciously indicate to be more fearful of learned threats, such as certain human non-verbal communications, and they are more regularly confronted with angry people than dangerous animals.However, people have an unconscious and evolutionary-driven fear of innate threats, such as spiders and snakes.Remarkably, several studies show that visual sensitivity to looming motion is already observed in infants during the first months of their life, which contributes significantly to the implementation of defensive responses (e.g.Náñez, 1988), such as defensive blinks (Ayzenberg et al., 2015).For instance, Hoehl et al. (2017) showed that infants react with increased physiological arousal to these animals.This suggests an evolved preparedness for developing a fear of ancestral animal threats.
Our findings indicate that the recall of negative emotional experiences toward looming threats can indeed be heavily biased (Thomas & Diener, 1990), further emphasizing the importance of facial expression recognition in understanding defensive responses.We suggest future studies to further delve into the differences between facial expressions and self-reported emotions to animal and human threats in motion, considering other potential, equally plausible underlying mechanisms such as imitation (e.g.Proverbio et al., 2009), facial mimicry (e.g.Rymarczyk et al., 2016;Sato & Yoshikawa, 2007), emotional contagion (e.g.Hess & Blairy, 2001;Olszanowski et al., 2020), or empathy (Sonnby-Borgström, 2002).
Our study has limitations.First, we acknowledge that our experiment is underpowered, so the reported findings should be interpreted with caution and we advise for replication studies with a larger sample size.Secondly, the study participants were not very expressive toward the threatening stimuli presented on a computer, potentially leading to floor effects for facial expressions of valence toward receding and static threats.Using more intense or lively settings or more emotionally laden (threat) stimuli might elicit stronger facial expressions, such as more brow action (Noordewier & van Dijk, 2018), resulting in higher levels of emotional expressions, which could limit the risks of floor effects.Thirdly, given the lack of correlations between disgust and the other negative emotions in our study, we advise future studies to consider excluding the potentially ambiguous emotion of disgust to allow for a more coherent pooled negative emotions measure.Lastly, we did not manually code facial expressions with FACS (Ekman et al., 2002), which could result in lower reliability of the coding of spontaneous or naturally occurring facial expressions.We recommend future studies employ other objective or reliable measures, by manually coding facial expressions or using electromyogram (EMG) measurements to capture more subtle changes in facial reactions to stimulus motion (Zeinstra et al., 2009).

Disclosure statement
No potential conflict of interest was reported by the author(s).