Eye image effect in the context of pedestrian safety: a French questionnaire study

1.1 Audience effect and social control

Our behavior is therefore influenced by the presence of others, which scientists also call ‘the audience effect’ (Filiz-Ozbay and Ozbay, 2014; Kurzban et al., 2007; Zuberbühler, 2008). Feeling observed by real persons or an imagined audience, therefore has an impact on human behavior. Embarrassment, for instance, is defined as the ‘acute state of flustered, awkward, abashed chagrin that follows events that increase the threat of unwanted evaluations from real or imagined audiences’ (Eller et al., 2011; Miller, 1996). Being observed also tends to make individuals more compliant (Dear et al., 2019; Rodriguez Mosquera et al., 2011). With this in mind, Bateson et al. (2006) conducted an experiment to test the effect of eye images on cooperative behavior. In the coffee area of a building at the University of Newcastle, an honesty box was used for people to make contributions to the coffee fund. The experiment consisted of placing pictures of eyes or flowers close to this box and assessing whether they led to differences in the contributions made. The authors found that in the presence of eye images, subjects paid on average 2.76 times more than when flowers images were displayed. Still comparing the effect of these flowers images as control to those of eye images, other studies have found similar prosocial effects during everyday events. For example, eye images had prosocial effects such as cooperation for the clearing of trays in a university cafeteria (Ernest-Jones et al., 2011) and for waste sorting at a bus stop (Francey and Bergmüller, 2012). Similar results were found in experiments in more specific contexts such as blood donations: While eye images on flyers did not result in differences in pledge with a logo as control, more “real” donations were made by students who got the flyers with eyes image (Sénémeaud et al., 2017). Other than the activation of a “sense of being seen” (Pfattheicher and Keller, 2015) or the desire to maintain a pro-social reputation (Bateson et al., 2006), it is important to note that humans possess neurons that respond to faces and eyes and activate such prosocial behaviors (Emery, 2000; Haxby et al., 2000). Bateson et al. (2013) suggest that eye images induce more pro-social behavior regardless of local norms, thus suggesting that the application of eye images could be a means to combat anti-social behavior by triggering a feeling of shame (Nugier et al., 2007).

The use of social control to produce more cooperative behaviors may positively influence road use and safety. The limitation of conflicts and accidents on road infrastructures is directly dependent upon the respecting of rules by the numerous pedestrians and drivers. However, more than 8000 pedestrians die in road accidents in Europe every year, 25% of whom die when using crosswalks (Guéguen et al., 2015, 2016). These lethal accidents are due to cars not stopping at signalized intersections but also to pedestrians crossing illegally at the red signal (Sueur et al., 2013). Past studies show that individuals do not cross illegally when other pedestrians are present (Pelé et al., 2017, 2019a, 2019b). Visual communication between two individuals can lead to a change in the receiver's behavior and this effect can be found in a road context. Past studies showed that other people's presence impacts signal compliance (Rosenbloom, 2009; Faria et al., 2010; Raoniar and Maurya, 2022). For example, eye contact is an important element. Some studies explored the effect of gaze and smile on the propensity of drivers to stop at signalized intersections and allow pedestrians to cross (Guéguen et al., 2015, 2016). A pedestrian waiting at the edge of an unmarked crosswalk has a greater likelihood to cross if s/he seeks to make visual contact with an approaching driver than if s/he is not looking towards the approaching car, with 67.7% of cars stopping versus 55.1%, respectively (Guéguen et al., 2015). If in addition to this visual contact the pedestrian smiles, 62.9% of drivers stopped compared to 50.1% if the pedestrian sought visual contact with a neutral face (Guéguen et al., 2016). These studies show that visual contact can modify the behavior and speed of drivers, and highlight that the facial expression of the pedestrian also has an impact (Ren et al., 2016). However, these studies are rare and more research is needed on how human facial expressions affect the probability that pedestrians will cross the road illegally. This research may have great potential in terms of applications in the field of road safety, especially regarding the regulation of pedestrian behaviors.

1.2 Factors influencing road-crossing behaviour

Various factors influence pedestrians' risky behavior of crossing at red lights and highlights the importance of understanding these factors to improve road safety. According to the ONISR (2023), pedestrians could be held responsible for 21% of accidents. Risky behaviors exhibited by pedestrians include crossing outside designated areas, disregarding signals, not crossing perpendicularly, and even walking on the road instead of pedestrian facilities (Ding et al., 2014).

Pedestrians may engage in risky behaviors such as crossing at red lights due to the belief that it saves time, but this exposes them to potential accidents. The perception of risk associated with crossing at red lights depends on environmental and individual factors. The number of rule violations decreases with an increase in the number of traffic lanes, and a sense of safety within a pedestrian group can increase risk-taking behavior (Pelé et al., 2017; Wang et al., 2011). Pedestrians have unique characteristics as road users. They are slow, vulnerable, and often make choices to optimize time and distance. Some pedestrians intentionally break the rules and gather information before crossing, while others may cross without sufficient information due to distractions or group movement (Nasar et al., 2008). Pedestrians who cross at the beginning or end of a red light may use the traffic lights of cars to anticipate changes, while those who cross in the middle without stopping pose a greater risk (Lipovac et al., 2013; Pelé et al., 2019a).

Individual factors, such as gender and age, also influence risk-taking behavior. Men tend to take more risks and commit more offenses, while young adults and older people exhibit specific risk patterns. Social factors, such as crossing alone or in groups, also play a role. Pedestrians crossing in groups are more visible to other road users and engage in less risky behavior (Pelé et al., 2019b; Rosenbloom, 2009; Zhang et al., 2019). Cultural differences impact pedestrian behavior as well. For example, Japanese pedestrians exhibit fewer illegal behaviors compared to French pedestrians, which can be attributed to cultural factors and a greater sensitivity to social cues (Pelé et al., 2019a, 2019b). Environmental factors, such as road design and pedestrian facilities, also influence pedestrian behavior. Increasing road traffic density or reducing lane visibility can reduce risk-taking behavior, and well-designed pedestrian facilities lead to fewer offenses (Mueller et al., 1990; Keegan and O'Mahony, 2003; Wang et al., 2011).

Road safety policies and pedestrian facilities should be adapted to the culture and country. Additionally, the social aspect of human behavior, such as eye contact and non-verbal communication, can influence pedestrian and driver interactions. Eye contact and positive facial expressions have been shown to increase the likelihood of drivers stopping for pedestrians (Guéguen et al., 2015, 2016).

1.3 Study aims

This study aims to test the effect of eye images on the behaviors of pedestrians crossing at the red light. We collected different images of eyes from five different persons (two men, two women and one child) expressing different facial expressions (neutral, friendly and angry) and one image of flowers for use as a control ( Figure 1). These images were used in an online questionnaire that asked participants what they would feel if they saw these images before crossing a road. Would they feel observed, scared or surprised? Would the images discourage or encourage them to cross at the red light? These questions were chosen with caution in order to make the questionnaire valid and are based on previous studies (Bateson et al., 2013; Ernest-Jones et al., 2011; Francey and Bergmüller, 2012; Saeed et al., 2020). We compared their answers to these questions with sociodemographic variables (gender, age, geographical zone and city size) and also with a previous well-known questionnaire called ‘Pedestrian Behavior Questionnaire’ (hereafter referred to as PBQ, Appendix A: Deb et al., 2017; Granié et al., 2013), which tested the propensity of pedestrians to violate rules, make errors or lapses, or show positive or aggressive behaviors. The angry eye images are expected to have a stronger emotional impact on participants and thus prevent them from crossing illegally (Bateson et al., 2006). We also expect a gender effect and an age effect, with a stronger impact of eye images on women and younger individuals. Indeed, women might feel more observed or scared than men when looking at the eye images; studies have shown women to be more empathic and thus more receptive in different situations (Flynn, 2000; Furnham et al., 2003; Kellert and Berry, 1987; Miller et al., 2009). In the context of road crossing, fewer illegal and risky behaviors are observed in women than in men (Holland and Hill, 2007; Pelé et al., 2017; Sueur et al., 2013; Tom and Granié, 2011), with more positive behaviors and fewer rule violations or aggressive behaviors (Deb et al., 2017; Tom and Granié, 2011). Studies show that younger people are less respectful of rules (Holland and Hill, 2007; Pelé et al., 2017; Pfeffer and Hunter, 2013). We also expect regional differences in reactions to eye images, as interregional differences were observed in road accidents (Eksler et al., 2008; Lassarre and Thomas, 2005). Correlations are expected between the responses to the eye image questionnaire and the PBQ. Indeed, participants showing higher scores for violations or aggressive behaviors in the PBQ (i.e. the less prosocial individuals) should be less affected by eye images than participants who show more positive behaviors (i.e., the more empathic and cooperative individuals).

Figure 1. Images used to understand the effect of eye images on pedestrian behaviors.

The flower image is a control. The effect of eye images was tested using three different expressions (friendly, neutral and angry respectively from left to right) in the eyes of five different persons. Five questions were asked about feelings for each eye image: Do you feel observed, scared, inhibited, surprised or encouraged? (See methods for details). The score for each feeling and each image is the average score for all participants. Image credits: Cédric Sueur.

2.1 Questionnaire

The questionnaire was designed in three steps: 1. Eye images, 2. Pedestrian behavior questionnaire and 3. Sociodemographic questions.

Eye images: To test the effect of eye images, pictures were taken of five different individuals (one child, two women and two men) with three different facial expressions (friendly, neutral and angry). A picture of flowers was used as a control. Each picture was taken in black and white with the same contrast and brightness calibration (see Figure 1). For each of the 16 pictures (15 eyes and 1 control), participants were asked to answer five questions about their feelings when looking at the picture:

These five questions were chosen with caution. In order to check the validity of the questionnaire, we chose the four first questions to evaluate the negative effects of the eyes images and the fifth question to evaluate a positive effect. This means that we expected to have the scores of the fourth first questions to be positively correlated between them but negatively correlated with the fifth question. The questionnaire was validated with results (Section 3.1.) following our predictions.

As five questions for 16 pictures results in a very long questionnaire of 80 questions, 4 images (with for each image the five questions) were randomly selected for the participants to answer. We checked for a homogeneous distribution of questions to participants by making packages of 20 questions (for 4 images).

Sociodemographic questions: Participants were asked to indicate their gender (female or male), age (as numeric) and French postal code. The French postal code provides the city population size (DataNova Opensource, 2014 census) and the Defense and Security Zones (French Homeland Ministry, see https://doi.org/10.5281/zenodo.5745446).

Pedestrian behavior questionnaire (PBQ): Participants completed a questionnaire of 20 questions about pedestrian behaviors by Deb et al. (2017). The Pedestrian Behaviour Scale (PBS) is a self-report questionnaire that distinguishes five dimensions of pedestrian behaviour. Most questions concern pedestrian crossing at signalised crosswalk but some not (Vandroux et al., 2022). We used a pedestrian behavior questionnaire (PBQ) with 20 questions, developed by (Deb et al., 2017) and taken from (Granié et al., 2013) to optimise time as some original questions gave correlated answers (Vandroux et al., 2022). The order of questions was randomly attributed. In parenthesis, the average±stdv of the score (from 1 – never - to 6 – always -) attributed by participants to each question (N=611).

Violations (V)

V1 I cross the street even though the pedestrian light is red. (3.73 ± 1.35)

V2 I cross diagonally to save time. (3.45 ± 1.42)

V3 I cross outside the pedestrian crossing even if there is one (crosswalk) less than 50 m away. (3.51 ± 1.49)

V4 I take passageways forbidden to pedestrians to save time. (2.35 ± 1.36)

Errors (E)

E1 I cross between vehicles stopped on the roadway in traffic jams. (3.34 ± 1.43)

E2 I cross even if vehicles are coming because I think they will stop for me. (2.36 ± 1.19)

E3 I walk on cycling paths when I could walk on the sidewalk. (1.91 ± 0.96)

E4 I run across the street without looking because I am in a hurry. (1.27 ± 0.63)

Lapses (L)

L1 I realize that I have crossed several streets and intersections without paying attention to traffic. (1.52 ± 0.82)

L2 I forget to look before crossing because I am thinking about something else. (1.65 ± 0.78)

L3 I cross without looking because I am talking with someone. (1.65 ± 0.85)

L4 I forget to look before crossing because I want to join someone on the sidewalk on the other side. (1.45 ± 0.70)

Aggressive Behaviors (A)

A1 I get angry with another road user (pedestrian, driver, cyclist, etc.), and I yell at him. (1.94 ± 1.17)

A2 I cross very slowly to annoy a driver. (1.45 ± 0.87)

A3 I get angry with another road user (pedestrian, driver, cyclist, etc.), and I make a hand gesture. (2.08 ± 1.20)

A4 I have gotten angry with a driver and hit their vehicle. (1.25 ± 0.70)

Positive Behaviors (P, Reverse-scaled items)

P1 I thank a driver who stops to let me cross. (5.55 ± 0.82)

P2 When I am accompanied by other pedestrians, I walk in single file on narrow sidewalks so as not to bother the pedestrians I meet. 4.05 ± 1.46)

P3 I walk on the right-hand side of the sidewalk so as not to bother the pedestrians I meet. (4.04 ± 1.48)

P4 I let a car go by, even if I have the right of way, if there is no other vehicle behind it. (3.60 ± 1.49)

Questions were presented in random order. PBQ was the first complete questionnaire to study a broad range of aspects of pedestrian behavior on the road for all age groups. This questionnaire was originally developed by Tom and Granié (2011) with 47 questions. We chose the short version, which is considered as reliable as the long version according to Deb et al. (2017) and Tom and Granié (2011), in order to avoid demotivating participants. The 20 questions are categorized into five items as followed:

The participants were required to answer the questions using a 6-point Likert scale (1-very infrequently or never, 2-quite infrequently, 3-infrequently, 4-frequently, 5-quite frequently, 6-very often or always).

2.2 Survey administration and participants

The survey was created using LimeSurvey (Engard, 2009; Jayasundara et al., 2010; LimeSurvey Project Team, 2012) and administered online to the French population through mails and social media (Facebook and Twitter).

The questionnaire was received by 1,447 participants, 610 of whom answered the entire questionnaire (all three steps; response rate: 42.15%). The resulting dataset was used in our analyses (N=610). The average completion time was 334 ± 310 seconds. 71% of participants were women, and the mean age was 35 ± 14 years (min = 16, max = 84). The geographic repartition of the population is as follows: 7 zones with Hauts-de-France (N=17), Ile-de-France (N=100), Ouest (N=92), Est (N=219), Sud-Ouest (N=18), Sud-Est (N=75), Sud (N=89). Population size of cities was 237535±252149 (min=46; max=2220445). These factors were included in statistical analyses to avoid selection biases selection.

As per power, computation of sample size using G*Power version 3.1.9.7, reported 262 participants were required for this research at an α value 0.05, power equal to 0.95 and with effect size 0.20.

2.3 Research ethics

All data were anonymous, and participants were given sequential numerical identities corresponding to the moment they answered the questionnaire. Participants could obtain information about the study and its results by contacting the authors via an email address provided at the end of the questionnaire. We followed the ethical guidelines of our institution (CNRS-IPHC, Strasbourg, France). This study received ethical approval from the road security direction of the French Homeland Ministry (Ref: CNRS190529).

Formal written agreement or parental consent was obtained from the five people photographed for the eye images. They were aware of, and agreed to, the intended use of the photographs in the questionnaire and in the publication.

2.4 Statistical analyses

We first analysed feeling scores about images, we then analysed the PBQ questionnaire and we lastly assessed the link between these two groups of variables.

Analyses of feelings about images

We first calculated the mean score for each question and each image. This score ranges from 0 to 1, where 0 indicates 100% of participants answering “No” to the question and 1 corresponds to 100% answering “Yes”. The scores concerning all three answers [Yes, maybe, No] and the scores concerning only two answers [Yes, No] are correlated at 95% (linear regression, P<0.0001, R² = 0.95, N = 80). For this reason and in order to simplify statistical analyses, we only used the [Yes, No] answers.

A Pearson correlation test was used to check the correlation between each feeling. We then performed a principal component analysis (PCA), followed by Hierarchical Clustering on Principal Components (HCPC) in order to assess which images resulted in higher scores. PCA is a statistical method of reduction of variables in new dimensions usually meaningful. Variables are automatically corrected to be comparable (mean and range). Dimensions (groups of different variables) with eigenvalue superior to 1 are kept. We examined loadings of each variable on each dimension. The loadings are interpreted as the coefficients of the linear combination of the initial variables from which the principal components are constructed. The loadings are equal to the coordinates of the variables divided by the square root of the eigenvalue associated with the component. The goal of HCPC is to identify groups (i.e. clusters) of similar objects within a data set of interest. Following these analyses, the incitation question was excluded from the next statistical tests (see results) and a mean score was calculated for each participant, combining all four remaining questions.

We assessed whether this score is influenced by the following dependent variables: the age (continuous variable), gender (factor: Male or Female), geographical zone of their city (factor with seven zones) or the city population size (as a factor categorized according to quartiles). A general linear model (glm) with a normal law was used with the R package “MultComp” for multiple comparisons (Bretz et al., 2016). A separate GLM tested the interactions between age/sex factors and the expression of eye images (i.e., neutral, angry, friendly). The conditions of application (normality and homoscedasticity of residuals) were graphically verified.

PBQ analyses

The 20 PBQ questions with values from 1 to 6 were analyzed using a PCA with a varimax rotation (Package R Psych; Revelle, 2011; Revelle and Revelle, 2015), following the procedure explained in Granié et al. (2011, 2013). Varimax rotation is used to simplify the expression of a particular subspace in terms of just a few major items each. This means that the Varimax rotation turns the variables on the dimensions in order to maximise the explained variance. In order to fit with the PCA axes of these studies, we set a maximum number of four loadings (as a preliminary analysis of five PCA dimensions shows a division of the positive behaviors in dimensions 4 and 5, we combined both dimensions as described by Granié et al., 2011, 2013). The coordinates of participants in each dimension (five with loadings higher than 1.00) were then compared with the eye images mean score using a Pearson correlation test. GLM analysis was also carried out to test coordinates with gender, age and city data of participants using. The four dimensions were scaled and normalized using the scale function in R. For the “zone” variable (n = 7), an Anova followed by a Tukey posthoc test (df = 6, 603, confidence level = 0.95) was performed on the GLM residuals.

Lastly, we identified correlations between the eye image questionnaire score and the four PBQ dimensions.

All tests were carried out with R 3.6 (R Development Core Team, 2009). The significance level was set at 0.05. Results are indicated with mean ± stdv.

3.1 What do participants feel when seeing the eye images?

Whatever the eye image, the mean score for the ‘Observation’ question is 0.33 ± 0.14 (meaning that 33% of participants answered ‘Yes’ to this question and therefore feel observed). The mean score for the ‘Fear’ question is 0.05 ± 0.07. The mean score for the ‘Inhibition’ question is 0.10 ± 0.08. The mean score for the ‘Surprise’ question is 0.26 ± 0.12. Finally, the mean score for the ‘Incitation’ question is 0.02 ± 0.01. Scores for each image and each question are provided in Table 1 and are shown in Figure 1.

Pearson’s correlation tests between scores ( Figure 2) showed a high correlation between all feelings (r > 0.8, p < 0.001) excluding incitation (r < 0.39, p > 0.05). Indeed, a PCA showed that dimension 1 was composed of four feelings (Observation, Fear, Inhibition and Surprise, r > 0.85 see https://doi.org/10.5281/zenodo.5745446 for details) and explained 72.73% of variance in the scores, whilst dimension 2, which was solely composed of the Incitation feeling (r = 0.89), explained 19.28% of variance. These results validate our questionnaire with participants answering coherently following our predictions. A HCPC following this PCA produced five clusters, without including picture 2 (Child’s friendly eyes). When the incitation question was removed from analyses, dimension 1 (r > 0.88) explained 87.06% of score variance whilst dimension 2 (r < 0.45) explained only 7.28%. The resulting HCPC produced four clusters, with picture 2 regrouped with another cluster. In view of these results and the aims of our study, we decided to discard the incitation question for the following analyses. The images were then grouped into four clusters. The first cluster includes the flower, and a man and a woman with neutral expressions. It has a lower influence on the scores (see green elements in Table 1; participants do not feel surprised, scared, observed or inhibited). The fourth and last cluster (see red elements in Table 1) is composed of the child and a man with an angry expression. This cluster shows the highest scores (i.e. participants felt observed, scared, surprised and inhibited). Gender is therefore equally distributed in these four clusters, but the picture of a child’s eyes has a strong effect on feelings (ranked 10, 14 and 15 on the 16 images, Table 1). Moreover, whilst the first cluster includes only ‘neutral’ images and the second cluster includes ‘friendly’ and ‘neutral’ images, all the angry images are found in the third and fourth clusters, meaning that these expressions lead to stronger feelings.

Figure 2. Chart of correlations between the feelings scores.

Values indicate the Pearson correlation. Stars indicate statistical significance (Absence = p > 0.05; *** = p < 0.001).

3.2 How did the influence of eye images on the feelings of participants vary according to their sociodemographic factors?

The mean score for each participant was influenced by gender and age ( Figure 3) but not by the geographical zone or the city population size (see Table 2 for statistical values). Men have a lower score than women, meaning that they feel less scared or observed by eyes. Age negatively influences the score, meaning that older participants have a lower score and feel less observed or scared than youngers. When we tested the effect on the mean score of the interactions of age and sex with the eye expressions (i.e., neutral, angry or friendly), we did not found any interaction between age and eye expression (|t-value| < 1.086, p > 0.277) or between sex and eye expression (|t-value|< 1.449, p > 0.147).

Figure 3. Influence of gender (a.) and age (b.) on the mean score of the feelings of participants on seeing the eye images.

3.3 Answers of participants to the pedestrian behavior questionnaire

The 20 PBQ questions were analyzed in the same way as those in referenced studies (Deb et al., 2017; Granié et al., 2013; Tom and Granié, 2011). Results are in accordance with the precited studies (see https://doi.org/10.5281/zenodo.5745446 for details). The results explain 49.7% of total variance. The Kaiser-Meyer-Olkin measure of sampling adequacy (indicating the proportion of variance in your variables that might be caused by underlying factors with high values, close to 1.0, generally indicate that a factor analysis may be useful with your data, Trujillo-Ortiz et al., 2006) was satisfactory and good (0.80) and the Bartlett’s test of sphericity was significant (p < 0.0001) indicating a homogeneity of variances. The first dimension explains 15.7% of variance and corresponds to “lapses” (following a loading above 0.4). The second dimension explained 14.7% of variance and corresponds to “transgressions”, meaning violations and errors. The third dimension explains 11.2% of variance and corresponds to “aggressive behaviors”. Finally, the fourth dimension explains 8.2% of variance and corresponds to “positive behaviors”.

3.4 How did the PBS axes influence participants according to their sociodemographic factors?

All statistical results are indicated in Table 3. Posthoc Tukey multiple comparisons are detailed in https://doi.org/10.5281/zenodo.5745446. The “lapses” dimension is only influenced by the city population size, with a larger city population size leading to higher occurrence of unintentional deviation from rules. The “transgressions” dimension (violations and errors) is influenced by age, with younger people making more transgressions, and also by geographical zones, with a higher transgressions score for the “Sud” (South) Zone than the “Est” (East) area (however, see the Tukey test for further details). The “aggressive behaviors” axis is influenced only by the zone but the posthoc test revealed no differences. Finally, the positive behaviors dimension is influenced by gender, with women showing more positive behavior than men do.

3.5 Link between feelings on seeing the eye images and PBQ axes

We did not identify any statistical correlation between the eye image questionnaire score and the four PBQ dimensions ( Figure 4). Lapses dimension: r = 0.05, t = 1.2843, df = 608, p-value = 0.1995; Transgressions dimension: r = 0.02, t = 0.47498, df = 608, p-value = 0.635, Aggressive behaviors dimension: r = -0.07, t = -1.8156, df = 608, p-value = 0.06993; Positive behaviors dimension: r = 0.05, t = 1.129, df = 608, p-value = 0.2593.

Figure 4. Relation between the eye image feeling score and the four dimensions of the pedestrian behavior questionnaire (PBQ).