Anchoring Effects in Facial Attractiveness

Introduction

Previous work on the evolutionary psychology of facial attractiveness has emphasised the importance of appearance in mate selection (for reviews, see Fink & Penton-Voak, 2002; Little, Jones, & DeBruine, 2011; Rhodes, 2006). This influential line of work has often drawn links between facial characteristics and genetic or developmental characteristics of individuals. For example, facial symmetry (Wade, 2010), averageness (DeBruine, Jones, Unger, Little, & Feinberg, 2007) and sexual dimorphism (Komori, Kawamura, & Ishihara, 2009) are hypothesised to strike us as attractive because they are markers of genetic health. Such stable aspects of appearance may account for between-person variability in attractiveness (Jones et al., 2001; Lorenzo, Biesanz, & Human, 2010; Morrison, Morris, & Bard, 2013; Rhodes et al., 2011), but they cannot account for within-person variability, across which biological makeup remains constant. Different photos of the same face often make contrasting impressions on the viewer, even when they are drawn from the natural range of the person’s variability. In an early demonstration of this, Jenkins, White, Van Montfort, and Burton (2011) obtained attractiveness judgements for 20 different, ambient photos for each of 20 unfamiliar faces. They found that within-person variability in attractiveness was large compared with between-person variability, such that photo selection could reverse underlying person-level preferences.

The large variability for each face raises the question of how stable facial impressions of a person are formed (Todorov, Olivola, Dotsch, & Mende-Siedlecki, 2015; Vernon, Sutherland, Young, & Hartley, 2014). Previous work has shown that impressions tend to be formed very quickly and that early impressions change little with additional exposure (Willis & Todorov, 2006). But that conclusion was based on additional exposure to the same image. Given that a person’s appearance changes from one image to the next, we asked how impressions from different photos of the same face are integrated during impression formation. One possibility is that integration is indifferent to order of encounter, such that impressions from each photo carry equal weight. This assumption is implicit in models of face learning that simply pool the available visual information (Burton, Jenkins, Hancock, & White, 2005; Burton, Kramer, Ritchie, & Jenkins, 2016; Jenkins & Burton, 2008). However, a plausible alternative is that order matters, such that impressions from early or recent photos dominate the impression that is formed. In rating a series of faces, order matters for identification (Liberman, Fischer, & Whitney, 2014) and for affective judgments (Taubert, Alais, & Burr, 2016). In facial attractiveness, previous studies observed positive serial dependency in that the evaluation of the current face is pulled towards the more recent items (Kondo, Takahashi, & Watanabe, 2012, 2013; Kramer, Jones, & Sharma, 2013; Taubert, Van der Burg, & Alais, 2016). Based on these findings, we would expect a similar effect for judgments of within-person variation. However, there is a clear difference between judging faces from different identities and judging the attractiveness of one individual based on different photos. The repeated judgment of the same person requires a process of impression formation, which is not the case in the studies on between-face comparisons. Such impression formation in face learning and person acquaintance is essentially different from rating faces of different people. Because negative and positive serial effects are both common in face perception (Taubert et al., 2016; Webster, Kaping, Mizokami, & Duhamel, 2004), serial dependency in person formation of facial attractiveness is a pending empirical question.

To empirically test order effects in impression formation, we appropriated the pioneering work based on character descriptions from Solomon Asch (1946). In his experiment, participants were presented with character descriptions in the form of one of two lists. List A began with positive descriptors (e.g. ‘intelligent’, ‘industrious’), and ended with negative items (e.g. ‘stubborn’, ‘envious’). In List B, the order of these items was reversed. Participants then rated the person’s character on a range of other traits. Key to this design is that Lists A and B contain exactly the same information. Only the order of presentation is changed. If participants simply pool the available information, then trait ratings for both lists should be the same. In fact, impressions of personality were anchored to the early items: A positive start left a positive impression and a negative start left a negative impression. Instead of character descriptions, we presented different photos of the same person in either ascending or descending order of attractiveness. We then probed the impression that participants formed in each case. We used unfamiliar faces in these studies, to ensure that the participants have no preconceptions about how attractive the person is. For comparison, we also included familiar faces, so that we could examine order effects in faces for which impressions have already been formed. In all cases, we focused solely on within-person effects, constructing image sequences from different photos of the same person. To ensure that these images sampled the natural range of variability for each person, we used ambient photos as stimuli throughout (Jenkins et al., 2011).

Experiment 1

The purpose of this experiment was to establish whether anchoring effects in facial attractiveness can be elicited using multiple images of the same person. We expected that a test image would be rated as more attractive when the preceding sequence was ordered from most-to-least attractive (descending order) and less attractive when the preceding sequence was ordered from least-to-most attractive (ascending order).

Method

Participants

Thirty-eight consenting undergraduates (30 women and 8 men, M_age = 21.68 years, age range: 18–27 years) took part in exchange for course credit.

Design and stimuli

We compiled a list of 14 UK national celebrities and 14 German national celebrities (50% women, 50% men; see Appendix for a list of names). The UK celebrities served as unfamiliar faces and the German celebrities as familiar faces. For each celebrity, we collected 15 different photos through an online search. We used only high quality images that showed the face in roughly frontal aspect and free from occlusions. In a pre-test, 39 undergraduates (26 women and 13 men; M_age = 25.44 years) rated each photo for attractiveness using a Likert scale from 1 (not attractive) to 7 (very attractive). Participants also rated the familiarity of the celebrities on a scale from 0 (unknown) to 4 (well known) (UK: M = 0.03; GER: M = 3.54). We then ordered the photos by attractiveness, setting aside the median photo of each celebrity for use as a test image. The mean value of the 28 test images ranged from 1.97 to 5.23 with a mean of 3.45 and a standard deviation of 0.88. From the remaining photos, we selected the six photos that covered the range from least attractive to most attractive with the most evenly spaced ratings. Our experimental manipulation was the order of this image sequence. The photos could be presented in either ascending order of attractiveness (from least to most attractive) or in descending order of attractiveness (from most to least attractive; see Figure 1). Participants were randomly assigned to the ascending or descending condition, and each participant rated all of the faces.

OPEN IN VIEWER

Figure 1.

Example photos of one face (not used in the experiment) in ascending and descending order of attractiveness.

Procedure

On each trial, participants passively viewed six photos of the same person. Each photo was presented for 2 seconds. This was followed by the test image, which was the median attractiveness image as determined in pre-test. The test image disappeared, and participants were instructed to rate ‘the attractiveness of this person’ using a Likert scale from 1 (not attractive) to 7 (very attractive). Participants were instructed that all seven photos in a given sequence (the six ordered photos and the test image) showed the same person. This identity aspect was additionally strengthened by presenting the name of the celebrity before each sequence and on the final rating screen. Participants completed 28 sequences in this manner and the order of the sequences was the same for each participant. At the end of the session, each participant rated each photo for familiarity and attractiveness, verifying the ratings from the pre-test phase. The entire test session lasted approximately 20 minutes.

Experiment 2

Although the results of Experiment 1 are consistent with anchoring effects, an alternative explanation could be contrast effects between the sixth and seventh (test) image. In the ascending condition, the test image could appear attractive relative to the final image in the sequence and vice versa for the descending condition. To better understand impression formation over the entire sequence, we next asked participants to rate the attractiveness of the person after each image. Anchoring predicts that the first item should influence ratings of all subsequent photos. The contrast account predicts that the transition from the ordered sequence to the test item is the locus of the effect.

Results and discussion

We ran a 2 × 2 (Image order × Familiarity) mixed-design analysis of variance with the mean across the sequence of the seven images as the dependent variable. We found a significant main effect of image order, F(1, 35) = 4.28, p = .046, η_p² = .11, Cohen’s d = 1.32, replicating the main finding from Experiment 1. Ratings were higher when images were presented in descending (M = 4.66, SD = 0.79) than when presented in ascending order (M = 4.06, SD = 0.8). Figure 2(b) shows that the effect of image order is present at each level of attractiveness. This pattern indicates that the effect is not a contrast effect between the penultimate photo and the final photo or any other pair of successive photos. Instead, the effect seems to be the result of an anchoring effect, as illustrated in Figure 2(a). Participants who started with the most attractive photo of a person provided higher attractiveness ratings for subsequent photos, and participants who started with the least attractive photo first provided lower ratings. Unlike in Experiment 1, we also found a main effect of familiarity, such that familiar faces (M = 4.07, SD = 1.06) received lower attractiveness ratings than unfamiliar faces overall (M = 4.58, SD = 1.02), F(1, 35) = 31.63, p < .001, d = 0.53, η_p² = .48, presumably due to the modified stimulus set. Importantly however, there was again no Image order × Familiarity interaction, F(1, 35) < 0.01, p = .938, indicating that anchoring effects for the two familiarity conditions were similar.

OPEN IN VIEWER

Figure 2.

(a) Mean overall attractiveness ratings across the image sequences (1–6) and test image (7), plotted for ascending (blue) and descending (red) order of attractiveness. The first rating serves as an anchor for subsequent evaluations. (b) The same data replotted to reverse the order of the first six images in the descending condition (for visualisation). When images were presented in descending order, the mean ratings were higher on all levels of attractiveness. Error bars show SE.

Our data also emphasise the large within-person variability seen for ratings of attractiveness (Jenkins et al., 2011). To estimate the relative contributions of within-person and between-person variability to ratings in this study, we ran two variance component analyses with random effects for participants, faces and celebrities, separately for familiar and unfamiliar faces (restricted maximum likelihood method; Hönekopp, 2006; Leder, Goller, Rigotti, & Forster, 2016). For familiar faces, we found a relative contribution of within-person variability of 22%, compared to 78% between-person. For unfamiliar faces, the relative contribution of within-person variability was higher with 36%, compared to 64% between-person. The smaller value for familiar faces could indicate that prior knowledge of appearance constrained ratings of the experimental images. Presumably, such knowledge would not eliminate within-person variation in perceived attractiveness, even when it does constrain the range. For example, viewers may be sensitive to changes in health or fatigue in familiar faces, and both of these factors influence perceived attractiveness (Axelsson et al., 2010; Rhodes et al., 2007). In our final experiment, we go beyond attractiveness ratings to ask whether the observed anchoring effects are strong enough to influence behaviour decisions.

Experiment 3

Likert scale ratings of attractiveness have a number of limitations (Hodge & Gillespie, 2003). First, in everyday social interactions, attractiveness is not assessed on a numeric scale. Indeed, processing of attractiveness seems to require no explicit judgement and only little awareness (Olson & Marshuetz, 2005; Sui & Liu, 2009). Second, subjective ratings expressed via Likert scales tend to reflect emotional value (liking) in evaluating attractiveness, rather than incentive value (wanting) (Berridge, 2004; Dai, Brendl, & Ariely, 2010). Yet incentive value seems to be especially important in the context of mate selection and sexual behaviour (cf. Rhodes, 2006). For these reasons, we next examined whether image order could affect dating decisions for unfamiliar faces.

Results and discussion

Table 1 shows the number of Yes and No responses for each image order. A χ² test revealed a significant association between image order and dating intentions, χ²(1, N = 1,046) = 4.95, p = .030. The odds ratio shows that the participants were 1.32 times more likely to consider dating the person when images were presented in descending order of attractiveness, compared with ascending order. Importantly, participants in this study were not informed that the study was about attractiveness and were never asked to provide attractiveness ratings. Nevertheless, their decisions were biased by the presentation order of images.

OPEN IN VIEWER

Table 1.

Contingency Table for Image Order and Participants’ Dating Intentions.

	Response
	Yes		No		Total
Image order	Count	%	Count	%	Count	%
Ascending	243	46.3	282	53.7	525	100
Descending	277	53.2	244	46.8	521	100
Total	520	49.7	526	50.3	1,046	100

General Discussion

In three experiments, we presented increasingly attractive or decreasingly attractive photos of a person’s face and assessed the impact of this order manipulation on measures of attractiveness. In all three experiments, impressions of attractiveness were more positive when the preceding images were presented in descending order of attractiveness (from most to least attractive) compared with ascending order of attractiveness (from least to most attractive). Experiment 1 established this anchoring effect in Likert scale ratings of attractiveness. Experiment 2 replicated the basic effect and ruled out contrasting attractiveness between the sequence and the test image as the mechanism. Experiment 3 showed that the effect is not limited to Likert scale ratings but can also influence relevant decisions.

Asch’s (1946) seminal work on anchor effects inspired a great deal of research on time-dependent decisions. The basic phenomenon has been replicated for various dimensions and in several domains (Hogarth & Einhorn, 1992; Hunt, 1941; Strack & Mussweiler, 1997; Tversky & Kahneman, 1974; Wilson, Houston, & Brekke, 1996). However, its implications for learning an individual face have not previously been noticed. A number of recent studies have examined effects of preceding faces on perception of a current face (Liberman et al., 2014; Taubert, Alais, et al., 2016) – including effects on perceived attractiveness (Taubert, Van der Burg, et al., 2016; Taubert & Alais, 2016). However, all of these previous studies have focused solely on between-person effects. Such studies can reveal sequential dependencies among individuals in a population, but they rely on comparisons across images of different people, whose faces differ in their stable traits (e.g. facial symmetry, averageness and sexual dimorphism). It is only relatively recently that within-person variability in appearance, including attractiveness, has been incorporated into theorising (Burton, 2013; Jenkins et al., 2011). And it is only when the magnitude of this variability is appreciated that ranking photos of the same face on social dimensions makes sense.

We submit that taking temporal order seriously has great potential to enhance understanding of individual face learning. Representations that are based on statistical summaries of images (Burton et al., 2005, 2016; Jenkins & Burton, 2011) do not currently account for the effects shown here. However, they could easily be modified to capture the time dimension. Doing so could help to explain how different viewers come to have different impressions of a particular face (Hönekopp, 2006). Even when viewers receive similar exposure to a particular face, we would expect each viewer to be disproportionately influenced by the images that they encountered first (Lewis, 1999; Moore & Valentine, 1998).

To inform theoretical development, it would be useful in future studies to establish the longevity of anchor effects and how easily they can be overcome by subsequent exposure. The observation that anchor effects apply to familiar faces, as well as to unfamiliar faces, implies that new images continue to inform our impressions of people long after initial impressions have settled. Presumably, tracking the impressions that familiar faces make allows us to detect relevant changes in appearance over time (e.g. looking tired, ageing gracefully). It is interesting to speculate whether anchoring effects in other domains (such as person descriptions in Asch’s original experiments) would show a similar pattern. Asch’s (1946) experiments did not include a familiarity manipulation, so we do not know. However, this is an empirical question and one that we think might be informative. Invariance to familiarity might suggest that anchoring effects reflect short-term impression tracking more than longer lasting impression formation. These findings have also real-life implications. With the rise of social media, person acquaintance based on a selection of images is increasingly important. Experiment 3 demonstrates how important decisions, like whether someone is considered a potential mate or not, are easily influenced by image choice.

Another interesting avenue would be to investigate the generality of anchor effects in face perception, and particularly whether they arise in the context of identification tasks. Different photos of the same face vary greatly in their rated likeness – the degree to which they capture the person’s true appearance (Bindemann & Sandford, 2011; Jenkins et al., 2011; Ritchie, Kramer, & Burton, 2018). Thus, an intriguing possibility is that a viewer’s ability to identify a particular face may depend on the order in which the learning images were presented (e.g. ascending versus descending order of likeness). Encountering the images in good-to-poor order of likeness may lead to better recognition than encountering the same images in poor-to-good order of likeness.

For now, we show that early items dominate impression formation, vindicating the aphorism that first impressions last. In this case, the very first impressions last a barrage of subsequent impressions and survive to influence relevant decisions.