University of Amsterdam
Tineke Slotegraaf, under supervision of
Andries van der Leij
Trait Judgments: How Correctly are
They?
Combining Trait Judgments With Real
Personalities
ABSTRACT
We judge other people by their appearances, but are our judgements correct? Earlier research (Dotsch and Todorov, 2011; Engell et al, 2007; Oosterhof and Todorov, 2008) focussed on the judgments of faces, without knowing anything about the people on the pictures. We tried to answer this question by letting 100 participants judge pictures of about 700 people who participated in the ID1000 project (this project was set up to see the differences in brains and behavior in a group representing the Dutch population). In the ID1000 project not only pictures were taken, they also collected data through questionnaires and a structural MRI scan. The judgements given to these people were correlated with the structural MRI scan, the significant brain area was extracted, and this area was used in a regression with the Eigenfaces we generated from the pictures. The last step was to regress the brain area to the trait judgements to see which traits correlated with the brain structures and faces. This was done for males and females separately. For females the significant trait was extraversion, and for males the significant traits were dominance and openness.This research was purely explorative. We did not know if we would find a brain structure or which one it would be. We did find a deeper cause for trait inference, but replication is needed to get better results and a better understanding of which facial features are used in trait inference.
INTRODUCTION
It is quite common to have an opinion about someone before knowing anything about that person. This is like judging a book by it’s cover. But is this judgement based on anything? Can we trust it to be correct?
The most expressive part of the human body is the face. It is not only used to express emotions, research also showed complex trait inferences can be made consistently by only looking at faces of people (Dotsch and Todorov, 2011; Engell et al, 2007; Oosterhof and Todorov, 2008; Penton-Voak et al, 2006; Said et al, 2010; Said and Todorov, 2011; Todorov and Engell, 2008; Todorov et al, 2011, 2013, Willis and Todorov, 2006). Oosterhof and Todorov (2008) published one of the most influential articles in this field, where they showed that the variance of the different traits can be explained by two axes: trustworthiness and dominance. These axes resemble those found by another study where they did not judged faces, but traits were clustered in eight groups(Wiggins, 1979). Trait inference
is even influencing our way of how to deal with people, like on who to vote in an election (Ballew and Todorov, 2007).
Specific facial characteristics influence our opinion regarding trustworthiness and dominance (Dotsch and Todorov, 2011; Said and Todorov, 2008; Todorov et al, 2011). However, these associations are between facial characteristics and the way these characteristics lead to trait inferences. There is little known about the basis of these specific facial characteristics influencing trait inferences; is it socially, or does the face really tell something about the personality behind it. Research of Penton-Voak et al (2006) suggests there is a basis that can play a role in all this, for females extraversion and for males openness inferences do correspond to the traits of the people judged.
In our research we focussed on the participants who have been judged, instead of the judgements of the faces. We checked correlations between the judgements of the faces, the brain structure and the eigenvalues of the faces - see Dotsch and Todorov
2
http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSLVBM), an optimised VBM protocol (Good et al., 2001) carried out with FSL tools (Smith et al., 2004). First, structural images were manually brain-extracted and grey matter-segmented before being registered to the MNI 152 standard space using non-linear registration (Andersson et al., 2007). The resulting images were averaged and flipped along the x-axis to create a left-right symmetric, study-specific grey matter template. Second, all native grey matter images were non-linearly registered to this study-specific template and “modulated” to correct for local expansion (or contraction) due to the non-linear component of the spatial transformation. The modulated grey matter images were then smoothed with an isotropic Gaussian kernel with a sigma of 1 mm. Finally, voxelwise GLM was applied using permutation-based non-parametric testing, correcting for multiple comparisons across space.”
This experiment
Participants
100 students from the University of Amsterdam and the Radboud University Nijmegen participated in this study. Of these student 31 were male. All participants had normal or corrected to normal vision. The age range of the participants was 17-40 (mean 20.59, standard deviation 2.97)
Design
Participants were instructed on paper that they would grade people on 1 of our 10 traits (Intelligence, trustworthiness, competence, dominance, attractiveness, neuroticism, extraversion, openness, conscientiousness, agreeableness) by seeing their picture and press a button (1-9). They were told to go with their gut feeling and respond as quickly as possible. The task was performed on a computer. The first dialogue box would tell the participant which question (about one of our traits) he had to answer during the experiment. A button press would start the experiment. Underneath each picture a scale of 1-9 was printed with the meaning of both ends of the scale.
The pictures were divided in 13 blocks of 50 or 51 pictures per block. The first block always had 50 pictures. The pictures of this first block were all graded twice in the experiment. Every new block would have some of these pictures (3 or 4) at random. This was done to see if people would give higher or lower grades on average later on in the experiment. After blocks 2, 5, 8, and 11 a break was implemented. The pictures were on the screen for as long as it took the participant to answer. Between pictures was a 200 ms break to make sure participants would see a blank screen between pictures and know it would be a new face to grade. The design is visualized in figure 1.
The blocks were put in 5 different orders to make sure any effect would not be because of the order participants got to grade the pictures. This way every order occurred twice for every question. Within the blocks the pictures were completely randomized every time the experiment was started. The experiment was programmed in Inquisit 3.
(2011), Said and Todorov (2008), and Todorov et al (2011) for more information about eigenfaces - to find if trait inferences gives us true information about someone’s personality.
The aims of our study are to find if the judgments of faces made by participants resemble earlier findings (Oosterhof and Todorov, 2008) and to compare the findings with the brain structures and eigenfaces of the participants judged in order to find a deeper cause for this trait inference. The unique rich database of about 700 faces, MRI scans, and extensive self-report of a sample representative of the Dutch population makes it possible to answer whether judgments from faces correlate with characteristics of the individual. If there is a correlation between judgments of faces and the true personality belonging to the person judged, it means we have developed a trustworthy tool throughout evolution to evaluate personality based on someone’s face.
METHODS
ID1000
Participants
990 (nine hundred ninety) Dutch participants representing the Dutch population were tested in the ID1000 program. All participants had normal or corrected to normal vision. The age range of the participants was 20-34 (mean 22.74, standard deviation 1.79). Subjects were recruited by Motivaction and are representative of the Dutch population in terms of educational level.
Image acquisition and preprocessing
Structural brain scans were acquired of each subject using the same Philips Achieva TX 3-Tesla scanner and protocol at the Spinoza Center of the University of Amsterdam. For each subject, three identical three-dimensional T1 turbo field echo (TFE) scans were carried out (repitition time(TR) = 8.124 ms, echo time(TE) = 3.71 ms, flip angle = 8$\,^{\circ}$, field of view = 256 x 256 x 160 mm, slice thickness = 1 mm, number of slices = 160, voxel size = 1 x 1 x 1 mm).
Pictures
Each of the 696 participants signed a new informed consent, so we can use their pictures in this experiment. The subset of 696 participants (350 Male) that gave consent were used for this study. The pictures were taken in a room with windows and a busy background. The camera used to take the pictures was of type Fujifilm finepix z70.
Design
Participants filled in some questionnaires at home and came to the university of Amsterdam to be tested for a day. During this day they performed tasks, filled out more questionnaires and got a structural MRI scan. The present study was approved by the ethical board of Psychology at the University of Amsterdam and all the subjects gave their informed consent in a written form.
Dataprocessing
Photo preprocessing
Each picture was loaded into photoshop separately. White balance was corrected by taking a spot in the pictures which would be visible in every picture and white (It would have worked better with white, black, and grey, but black and grey were not visible in every picture). This spot was highlighted and blurred to get the best possible white in the picture. Photoshop changed the colours in the picture so our square would be pure white and the function autotone made the colours look a bit more natural again.
The background was deleted by selecting everything but the person with the quick select tool, reverse the selection, and click on refine edge. Refine edge is used to make sure the hair in not cut of but we keep a bit of the fluff this way. The background was deleted and replaced by a grey colour (220). We cropped the pictures to 900 px, because otherwise starting up the experiment was too slow. These pictures were used in the experiment.
To be able to make a mean of the faces, the faces had to be on the same place in every picture. We used Matlab code from Oliver Langner (also used for Langner et al (2010)) an adapted it to our situation. Oliver’s scripts were made for several models whom were all photographed from different angles and with different emotions. We only had one picture for each model and therefore had to change all the scripts to our situation. The pictures were cropped by clicking the top (within the hair, almost in the background) and bottom (the chin) of each head to make sure all faces had the same height. After that we clicked each left eye of the face in the pictures and the computer aligned the faces from this starting point. All faces were corrected by hand afterwards.
When we had all the faces it was time to make the eigenfaces
using the Matlab code on http://www.cs.princeton.edu/~cdecoro/ eigenfaces/. This is basically done by doing a factoranalysis on all the pictures-mean of all the pictures. There will than be as many eigenvectors as there are pictures. We did this for Men and Women separately.
Initial analysis
We used SPSS to look for outliers in our data. We used for each participant the times a particular response was given (i.e. how often gave the participant a picture a grade of 5) for each possible response, the mean of the response as well as the max and variance of the response, the mean reaction time, and the r2 for the pictures
which were graded twice. The r2 was calculated over the difference
in response or reaction time and was calculated both normally and absolute.
SPSS calculated for all of these different inputs (there were 17) if there were outliers. We found a lot of outliers. We graded them with 1 or 3 points respectively of how significant the outlier was. If a participant scored 5 points or higher he/she was removed from further analysis.
We also removed individual scores if participants responded faster than 120 ms (they couldn’t have made an opinion in that time) or longer than 10 seconds (it wouldn’t be a first opinion anymore).
After deleting data we calculated the mean score for each picture for each trait and did separate t-tests for each trait to check for gender differences.
Brain analysis
We first did a factor analysis on the data we gathered in this experiment. The factors were use in a GLM analysis as well as the
Figure 1: The setup of the experiment. The pictures of neutral looking faces were presented to participants in the middle of the screen. After a response (1-9) was given, the program continued with a blank screen, in the same colour as the background of the pictures, for 200 ms. The 200 ms were used to make sure the next picture was received as new stimuli by the participants.
4
Figure 2: Split by gender. We did separate t-tests for all mean scores of the ten traits check for gender differences. All traits, but extraversion, are found to be significantly different for males and females. Therefore we can split the data by gender in the rest of the analysis. * = p < 0.05, ** = p < 0.01
Figure 3: The analysis. For the analysis we started with the grades that were given to the faces in the pictures. Every picture was graded by ten participants per trait, so for all pictures a mean grade for every trait was put in the analysis. After a factor analysis only three factors remained, which were then put into a VBM analysis, together with the scans of the brains of the people in the pictures and information about age, length and brain size. We performed the whole analysis for males and females separately, for both we found one significant brain area. We found for which particular traits this area was significant by using a regression analysis on the extracted brain area and visualized the faces using Eigenfaces (eigenvectors for face images) for these significant traits.
variables age, length, and brain size to sort out the effect of those three variables on the data. To find a significant area of the brain we did an F-test in the GLM over the factors.
From the Glm analysis a place in the brain was found. We used FSLview to visualize the significant areas. We used a Matlab code to extract the value of grey matter in this Region Of Interest (ROI) for each brain. This value was used for further analysis. The first regression was used to get out the influences of age, length, and brain size on this ROI. We continued by putting the residuals of this analysis in a new regression analysis to find which EigenVectors from facespace could explain the ROI-value. After this new analysis we got a Vector with gradings for all EigenVectors. With a regression of the brain data on the 10 traits, we found which traits were significant. We multiplied the new Vector with the EigenVectors for these significant traits and visualized it. This way we can see what changes in a face if it is graded high on for instance extraversion. The flowchart of this analysis is shown in figure 2.
RESULTS
Preliminary results
The faces were judged on ten traits. We first used an outlier analysis and deleted participants who were significantly different on the mean response, mean reaction time, R-squared of response, R-squared of reaction time, maximum response, variance of response or how often each number was given. We rated the significantly different participants with 1 or 3 points, dependent on how significantly different they were, and deleted the participants with the grade of a 5 or higher.
Over the rest of the data we calculated the mean for each trait for each face. These were the numbers we used in further analysis. We performed t-tests on each trait to check if males and females were judged the same way. Results of these t-tests can be found in figure 3. Almost every trait is significant, so the grades for females are different (mostly towards more friendly) than the grades for males. Therefore we decided to do the rest of the analysis for males
and females separately.
The results of the factor analysis over the ten traits can be seen in figure 4 and table 1. Figure 4 shows how much of the data the factors explain. It is normal to pick the number of factors from the ‘knee’of the graph, so we picked 3 factors for both genders. Adding more factors does not add enough explanation of the data compared to the power we would lose. In table 1 we show which traits are represented by which factors.
Brain analysis
The factors were used in a VBM analysis. The results of this analysis are shown in figure 5. For females we found a region in the top left of the brain. This region includes the middle frontal gyrus and the superior frontal gyrus. For males we found a region more in the middle of the brain with a lot of white matter, but also the Caudate and Thalamus.
EigenFaces
The gray matter of these regions found for males and females was extracted and the significant traits with this gray matter were extraversion for females, and openness and dominance for males, as can be seen in table 2. The visualizations of these traits can be found in figure 6. In these visualization we show that most people seem to be judging faces on areas around the eyes and mouth.
Figure 4: Factor analysis males and females. After three factors most data is explained, adding another factor, will not explained that much more data that it is useful. So we used three factors in further analysis to explain our data. The left graph is the factor analysis for males, the right graph is the factor analysis for females.
Table 1: Rotated component matrix Males and Females. The tables show how the traits are devided into three factors for both sexes.
6
DISCUSSION
Conclusion
We did find a deeper cause for trait inference. We found brain areas that were significant given the factors we derived from our ten traits and we found changes in the mean faces for the traits (extraversion, openness, and dominance) which were significant with these brain areas.
Significant traits
The trait extraversion was significant in females, and the traits openness and dominance were significant in males. This correspond
to the findings of Penton-voak et al (2006) who looked to self report, while we looked at brainstructures.
Oosterhof and Todorov (2008) found trustworthiness and dominance to tell most and also explain the other traits judged. We did find the trait dominance to be significant for males. This trait does group together with extraversion, which is significant for females, in our factor analysis. So we can conclude we found the trait dominance in both our groups. Which gives us the information that facial features do show partly what is going on in the brain and devines our personality.
The other axis, trustworthiness, and the line between
Figure 5: VBM analysis. (A) The area found by applying VBM on the brains of the male subjects. The coordinates of this area are: x= 45 y = 65 z = 37 (B) The area found by applying VBM on the brains of the female subjects. The coordinates of this area are: x = 31, y = 78 z= 61
Table 2: Regression on brain data. The tables show stich traits are put into the repression analysis and which are left out. Left is the regression score from each trait on the significant brain area for males. Right is the regression score from each trait on the significant brain area for females.
Figure 6: Visualization of the significant traits. The visualizations show the mean face of females or males with added and extracted the significant eigenface values for the trait. This way we can see best what changes in a face given a trait. (A) Females with the trait extraversion. (B) Males with the trait openness. (C) Males with the trait dominance.
8
the two axes, competence, was not found to be significant in the grey matter of the brainstructure. This does not mean there is no underlying brainstructure defining trustworthiness and competence, but just that this is not the grey matter. It could be based on other brain measures of DNA, further research is necessary to check if there is another biological basis for these traits and the facial features used for trait inference. It could also be possible there is no biological basis for these traits, but it is the way we use our brain that defines them.
Brain areas
The middle frontal gyrus and superior frontal gyrus, which were found to be significant for females, are involved in judging one\ textquoteright s own personality (Zhu et al, 2007). In a study of Canli et al (2001) they found this region to be significant in women for the trait extraversion. We replicated this finding. They explain this region is associated with emotion. Extraverts tend to have more positive emotion and hence in a structural scan they have a significant difference in the frontal gyrus.
The thalamus and caudate which were found to be significant in males both belong to the dopaminergic system. This system plays a role in the regulation of anxiety and is associated with novelty seeking behavior and correlates with the introversion/extraversion scale(Laakso et al, 2003, Johnson et al, 1999). Openness is about liking new experiences, which is similar to seeking novelty and Ebstein et al (2000) explains extraversion is also a positive aspect of novelty seeking behavior. Since dominance and extraversion shared a factor in our analysis, we can state that they were judged quite similar. Therefore we found brain areas which could be better explained with extraversion than with dominance as the significant trait.
The caudate was found to correlate with anxiety on a personality test (Laakso et al, 2003). This could be a reverse correlation with openness, since someone who is anxious about an event will probably not be open for the new experience that event will provide. It is also an aspect of introverts and thus a reverse correlation with extraversion. Introverts feel more anxious when in new experiences and around people than extraverts, who seem to be seeking out these situations.
Pictures
In previous research (Willis and Todorov, 2006; Engell et al, 2007; Oosterhof and Todorov, 2008; Langner et al, 2010) taking the pictures was the main goal to ask participants to come to the lab. In our research it was added at the end of an already full program. This means we did not have a full protocol of how people should be photographed other than with a neutral look.
Others used protocols to make sure a face was photographed and not a whole personality. They did this by making sure people did not wear makeup or eyeglasses, were wearing their hair back, and were photographed in the same color tshirt. In our pictures, people were wearing their own clothes, hairstyles and sometimes makeup. By having all these elements in the pictures it is possible people
judged not only based on the face, but also based on the clothing- and hairstyles of the people in the pictures.
Another improvement on the pictures could be to judge them first on emotion. Previous research (Engell et al, 2007) did this to make sure that the right emotion was shown. They did want multiple emotions and made sure participants practised these emotions in their face before taking their picture. We only wanted neutral looks, but our selection of pictures could have been improved if we had first tested them on emotion to see if their look was neutral.
Visualization
Our visualizations are hard to interpret. This is partly because of the noisy pictures we used in the experiment. By having different hairstyles the mean pictures will become blurry where the hair is in the pictures. This means the visualization will already be a lot better if we would have had a full protocol for how the pictures should be made.
Another possibility to make the visualization better is if we could translate our faces to the facespace of Oosterhof and Todorov (2008). This face space can be used to make visualizations of all kinds of faces in all directions of the 50 dimensions. If we could use a space like that we could get clear faces and show the difference from a person who is extravert compared to the average face. Than we could also better explain what facial features change and how we see traits in a face.
Follow-up research
Future research can try to replicate our findings according to the brain area. A possibility to improve our research would be to use better pictures or pictures that can be translated to a clear face space, like the one from Oosterhof and Todorov (2008). That way the findings of which facial features represent a deeper brain structure which causes a person to have a certain trait would be clearer.
Another idea would be to also translate these findings to the personality of the participants judged. To see if we didn’t only find a deeper cause for trait inference, but also if the person really has that particular trait. If that can be found we can be sure that we judge people quite correctly on some traits and maybe incorrectly on other traits.
In this paper we found a deeper cause for trait inference for the traits extraversion, openness, and dominance, but future research is needed to replicate our findings and give a clearer view of which facial features give us information about these personality traits.
REFERENCES
Andersson, M. Jenkinson and S. Smith (2007) Non-linear registration, aka Spatial normalisation. FMRIB technical report TR07JA2 from www.fmrib.ox.ac.uk/analysis/techrep
Ballew, C. C., & Todorov, A. (2007). Predicting political elections from rapid and unreflective face judgments. Proceedings of the National Academy of Sciences, 104(46), 17948-17953.
Canli, T., Zhao, Z., Desmond, J.E., Kang, E., Gross, J., Gabrieli, J.D.E. (2001) An fMRI Study of Personality Influences on Brain Reactivity to Emotional Stimuli. Behavioral Neuroscience 115-1: 33-42
Dotsch, R., Todorov, A. (2011). Reverse Correlating Social Face Perception. Doi: 10.1177/1948550611430272
Douaud G, Smith S, Jenkinson M, Behrens T, Johansen-Berg H, Vickers J, James S, Voets N, Watkins K, Matthews PM, James A (2007) Anatomically related grey and white matter abnormalities in adolescent-onset schizophrenia. Brain 130:2375-2386. Ebstein, R.P., Benjamin, J., Belmaker, R.H. (2000) Personality
and polymorphisms of genes involved in aminergic neurotransmission. European Journal of Pharmacology 410: 205-214
Engell, A.D., Haxby, J.V., Todorov, A. (2007) Implicit Trustworthiness Decisions: Automatic Coding of Face Properties in the Human Amygdala. Journal of Cognitive Neuroscience 19-9: 1508-1519 Good CD, Johnsrude IS, Ashburner J, Henson RN, Friston KJ,
Frackowiak RS (2001) A voxel-based morphometric study of ageing in 465 normal adult human brains. NeuroImage 14: 21-36.
Johnson, D.L., Wiebe, J.S., Gold, S.M., Andreasen, N.C., Hichwa, R.D., Watkins, L., Boles Ponto, L.L. (1999). Cerebral Blood Flow and Personality: A Positron Emission Tomography Study Am J Psychiatry 156: 252-257
Laakso, A., Wallius, E., Kajander, J., Berman J., Eksola, O., Solin, O., Ilonen, T., Salokangas, R.K.R., Syvälahti, E., Hietala, J. (2003) Personality Traits and Striatal Dopamine Synthesis Capacity in Healthy Subjects. Am J Psychiatry; 160:904-910. doi: 10.1176/appi.ajp.160.5.904
Langner, O., Dotsch, R., Bijlstra G., Wigboldus D.H.J Hawk, S.T., van Knippenberg, A (2010) Presentation and validation of the Radboud Faces Database. Cognition & Emotion 24-8: 1377-1388
Oosterhof, N.N., Todorov, A. (2008) The functional basis of face evaluation. PNAS 105-32: 11087-11092
Penton-Voak, I.S., Pound, N., Little, A.C., Perrett, D.I. (2006) Personality Judgments from Natural and Composite Facial Images: More Evidence for a “Kernel of Truth” in Social Perception. Social Cognition 24-5: 607-640
Said, C.P., Dotsch, R., Todorov, A. (2010) The Amygdala and FFA track both social and non-social face dimensions. Neuropsychologia 48: 3596-3605
Said, C.P., Todorov, A. (2011) A Statistical Model of Facial Attractveness. Doi: 10.1177/0956797611419169
Smith SM, Jenkinson M, Woolrich MW, Beckmann CF, Behrens TE, Johansen-Berg H, Bannister PR, De Luca M, Drobnjak I, Flitney DE, Niazy RK, Saunders J, Vickers J, Zhang Y, De Stefano N, Brady JM, Matthews PM (2004) Advances in functional and
structural MR image analysis and implementation as FSL. NeuroImage 23 Suppl 1: S208-219.
Todorov, A., Engell, A.D. (2008) The role of the amygdala in implicit evaluation of emotionally neutral faces. Doi: 10,1093/scan/ nsn033
Todorov, A., Dotsch, R., Wigboldus, D.H.J., Said, C.P. (2011) Data-driven Methods for Modeling Social Perception. Social and Personality Psychology Compass 5/10: 775-791
Todorov, A., Mende-Siedlecki, P., Dotsch, R., (2013) Social judgments form faces. Current Opinion in Neurobiology 23: 1-8 Wiggins, J.S. (1979) A Psychological Taxonomy of Trait-Descriptive
Terms: The Interpersonal Domain. Journal of Personality and Social Psychology 37-3: 395-412
Willis, J., Todorov. A. (2006) First Impressions Making Up Your Mind After a 100-Ms Exposure to a Face. Psychological Science 17-7: 592-598
Zhu, Y., Zhang, L., Fan, J. (2007) Neural basis of cultural influence on self-representation. NeuroImage 34-3: 1310-1316
