Supplementary Information
1
Lin et al.
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
2
Supplementary Figures
1
2
3
Supplementary Fig
ure
1
:
Verifying the representativeness of the 100 selected traits and
4
face
image
s.
5
a,
Word cloud of the 973 freely generated descriptions of
the
100
selected
face
image
s
during
6
spontaneous face judgments
(see Methods). All words that appeared at least twice are shown in
7
the figure (words that appeared only once were excluded, as they were comprised mainly of
8
misspelled words or words not included in the FastText vocabulary
1
). The scale indicates
the
9
frequency
the word was mentioned
(ranging from 2 to 306 times).
b,
Uniform Manifold
10
Approximation and Projection (UMAP
2
) of
the
100 selected face
image
s
(stars)
,
frontal, neutral,
11
white
face images from multiple databases (dots that are not light blue
,
N = 632
)
,
and ambient
12
3
fac
e
images
in real world contexts
3,4
(
light
blue dots; with various angles, gazes, facial
1
expressions, lighting, backgrounds, etc.; N =
4744
). All faces were represented by the 128
2
computationally extracted feat
ures used by a state
-
of
-
the
-
art neural network for facial
3
recognition
5
.
4
5
4
1
Supplementary Figure
2:
Variance and factorizability of ratings across 100 traits.
2
a,
Distribution of aggregate
-
level trait ratings.
Each row plots the average ratings across
3
participants for the 100 faces on a trait (grey dots), with the median (line in the box), the first
4
quartile (left edge of the box), the third quartile (right edge of the box), and outliers that are more
5
extreme th
an 3/2 times of the quartiles (open dots).
We supplemented the
94 traits selected from
6
the literature
with additional words for which we believed there was no equivalent in the initial
7
list but would reflect vocabulary used to describe first impressions, i
ncluding words that describe
8
5
sexual orientation (homosexual), traits associated with a neurodevelopmental disorder (autistic),
1
perceived demographics (high
-
income, well
-
educated), and words that are considered derogatory
2
terms for an individual’s intellect
ual or social ability (idiot, loser)
.
b,
Factorability of aggregate
-
3
level trait ratings.
Each row plots the mean (dot), median (triangle), and maximum (square)
4
absolute correlations a trait has with all the other 99 traits (across faces, averaged over
5
part
icipants). The vertical dashed line indicates
r
= 0.30, which describes an inflection point in
6
the curve of mean absolute correlations. The eight traits at the bottom (in bold) were excluded
7
from EFA because of their low average correlations with all other
traits (i.e. low factorizability)
;
8
including these eight traits did not change the dimensions we eventually found.
Source data are
9
provided as a Source Data file.
10
6
1
Supplementary Figure
3: Scree plots of data across eight samples.
2
7
a,
Study 1 sample.
b
-
h,
Study 2 samples from
North America
(b)
, Latvia
(c)
, Peru
(d)
, the
1
Philippines
(e)
, India
(f)
, Kenya
(g)
, and
Gaza
(h)
. Circles plot the eigenvalues (the fraction of
2
total common variance in the data as explained by each factor) of the original data
across factors,
3
ordered from the largest to the smallest. Triangles plot the 95th percentile of the eigenvalues of
4
the simulated data from parallel analysis. The optimal number of factors to retain as
5
recommended by each of the five methods is shown. Paral
lel analysis retains factors with
6
eigenvalues (circles) greater than those from the simulated data (triangles) from 5,000 Monte
7
Carlo simulations (see the close
-
up image for a clearer comparison). Cattell’s scree test retains
8
factors to the left of the poi
nt from which the plotted ordered eigenvalues could be approximated
9
with a straight line (i.e., “above the elbow”). The optimal coordinates index provides a non
-
10
graphical solution to Cattell’s scree test based on linear extrapolation. Empirical Bayesian
11
in
formation criterion (eBIC) retains factors that minimize the overall discrepancy between the
12
population’s and the model’s predicted covariance matrices while penalizing model complexity
13
(purple dots in inset graphs). Velicer’s minimum average partial (MAP)
test is “most appropriate
14
when component analysis is employed as an alternative to, or a first
-
stage solution for, factor
15
analysis”
8
. It is also included in our present study due
to its popularity. MAP retains components
16
by partialing those that resulted in the lowest average squared partial correlation. The MAP test
17
gave variable numbers of components greater than 4; it is not plotted but the results are
18
numerically provided in th
e legend inset.
Source data are provided as a Source Data file.
19
20
21
8
1
9
1
10
Supplementary Figure
4: Four dimensions from EFA in Study 1.
1
a,
Factor loadings of trait ratings on the four dimensions from EFA. Each column plots the
2
strength of the factor loadings (
x
-
axis, absolute value) across traits (y
-
axis). Color indicates the
3
sign of the loading (red for positive and blue for negative); more saturated colors for higher
4
absolute values.
b,
Distributions of the 92 traits along each pair of dimensions based on their
5
factor loadings on the four dimensions.
Source data are provided as a Source Data file.
6
7
11
a
1
2
3
4
5
12
b
1
2
Supplementary Figure
5:
Comparison with existing dimensional frameworks
.
3
13
a,
Four dimensions from PCA in Study 1. Columns plot the strength of the loadings (x
-
axis,
1
absolute value) on the first four varimax rotated principal components across all 92 traits (y
-
2
axis). Colors indicate the sign of the loading (red for positive and blue
for negative); more
3
saturated colors for higher absolute values. The first four principal components without rotation
4
accounted for 52%, 21%, 7%, and 5% of the variance in our data, 86% in total; the fifth
5
accounted for 2%.
b,
Predicting trait judgments u
sing different dimensional frameworks.
6
Regressors were linear combinations of the traits that showed highest loadings in each
7
dimensional framework (two traits for each dimension because there were only two traits that
8
loaded highest on one of the 3D
-
frame
work’s dimensions
9
; for example, for our 4D
-
framework,
9
the model consisted of eight regressors). Each row indicates three different models that regressed
10
the ratings of a targeted trait (row name; which was not one of the regressors) on the three
11
different sets of regressor
s from the three frameworks, and plots the adjusted
R
-
squared
.
Source
12
data are provided as a Source Data file.
13
14
14
a
1
2
3
15
b
1
2
16
c
1
2
17
d
1
2
18
e
1
2
19
f
1
2
20
g
1
2
21
Supplementary Figure 6
: Four factors extracted from aggregated data in Study 2.
1
The seven panels plot results for samples from
a,
North America,
b,
Latvia,
c,
Peru,
d,
the
2
Philippines,
e
, India,
f,
Kenya, and
g,
Gaza. Each column plots the strength of the factor loadings
3
acros
s the 80 traits
(20 of the 100 traits were excluded in the present study for low correlations
4
with other traits [see Supplementary Figure 2], ambiguity or similarity in meaning [trustful,
5
natural, passive, reasonable, strict, enthusiastic, affectionate, an
d sincere], and potential
6
offensiveness in some cultures
[
idiot, loser, criminal, and abusive
]
)
. The color of the bar
7
indicates the sign of the loading (red: positive; blue: negative); the length and saturation of the
8
bar indicates the magnitude of the loa
ding.
Source data are provided as a Source Data file.
9
10
22
a1
1
2
3
4
23
a2
1
2
3
4
24
a3
1
2
3
4
25
1