How trait impressions of faces shape subsequent mental state inferences

Nature Human Behaviour

| Volume 9 | January

2025 | 208–226

208

nature human behaviour

Registered Report

https://doi.org/10.1038/s41562-024-02059-4

How trait impressions of faces shape

subsequent mental state inferences

Chujun Lin

, Umit Keles

, Mark A. Thornton

3,4

& Ralph Adolphs

2,4

People form impressions of one another in a split second from faces.

However, people also infer others’ momentary mental states on the basis

of context—for example, one might infer that somebody feels encouraged

from the fact that they are receiving constructive feedback. How do trait

judgements of faces influence these context-based mental state inferences?

In this Registered Report, we asked participants to infer the mental states of

unfamiliar people, identified by their neutral faces, under specific contexts.

To increase generalizability, we representatively sampled all stimuli from

inclusive sets using computational methods. We tested four hypotheses:

that trait impressions of faces (1) are correlated with subsequent mental

state inferences in a range of contexts, (2) alter the dimensional space that

underlies mental state inferences, (3) are associated with specific mental

state dimensions in this space and (4) causally influence mental state

inferences. We found evidence in support of all hypotheses.

Humans spontaneously form impressions of other people’s general

characteristics upon seeing their faces

. For instance, we judge

whether people are trustworthy, intelligent or feminine on the basis

of how their faces look

–

. These trait judgements show high consensus

across perceivers in different age groups and from different cultures

–

The accuracy of trait judgements from faces is debated: although some

suggest that personality can be accurately judged from static face

images

–

, many argue that trait judgements from faces merely reflect

perceivers’ biases and stereotypes

–

. However valid or invalid they

may be, these judgements shape consequential decisions in the real

world

–

. Such influences are most prominent in situations where

understanding a person’s general traits plays an important role, such

as evaluating which candidate might be a good political leader

which individual on a dating site might be the best long-term partner

Understanding other people’s enduring dispositions is only one

contributing factor that guides social judgements and decisions. More

often, to successfully navigate the complex social world, it is critical

also to understand a person’s context-dependent mental states in the

moment

: what is the other person currently thinking about, feeling

or intending? For instance, our ability to tell whether a friend is joking

or being serious would make all the difference in selecting appropri-

ate behaviour towards them in that particular situation. As with trait

inferences from faces, people also make inferences about others’ men

tal states rapidly and automatically

–

. This ability develops early on,

with evidence suggesting that infants are able to infer goals and inten

tions from six months of age

. Inferences of momentary mental states

are based on a range of cues, such as facial expressions, body postures

and gestures, together with situational information

–

Little is known about how judgements of relatively stable traits

from faces might bias or influence judgements of more transient men

tal states. Studies on the recognition of facial expressions show that

people perceive faces that are digitally manipulated to look untrust

worthy as displaying more negative emotions such as anger

. This

finding suggests that trait judgements from faces may shape emotion

judgements of isolated faces. However, whether this effect generalizes

to judging a broader set of mental states (beyond basic emotions) in

more realistic settings (for example, with situational information) is

unclear. Studies investigating the relation between a wider range of

traits and mental states in more naturalistic settings show that trait

knowledge does shape mental state inferences

. However, those

studies focus on more reliable trait knowledge, such as trait inferences

of participants’ friends and family members, and famous people about

whom participants already have substantial biographical and contex

tual information. It remains unknown whether people rely on trait

Received: 11 August 2020

Accepted: 10 October 2024

Published online: 2 December 2024

Check for updates

Department of Psychology, University of California, San Diego, San Diego, CA, USA.

Division of the Humanities and Social Sciences, California Institute

of Technology, Pasadena, CA, USA.

Department of Psychological and Brain Sciences, Dartmouth College, Hanover, NH, USA.

These authors jointly

supervised this work: Mark A. Thornton, Ralph Adolphs.

e-mail:

chujunlin@ucsd.edu

Nature Human Behaviour

| Volume 9 | January

2025 | 208–226

209

Registered Report

https://doi.org/10.1038/s41562-024-02059-4

the location of an object in space. When more information about the

object is available, we may be able to use fewer, the same, or more and/

or different dimensions to locate the object. For instance, if the new

information identifies the object’s distance to Earth, then we need only

two dimensions; if the new information identifies a moving object, then

we need a fourth dimension (time).

Third, we asked whether (and to what extent) the mental state

dimensions are associated with trait impressions from faces (Table

Q3). For example, recent work has shown that how frequently people

judged the targets to experience positive or negative mental states in

various situations was most closely related to the trait dimension of

warmth

. However, that study investigated trait judgements based on

more information than just faces. It remains an open question how trait

judgements merely from briefly seeing an unfamiliar face might be asso

ciated with mental state dimensions. Understanding the association

between trait impressions and core mental state dimensions (beyond

individual mental state judgements) would allow for further general

izability of our findings. Specifically, it would allow us to predict how

inferences of any mental state—beyond the 60 mental states for which

we collected data—would be associated with trait impressions of faces.

Finally, we asked whether the associations between trait impres-

sions from faces and mental state inferences might be causal (Table

Q4). Prior research has shown that changing the trait impressions of

emotional faces shifted people’s perception of emotions

. Further

more, changing the trait impressions of target people (beyond faces)

has a causal effect on people’s mental state inferences of those targets

across a range of situations

. These results suggest a causal link from

trait to mental state inferences. However, it remains unclear whether

changing the trait impressions formed merely on the basis of faces

would be sufficient to cause people to change their mental state attribu

tions. We tested this hypothesis by digitally manipulating faces so that

the same individual generated multiple stimulus images that exhibited

different traits. We then measured how participants attributed dif

ferent mental states to the same individual in the same context as a

function of the experimental manipulation of that individual’s face.

Understanding the causal effects (Q4) beyond correlations (Q1–Q3) is

essential to determining the nature of the relation between trait impres

sions from faces and mental state inferences. Findings of correlations

without causation would suggest that the observed correlations were

driven by third variables (for example, inferred social roles that shape

both inferences of traits and mental states).

Protocol registration

The Stage 1 protocol for this Registered Report was accepted in princi

ple on 30 March 2022. The protocol, as accepted by the journal, can be

found at

https://doi.org/10.6084/m9.figshare.19664316.v1

Results

As mentioned in Methods (‘Deviations from protocol’), one devia

tion from the approved registered Stage 1 protocol occurred in the

participant recruitment sources for the cross-world-region data. In

the approved registered Stage 1 protocol, we had planned to recruit

participants in all five world regions (the USA, Africa, Asia, Europe and

South America) using the MTurk Toolkit via CloudResearch. We suc

cessfully collected all data from the USA (

= 5,260), Asia (

= 961) and

Europe (

= 1,153) as planned. Due to the limited number of participants

in Africa and South America on MTurk Toolkit, we collected the data in

these two regions using an additional recruitment option offered on

CloudResearch, Prime Panels, together obtaining the planned amount

of data in Africa (

= 1,040) and South America (

= 1,171). We obtained

permission from the Editors before carrying out the above data collec

tion. Except for the deviation mentioned above, we adhered precisely

to the approved registered experimental procedures, data analysis

procedures and result interpretation procedures as detailed in Table

and Methods.

information to the same extent when it is drawn from solely superficial

facial judgements.

In the present investigation, we ask: are the diverse mental states

that people attribute to different individuals in specific situations biased

by the trait judgements of those individuals’ neutral faces? Answering

this question would advance our understanding of how people make

sense of others’ momentary and enduring features, two types of informa

tion critical for social navigation

. If the answer is yes, then the biases

and stereotypes in trait impressions from faces

would probably be

carried over to shape inferences of various mental states in a wide range

of situations. This may help explain, for instance, why Black males whose

faces are stereotypically perceived to be aggressive

are wrongly

attributed the mental state of intending to harm more often in various

situations even without any evidence

. If trait impressions of faces do

influence mental state inferences, then this would also suggest that the

impacts of spontaneous trait judgements of faces are much broader than

previously thought

. They not only would influence decisions in situ-

ations where temporally stable trait information is patently important

but also could influence moment-to-moment decisions we make in the

course of social interactions. Such broader influence could also help

explain why trait judgements of faces are sometimes accurate

–

. For

instance, individuals whose faces look like they are introverted may

be attributed the mental state of being unwilling to interact with other

people, leading others to reduce interaction with these individuals and

in turn exacerbating the social isolation of these individuals.

Our present research tested four main hypotheses. First, when a

photograph of a person’s neutral face is available, we asked whether

inferences of this individual’s mental states in specific situations are

associated with the trait impressions that are based on the neutral face

(Table

, Q1). As mentioned above, prior research shows that emotion

recognition from faces (for example, anger, fear or happiness) is associ

ated with trait judgements of those faces (for example, sociable, domi-

nant or trustworthy)

. However, in real life, we do not judge others’

mental states from their faces in isolation, as participants in most prior

studies have done. Instead, we make sense of people’s mental states

in specific situations. To test this hypothesis, we asked participants

to infer how much different specific people, whose neutral faces the

participants saw, would feel a certain mental state in a given situation

(scenario-state task; Methods). We linked these mental state inferences

to the first impressions formed from those specific people’s neutral

faces on a range of traits (face-trait task; Methods).

To increase the comprehensiveness and generalizability of our

study, we representatively sampled mental state terms using deep

neural networks and the maximum variation procedure from a com

prehensive list of putative mental states (Fig.

1a–d

). We verified that

our final selected set of 60 mental states were representative of the

terms laypeople use in everyday life to describe others’ internal,

non-pathological, specific mental states (Fig.

2a–c

). For each mental

state, we selected a situation that people thought co-occurred with

the mental state in real life (Methods). We representatively sampled

trait terms (Fig.

1e–h

) that people spontaneously use to describe faces

(Fig.

2d–f

), and faces that were diverse with respect to gender, race and

age (Fig.

1i–l

) and that populate the facial geometry that people see in

everyday life (Fig.

2g–i

Second, we hypothesized that the underlying psychological

dimensions people use to represent others’ mental states differ when

face images are available compared with when they are not (Table

, Q2).

Prior research has shown that people use three dimensions to represent

mental states (the 3-D Mind Model dimensions: valence, rationality

and social impact)

. Those studies did not include face information

when participants made mental state inferences. However, in real life,

we can usually see the individuals to whom we attribute mental states.

It remains an open question how adding the information from faces

might modify the psychological dimensions that characterize mental

state inferences. By analogy, we need three dimensions to represent

Nature Human Behaviour

| Volume 9 | January

2025 | 208–226

210

Registered Report

https://doi.org/10.1038/s41562-024-02059-4

Table 1 | Design table

Question

Hypothesis

Sampling plan

Analysis plan

Interpretation given to different

outcomes

Q1. Are trait

impressions

from neutral

faces associated

with mental

state inferences

about those

same people in

given scenarios?

H1a. We predicted that

mental state inferences

of unfamiliar others

in given scenarios

(scenario-state) would

be associated with

the trait impressions

formed from those

individuals’ neutral

faces when shown in

isolation (face-trait).

H1b. We predicted that

H1a would hold even

when we controlled

for the mental states

that the neutral faces

displayed (face-state).

Please refer to ‘Sampling plan’ in

Methods for the details.

H1a. We determined the sample

size for the scenario-state task

to be

= 50 participants per

mental state. This sample size

was estimated empirically on

the basis of our pilot data via the

jackknife resampling procedure.

We determined the sample

size for the face-trait task to

= 38 participants per trait.

This sample size was estimated

empirically on the basis of

sequential resampling by prior

research

H1b. We determined the sample

size for the face-state task to be

= 31 participants per mental

state. This sample size was

estimated empirically on the

basis of sequential resampling

by prior research

Please refer to ‘Analysis plan’ in Methods

for the details.

H1a. We tested this hypothesis using ridge

regression with cross-validations for each

of the 60 mental states.

Each model regressed the average ratings

given to the faces for a mental state under

the specific scenario (scenario-state) on

the ratings given to the faces for 13 traits

(face-traits).

Multiple comparisons across the 60

mental states were corrected for via

maximal statistic permutation tests.

H1b. We tested this hypothesis using

variance partition analyses.

We assessed whether the unique variance

in the scenario-state ratings explained

by face-traits, when controlling for

face-states, is significantly greater than

zero across cross-validation resampling.

H1a. If >80% of the scenario-states

are significantly predicted by

face-traits, the evidence for H1a is

very broad; if 60–80% are predicted,

the evidence is broad; if 40–60%

are predicted, the evidence is

moderately broad; if 20–40% are

predicted, the evidence is narrow;

and if <20% (but at least one mental

state) are predicted, the evidence is

very narrow. Otherwise, there is no

evidence for H1a.

H1b. If the unique variance

explained by face-traits is

greater than zero in >80% of the

scenario-states, the evidence for

H1b is very broad; if greater than

zero in 60–80%, the evidence

is broad; if greater than zero

in 40–60%, the evidence is

moderately broad; if greater than

zero in 20–40%, the evidence is

narrow; and if greater than zero

in <20% (but at least one mental

state), the evidence is very narrow.

Otherwise, there is no evidence for

H1b.

Q2. What are the

dimensions that

underlie mental

state inferences

of others when

we also see their

faces?

H2a. We predicted that

mental state inferences

(scenario-states)

could be represented

by a small number

of dimensions (<10)

even when faces were

available.

H2b. We predicted

that the mental state

dimensions in H2a

would at least partially

overlap with previously

found mental state

dimensions when no

face was available (the

3-D Mind dimensions:

rationality, social

impact and valence)

We tested H2a and H2b using

the same set of data collected

for testing H1a.

Please refer to ‘Analysis plan’ in Methods

for the details.

H2a. Since no single method is regarded

as the best method for determining the

optimal number of dimensions, we applied

five distinct methods: Horn’s parallel

analysis, the optimal coordinate index, the

empirical Bayesian information criterion,

Velicer’s minimum average partial test and

bi-cross-validation.

The optimal number of dimensions was

the number that most methods agreed

on; or, if all methods disagreed, it was the

minimum number that generated the most

interpretable dimensions that accounted

for ≥75% variance in the data in exploratory

factor analysis.

H2b. We measured the Spearman

correlation between the dimensions in

our data and the 3-D Mind dimensions

using two different methods: one based on

factor loadings and scores, and the other

based on participants’ ratings of meaning

similarity.

We deemed an absolute correlation of 0.2–

0.39, 0.4–0.59 or ≥0.6 to be an indication

of weak, moderate or strong similarity.

H2a. If the optimal number of

dimensions was <10, we concluded

that mental state inferences are

represented by a small number of

dimensions even when faces are

available.

Otherwise, we concluded that

mental state inferences are no

longer represented by a small

number of dimensions when faces

are available.

H2b. If any dimension in our data

was at least moderately correlated

(

≥ 0.4) with any 3-D Mind dimension

on the basis of both methods,

we concluded that mental state

dimensions when faces are available

partly overlap with previously found

mental state dimensions when no

face was available.

Otherwise, we concluded that there

is no strong evidence for H2b.

Q3. Are the

dimensions that

underlie mental

state inferences

of others

when faces

are available

associated with

trait impressions

formed from

those faces?

H3a. We predicted

that the dimensions

of mental state

inferences when

faces were available

(scenario-state

dimensions) would be

associated with trait

impressions formed

from those faces

(face-trait).

H3b. We predicted that

H3a would hold even

when we controlled for

the mental states that

those neutral faces

displayed (face-state).

We tested H3a and H3b using

the same set of data collected

for testing H1a and H1b.

Please refer to ‘Analysis plan’ in Methods

for the details.

H3a. We tested this hypothesis using ridge

regression with cross-validations as in H1a.

The only difference is that each model

here corresponded to each core mental

state dimension in Q2.

Each model regressed the factor scores

for a mental state dimension across the

faces (factor scores of scenario-states) on

the ratings given to the faces for 13 traits

(face-traits).

H3b. We tested this hypothesis using

variance partition analyses as in H1b. The

only difference is that the dependent

variable here is the factor scores for a

mental state dimension across the faces.

H3a. If any mental state dimension

was significantly predicted by

face-traits, we concluded that

the dimension(s) of mental state

inferences when faces are available

is (are) associated with trait

impressions from those faces.

Otherwise, we concluded that there

is no evidence for H3a.

H3b. If for any mental state

dimension, the unique explained

variance of face-traits was

significantly greater than zero, we

concluded that there is evidence

for H3b.

Otherwise, we concluded that there

is no evidence for H3b.

Nature Human Behaviour

| Volume 9 | January

2025 | 208–226

211

Registered Report

https://doi.org/10.1038/s41562-024-02059-4

Associations between mental state and trait inferences

We found very broad evidence that mental state inferences of unfamiliar

others in given scenarios were associated with trait impressions formed

from those individuals’ faces (Table

, H1a). Inferences of every one of

the 60 mental states were significantly predicted by inferences of the

13 traits: ridge regression analysis with cross-validations (for increasing

generalizability; Methods) showed that the prediction accuracy for the 60

mental state inferences ranged from

= 0.64 to

= 0.97, with mean

= 0.88

(multiple comparisons across 60 mental states were corrected for using

maximal statistic permutations; corrected

values ranged from 0.0005

to 0.047; see Supplementary Table 2 for the results with detailed statistics

of all 60 mental states). These results suggest that given the same context,

how people infer different individuals’ mental states is associated with

the trait impressions inferred from those individuals’ faces.

The strong associations between context-specific mental state

inferences and trait impressions from faces remained robust even when

controlling for the context-irrelevant affective and cognitive states

inferred from the faces (captured when the photos were taken; Table

H1b). Using variance partition analysis (three models were fitted per

mental state with different predictors; Methods), we found that trait

impressions contributed a significant amount of unique explained

variance for 47 of the 60 mental state inferences (Fig.

). Across those

47 mental states, face-trait impressions on average contributed

= 0.39

explained variance (the lower bound across 2,000 cross-validation

iterations (that is, the 2.5th percentile) ranged from 0.005 to 0.597

across mental states; see Fig.

for detailed statistics) beyond the vari

ance that was commonly explained by both face-trait and face-state

impressions (see Supplementary Table 2 for the results of all 60 mental

states). These results suggest that given the same context, how people

infer different individuals’ mental states is influenced by first impres-

sions from faces that are specifically about the individual’s stable

characteristics beyond momentary states such as emotions (see Sup

plementary Fig. 3 for how each mental state was differently influenced

by different trait impressions; we validated this interpretation of the

ridge regression coefficients with three additional analysis methods

such as ordinary least squares regressions and LASSO regressions;

Supplementary Methods and Supplementary Fig. 4).

Trait inferences of faces do not merely influence an isolated mental

state judgement when that face is seen in a particular context. Trait

inferences also change the psychological space that characterizes the

relationships among mental state judgements (Table

, H2a). We inves

tigated the dimensions underlying these relationships by analysing

which mental state inferences were highly correlated with one another

across different target faces. As preregistered, we first determined

how many dimensions optimally summarized the common variance

in the judgements of the 60 mental states. Three of the five planned

methods indicated that four dimensions optimally represented men-

tal state judgements (agreed by Horn’s parallel analysis, the optimal

coordinate index and the empirical Bayesian information criterion;

2 of the 60 mental states were excluded for low factorability: ‘bored’

and ‘indecisive’).

The interpretation of these four mental state dimensions is most

clearly obtained by examining which types of targets were more (and

less) often attributed the mental states associated with each dimension

(Figs.

and

). We interpreted the four dimensions as describing senti

mental mental states (mental states that are stereotypically associated

with exaggerated or self-indulgent emotions), youthful mental states

(mental states that are stereotypically associated with youthful peo

ple), empathetic mental states (mental states that are stereotypically

associated with people who understand others’ feelings) and compe-

tent mental states (mental states that are stereotypically associated

with competent people). These four mental state dimensions together

explained 83% of the common variance in the context-specific mental

state inferences (each explaining 39%, 19%, 15% and 10%). Since our

dimension analysis method (exploratory factor analysis) does not force

the dimensions to be orthogonal, it reveals the natural relationships

between the dimensions. The sentimental mental state dimension and

the empathetic mental state dimension were moderately correlated

(

= 0.47;

= 5.32;

= 6.57 × 10

−7

; 95% confidence interval (CI), (0.31,

0.61)); correlations between the other mental state dimensions were

weak (all

≤ 0.25).

We compared the four mental state dimensions uncovered when

targets’ faces were available with the three mental state dimensions

(valence, rationality and social impact) from prior theory

(Table

Question

Hypothesis

Sampling plan

Analysis plan

Interpretation given to different

outcomes

Q4. Are mental

state inferences

of others in a

given scenario

causally

influenced

by the trait

impressions

formed

from those

individuals’

neutral faces?

H4. We predicted

that trait impressions

formed from neutral

faces (face-trait)

causally shape mental

state inferences

of those people in

specific scenarios

(scenario-state).

Please refer to ‘Sampling plan’ in

Methods for the details.

H4. We used a set of

= 272

face images to detect the

causal effect. This sample size

was determined via formal

power analysis, with a paired

one-sided

-test. See ‘Stimuli:

Trait-manipulation of face

stimuli’ in Methods.

We checked each trait

manipulation. Each subset

of facial identities with their

manipulated images were

rated by

= 38 participants per

manipulated trait, using a similar

procedure as in the face-trait

task in H1a.

We tested causality using the

trait-manipulated faces via

a similar procedure as in the

scenario-state task. Each subset

of facial identities with their

manipulated images were rated

= 50 participants per mental

state as in H1a.

Please refer to ‘Analysis plan’ in Methods

for the details.

H4. We tested causality for the mental

states that were strongly correlated

with trait impressions in H1 (for example,

the top predicted state(s) from each

dimension; targeting around six states).

For each mental state, we identified a

different, strongly correlated trait. For

each trait, we digitally manipulated each

face to enhance and reduce that trait,

generating two versions of face images.

We tested the causal effect for each

state-trait pair using two methods: one

based on aggregate scenario-state ratings,

using paired one-sided

-tests between

the two versions of faces; and another

based on individual scenario-state ratings,

using linear mixed modelling to regress

the ratings on the face versions while

controlling for the random effects of

participants and face identities.

H4. For each state–trait pair, if we

found a significant effect in the

expected direction (as discovered

in H1) using both methods, we

concluded that there is strong

causal evidence for that state–trait

pair. If only one method indicated

a significant effect, the evidence is

weak. If neither method indicated a

significant effect, there is no causal

evidence for that state–trait pair.

Across all tested state–trait pairs, we

concluded that the evidence for H4

is very broad if >80% pairs showed

strong evidence; broad if 60–80%

pairs showed strong evidence;

moderately broad if 40–60% pairs

showed strong evidence; narrow

if 20–40% pairs showed strong

evidence; and very narrow if <20%

but at least one pair showed strong

evidence. Otherwise, there is no

strong evidence for H4.

Table 1 (continued) | Design table