Abstract Thinking Facilitates Aggregation of Information

Abstract

Many situations in life (such as considering which stock to invest in, or which people to befriend) require averaging across series of values. Here, we examined predictions derived from construal level theory, and tested whether abstract compared with concrete thinking facilitates the process of aggregating values into a unified summary representation. In four experiments, participants were induced to think more abstractly (vs. concretely) and performed different variations of an averaging task with numerical values (Experiments 1-2 and 4), and emotional faces (Experiment 3). We found that the induction of abstract, compared with concrete thinking, improved aggregation accuracy (Experiments 1-3), but did not improve memory for specific items (Experiment 4). In particular, in concrete thinking, averaging was characterized by increased regression toward the mean and lower signal-to-noise ratio, compared with abstract thinking. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Full text

Journal of Experimental Psychology: General
Abstract Thinking Facilitates Aggregation of Information
Britt Hadar, Moshe Glickman, Yaacov Trope, Nira Liberman, and Marius Usher
Online First Publication, December 20, 2021. http://dx.doi.org/10.1037/xge0001126
CITATION
Hadar, B., Glickman, M., Trope, Y., Liberman, N., & Usher, M. (2021, December 20). Abstract Thinking Facilitates
Aggregation of Information. Journal of Experimental Psychology: General. Advance online publication.
http://dx.doi.org/10.1037/xge0001126

BRIEF REPORT
Abstract Thinking Facilitates Aggregation of Information
Britt Hadar
1, 2, 3
, Moshe Glickman
4, 5
, Yaacov Trope
6
, Nira Liberman
1
, and Marius Usher
1, 7
1
School of Psychological Sciences, Tel-Aviv University
2
Department of Psychology, Princeton University
3
Princeton School of Public and International Affairs, Princeton University
4
Department of Experimental Psychology, University College London
5
Max Planck University College London Centre for Computational Psychiatry and Ageing Research, London, United Kingdom
6
Department of Psychology, New York University
7
Sagol School of Neuroscience, Tel-Aviv University
Many situations in life (such as considering which stock to invest in, or which people to befriend)
require averaging across series of values. Here, we examined predictions derived from construal level
theory, and tested whether abstract compared with concrete thinking facilitates the process of aggregat-
ing values into a unified summary representation. In four experiments, participants were induced to
think more abstractly (vs. concretely) and performed different variations of an averaging task with nu-
merical values (Experiments 1–2 and 4), and emotional faces (Experiment 3). We found that the induc-
tion of abstract, compared with concrete thinking, improved aggregation accuracy (Experiments 1–3),
but did not improve memory for specific items (Experiment 4). In particular, in concrete thinking, aver-
aging was characterized by increased regression toward the mean and lower signal-to-noise ratio, com-
pared with abstract thinking.
Keywords: construal level theory, abstraction, numerical averaging, emotion perception
Supplemental materials: https://doi.org/10.1037/xge0001126.supp
Imagine yourself in a new workplace, trying to figure out how
friendly are your new coworkers. To form impressions about their
dispositions, you sample their behavior (e.g., whether they greeted
you in the elevator or invited you to lunch), and average those sam-
ples into a general value-representation (Anderson, 1981). Similarly,
when watching products displayed in a shop (YamanashiLeibetal.,
2020), or a rapid stream of stock returns (Betsch et al., 2001;Vanunu
et al., 2019), one might extract an overall value estimation to guide
behavior. In all these cases, different exemplars that span across time
and space are collapsed into one representative value (e.g., “I like a
lot this coworker, this shop is expensive, or the stock is profitable”).
Averaging is a common and intuitive way of extracting a representa-
tive value from a stream of exemplars (Anderson, 1971;Kahneman,
2011). Indeed, the rapid extraction of ensemble information has been
documented in a wide array of ensemble properties ranging from
simple features such as size (Ariely, 2001;Chong et al., 2008;Chong
& Treisman, 2003;2005), orientation (Dakin & Watt, 1997;Parkes
et al., 2001), color (Maule et al., 2014), motion direction (Watama-
niuk & McKee, 1998;Watamaniuk et al., 1989),andmotionspeed
(Watamaniuk & Duchon, 1992); to more complex properties, such as
gender (Haberman & Whitney, 2007), identity (de Fockert & Wolf-
enstein, 2009;Neumann et al., 2013), emotional facial expressions
(Haberman et al., 2009), animacy (Yamanashi Leib et al., 2016),
gaze direction (Florey et al., 2016;Sweeny & Whitney, 2014), attrac-
tiveness (Post et al., 2012;van Osch et al., 2015;Walker & Vul,
2014), traits (Asch, 1946;Eyal et al., 2011;Hamilton & Sherman,
1996), and even when deciding whether a basketball player’scareer
earns him a place in the Hall of Fame (Brusovansky et al., 2019). For
Britt Hadar https://orcid.org/0000-0003-0863-8650
Britt Hadar and Moshe Glickman shared equal contribution. Nira
Liberman and Marius Usher joined senior authorship.
Data is available at: https://osf.io/y4uxd/.
This research was supported by a United States-Israel Binational
Science Foundation (BSF) Grant 2016090 to Nira Liberman and Yaacov
Trope. Nira Liberman and Britt Hadar were supported by Grant 524/17
by the Israel Science Foundation (ISF). Marius Usher was supported by
an ISF Grant 1413/17. The authors would like to thank Talya Rapp and
Yasmeen Shamshoum for their help in running the experiments.
Correspondence concerning this article should be addressed to Britt
Hadar, Department of Psychology, Princeton University, South Drive,
Princeton, NJ 08540, United States, or Moshe Glickman, Department of
Experimental Psychology, University College London, 26 Bedford Way,
London, WC1H 0AP, United Kingdom. Email: britt.hadar@gmail.com or
mosheglickman345@gmail.com
1
Journal of Experimental Psychology: General
©2021 American Psychological Association
ISSN: 0096-3445 https://doi.org/10.1037/xge0001126
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

an extensive review of ensemble perception, the reader is referred to
Whitney and Yamanashi Leib (2018).
Recent research has supported the idea that average extraction is
automatic, or at least nonintentional (Betsch et al., 2001;Brusovan-
sky et al., 2018;Khayat & Hochstein, 2018), and is based on a popu-
lation-coding mechanism (Baek & Chong, 2020;Brezis et al., 2016;
Brezis et al., 2018). Several factors, including set size (Brezis et al.,
2015;Robitaille & Harris, 2011), and variance (Brezis et al., 2018;
Solomon, 2010) were found to affect the accuracy of averaging.
However, little is known about the impact of thinking-modes on the
accuracy of intuitive (i.e., without explicit symbolic calculation) aver-
aging. We specifically focus on concrete versus abstract thinking.
Research conducted within construal level theory (CLT), shows that
people form representations at varying degrees of abstraction (Gilead
et al., 2020;Liberman & Trope, 2014;Trope & Liberman, 2010). A
process of abstraction, in this view, occurs when distinctions between
exemplars are disregarded, and a dimension of similarity is high-
lighted that places them into a category. For example, categorizing
sandals and boots as footwear disregards “waterproof”and highlights
“worn on feet.”Averaging bears inherent relation to abstraction
because when a stream of exemplars is represented by an average,
differences between specific exemplars are disregarded, and their
commonality (their origin in a common distribution) is emphasized.
We reasoned that because averaging is a process of abstraction
wherein exemplars are integrated (with equal weights) into a single
representative value, then abstract thinking (compared with concrete
thinking) should enhance the accuracy of intuitive average estimation.
We tested this prediction in three experiments (Experiments 1–3) that
manipulated, within participants, abstract versus concrete thinking via
the “why/how”mindset manipulation, which was originally developed
by Freitas et al. (2004), and has been widely used since then. Indeed, a
review by Burgoon et al. (2013) mentions 11 published articles that
used this manipulation, and many more used it since then (Ding &
Keh, 2017;Efrat-Treister et al., 2020;Gilead et al., 2014;Hadar et al.,
2019;Hansen & Trope, 2013;Kille et al., 2017;Napier et al., 2018;
Spunt et al., 2016;Stillman et al., 2017;Yudkin et al., 2020). Experi-
ment 1 examined accuracy in a numerical averaging task (Brezis et al.,
2015;Malmi & Samson, 1983;Rosenbaum et al., 2021), and Experi-
ment 2 was a preregistered replication. Experiment 3 tested the effect
of abstract thinking on averaging of emotional faces (which prevents
the use of rule-based strategies). Experiment 4 was a control experi-
ment that tested the effect of abstract mindset on memory for specific
details. We predicted that abstract (compared with concrete) thinking
would improve accuracy of average extraction (Experiments 1–3), but
not memory for specific details (Experiment 4). We examined two
mechanisms that may explain the effect of abstract thought: signal-to-
noise ratio and temporal weighting biases.
Experiments
Mindset Manipulation
In all four experiments
1
we used the mindset induction manipu-
lation developed by Freitas et al. (2004); see Supplemental
Information (SI) for a detailed description of the manipulation. In
the abstract mindset condition, participants were presented with
four boxes, in which they answered four consecutive “why”ques-
tions, starting from, for example, “Why maintain good physical
health”. In the concrete mindset condition, participants answered
four consecutive “how”questions, starting with the same behavior,
for example, “How to maintain good physical health”.(Figure 1A
presents the manipulation).
Averaging Task
In all three experiments, following each mindset induction, the
participants estimated the average of rapidly presented sequences of
values (Brezis et al., 2015;Malmi & Samson, 1983). The estimation
precision was measured via the Pearson correlation between the
actual and the estimated averages across all trials (for each participant
and mindset manipulation). Relying on correlation allows testing sen-
sitivity to relative changes in the variable, while allowing for general
biases.
2
In the General Discussion section, we also report analyses of
actual deviations (i.e., root mean square deviation [RMSD]).
Experiment 1
Method
Participants
Fifty-six Tel-Aviv University students (35 women, M
age
=
22.80, SD = 4.12) participated for a course credit. One participant
left before finishing the task and was excluded from analysis.
Materials
In the numerical averaging task, each trial began with a central
fixation cross (250 ms), after which a sequence of eight two-digit
numbers (between 10 and 90) was presented at a rate of 2 Hz and
participants were required to enter their estimation of the sequence
average on a visual analog scale ranging from 0–100 (see Figure
1B, and SI for additional details). The numbers were drawn from a
Gaussian distribution (l=50þk, r= 20, k U[–15,15]), with
no successive repetitions. Participants completed 20 practice trials
and 240 experimental trials.
Procedure
Participants first completed the practice trials of the numerical
averaging task, then the mindset manipulation (either the concrete
“how”or the abstract “why”). Then 60 trials of the numerical aver-
aging task, such that two “why”blocks were completed consecu-
tively and then two “how”blocks were completed consecutively by
each participant. Whether a participants started with “why”or
“how”blocks was counterbalanced between participants.
Results
Performance in the abstract and concrete thinking conditions was
quantified by computing the Pearson correlation between partici-
pants’estimations and actual mean of each sequence (see Figure
2A–Bfor an illustration of two representative participants). Pearson
1
Data is available at: https://osf.io/y4uxd/, see Hadar (2021).
2
For example, if the average estimations are accurate but shifted by a
constant, this would reduce the deviation score but preserve a high
correlation between real and the estimated averages. Such a situation
indicates a high discrimination sensitivity of the sequence-average.
2HADAR, GLICKMAN, TROPE, LIBERMAN, AND USHER
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

correlations were transformed into Fisher’s Z-scores (z=arctanh
(r)) and submitted to 2 (mindset: abstract vs. concrete) 32(order:
abstract block first vs. concrete block first) mixed-design ANOVA
with mindset as a within-participants factor. The analysis revealed
the predicted effect of mindset, whereby the Fisher-transformed
correlations were significantly higher in the abstract mindset condi-
tion (M=.911,SD= .221) compared with the concrete mindset
condition (M=.859,SD =.214),F(1, 53) = 5.411, p=.024,h
p
2
=
.093. A main effect of order, F(1, 53) = 8.210, p=.006,h
p
2
=.134,
indicated that correlations were higher for participants who com-
pleted the concrete mindset condition first compared to those who
completed the abstract mindset condition first. The interaction was
not significant, F(1, 53) = .340, p= .562, indicating that the effect
of abstract thinking obtained irrespective of order (see Figure 2C
for the group-level [z-transformed] Pearson correlations collapsed
across order conditions). We confirmed that the difference between
the abstract and concrete conditions remain significant if we use a
nonparametric test to compare the raw correlations (i.e., without
using Fisher z-transformation). To this end, we conducted a two-
tailed paired-samples permutation test with 10,000 shuffles, which
yielded a significant result (p=.023).
Experiment 2
Experiment 2 was a preregistered replication
3
of Experiment 1
with several changes. First, we cut the number of trials by half.
Second, to generalize our finding across different underlying dis-
tributions, the values in Experiment 2 were sampled from a uni-
form distribution (in which all the numbers in the designated range
have the same probability to be sampled). This makes the estima-
tion more challenging by rendering inefficient heuristic strategies
(e.g., basing the average on only two to three random samples.)
Method
Participants
Sixty-two Tel-Aviv University students (37 women, M
age
=
23.69, SD = 3.27) participated in the experiment for course credit.
Two participants were excluded; one was not a native Hebrew
speaker, and one did not complete the manipulation forms.
Materials and Procedure
The procedure was the same as in Experiment 1, except that
participants completed only one block of 60 trials under each
mindset condition and the distribution from which numbers in
each sequence were sampled was uniform (U[45 þk, 55 þk],
kU[15, 15]) rather than Gaussian.
Results
The same analysis as in Experiment 1 revealed the predicted
effect of mindset, whereby Fisher-transformed correlations
were higher in the abstract mindset condition (M=1.091,SD=
.234) compared with the concrete mindset condition, (M=
1.041, SD = .278), F(1, 58) = 4.572, p= .037, h
p
2
= .073. Nei-
ther order F(1, 58) = .964, p= .330, nor the interaction of mind-
set and order were significant, F(1, 58) = .185, p=.669(see
Figure 2C). The same pattern of results was obtained if we ana-
lyzed the data using a paired-sampled permutation test similar
to that used in Experiment 1 (p= .013).
Experiment 3
The aim of Experiment 3 was to generalize the effect of abstrac-
tion on aggregation of information to more ecological and real-life
Figure 1
Event Sequence in Experiments 1 and 2
Note. (A) Mindset induction manipulation: Participants complete the abstract mindset manipulation (“Why?”) or the concrete
mindset manipulation (“How?”). (B) Participants then observe a sequence of eight numbers and estimate their average on an ana-
log scale (60 in each block). In Experiment 1 there were two blocks after a concrete mindset induction, two blocks after an
abstract mindset induction. In Experiment 2 there was only one block after each manipulation.
3
http://aspredicted.org/blind.php?x=bd5245
ABSTRACTION FACILITATES AGGREGATION OF INFORMATION 3
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

stimuli, which also prevents the use of rule-based computations. We
chose extraction of the mean emotion from a set of emotional faces
for several reasons. First, people may face a similar task on a daily
basis when trying to understand their partners’or coworkers’
feelings. Second, finding a common process that applies to both sim-
ple stimuli (numbers in Experiments 1 and 2) as well as to complex
social stimuli (emotional faces) is not trivial and can shed light on a
broader spectrum of instances in which abstraction plays a role.
Figure 2
Results of Experiments 1–3
Note.(A–B) Correlations between the sequence actual mean and the estimated responses of two representative participants
in the abstract (blue) and concrete (orange) thinking conditions. (C) Pearson correlations (z-transformed) by mindset condi-
tion (abstract vs. concrete) in Experiments 1–3. Blue and orange circles correspond to the correlations of each participant in
the abstract and concrete conditions, respectively. Gray circles correspond to the conditions means. The central mark on
each box-plot indicates the median, and the bottom and top edges indicate the 25th and 75th percentiles, respectively. *p,
.05. See the online article for the color version of this figure.
4HADAR, GLICKMAN, TROPE, LIBERMAN, AND USHER
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Method
Participants
Sixty-two Tel-Aviv University students (38 women, M
age
=
22.88, SD = 2.54) participated in the experiment for course credit.
Two participants did not complete the manipulation forms and
were excluded.
Materials
Mindset Manipulation. We used the same mindset manipula-
tion as in Experiments 1 and 2.
Emotion Averaging Task. We adopted the morphed faces
from Haberman et al. (2009). We used 50 different faces that were
created by linearly interpolating between two emotion extremes of
the same actor, taken from the Ekman gallery (Ekman & Friesen,
1976), such that each morphed face was one emotional unit sadder
than the one before it. Thus, the values of the faces ranged
between 1 (happiest) and 50 (saddest; see Figure 3). On each trial,
faces were randomly sampled from a uniform distribution (with
repetition).
Procedure
Participants were told that they will be presented with sequences
of emotional faces (all of the same person, “Rachel”). Each sequence
represents Rachel’s emotions on a given day. Participants were asked
to estimate Rachel’s mean emotion during each day. Participants
completed five practice trials, then completed the mindset induction
task, and then one block of the averaging task (68 trials). Participants
then completed the second mindset induction task (order was
counterbalanced between participants) and performed the second and
last block of the averaging task. Overall, participants completed 136
critical trials.
Results
The same analysis as in Experiments 1 and 2 revealed the pre-
dicted effect of mindset, wherein Fisher-transformed correlations
were higher in the abstract mindset condition (M= 1.008,SD=
.196) compared to the concrete mindset condition (M= .955, SD =
.191), F(1, 58) = 5.514, p= .022, h
p
2
= .087, (Figure 2C). Neither
order F(1, 58) = .607, p= .439, nor the interaction of mindset and
order were significant, F(1, 58) = .914, p= .343. The same pattern
of results was obtained if we analyze the data using a paired-
sampled permutation test (p= .035).
Computational Modeling
We begin by examining whether the participants’estimations
were more influenced by earlier items (i.e., primacy bias) or by
later items (i.e., recency bias). Previous studies have reported the
latter both for numerical values averaging task (Brezis et al., 2015;
Spitzer et al., 2017), and emotional faces (Hubert-Wallander &
Boynton, 2015). However, because different mechanisms may
underlie these two tasks (Haberman et al., 2015;Hubert-Wallander
& Boynton, 2015;Whitney & Yamanashi Leib, 2018), we ana-
lyzed the temporal weighting in each experiment separately. For
each experiment, we performed a temporal regression analysis, in
Figure 3
Illustration of the Emotion Averaging Task in Experiment 3
Note.Afixation cross is presented for 500 ms, followed by a sequence of eight faces presented
for 500 ms each. Participants then estimate the average emotional expression using an analog scale,
presented until response. The authors adopted the morphed faces from Haberman et al. (2009).
ABSTRACTION FACILITATES AGGREGATION OF INFORMATION 5
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

which we predicted the response in each trial for Experiments 1–3
(Y), using the samples (X
i
) ranked by their temporal order.
Y¼woþX
n
i¼1
wiXi(1)
As shown in Figure 4, across Experiments 1–3 the mean normal-
ized weights of Samples 5–8 was higher than that of Samples 1–4
(Experiment 1: t(54) = 7.49, p,.001; Experiment 2: t(59) = 6.97,
p,.001; Experiment 3: t(59) = 9.11, p,.001), indicating a recency
bias. This motivated us to include a leak term (a parameter that con-
trols the extent to which earlier values are given less weight; Usher &
McClelland, 2001;Vanunu et al., 2019)insomeofourmodels.
We examined several models to gain better understanding of the
effects in Experiments 1–3. All models assume that the mean is
estimated based on noisy representations of the magnitudes of the
numbers (Experiments 1 and 2) or facial expressions (Experiment
3) presented in each sequence. Models are based on Equation 1
(with a Gaussian noise term), and approximate the neural popula-
tion-averaging model suggested by Brezis et al. (2015,2018; see
also Rosenbaum et al., 2021 for simulations showing the equiva-
lence between models based on Equation 1 and population-averag-
ing models). Because similar patterns of results were obtained in
Experiments 1–3, we collapsed them together.
In particular, we examined four models based on the mecha-
nism presented in Equation 1 (see Table 1). The first model
serves as a baseline, and includes equal temporal weights (i.e.,
no-leak) and a direct (unbiased) map of estimated values to the
response scale. The second model also assumes equal weights,
but includes two free-parameters: intercept and slope parameter
(b0and b1) which map between the participant’s internal
estimation to the external response scale; for slope parameters
that are smaller than 1, this reflects a regression to the mean.
Models 3–4 are similar to Models 1–2, but introduce leak
(unequal temporal weights). For each participant, the models
were fitted to the data of the abstract and concrete mindset condi-
tions separately, and were compared with each other using the
Akaike information criterion (AIC; Akaike, 1974).
As shown in Table 1, Model 4 decisively outperformed all other
models both in the concrete and abstract mindset conditions, sug-
gesting that the integration process was characterized by decreas-
ing weight to earlier-presented samples (recency bias), as well as
by biases in the mapping between presented numbers and the
response scale. To examine whether these processes varied
between conditions, we compared between conditions the best-fit-
ted leak, intercept and slope parameters of Model 4.
Toward this end, we performed three t-tests with mindset
(abstract vs. concrete) as independent variable, and leak, inter-
cept, and slope as dependent variables.
4
The first ttest revealed a
significant difference, t(174) = 3.48, p,.001, between the slope
of the abstract mindset condition (b1= .98) and the concrete
mindset condition (b1= .93), indicating a more accurate, less re-
gressive estimation mapping in the abstract condition. Similar
results were obtained for the intercept parameter, t(174) = –3.39,
p,.001, revealing that the intercept in the abstract mindset con-
dition (b0= .54) was less biased than the intercept in the con-
crete mindset condition (b0= 2.69). Finally, while the leak was
numerically lower in the concrete compared with the abstract condi-
tion, this difference was not statistically significant, t(174) = –1.59,
Figure 4
The Temporal-Weighting Profile (Standardized Regression Weights) of the Samples in Experiments
1–3
Note. In all experiments the evaluations of the participants were more influenced by the recent items in the
sequence (i.e., recency bias). Error bars correspond to within-subjects SE. See the online article for the color
version of this figure.
4
For simplicity, the data in the analyses presented here were collapsed
across Experiments 1–3, see SI for a full two-way ANOVAs using mindset
and experiment as independent factors.
6HADAR, GLICKMAN, TROPE, LIBERMAN, AND USHER
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

p= .12. For each condition we also computed a measure of signal-
to-noise ratio, by dividing the slope by the squared root of the
error-term ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn
i¼1
ypredictedyobserved
ðÞ
2
n1
s
0
@1
A, which provides an estima-
tion of the scatter around the regression line (Figure 2A–B). We
find that this measure of signal-to-noise ratio was higher in the
abstract mindset condition than in the concrete mindset condition, t
(174) = 3.83, p,.001.
Experiment 4
In three experiments we have shown that abstract mindset (com-
pared with concrete mindset) improves the accuracy by which infor-
mation is aggregated in average estimations. A remaining question,
however, is whether this improvement is unique to averaging, or is
rather a mere reflection of a general performance improvement.
Experiment 4 is a preregistered
5
examination of this question. We
modified the task from average-estimation to memory for individual
items. We hypothesized that abstract mindset would not improve per-
formance in this task, and might even impair performance, compared
with concrete mindset (which, compared with the abstract mindset,
may confer an advantage to the processing of individual elements).
Method
Participants
One-hundred and 57 participants were recruited from Prolific
6
(79 women; M
age
= 23.60, SD = 5.31), in exchange for $4. Partici-
pants were prescreened to those who currently hold a student sta-
tus (in order to keep the sample relatively comparable to the
previous lab-based experiments). Six participants were excluded,
five due to below-chance performance (i.e., accuracy rate ,.5),
and one for not completing the manipulation forms.
Materials and Procedure
Procedure generally followed that of Experiment 1 with few
changes. First, instead of asking about the mean value of the
sequence participants were presented after each sequence of num-
bers with two probe-numbers, and were asked to choose which one
of them has been presented in the sequence. One of these numbers
was randomly sampled from the eight numbers that were presented
in that sequence, and the other number was randomly sampled from
the same Gaussian distribution, but we made sure it would not be
any of the numbers in the sequence. Participants completed 10 prac-
tice trials, and 80 critical trials in each mindset condition (i.e., a
total of 160 critical trials), see Figure 5 for an illustration.
Second, following each critical block, participants reported their
current motivation (“I was motivated to succeed in the numerical
task”), stress (“I feel stressed at the moment”), and mood (“I have
a positive mood at the moment”), on a 5-point scale ranging from
1(strongly disagree)to5(strongly agree).
Results
Manipulation Check
In order to validate that the manipulation elicited the intended
mindset in each condition, we used an automated text analysis tool
(TATE; Simchon, 2019), which rates the concreteness level of words
accordingtoanormeddictionary(Brysbaert et al., 2014). Higher
scores indicate higher concreteness level. Concreteness scores were
submitted to a repeated measures ANOVA, and revealed that as
expected, texts’concreteness level in the concrete mindset condition
was higher (M= 2.943, SD = .321) than in the abstract mindset con-
dition (M=2.483,SD = .312), F(1, 149) = 217.268, p,.001, h
p
2
=
.593. This analysis confirms that the experimental manipulation
made participants use abstract versus concrete language, in the
abstract versus concrete conditions as intended.
Accuracy
Memory-accuracy was submitted to the same analyses as in
Experiments 1–3. In contrast to our previous experiments (and in
contradiction to a general improvement account of our previous
results), there was no advantage for the abstract condition relative
to the concrete condition. Rather, the analysis revealed a marginal
effect of mindset favoring the concrete mindset condition.
Table 1
Model Comparison
Model number Model Abstract mindset (aggregated AIC) Concrete mindset (aggregated AIC)
1
yi¼Xn
i¼1
Xi
nþei
97,902 98,942
2
yi¼b0þb1Xn
i¼1
Xi
nþei
94,458 94,920
3
yi¼Xn
i¼1
1k
ðÞ
niXi
Xn
i¼1
1k
ðÞ
ni
þei
97,527 98,545
4
yi¼b0þb1Xn
i¼1
1k
ðÞ
niXi
Xn
i¼1
1k
ðÞ
ni
þei
93,927 94,394
Note. Note that AIC differences higher than 10 are considered decisive evidence in favor of the model with the lower numerical value (bold values indi-
cate the best fits).
5
https://aspredicted.org/blind2.php
6
https://prolific.co/
ABSTRACTION FACILITATES AGGREGATION OF INFORMATION 7
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Accuracy in the abstract mindset condition was slightly lower
(M= .709,SD= .100) than in the concrete mindset condition (M=
.721,SD= .093), F(1, 149) = 3.78, p= .054, h
p
2
= .025, see Figure
5The effect of order was not significant, F(1, 149) = .414, p=
.521, but the interaction of mindset and order was significant, F(1,
149) = 9.194, p= .003, h
p
2
= .058. In addition, motivation (p= .639),
stress (p= .425), and mood (p= .189) did not significantly differ
between the two mindset conditions.
Taken together, these results indicate that the improved per-
formance in the abstract mindset condition in Experiments 1–3
is unlikely to reflect a general effect on any type of performance
(e.g., due to higher motivation, better attention, or lesser fa-
tigue). Rather, as we discuss in the General Discussion, this pat-
tern of findings supports the notion that abstract versus concrete
processing differentially favor gist extraction versus attention to
details.
General Discussion
In three experiments, we examined how abstract versus con-
crete thinking affected aggregation of information, and in the
fourth experiment we examined how it affects memory for spe-
cific items. As predicted, we found that abstract thinking
improved aggregation accuracy compared with concrete think-
ing, whereas it did not improve accuracy in the memory task. In
Experiments 1 and 2, participants performed a numerical averag-
ing task adopted from Brezis et al. (2015), whereas in Experi-
ment 3 they performed a variation of the averaging task with
emotional faces (Haberman et al., 2009). The results were tested
using Pearson correlations as a measure of estimation precision,
but similar results obtained with deviations between correct and
estimated values (RMSD).
7
Finally, in Experiment 4 we ruled
out a general (e.g., motivational or attentional) account for the
improved performance in Experiments 1–3. When the task was
changed to remembering individual items, the abstract mindset
condition showed no advantage but rather a small decrement rel-
ative to the concrete mindset condition.
To better understand the processes that drive the improved aver-
age estimations in the abstract mindset, we examined several com-
putational models, assuming that the estimation of the mean is
based on transformation of numerical values (Dehaene, 1992;
Dehaene & Cohen, 1991), or emotional faces (Holmes & Lourenco,
2011) into analog/magnitude representations. We found an overall
recency effect—this did not differ significantly between the abstract
and concrete conditions. Rather, our best fitting model indicated
that concrete thinking resulted in higher levels of regression toward
the mean (lower slopes) compared with abstract thinking. One way
to understand the increased regression to the mean in the concrete
condition is viewing it as a compensatory strategy in face of uncer-
tainty or difficulty. This possible explanation assumes that partici-
pants in the concrete condition experienced more difficulty and/or
uncertainty in their estimations and as a result provided a response
that was closer to the mean of the entire block (Anobile et al., 2012;
Figure 5
Task and Results of Experiment 4
Note. (A) After each sequence of numbers, participants were presented with two probe-numbers and asked to
choose which one of them has been presented in the sequence. (B) Memory accuracy rates by mindset condition.
Blue and orange circles correspond to the accuracy rates of each participant in the abstract and concrete condi-
tions, respectively. Gray circles correspond to the conditions means. The central mark on each box-plot indicates
the median, and the bottom and top edges indicate the 25th and 75th percentiles, respectively. †p,.10. See the
online article for the color version of this figure.
7
Across the three experiments, the RMSD for the abstract thinking was
smaller than that of the concrete thinking, t(174) = 2.86, p= .004. In
addition, there was a high correlation between the two precision measures,
Pearson correlation and RMSD, across participants and conditions (r=
–.61, p,.001)
8HADAR, GLICKMAN, TROPE, LIBERMAN, AND USHER
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Jazayeri & Shadlen, 2010). Indeed, previous work suggests that
participants have access to their averaging uncertainty as their esti-
mation confidence decreases with the variance of the sequence
(Rosenbaum et al., 2021). Future studies could examine the specific
strategy that gives rise to more regressive responses in the concrete
processing condition: Is it the case that participants in this condition
represent (a limited number of) individual samples rather than keep-
ing track of a “running average”? Do they fail to keep track of the
boundaries between different trials? Do they provide a “default”re-
gressive middle-of-the-scale response on a larger proportion of the
trials? All of these (nonmutually exclusive) strategies would pro-
duce more regressive responses in the concrete than the abstract
condition.
Finally, the measure of signal-to-noise ratio (slope/error-term)
was higher for the abstract condition, indicating that abstract think-
ing created a less noisy representation of the sequences of values in
each trial. We find this pattern of results noteworthy, because it indi-
cates that abstract processing gave rise to better, more accurate inte-
gration, that at the same time was not characterized by increased
information loss (as would be predicted, e.g., if abstraction would
lead to heuristic, less elaborate processing).
Another prediction derived from CLT is that an abstract mindset
would improve performance when gist extraction or filtering-out
irrelevant information is beneficial (e.g., Hadar et al., 2019), but not
when one needs to retain the exact details. The results of Experiment
4 are consistent with this prediction. However, future research with
a larger sample of participants and stimuli types will be needed to
examine this hypothesis more closely.
In conclusion, people may face a need to average series of val-
ues when they consider which stock to buy, which people to
befriend, or which student shows more promising performance.
Understanding the contextual features that contribute to the accu-
racy of the averaging process could thus inform not only theories
of decision-making and (social) cognition, but also educators
and policymakers.
References
Akaike, H. (1974). A new look at the statistical model identification. IEEE
Transactions on Automatic Control,19(6), 716–723. https://doi.org/10
.1109/TAC.1974.1100705
Anderson, N. H. (1971). Integration theory and attitude change. Psycholog-
ical Review,78(3), 171–206. https://doi.org/10.1037/h0030834
Anderson, N. H. (1981). Foundations of information integration theory
(Vol. 578). Academic Press.
Anobile, G., Cicchini, G. M., & Burr, D. C. (2012). Linear mapping of
numbers onto space requires attention. Cognition,122(3), 454–459.
https://doi.org/10.1016/j.cognition.2011.11.006
Ariely, D. (2001). Seeing sets: Representation by statistical properties.
Psychological Science,12(2), 157–162. https://doi.org/10.1111/1467
-9280.00327
Asch, S. E. (1946). Forming impressions of personality. JournalofAbnormal
and Social Psychology,41,258–290. https://doi.org/10.1037/h0055756
Baek, J., & Chong, S. C. (2020). Ensemble perception and focused atten-
tion: Two different modes of visual processing to cope with limited
capacity. Psychonomic Bulletin & Review,27(4), 602–606. https://doi
.org/10.3758/s13423-020-01718-7
Betsch, T., Plessner, H., Schwieren, C., & Gütig, R. (2001). I like it but I
don’t know why: A value-account approach to implicit attitude
formation. Personality and Social Psychology Bulletin,27(2), 242–253.
https://doi.org/10.1177/0146167201272009
Brezis, N., Bronfman, Z. Z., & Usher, M. (2015). Adaptive spontaneous
transitions between two mechanisms of numerical averaging. Scientific
Reports,5(1), 10415. https://doi.org/10.1038/srep10415
Brezis, N., Bronfman, Z. Z., Jacoby, N., Lavidor, M., & Usher, M. (2016).
Transcranial direct current stimulation over the parietal cortex improves
approximate numerical averaging. Journal of Cognitive Neuroscience,
28(11), 1700–1713. https://doi.org/10.1162/jocn_a_00991
Brezis, N., Bronfman, Z. Z., & Usher, M. (2018). A perceptual-like pop-
ulation-coding mechanism of approximate numerical averaging. Neu-
ral Computation,30(2), 428–446. https://doi.org/10.1162/neco_a
_01037
Brusovansky, M., Glickman, M., & Usher, M. (2018). Fast and effec-
tive: Intuitive processes in complex decisions. Psychonomic Bulletin
&Review,25(4), 1542–1548. https://doi.org/10.3758/s13423-018
-1474-1
Brusovansky, M., Vanunu, Y., & Usher, M. (2019). Why we should quit
while we’re ahead: When do averages matter more than sums? Decision,
6(1), 1–16. https://doi.org/10.1037/dec0000087
Brysbaert, M., Warriner, A. B., & Kuperman, V. (2014). Concreteness ratings
for 40 thousand generally known English word lemmas. Behavior Research
Methods,46(3), 904–911. https://doi.org/10.3758/s13428-013-0403-5
Burgoon, E. M., Henderson, M. D., & Markman, A. B. (2013). There are
many ways to see the forest for the trees: A tour guide for abstraction.
Perspectives on Psychological Science,8(5), 501–520. https://doi.org/10
.1177/1745691613497964
Chong, S. C., & Treisman, A. (2003). Representation of statistical proper-
ties. Vision Research,43(4), 393–404. https://doi.org/10.1016/S0042
-6989(02)00596-5
Chong, S. C., & Treisman, A. (2005). Statistical processing: Computing
the average size in perceptual groups. Vision Research,45(7), 891–900.
https://doi.org/10.1016/j.visres.2004.10.004
Chong, S. C., Joo, S. J., Emmanouil, T. A., & Treisman, A. (2008). Statis-
tical processing: Not so implausible after all. Perception & Psychophy-
sics,70(7), 1327–1334. https://doi.org/10.3758/pp.70.7.1327
Dakin, S. C., & Watt, R. J. (1997). The computation of orientation statis-
tics from visual texture. Vision Research,37(22), 3181–3192. https://doi
.org/10.1016/S0042-6989(97)00133-8
de Fockert, J., & Wolfenstein, C. (2009). Rapid extraction of mean identity
from sets of faces. Quarterly Journal of Experimental Psychology:
Human Experimental Psychology,62(9), 1716–1722. https://doi.org/10
.1080/17470210902811249
Dehaene, S. (1992). Varieties of numerical abilities. Cognition,44(1-2),
1–42. https://doi.org/10.1016/0010-0277(92)90049-N
Dehaene, S., & Cohen, L. (1991). Two mental calculation systems: A case
study of severe acalculia with preserved approximation. Neuropsycholo-
gia,29(11), 1045–1054. https://doi.org/10.1016/0028-3932(91)90076-K
Ding, Y., & Keh, H. T. (2017). Consumer reliance on intangible versus
tangible attributes in service evaluation: The role of construal level.
Journal of the Academy of Marketing Science,45(6), 848–865. https://
doi.org/10.1007/s11747-017-0527-8
Efrat-Treister, D., Daniels, M. A., & Robinson, S. L. (2020). Putting time
in perspective: How and why construal level buffers the relationship
between wait time and aggressive tendencies. Journal of Organizational
Behavior,41(3), 294–309. https://doi.org/10.1002/job.2433
Ekman, P., & Friesen, W. V. (1976). Measuring facial movement. Environ-
mental Psychology and Nonverbal Behavior,1(1), 56–75. https://doi
.org/10.1007/BF01115465
Eyal, T., Hoover, G. M., Fujita, K., & Nussbaum, S. (2011). The effect of
distance-dependent construals on schema-driven impression formation.
Journal of Experimental Social Psychology,47(1), 278–281. https://doi
.org/10.1016/j.jesp.2010.10.007
ABSTRACTION FACILITATES AGGREGATION OF INFORMATION 9
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Florey, J., Clifford, C. W., Dakin, S., & Mareschal, I. (2016). Spatial limi-
tations in averaging social cues. Scientific Reports,6(1), 32210. https://
doi.org/10.1038/srep32210
Freitas, A. L., Gollwitzer, P., & Trope, Y. (2004). The influence of abstract
and concrete mindsets on anticipating and guiding others’self-regula-
tory efforts. Journal of Experimental Social Psychology,40(6),
739–752. https://doi.org/10.1016/j.jesp.2004.04.003
Gilead, M., Liberman, N., & Maril, A. (2014). From mind to matter: Neural
correlates of abstract and concrete mindsets. Social Cognitive and Affective
Neuroscience,9(5), 638–645. https://doi.org/10.1093/scan/nst031
Gilead, M., Trope, Y., & Liberman, N. (2020). Above and beyond the con-
crete: The diverse representational substrates of the predictive brain. Be-
havioral and Brain Sciences,43,e121.https://doi.org/10.1017/S0140525
X19002000
Haberman, J., Brady, T. F., & Alvarez, G. A. (2015). Individual differen-
ces in ensemble perception reveal multiple, independent levels of en-
semble representation. Journal of Experimental Psychology: General,
144(2), 432–446. https://doi.org/10.1037/xge0000053
Haberman, J., & Whitney, D. (2007). Rapid extraction of mean emotion
and gender from sets of faces. Current Biology,17(17), R751–R753.
https://doi.org/10.1016/j.cub.2007.06.039
Haberman, J., Harp, T., & Whitney, D. (2009). Averaging facial expression
over time. Journal of Vision,9(11), 1. https://doi.org/10.1167/9.11.1
Hadar, B., Luria, R., & Liberman, N. (2019). Concrete mindset impairs fil-
tering in visual working memory. Psychonomic Bulletin & Review,
26(6), 1917–1924. https://doi.org/10.3758/s13423-019-01625-6
Hadar, B. (2021, June 2). Abstract thinking facilitates aggregation of infor-
mation.osf.io/y4uxd/
Hamilton, D. L., & Sherman, S. J. (1996). Perceiving persons and groups.
Psychological Review,103(2), 336–355. https://doi.org/10.1037/0033
-295X.103.2.336
Hansen, J., & Trope, Y. (2013). When time flies: How abstract and con-
crete mental construal affect the perception of time. Journal of Exper-
imental Psychology: General,142(2), 336–347. https://doi.org/10
.1037/a0029283
Holmes,K.J.,&Lourenco,S.F.(2011).Commonspatialorganization
of number and emotional expression: A mental magnitude line. Brain
and Cognition,77(2), 315–323. https://doi.org/10.1016/j.bandc.2011
.07.002
Hubert-Wallander, B., & Boynton, G. M. (2015). Not all summary statis-
tics are made equal: Evidence from extracting summaries across time.
Journal of Vision,15(4), 5. https://doi.org/10.1167/15.4.5
Jazayeri, M., & Shadlen, M. N. (2010). Temporal context calibrates inter-
val timing. Nature Neuroscience,13(8), 1020–1026. https://doi.org/10
.1038/nn.2590
Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus & Giroux.
Khayat, N., & Hochstein, S. (2018). Perceiving set mean and range: Auto-
maticity and precision. Journal of Vision,18(9), 23. https://doi.org/10
.1167/18.9.23
Kille, D. R., Eibach, R. P., Wood, J. V., & Holmes, J. G. (2017). Who
can’t take a compliment? The role of construal level and self-esteem in
accepting positive feedback from close others. Journal of Experimen-
tal Social Psychology,68(1), 40–49. https://doi.org/10.1016/j.jesp
.2016.05.003
Liberman, N., & Trope, Y. (2014). Traversing psychological distance.
Trends in Cognitive Sciences,18(7), 364–369. https://doi.org/10.1016/j
.tics.2014.03.001
Malmi, R. A., & Samson, D. J. (1983). Intuitive averaging of categorized
numerical stimuli. Journal of Verbal Learning and Verbal Behavior,
22(5), 547–559. https://doi.org/10.1016/S0022-5371(83)90337-7
Maule, J., Witzel, C., & Franklin, A. (2014). Getting the gist of multiple
hues: Metric and categorical effects on ensemble perception of hue. Jour-
nal of the Optical Society of America. A, Optics, Image Science, and
Vision,31(4), A93–A102. https://doi.org/10.1364/JOSAA.31.000A93
Napier,J.L.,Luguri,J.B.,Dovidio,J.F.,&Oltman,K.A.(2018).
Construing the essence: The effects of construal level on genetic
attributions for individual and social group differences. Personality
and Social Psychology Bulletin,44(9), 1395–1407. https://doi.org/10
.1177/0146167218768799
Neumann, M. F., Schweinberger, S. R., & Burton, A. M. (2013). Viewers
extract mean and individual identity from sets of famous faces. Cognition,
128(1), 56–63. https://doi.org/10.1016/j.cognition.2013.03.006
Parkes, L., Lund, J., Angelucci, A., Solomon, J. A., & Morgan, M. (2001).
Compulsory averaging of crowded orientation signals in human vision.
Nature Neuroscience,4(7), 739–744. https://doi.org/10.1038/89532
Post, R. B., Haberman, J., Iwaki, L., & Whitney, D. (2012). The frozen
face effect: Why static photographs may not do you justice. Frontiers in
Psychology,3(1), 22. https://doi.org/10.3389/fpsyg.2012.00022
Robitaille, N., & Harris, I. M. (2011). When more is less: Extraction of
summary statistics benefits from larger sets. Journal of Vision,11(12),
18. https://doi.org/10.1167/11.12.18
Rosenbaum, D., Glickman, M., & Usher, M. (2021). Extracting sum-
mary statistics of rapid numerical sequences. Frontiers in Psychol-
ogy,12, 693575. https://doi.org/10.3389/fpsyg.2021.693575
Simchon, A. (2019). TATE: Text analysis tools for English (R package
version 3.0) [Computer software]. https://github.com/almogsi/TATE
Solomon, J. A. (2010). Visual discrimination of orientation statistics in
crowded and uncrowded arrays. Journal of Vision,10(14), 19. https://
doi.org/10.1167/10.14.19
Spitzer, B., Waschke, L., & Summerfield, C. (2017). Selective over-
weighting of larger magnitudes during noisy numerical comparison.
Nature Human Behaviour,1(8), 145. https://doi.org/10.1038/s41562
-017-0145
Spunt, R. P., Kemmerer, D., & Adolphs, R. (2016). The neural basis of
conceptualizing the same action at different levels of abstraction. Social
Cognitive and Affective Neuroscience,11(7), 1141–1151. https://doi
.org/10.1093/scan/nsv084
Stillman, P. E., Medvedev, D., & Ferguson, M. J. (2017). Resisting tempta-
tion: Tracking how self-control conflicts are successfully resolved in
real time. Psychological Science,28(9), 1240–1258. https://doi.org/10
.1177/0956797617705386
Sweeny, T. D., & Whitney, D. (2014). Perceiving crowd attention: Ensem-
ble perception of a crowd’s gaze. Psychological Science,25(10),
1903–1913. https://doi.org/10.1177/0956797614544510
Trope, Y., & Liberman, N. (2010). Construal-level theory of psychological
distance. Psychological Review,117(2), 440–463. https://doi.org/10.1037/
a0018963
Usher, M., & McClelland, J. L. (2001). The time course of perceptual
choice: The leaky, competing accumulator model. Psychological Review,
108(3), 550–592. https://doi.org/10.1037/0033-295X.108.3.550
vanOsch,Y.,Blanken,I.,Meijs,M.H.,&vanWolferen,J.(2015).A
group’s physical attractiveness is greater than the average attractive-
ness of its members: The group attractiveness effect. Personality and
Social Psychology Bulletin,41(4), 559–574. https://doi.org/10.1177/
0146167215572799
Vanunu, Y., Pachur, T., & Usher, M. (2019). Constructing preference from
sequential samples: The impact of evaluation format on risk attitudes.
Decision,6(3), 223–236. https://doi.org/10.1037/dec0000098
Walker, D., & Vul, E. (2014). Hierarchical encoding makes individuals in
a group seem more attractive. Psychological Science,25(1), 230–235.
https://doi.org/10.1177/0956797613497969
Watamaniuk, S. N., & Duchon, A. (1992). The human visual system aver-
ages speed information. Vision Research,32(5), 931–941. https://doi
.org/10.1016/0042-6989(92)90036-I
Watamaniuk, S. N., & McKee, S. P. (1998). Simultaneous encoding of
direction at a local and global scale. Perception & Psychophysics,60(2),
191–200. https://doi.org/10.3758/BF03206028
10 HADAR, GLICKMAN, TROPE, LIBERMAN, AND USHER
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Watamaniuk, S. N., Sekuler, R., & Williams, D. W. (1989). Direction per-
ception in complex dynamic displays: The integration of direction infor-
mation. Vision Research,29(1), 47–59. https://doi.org/10.1016/0042
-6989(89)90173-9
Whitney, D., & Yamanashi Leib, A. (2018). Ensemble perception. Annual
Review of Psychology,69(1), 105–129. https://doi.org/10.1146/annurev
-psych-010416-044232
Yamanashi Leib, A., Chang, K., Xia, Y., Peng, A., & Whitney, D. (2020).
Fleeting impressions of economic value via summary statistical representa-
tions. Journal of Experimental Psychology: General,149(10), 1811–1822.
https://doi.org/10.1037/xge0000745
Yamanashi Leib, A., Kosovicheva, A., & Whitney, D. (2016). Fast ensem-
ble representations for abstract visual impressions. Nature Communica-
tions,7(1), 13186. https://doi.org/10.1038/ncomms13186
Yudkin, D. A., Liberman, N., Wakslak, C., & Trope, Y. (2020). Better off
and far away: Reactions to others’outcomes depends on their distance.
Organizational Behavior and Human Decision Processes,156,13–23.
https://doi.org/10.1016/j.obhdp.2019.10.004
Received September 7, 2020
Revision received June 1, 2021
Accepted July 9, 2021 n
ABSTRACTION FACILITATES AGGREGATION OF INFORMATION 11
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.