Content uploaded by Michele Vicovaro
Author content
All content in this area was uploaded by Michele Vicovaro on Feb 01, 2024
Content may be subject to copyright.
Does perceptual disfluency affect the illusion of causality?
Preprint Version
Stefano Dalla Bonaa and Michele Vicovaroa
September, 2023
This document constitutes the
Preprint
version of our article, accessibile via the following
DOI link:
https://doi.org/10.1177/17470218231220928
a Department of General Psychology, University of Padova, Italy
1
0. ABSTRACT
When a subjective experience of difficulty is associated with a mental task, people tend to engage in
systematic and deliberative reasoning, which can reduce the usage of intuitive and effortless
thinking that gives rise to cognitive biases. One such bias is the illusion of causality, where people
perceive a causal link between two unrelated events. Díaz-Lago and Matute (2019a) found that a
superficial perceptual feature of the task could modulate the magnitude of the illusion (i.e., a hard-
to-read font led to a decrease in the magnitude of the illusion). The present study explored the
generalizability of the idea that perceptual disfluency can lead to a decrease in the magnitude of the
illusion. In the first experiment, we tested whether a physical-perceptual manipulation of the
stimuli, specifically the contrast between the written text and the background, could modulate the
illusion in a contingency learning task. The results of the online experiment (N=200) showed no
effect of contrast on the magnitude of the illusion, despite our manipulation successfully induced
task fluency or disfluency. Building upon this null result, our second experiment (N=100) focused
on manipulating the font type, in the attempt to replicate the results obtained by Díaz-Lago and
Matute (2019a). In contrast to their findings, we found no discernible effect of font type on the
magnitude of the illusion, even though this manipulation also effectively induced variations in task
fluency or disfluency. These results underscore the notion that not all categories of (dis)fluency in
cognitive processing wield a modulatory influence on cognitive biases, and they call for a re-
evaluation and a more precise delineation of the (dis)fluency construct.
Keywords: Causality bias, Illusion of causality, Process fluency, Causal cognition.
2
1. INTRODUCTION
The term cognitive illusion (or cognitive bias) refers to a wide range of phenomena, which together
show how thinking, making judgments, and/or remembering deviate from an objectively correct
standard. The trigger for the interest in the field of cognitive biases can be traced back to the early
70s (Tversky & Kahneman, 1996), with the introduction of a research program on judgment under
uncertainty (Tversky & Kahneman, 1974).
Over the years, some authors have made an effort to identify and list all cognitive biases (e.g.,
Benson, 2016), whereas others have argued that no accurate and unitary definition of cognitive bias
is actually possible (Caverni et al., 1990). According to Roediger (1996), an analogy can be drawn
between cognitive biases and optical illusions. Just as sensory processing can lead to the
misperception of a physical stimulus, processes of codification, elaboration, and/or retention of
information can lead to a plethora of judgment errors. Due to the limited processing capacity of the
cognitive system, in many occasions of judgment under uncertainty the human mind employs a
small set of rules of thumb that can lead to severe and systematic errors (Tversky & Kahneman,
1974). Nevertheless, Gigerenzer (2008) underscored that the influence of intuitive and effortless
cognitive processes is not inherently detrimental. In fact, they enable the attainment of reasonably
accurate outcomes while conserving valuable cognitive resources. Recently, Pohl (2022) highlighted
five analogies between cognitive biases and optical illusions: (i) deviation from reality, (ii) systematic
deviation from the standard, (iii) involuntary production of the illusion, (iv) impossibility to avoid
the illusion, (v) universal appearance of the illusion. As will become evident in the ensuing
discussion, certain authors have raised objections to the concept of these biases being universally
inescapable. Instead, they emphasize that these biases can be mitigated under specific conditions.
The present study focuses on a specific cognitive bias called the illusion of causality (or
causality bias). According to the classification proposed by Pohl (2022), this illusion can be included
in the illusion of thinking category (the other two categories are illusion of memory and illusion of
3
judgment). Illusions of thinking are characterized by the fact that, in a given scenario, the correct
solution requires the precise application of a rather simple mathematical or/and logical rule derived
from a normative model. Typically, the results show a discrepancy between human performance
and the predictions from the normative model.
1.1 THE ILLUSION OF CAUSALITY
Causality is the relation between two events, one of which is the consequence (or effect) of the other
(cause). The understanding of the cause-effect relations is an ability in which humans clearly
outperform any other species (Bender, 2020). Since the first modern systematic philosophical
explanation by David Hume (Wasserman, 1990), causality has been a traditional research topic in
psychology. It has been explored from various perspectives, including comparative cognition
(Blaisdell et al., 2006), visual perception (Michotte, 1963), psychology of reasoning (Waldmann et
al., 2006), and psychology of learning (Dickinson et al., 1984). As for the latter perspective, which
lays the foundation for our study paradigm (Wasserman, 1990), a series of studies have provided
some evidence for the idea that human judgments of contingency between two events can be
analysed from an associationist perspective, and more particularly that contingency judgment can
be described in terms of animal conditioning (Shanks et al., 1996).
The illusion of causality occurs when the subject develops the belief that there is a causal
connection between two events that are actually unrelated. It refers to the perception that one event
A, called cue or potential cause, is causally linked to another event B, called outcome or effect, when
there is just a coincidence between them. Generally, humans infer the presence of a causality link
through the (single or multiple) contingencies between two events (Matute et al., 2019), often
showing great accuracy in the detection of causal links that are actually present in the environment.
This ability is critical for survival, as it underlies the capacity to make accurate predictions about the
future states of the world. However, contingency learning can sometimes drag to an overestimation
4
of the degree of a causal link, when in fact the two events are independent (i.e., the probability of
A is substantially independent of the probability of B), the so-called over-estimation of zero-
contingencies (Blanco et al., 2014).
If we take two events A (cue event) and B (outcome), their possible presence or absence can
give rise to four hypothetical scenarios (Tab. A): (a) the cue and the outcome are present, (b) only
the cue is present, (c) only the outcome is present, (d) neither the cue nor the outcome is present
(Vadillo & Matute, 2007). One of the most widely used paradigms to explore the illusion of
causality is the contingency learning task (CLT), in which a series of events characterized by the
presence or absence of a cue and an outcome are presented to participants. Each trial corresponds
to one of the four possible scenarios represented in Tab. A. Participants act as outsider observers of
sequences of scenarios presented trial by trial. After all the trials have been presented to the
participants, they are asked to judge the strength of the causal link between the cue and outcome.
Outcome present
Outcome absent
Cue present
a
b
Cue absent
c
d
Tab. A: Contingency Table. This table shows the four possible combinations that can occur from
the presence/absence of the cue event and the potential outcome.
In the literature, different normative models of causal induction have been proposed
(Perales & Shanks, 2007). The most widely used method to measure contingency (Díaz-Lago &
Matute, 2019a) is the ΔP contingency index (Allan, 1980), a normative model for human causal
learning (Matute et al., 2022). To compute the ΔP, the probability to observe the outcome when
the cue event is not present (i.e.,
!"#$%&'
) is subtracted from the probability to observe the
outcome when the cue is present (i.e.,
!"#$&'
) (Jenkins & Ward, 1965):
5
!"#$&' ( !"#$%&' ) * ) +
"+ , -' (.
". , /'
Equation 1.
where a, b, c, and d are the observed frequencies of the four scenarios represented in Tab. A.
If ΔP is equal to zero (i.e.,
!"#$&' 0 !"#$%&'
), then the contingency between the cue and
the outcome event is null and there is no causal link between them. A causal link is present if ΔP >
0 or ΔP < 0 (if ΔP < 0, then it means that the cue event has an inhibitive effect on the outcome).
Theoretically, the overestimation of the extent to which A and B are causally related can
occur in any of these cases (i.e., when ΔP is negative, null, or positive). However, the causality bias
has been appreciated and studied mostly in the case of the null contingency (Allan & Jenkins, 1980).
In line with Díaz-Lago & Matute (2019a), in the present study we specifically focus on the
Outcome-Density Bias (Matute et al., 2015), which occurs when the frequencies of the scenarios in
which the outcome is present (cells a and c in Tab. A) are larger with respect to the frequencies of
the scenarios in which the outcome is absent (cells b and d in Tab. A), despite the fact that the ΔP
index is equal to zero (Alloy & Abramson, 1979). Though not directly relevant to our present study,
it is important to note that a causality bias can also be obtained by increasing the frequency of
scenarios a and b relatively to scenarios c and d (Cause-Density Bias; Matute et al., 2015) and by
increasing the frequency of a relatively to the other scenarios, while keeping the ΔP index fixed to
zero (Blanco, 2017; Blanco et al., 2013).
1.2 DEBIASING THROUGH PROCESS (DIS)FLUENCY
The illusion of causality is a pervasive phenomenon that can have important consequences in
several life areas such as health and politics (Matute et al., 2022). Many social judgments and
behaviours depend on intuitive judgments about causal relationships between events (Crocker,
6
1981), and it has been suggested that pseudoscientific thinking and beliefs can be considered as
instances of the illusion (Matute et al., 2011). Griffiths et al. (2019) have shown that people prone to
superstitious beliefs are also prone to the causality bias.
In light of these considerations, it is no surprise that researchers have devoted some efforts
to the development of possible strategies to reduce the causality bias (Matute et al., 2022). On one
hand, psychoeducational interventions have been developed to eliminate cognitive biases or
diminish their intensity and frequency (Lilienfeld et al., 2009; but see Arkes, 1981 for the critics on
the usage of debiasing techniques; see also Barberia et al.,2013 and Barberia et al., 2018 for some
interventions focusing directly on the reduction of the causality bias). On the other hand, recent
studies have found that a reduction of the causality bias can be obtained through the manipulation
of the information presented in classic CLT tasks. For instance, a reduced causality bias was found
when the information was presented in a foreign language (Díaz-Lago & Matute, 2019b). This
result is coherent with the foreign language effect (Costa et al., 2014), a well-known phenomenon
in the bilingual study field, in which a modification in the cognitive processes responsible for
judgment and decision-making is observed when participants have to actively think in a non-native
language (McFarlane et al., 2020).
Interestingly, the results of a recent study by Díaz-Lago and Matute (2019a) suggest that a
rather superficial feature of the information presented in a CLT, namely the text font, can also
affect the causality bias. In particular, in one condition of the experiment, a sample of participants
were presented with a null contingency CLT characterized by an easy-to-read font. In a second
experimental condition, a different sample of participants were presented with the same null-
contingency CLT characterized by a hard-to-read font. The results showed a reduced magnitude of
the causality bias in the hard-to-read font condition with respect to the easy-to-read font condition.
As a possible general explanation of the influence of language and font type on the illusion
of causality, it has been suggested that the magnitude of illusion is directly related to processing
7
fluency. Processing fluency is the subjective experience of the ease with which people process
information (Oppenheimer, 2008) and acts like a metacognitive cue that plays an important role in
human judgment (Alter & Oppenheimer, 2009). It pertains to the ease or difficulty with which
new information can be processed (Schwarz, 2004), and it can be conceptualized as a feeling of ease
associated with a cognitive operation. The processing fluency of a task can be manipulated in
various forms depending on the type of cognitive process involved in the task itself. This makes
processing fluency a difficult construct to define.
Processing fluency can be captured with objective as well as with subjective measures, and
the effects of its experimental manipulation appear to extend beyond the boundaries of the illusion
of causality. Indeed, effects of processing fluency manipulation have been documented in various
domains, such as affective judgment (Reber et al., 1998) and the judgment of true statement (Reber
& Schwarz, 1999). Furthermore, processing fluency has an impact on classical tasks that are used to
measure cognitive biases. For instance, it was shown that a disfluency increment can lead to an
attenuation of the Moses illusion (Song & Schwarz, 2008) and to a reduction of the framing effect
(Korn et al., 2018).
The well-documented relationship between processing fluency and cognitive biases has
been taken as evidence in support for a dual process approach to reasoning (Oppenheimer, 2008).
According to this view, the task fluency, manipulated potentially via numerous types of variables,
can influence the strategy adopted by people in the reasoning process. A disfluent manipulation
seems to cause the engagement of more effortful processing strategies, shifting from an effortless,
automatic, and heuristic kind of reasoning (i.e., system one), to a more deliberative, systematic, and
analytical kind of reasoning (i.e., system two; Kahneman & Frederick, 2002). However, this type of
explanation does not differentiate between different processing types, as processing fluency is
intended in terms of a general perceived difficulty of how the process feels.
8
It can be argued that a flaw of the processing fluency construct is its multidimensional
nature. It appears to function as an umbrella construct, as various types of manipulations that affect
the subjective level of difficulty of information processing are expected to produce similar effects
on cognitive biases (Alter & Oppenheimer, 2009). However, it is important to recognize that
different types of fluency manipulations are likely to involve distinct underlying cognitive
processes. For instance, manipulating language fluency (such as the reader’s mother language vs. an
unbalanced foreign language) presumably affects the fluency of syntactic and semantic information
processing, which occurs at a later stage of text comprehension. In contrast, manipulating font type
affects only the perceptual processing of written text, which is an earlier stage of text
comprehension. Therefore, the idea that these two forms of processing fluency manipulations will
have similar effects on the level of engagement of different information processing strategies (i.e.,
system one vs. system two), despite the fact that they involve different cognitive processes, requires
further empirical investigation.
The results of the study by Díaz-Lago and Matute (2019a) indicate that perceptual
disfluency, manipulated through font type, can lead to a reduction in the magnitude of the illusion
of causality. Our first experiment sought to test the generalizability of the idea that perceptual
disfluency can lead to a reduction in the magnitude of the illusion of causality.
9
2. EXPERIMENT 1
In the context of a classic CLT paradigm, we manipulated the physical-perceptual characteristics of
the stimuli by manipulating the contrast of the written stimuli with the background, using
different colours (i.e., blue text on a white background vs. yellow test on a white background). In
so doing, we manipulated the difficulty with which participants could perceive the target stimuli
(i.e., perceptual fluency). Contrast manipulation has already been found to be a reliable source of
processing fluency variation (Reber & Schwarz, 1999), e.g., in judgments of agent competence
(Thompson & Ince, 2013). More specifically, we used an online version of a standard CLT in two
conditions: high contrast (HC) of the written stimuli with the background (blue on white), and
low contrast (LC) of the written stimuli with the background (yellow on white).
To test if the contrast manipulation had an effect on the magnitude of the causality bias,
the two contrast conditions were combined with two different contingency scenarios, according to
a 2 (Contrast: HC vs LC) x 2 (Contingency: True Contingency vs Null Contingency) between-
subject factorial design. The true contingency scenario was characterized by positive ΔP and thus
by the actual existence of a causal link between the cue and the outcome event. The null
contingency scenario was instead characterized by null ΔP and thus by the lack of a real causal link
between the cue and the outcome event. An outcome-density scenario was used because, according
to previous studies (e.g., Díaz-Lago & Matute, 2019a, 2019b), this scenario should give rise to a
robust illusion of causality.
If perceptual disfluency (i.e., low contrast) reduces the magnitude of the illusion of
causality, then, in the null contingency scenario, lower causality ratings are expected to emerge in
the case of low contrast than in the case of high contrast. In line with the results from previous
studies (Díaz-Lago & Matute, 2019a, 2019b), no effect of perceptual disfluency is expected to
emerge in the true contingency scenario.
10
The pre-registered hypotheses and methods are available on Open Science Framework
(OSF) at the following link:
https://osf.io/74d6g
. The codes for the experiment, the raw data and
the script used for the main analyses are available on OSF at the following link:
https://osf.io/c26qa
.
2.1 PARTICIPANTS
To establish the sample size, we conducted a power analysis using the pwr package (Champely,
2020) in R (R-Core-Team, 2023). We relied on the results of the previous study by Díaz-Lago &
Matute (2019a) who found, in the null contingency scenario, a medium effect size (d=0.58) for the
difference between the rated strength of cause-effect relationships in the two conditions (i.e., easy-
to-read font vs. hard-to-read font). Our goal was to obtain a 0.8 power to detect a medium effect
size of 0.58 at the standard .05 alpha error probability. The result of the analysis was N = 37, with
N referring to the number of participants in each group. Because there were four groups (i.e., 2
Contingency × 2 Contrast), testing a minimum of 148 participants was necessary. However, to
increase further the statistical power of the critical comparison between the two contrast conditions
(HC vs. LC) in the null contingency scenario, we decided to collect 60 participants for each of the
two groups, raising the statistical power above 0.9. Forty participants were instead recruited for
each of the two true contingency groups (HC vs. LC), for a total of 200 participants. To sum up,
participants were randomly assigned to each of the four conditions according to a 3 (Null
Contingency, LC):3 (Null Contingency, HC):2 (True Contingency, LC):2 (True Contingency,
HC) proportion.
Participants were recruited through various forms of advertisement on social
networks. People who agreed to participate had a chance to win 25 euros through a lottery. A total
of 209 participants took part in the experiment, however the data of nine participants were
11
excluded from the analyses according to the following criteria: (i) doing the experiment twice (five
participants); (ii) reading the fictitious story in less than 10 seconds (four participants); (iii)
completing the experiment in less than 180 seconds (zero participants), (iv) completing the trial
section in less than 160 seconds (zero participants), (v-vi) responding to each of the two final
questions in less than 2 seconds (zero participants). The final sample was constituted by 142 females
and 58 males. The average age of participants was 25.92 years (SD=9.59). A Pearson’s chi-square test
of independence showed that the frequencies of the two sexes did not differ significantly across the
four groups, χ2(3) = 3.99, p = 0.26. A Kruskal-Wallis test also showed no systematic differences in
age across the four groups, χ2(3) = 3.20, p = 0.36.
2.2 STIMULI AND PROCEDURE
Figure 1. CLT structure and contrast differences. This figure shows the steps of the procedure
that together composed the experimental online adaptation of the CLT.
12
Before starting the experiment, participants read the online informed consent form approved by
the Ethics Committee for Psychological Research at the University of Padova (protocol number
5010, November 3rd, 2022) and then gave their consent to participate through a response key. Then,
they were automatically assigned to one of four conditions by an online program
1
, which led to the
start of the experiment in the Pavlovia platform
2
. The experiment could only be launched from a
computer or laptop, no other devices could be used. The screen background was set to white.
Participants were asked for three times to position themselves in a sufficiently illuminated room
with no direct light on the screen (i.e., when they were provided with the recruitment link, in the
informed consent form, and in the first instructions screen).
The experiment consisted of a standard CLT (Fig. 1), using an adaptation of the allergy task
(Wasserman et al., 1990). The task was presented in Italian language. In the first phase of the
experiment (Instructions, see Fig 1a), a fictitious story was presented to the participants.
Participants impersonated an emergency room personnel and they were instructed to determine if
there was a causal relationship between the presence of a medicine “Batatrim” (cue) and the healing
of the disease “Lindsay Syndrome” (outcome). Then, in the second phase (Patient records, see Fig.
1b), participants were exposed to a succession of 40 patient records (i.e., 40 trials, ITI= 1 sec), in a
random order. Each record described one of four different scenarios, given by the possible presence
or absence of the cue (i.e., the patient had or had not taken the medicine to recover from the disease)
and the possible presence or absence of the outcome (i.e., the patient had or had not recovered).
Through the manipulation of the frequency of the four scenarios (see Tab. A), we created two
different contingency conditions, namely a null contingency condition, in which ΔP = 0, and a true
contingency condition, in which ΔP =0.60. The exact frequencies of the four scenarios in each
condition are reported in Tab. B. It should be noticed that, in the null contingency condition, the
1
For the informed consent form and the self-balancing randomization, we used the VESPR online studies portal (Morys-
Carter, 2023).
2
The experiment has been programmed from the ground up using PsychoPy (Peirce et al., 2019). The program has been
hosted on Pavlovia website (
https://pavlovia.org
).
13
probability of the presence of the outcome (i.e., P = .75) was much higher than the probability of
the absence of the outcome (i.e., P = .25). According to the results from previous studies, this should
lead to a Cause-Density Bias (Matute et al., 2015). Each patient record was composed of three
horizontal panels. The upper panel remained visible for the whole duration of the trial, and
informed the participant about the presence/absence of the cue (e.g., “The patient has taken the
Batatrim”). The middle panel remained visible for the whole duration of the trial as well, and it
presented a predictive question, to maintain the attention on the task. The participant was asked
about whether the participant will heal after taking the medicine, by clicking with the mouse on
one of the two buttons (see Fig. 1b). No time limits were provided for the response. After the
response was recorded, a third panel appeared below the middle one, which informed the
participant about whether the patient had recovered or not. It is important to notice that the
response provided by the participant through the mouse click had no influence on the information
provided in the third panel. The three panels disappeared from the screen after 2 seconds, and then
a new patient record was presented. As in Díaz-Lago & Matute (2019a)’s study, we avoided
including any pictures of the drug and the patient, to force the participants to not rely on shortcuts.
In the third phase of the procedure (Causality rating, see Fig. 1c), participants were asked to
judge the strength of the causal relationship between the two events (i.e., “To what extent do you
think that Batatrim was effective in healing the crises of the patients you have seen?”), using a visual
analog scale from 0 (“Definitely Not”) to 100 (“Definitely Yes”). Once participants clicked on the
scale, a cursor appeared, and participants could drag the cursor along the entire range between 0
and 100 to pick the exact judged discrete number. A numeric feedback was presented under the
visual scale.
In the last phase (Reading difficulty rating, see Fig. 1d), participants were asked to judge the
disfluency of the task through a single question (“How difficult have you found the reading activity
during this experiment?”), using a 7-point Likert scale (1= “Very Easy”; 7= “Very Difficult”) that
14
was similar to that used for the causality rating. We used a single item because, in this specific
domain, the application of a single question has been shown to be robust from a psychometric
standpoint and more understandable for participants with respect to multi-item scales (Graf et al.,
2018).
As for the manipulation of perceptual (dis)fluency, in the HC condition, the text for the
instructions, patients records, and causality rating was presented in blue on a white background
(HEX #000063; 17.79 contrast ratio; see the upper half of Fig 1), whereas in the LC condition the
text was presented in yellow on a white background (HEX #FFFF73; 1.07 contrast ratio; see the
lower half of Fig 1)
3
. In the reading difficulty rating phase (Fig 1d), a black text on white background
(HEX #000000; 21 contrast ratio) was used both in the HC and in the LC condition. In each phase
of the experiment an ‘Arial’ font was used, scaled to 0.03 height (i.e, the maximum height of any
letter did not exceed 3% of the height of the screen).
Null Contingency
True Contingency
Outcome
present
Outcome
absent
Probability
Outcom
e present
Outcome
absent
Probability
Cause
present
15
5
1"2$3' =0.75
15
5
1"2$3'=0.75
Cause
absent
15
5
1"2$%3' =0.75
3
17
1"2$%3'=0.15
ΔP =
1"2$3' ( 1"2$%3')
=0.75-0.75=0
ΔP =
1"2$3' ( 1"2$%3')
=0.75-0.15=0.60
Tab. B: Frequency of each scenario and corresponding conditional probabilities in the null
contingency and in the true contingency conditions.
3
The colours have been chosen on the basis of the processing fluency literature (Thompson & Ince, 2013; Reber & Schwarz,
1999). Yellow text on white background has been demonstrated to produce a clear disfluency, whereas dark blue text on white
background has been shown to be easy-to-read. Based on the human-computer interaction literature (Zuffi et al., 2007; Hall
& Hanna, 2004), for the LC condition we had initially opted for a HEX #FFFF00 yellow, which seemed to be sufficiently
disfluent to make the text hard-to-read. However, because the actual contrast in different output devices showed some
variability, we decided through empirical observations to further decrease the contrast ratio, using a HEX #FFFF73 yellow.
15
2.3 RESULTS
We first present the results from the analysis of processing fluency data, and then the results from
the analysis of the causality rating task.
2.3.1 PROCESSING FLUENCY
Fig. 2: Subjective evaluation of reading difficulty in Experiment 1. Histograms and boxplots
showing that the LC conditions (yellow) resulted in a greater subjective difficulty evaluation
compared with the HC conditions (blue). These and the following charts have been created using
the ggplot2 package in R (Wickham, 2016).
First, we expected the LC condition to produce a disfluency effect on participants. The processing
fluency was measured with both subjective and objective indexes. As for the subjective measure, the
results from the single-item question on reading difficulty are represented in Fig. 2. A two-way
16
between-subject analysis of variance (ANOVA) with factors contingency (Null vs. True) and
contrast (HC vs. LC) showed a significant main effect of contrast, F(1,196) = 184.99, p <.001, ηp2
=.48. Consistent with expectations, reading difficulty was rated higher in the LC condition (M =
4.98, SD = 1.68) than in the HC condition (M = 1.96, SD = 1.46). The main effect of contingency
and the two-way interaction were not statistically significant [F(1,196) = 0.179, p =.67, ηp2 = .0004;
F(1,196) = 1.704, p = .19, ηp2 = .004], suggesting that contingency had no effect on perceived reading
difficulty (M = 3.51, SD = 2.14 in the null contingency condition and M = 3.41, SD = 2.23 in the true
contingency condition).
Fig. 3: Experiment time in Experiment 1. Boxplots and violin plots showing how the LC conditions
(yellow) resulted in a longer experiment time compared with the HC conditions (blue).
17
As for the objective measure, we considered the total time to complete the whole experimental
procedure (see Fig. 3). Experiment time was analysed in the same manner as the rated reading
difficulty. The main effect of contrast was statistically significant F(1,196) = 4.61, p = .033, ηp2 = .02,
due to longer experiment time in the LC condition (M = 373.25 sec, SD = 105.68 sec) than in the
HC condition (M = 341.03 sec, SD = 106.26 sec). The main effect of contingency and the two-way
interaction were not statistically significant [F(1,196) = 1.27, p = .26, ηp2 = .006; F(1,196) = 0.34, p =
.56, ηp2 = .002], suggesting that contingency had no effect on experiment time (M = 350.25 sec, SD
= 96.26 sec in the null contingency condition and M = 367.48 sec, SD = 121.09 sec in the true
contingency condition)
4
.
2.3.2 CAUSALITY RATINGS
Our main goal was to test if perceptual disfluency, induced by a low contrast of the written text
with the background, was able to produce the engagement of a more systematic and deliberative
kind of reasoning (i.e., system two), leading to a reduction in the illusion of causality induced by
the outcome-density null contingency scenario. The causality ratings are represented in Fig. 4. A
two-way between-subject ANOVA with factors contingency (Null vs. True) and contrast (HC vs.
LC) showed a statistically significant main effect of contingency, F(1,196) = 12.98, p < .001, ηp2 =.06.
As expected, the causality ratings were larger in the true contingency condition (M = 69.95, SD =
15.82) than in the null contingency condition (M = 60.82, SD = 18.62). However, it is worth
highlighting the large mean value observed in the latter condition, which confirms the presence of
a robust illusion of causality.
4
The same analysis was also performed on log-transformed experiment time, to reduce the possible impact of the skewed
distributions (see Fig. 3). The same pattern of results emerged, which supports the reliability of the main conclusions.
18
The main effect of contrast was not statistically significant, F(1,196) = 0.86, p = .36, ηp 2 =.004, as
the means of the causality ratings in the HC and the LC condition were similar to each other (HC:
M = 63.32, SD = 18.25; LC: M = 65.62, SD= 17.93). Crucially, the two-way interaction was not
statistically significant F(1,196) = 1.14, p = .29, ηp 2 = .005, which is at odds with the hypothesis that
perceptual disfluency induced by low contrast can lead to a decrease in the magnitude of the illusion
of causality in the null contingency scenario.
In line with the pre-registered analysis plan
5
, we also conducted a classic one-tailed
independent sample t-test and a one-sided Bayesian t-test to test if, in the null contingency
5
The hypotheses and the analysis plan were pre-registered on Open Science Framework on November 25, 2022
(
https://osf.io/74d6g
).
Fig.4: Causality ratings in Experiment 1. The raincloud plots summarize the main results of the
experiment, showing the judged strength of causality for the four experimental groups. The red
dots and lines on the charts indicate the means in the experimental groups.
19
condition, the causality ratings in the HC condition were larger than the causality ratings in the LC
condition. The results of the classic t-test were not statistically significant t(118) = -1.32, p =.90, d =
-0.24. It is worth noting that the difference Is in a direction opposite to that hypothesized, as the
causality ratings in the HC condition (M = 58.58, SD = 18.51) were slightly smaller than those in the
LC condition (M=63.05, SD=18.64).
The Bayesian t-test, performed with JASP (JASP-Team, 2023), was characterized by a half-
Cauchy prior distribution with a standard width parameter of 0.707 for the alternative hypothesis
(i.e., we assumed a probability of .50 that the effect size lied between 0 and 0.707). A point-null
hypothesis was instead used as the prior for H0. The results showed that the observed data were
over 11 times more likely under the null hypothesis than under the alternative hypothesis, BF01=
11.07 (see also the Supplementary Figure on OSF).
Based on these results, we can conclude that increased perceptual disfluency, obtained
through the presentation of the CLT experiment with low-contrast written text, did not elicit a
reduction in the magnitude of the illusion of causality.
20
3. EXPERIMENT 2
Our first experiment has been conceptualized as a generalization attempt of the findings by Díaz-
Lago and Matute (2019a), to provide evidence for the illusion-reduction hypothesis via the
manipulation of perceptual characteristics of stimuli in a CLT paradigm. Despite successfully
inducing task fluency or disfluency through contrast manipulation, our attempt at generalization
did not yield the expected results. This null outcome suggests that not all forms of fluency
manipulations can effectively modulate the magnitude of the illusion. Consequently, this result
alone highlights the need to revise the general hypothesis that processing disfluency inevitably
triggers a deliberative and effortful mode of thinking (referred to as system two).
To gain a more comprehensive understanding of the impact of perceptual disfluency on the
causality bias, Experiment 2 was conducted to determine if the original effect reported by Díaz-
Lago and Matute (2019a) would replicate in a highly similar experimental context. Our intention
is to delve deeper into the potential impact of font type on the extent of the causality bias by
conceptually replicating the font study conducted by Díaz-Lago and Matute (2019a).
Based on the results of the reading difficulty task (Section 2.3.1), it is important to recognize
that both the HC true contingency condition and HC null contingency condition of Experiment 1
can be classified as representing easy-to-read font conditions. On a 1-7 Likert scale, the mean rated
reading difficulty was just 1.96 (SD = 1.46; see Fig. 2). In Experiment 2, we conducted a comparison
between the already collected data from these two conditions, with the fresh data from two new
groups of participants. These new groups were tested in a hard-to-read font (Uppercase Brush
Script MT, hereafter abbreviated as Brush) under either a true contingency scenario (ΔP = 0.60) or
a null contingency scenario (ΔP = 0). Consequently, Experiment 2 utilized a 2 (Font: Arial, easy-to-
read vs Brush, hard-to-read) x 2 (Contingency: True Contingency vs Null Contingency) between-
subject factorial design. Instructions with text written in Brush font are represented in Fig. 5.
21
If the use of a hard-to-read font induces perceptual disfluency and subsequently diminishes the
magnitude of the causality illusion, we expect lower causality ratings in the Brush font condition
compared to the Arial font condition specifically within the null contingency scenario. Conversely,
no significant effect of font type is anticipated in the true contingency scenario.
As a technical clarification, it is important to note that the Brush font employed in our
Experiment 2 differs from the Mistral font utilized in the study conducted by Díaz-Lago and
Matute (2019a). While visually similar to Mistral, Brush offers the advantage of being more suitable
for online studies as it does not require preliminary downloading and installation on participants'
Fig.5: Uppercase Brush Script MT instruction in Experiment 2. This figure shows the experimental
instructions presented in Experiment 2. Text and background colours were maintained equal to
the HC conditions of Experiment 1 (see the upper panels in Fig. 1).
22
devices
6
. We maintained consistency with every aspect of the experimental procedure employed in
Experiment 1 to facilitate a direct comparison between the results of the two experiments.
The pre-registered hypotheses and methods are available on OSF at the following link:
7
https://osf.io/4tdcy
. The codes for the experiment, the raw data and the script used for the main
analyses are available on OSF at the same link used for the previous experiment:
https://osf.io/c26qa
.
3.1 PARTICIPANTS
The sample size was determined according to the same criteria of Experiment 1. As we had tested
100 participants in the easy-to-read Arial font condition of Experiment 1 (60 participants in the null
contingency scenario and 40 participants in true contingency scenario), in Experiment 2 we
recruited a new sample of 100 participants, randomly assigned either to a null contingency scenario
(60 participants) or to a true contingency scenario (40 participants).
Participants (native Italian speakers) were recruited through the online platform Prolific
(
https://www.prolific.co/
)
8
. The new sample was constituted of 60 females and 40 males. The
average age of participants was 32.19 years (SD=10.67). No data were excluded from the analyses,
according to the criteria already applied in Experiment 1 (see Section 2.1). A Pearson’s chi-square
test of independence showed that the frequencies of the two sexes did not differ significantly across
the four groups in the new experiment design, χ2(3) = 6.78, p = 0.08. Instead, a Kruskal-Wallis test
revealed the presence of statistically significant differences in age between the four groups, χ2(3) =
6 To retrieve more information about this font, the interested reader might consider consulting the w3schools web page
(https://www.w3schools.com/css//css_font_websafe.asp). The Brush Script MT font has been designed to mimic
handwriting, a feature that makes it similar to the Mistral font. Furthermore, it has already been used as a hard-to-read
font for inducting a cognitive bias reduction (e.g., Song & Schwarz, 2008). We decided through empirical observations
to use an Uppercase variant of the font, in order to make sure that we increased the reading difficulty on the online task
7 The hypotheses and the analysis plan were pre-registered on Open Science Framework on June 20, 2023
8 It must be noticed that the sample in this second experiment differs among some demographic features with respect to
the original study by Díaz-Lago and Matute (2019a): in the original study participants were undergraduate students that
conducted the experiment in person in a computer room.
23
42.27, p < 0.001, due to the fact that the mean age of the new sample was higher than that of the
participants in the HC conditions of Experiment 1 (M = 25.21 years, SD=8.21).
Regarding these demographic differences, prior research on the illusion of causality has
consistently demonstrated its manifestation regardless of confounding variables such as intelligence
and personality (Matute et al., 2022). No existing literature claims a consistent demographic
distinction in the occurrence of the causality illusion. To empirically support this statement, we
combined the data of the 300 participants from both experiments and conducted a Bayes Factor
general linear model analysis utilizing the BayesFactor R package (Morey & Rouder, 2022), to
examine the potential effects of age and sex on causality judgments. The results revealed that the
null model had at least 7.37 times greater support compared to alternative models that included age,
sex, or both as predictors
9
. The analysis script and detailed results can be found on OSF at the
following link:
https://osf.io/c26qa
. Consequently, we can confidently assert that the comparison
between the hard-to-read font condition and the easy-to-read font condition remains reliable
despite the observed demographic differences.
3.2 STIMULI AND PROCEDURE
Everything remained consistent with Experiment 1, except for the following specific modifications.
Instructions, patient records, and causality rating were presented in a blue hard-to-read font (font:
Uppercase Brush Script MT, color: HEX #000063, 17.79 contrast ratio, 0.03 height).
Upon agreeing to take part in the study through Prolific, participants were directed to
initiate the experiment on the Pavlovia website. Prior to commencing the actual study, participants
were presented with the approved online informed consent form, authorized by the Ethics
9 We included sex as a potential predictor because the new sample exhibited a more balanced distribution of sexes (60
females and 40 males) compared with the HC conditions of Experiment 1 (70 females and 30 males), although the p-
value for the chi-square test was slightly above the threshold of statistical significance.
24
Committee for Psychological Research at the University of Padova (protocol number 5010,
November 3rd, 2022). Once participants provided their consent to participate by simply pressing a
key, the experiment commenced.
3.3 RESULTS
3.3.1 PROCESSING FLUENCY
As for the subjective measure, the results from the single-item question on reading difficulty are
represented in Fig. 6. A two-way between-subject ANOVA with factors contingency (Null vs.
True) and font (Arial vs. Brush) showed a significant main effect of font, F(1,196) = 99.84, p <.001,
ηp2 =.34. Consistent with expectations, reading difficulty was rated higher with the Brush font (M
= 4.48, SD = 2.07) than with the Arial font (M = 1.96, SD = 1.46). The main effect of contingency
and the two-way interaction were not statistically significant [F(1,196) = 0.03, p =.85, ηp2 = .0002;
F(1,196) = 2.95, p = .08, ηp2 = .014], suggesting that contingency had no effect on perceived reading
difficulty (M = 3.2, SD = 2.14 in the null contingency condition and M = 3.25, SD = 2.27 in the true
contingency condition).
25
As for the objective measure (Fig. 7), experiment time was analysed in the same manner as the rated
reading difficulty. The main effect of font was statistically significant F(1,196) = 4.97, p = .03, ηp2 =
.02, due to longer experiment time in the Brush condition (M = 376.81 sec, SD = 122.37 sec) than in
the Arial condition (M = 341.03 sec, SD = 106.26 sec). The main effect of contingency and the
interaction were not statistically significant [F(1,196) = 2.13, p = .15, ηp2 = .01; F(1,196) = 3.86, p =
.051, ηp2 = .02], suggesting that contingency had no effect on experiment time (M = 344.58 sec, SD
= 97.23 sec in the true contingency condition and M = 368.47 sec, SD = 126.03 sec in the null
contingency condition). It is worth noting that the outlier shown in the left panel of Fig. 7
Fig. 6: Subjective evaluation of reading difficulty in Experiment 2. Histograms and boxplots
showing that the hard-to-read font conditions (orange) resulted in a greater subjective
difficulty evaluation compared with the easy-to-read font conditions (green). The green bars
and boxplots represent the same data as the blue bars and boxplots in Fig. 2.
26
(experiment time longer than 900 sec) was associated with an acceptable Cook’s distance (0.11), and
the main results did not change even after removing that outlier from the dataset
10
.
3.3.2 CAUSALITY RATINGS
The causality ratings are represented in Fig. 8. A two-way between-subject ANOVA with factors
contingency (Null vs. True) and font (Arial vs. Brush) showed a statistically significant main effect
of contingency, F(1,196) = 30.11, p < .001, ηp2 =.13. As expected, the causality ratings were larger in
the true contingency condition (M = 72.74, SD = 16.72) than in the null contingency condition (M
10As in the previous experiment, the same analysis was also performed on log-transformed experiment time, to reduce
the impact of skewed data (see Fig. 7). This analysis confirmed our previous findings.
Fig. 7: Experiment time in Experiment 2. Boxplots showing how the Brush, hard-to-read font
conditions (orange) resulted in a longer experiment time compared with the Arial, easy-to-read
font conditions (green). The green boxplots and violin plots represent the same data as the blue
boxplots and violin plots in Fig. 3.
27
= 58.13, SD = 19.44). The main effect of font was not statistically significant, F(1,196) = 0.25, p = .61,
ηp 2 =.001, as the means of the causality ratings with the hard-to-read and the easy-to-read fonts were
almost identical to each other (Brush: M = 64.63, SD = 21.14; Arial: M = 63.32, SD= 18.25). The two-
way interaction was not statistically significant F(1,196) = 1.08, p = .30, ηp 2 = .005, which is at odds
with the hypothesis that perceptual disfluency induced by a hard-to-read font can lead to a decrease
in the illusion of causality.
Fig. 8: Causality ratings in Experiment 2. The raincloud plots show the judged strength of
causality for the four experimental groups. The red dots and lines on the charts indicate the
means in the experimental groups. The green elements represent the same data as the blue
elements in Fig. 4.
28
In line with the pre-registered analysis plan, we also conducted a one-tailed independent sample t-
test and a one-sided Bayesian t-test to test if, in the null contingency condition, the causality ratings
in the disfluency condition were larger than the causality ratings in the fluency condition. The
results of the t-test showed that the causality ratings associated with the disfluent Brush font (M =
57.68, SD = 20.48) were not significantly smaller than those associated with the fluent Arial font
(M = 58.58, SD = 18.51), t(118) = 0.25, p =.40, d = .05. The Bayesian t-test showed that the observed
data were over four times more likely under the null hypothesis than under the alternative
hypothesis, BF01= 4.203, thereby confirming the lack of an effect of font type on the magnitude of
the illusion.
29
4. MODELS COMPARISON
After merging the data from the two experiments (N = 300), we conducted further analyses to
investigate the potential impact of (dis)fluency on causality ratings. Fig. 9 provides a visual
representation of the effects of disfluency, obtained through contrast or font manipulations, on
causality ratings.
Fig. 9: Relationship between fluency, contingency, and rated causality across the two
experiments. The dot plot shows the judged strength of causality for the six experimental
groups. Each dot represents the mean of the corresponding group. Each line represents the
standard deviation of the corresponding group. The low-contrast boxes show the overall means
(central line) and standard deviations (ends of the boxes) for the Null contingency and the True
contingency conditions.
30
We constructed four linear models (see Tab. C) to assess the relationship between judged strength
of causality, contingency, and fluency. The dependent variable was rated causality, and the
predictors were as follows: (1) contingency, (2) fluency, (3) contingency and fluency without
interaction, and (4) contingency and fluency with interaction. Note that fluency was treated as a
categorical predictor with three levels (see Fig. 9). To determine the strength of evidence supporting
each model, we computed the Bayes Factor (BF10) using a Markov Chain Monte Carlo (MCMC)
procedure via the BayesFactor package (Morey & Rouder, 2022). Notably, Model 1, which
excluded fluency as a predictor, exhibited the highest explanatory power. This model was found to
be 19 to 3×105 times more likely than the models that included fluency as a predictor. In essence, the
inclusion of the fluency predictor alongside the contingency predictor weakened the evidence
against the null model. Furthermore, to validate these findings, we calculated Akaike weights
(Akaike Information Criterion) and Bayesian weights (Bayesian Information Criterion) using the
MuMIn package (Bartoń, 2023), which confirmed the consistent pattern of results favouring
Model 1 (Tab. C).
Predictor(s)
45
!"
AIC
Weight
BIC
Weight
Model 1
Contingency
156163.60
0.69
0.99
Model 2
Fluency
0.05
0.17
<0.01
Model 3
Contingency +
Fluency
7815.36
0.14
<0.01
Model 4
Contingency *
Fluency
3690.19
<0.01
<0.01
Tab. C: Models description and associated Bayesian indices.
31
5. GENERAL DISCUSSION
Through two different effective manipulations of perceptual fluency, namely text-background
contrast and font type, our present study found null results that support the idea that perceptual
disfluency does not modulate the illusion of causality, at least for those manipulations where a
larger magnitude in the disfluency construct is provoked (See Section 5, next paragraph). These
results can also support the general conclusion that the effects of processing fluency on cognition
are unclear (Meyer et al., 2015). It is important to emphasize that the samples size was defined
through an a-priori power analysis (see Section 2.1), therefore the null result cannot be attributed
to a lack of statistical power. Furthermore, the null effect cannot be attributed to an ineffective
experimental manipulation, because the subjective index of difficulty rating and the objective index
of experimental time indicated that low contrast and hard-to-read font were associated with lower
processing fluency than high contrast and easy-to-read font, respectively. Below are two possible
explanations for the null results observed in our study. These explanations are tentative and require
further research to confirm their validity.
Our first hypothesis posits that the manipulation of perceptual fluency has the potential to
engage a more deliberative and effortful kind of thinking (i.e., system two). However, we propose
that the relationship between processing fluency and the illusion of causality may follow a non-
linear U-shaped function, whereby moderate disfluency can improve performance by engaging
system two, but excessive disfluency may overload and saturate the system's capacities, thereby
masking or reducing the engagement of system two. If our hypothesis is correct, it suggests that the
level of disfluency induced by low contrast and hard-to-read font in the current study may already
be excessive for the capacities of system two.
Tentative support for this hypothesis comes from the comparison between the disfluency
measures in our two experiments and those in Díaz-Lago & Matute's (2019a) study. In that study,
fluency was assessed using three 7-point Likert scales that evaluated easiness of reading, fluency of
32
the task, and perceived duration. While no significant results emerged for perceived task duration,
font type significantly influenced the easiness of reading and the task fluency, with a mean
difference of approximately 1.4 points between the two font types on the easiness of reading scale
11
.
In our experiments, we found instead a mean difference of approximately 3 points (Experiment 1)
and 2.5 points (Experiment 2) in reading difficulty between the disfluent and the fluent conditions.
These observations appear to align with the tenets of the presented hypothesis. In fact, in
accordance with the hypothesis, Díaz-Lago and Matute (2019a) may have encountered a mitigated
causality bias due to the introduction of low to moderate levels of disfluency within their
manipulations. Conversely, our experiments failed to replicate such bias attenuation, possibly due
to the imposition of higher levels of disfluency. While caution must be exercised when making this
comparison, it is worth noting that the use of a 7-point Likert scale in both studies facilitates the
comparison itself.
Beyond these qualitative observations, a more rigorous test of the hypothesis of a U-shaped
relationship between processing fluency and the illusion of causality is made possible by the fact
that materials and raw data from the study by Díaz-Lago and Matute (2019a) are available to consult
and download on OSF (
https://osf.io/vrukz/
). We merged data from our experiments with data
from the original font study, focusing on the relationship between the causality ratings and the
perceived difficulty of the task
12
in null contingency conditions (see Fig. 10). We constructed two
statistical models, one characterized by perceived difficulty of the task as a linear predictor and the
other characterized by perceived difficulty of the task as a quadratic predictor. To determine the
strength of the evidence supporting each model, we computed the Bayes Factor using an MCMC
11 In Díaz-Lago & Matute's (2019a) study, the easiness of reading was found to be higher for the easy-to-read font (M =
5.72, SD = 1.11 in the null contingency condition; M = 5.62, SD = 1.12 in the true contingency condition) with respect to
the hard-to-read font (M = 4.35, SD = 1.20 in the null contingency condition; M = 4.28, SD = 1.08 in the true contingency
condition).
12 The “easiness of reading” item (7-point Likert scale) from the study by Díaz-Lago and Matute (2019a) has been
inverted in order to express the perceived difficulty of the task instead of the perceived easiness. This makes it directly
comparable with the ratings from our two experiments.
33
procedure via the BayesFactor package (Morey & Rouder, 2022). The results showed that the null
model was 9.66 times more likely than the linear model and 4.33 times more likely than the
quadratic model. These results run against the hypothesis of a linear or non-linear relationship
between perceived difficulty and judged strength of causality.
Fig.4: Causality ratings in Experiment 1. The raincloud plots summarize the main results of the
experiment, showing the judged strength of causality for the four experimental groups. The red
dots and lines on the charts indicate the means in the experimental groups. Fig. 10: Relationship
between perceived difficulty and causality ratings in null contingency scenarios (combined data
from our two experiments and Díaz-Lago & Matute's (2019a) experiment). The line represents the
best fitting model, which corresponds to the null model (no relationship between difficulty and
rated causality).
34
It is important to emphasize that the experiments under consideration did not specifically target
the examination of a U-shaped correlation between task difficulty perception and causality ratings
and the original experiment by Díaz-Lago & Matute (2019a) provided a smaller sample (63
participants) compared to samples numerosity from our two experiments (180 participants).
Therefore a small partition of our merged data has been exposed to a less disfluent condition
compared to the large and complementary subset. Consequently, it is prudent to interpret these
results with circumspection. It is recommended that future studies adopt a systematic and
quantitative approach, which could facilitate the development of a mathematical description of the
potential non-linear relationship between perceptual fluency and the magnitude of the illusion.
Our second hypothesis builds upon the absence of any discernible systematic connection
between perceptual fluency and causality bias, as evidenced by the outcomes of our two
experiments. The results from our experiments and Díaz-Lago and Matute's (2019a) study can be
interpreted as a clash of evidence, in which we can apply a standard NHST type of reasoning: in
our study, the probability for at least one of two critical statistical comparisons to correctly reject
the null point hypotheses and so detecting the significative presence of an effect if it was truly
present, given a fixed hypothesized effect of .58 Cohen’s d, was >99%. On the other hand, the results
obtained by Díaz-Lago and Matute (2019a) could indicate a false positive with a standard alpha
probability of 5%, making our study results more likely from a statistical point. This hypothesis also
takes into account the foreign language effect (FLE) on the causality illusion (Díaz-Lago & Matute,
2019b). Although the literature on the FLE is inconclusive in explaining this phenomenon (Circi et
al., 2021), many functional bilingual models are centred on cognitive control mechanisms
(Schwieter & Ferreira, 2016). We hypothesize that not all types of processing disfluency are equally
effective in activating the engagement of system two. It is possible that the activation of system two
is related to high-order lexical and semantic processes that are involved in the processing of a
disfluent foreign language, but remain unaffected by the manipulation of superficial perceptual
35
features of the task, such as contrast or font type. This may explain why presenting the CLT in a
disfluent foreign language reduces the causality bias while presenting the CLT in a low-contrast
hard-to-read version does not. This second hypothesis can be considered antithetical with respect
to the first hypothesis and to the results obtained by Díaz-Lago and Matute (2019a), because it
suggests that a manipulation of the perceptual features of the information cannot lead to a reduced
magnitude of the illusion of causality.
Another possible explanation for our results can be inevitably attributed to the different
composition of the statistical samples: (i) Díaz-Lago and Matute’s (2019a) experiment has been
presented in a native Spanish language, whereas ours were presented in a native Italian language
and (ii) our experiment has been conducted remotely on an online-recruited sample, whereas Díaz-
Lago and Matute’s (2019a) experiment were conducted in person on a computer room on an
undergraduate-student sample. Even though it seems quite improbable that such ancillary variables
could provide a comprehensive explanation for our results, since we applied strict measures for
analysis inclusion (see Section 2.1) and control of the conduction of the experiment (see Section
2.2), these specific differences have to be seriously acknowledged and considered as potential sources
of data variability.
To conclude, fluency functions as a versatile overarching construct, given its role in gauging
the perceived difficulty of a task, a facet intricately linked with the load on different cognitive
processes. Nevertheless, the outcomes derived from our two experiments cast uncertainty upon the
all-encompassing explanatory prowess of the fluency construct. These findings prompt a
compelling need for a more profound and nuanced exploration of the intricate cognitive processes
that underlie the modulation of cognitive biases. Aligned with this perspective, there emerges a
distinct call for further investigations, aimed at establishing with heightened precision the specific
manipulations of processing fluency capable of engendering a reduction in the magnitude of the
illusion of causality.
36
ACKNOWLEDGMENTS
We thank all the people who helped us recruit participants for the experiment through the
sharing of the recruitment form.
CONFLICT OF INTEREST
The Authors declare that there is no conflict of interest to disclose.
37
REFERENCES
Allan, L. G. (1980). A note on measurement of contingency between two binary variables in
judgment tasks. Bulletin of the Psychonomic Society, 15(3), 147-149.
https://doi.org/10.3758/BF03334492
Allan, L. G., & Jenkins, H. M. (1980). The judgment of contingency and the nature of the response
alternatives. Canadian Journal of Psychology/Revue canadienne de psychologie, 34(1), 1.
https://doi.org/10.1037/ h0081013
Allan, L. G., & Jenkins, H. M. (1983). The effect of representations of binary variables on judgment
of influence. Learning and Motivation, 14(4), 381-405. https://doi.org/10.1016/0023-
9690(83)90024- 3
Alloy, L. B., & Abramson, L. Y. (1979). Judgment of contingency in depressed and nondepressed
students: Sadder but wiser? Journal of experimental psychology: General, 108(4), 441.
https://doi.org/10.1037/0096-3445.108.4.441
Alter, A. L., & Oppenheimer, D. M. (2009). Uniting the tribes of fluency to form a metacognitive
nation. Personality and social psychology review, 13(3), 219-235.
https://doi.org/10.1177/1088868309341564
Arkes, H. R. (1981). Impediments to accurate clinical judgment and possible ways to minimize their
impact. Journal of consulting and clinical psychology, 49(3), 323.
https://doi.org/10.1037/0022-006X.49.3.323
Barberia, I., Blanco, F., Cubillas, C. P., & Matute, H. (2013). Implementation and assessment of an
intervention to debias adolescents against causal illusions. PLoS One, 8(8).
https://doi.org/10.1371/journal.pone.0071303
Barberia, I., Tubau, E., Matute, H., & Rodríguez-Ferreiro, J. (2018). A short educational
intervention diminishes causal illusions and specific paranormal beliefs in undergraduates.
PLoS One, 13(1). https://doi.org/10.1371/journal.pone.0191907
38
Bartoń, K. (2023). MuMIn: Multi-Model Inference. R package version 1.47.5, https://CRAN.R-
project.org/package=MuMIn
Bender, A. (2020). What is causal cognition? Frontiers in psychology, 11, 3.
https://doi.org/10.3389/fpsyg.2020.00003
Benson, B. (2016). Cognitive Bias Cheat Sheet. Retrieved from
https://betterhumans.pub/cognitive-bias-cheat-sheet-55a472476b18
Blaisdell, A. P., Sawa, K. L., & Waldmann, M. R. (2006). Causal reasoning in rats. Science, 311(5763),
1020-1022. https://doi.org/10.1126/science.1121872
Blanco, F. (2017). Positive and negative implications of the causal illusion. Consciousness and
Cognition, 50, 56-68. https://doi.org/10.1016/j.concog.2016.08.012
Blanco, F., Barberia, I., & Matute, H. (2014). The lack of side effects of an ineffective treatment
facilitates the development of a belief in its effectiveness. PloS one, 9(1), 84.
https://doi.org/10.1371/journal.pone.0084084
Blanco, F., Matute, H., & Vadillo, M. A. (2013). Interactive effects of the probability of the cue and
the probability of the outcome on the overestimation of null contingency. Learning &
Behavior, 41, 333-340. https://doi.org/10.3758/s13420-013-0108-8
Caverni, J. P., Fabre, J. M., & Gonzalez, M. (1990). Cognitive biases. North-Holland. ISBN-13:
9780080867229
Champely, S. (2020). pwr: Basic Functions for Power Analysis. Retrieved from https://CRAN.R-
project.org/package=pwr
Circi, R., Gatti, D., & Russo, V. (2021). The foreign language effect on decision-making: A meta-
analysis. Psychon Bull Rev, 28, 1131–1141. https://doi.org/10.3758/s13423-020-01871-z
Costa, A., Foucart, A., Arnon, I., Aparici, M., & Apesteguia, J. (2014). “Piensa” twice: On the
foreign language effect in decision making. Cognition, 130(2), 236-254.
https://doi.org/10.1016/j.cognition.2013.11.010
39
Crocker, J. (1981). Judgment of covariation by social perceivers. Psychological Bulletin, 90(2), 272.
https://doi.org/10.1037/0033-2909.90.2.272
Díaz-Lago, M., & Matute, H. (2019a). A hard to read font reduces the causality bias. Judgment and
Decision Making, 14(5), 547-554. https://doi.org/10.1017/S1930297500004848
Díaz-Lago, M., & Matute, H. (2019b). Thinking in a Foreign language reduces the causality bias.
Quarterly Journal of Experimental Psychology, 72(1), 41-51.
https://doi.org/10.1177/1747021818755326
Dickinson, A., Shanks, D., & Evenden, J. (1984). Judgement of act-outcome contingency: The role
of selective attribution. The Quarterly Journal of Experimental Psychology, 36(1), 29-50.
https://doi.org/10.1080/14640748408401502
Gigerenzer, G. (2008). Why heuristics work. Perspectives on psychological science, 3(1), 20-29.
https://doi.org/10.1111/j.1745-6916.2008.00058.x
Graf, L. K., Mayer, S., & Landwehr, J. R. (2018). Measuring processing fluency: One versus five
items. Journal of Consumer Psychology, 28(3), 393-411. https://doi.org/10.1002/jcpy.1021
Griffiths, O., Shehabi, N., Murphy, R. A., & Le Pelley, M. E. (2019). Superstition predicts the
perception of illusory control. British Journal of Psychology, 110(3), 499-518.
https://doi.org/10.1111/bjop.12344
Hall, R., & Hanna, P. (2004). The impact of web page text-background colour combinations on
readability, retention, aesthetics and behavioural intention. Behaviour & information
technology, 23(3), 183-195. https://doi.org/10.1080/01449290410001669932
JASP-Team. (2023). JASP (Version 0.17). Retrieved from https://jasp-stats.org/
Jenkins, H. M., & Ward, W. C. (1965). Judgment of contingency between responses and outcomes.
Psychological monographs: General and applied, 79(1), 1.
https://doi.org/10.1037/h0093874
40
Kahneman, D., & Frederick, S. (2002). Representativeness revisited: Attribute substitution in
intuitive judgment. Heuristics and biases. The psychology of intuitive judgment, 49(49-81),
74. Retrieved from http://bear.warrington.ufl.edu/brenner/mar7588/papers/kahneman-
frederick-2002.pdf
Korn, C. W., Ries J., Schalk L., Oganian, Y., & Saalbach, H. (2018). A hard-to-read font reduces the
framing effect in a large sample. Psychonomic bulletin & review, 25, 696-703.
https://doi.org/10.3758/s13423-017-1395-4
Lilienfeld, S. O., Ammirati, R., & Landfield, K. (2009). Giving debiasing away: Can psychological
research on correcting cognitive errors promote human welfare? Perspectives on
psychological science, 4(4), 390-398. https://doi.org/10.1111/j.1745-6924.2009.01144.x
Matute, H., Blanco, F., & Díaz-Lago, M. (2019). Learning mechanisms underlying accurate and
biased contingency judgments. Journal of Experimental Psychology: Animal Learning and
Cognition, 45(4), 373. https://doi.org/10.1037/xan0000222
Matute, H., Blanco, F., & Moreno-Fernández, M. M. (2022). Causality bias. Cognitive Illusions:
Intriguing Phenomena in Thinking, Judgment, and Memory, 108-123. ISBN-13:
9780367724245
Matute, H., Blanco, F., Yarritu, I., Díaz-Lago, M., & Vadillo, M. A. (2015). Illusions of causality:
how they bias our everyday thinking and how they could be reduced. Frontiers in
psychology, 6, 888. https://doi.org/10.3389/fpsyg.2015.00888
Matute, H., Yarritu, I., & Vadillo, M. A. (2011). Illusions of causality at the heart of pseudoscience.
British Journal of Psychology, 102(3), 392-405. https://doi.org/10.1348/000712610X532210
McFarlane, S., Cipolletti, P. H., & Weissglass, C. (2020). Thinking in a Non-native Language: A
New Nudge? . Front. Psychol., 11. https://doi.org/10.3389/fpsyg.2020.549083
41
Meyer, A., Frederick, S., Burnham, T. C., Guevara Pinto, J. D., Boyer, T. W., Ball, L. J., & Schuldt,
J. P. (2015). Disfluent fonts don’t help people solve math problems. Journal of Experimental
Psychology: General, 144(2), 16. https://doi.org/10.1037/xge0000049
Michotte, A. (1963). The Perception of Causality. Taylor & Francis Ltd (2017, originally published
in 1963). https://doi.org/10.4324/9781315519050
Morys-Carter, W. (2023). VESPR study portal. Retrieved from
https://moryscarter.com/vespr/portal
Morey, R., Rouder, J. (2022). BayesFactor: Computation of Bayes Factors for Common Designs.
R package version 0.9.12-4.4. https://CRAN.R-project.org/package=BayesFactor
Oppenheimer, D. M. (2008). The secret life of fluency. Trends in cognitive sciences, 12(6), 237-241.
https://doi.org/10.1016/j.tics.2008.02.014
Peirce, L., Gray, J., Simpson, S., MacAskill, M., H"ochenberger, K., Sogo, H., Lindelov, J. (2019).
PsychoPy2: Experiments in behavior made easy. Behavior Research Methods, 51(1), 195–203.
https://doi.org/10.3758/s13428-018-01193-y
Perales, J. C., & Shanks, D. (2007). Models of covariation-based causal judgment: A review and
synthesis. Psychonomic bulletin & review 14, 577-596. https://doi.org/10.3758/BF03196807
Pohl, R. (2022). Cognitive Illusions: Intriguing Phenomena in Thinking, Judgment and Memory.
Routledge. ISBN-13: 9780367724245
R-Core-Team. (2023). R: A language and environment for statistical computing. Retrieved from
https://www.R-project.org/
Reber, R., & Schwarz, N. (1999). Effects of perceptual fluency on judgments of truth.
Consciousness and cognition, 8(3), 338-342. https://doi.org/10.1006/ccog.1999.0386
Reber, R., Winkielman, P., & Schwarz, N. (1998). Effects of perceptual fluency on affective
judgments. Psychological science, 45-48. https://doi.org/10.1111/1467-9280.00008
42
Roediger III, H. L. (1996). Memory illusions. Journal of memory and Language 35(2), 75-100.
https://doi.org/10.1006/jmla.1996.0005
Schwarz, N. (2004). Metacognitive experiences in consumer judgment and decision making.
Journal of consumer psychology, 14(4), 332-348.
https://doi.org/10.1207/s15327663jcp1404_2
Schwieter, J., & Ferreira, A. (2016). Effects of cognitive control, lexical robustness, and frequency
of codeswitching on language switching. Retrieved from
http://digital.casalini.it/9789027266729
Shanks, D. R., Lopez, F. J., & Dickinson, A. (1996). Distinguishing associative and probabilistic
contrast theories of human contingency judgment. Psychology of learning and motivation
(Vol. 34), 265-311. https://doi.org/10.1016/S0079-7421(08)60563-0
Song, H., & Schwarz, N. (2008). Fluency and the detection of misleading questions: Low
processing fluency attenuates the Moses illusion. Social cognition, 26(6), 791-799.
https://doi.org/10.1521/soco.2008.26.6.791
Thompson, D. V., & Ince, E. C. (2013). When disfluency signals competence: The effect of
processing difficulty on perceptions of service agents. Journal of Marketing Research, 50(2),
228-240. https://doi.org/10.1509/jmr.11.0340
Tversky, A., & Kahneman, D. (1974). Judgment under Uncertainty: Heuristics and Biases: Biases
in judgments reveal some heuristics of thinking under uncertainty. Science, 185(4157), 1124-
1131. https://doi.org/10.1126/science.185.4157.1124
Tversky, A., & Kahneman, D. (1996). On the reality of cognitive illusions. Psychological Review,
103(3), 582-591. https://doi.org/10.1037/0033-295X.103.3.582
Vadillo, M. A., & Matute, H. (2007). Predictions and causal estimations are not supported by the
same associative structure. Quarterly Journal of Experimental Psychology, 60(3), 433-447.
https://doi.org/10.1080/17470210601002520
43
Waldmann, M. R., Hagmayer, Y., & Blaisdell, A. P. (2006). Beyond the information given: Causal
models in learning and reasoning. Current Directions in Psychological Science, 15(6), 307-
311. https://doi.org/10.1111/j.1467-8721.2006.00458.x
Wasserman, E. A. (1990). Detecting response-outcome relations: Toward an understanding of the
causal texture of the environment. Psychology of learning and motivation (Vol. 26), 27-82.
https://doi.org/10.1016/S0079-7421(08)60051-1
Wasserman, E. A., Dorner, W. W., & Kao, S. F. (1990). Contributions of specific cell information
to judgments of interevent contingency. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 16(3), 509-521. https://doi.org/10.1037/0278-7393.16.3.509
Wickham, H. (2016). ggplot2: Elegant Graphics for Data Analysis. New York: Springer-Verlag.
Retrieved from https://ggplot2.tidyverse.org
Zuffi, S., Brambilla, C., Beretta, G., & Scala, P. (2007). Human computer interaction: Legibility
and contrast. 14th International Conference on Image Analysis and Processing (ICIAP
2007), 241-246. https://doi.org/10.1109/ICIAP.2007.4362786
