=Paper=
{{Paper
|id=Vol-1751/AICS_2016_paper_14
|storemode=property
|title=Illusory Correlations: An Investigation in a Social Media Setting
|pdfUrl=https://ceur-ws.org/Vol-1751/AICS_2016_paper_14.pdf
|volume=Vol-1751
|authors=Abeba Birhane,Roisin Fitzpatrick,Caoimhe Kavanagh,John Rogers
|dblpUrl=https://dblp.org/rec/conf/aics/BirhaneFKR16
}}
==Illusory Correlations: An Investigation in a Social Media Setting==
https://ceur-ws.org/Vol-1751/AICS_2016_paper_14.pdf
Illusory Correlations: An Investigation in a Social Media
Setting
Abeba Birhane, Róisín Fitzpatrick, Caoimhe Kavanagh, John Rogers
Computer Science and Informatics, University College Dublin, Ireland
abeba.birhane@ucdconnect.ie, roz2030@hotmail.com, Caoimhe.Kavanagh@ucdconnect.ie,
John.Rogers.1@ucdconnect.ie
Abstract. This experiment investigated whether illusory correlations (ICs) exist
within social media settings, and the likelihood of the occurrence of marked dif-
ferences between individuals in their IC rates. Eighty individuals participated in
the study, with 71 used for analysis. The experiment consisted of three tasks: a
learning phase and two test phases, A and B. Participants were first presented
with Facebook statuses from two groups, Purple or Orange, followed by ques-
tions on various aspects in phases A and B. Statistical analysis of the results
showed no significant difference between the groups, however further analysis
of subdivided groups showed higher IC rates in some groups. Confidence levels
were found to correlate with the emergence of IC. Additionally, the results
demonstrated the emergence of a ‘reverse’ IC that has been previously unre-
ported in the literature. Qualitative analysis reflected a significant overall IC be-
tween the groups.
1 Introduction
Social discrimination or stereotyping has substantial and far reaching implications
for society, and understanding the cognitive aspects that give rise to this phenomenon
is crucial. In any society we would expect desirable and pro-social behaviours to be
the norm, whereas behaviours on the opposite ends of those spectrums the exceptions.
A problem arises when we realise that individuals, without evidence to the contrary,
have a cognitive-based tendency [4] to attribute more frequently occurring behaviours
with majority groups and less frequent behaviours with minority groups. This
tendency where an individual perceives that a statistical relationship exists between
two variables, where in fact none exists, is referred to as an illusory correlation (IC)
[2]. This phenomenon is exemplified by an instance in which a false association is
formed between generally uncommon negativebehaviours and a minority group –
contributing to the formation of stereotypes [7]. Traditional social cognition
approachs argues that we form stereotypes by means of our cognitive processes which
rely on strategies such as the availability heuristic [7]. The term illusory correlation,
first coined by Chapman in his 1967 landmark paper, refers to the systematic errors
which are commonly made in the estimated correlation between two classes of events
[2]. ICs were then first demonstrated empirically by Hamilton and Gifford in 1976,
along with the first attempts at modelling the phenomenon. Despite the fact that the
IC phenomenon has been shown to be robust [3], [4], [9], there has not been a
satisfactory explanation as to why it occurs or the cognitive mechanisms behind it.
Nonetheless, a number of recognised, yet conflicting theories have been offered as
explanations for this phenomenon.
The distinctiveness-based explanation (DBE), first proposed by [4] has been one of
the most dominant theories to account for ICs. According to this explanation, an
individual will estimate a higher rate of co-occurrence between two events when the
events in question are unusual, rarely occurring, or are otherwise salient. This model
proposes that the increased salience of the events in question leads to information
concerning them being memorised and then recalled more easily, with a higher
assumption of co-occurrence. Ethnic or social minorities are by definition less
frequently occurring in a population than the respective majorities, and negative or
anti-social behaviour is typically perceived as deviant and less prevalent than pro-
social behaviour. The DBE asserts that two distinct pieces of information become
implicitly linked during the encoding phase, resulting in the recall of either piece of
information involving the retrieval of the paired ‘unit’ of events [9]. However,
empirical evidence for distinctiveness based IC seems scarce. According to [6], there
is no direct evidence for the idea that we have enhanced memory as a result of biased
learning of infrequent pairings. Furthermore, the indirect evidence does not stand up
to scrutiny.
Memory for rare events was not enhanced, and if anything it was shown to
be impaired [6] according to their incomplete learning account (ILA). This account
proposes that the assumption of selective memory or bias is unnecessary. The ILA
instead argues that there is an increased number of learning instances for both the
majority group and the positive characteristics, which results in these pieces of
information being learnt more thoroughly. As a result the relative lack of information
concerning the minority group and negative characteristics prevent the observer from
learning the frequency of co-occurrence accurately, resulting in the over-estimations
known as Ics [6].
A connectionist approach which uses autoassociative networks that mimic learning
processes proposed by [10] proposes that observers construct a mental representation
which consists of connections between social groups, episodic behavioural
information, and an overall evaluative impression. Observers are considered to
gradually develop prototypes of both majority and minority groups that include not
only evaluative information, such as positive or negative feelings, but also store
specific behavioural information. With every encounter the prototype is updated via a
connectionist learning algorithm. In this way information concerning social groups
and their behaviours are encoded twice: first as an overall impression of the group,
and second as a specific episodic memory about the group or its members’ behaviour.
As the connections between the majority social group and overall impression are
strengthened, the connections between these events and the other areas are weakened:
those being the majority group and episodic memories, and the overall impression and
minority group. Accordingly, this shift of potency results in the connections between
the two smaller groups (minority event and episodic memory) being unaffected and
remaining intact. ICs are then formed as an unintentional by-product of the cognitive
mechanisms updating the mental representations [10].
On the other hand [3] have argued that IC is effective at all stages of information
processing and not a product or by-product of a single stage of cognitive functioning.
Following their examination of ICs between person types (a student versus a clerk)
and educational attitudes (liberal versus authoritarian) under conditions where
selective recall processes are highly unlikely, they argued that ICs are effective when
the stimulus is perceived and encoded; when information is constructed on cue recall
tests; when statistical properties are estimated; and when the impressions of the target
persons are expressed within an adjective space. They found that there was a marked
difference between individuals in their ICs and that the group effect was due to the
overly biased response of a dozen participants. This implies that ICs, even though
observed universally, may not be inherently dictated by the laws of cognitive
processing, but rather are a product of arbitrarily chosen or learned strategies.
The concept of using social media, Facebook (FB) specifically, as a medium to
investigate ICs is quite novel. In a recent study [6] have used Facebook-like personal
profiles to investigate the mechanisms behind ICs in the context of attention shift in
category learning. They proposed an inverse base-rate effect (IBRE) as a mechanism
by which IC arises when frequent categories are paired with corresponding common
attributes and infrequent categories with corresponding rare attributes. However, no
evidence was found for IBRE as an explanation of ICs: although FB was used in this
study, IC was not directly investigated.
1.1 The Current Experiment
In contemporary society, a large amount of the social information that an individual
gains originates from a social media platform. The aim of the current study was to
investigate whether ICs can also be found in individuals perceptions of FB statuses.
The formation of ICs on social media is of great importance and relevance as an
increasing amount of our social interaction takes place online [5]. Previous literature
has shown that the IC phenomenon has been demonstrated to be robust across a
variety of methods and in a variety of situations.
One of the main aims of the experiment was to investigate whether ICs will form
based on the evaluative information that is present in FB statuses. Given that ICs have
been implicated in a variety of social-cognitive settings (e.g. stereotyping), it is
appropriate that empirical investigation takes place in a social media setting.
Additionally, with the rising interest in social media among scientists, the current
experiment adds to a rapidly building body of literature in this area.
Finally, previous studies in ICs tend to use vignettes rather than information
presented in a more ecologically valid setting, and the current study aims to test the
strength of the phenomenon outside of the lab. Given previous research suggests that
generally, the majority group is judged more favourably compared to minority group,
the current experiment consists of the following hypotheses: 1) Desirability and
likeability scores for the majority group will be significantly higher than the minority
group’s scores for positive test items. 2) Negative test items will be disproportionately
assigned to the minority group. 3) Likeability scores for the neutral statuses will be
significantly higher when they are presented as made by the majority group than they
were originally rated. We also further investigated [3]‘s claim that there are marked
differences between individuals in their IC rates.
2 Method
2.1 Participants
Eighty participants took part and nine were excluded from analysis due to incomplete
responses, resulting in a final N = 71 (32 males and 39 females). Participants were
recruited by convenience and referral sampling, predominantly through FB. Since FB
status updates were used as a medium for inducing ICs, it was a necessary criterion
that all participants were active FB users. The majority of participants (54%) fell in
the age range 18-25, with a further 21% between 26 and 35 and the remaining 24%
being 36+. The majority of participants reported active FB usage, with 68% reporting
that they use the site daily, and only 8% reporting they use the site once or twice a
week. The remaining 24% used FB between two to five times per week.
2.2 Materials
2.2.1 Status Selection. The status updates used in this experiment were gathered from
FB Newsfeeds. These statuses were edited to remove identifiable information such as
names of people or places as well as colloquial language in order to illuminate already
formed biases associated with particular names or places. For example, ‘I’m loving
the Irish weather’ was edited as ‘I'm loving this weather’ and the name and image of
the person was replaced by a white FB template with Orange background (see Fig 1.).
In order to avoid experimenter bias, the statuses were rated independently by 76
participants in a precursor ratings task by means of distributing these statuses online
using Google Forms. Over 100 statuses were selected and distributed online through
FB. The perceived desirability of the author of each status was recorded on a Likert-
style five-point scale from extremely undesirable to extremely desirable. Desirability
was defined as how positive or negative an impression the reader was left with of the
author after reading their status. A neutral status was defined as one which made no
impact on the reader, and would be scrolled past without a second thought.
The average rating of each status was then used to divide the statuses into four
distinct groups: positive statuses were defined as those which fell between a score of
3.5 and 4.5, negative between 1.5 and 2.5, neutral between 2.5 and 3.5, and extreme
statuses fell below 1.5 or above 4.5. Extreme statuses were excluded from further use,
with the justification that if a status is regarded as too positive or negative, it becomes
inherently memorable and would be unsuitable for this study [10].
2.2.2 Stimuli. FB newsfeed templates which resembled the real newsfeed were
created and made anonymous by removing and censoring identifiable information,
such as name, profile picture, and FB group membership. The statuses were then
inserted into these templates and divided randomly between the majority and minority
group: each group had twice as many positive statuses as they had negative, with the
majority group having twice as many statuses overall as the minority group (20:10,
10:5 status ratio of positive: negative for each group). Consequently there were 45
total status images for the learning phase of this project. The two groups were
distinguished using differently coloured profile pictures, with orange representing the
majority group and purple representing the minority group (Fig 1). Participants were
unaware of any existing differences between the two groups. The aim of the coloured
profile pictures was to eliminate possible pre-existing biases that a participant may
have regarding names, sex, race, age, place of birth, schools attended, etc. However
researchers also wanted to use somewhat established group dividers as the use of
arbitrary group signifiers, such as ‘Group A and Group B’ has been seen to
exaggerate IC effects [1]. Using a coloured but anonymous profile picture as well as
individual but censored names for each status enables the researchers to control for
these social biases, whilst still demonstrating to participants that each status was
posted by a distinct person belonging to one of two relevant groups.
Fig. 1. Stimuli used to denote group membership to the Purple or Orange group, respectively.
The task consisted of three parts: the learning phase and two testing phases, hereafter
referred to as test phase A and B. In the learning phase statuses were presented in a
random order, featuring a group’s icon to denote authorship. Participants were re-
quired to read each status and get impressions of the statuses associated with each
group. In test phase A, a total number of 20 (ten positive and ten negative) statuses
were presented without group denotations. Participants then predicted the group
membership of that status’s author; rated their confidence in their group assignment
choice; their perceived desirability of the author upon reading the status; and finally
how likely they would have been to ‘like’ that status had it appeared on their own
Newsfeed. In test phase B participants were presented with ten statuses which were
previously evaluated as neutral and non-impactful. These statuses featured an Orange
(posted by a member of the majority group) profile picture. Participants again rated
their likelihood to have ‘liked’ that status if they had seen it on their own Newsfeed.
2.2.3 Software. The project was built and presented in Python code using the program
PsychoPy, [8]. This enabled the display of experimental statues to appear exactly like
real FB status updates while excluding some information like names that participants
might have already formed biases on. All of the FB statuses and instruction pages
were formatted in JPEG images. Participants performed the task on researchers’ lap-
tops. Prior to each phase participants were provided with clear instructions on the
task. Each instruction screen included a five second minimum delay before the partic-
ipant could move to the next screen by pressing the ‘Enter’ key: this prevented acci-
dental skipping. The program then displayed each of the learning phase statuses in a
randomised order, featuring a two second minimum delay before the participant could
progress, for the same reason. The ratings for the test phases were recorded on a Lik-
ert-type scale. In order to familiarise participants with the experiment, they were pre-
sented with a practice phase prior to the test phases. After selecting their desired val-
ue, they had the option to change the value or confirm it via button click. The group
assignment question was a two-point scale, while the other questions featured a five-
point scale.
2.2.4 Other materials. Participants were provided with an information sheet describ-
ing the nature of the study and their task. Relevant information, such as age, gender,
and the frequency of FB use were also recorded. We also had a question to be an-
swered only after the task was completed: whether participants had a particular im-
pression of the Purple and the Orange group upon reading such statuses.
2.3 Procedure
The task was conducted in residential environments. Participants were briefed and
instructed on how to complete the various phases of the study. They were given the
information sheet explaining what the study was concerned with, and what their in-
volvement entailed and signed a consent form prior to their engagement. The task of
the researcher was to initialise the experiment and give the participant space to com-
plete the task. Upon completion of the task, participants were debriefed with further
discussion of the subject of the experiment, and requested to fill out the demographic
sheet. They were then thanked for their contribution.
2.4 Data analysis
Data was organised according to the group assignment made by participants in test
phase A. As the statuses in test phase B were previously unseen, participants were
using prototypes of the Orange and Purple group formed in the learning phase to as-
sign groups to statuses The data was first analysed by examining the frequency of
group assignment made to each test item: as each participant (N = 71) engaged in 20
test items (10 +ve and 10 -ve) in test phase A, a total of 1391 instances, excluding 29
missing due to technical difficulty, were used for this analysis. Subsequently, partici-
pants were separated into three subgroups for further analysis, which will be elaborat-
ed on below. Finally, qualitative analysis was carried out.
3 Results
3.1 Statistical
The analysis’ point of departure was the desirability and likeability scores for each
test item. During the initial testing phase, participants rated 20 test items consisting
of equal numbers of positive and negative statuses (freq = 1390) as well as assigning
them to a group. Assignment to the orange group occurred 51% of the time overall,
47% of the time for negative statuses and 55% for positive statuses. A battery of t-
tests also showed that there were no significant differences in likability or desirability
between Orange+ and Purple+ groups, or Orange- and Purple- groups. Thus no evi-
dence was found for the first hypothesis.
Self-reported confidence measures were then controlled for two reasons. Firstly, as
there was no objective difference between groups in terms of their positive and nega-
tive statuses (except presentation ratios) it is impossible for a participant to say, with
any legitimate certainty, which group an unseen status, belongs to. It was expected
that individuals with low confidence scores were more likely to engage in random
guessing in their group assignments. Secondly, by taking into account individual dif-
ferences and how certain individuals may be particularly susceptible to IC, it was
deduced that individuals with higher confidence levels may demonstrate a higher
level of IC. High confidence responses were taken as those scoring either 4 or 5. An
independent samples t-test examined differences in desirability for Orange- and Pur-
ple-, (of the high confidence responses) and found that Purple- (M = 1.72, SD = .95)
scores were not significantly lower than Orange- (M =1.91, SD = 0.97) scores: t (290)
= 1.63, p = 0.052.
An independent samples t-test showed that likelihood scores were significantly
[t(296)= 2.49, p < 0.01] lower for Purple- (M = 1.39, SD = .92) than Orange- (M
=1.68, SD =1.07). It was found that individuals who are confident in their group as-
signments were less likely to ‘like’ the negative statuses of individuals belonging to
the minority group supporting the prediction that higher confidence is likely to be
associated with higher level of IC.
For low confidence individuals, significant differences [t(112) = -2.94, p < 0.01]
were found between likeability scores for Orange+ (M = 2.80, SD =1.45) and Purple+
(M = 3.44, SD = 1.47). Additionally low confidence individuals rated their likability
of assigned Purple- statuses (M = 1.72, SD = 1.06) as higher than those of Orange-
(M = 1.48, SD = 0.83); [t(188) = -1.66, p = 0.048). An analysis investigating the
number of instances where individuals favoured selecting one group over the other for
either positive or negative statuses was then conducted and is shown in table 1.
Table 1. Frequency of participants leaning towards one group
Individual Orange+ Orange- Purple+ Purple-
preference
70% 27 22 24 31
80% 23 15 12 20
Following this, the probability of whether these individuals would assign Pur-
ple/Orange to either a positive or negative status in test stage A, was calculated. Indi-
viduals who rated the neutral statuses as positive in test phase B were 10% more like-
ly to assign positive statuses to the Orange group in the test phase A. Likewise, indi-
viduals who voted the neutral statuses as negative were 10% more likely assign nega-
tive statuses to the Purple group.
When examining the complete dataset there were few statistically significant re-
sults. However upon closer inspection, it became clear that the results were not all
falling into a neutral range, but rather the majority of responses were on opposite ends
of the scale and were cancelling each other out (see Fig 2). After this initial analysis,
it became evident that participants fell into one of three subgroups: participants that
formed the IC as expected (Expected Illusory Correlation, hereafter ExpIC; n = 31),
participants that displayed no preference or aversion to either group (No Illusory Cor-
relation, hereafter NoIC; n = 16), and participants that formed the ‘Unexpected’ IC
(Opposite Illusory Correlation, hereafter OpIC; n = 24). ExpIC were participants that
assigned ≥70% of negative statuses to the Purple group, as hypothesised in this exper-
iment and previous research. OpIC assigned ≥ 70% of negative statuses to the Orange
group, which is a trend unprecedented in and unexpected by previous literature.
Fig 2. Group assignment frequency % of all test items
It appears that this subgroup formed a ‘reverse’ or ‘opposite’ IC, and this finding is
discussed in detail below. Finally NoIC indiscriminately chose either Orange or Pur-
ple for negative test items, indicating that they formed no IC. The participants in this
subgroup did not differ particularly from those in the other subgroups in terms of age,
sex or ethnicity. This supported [3]’s assertion that there are individual differences in
the susceptibility to IC formation and is discussed further below. An independent
samples t-test found that ExpIC (M = 3.12, SD = 1.3) scored the neutral statuses sig-
nificantly higher [t (414) = 2.43, p< 0.01] than OpIC (M = 2.83, SD = 1.25). NoIC did
not differ significantly for neutral scores indicating that individuals who formed the
Expected IC also demonstrated an exaggerated preference for statuses when they
believed them to be authored by the Orange group.
Table 2. Differences between ExpIC, OpIC and NoIC for desirability and likeability for all test items.
DesO+ LikO+
Var 1 Var 2 Mean Var1 Mean Var2 T Df sig Mean Var1 Mean Var2 T df Sig
ExpIC OpIC 3.70 3.47 1.61 68 0.06 3.36 2.70 3.22 63 0.01**
OpIC NoIC 3.47 3.28 1.09 112 0.14 2.70 3.02 -1.33 100 0.09
ExpIC NoIC 3.70** 3.28** 3.31 188 0.01** 3.36* 3.02* 2.12 186 0.02*
DesP+ LikP+
Var 1 Var 2 Mean Var1 Mean Var2 T Df sig Mean Var1 Mean Var2 T df Sig
ExpIC OpIC 3.26** 3.77* -2.94 73 0.01** 2.9** 3.56** -3.06 74 0.01**
OpIC NoIC 3.77* 3.47* 2.17 149 0.02* 3.56* 3.18* 2.31 170 0.02*
ExpIC NoIC 3.26 3.47 -1.03 112 0.20 2.87 3.18 -1.24 100 0.15
DesO- 2 D LikO- Column2
Var 1 Var 2 Mean Var1 Mean Var2 T Df sig Mean Var1 Mean Var2 T df Sig
ExpIC OpIC 2.05* 1.88* 0.89 52 0.2* 1.90 1.88 1.42 51 0.08
OpIC NoIC 1.88 1.97 -0.79 211 0.22 1.88** 1.52** 3.25 215 0.01**
ExpIC NoIC 2.05 1.97 t Stat 64 0.35 1.3* 1.52** 1.82 59 0.04*
DesP- LikP-
Var 1 Var 2 Mean Var1 Mean Var2 T Df sig Mean Var1 Mean Var2 T df Sig
ExpIC OpIC 1.80** 2.15** -2.89 157 0.01** 1.49* 1.71* -1.75 157 0.04*
OpIC NoIC 2.15** 1.71** 2.84 193 0.01** 1.71 1.71 0.75 150 0.23
ExpIC NoIC 1.80** 1.71** -2.89 157 0.01** 1.49 1.71 0.75 150 0.23
A series of t-tests was then conducted in order to compare the differences between the
three groups (Table 2.). The ExpIC group had significantly higher means for both
desirability and likeability for Orange+ test items than both OpIC and NoIC groups.
Additionally, the ExpIC group had a significantly higher mean than the NoIC group
for both desirability and likeability for Purple- items. However, there was no signifi-
cant difference between ExpIC and NoIC in relation to likeability scores for Purple-.
For the OpIC groups, the mean for both desirability and likeability for Purple+ items
were significantly higher than both NoIC and ExpIC groups. OpIC had a significantly
higher mean for likability, but not desirability, than both ExpIC and NoIC groups for
orange- test items. Contrary to previous findings, OpIC group rated positive purple
items as being both more likeable and desirable than the other two groups.
3.2 Qualitative Analysis
Participant’s qualitative responses, analysed in conjunction with their task responses,
showed support for IC and in line with previous findings. Of the fifty-one participants
that provided their characterisations of the groups, there was an overall tendency to
favour the majority group. In characterising the Orange group, positive/desirable at-
tributes were given 27 times, whereas undesirable/negative impressions were given 24
times. The Purple group however was ascribed with negative/undesirable characteris-
tics in 34 incidents while only attributed 19 positive/desirable characteristics. Several
participants characterised the Orange group as braggers, attention-seekers, over-
sharers, or hormonal posters.
One common theme among the majority of participants was their general negative
attitude towards FB, regardless of the experimental content or the group presented.
This was reflected by the greater number (58) of undesirable/negative comments
compared to (46) positive/desirable comments, made by participants when asked to
provide their impression of each group. The most prevalent distinguisher between the
Purple and Orange groups that was reported by participants was a believed difference
in age. Although statuses were randomly divided between the groups during, partici-
pants felt the Orange group was older and more mature than the Purple group.
4 Discussions
4.1 Present Findings
Even-though the initial hypotheses had to be rejected, analysis of the results found
considerable evidence to suggest that the ICs exist in social media and online interac-
tions. Analysis of subgroups showed that ICs were evident. For instance, individuals
who felt confident in their group assignments attributed not only a higher frequency
of negative statuses to Purple but also relative lower/higher likeability and desirability
scores for both the Purple- and Orange+ items. This suggests an IC had occurred.
Additionally, low confidence individuals rated Purple- items as higher and Orange+
as lower; providing a possible insight into why OpIC may have occurred. Perhaps the
core substance of the experiment comes from the separating of participants into three
distinct groups. We hypothesised that after analysis, two distinct groups would
emerge (ExpIC, NoIC). An unexpected and substantial OpIC group mirrored the Ex-
pIC group; nullifying their IC effect. Had our initial hypotheses been applied to each
of the three groups, rather than conflating these distinct subgroups, our first hypothe-
sis must be accepted for the ExpIC group; likewise, the opposite hypotheses would
have been true for the OpIC group. What is evident from the comparison of the ExpIC
and the NoIC group is that individual differences among participants can affect IC
which is in line with [3]’ findings that some individuals are more susceptible to ICs
than others.
The OpIC group was an unexpected finding, with no previous literature discover-
ing an opposite IC in an IC experiment. Confidence appeared to be a correlational
factor in that high confidence individuals were more likely to rate desirability higher
for Purple- and lower for Orange+, and vice versa. The qualitative data gathered by
members of this group seemed to elucidate some of the cause of this group’s appear-
ance: although the most substantial block of statuses shown was Orange+, the Or-
ange- block was measurably higher than the Purple-. IC literature is based on the evi-
dence that humans disregard this objective information and form erroneous opinions
concerning minority groups and behaviours; however this did not occur with the
OpIC group. This might suggests that specifically in a FB setting, quantity can matter
more than quality: participants felt that the Orange was the more negative of the
groups, even though this itself is still an IC.
Testing phase B also yielded significant results. ExpIC individuals rated the neutral
statuses as positive, whereas OpIC individuals tended to vote for the neutral statuses
as being more negative. Given the neutral statuses were presented as posted by the
Orange group, we can deduce that the formation of ICs, whichever direction it took,
may have influenced individuals perceived likeability of the statuses. IC research does
not often investigate this inverse trend: while it is typical to examine an aversion to
the minority group, most researchers do not consider if the participant also displays a
preference to the majority group. By asking participants to rate neutral statuses as-
signed to the majority group we found that this converse preference existed in the
expected direction for both ExpIC and OpIC, with the NoIC group continuing to
evaluate them as neutral. This was supported by the probability estimates discussed
previously. The potential implications of this inverse finding suggests that frequency
of exposure to a minority and majority group can affect approval ratings of not only
the minority group but also the majority group.
Taken together, the results suggest that individuals who were confident in their an-
swers were more likely to rate Purple- items worse than negative items they deemed
as belonging to the Orange group. Additionally individuals who were uncertain in
their answers were more likely to attribute higher likeability scores to Purple groups.
As the overall qualitative characterisation of the Purple group was more negative
when compared to the Orange group, the present analysis supports previous IC re-
search findings [4],[9].
Even-though there were more positive statuses posted by the Orange group, some
participants indicated that this led to a negative impression of the Orange group as a
whole, which contradicts previous findings. Several participants characterised the
Orange group as braggers, attention-seekers, over-sharers, or hormonal posters. This
highlights a key difference in the design of this experiment when compared to less
recent studies: whilst this information would usually have been conveyed in a vi-
gnette, the present study presented the same information as being posted directly from
the individual. Accordingly, that which is usually seen as a positive characteristic can
have negative connotations when self- presented in a FB setting.
4.2 Limitations and Future Research
This experiment was carried out mainly at residences limiting the variety within the
sample of participants. Most participants had connections to the researches either
through university or family and friends. Given that FB users is arguably one of the
largest populations worldwide, with over 1.49 billion users and regularly being ranked
in the top most trafficked sites on the internet, the population sample used in this
study is unlikely to represent the general FB population. This experiment could be
improved by conducting the study online: this affords the chance to gather data from
participants from a more varied demographic background, likely to be more repre-
sentative of the FB user population. Nonetheless, the findings from this research are
valid within the context in which it was carried out.
As the experiment was using status updates, i.e. statements with values attached, it
is inevitable that individual’s values differ, despite the fact that the statuses were rated
for a consensus regarding its inherent positivity or negativity. These statuses could
potentially mean different things to different participants and this was evident occa-
sionally during debriefing of participants. ‘Likeability’, for example, was subjective
and dependent on the kind of relationship one has with the person issuing status. A
FB user is likely to ‘Like’ a status such as ‘I can’t wait to go on my holiday’ if it was
posted by a close friend or family member; the likelihood for ‘Liking’ that same sta-
tus is much lower when it is posted by a person that they barely knew. It is recom-
mended that future research takes a note of this important aspect and incorporate it
when designing experiments. Similarly, complacency with or negative opinions to-
wards FB were reported by certain participants, which complicates the findings of this
research.
After the statuses were assigned as positive, negative, or neutral, they were divided
randomly between the Orange and Purple groups to the desired frequencies. For the
most part, excluding references to one group being more positive or negative than the
other, the characteristics assigned to the groups by participants were arbitrary and
reflected their general opinion and not the content of the statuses themselves. The
only somewhat common theme was a perceived difference in age between the Orange
and Purple groups. While this was not apparent in the statuses, perhaps there were age
cues which were implicit in the statuses that happened to divide between the groups
after random assignment. This again is an aspect for future research which may be
considered. During debriefing participants claimed that it was strange to see an Or-
ange member post such a “Purple-typical” status, indicating that matching statuses
may not have had a significant effect.
A few technical issues arose during this study. The slider range used for test phase
A and B featured five points, which could have been lengthened for increased sensi-
tivity. Furthermore, the demographic information gathered failed to capture any of
the individual differences which effected IC formation, or lack thereof, which oc-
curred in this experiment and is proposed by [3]. More detailed demographic infor-
mation may have shed light on the underlying grounds of these individual differences.
Previous research has shown that using arbitrary (e.g. Group A/Group B) instead of
real-world (e.g. Class of 1989/Class of 1992) groupings for IC research leads to exag-
gerated or heightened ICs [1]. This implies that when a participant is not given a dis-
tinguishing feature between two groups, they will search for one or perhaps even
create one. This is a confounding factor when trying to specifically examine ICs as
they form naturally. However, in using real names or profile pictures social biases are
introduced. The design of this experiment endeavoured to control for both extraneous
social biases and these categorisation processes by using coloured profile pictures and
censored names. However through debriefing participants it became clear that many
were still focused on trying to find a difference between the groups, rather than ac-
cepting that the groups were already distinguished. This could potentially have led to
exaggerated IC effects, as is seen in other papers. Future research may wish to seek
other means to differentiate between their groups while still controlling for social
biases.
References
1. Berndsen, M., McGarty, C., Pligt, J., & Spears, R. (2001). Meaning‐seeking in
the illusory correlation paradigm: The active role of participants in the categori-
zation process. British Journal of Social Psychology, 40(2), 209-233.
2. Chapman, L. J. (1967). Illusory Correlation in Observational Report. Journal of
Verbal Learning and Verbal Behaviour, 6(1), 151-155.
3. Fiedler, K., Hemmeter, U., & Hofmann, C. (1984). On the origin of illusory cor-
relations. European Journal of Social Psychology, 14(2), 191-201.
3. Fiedler, K. (2000). Illusory correlations: A simple associative algorithm provides
a convergent account of seemingly divergent paradigms. Review of General Psy-
chology, 4(1), 25-58.
4. Hamilton, D., & Gifford, R. (1976). Illusory correlation in interpersonal percep-
tion: A cognitive basis of stereotypic judgements. Journal of Experimental Social
Psychology, 12(4), 392-407.
5. Hunsinger, J., & Senft, T. M. (2013). The social media handbook. Routledge.
6. Kutzner, F. L., & Fiedler, K. (2015). No correlation, no evidence for attention
shift in category learning: Different mechanisms behind illusory correlations
and the inverse base-rate effect. Journal of Experimental Psychology: General,
144(1), 58-75.
7. Matlin, M. (2005). Cognition (6th ed.). New Jersey: Wiley.
8. Peirce, J. W. (2007). PsychoPy—psychophysics software in Python. Journal
of neuroscience methods, 162(1), 8-13.
9. Risen, J. L., Gilovich, T., & Dunning, D. (2007). One- shot illusory correlations
and stereotype formation. Personality and Social Psychology Bulletin, 33(11),
1492-1502.
10. Van Rooy, D., Van Overwalle, F., & Vanhoomissen, T. (2013). Illusory correla-
tion, group size and memory. Journal of Experimental Social Psychology, 49(6),
1159-1167.