Escolar Documentos
Profissional Documentos
Cultura Documentos
Making (2014)
Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/bdm.1849
INTRODUCTION
Throughout our everyday life, we prospectively estimate the
likelihood of possible events and make decisions based on
these judgments. However, probability assessments do not represent an objective and formally realistic processing of available evidence but are inherently subjective and intuitive and
often systematically biased (Tversky & Kahneman, 1982a).
When we estimate our chances of experiencing desirable
events or our risk for adverse experiences, we tend to
overestimate the former and underestimate the latter, thereby
promoting optimistic future outlooks (Shepperd, Klein,
Waters, & Weinstein, 2013; Weinstein, 1980). This optimism
bias has been demonstrated to be prevalent in the whole
populationregardless of age, sex, education, and occupationand for a broad spectrum of domains including health,
nance, career, personal relationships, or self-evaluation
(Dunning, Meyerowitz, & Holzber, 1989; Weinstein, 1980,
1987). Critically, the manner in which lay people and experts
evaluate general and individual risk factors to assess probabilities has direct implications for efforts at personal, public, and
environmental protection (Slovic, Fischhoff, & Lichtenstein,
1982; Weinstein, 1987). For instance, crises in the nancial
markets may be critically promoted by systematic judgmental
limitations and optimism biases (Puri & Robinson, 2007;
Sharot, Korn, & Dolan, 2011).
*Correspondence to: Bojana Kuzmanovic, INM-8, Research Center Juelich,
52425 Jlich, Germany. E-mail: b.kuzmanovic@fz-juelich.de
identication of mechanisms, which may explain the maintenance of optimism despite conicting evidence while taking
into account inter-individual differences in trait optimism.
Stimulus material
We used 88 German descriptions of adverse life events as stimuli (see Supporting Information S1 for the full list). A wide
range of events relating to different life domains (e.g., arthritis,
unemployment, or pest infestation in the home) were included
and randomly assigned to the experimental conditions anew
for each participant in order to attain equal distributions of
event characteristics. Event characteristics that have been suggested to modulate the optimism bias include the base rate,
severity (or event valence), arousal, controllability, and personal experience (Rose, Endo, Windschitl, & Suls, 2008;
Weinstein, 1980, 1987). We tested for relations between event
characteristics and task performance to rule out that these
variables could constitute confounding variables.
B. Kuzmanovic et al.
Figure 1. Examples of experimental trials with varying adverse life events. (A) Condition self_p, self as target person, positive base rate: A
participant estimates his or her probability of suffering from arthritis at least once in a lifetime and is confronted with a lower and thus desirable
base rate. In addition, durations of single events in a trial are provided. (B) Condition other_n, other as target person, negative base rate: A
participant estimates the probability that a similar other will be unemployed at least once in a lifetime and is confronted with a higher and thus
undesirable base rate. Positive and negative base rates were generated by subtracting or adding varying values (maximal 25) from the rst estimate. Updates were computed as the difference between the rst and second estimates
Measures
For each participant, all three experimental measures were
computed separately for each of the four conditions (self_p,
self_n, other_p, and other_n). First, updates were dened as
the differences between the rst and second estimates in each
trial. When the base rate was lower than the rst estimate, it
was expected that the estimate would tend to be adjusted
toward this lower value; thus, update = rst estimate second
estimate. Conversely, when the base rate was higher than the
rst estimate, higher second estimates were expected; thus,
update =second estimate rst estimate. Second, in order to
obtain a measure for the amount of the categorical neglect of
the presented base rates, the percentage of trials with no
difference between the rst and second estimates (update = 0,
update0) was computed. And nally, in order to take into
account predictions of formal learning theories that the amount
of learning should depend on the error size (Garrison et al.,
2013), Pearsons correlation coefcients between the update
and the estimation error (update_error) were calculated.
In order to obtain inter-individual measures of the strength
of the optimism bias, we computed the difference between
the mean update for positive and negative trials, separately
for self and other and individually for each participant.
Thus, the self-related optimism bias (biasself) was computed
as biasself = mean updateself_p mean updateself_n; the otherrelated optimism bias (biasother) was computed as
Copyright 2014 John Wiley & Sons, Ltd.
Statistical analyses
All analyses were conducted using SPSS (IBM SPSS
Advanced Statistics 20.0, International Business Machines
Corporation, Armonk, NY). In order to test for an optimism
bias within the update experiment, we conducted a multivariate,
repeated-measures analysis of variance (rmANOVA) with
within-subject independent variables, target person (self vs.
other) and valence (positive vs. negative), and three dependent
variables, update, update0, and update_error. Self-related
and other-related mean rst estimates were compared by
using the dependent t-test. Pearsons correlation coefcient
was used as an effect size measure (all F-values had one
degree of freedom).
In order to test for relationships between the optimism bias
values (biasself, biasother, and biasself_other) and trait optimism,
Spearmans correlation coefcients (rs) were calculated because questionnaire scores represent ordinal variables. We expected that biasself and biasself_other would positively correlate
with trait optimism (one-tailed testing), but we did not have
any hypotheses about the relationship between biasother and
LOT-R scores (two-tailed testing).
The alpha level of .05 was used for all statistical tests. We
adjusted this threshold for multiple comparisons for the
three correlation analyses by using the Bonferroni correction
(p = .050/3 = .017). Furthermore, events that were experienced at the moment of participation (e.g., when a participant was already suffering from hay fever and had to
estimate the likelihood of suffering from hay fever) were
excluded from all analyses when they occurred in self-trials
by using the ratings of personal experience (rating currently
affected; Supporting Information S2).
J. Behav. Dec. Making (2014)
DOI: 10.1002/bdm
B. Kuzmanovic et al.
Results
Update experiment
The effects of target person and valence of new information
on mean update, update0, and update_error are presented in
Table 1 and Figure 2. The results for these three dependent
variables relating to the amount of learning were fully comparable. The signicant main effect of target person indicates
that belief updating was generally greater during judgments
about others relative to those about oneself. The main effect
of valence shows that participants generally updated less in
response to undesirable than desirable information. And
nally, the signicant interaction effect demonstrates that
this effect of decreased undesirable updating was stronger
in self-related than in other-related judgments. In accordance,
pairwise comparisons revealed signicant differences between positive and negative trials for self, but not for other.
The overall means (with standard deviations in parentheses)
Table 1. Effects of target person and valence of new information on three measures of belief updating from Study 1: mean update,
categorical resistance to update (update0), and correlation between update and estimation error (update_error)
Source
SD
F/t
4.61/6.59
6.87/4.79
2.44/1.67
2.18/1.93
17.26**
18.61**
12.41**
.001
.000
.002
.69
.70
6.65/3.32
7.10/6.24
3.01/2.31
1.88/2.26
5.16**
1.56
.000
.135
.46
.28
40.39/23.42
23.64/38.20
19.09/15.78
16.42/18.21
20.96**
13.29**
8.78**
.000
.002
.008
.72
.64
26.33/50.56
20.85/26.13
20.73/24.43
15.52/18.78
3.64**
1.77
.002
.094
.40
.29
0.42/0.67
0.67/0.46
0.18/0.18
0.17/0.14
30.83**
26.41**
6.21*
.000
.000
.022
.79
.76
0.62/0.34
0.72/0.64
0.19/0.21
0.20/0.21
4.65**
1.82
.000
.084
.44
.30
Mean update
Target person: self/other
Valence: p/n
Target person valence
Pairwise comparisons
Self_p/self_n
Other_p/other_n
Update0 (%)
Target person: self/other
Valence: p/n
Target person valence
Pairwise comparisons
Self_p/self_n
Other_p/other_n
Update_error (r)
Target person: self/other
Valence: p/n
Target person valence
Pairwise comparisons
Self_p/self_n
Other_p/other_n
All dfs were 1, 19. _p and _n, positive and negative valence of base rates.
*p < .05,
**p < .01.
Figure 2. Results of Study 1 demonstrating optimistically biased updating in self-referential risk estimations based on three measures: mean
updates (left), the categorical resistance to update (update = 0, middle), and the correlation between the update and the estimation error (i.
e., the difference between the rst estimate and the base rate; update_error, right). All measures show that belief updating was lower in response to undesirable base rates and that this effect was greater for judgments relating to oneself than to similar others. Error bars show
95% condence interval
1.21/8.28
4.17/5.86
2.24/12.45
13/24
0/11
26/48
26/49
Study 2, N = 26
M (SD)
3.34 (2.89)
0.86 (2.46)
2.48 (3.15)
17.95 (2.61)
3.75 (3.14)
33.25 (5.89)
32.20 (7.54)
Min/max
1.88/6.52
8.17/6.57
5.81/9.76
10/23
0/16
24/56
32/58
M (SD)
2.60 (2.50)
0.70 (3.31)
1.90 (3.51)
17.15 (3.79)
5.15 (3.99)
34.42 (9.23)
34.38 (8.98)
Biasself = mean updates after desirable information mean updates after undesirable information, in self-trials; Biasother = mean updates after desirable
information mean updates after undesirable information, in other-trials; Biasself_other = biasself biasother. BDI, Beck Depression Inventory; LOT-R, Life Orientation Test; STAI, StateTrait Anxiety Inventory.
updates after negative base rates for self and other, highTO
updated less after negative base rates only in self-related
judgments, but not in other-related judgments. Because of
the low and unbalanced sizes of subsamples, we did not compute an analysis of variance (ANOVA) with trait optimism as
a between-subject factor.
Discussion
The primary purpose of Study 1 was to investigate
whether there is a self-specic optimism bias in updating
of beliefs relating to future outcomes. Furthermore, Study
Figure 3. Exploration of the relationship between the update behavior and trait optimism in Study 1 (A) and Study 2 (B). The overall samples
from both studies were divided into a low-trait-optimism subsample and a high-trait-optimism subsample by using the median Life Orientation
Test score (both Mdns = 18) in order to compare the patterns of update behavior. In both studies, the participants with low trait optimism demonstrated a decreased updating after undesirable new information for both self and other, whereas the participants with high trait optimism
showed this bias selectively for self-related judgments. Error bars show SD
B. Kuzmanovic et al.
1 explored the relationship between task performance and
trait optimism.
Method
Participants
A total of 29 individuals were recruited online at the
University of Cologne, Germany. Three persons were
excluded from the analyses because they had missing
responses in more than one third of the trials per condition. Thus, data from 26 participants were included in
Copyright 2014 John Wiley & Sons, Ltd.
Measures
The following measures, which were computed as described
in Study 1, were used: update, update0, update_error, and
bias measures biasself, biasother, and biasself_other (Measures
in Study 1). Additionally, these six measures were computed
after identication and exclusion of trials with apparent
valence misclassications (corrected, corr). Trials with apparent valence misclassication were those where the
valence classication based on the rst estimate and that
based on the subjective base rate estimate differed (see example in the introduction of Study 2). In the latter, we simply
compared the subjective estimates of base rates with the
presented base rates in order to determine whether the presented rates were better or worse than assumed.
Furthermore, we report correlations between the optimism
bias measures and trait optimism scores, as well as an analysis with the between-subject factor trait optimism. For this
analysis, the overall sample was divided by the median of
the LOT-R scores as described in Study 1 (Mdn = 18;
highTO, n = 14; lowTO, n = 12).
Statistical analyses
We tested for the optimism bias before and after correction for
apparent misclassications by using a multivariate rmANOVA
with two within-subject independent variables, target person
and valence, and six dependent variables update, updatecorr,
update0, update0corr, update_error, and update_errorcorr.
Furthermore, we tested for differences between biasself and
biasselfcorr, biasother and biasothercorr, and biasself_other and
biasself_othercorr.
In order to test for relationships between the measures of
the optimism bias and trait optimism, Spearmans correlation
coefcients (rs) were calculated (p < .017, corrected for
J. Behav. Dec. Making (2014)
DOI: 10.1002/bdm
B. Kuzmanovic et al.
Results
Update experiment
Table 3 and Figure 4 show results revealed by the multivariate rmANOVA on six dependent variables: update, update0,
and update_error, with and without correction, respectively.
Taken together, most of the results of Study 1 could be
replicated, independently of correction, with the following
Biases
There were no signicant differences between bias measures
without and with correction: biasself, t(25) = 1.12, p = .274,
r = .21, M = 2.60, SD = 2.50, Mcorr = 2.26, SDcorr = 3.05; biasother,
t(25) = 1.66, p = .109, r = .25, M = 0.70, SD = 3.31, Mcorr = 0.35,
Table 3. Comparison of corrected and uncorrected effects of target person and valence on three measures of belief updating from
Study 2: update, categorical resistance to update (update0), and correlation between update and estimation error (update_error)
Source
SD
F/t
5.34/7.12
6.71/7.74
7.13/5.50
7.67/6.67
2.24/2.74
2.81/3.01
2.40/2.41
2.60/3.18
9.89**
3.94
12.86**
6.77*
7.63*
4.96*
.004
.058
.001
.015
.011
.035
.53
.37
.58
.46
6.69/4.09
7.55/5.30
7.47/6.77
7.82/7.47
2.92/2.13
3.35/2.92
2.78/3.58
2.91/4.14
5.30**
3.77**
1.08
0.49
.000
.001
.290
.630
.42
.36
.20
.14
42.14/28.78
35.24/25.29
28.28/41.38
24.89/37.16
20.77/18.71
19.89/16.75
17.31/20.76
16.19/20.52
17.91**
10.50**
24.68**
30.06**
5.50*
4.88*
.000
.003
.000
.000
.027
.037
.65
.54
.70
.74
32.85/50.91
27.53/47.89
24.66/32.77
22.81/30.39
21.26/22.47
20.77/22.24
17.93/23.61
16.80/25.12
6.21**
5.84**
2.13*
1.81
.000
.000
.043
.083
.45
.44
.28
.26
0.31/0.54
0.32/0.54
0.54/0.37
0.54/0.34
0.18/0.17
0.21/0.19
0.17/0.12
0.20/0.18
17.93**
17.49**
17.14**
15.69**
1.47
3.97
.000
.000
.000
.001
.237
.057
.65
.64
.64
.62
0.47/0.27
0.49/0.19
0.61/0.52
0.60/0.51
0.26/0.22
0.30/0.32
0.18/0.24
0.21/0.27
3.76**
3.79**
1.54
1.32
.001
.001
.137
.199
.36
.36
.24
.22
Mean update
Target person: self/other
Valence: p/ n
Target person valence
Pairwise comparisons
Self_p/self_n
Other_p/other_n
Update0 (%)
Target person: self/other
Valence: p/n
Target person valence
Pairwise comparisons
Self_p/self_n
Other_p/other_n
Update_error (r)
Target person: self/other
Valence: p/n
Target person valence
Pairwise comparisons
Self_p/self_n
Other_p/other_n
Note. Data with correction for apparent misclassications of valence are in bold, below the original data. All dfs were 25. _p and _n, positive and
negative valence of base rates.
*p < .05,
**p < .01.
Figure 4. Results of Study 2 validating the update paradigm. The data without correction represent the original analysis, and the data with correction represent the analysis after exclusion of trials with apparent valence misclassications. The successful replication of the critical ndings irrespective of correction demonstrates that valence misclassications do not systematically inuence the effects indicating an optimism
bias in belief updating. Error bars show 95% condence interval
Frequencies
On average, 23.04 (SD = 5.38) out of 88 trials in the experiment met the criteria of apparent valence misclassication.
However, only in 1.46 trials (SD = 1.30) out of these was
the update made in a direction that one would not expect
given the rst estimate and the base rate (compared with
the mean of 1.19 trials from the non-misclassied trials).
By expected direction, we mean an increased second
estimate if the base rate was higher than the rst estimate
and a decreased second estimate if the base rate was lower
than the rst estimate.
First estimates
Mean rst estimates were lower for judgments relating to self
(M = 30.96, SD = 9.89) than other (M = 37.37, SD = 10.52),
t(25) = 3.83, p = .001, r = .61.
Discussion
Study 2 successfully replicated the optimism bias in belief
updating, supporting the validity of the paradigm. Participants
were generally less likely to update their beliefs in response to
undesirable information, and this to a greater extent during
self-related judgments. This optimism bias was indicated by
the estimation-error-independent measures, mean update and
the percentage of the base rate neglect (update0), as well as
by the individual bias measures. Critically, these results were
demonstrated both without any corrections and in the analyses
excluding trials for which the assessment of participants
subjective base rates suggested a misclassication of the
desirability of new information.
J. Behav. Dec. Making (2014)
DOI: 10.1002/bdm
B. Kuzmanovic et al.
Taken together, valence misclassication cannot account
for the demonstrated bias effects. On the basis of participants assessments of the general base rate, in approximately
one-quarter of trials, apparent misclassications occurred.
Notably, such misclassications would be expected to lead
to a different direction of updating, resulting in negative
update values (because participants do not adjust their rst
estimates toward the presented base rate). However, this only
occurred extremely rarely, indicating that, during the experiment, subjects focused on the presented base rates instead of
on their own prior base rate estimations. At the same time,
however, this neglect of own prior beliefs about base rates
contradicts the general tendency to hold on subjective beliefs
(Slovic et al., 1982). Therefore, we suggest that the neglect
of own prior belief reects memory effects in the sense that
it was not possible for the participants to remember what they
assumed about each of the 88 base rates. Thus, in future studies, the inuence of subjective base rates should be treated
with greater precision within experimental designs (see our
suggestion in Limitations). Along with the replication of effects irrespective of correction, this interpretation relating to
memory effects suggests that subjective base rates played
only a minor role within the update task and are thus not
likely to systematically inuence the observed effects.
However, for the measure relating to the correlation between
updates and estimation errors, the critical interaction effect
between target person and valence did not reach signicance, irrespective of correction for apparent valence misclassications.
This may be attributable to the increased measurement error
within the assessment of the true, that is, subjective
estimation errors, because the potentially relevant personal
vulnerabilities were unknown for the individual participants
and the different future events. If it is possible for the true
subjectively perceived error to be unequal to the exact difference between the own rst estimate and the presented base
rate, then this measurement error may attenuate the size of
related correlations, so that their use as a dependent variable
may lead to inconsistent and unreliable results.
Alternatively, moderator effects of trait optimism as
reported for the measure update might be responsible for the
non-signicant interaction effect. Indeed, similar to the analysis on updates, an equivalent mixed ANOVA on the measure
update_error revealed strikingly differential response patterns
for the subsamples with low and high trait optimism, which
tend to cancel each other out when collapsed (the results of
the mixed ANOVA are shown in Supporting Information
S4). However, for the measure update, the interaction between
target person and valence was signicant despite the moderator
effect of trait optimism, so that it appears plausible that both
explanationsthe problems in measuring estimation errors
and the moderator effect of trait optimismmay have contributed to the unreliable results of this measure.
With regard to the relationship between trait optimism and
the optimism bias in belief updating, Study 2 replicated the ndings of Study 1 with a bigger and more balanced sample. The
signicant correlation between biasself_other and trait optimism
conrms that the higher participants scored on trait optimism,
the higher their optimism bias in belief updating for self relative
to other. Thus, there seems to be a good reliability of the
Copyright 2014 John Wiley & Sons, Ltd.
LIMITATIONS
Although the optimism bias has been demonstrated for both
positive and negative events, in the present study, we
included only negative events. The reason for this procedure
is that the optimism bias tends to be greater for negative
events (Gold & de Sousa, 2011). Furthermore, there is much
more consensus about what constitutes a negative event in
the sense that most people would want to avoid the experience of a disease or an accident, while people may differ to
a larger extent in their evaluations of positive events
(Weinstein, 1980). In addition, it is difcult to nd positive
events that show similar frequencies over the lifetime as the
negative ones (e.g., eat at your favorite restaurant is likely
to occur with a likelihood of 100%). And nally, participants
may be well aware that no ofcial statistics exist for many
base rates of the positive events.
Second, because we wanted to keep the relevance of the
base rate information similar for self and other, we decided
to use base rate statistics relating to the general population.
Although base rates relating to similar others would have
been more informative for self-referential judgments and
thus might have achieved a greater impact, they would have
rendered estimations of risks for similar others trivial
(Harris & Hahn, 2011). Hence, the use of population base
rates can be justied, because it enables an adequate
comparison between self and a similar other.
Third, the use of the linear mixed-effects modeling may
allow for a more appropriate adjustment for and incorporation of correlated data because of repeated measurements
and multiple related outcome measures in future research
(McCulloch & Searle, 2000).
Finally, within the trials with apparent misclassication of
valence in Study 2, we could not reveal a pronounced
inuence of subjectively assumed base rates on the update
behavior (Discussion of Study 2). Rather, participants
adjusted their estimates mostly toward the presented base
rates, possibly because of difculties in remembering their
own base rate estimates. This memory effect could be
avoided by including the assessment of subjective base rates
into the trial of the update experiment. An additional advantage of this design would be the precise assessment of
subjective estimation errors. Moreover, the difference
between the subjective base rate and the rst estimate in such
a design would allow for inferences about participants
estimates of their personal vulnerabilities.
ACKNOWLEDGEMENTS
We thank Gloria Ache for data acquisition in Study 2 and
Jan-Hendrik Heinrichs, Holger Schtz, the anonymous
reviewers, and Tali Sharot for their comments on previous
versions of the manuscript.
J. Behav. Dec. Making (2014)
DOI: 10.1002/bdm
B. Kuzmanovic et al.
Authors biographies:
Bojana Kuzmanovic is a postdoctoral researcher at the Institute of
Neuroscience and Medicine Ethics in the Neurosciences at the
Research Center Juelich. Her research focuses on social cognition
and decision making using behavioral, clinical and neuroscientic
approaches in an interdisciplinary setting.
Anneli Jefferson is a visiting research fellow at the Philosophy
Department at Kings College London and was previously a postdoctoral researcher at the Institute of Neuroscience and Medicine
Ethics in the Neurosciences at the Research Center Juelich.
Her research focus is on philosophy of psychology and ethics,
especially on topics at the intersection of these elds.
Kai Vogeley is a professor for Psychiatry at the University Hospital
Cologne and head of the research group Social Cognition at the
Institute of Neuroscience and Medicine Cognitive Neuroscience
at the Research Center Juelich. His research interests comprise
social cognition and social neuroscience with a recent focus on
nonverbal communication. His clinical interest focuses on highfunctioning autism in adulthood.
Authors addresses:
Bojana Kuzmanovic, Ethics in the Neurosciences (INM-8), Institute
of Neuroscience and Medicine, Research Center Juelich, Jlich,
Germany.
Anneli Jefferson, Department of Philosophy, Kings College London,
London, UK.
Kai Vogeley, Cognitive Neuroscience (INM-3), Institute of Neuroscience
and Medicine, Research Center Juelich, Jlich, Germany.