ArticlePDF Available

The Mixed Blessings of Self-Knowledge in Behavioral Prediction: Enhanced Discrimination but Exacerbated Bias


Abstract and Figures

Four experiments demonstrate that self-knowledge provides a mixed blessing in behavioral prediction, depending on how accuracy is measured. Compared with predictions of others, self-knowledge tends to decrease overall accuracy by increasing bias (the mean difference between predicted behavior and reality) but tends to increase overall accuracy by also enhancing discrimination (the correlation between predicted behavior and reality). Overall, participants' self-predictions overestimated the likelihood that they would engage in desirable behaviors (bias), whereas peer predictions were relatively unbiased. However, self-predictions also were more strongly correlated with individual differences in actual behavior (discrimination) than were peer predictions. Discussion addresses the costs and benefits of self-knowledge in behavioral prediction and the broader implications of measuring judgmental accuracy of judgment in terms of bias versus discrimination.
Content may be subject to copyright.
The Mixed Blessings of Self-Knowledge
in Behavioral Prediction:
Enhanced Discrimination but Exacerbated Bias
Nicholas Epley
University of Chicago
David Dunning
Cornell University
Four experiments demonstrate that self-knowledge provides a
mixed blessing in behavioral prediction, depending on how
accuracy is measured. Compared with predictions of others, self-
knowledge tends to decrease overall accuracy by increasing bias
(the mean difference between predicted behavior and reality) but
tends to increase overall accuracy by also enhancing discrimina-
tion (the correlation between predicted behavior and reality).
Overall, participants’ self-predictions overestimated the likeli-
hood that they would engage in desirable behaviors (bias),
whereas peer predictions were relatively unbiased. However, self-
predictions also were more strongly correlated with individual
differences in actual behavior (discrimination) than were peer
predictions. Discussion addresses the costs and benefits of self-
knowledge in behavioral prediction and the broader implica
tions of measuring judgmental accuracy of judgment in terms of
bias versus discrimination.
Keywords: accuracy; bias; prediction; calibration; discrimination
Echoing what is likely the consensual confidence in
the importance of self-knowledge, C. S. Lewis (1952)
There is one thing, and only one, in the whole universe
which we know more about than we could learn from
external observation. That one thing is ourselves. . . . In
this case we have, so to speak, inside information; we are
in the know. (p. 33)
Without question, most people feel they are in the know
about themselves, particularly when it comes to predict
ing their future outcomes and achievements (Dunning,
Although self-knowledge serves as an obvious guide
to predictions of future behavior, its correspondence
with reality can be surprisingly tenuous. People’s beliefs
about their traits and abilities often correlate modestly, if
at all, with objective measures (Dunning, Heath, & Suls,
2004; Hansford & Hattie, 1982; Mabe & West, 1982).
Similarly, people’s predictions of their own behavior are
sometimes less accurate than predictions made by oth-
ers. Compared with predictions from peers, people are
less able to predict whether they will be promoted into a
leadership position (Bass & Yammarino, 1991), score
high on a test of surgical skill (Risucci, Tortolani, &
Ward, 1989), or suffer a break-up of their current ro-
mance (MacDonald & Ross, 1999). Although it would
seem that people acquire large amounts of diagnostic in
formation about themselves, this information does not
necessarily result in more calibrated predictions about
the future.
This difficulty in accurately predicting one’s own
behavior—but relative accuracy in predicting others’
behavior—is best illustrated in expectations concerning
Authors’ Note: This research was supported financially by National Sci
ence Foundation Grant SES0241544 awarded to Epley and National In
stitute of Mental Health Grant RO1 56072 awarded to Dunning. We
thank Sabiha Barot, Celeste Beck, Ecaterina Burton, David Chen,
Kevin Dugan, Samantha Franklin, Lydia Gardiner, Richard Lonsdorf,
Erin Rapien, Aram Seo, Regina Slejko, and Seema Saifee for assistance
conducting these experiments. Correspondence concerning this arti
cle should be addressed to Nicholas Epley at the University of Chi
cago, 5807 South Woodlawn Avenue, Chicago, IL 60637; e-mail:
PSPB, Vol. 32 No. 5, May 2006 641-655
DOI: 10.1177/0146167205284007
© 2006 by the Society for Personality and Social Psychology, Inc.
socially or morally desirable behavior. People tend to
overestimate, for example, how likely they are to donate
time or money to charitable causes, complete important
tasks ahead of schedule, and maintain their current ro-
mantic relationship (Buehler, Griffin, & Ross, 1994;
Epley & Dunning, 2000; MacDonald & Ross, 1999). In
each case, people believe that they were more likely than
others to behave in an ethical or desirable fashion, but
predictions of others’ behavior were consistently more
accurate than people’s predictions about their own be-
havior. These results suggest that the increased informa-
tion people have about themselves compared with oth-
ers may have a perverse impact on their ability to forecast
their futures, decreasing rather than increasing the ac
curacy of self-predictions. In this case, people appear to
know others better than they know themselves.
These research findings clearly contradict peoples
intuitions that they know themselves better than other
people do, and this research was designed, at least in
part, to reconcile this contradiction between intuition
and the empirical data about self-prediction. We suggest
that people’s intuitions about the accuracy of self-assess
ments and the existing empirical research appeal to two
very different, but commonly confused, forms of accu
racy (Green & Swets, 1966; see also Murphy, 1973; Yaniv,
Yates, & Smith, 1991).
One form is discrimination—the
extent to which predictions accurately discriminate be
tween who is likely to engage in a particular behavior and
who is not—and would be indexed by the correlation be
tween predicted behavior and actual behavior. For ex
ample, discrimination accuracy would index the extent
to which predictions of whether a particular person will
vote in an upcoming election correlates with whether
that individual actually votes. The second form is bias—
the extent to which the average prediction corresponds
to the average of actual behavior. For example, bias
would arise if the overall percentage of people who pre
dicted that they would vote over- or underestimated the
overall percentage who actually do vote. This critical
distinction between discrimination and bias is further
clarified in Table 1.
Although both discrimination and bias are indexes of
accuracy, they measure accuracy of very different sorts.
As shown in Table 1, discrimination and bias can be quite
discrepant within any particular set of predictions
(Epley, Savitsky, & Gilovich, 2001; Gagné & Lydon, 2004;
Gilovich, Medvec, & Savitsky, 2000; Kenny & DePaulo,
1993) and are statistically independent such that the
magnitude of discrimination can be unrelated to the
magnitude of bias (Funder & Colvin, 1997).
We propose that self-knowledge tends to simulta
neously increase both discrimination (and thus accu
racy) and bias (and thus inaccuracy) in behavioral pre
diction. Self-knowledge therefore provides both benefits
and costs to the overall accuracy of predictions com
pared to predictions about others. This proposal helps
to reconcile peoples intuitions about the validity of their
self-knowledge with the scientific research that suggests
major flaws in self-knowledge because we suspect that
people’s intuitions appeal to discrimination accuracy,
whereas scientific research has typically addressed bias.
In addition, this proposal helps to reconcile more gen
eral scientific and social debates about the overall accu
racy of human judgment (Krueger & Funder, 2004). Al
though psychologists have known for some time that the
accuracy of any particular judgment may be measured in
a variety of different ways (Cronbach, 1955), many ex
perimental investigations report only one measure of ac
curacy, even when multiple measures could be calcu
lated. Such simple treatments that highlight one
TABLE 1: Illustration of Discrimination and Calibration in Predicted Versus Actual Completion Times of Five Separate Tasks by Two Hypotheti
cal Predictors
John George
Task Predicted Days Actual Days Predicted Days Actual Days
Task 1 24 20 5 20
Task 2 24 20 5 20
Task 3 22 25 10 25
Task 4 21 25 10 25
Task 5 26 30 15 30
Discrimination r = .21 r = 1
Bias Mean difference = .6 days Mean difference = 10 days
NOTE: In this example (adapted from Campbell & Kenny, 1999), two predictors (John and George) have five separate tasks they are supposed to
complete in the upcoming month and both John and George predict how many days it will take them to complete each task. The correlation be
tween predicted and actual completion times of these tasks is an index of discrimination, and George shows considerably higher discrimination
than John (with a perfect 1.0 correlation between predicted and actual completion times compared against a correlation of .21). The average sim
ple difference between predicted and actual completion times, in contrast, is an index of bias. And here, John is clearly less biased than George
(overly optimistic by an average of 1 day rather than 10 days).
measure of accuracy and overlook others may not pro
vide a full account of the accuracy of human judgment.
As we will report, circumstances that promote one form
of accuracy can simultaneously detract from another
form. In the realm of self-predictions, whether people’s
judgments appear fundamentally accurate or
fundamentally flawed depends in large part on how
accuracy is measured.
The Costs of Self-Knowledge:
Bias in Self-Prediction
There are at least two major reasons that self-
knowledge would increase bias in behavioral prediction.
First, self-knowledge provides case-based (or individuat
ing) information that leads predictors to disregard pop
ulation base rates that would otherwise produce less-
biased predictions. Second, self-knowledge is often
optimistically biased, such that the individuation infor
mation people consult when predicting their own behav
ior has been enhanced in ways that lead to optimistically
biased predictions. We address each of these reasons in
Case-based information. As Kahneman and Tversky
(1979) noted in their analysis of behavioral forecasting,
people generally have two types of information on which
to base a future prediction: case-based and distribu-
tional. Case-based information involves evidence rele-
vant to the specific actor or action being predicted,
whereas distributional evidence involves information
about the long-run distribution of behavior (i.e., base
rates about the actor’s behavior over time of the fre-
quency of actions across the population). Judgments
that utilize both types of information tend to be more
accurate than those that utilize only one (Dunning,
Griffin, Milojkovic, & Ross, 1990; Kahneman & Tversky,
1973). However, people have an overwhelming prefer
ence to base their predictions on case-based evidence
when they have it (Buehler et al., 1994; Kahneman &
Tversky, 1979; Koehler, 1996). This preference exists de
spite people’s well-documented ability to accurately
code both the frequency as well as the underlying dis
tribution of behaviors in the population (Hasher, Zacks,
Rose, & Sanft, 1987; Nisbett & Kunda, 1985).
There is, of course, a fundamental asymmetry be
tween the kind of information people possess about
themselves and others that may explain why people can
sometimes predict others’ behavior better than their own.
People have a wealth of case-based information about
themselves, but they have only automatically encoded
base rates when considering unknown others—general
intuitions about how most people behave in a given situ
ation. This asymmetry means that self-predictions are
more likely to be guided by case-based evidence than by
base rates, whereas predictions of others are more likely
to be guided by intuitive base rates than by case-based
evidence (Buehler et al., 1994; Epley & Dunning, 2000).
Self-knowledge may therefore lead people to ignore use
ful distributional information that they utilize when pre
dicting others’ behavior.
Optimistic biases. An excessive focus on case-based
information in self-predictions would not be problem
atic if such information was unbiased, but self-
knowledge is rarely generated dispassionately. Instead,
self-knowledge is often massaged or molded in just the
right ways to maintain positive images of the self
(Dunning, 1999; Kunda, 1990; Taylor, 1989). As a conse
quence, the case-based information people use to make
self-predictions is often optimistically distorted. This dis
tortion would obviously leave the individual with infor
mation that would mislead self-predictions. The same in
formation applied to others is unlikely to be colored by
the same kind of optimistic distortions, meaning that
predictions about random others are less likely to be
optimistically biased (Epley & Dunning, 2000).
Implications. A preference for case-based information
coupled with selective self-enhancement makes peoples
ability to more accurately predict others’ behavior than
their own somewhat clearer. When predicting a random
stranger, people rely on intuitive base rates for the rele-
vant behavior. Because these intuitive base rates tend to
be at least reasonably calibrated, people’s predictions
are also reasonably calibrated. When predicting the self,
however, the combination of case-based knowledge and
the desire to maintain positive self-images leads people
to give little, if any, weight to distributional information.
These two factors together leave self-predictions more
prone to optimistic biases in comparison to reality than
peer predictions.
The Benefits of Self-Knowledge:
Discrimination in Self-Prediction
The preceding description of self-knowledge paints a
rather dim picture of peoples ability to “know them
selves,” but it does so only with respect to bias. A consid
erably different picture is likely to emerge when accu
racy is measured via discrimination. After all, peoples
self-knowledge is likely to increase discrimination accu
racy to the extent that this individuating information is
diagnostic of actual behavior. Indeed, self-predictions of
ten contain just as much discriminant accuracy as objec
tive measures (e.g., test scores) when predicting perfor
mance across a number of different domains, including
intellectual achievement, vocational choice, and future
preferences (Shrauger & Osberg, 1981, 1982). Predic
tions of others, in contrast, are likely to show little
Epley, Dunning / SELF-PREDICTION 643
discrimination due to the absence of diagnostic
individuating information.
Overview of Studies
We therefore predicted that self-knowledge would
confer both costs and benefits to behavioral predictions.
Self-predictions would show better discrimination than
peer-predictions but would simultaneously show more
bias as well. We tested this predicted pattern of discrimi
nation and bias in four experiments, the first two involv
ing predictions of voting and the last two involving pre
dictions of the longevity of romantic relationships.
Studies 2 through 4 also allowed us to explore the
mechanisms by which self-prediction differs from social
prediction. Because self-knowledge is difficult to manip
ulate, Studies 2 through 4 instead manipulated the kind
of information people had about others. This allowed
us to test our model of the costs and benefits of self-
knowledge in prediction by testing whether gaining
knowledge about another person produced costs and
benefits similar to those we attribute to self-knowledge.
We predicted that providing participants with relevant
information about another person would increase dis-
crimination (Study 2) but lead people to disregard rele-
vant base rates that they otherwise would attend to in the
absence of individuating information (Study 3), pro-
ducing bias when the relevant information is construed
optimistically (Study 4).
Approximately 1 month before the 2000 U.S. presi-
dential election, participants predicted whether they
and a randomly selected student from their class would
vote. Two days after the election, these participants were
contacted again and asked whether they actually voted.
We expected to observe greater discrimination as well as
bias in self-predictions than peer predictions. In addi
tion, participants were given a small amount of individu
ating information about another person in the experi
ment and were asked to predict his or her behavior as
well. This prediction target allowed us to begin explor
ing the effects of individuating information per se on dis
crimination and bias.
One-hundred ninety Cornell undergraduates were
told that they would be asked to think about their own
and others’ behavior in an upcoming event. Before do
ing so, participants were asked to provide a short self-
description that they were told would be used later in the
experiment. This self-description served as the individu
ating information that would later be provided to an
other participant in the experiment, and it required
them to write five words that best described their person
ality as well as a short description of “who they are.”
When finished, the experimenter exchanged the partici
pants’ self-description with a description completed by
another participant. They were told to read this new self-
description carefully because they would be asked to
make some predictions about this person’s behavior
later in the experiment.
After several minutes, the experimenter gave partici
pants a questionnaire that included a reminder of the
upcoming election and asked them to predict whether
they, the participant they received information about
(hereafter referred to as the individuated peer), and a
randomly selected person from this experiment (hereaf
ter referred to as the random peer) would vote in the up
coming election. The order in which participants pre
dicted the different targets was counterbalanced in a
Latin-square design.
Participants were contacted again 1 day following the
election by e-mail (or telephone, if necessary) and were
asked whether they were eligible to vote in the election
and, if so, whether they voted.
Results and Discussion
Participants were not screened prior to the experi-
ment to determine whether they were eligible to vote.
Because participants were predicting the behavior of an-
other participant in the experiment, some predicted the
behavior of an ineligible voter. Ineligible voters, or par-
ticipants who were considering ineligible voters, were
therefore excluded from the following analyses. Partici-
pants who could not be contacted after the election, or
who were predicting someone who could not be con
tacted, also were excluded. This left 104 participants
with complete data. Including participants with incom
plete data, however, does not alter the significance levels
of any of the following analyses.
As seen in Table 2, participants predicted that they
were more likely to vote than both the randomly selected
and individuated peer, paired t(103) = 3.28, p < .01, d =
.65. This difference reflected bias in self-predictions.
Participants overestimated the likelihood that they
would vote, t(103) = 7.28, p < .01, d = 1.43, but were rea
sonably calibrated when predicting both the individu
ated and random peer, ts(103) = 1.35 and .21, both ps >
.17, ds = .27 and .04.
As predicted, this increased bias in self-predictions
was offset by increased discrimination. Overall, 75% of
participants correctly predicted their own behavior,
whereas participants would have been right only 65% of
the time if they had been predicting themselves ran
domly (based on base rates of predicted and actual be
havior). Only 61% and 60% were correct when predict
ing the individuated and random peer, respectively, with
the associated chance accuracy levels for the individu
ated group of 58%.
Of course, these hit rates are influenced by the actual
(high) base rate of voting and so simply predicting that
any person would vote also would increase one’s hit rate
without actually increasing discrimination. The better
measure of discrimination is therefore the correlation
between predicted and actual behavior. This correlation
shows that self-predictions were significantly related to
actual voting behavior, φ = .35, p < .001, whereas individu-
ated peer predictions were not, φ = .11, p = .18. The dif-
ference between correlations is marginally significant,
z = 1.81, p = .07. Furthermore, logistic regressions indi-
cated that participants who predicted they would vote
were more likely to do so (M = 74.47%) than were those
who predicted they would not vote (M = 20%), β = 2.45,
p < .01. In contrast, individuated peers who were pre-
dicted to vote were not more likely to do so (M = 69.86%)
than were those who were predicted not to vote (M =
61.29%), β = .38, ns.
That discrimination was somewhat lower for individu
ated peer predictions suggests that the individuating
information participants received did not provide the
same discriminant accuracy as self-knowledge. This is
not very surprising because participants received no in
dividuating information relevant to the likelihood of vot
ing. These results suggest that it is not the mere presence
of case-based information that produces the patterns of
discrimination and bias observed in self-predictions but
rather the quality of that information. We explored this
issue further in Study 2.
Participants in Study 1 demonstrated both greater dis
crimination and greater bias when predicting their own
behavior than when predicting randomly selected oth
ers. We predicted this pattern would arise because self-
knowledge contains at least some relevant and diagnos
tic information that increases discrimination, but with a
cost of leading people to ignore relevant distributional
information. Study 2 addressed this explanation more
directly by providing participants either relevant or irrel
evant case-based information about one of their peers. If
self-knowledge increases discrimination because it con
tains relevant diagnostic information, then providing
relevant (and diagnostic) information about another
person, as opposed to irrelevant information, should
increase discrimination accuracy.
We predicted that people are less biased when pre
dicting others’ behavior in the absence of relevant indi
viduating information, however, because they are more
likely to utilize relevant distributional evidence than
when predicting one’s own behavior. To test this idea di
rectly, Study 2 also manipulated the distributional evi
dence available to participants. We expected that partici
pants’ predictions of others’ behavior would be strongly
influenced by this information when they did not possess
relevant case-based information (i.e., when predicting
a random peer or an irrelevant individuated peer),
whereas participants’ predictions of their own or the rel
evant individuated peer’s behavior would not.
Finally, Study 2 included two important methodologi-
cal improvements. First, participants made their predic-
tions on a continuous as well as a dichotomous measure.
This was done to rule out concerns that the dichotomous
measure used in Study 1 was too insensitive to detect sub-
tle differences in accuracy. Second, Study 2 included a
no-prediction control group to address concerns that
predicting one’s own behavior may alter the behavior
being predicted (Sherman, 1980).
Participants. One hundred and thirty-eight Harvard
University undergraduates approached in one of several
campus dormitories received $5 for their participation.
Efforts were made by the experimenters to target stu
dents who appeared to be U.S. citizens to reduce the
number of ineligible voters in the sample.
Procedure. All participants were told that this experi
ment investigated predictions of one’s own and others’
behavior and that this experiment would require com
pleting a series of questionnaires as well as responding to
a brief e-mail questionnaire approximately 1 month
later. All participants agreed to answer the e-mail
Participants were next told that we needed to collect
some background information and were given two ques
tionnaires. One included five items directly relevant to
the upcoming presidential election, whereas the other
included five items that were relatively irrelevant. For in
stance, the relevant questionnaire asked, “How inter
ested would you say you are in the upcoming presiden
Epley, Dunning / SELF-PREDICTION 645
TABLE 2: Predicted Versus Actual Voting Behavior for the Self, Indi
viduated Peer, and Random Peer (Study 1)
Percentage Voting
% % Accurate
Target Predicted Actual Predictions
Self 90 69 75
Individuated peer 70 70 61
Random peer 75 69 60
NOTE: Actual behavior for the individuated peer varies because a
small number of participants were predicted multiple times.
tial election?” and “How pleased do you think you will be
if your preferred candidate wins the presidential elec
tion?” whereas the irrelevant questionnaire asked, “How
interested would you say you are in modern art?” and
“How pleased do you think you would be if you won an
all-expenses paid trip for 1 week to Florida?” Control
participants were then excused from this first phase of
the experiment.
Participants in the prediction conditions then re-
ceived either the relevant or irrelevant information
questionnaire completed by another participant in the
experiment, along with the questionnaire asking them
to predict their own and others’ behavior. The first page
began with a sentence noting the upcoming presidential
election and continued with some ostensible back-
ground information about the typical percentage of stu-
dent voters from Ivy League universities—a relevant
comparison group for these Harvard University stu
dents. Participants in the low base rate condition read
that “voting among Ivy League college students has been
historically quite low, with only 15% of students enrolled
in Ivy League colleges voting, on average, in the presi
dential election over the last 30 years.” Participants in
the high base rate condition, in contrast, read that “vot
ing . . . has been historically quite high, with 70% of stu
dents . . . voting . . . over the last 30 years.”
Participants in the prediction conditions were then
asked to predict whether they, a randomly selected Har
vard student, and the individuated peer would vote in
the upcoming election (yes or no) as well as the likeli
hood that each of these targets would vote on a scale
ranging from 0% (not at all likely) to 100% (absolutely cer
tain). The order of these predictions was counterbal
anced in a Latin-square design. All participants were
contacted again 2 days after the election by e-mail and
asked to indicate whether they had voted in the presi
dential election and whether they were U.S. citizens (to
determine their voting eligibility).
Participants who were ineligible to vote or who were
predicting the behavior an ineligible voter were re
moved from the analyses. This left 131 participants in the
final analysis (105 in the prediction conditions and 26 in
the control condition). Including the partial data from
excluded participants does not alter the significance
levels of any of the following analyses.
Participants predicted their own and others’ voting
behavior once as a dichotomous judgment and once as a
continuous judgment. Not surprisingly, these measures
were largely redundant. To highlight the unique contri-
butions of Study 2 and to simplify presentation, we dis-
cuss only results from the continuous likelihood judg-
ments below.
Accuracy. All relevant means and correlations for be-
havioral predictions and actual voting behavior are pre
sented in Table 3. As in Study 1, participants predicted
that they would be more likely to vote (M = 89.9%) than
either their individuated (M = 64.9%) or random peers
(64.8%), paired ts(104) = 9.17 and 9.68, respectively, ps <
.001, ds = 1.80 and 1.90. This reflected significant bias in
self-prediction because participants significantly over
estimated the likelihood that they would vote, paired
t(104) = 5.82, p < .001, d = 1.14. The actual rate of voting
observed in these prediction conditions did not seem to
be influenced by the act of making a prediction because
the voting rate in the no-prediction control was statisti
cally identical to the prediction conditions, χ
< 1.
As in Study 1, the increase in bias for self-predictions
compared to peer predictions was offset by a corre
sponding increase in discrimination. As seen in Table 3,
there was a significant positive correlation between pre
dictions and reality for self-predictions and also for rele
vant individuated peer predictions. These correlations
were both significantly higher than the analogous (non
significant) correlation observed for irrelevant peer pre
dictions, zs = 2.91 and 2.23, respectively, both ps .01.
TABLE 3: Predicted Likelihood of Voting Versus Actual Voting for the Self, Individuated Peer, and Random Peer Among Participants Provided
With High Versus Low Base Rate Information (Study 2)
Base Rate Condition
Target High Low Difference Actual r (Prediction/Actual)
Self 89% 90% –1% 66% .51*
Individuated peer
Relevant information 68% 65% 3% 61% .48*
Irrelevant information 75% 51% 24%* 69% .07
Random peer 76% 53% 23%* 66%
Control (no prediction) 69%
NOTE: Participants predicted the behavior of three targets: self, a randomly selected peer, and an individuated peer about whom they received ei
ther relevant or irrelevant information. Actual behavior varies because a small number of participants were predicted multiple times in the individu
ated peer predictions.
*p < .05.
There are two interesting points worth noting. First,
discrimination for the relevant individuated peer pre
diction was not only higher than it was for the irrelevant
peer prediction but it also was just as high as discrimina
tion in self-predictions. Self-knowledge provided bene-
fits to discrimination that were relatively easy to mimic in
predictions of other individuated peers. Second, the
most accurate predictions overall—that is, with the most
discrimination and least bias—were for the relevant indi-
viduated peer. These predictions contain the same kind
of diagnostic information as self-knowledge without the
accompanying optimistic bias, exactly as our theory
would predict.
Use of distributional evidence. Table 3 also shows that par
ticipants’ self-predictions were not significantly influ
enced by the base rates of voting behavior we provided,
and neither were predictions of the individuated peer
when we provided relevant information, both ts < 1. The
base rate information did, however, significantly influ
ence predictions of the random peer, t(103) = 7.23, p <
.001, d = 1.42, and of the individuated peer with irrele
vant information, t(54) = 4.63, p < .001, d = 1.26. Because
the relevant versus irrelevant information conditions ap
plied only to predictions of the individuated peer, we
tested this overall pattern of results in two separate
ANOVAs. The first compared the effect of the base rate
manipulation on predictions of self and the random
peer in a 2 (base rate: high vs. low) × 2 (target: self vs. ran
dom peer) mixed-model ANOVA with repeated mea
sures on the last factor. This analysis yielded the pre
dicted main effect for target, F(1, 103) = 118.35, p < .001,
= .53, qualified by the predicted interaction, F(1, 103) =
27.24, p < .001, η
= .21. This indicated that the base rate
manipulation influenced random peer predictions but
not self-predictions. The second ANOVA tested the im
pact of background information and base rates on pre
dictions of the individuated peer only in a 2 (base rate:
low vs. high) × 2 (background information: relevant vs.
irrelevant) ANOVA. This analysis yielded a significant
main effect of base rate, F(1, 101) = 10.66, p < .001, η
.10, again qualified by the predicted interaction, F(1,
101) = 6.35, p < .05, η
= .06. This indicates that the base
rate information influenced predictions of the individu
ated peer only when provided with irrelevant individuat
ing information, consistent with our theoretical
Use of case-based information. Instead of using distribu
tional evidence, participants used relevant individuating
case-based evidence to predict their own and the rele
vant individuated peer’s behavior. As shown in the first
row of Table 4, the relevant individuating information
about political interests was significantly correlated with
participants own predicted likelihood of voting,
whereas the correlation with irrelevant information
about personal interests was not. The difference be
tween these two correlations was significant, z = 2.94, p <
.01. A similar pattern emerged for predictions of the in-
dividuated peer, with a significant correlation observed
between the individuating information and the likeli-
hood of voting for the relevant individuated peer, but
not for the irrelevant individuated peer. Again, the dif-
ference between these two correlations was significant,
z = 3.47, p < .001.
Summary. Overall, these results are consistent with our
analysis that participants disregard distributional evi-
dence when relevant case-based evidence is available, in-
creasing discrimination but decreasing the tendency to
consider relevant base rates that could reduce bias.
These results make it clear that participants are not sim
ply using whatever information is deliberately provided
to them but are using it selectively. Participants utilized
relevant case-based information when they had it to
make their predictions and relied on base rates when
they did not. Discrimination accuracy can be enhanced
by attending to relevant case-based information,
whereas bias can be reduced by attending to relevant
base rates.
The main goal of Study 3 was to extend the results of
Study 2 to a different domain where predictions of the
future are of obvious practical importance. In particular,
we investigated predictions of the longevity of romantic
relationship by asking participants to predict the likeli
hood that their own and others’ current romantic rela
tionships would be intact 1, 3, and 6 months into the
Epley, Dunning / SELF-PREDICTION 647
TABLE 4: Correlations Between Individuating Background Infor
mation for Predictions of Self and Individuated Peer
(Study 2)
Background Information
Relevant Irrelevant
Target Information Information
Self .51** .15
Individuated peer
Relevant information .66** (.09)
Irrelevant information (.11) .09
NOTE: Participants completed both relevant and irrelevant back
ground information, meaning that correlations can be calculated for
both kinds of information on self-predictions. However, participants
received only relevant or irrelevant information about an individuated
peer. Correlations for individuated peers presented in parentheses are
with information participants did not receive.
*p < .01.
In contrast to Study 2, predictors in this experiment
were not asked to provide the relevant and irrelevant in
dividuating information before making their prediction
to control for an alternative interpretation that doing so
altered participants behavior and artifactually in
creased discriminant accuracy (Sherman, 1980). In
stead, relevant and irrelevant individuating information
was collected from a different set of participants who
served as a behavioral baseline for these individuated
peer predictions.
Participants. Sixty-four Harvard University undergrad
uates involved in a dating relationship received $6 in ex
change for their participation. Of these, 22 served as a
behavioral baseline for the individuated peer predic
tions. The remaining participants were in the prediction
condition. Participants were recruited in groups of up to
three but completed all materials in private cubicles. All
participants completed this experiment during a 3-week
period, but participants in the baseline condition were
run first to keep participants in the prediction condition
from believing they were participating simultaneously
(and therefore potentially knew) with the targets who
completed the relevant and irrelevant information
Procedure. As part of an unrelated experiment, par-
ticipants in the baseline behavior condition were asked
to provide both relevant and irrelevant information
about their dating relationship that would later be used
by prediction participants. The relevant information
questionnaire asked participants to indicate the extent
to which five adjectives described their current dating
relationship—honest, trusting, compatible, exciting, and
enjoyable—and to answer three open-ended questions—
”How many hours per week would you say you spend to
gether?” “How many arguments would you estimate you
have each week (and what is the most common cause of
them)?” and “What hobby or activity do you and your
partner most enjoy doing together (and how often do
you engage in this activity)?” The irrelevant information
questionnaire followed the same format but asked about
information that was transparently irrelevant to the lon
gevity of the relationship. Participants were asked to in
dicate the extent to which five irrelevant adjectives de
scribed their relationship—athletic, intellectual, musical,
artistic, and creative—and to answer three open-ended
questions—”Have you, at any point in the last year, run
together in a competitive race?” “Have you, at any point
in the last year, worked together in a staff position at
a public library?” and “Do you think you will, at any
point in the upcoming year, have dinner together in a
Participants in the prediction conditions, in contrast,
were told upon arrival to the lab that this study would re
quire them to make some predictions about their own
and others’ future behavior. Participants were then
given either a relevant or irrelevant background ques
tionnaire completed by one of the participants in the
baseline condition, were told that they would be asked to
make some judgments about this person later in the ex
periment, and were given several minutes to look over
the questionnaire.
When finished, participants received a questionnaire
informing them that they would be asked to make some
predictions about their own and others’ dating relation
ships at several points in the future. Participants read
that they were to provide their most accurate predic
tions, rather than their most hopeful or optimistic, and
were told that their responses would remain strictly con
fidential. The first page of this questionnaire also in
cluded a base rate manipulation similar to that used in
Study 2. In particular, participants in the low base rate
condition read that “a survey conducted on campus last
year indicated that Harvard dating relationships last, on
average, approximately 2 months.” Participants in the
high base rate condition read that . . . relationships last,
on average, approximately 8 months.”
All participants then predicted (by checking yes or
no) whether they, a randomly selected participant in the
experiment, and the person about whom they received
some information would be dating their current partner
1, 3, and 6 months into the future.
One, 3, and 6 months after this initial session, all par-
ticipants were contacted again by e-mail and asked if they
were currently involved in the relationship they consid
ered during the initial session (yes or no). Participants
who could not be contacted at one of these points were
asked to recall whether they were still dating at the previ
ous time interval. One participant in the baseline condi
tion and 3 participants in the prediction condition could
not be contacted at any time after the initial session, and
data involving these participants are therefore missing
from the following analyses.
Participants were less optimistic in their predictions
as the length of the relationship increased, but optimism
waned much faster in predictions of others’ relation
ships (see Table 5). More important, pessimistic base
rates were utilized only when predicting targets about
whom little was known—the irrelevant individuated and
randomly selected peer. Base rates were completely ne
glected when participants predicted their own and the
relevant individuated peer’s relationship.
We submitted participants’ predictions to a 4 (target:
self, relevant individuated peer, irrelevant individuated
peer, random peer) × 3 (time: 1, 3, and 6 months) × 2
(base rate provided: 2 months vs. 8 months) ANOVA
with repeated measures on the first two factors. This
analysis uncovered significant main effects for time, F(2,
32) = 7.67, p < .05, η
= .32, and target, F(3, 31) = 7.67, p <
.05, η
= .43, qualified by three interactions. A Time ×
Target interaction indicated that participants’ predic-
tions became more pessimistic about random and irrele-
vant individuated peers as time increased, F(6, 28) =
4.66, p < .05, η
= .50, and a Time × Base Rate interaction
indicated that base rates influenced predictions more
strongly as time increased, F(2, 32) = 3.81, p < .05, η
.19. More important, a Base Rate × Target interaction in
dicated that participants completely ignored base rates
when they possessed relevant case-based information,
F(2, 32) = 3.06, p < .05, η
= .16. As can be seen in the far
right column of Table 5, 6-month predictions showed no
impact of the base rate manipulation for either partici
pants’ own or the relevant individuated peer’s relation
ship, both ts < 1, both ps > .9, but showed significant ef
fects for predictions of both the irrelevant individuated
and random peer’s relationships, ts (40) = 2.20 and 3.63,
respectively, both ps < .05, ds = .70 and 1.15. Had these
base rates actually been accurate, participants would
have again been much more calibrated predicting the
behavior of complete strangers than predicting one’s
own behavior. In reality, 58% of participants who
predicted their own behavior were still dating after 6
months, as were 62% of relevant individuated peers and
53% of irrelevant individuated peers.
Although optimistically biased about the longevity of
their relationships, Table 6 indicates that discrimination
was again increased by relevant individuating informa
tion. Predictions of participants’ own and the relevant
individuated peer’s behavior correlated more strongly
with actual behavior than predictions of the irrelevant
individuated peer’s relationship. To test the statistical
significance of these patterns, we created a new variable
that coded a correct prediction as 1 and an incorrect pre
diction as 0 for each target at each time period. These hit
rates were submitted to a 3 (target: self, relevant individ
uated peer, irrelevant individuated peer) × 3 (time: 1, 3,
and 6 months) × 2 (base rate provided: 2 months vs. 8
months) ANOVA with repeated measures on the first
two factors. This analysis yielded only a marginally signif
icant main effect for target, F(2, 32) = 3.13, p = .06, η
.16, and a significant main effect for time, F(2, 32) =
18.31, p < .001, η
= .53, qualified by the predicted Target
× Time interaction, F(4, 30) = 4.51, p < .01, η
= .38.
Discriminant accuracy was lowest for predictions involv-
ing the irrelevant individuated peer, φ = –.14, ns, for 6-
month predictions. As in Study 2, providing relevant in-
formation about another person increased discrim-
ination to levels identical to self-predictions, F(1, 34) < 1,
p > .5, φs = .29 and .23, both ns, two-tailed, for self-
predictions and relevant information predictions at 6
months, respectively.
As in the previous studies, discriminant accuracy in
creased with the presence of relevant individuating infor
mation and decreased with its absence. Of interest, across
Studies 2 and 3, we find that peer predictions with rele
vant information are just as accurate as self-predictions.
Overall, peer predictions with relevant information were
68% accurate—the exact rate achieved by self-predic
tions (and in contrast to the rate of 50% achieved by
peers with irrelevant information). Expressed in terms
of effect size, the discrimination achieved across both
studies by self-predictions, r = .40, was only slightly higher
than that achieved by peers with relevant information, r
= .37. Peer predictions with irrelevant information
showed no such discrimination, r = –.04. The discrimina
tion effect size of irrelevant peer predictions was signifi
cantly lower than self-predictions and relevant informa
tion ones, both Zs > 2.9, p < .01.
Studies 2 and 3 provide an explanation for the ob
served patterns of discriminant accuracy in predictions
Epley, Dunning / SELF-PREDICTION 649
TABLE 5: Predicted Percentage of Intact Relationships for the Self,
Random Peer, and Individuated Peer at 1-, 3-, and 6-Month
Time Periods for Those Provided Low (2-Month Average)
Versus High (8-Month Average) Base Rates (Study 3)
Predicted 1 Month 3 Months 6 Months
Low base rate 91% 75% 75%
High base rate 95% 90% 76%
Relevant individuated
Low base rate 100% 80% 55%
High base rate 95% 82% 55%
Irrelevant individuated
Low base rate 90% 55% 15%
High base rate 95% 64% 45%
Random peer
Low base rate 95% 20% 0%
High base rate 95% 64% 41%
TABLE 6: Percentage of Correct Predictions for the Self, Relevant
Individuated Peer, and Irrelevant Individuated Peer for 1-,
3-, and 6-Month Time Periods (Study 3)
Target 1 Month 3 Months 6 Months
Self 85% 82% 64%
Relevant individuated peer 93% 69% 62%
Irrelevant individuated peer 87% 58% 42%
of the self and undifferentiated others—self-knowledge
contains diagnostic accuracy that increases discrimina
tion compared against predictions of typical others.
They do not, however, fully explain why self-predictions,
at the same time, tend to be optimistically biased com
pared to predictions of others. We have argued that this
bias occurs because self-knowledge is optimistically bi
ased (Dunning, 1999; Kunda, 1990; Taylor, 1989) and
attention to this individuating information leads people
to disregard an intuitive assessment of base rates that
would otherwise debias their predictions. Such an opti
mistic construal of relevant data would likely lead to
accurate discrimination but optimistic bias compared
to predictions of others. We tested this hypothesis in
Study 4.
Participants in this experiment were again asked to
predict the longevity of dating relationships, but this
time only others’ relationships and not their own. Partic
ipants received information about another’s relation
ship, either relevant to its longevity or irrelevant as in
Study 3. To mimic the optimistic bias that we believe per
meates self-prediction, some participants were asked to
interpret this information in the best possible light, con-
struing the information as optimistically as they could
without knowingly disregarding or blatantly distort-
ing information. Other participants were asked to in-
terpret this information as objectively as possible. We
made two predictions. First, discriminant accuracy
would be higher for participants who received relevant,
as opposed to irrelevant, individuating information,
regardless of their construal condition. Second, and
more important, that optimistic biases would arise in
only one of the four experimental conditions
participants asked to optimistically interpret relevant
individuating information.
Participants. One hundred thirty-two Harvard Univer
sity undergraduates participated in this experiment in
exchange for either extra course credit in their psychol
ogy courses or $6. All of these participants predicted the
longevity of another person’s relationship, and these tar
gets included the 22 baseline behavior participants from
Study 3 plus an additional 11 participants who provided
the same information shortly after Study 3 was com
pleted and were also contacted 1, 3, and 6 months after
the initial session to assess the status of their relationship.
Procedure. All participants completed the prediction
questionnaires individually. Because participants were
not predicting their own relationships in this experi
ment, participants did not need to be currently involved
in a dating relationship.
Participants were told that they would receive some
information about another person’s dating relationship
and then make some predictions about whether this per
son would still be dating his or her partner 1, 3, and 6
months from the time this information was collected.
Participants were then given a questionnaire containing
the optimistic versus objective construal manipulation,
the individuating information questionnaires com
pleted by previous participants, and the prediction
All participants read, “As with any information, it is
possible to interpret the relationship description you are
about to read in a variety of different ways.” Participants
in the optimistic condition then read,
One possibility is to interpret the description in an opti
mistic light, looking for the most positive spin on the
information as it relates to the quality and likely longev
ity of this person’s dating relationship. Putting a positive
spin on the information does not mean knowingly ignor
ing or misrepresenting the information you are given
but rather interpreting the information in the most posi
tive way you plausibly can—something similar to a “best-
case scenario.”
In contrast, participants in the objective condition read,
One possibility is to interpret the description in an objec-
tive light, trying to interpret the information presented
in the description as objectively and rationally as you
can. This does not mean interpreting information in a
negative light or ignoring feelings and emotions pre-
sented in the description but rather interpreting the
description as realistically as you possibly can
something similar to a “most likely scenario.”
All participants then read either relevant or irrelevant in
dividuating information about another person’s dating
relationship using the materials as in Study 3. Finally,
participants predicted whether their target would still be
involved in his or her dating relationship 1, 3, and 6
months from the point at which the individuating
information was provided.
Results and Discussion
Participants in this experiment received either rele
vant or irrelevant information about a target that they
were asked to construe in either an optimistic or objec
tive fashion. Four separate participants—one in each of
the experimental conditions—therefore made predic
tions about each target. However, because the relevant
and irrelevant information differed so dramatically for
each target, we opted for a more conservative approach
and analyzed all responses at the level of the individual
participant rather than the target. One target, as men
tioned in Study 3, could not be contacted after the ini
tial experiment. Predictions involving this target are
retained in the following analyses but are missing from
all analyses involving comparisons with actual behavior.
Two main effects emerged from participants’ predic-
tions (see Figure 1). First, participants asked to construe
the individuating information optimistically made more
optimistic predictions about the target’s relationship
than those asked to construe the individuating informa
tion objectively. Second, participants given relevant indi
viduating information about the target’s relationship
were more optimistic than those given irrelevant infor
mation, replicating a similar effect observed in Study 3.
As predicted, only participants who construed relevant
individuating information optimistically were optimisti
cally biased. In fact, 76% of participants believe the tar
get would still be dating their current partner 6 months
later, the very same percentage as participants predicted
for their own relationships in Study 3. Asking partici
pants to interpret relevant individuating information
positively made their prediction look exactly like those
based on self-knowledge.
To analyze the statistical significance of these re
sponses, predictions were submitted to a 3 (time: 1, 3,
and 6 months) × 2 (information: relevant vs. irrelevant)
× 2 (construal: optimistic vs. objective) ANOVA with re
peated measures on the first factor. All three main effects
in this analysis were significant. Participants predicted
that relationships were more likely to break up (a) as
time passed, F(2, 127) = 74.22, p < .05, η
= .53, (b) when
they received irrelevant as opposed to relevant individu-
ating information, F(1, 128) = 24.29, p < .001, η
= .16,
and (c) when they interpreted the individuating infor
mation objectively rather than optimistically, F(1, 128) =
8.98, p < .01, η
= .07. These main effects were qualified
by a significant Time × Information interaction, F(2,
127) = 8.52, p < .05, η
= .12, indicating that predicted
likelihood of a break-up increased faster over time when
participants received irrelevant individuating informa
tion, F(2, 130) = 64.03, p < .001, η
= .50, than when they
received relevant individuating information, F(2, 130) =
23.70, p < .001, η
= .27. Finally, a marginally significant
Time × Construal interaction, F(2, 127) = 2.47, p = .09,
= .04, indicated that the predicted likelihood of break-
up also increased faster over time when participants in
terpreted the individuating information objectively, F(2,
130) = 54.25, p < .001, η
= .45, than when they inter
preted it optimistically, F(2, 130) = 28.77, p < .001, η
.31. No other interactions were significant (all Fs < 1).
A secondary prediction was that participants
discriminant accuracy would show a somewhat different
pattern, showing an increase in accuracy only for rele
vant information. To test this prediction, we calculated
whether each prediction was accurate and submitted
Epley, Dunning / SELF-PREDICTION 651
1 Month 3 Months 6 Months
Actual %
Figure 1 Predicted versus actual percentage of intact relationships (i.e., calibration) among participants who receive relevant or irrelevant individ-
uating information that they interpret optimistically or objectively (Study 4).
these scores to a 3 (time: 1, 3, and 6 months) × 2 (infor-
mation: relevant vs. irrelevant) × 2 (construal: optimistic
vs. objective) ANOVA with repeated measures on the
first factor. This analysis revealed a significant main ef-
fect for time, indicating that discriminant accuracy de-
creased as the length of prediction increased, F(2, 127) =
30.52, p < .01, η
= .32. In addition, a marginally signifi
cant main effect for information indicated that partici
pants who received relevant information were somewhat
more accurate than were participants who received irrel
evant information, F(1, 124) = 2.84, p = .09, η
= .02. As
can be seen in Figure 2, participants provided with rele
vant information tended to be more accurate than those
provided with irrelevant information, especially as the
length of time involved in the prediction increased. In
contrast, an optimistic versus objective construal had no
influence on discrimination accuracy, F < 1.
These results are consistent with Studies 1 through 3
and provide further evidence that discrimination is a
function of the diagnosticity of relevant information.
Bias, in contrast, appears to be a function of the way that
information is construed. Participants in this experi
ment asked to interpret relevant information optimisti
cally were biased to the same degree, and in the same di
rection, as self-predictions in Study 3. Self-knowledge
provides a mixed blessing, increasing discrimination but
also bias in predictions of ethical or desirable behavior.
Predictions of the future are important because they
motivate action and influence decisions. At least in part,
people choose to marry because they believe their cur-
rent feelings will be long lasting, change careers because
they believe their life will improve, or tackle projects be
cause they seem manageable. Making these predictions
accurately is therefore a critical prerequisite for making
a variety of good decisions. Despite this importance,
many predictions contain reliable biases. In short, there
is ample confirmation of Yogi Berra’s hypothesis that
predictions are difficult to make, especially about the
The research reported in this article explores one rea
son why, at times, predictions are difficult to make, espe
cially about the self. This reason focuses on the approach
people take when predicting their own versus others’ be
havior. When predicting one’s own behavior, people
seem to rely on specific self-knowledge they possess
about themselves, leading them to disregard intuitive as
sessments of population base rates that might otherwise
enhance the accuracy of their predictions. The opposite
happens, however, when predicting random others sim
ply because no case-based evidence is available. This
predicts that self-assessments will sometimes be less ac
curate, by exhibiting more mean-level bias, than peer-
Figure 2 Percentage of correct predictions (i.e., discrimination) among participants who receive relevant or irrelevant individuating information
that they interpret optimistically or objectively (Study 4).
predictions. This prediction was confirmed in Studies 1
through 3, in which participants’ self-predictions were
optimistically biased when compared to actual behavior.
Peer predictions, in contrast, were unbiased. In this re
spect, people appear to have considerably more cali
brated insight into others’ behavior than into their own.
But this research also suggests that the relationship be
tween self-knowledge and accurate predictions is some
what complicated. Although self-knowledge may be reli
ably biased, it may still afford the ability to discriminate
between likely and unlikely behavior at the individual
level. This discriminant accuracy would by definition be
nonexistent when predicting random others for which
no case-based information is available. Studies 1 through
3 also found evidence for this relative difference in
discriminant accuracy. Despite optimistic biases overall,
participants were more accurate predicting their own
behavior than predicting a random other as well as more
accurate than a simple base rate prediction would have
produced. Evidence consistent with this analysis also co
mes from the individuated participants in both Studies 1
and 2. Providing people with inside information about
another person increased accuracy only when the infor-
mation was directly relevant to the prediction. Indeed,
peer predictions involving relevant information became
roughly as accurate as self-predictions. The reason it did
not simultaneously produce the same kinds of biases as
self-knowledge, we suspect, is because it was not as op-
timistically biased as self-knowledge in these domains.
Beyond providing a clearer account of accuracy in
self- and social predictions, the findings reported in this
article also help to bring together two lines of research
that have long operated in relative isolation. On one side
are researchers primarily interested in the determinants
of accurate social judgments. This interest goes back at
least as far as Darwin, who ignited research in emotion
recognition that has continued ever since (Zajonc,
1998). The interest in accuracy has expanded to include
everything from judgments about clinical treatment
(Dawes, Faust, & Meehl, 1989) to self-presentation
(Kenny & DePaulo, 1993) to sexual orientation (Ambady,
Hallahan, & Conner, 1999). Generally, the methods
used in this research are correlational and investigate
whether variability in the world is tracked by the variabil
ity in intuitive judgments.
On the other side are researchers generally interested
in the determinants of biases in social judgment. This re
search began in earnest with a monograph by Icheiser
(1949) that foreshadowed the development of the
heuristics and biases approach to social judgment
(Gilovich, Griffin, & Kahneman, 2002; Kahneman,
Slovic, & Tversky, 1982; Ross, 1977). This research was
less focused on whether judgments tracked the variabil
ity in real life but rather how any one judgment matched
up with reality, often indexed as a difference score. Psy
chologists in this tradition learned, among many other
things, that people are prone to overestimate the power
of dispositions in causing behavior (Gilbert & Malone,
1995), to reason egocentrically (Nickerson, 1999), and
to utilize simple rules and shortcuts that occasionally
lead to systematic errors (Kahneman et al., 1982). Al
though this research was (and is) motivated by an inter
est in the underlying mechanisms that guide human
judgment rather than a crusade in pursuit of error and
human shortcomings (Krueger & Funder, 2004), these
mechanisms are generally revealed only through the
demonstration of predictable biases in judgment.
At first glance, these two literatures seem diametri
cally opposed—either judgments are right or they are
wrong and the study of accuracy may bear little on the
study of error. However, the two are often close traveling
companions, and the findings presented here suggest
that whether researchers uncover a picture of accuracy
or error often depends on how they look. These results
demonstrate that the same information may produce, at
the same time, both accuracy and error for different but
related reasons. We suspect similar mechanisms under-
lie other observations of correlational accuracy in the
face of overall error or bias, including those involving
perceptions of others’ impressions (Epley et al., 2002;
Savitsky, Epley, & Gilovich, 2001), affective forecasting
(Wilson & Gilbert, 2003), and the planning fallacy
(Buehler et al., 1994). Although the common trajectory
of ideas in science is toward ever-finer distinctions and
specializations, we think the study of accuracy and error
in human judgment is ready to move in reverse, bringing
these separate camps under the same umbrella and
commencing integration.
The research reported here suggests that a central
component of this integration may involve simulta
neously studying multiple indexes of accuracy. Research
ers cannot examine only one form of accuracy (e.g., a
correlation coefficient) and then assume that the accu
racy displayed by that one indicator is an adequate proxy
for all. Instead, researchers must think through the vari
ous forms of accuracy that might be relevant to the cir
cumstances being studied. We have focused in this re
search on two forms (discrimination and bias), but
others have described additional variants of accuracy
that may be important to consider as well. Cronbach
(1955), for instance, described four different measures
of accuracy in his classic article. In more recent work,
Kenny (2004) has described how accuracy in social judg
ment can be measured from the perspective of the
perceiver, the perceived, and the interaction between
the two. Thus, the findings described herein suggest that
Epley, Dunning / SELF-PREDICTION 653
future researchers interested in accuracy and error
should be mindful of the specific forms of accuracy they
are measuring and whether they are being comprehen
sive in their treatment of the concept.
One major issue facing this integration, however, that
the present research leaves unaddressed is the func
tional benefit of different forms of accuracy in daily life.
Patterns of accuracy may differ between self- versus social
predictions because they serve different functions. Self-
predictions may serve to regulate moods, enhance cre
ativity, or facilitate goal attainment, and optimistic biases
constrained by discrimination accuracy may prove the
most beneficial in satisfying these functions (e.g., Taylor,
1989). Peer predictions, however, may serve to antici
pate others’ actions as accurately as possible, and unbi
ased estimates may prove to be the most beneficial in the
absence of any other information. However, the actual
functional benefits of these patterns of accuracy remains
largely unexplored and this issue deserves further study.
Whether people are truly benefited by optimistic dis
tortions about the self but unbiased assessments of
others remains to be seen (Dunning, 2005).
Regardless of the adaptive consequences of different
forms of accuracy in self versus social predictions, our re-
search suggests several practical strategies for those in-
terested in accurately predicting future behavior. A per-
son interested in accurately predicting how a group of
people will respond overall to a social event, such as a
charity drive or an upcoming election, would be well ad-
vised to ask people to predict how others would behave.
A person interested in identifying how specific groups of
individuals will behave relative to other groups, however,
should ask people to predict their own behavior. No per
son will ever be able to predict the future as accurately
as they can report on the present, but being mindful of
the two forms of accuracy studied here, and the extent
to which self- versus social predictions contain each
type of accuracy, will make the future as clear as it can
possibly be.
1. What we call “discrimination” also has been referred to as both
“resolution” (e.g., Brier, 1950; Liberman & Tversky, 1993; Murphy,
1973) and simply “accuracy” (e.g., Ambady, Bernieri, & Richeson,
2000; Gagné & Lydon, 2004; MacDonald & Ross, 1999). The latter term
is unfortunate because accuracy in everyday discourse is used to
describe both high levels of discrimination and low levels of bias, rather
than the precise statistical definitions used here.
2. Our hypotheses required a longitudinal design comparing pre
dictions of self and others to actual behavior and therefore introduced
the potential problem that predicting one’s behavior may lead people
to behave in a manner consistent with their prediction (Greenwald,
Carnot, Beach, & Young, 1987; Sherman, 1980). We took two steps to
address this influence. First, predictions were made far in advance of
actual behavior to diminish the likelihood of contamination. Second, a
supplemental study allowed us to investigate directly the extent to
which predictions might have influenced behavior in Studies 1 and 2.
One month before the 2000 presidential election, 141 eligible voters in
an introductory psychology class were asked to predict whether they
would vote. Although 85% predicted they would vote, only 64%
reported voting 2 days after the election. An additional 32 eligible vot
ers, who for whatever reason did not provide predictions, also reported
whether they voted after the election. Of these, 62% voted in the elec
tion, a figure that does not differ statistically from the percentage of
people who voted after predicting their behavior, χ
< 1.
Ambady, N., Bernieri, F., & Richeson, J. A. (2000). Towards a histology
of social behavior: Judgmental accuracy from thin slices of behav
ior. In M. Zanna (Ed.), Advances in experimental social psychology
(Vol. 32). San Diego, CA: Academic Press.
Ambady, N., Hallahan, M., & Conner, B. (1999). Accuracy of judg
ments of sexual orientation from thin slices of behavior. Journal of
Personality and Social Psychology, 77, 538-547.
Bass, B. M., & Yammarino, F. J. (1991). Congruence of self and others
leadership ratings of naval officers for understanding successful
performance. Applied Psychology, 40, 437-454.
Brier, G. W. (1950). Verification of forecasts expressed in terms of
probability. Monthly Weather Review, 75, 1-3.
Buehler, R., Griffin, D., & Ross, M. (1994). Exploring the “planning
fallacy”: Why people underestimate their task completion times.
Journal of Personality and Social Psychology, 67, 366-381.
Campbell, D. T., & Kenny, D. A. (1999). A primer of regression artifacts.
New York: Guilford.
Cronbach, L. J. (1955). Processes affecting scores on “understanding
of others” and “assumed similarity.” Psychological Bulletin, 52, 177-
Dawes, R. M., Faust, D., & Meehl, P. E. (1989, March 31). Clinical ver-
sus actuarial judgment. Science, 243, 1668-1674.
Dunning, D. (1999). A newer look: Motivated social cognition and
the schematic representation of social concepts. Psychological
Inquiry, 10, 1-11.
Dunning, D. (2005). Self-insight: Roadblocks and detours on the path to
knowing thyself. New York: Psychology Press.
Dunning, D., Griffin, D. W., Milojkovic, J. D., & Ross, L. (1990). The
overconfidence effect in social prediction. Journal of Personality
and Social Psychology, 58, 568-581.
Dunning, D., Heath, C., & Suls, J. M. (2004). Flawed self-assessment:
Implications for health, education, and the workplace. Psychologi
cal Science in the Public Interest, 5, 69-106.
Epley, N., & Dunning, D. (2000). Feeling “holier than thou”: Are self-
serving assessments produced by errors in self or social psychol
ogy? Journal of Personality and Social Psychology, 79, 861-875.
Epley, N., Savitsky, K., & Gilovich, T. (2002). Empathy neglect: Recon
ciling the spotlight effect and correspondence bias. Journal of Per
sonality and Social Psychology, 83, 300-312.
Funder, D. C., & Colvin, C. R. (1997). Congruence of others’ and self-
judgments of personality. In R. Hogan, J. Johnson, & S. Briggs
(Eds.), Handbook of personality psychology (pp. 617-647). San Diego,
CA: Academic Press.
Gagné, F. M., & Lydon, J. E. (2004). Bias and accuracy in close rela
tionship: An integrative review. Personality & Social Psychology Re
view, 8(4), 322-338.
Gilbert, D. T., & Malone, P. S. (1995). The correspondence bias. Psy
chological Bulletin, 117, 21-38.
Gilovich, T., Griffin, D., & Kahneman, D. (2002). Heuristics and biases:
The psychology of intuitive judgment. Cambridge, UK: Cambridge
University Press.
Gilovich, T., Medvec, V. H., & Savitsky, K. (2000). The spotlight effect
in social judgment: An egocentric bias in estimates of the salience
of one’s own actions and appearance. Journal of Personality and
Social Psychology, 78, 211-222.
Green, D. M., & Swets, J. A. (1966). Signal detection theory and
psychophysics. New York: John Wiley.
Greenwald, A. G., Carnot, C. G., Beach, R., & Young, B. (1987).
Increasing voting behavior by asking people if they expect to vote.
Journal of Applied Psychology, 72, 315-318.
Hansford, B. C., & Hattie, J. A. (1982). The relationship between self
and achievement/performance measures. Review of Educational
Research, 52, 123-142.
Hasher, L., Zacks, R. T., Rose, K. C., & Sanft, H. (1987). Truly inciden
tal encoding of frequency information. American Journal of Psychol
ogy, 100, 69-91.
Icheiser, G. (1949). Misunderstandings in human relations: A study
in false social perception. American Journal of Sociology, 55(Part 2),
Kahneman, D., Slovic, P., & Tversky, A. (1982). Judgment under uncer
tainty: Heuristics and biases. Cambridge, UK: Cambridge University
Kahneman, D., & Tversky, A. (1973). On the psychology of predic
tion. Psychological Review, 80, 313-327.
Kahneman, D., & Tversky, A. (1979). Intuitive prediction: Biases and
corrective procedures. TIMS Studies in Management Science, 12, 313-
Kenny, D. A. (2004). PERSON: A general model of interpersonal per
ception. Personality and Social Psychology Review, 8, 265-280.
Kenny, D. A., & DePaulo, B. M. (1993). Do people know how others
view them? An empirical and theoretical account. Psychological
Bulletin, 114, 145-161.
Koehler, D. J. (1996). The base rate fallacy reconsidered: Descriptive,
normative, and methodological challenges. Behavioral and Brain
Sciences, 19, 1-53.
Krueger, J. I., & Funder, D. C. (2004). Towards a balanced social
psychology: Causes, consequences, and cures for the problem-
seeking approach to social behavior and cognition. Behavioral &
Brain Sciences, 27(3), 313-327.
Kunda, Z. (1990). The case for motivated reasoning. Psychological Bul-
letin, 108, 480-498.
Lewis, C. S. (1952). Mere Christianity. New York: Macmillan.
Liberman, V., & Tversky, A. (1993). On the evaluation of probability
judgments: Calibration, resolution, and monotonicity. Psychologi-
cal Bulletin, 114, 162-173.
Mabe, P. A., & West, S. G. (1982). Validity of self-evaluation of ability:
A review and meta-analysis. Journal of Applied Psychology, 67, 280-
MacDonald, T. K., & Ross, M. (1999). Assessing the accuracy of pre-
dictions about dating relationships: How and why do lovers’ pre-
dictions differ from those made by observers. Personality and Social
Psychology Bulletin
, 25, 1417-1429.
Murphy, A. H. (1973). A new vector partition of the probability score.
Journal of Applied Meteorology, 12, 595-600.
Nickerson, R. S. (1999). How we know—and sometimes misjudge—
what others know: Imputing one’s own knowledge to others. Psy
chological Bulletin, 125, 737-759.
Nisbett, R. E., & Kunda, Z. (1985). Perception of social distributions.
Journal of Personality and Social Psychology, 48, 297-311.
Risucci, D. A., Tortolani, A. J., & Ward, R. J. (1989). Ratings of surgical
residents by self, supervisors, and peers. Surgery, Gynecology, and
Obstetrics, 169, 519-526.
Ross, L. (1977). The intuitive psychologist and his shortcomings: Dis
tortions in the attribution process. In L. Berkowitz (Ed.), Advances
in experimental social psychology (Vol. 10, pp. 174-221). New York:
Academic Press.
Savitsky, K., Epley, N., & Gilovich, T. (2001). Is it as bad as we think?
Overestimating the impact of our failures, shortcomings, and mis
haps. Journal of Personality and Social Psychology, 81, 44-56.
Sherman, S. J. (1980). On the self-erasing nature of errors of predic
tion. Journal of Personality and Social Psychology, 39, 211-221.
Shrauger, J. S., & Osberg, T. M. (1981). The relative accuracy of self-
predictions and judgments by others in psychological assessment.
Psychological Bulletin, 90, 322-351.
Shrauger, J. S., & Osberg, T. M. (1982). Self-awareness: The ability to
predict one’s future behavior. In G. Underwood (Ed.), Aspects of
consciousness (Vol. 3, pp. 267-313). London: Academic Press.
Taylor, S. E. (1989). Positive illusions: Creative self-deception and the
healthy mind. Hillsdale, NJ: Lawrence Erlbaum.
Wilson, T. D., & Gilbert, D. T. (2003). Affective forecasting. In
M. Zanna (Ed.), Advances in experimental social psychology (Vol. 35,
pp. 345-411). New York: Elsevier.
Yaniv, I., Yates, J. F., & Smith, J. E. K. (1991). Measures of discrimina-
tion skill in probabilistic judgment. Psychological Bulletin, 110, 611-
Zajonc, R. B. (1998). Emotions. In D. Gilbert, S. T. Fiske, & G. Lindzey
(Eds.), Handbook of social psychology (4th ed., Vol. 1, pp. 591-632).
New York: McGraw-Hill.
Received April 13, 2005
Revision accepted August 25, 2005
Epley, Dunning / SELF-PREDICTION 655
... Such finer-grained time-indexed trajectories of optimism may tell us more about an individual's or group's future-oriented attitudes and behaviors. However, asking people to self-report how their optimism differs towards next week, next month, and next year may not produce behaviorally accurate or stable results due to common validity and reliability issues of self-report scales (Dunning et al., 2004;Epley & Dunning, 2006). Here, we aim to measure and understand the optimism trajectory of large populations by addressing these challenges through the use of big data from social media that captures multiple time-horizons without self-report. ...
Individuals can hold contrasting views about distinct times: for example, dread over tomorrow's appointment and excitement about next summer's vacation. Yet, psychological measures of optimism often assess only one time point or ask participants to generalize about their future. Here, we address these limitations by developing the optimism curve, a measure of societal optimism that compares positivity toward different future times that was inspired by the Treasury bond yield curve. By performing sentiment analysis on over 3.5 million tweets that reference 23 future time points (2 days to 30 years), we measured how positivity differs across short-, medium-, and longer-term future references. We found a consistent negative association between positivity and the distance into the future referenced: From August 2017 to February 2020, the long-term future was discussed less positively than the short-term future. During the COVID-19 pandemic, this relationship inverted, indicating declining near-future- but stable distant-future-optimism. Our results demonstrate that individuals hold differentiated attitudes toward the near and distant future that shift in aggregate over time in response to external events. The optimism curve uniquely captures these shifting attitudes and may serve as a useful tool that can expand existing psychometric measures of optimism.
... In peer estimation questions, we asked participants to make inferences for their average peer, as opposed to giving specific direction regarding the type of peer they should consider. This framing is used in existing work examining social comparisons (Epley & Dunning, 2006) and avoids the potential influence of individual (e.g., selection of peers more or less concerned about social issues) and interpersonal (e.g., closeness) differences on the appraisal of others' behavior (Kelley, 1973). All code, deidentified data, supplemental analyses, and measures are publicly available1. ...
Preventing the negative impacts of major, intersectional social issues hinges on personal concern and willingness to take action. This research examines social comparison in the context of climate change, racial injustice, and COVID‐19 during Fall 2020. Participants in a U.S. university sample (n = 288), reported personal levels of concern and action and estimated peers' concern and action regarding these three issues. Participants estimated that they were more concerned than peers for all three issues and took more action than peers regarding COVID‐19 and climate change. Participants who reported higher levels of personal concern also estimated that they took greater action than peers (relative to participants who reported lower levels of concern). Exploratory analyses found that perceived personal control over social issues were associated with greater concern and action for racial injustice and climate change but not for COVID‐19. This indicates that issue‐specific features, including perceived controllability, may drive people to differently assess their experiences of distinct social issues.
... After all, if you simply do not know what you do not know, what role could there be for individual differences to shift when and to whom the DKE happens? While metacognitive processes alone are typically presumed to explain much of the DKE, careful experimental work has demonstrated that other processes are also at play (Epley & Dunning, 2006;Sheldon, Dunning, & Ames, 2014;Simons, 2013;Williams et al., 2013. Williams et al. (2013 demonstrated that people who were experimentally induced to conceive of their cognitive approach as rational and consistent were more likely to display the DKE. ...
Two studies examined whether Intellectual Humility moderates the Dunning-Kruger Effect – the tendency of those with little relative knowledge/expertise to overclaim their ability. In study 1, participants completed a measure of intellectual humility and were asked to predict their relative performance prior to a test of fluid intelligence, and then their actual performance was compared to their initial prediction and ranked relative to the other performances. Low performers tended to over-estimate their performance (i.e., the Dunning-Kruger effect), but less intellectually humble people demonstrated a greater susceptibility to this effect. Intellectual humility was unrelated to actual performance. These results were replicated in study 2, with a test of general knowledge in place of problem-solving ability. Additionally, in Study 2 more measures of intellectual humility were added to clarify potential relationships. Implications are discussed.
... Studies find that our comparative estimates of our chances of experiencing particular life events are strongly correlated with the valence of the event in question. For example, college-aged students estimate their chances of owning a home to be 44% higher than the average chance of their classmates and their chance of getting divorced as 48% lower than their average classmate (Eppley andDunning 2006). These differences are significantly correlated neither with the general probability of the event in question, the perceived general probability of the event, nor the subject's "personal experience with the type of event in ...
Epistemologists frequently claim that the question “What should I believe?” demarcates the field of epistemology. This question is then compared to the question asked in ethics: “What should I do?” The question and the ensuing comparison, it is thought, specify both the content and the normativity at stake in epistemology. I argue that both of the assumptions embedded in this demarcation are problematic. By thinking of epistemology’s focal question in this light, first, we risk importing our assumptions about the epistemic domain into our understanding of the nature and normativity of the belief state, and second, we come to have a false picture of the normativity that supposedly underlies the domain. In Chapter 1, “The Doxastic Account of the Epistemic”, I explore a range of views that assume there to be an essential connection between belief and truth. I look at views that treat all beliefs as attempts to believe the truth, views that consider belief’s biological function to be accurate representation, and views that hold that the very concept of belief is a normative concept. I go on to explore instrumentalist conceptions of belief’s truth connection and conduct an inquiry into the value of true belief. I conclude that neither the value of true belief nor an essential connection between belief and truth can explain epistemic normativity. In Chapter 2, “Evidential Exclusivity, Correctness, and the Nature of Belief” I note that epistemologists have recently argued that the best explanation for the apparent truth of a pair of claims - “Transparency” and “Exclusivity” – is that belief is subject to a standard of correctness such that a belief that p is correct if and only if p is true. I argue that the proposed explanation unduly privileges one part of belief’s full functional profile – its role in deliberation – and that a more complete consideration of belief’s role in cognition suggests an alternative explanation for Exclusivity and Transparency but denies belief’s standard of correctness. In Chapter 3, “Tradeoffs and Epistemic Value”, I look at a debate about whether epistemic norms are teleological. Though it’s standard to assume in keeping with teleology that certain goals or values explain the content of our norms, a collection of recent papers have aimed to show that this assumption can’t be correct because teleological norms countenance tradeoffs but epistemic norms don’t countenance tradeoffs. I note that the kind of non-teleological view that countenances no tradeoffs whatsoever is actually quite extreme and virtually unheard of in ethics. I go on to make the case that norms that license no tradeoffs can’t reasonably be taken to be grounded in value at all, and thus can’t be understood to give rise to necessary normativity. I conclude by suggesting that we broaden our conception of the epistemic domain to recognize teleological norms that provide recommendations for methods of inquiry or pursuit of significant truth or knowledge.
... Systematic miscalibration between the expressers' expectations and the recipients' actual experiences does not necessarily mean that the expressers have no insight in their recipients' experience. They could still have some above-chance understanding of how their own recipient will respond compared to other recipients, resulting in a strong correlation between expressers' and recipients' ratings even if they are systematically underestimating the overall positivity of their recipient's experience (Epley & Dunning, 2006;Gagne & Lydon, 2004). However, the correlations between expressers' and recipients' ratings provided only modest support for such discrimination accuracy among expressers. ...
Full-text available
Compliments increase the well-being of both expressers and recipients, yet people report in a series of surveys giving fewer compliments than they should give, or would like to give. Nine experiments suggest that a reluctance to express genuine compliments partly stems from underestimating the positive impact that compliments will have on recipients. Participants wrote genuine compliments and then predicted how happy and awkward those compliments would make recipients feel. Expressers consistently underestimated how positive the recipients would feel but overestimated how awkward recipients would feel (Experiments 1-3, S4). These miscalibrated expectations are driven partly by perspective gaps in which expressers underestimate how competent—and to a lesser extent how warm—their compliments will be perceived by recipients (Experiments 1-3). Because people’s interest in expressing a compliment is partly driven by their expectations of the recipient’s reaction, undervaluing a compliment creates a barrier to expressing them (Supplemental Experiments S2, S3, S4). As a result, directing people to focus on the warmth conveyed by their compliment (Experiment 4) increased interest in expressing it. We believe these findings may reflect a more general tendency for people to underestimate the positive impact of prosocial actions on others, leading people to be less prosocial than would be optimal for both their own and others’ well-being.
Individuals' psychological distress is associated with disinhibited eating (external and emotional eating). The aim of the current study was to examine the moderating associations of COVID-19-related stress on parents' psychological distress (anxiety, hostility, depression) and external and emotional eating. One hundred and sixty U.S. parents of three- to five-year-old children (Mage = 34.08, SD = 6.76; 89 females) completed an online survey. After accounting for participant characteristics (i.e., age, BMI, sex), regression analyses showed that COVID-19 stress moderated the effects of anxiety, hostility, and depression on external eating. Additionally, findings showed that COVID-19 stress moderated hostility (but not anxiety or depression) on emotional eating. These findings suggest that unexpected stressors from the COVID-19 pandemic may exacerbate disinhibited eating among those individuals who experience psychological distress. This presents support for providing interventions that focus on healthy coping strategies and family well-being, support groups, and community resources (e.g., financial assistance) to alleviate external pressures during unprecedented times.
We analyze positional concerns in the unethical domain. We introduce an original distinction between “selective” positionality—where individuals prefer behaving unethically but to a lesser extent than peers—and “ego” positionality—where they prefer behaving unethically but to a higher extent than peers, regardless of the absolute level. We also report the results of an exploratory survey in Algeria that exploits the counterintuitive insight that people are better at predicting others’ behaviors than their own behaviors. We increase the finding’s generalizability by conducting the same survey among a similar sample in France. Our findings are twofold: first, the majority of participants attributes to others preferences for ethical (i.e., where everyone is honest) and unethical egalitarian (where all are similarly dishonest) situations. Second, a non-negligible proportion of respondents attributes to the average individual preferences for either selective or ego-positionality in unethical behaviors.
Behavioral Economics: Evidence, Theory, and Welfare provides an engaging and accessible introduction to the motivating questions, real-world evidence, theoretical models, and welfare implications of behavioral economics concepts. Applications and examples illustrate the broad relevance of behavioral economics for consumers, firms, markets, and policy makers alike. The book highlights the process by which economists evaluate evidence and disentangle theories with different social welfare implications. Accessible to students from diverse economic backgrounds, this textbook is an ideal resource for courses on behavioral economics, experimental economics and related areas.
Self-knowledge includes not only beliefs about one’s own traits and abilities, but beliefs about how others view the self. Are such metaperceptions accurate? This article identifies two distinct standards used to determine meta-accuracy. The correlational approach tests whether metaperceptions correlate with an accuracy criterion (i.e., social perceptions). The mean-level approach instead asks whether metaperceptions tend to err in a systematic direction. This article reviews complementary lessons gleaned from research taking one approach or the other: whether metaperceptions merely reflect self-perceptions, whose metaperceptions are more or less accurate, and what psychological processes impede meta-accuracy, among others. Ultimately, neither approach is endorsed as unconditionally superior. Instead, which approach offers the proper accuracy standard should depend on the decisions those metaperceptions will guide.
Full-text available
To communicate effectively, people must have a reasonably accurate idea about what specific other people know. An obvious starting point for building a model of what another knows is what one oneself knows, or thinks one knows. This article reviews evidence that people impute their own knowledge to others and that, although this serves them well in general, they often do so uncritically, with the result of erroneously assuming that other people have the same knowledge. Overimputation of one's own knowledge can contribute to communication difficulties. Corrective approaches are considered. A conceptualization of where own-knowledge imputation fits in the process of developing models of other people's knowledge is proposed.
Many decisions are based on beliefs concerning the likelihood of uncertain events such as the outcome of an election, the guilt of a defendant, or the future value of the dollar. Occasionally, beliefs concerning uncertain events are expressed in numerical form as odds or subjective probabilities. In general, the heuristics are quite useful, but sometimes they lead to severe and systematic errors. The subjective assessment of probability resembles the subjective assessment of physical quantities such as distance or size. These judgments are all based on data of limited validity, which are processed according to heuristic rules. However, the reliance on this rule leads to systematic errors in the estimation of distance. This chapter describes three heuristics that are employed in making judgments under uncertainty. The first is representativeness, which is usually employed when people are asked to judge the probability that an object or event belongs to a class or event. The second is the availability of instances or scenarios, which is often employed when people are asked to assess the frequency of a class or the plausibility of a particular development, and the third is adjustment from an anchor, which is usually employed in numerical prediction when a relevant value is available.
82 female undergraduates were assigned to 1 of 4 experimental groups--predict-request, information-request, predict only, and request only--in which requested tasks involved writing a counterattitudinal essay or singing over the telephone. In 3 experiments, Ss overpredicted the degree to which their behavior would be socially desirable and these errors of prediction proved to be self-erasing. Having mispredicted a given behavior, Ss were likely to have these predictions confirmed in later behavior, indicating that prediction of a behavioral sequence evokes a specific cognitive representation of that sequence which is subsequently accessed. Results demonstrate the strong effects on behavior of engaging in prebehavioral cognitive work. (27 ref) (PsycINFO Database Record (c) 2006 APA, all rights reserved).
In two longitudinal studies, university students, their roommates, and parents assessed the quality and forecast the longevity of the students’ dating relationships. The longitudinal nature of this research allowed assessment of the relative accuracy of predictions offered by students and observers. Students assessed their relationships more positively, focusing primarily on the strengths of their relationships, and made more optimistic predictions than did parents and roommates. Although students were more confident in their predictions, their explicit forecasts tended to be less accurate than those of the two observer groups. Students, however, possessed information that could have yielded more accurate forecasts: In comparison to parents’ and roommates’ evaluations of relationship quality, students’ assessments of relationship quality were more predictive of stability at 1 year. Implications of these findings for understanding biases and accuracy in prediction are discussed.