zyxwvu
zyx
zy
zy
zyxw
zyx
Journal of Behavioral Decision Making, Vol. I, 29-41 (1988)
Confidence in Judgments Based on
Incomplete Information: An Investigation
Using Both Hypothetical and Real Gambles
IRWIN P. LEVIN AND DANIEL P. CHAPMAN
The University of Iowa. U.S.A.
RICHARD D. JOHNSON
The University of Alberta, Canada
ABSTRACT
This study assessed how confidence in judgments is affected by the need to
make inferences about missing information. Subjects indicated their likelihood of taking each of a series of gambles based on both probability and
payoff information or only one of these sources of information. They also
rated their confidence in each likelihood judgment. Subjects in the Explicit
Inference condition were asked to explicitly estimate the values of missing
information before making their responses while subjects in the Implicit
Inference condition were not. The manner in which probability information
was framed was also manipulated. Experiment 1 employed hypothetical
gambles and Experiment 2 employed gambles with real money. Expressed
likelihood of taking gambles was higher when probability was phrased in
terms of ‘% chance of winning’ rather than ‘% chance of losing’, but this
difference was somewhat less with real gambles than with hypothetical
gambles. Confidence ratings in each experiment were actually higher on
incomplete information trials than on complete information trials in the
Explicit Inference condition. Results were related to the general issue of
confidence in judgments.
zyxwvu
KEY WORDS
Confidence in judgments
Framing effects
Risky decision making
Inferences
Most judgments and decisions in the real world must be made with limited information. It is rarely the
case that values are known for all the important attributes. Earlier work in our laboratory and in the
laboratories of others suggests that the missing information is not always ignored (Huber and McCann,
1982; Johnson and Levin, 1985; Levin, Johnson and Faraone, 1984; Singh, 1984; Yamagishi and Hill,
1981). Decision makers can use their prior knowledge to infer the value of missing information. For
example, when the advertisements for a product stress its quality features without mentioning cost,
most of us would infer a high price tag.
Studies using the ‘missing information paradigm’ have potential use to both Psychology and
Marketing Science. They can help us gain insight into inference processes evoked when decision makers
zyxwvuts
*The authors would like to thank Mary Snyder, Karen Rego, Elizabeth Vera, and Ross Dickerson for their assistance.
0894-32571 88/010029%13$06.50
0 1988 by John Wiley & Sons, Ltd.
Received 30 May 1987
Revised 27 October 1987
30
zyxwvutsrqp
zyx
zyxw
zyx
Journal of Behavioral Decision Making
Vol. I, Iss. No. I
are faced with insufficient information to make a rational decision. They can also show the extent to
which consumers are affected by the incomplete disclosure of information that is found in many
advertisements.
The present study departs from the earlier studies in three ways. First, the inference process is
brought under experimental scrutiny by including a condition where subjects are asked to explicitly
assign values to missing information before making their judgments. Second, because the need to deal
with missing information may affect confidence in judgments, confidence ratings are added to the
judgment task. Third, tests of the external validity of laboratory results are included.
The use of confidence ratings serves to link the study of inference processes in judgment to the more
general area of confidence in judgment. Koriat, Lichtenstein and Fischhoff (1980) report that
confidence is affected by the amount and strength of information supporting the decision made.
However, overconfidence in judgments is a common finding (Fischhoff, Slovic, and Lichtenstein, 1977).
In particular, overconfidence in responding to general knowledge questions may be the result of
accepting one’s own fallible associations or inferences linking aspects of the question to the selected
answer. For example, consider a typical question used by Fischhoff et a1 (1977): are potatoes native to
Ireland or Peru? The association ‘Irish-potato’ leads to high confidence in what is, in fact, an incorrect
response. The present study provides a direct look at how the need to make subjective inferences affects
confidence in judgments.
One additional variable shown to be important in earlier studies of judgments and decisions based on
varying amounts of information is also included. Levin, Johnson, Russo and Deldin (1989, and Levin,
Johnson, Deldin, Carsten, Cressey and Davis (1986) showed that the manner in which information is
framed - for example, whether gambles are described in terms of probability of winning or probability
of losing - affects the evaluation of individual choice options and can lead to preference reversals in
choosing between choice options with complete and incomplete information. While inclusion of the
framing manipulation is of only secondary interest in the present study, it does represent the first
attempt to examine how confidence in judgments is affected by this important contextual variable.
The goals of the present study can be summarized as follows: (1) to examine how confidence in
judgments is affected by the absence of a key source of information; (2) to compare conditions where
subjects are or are not required to make explicit inferences of missing information values before making
their judgments; (3) to replicate and extend earlier results of how information frame affects judgments
with complete and incomplete information; and (4) to examine the external validity of laboratory
results by including a ‘real-play’ procedure where subjects actually gamble for money.
An earlier study by Levin and Johnson (1987) of price-quality tradeoffs in consumer judgments
addresses some of these issues. Some subjects were asked t o impute values to missing quality
information before rating the desirability of hypothetical purchases of ground (minced) beef. In
contrast to control subjects, those subjects who made explicit inferences showed a tendency to give
higher desirability ratings to high price than to low price purchases when information about quality was
missing. The tendency observed in the Explicit Inference condition was mediated by the perception that
higher prices go with higher quality, and that quality is more important than price for these purchases.
It appears from these results that when subjects were asked to estimate missing values, they placed great
confidence in their estimates and treated them as if they were veridical.
Results of the Levin and Johnson (1987) study lead us to predict that confidence in judgments will
not decrease on missing-information trials when subjects are explicitly asked to estimate missing values.
Moreover, to the extent that subjects make inferences without being explicitly asked and they accept
these inferences, confidence may be high on missing information trials even when explicit inferences are
not required. The current study examines more directly the issue of how the absence of a key source of
information affects subjects’ confidence in their judgments. However, the price-quality judgment task of
the Levin and Johnson study produced such a strongly perceived interdimensional relationship
zyxwvutsrq
zyxw
zyxwv
I. P. Levin et al.
Confidence in Judgments
zy
zy
31
mediating the imputation of missing information values that the generality of its results may be limited.
Hence, a different task was used in the current study where subjects would not presumably have as
strong a basis for imputing values to missing information. A gambling task was used, both in
Experiment 1 where hypothetical investments and payoffs were employed, and in Experiment 2 where
subjects were gambling for real money.
EXPERIMENT 1
Method
zyxwvutsrq
Experimental design
Subjects were asked to evaluate each of a series of 38 gambles presented in a booklet. Evaluations
included both ajudgment of the likelihood of taking each gamble and a rating of the confidence in that
judgment. The gambles were described by both probability and payoff information or by only one of
these sources of information. Within each of two randomly-ordered replications of the experimental
design, 12 gambles were formed by providing all combinations of four levels of probability and three
levels of payoff. The remaining seven gambles were described by only one piece of information: one of
the four levels of probability or one of the three levels of payoff. Some subjects were instructed to
estimate the value of missing information on these latter trials (the Explicit Inference condition) and
others were not so instructed (the Implicit Inference condition).
Probability information was presented as ‘% chance of winning’ in the Positive condition and as ‘%
chance of losing’ in the Negative condition. Percentages ranged from 5% to 20% chance of winning in
the Positive condition or 80% to 95% chance of losing in the Negative condition. Payoff amount varied
from $100 to $200 (see Exhibit I for exact factor levels and combinations).
zyxwvu
zyxw
Instructions and procedure
Subjects were told that each gamble would require a $15 investment; if they decide to take a gamble,
that means that they would invest the $15 to possibly win a larger amount; if they decide to refuse a
gamble, that means that they would keep the $15 and not invest it. Subjects were asked to rate how
likely they would be to take or refuse each gamble, using a seven-category checklist ranging from ‘very
likely to refuse gamble’ to ‘very likely to take gamble’. For data analysis purposes, responses were
scored on a scale from -3 to +3.
Subjects in the Explicit Inference condition were given the following additional instructions:
‘Consider the cases where only the amount you might win or only the chance of winning (losing) is
given. Before making your rating response in these cases, we want you to write down a value for the
missing information that would be reasonable, based on the presented information’. Subjects in the
Implicit Inference condition were not given these instructions, but were merely asked to respond on the
basis of the information presented. These two sets of instructions were designed to provide two
contrasting conditions for varying the extent to which subjects rely on their own inferences in making
judgments.
Subjects in both conditions were given the following instructions concerning the rating of confidence
in each likelihood judgment: ‘Regardless of whether your response was favorable, unfavorable, or
neutral, there may be times when you feel relatively confident in that response and times when you feel
relatively unconfident in that response. We want you to indicate the confidence you have in each of your
responses by indicating the extent to which you feel that you made the most apropriate response to that
gamble’.’ Subjects were told to use a scale of I to 20 to express their confidence.
32
zyxwvut
zyxwvutsrqp
zyx
zyxwvutsrq
zyxwvut
zyxwvuts
Journal of Behavioral Decision Making
Vol. I , Iss. No. I
A typical response booket page, illustrating stimulus presentation and response scales, is shown
below. This one is for a gamble in the Positive Framing condition with missing payoff information.
(The equivalent gamble in the Negative condition substitutes ‘Chance of losing: 90%’ for ‘Chance of
winning: 10%’).Subjects in the Explicit Inference condition filled in the blank space next to ‘Amount to
be won’ with their own estimate.
Amount of investment: $15
Amount to be won:
$-Chance of winning:
10%
Your response (check one):
__ very likely to take gamble
likely to take gamble
likely to take gamble
-- equally likely to take or refuse gamble
-- somewhat likely to refuse gamble
-- likely to refuse gamble
-- very likely to refuse gamble
--
-- somewhat
Degree of confidence in your response (circle one number):
1 2 3
very
unconfident
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
very
confident
Subjects
Subjects were 80 students in introductory psychology courses at The University of Iowa, 20 in each of
the four combinations of Inference condition and Framing condition. Subjects were tested in groups of
about five at a time. Trials in Replication 1 were considered to represent practice with the response
scales and familiarity with the range of stimuli. Only results for Replication 2 are reported.
Results and discussion
Likelihood of taking gambles
Exhibit 1 shows how the ratings of likelihood of taking gambles vary across framing and inference
conditions. Each combination of framing and inference condition is represented by a different panel in
Exhibit 1. Within each panel the solid lines plot the mean likelihood of taking gambles described by
both payoff and probability information. The dotted line plots mean likelihood of taking gambles for
which only probability information was given (amount of payoff missing) and the points to the right of
each panel plot mean likelihood of taking gambles for which only payoff amount information was given
(probability missing).
Across panels it can be seen that responses are predominantly negative, indicating a tendency to
refuse gambles, even though with a $15 investment most of the gambles have positive expected value.
An information framing effect can be seen by comparing the left and right panels. Ratings of the
likelihood of taking gambles were significantly less negative in the Positive Framing condition than in
the Negative Framing condition for both probability-payoff combinations (solid lines) and probabilityonly trials (dotted lines), F(1,76) = 5.73,p <0.05, and F(1,76) = 9.32, p <0.01, respectively. These
differences were somewhat greater in the Explicit Inference condition than in the Implicit Inference
condition but the Framing by Inference condition interaction did not reach statistical significance for
zyxwvut
zy
zyxwvutsrqp
zyxwvutsrq
zy
zyxwvut
zyxwv
zyxwvuts
Confidence in Judgments
I. P. Levin et al.
IMPLICIT COND: NEGATIVE FRAME
IMPLICIT COND: POSITIVE FRAME
I
,
I
I
6
10
16
20
33
96
Arnt-
only ($1
1
90
86
80
Arnt-
only ( S )
Chance of Winning ( % )
Chance of Losing ( % )
EXPLICIT COND: POSITIVE FRAME
EXPLICIT COND: NEGATIVE FRAME
2-
, 160
200
'
100
1-
0-
zyxwvutsrqpo
zyxwvutsrq
!
150
200
100
Prob.
-I-2-
150
100
200
160
100
-3-
I
r
1
6
10
16
20
Arntonly ($1
Chance of Winning (%)
96
90
86
80
Amt-
only ($1
Chance of Losing ( % )
Exhibit I . Mean likelihood of taking gambles as a function of probability and payoff information for each
combination of inference condition and framing condition in Experiment 1
either probability-payoff combinations or for probability-only trials. Differences between framing
conditions on payoff amount-only trials were neither predicted nor obtained because these trials did not
manipulate levels of the framing variable.
A comparison of the upper and lower panels of Exhibit 1 reveals how the likelihood of taking
gambles varies as a function of inference condition. There was no significant effect of inference
condition on these ratings for probability-payoff combinations. Of primary interest are comparisons
between trials with complete information (probability-payoff combinations) and trials with missing
information. In the Explicit Inference condition, ratings were significantly higher on missinginformation trials than on complete-information trials; t(39) = 3.64 for comparing mean rating averaged
over all probability-only trials with mean rating averaged over all probability-payoff trials, and t(39) =
6.04 for comparing mean rating averaged over all payoff amount-only trials with mean rating averaged
over all probability-payoff trials, p <0.01 in each case. Conversely, in the Implicit Inference condition
ratings were significantly lower on probability-only trials than on complete-information trials, t(39) =
3.76, p <0.01; and ratings were not significantly different on payoff amount-only trials and completeinformation trials.
34
zyxwvutsrqp
zyx
zy
zyx
Journal of Behavioral Decision Making
Vol. 1, Iss. No. 1
The effects in the Explicit Inference condition are mediated by optimistic imputations - the imputed
values of missing probability and payoff amount information tended to be more favorable than the
levels presented on the complete-information trials. However, neither the imputations of probability
from amount nor the imputations of amount from probability showed that subjects perceived either a
strong negative relationship or a strong positive relationship between probability and payoff. The near
parallelism of the dotted and solid lines in each panel of Exhibit I shows that the imputations of missing
values provide an additive constant to the evaluation of presented information.
Confidence ratings
Exhibit 2 compares confidence ratings for Explicit and Implicit Inference conditions, averaged over
framing conditions (which was not a significant factor nor did it enter into systematic interactions with
other factors). Confidence ratings average about 16 on a scale where 1 represents ‘very unconfident’ and
20 represents ‘very confident’. The highest levels of confidence tended to occur when the likelihood
ratings were the most extreme, which usually meant a response of likely or very likely to refuse gamble
when chances of winning were low.
While the mean confidence rating did not differ significantly across inference conditions, the key
comparison of confidence in judgments based on complete and incomplete information did reveal
differences between conditions. In each condition the mean confidence rating averaged over all
complete information trials (solid lines) was compared to the mean confidence rating averaged over all
19
160
100
zyx
IMPLICIT INFERENCE CONDITION
EXPLICIT INFERENCE CONDITION
200
1
zyxwvut
l5I
100
200
160
Prob.
Prob.
only
14
only
13
i
’*1
1
I
I
6/96
10190 16186 20180
Amtonly (tt
Chance of WinninglLosing (%)
zy
6/96
10/90
16/86 20180
Chance of WinningiLosing
Amtonly ( $ )
(%I
Exhibit 2. Mean confidence ratings as a function of probability and payoff information for each inference
condition in Experiment 1
zyxwvuts
zyxwvutsrqpo
zyxwvutsrq
zyx
zyxwvu
zyxw
zy
I. P. Levin et al.
Confidence in Judgments
35
payoff amount-only trials (points to the right of each panel). Confidence ratings were slightly - but not
significantly - lower on payoff amount-only trials than on complete-information trials in the Implicit
condition. Confidence ratings were significantly higher for payoff amount-only trials than for completeinformation trials in the Explicit condition, t(39) = 3.18,p<0.01. Mean confidence ratings on completeinformation trials did not differ significantly from mean confidence ratings on probability-only trials
(dotted lines) in either condition.
A follow-up analysis was conducted of the effects of framing condition, inference condition and
amount of information on confidence ratings, using each subject’s mean absolute likelihood rating on
complete and incomplete information trials as a covariate. Results conformed closely to those reported
above. None of the factors or their interactions was significant when comparing confidence on
complete-information trials with confidence on probability-only trials. When comparing confidence on
complete-information trials with confidence on amount-only trials, the interaction of amount of
information presented and inference condition was statistically significant, F( 1,75) 4.33, p <0.05, with
adjusted confidence ratings being higher for incomplete- than for complete-information trials in the
Explicit Inference condition but not in the Implicit condition.
Experiment 1 served to replicate the results of earlier studies showing the effect of information frame
on judgments and to provide new data concerning confidence in judgments based on incomplete
information. Of particular interest is the result that confidence is sometimes higher under conditions of
incomplete information than with complete information. However, these results may be confined to
hypothetical gambles where subjects have nothing to win or lose. Experiment 2 was thus conducted
using a partial replication of the design of Experiment 1, but employing real monetary gambles.
EXPERIMENT 2
The purpose of Experiment 2 is to see if the results of Experiment 1 concerning the effects of
information frame, amount of information and inference condition hold if gambles are actually played
for money.
Method
Experimental design
The design of stimulus materials was similar in Experiment 2 to Experiment 1, with two exceptions. The
monetary levels of investment and payoff were scaled back to meet budgetary constraints, and the
number of levels of probability and payoff were reduced because of the increased length of the
experimental session when monetary gambles were actually played. Within each of two randomlyordered replications of the experimental design six gambles were formed by providing all combinations
of three levels of probability and two levels of payoff, and five additional gambles were described by
only one piece of information: one of the three levels of probability or one of the two levels of payoff.
As in Experiment 1, probability information was presented as ‘% chance of winning’(in this case, 5 ,
10 or 15) in the Positive Framing condition and as ‘% chance of losing’(95,90 or 8 5 ) in the Negative
condition. Initial investment for each gamble was e l 5 and payoff amount was either $1.00 or $2.00. The
Explicit and Implicit Inference conditions were the same as in Experiment 1.
Instructions and procedure
Instructions and procedure concerning the use of response scales for indicating likelihood of taking
gambles and expressing confidence in judgments were the same in Experiment 2 as in Experiment 1.
36
Journal of Behavioral Decision Making
zyx
zyxw
Vol. I , Iss. No. I
Subjects in Experiment 2 were given additional instructions concerning the actual playing of gambles.
Subjects were told that they would be given money to gamble with (el5 per gamble) and that they
would be allowed to keep whatever money they have at the end of the session. They were told that when
they rated the likelihood of taking a gamble as ‘somewhat likely’, ‘likely’ or ‘very likely to refuse
gamble’, then they would keep the e l 5 and not invest it. When they rated the likelihood of taking a
gamble as ‘somewhat likely’, ‘likely’ or ‘very likely to take gamble’, then they would draw a chip from a
bag to determine if they win or lose. When they responded ‘equally likely to take or refuse gamble’, a
coin flip would determine whether or not they play the gamble.
In order to eliminate the influence of ‘chance’ factors (i.e. outcomes based on random events), the
following procedure was used for playing the monetary gambles. Subjects were asked to indicate their
responses to each gamble in their booklet before starting to draw chips. Subjects knew that their
responses would lead to actual monetary investments and possible payoffs. By waiting until the end to
play out the gambles, any given response could not be affected by whether or not previous gambles
happened to result in wins or losses, or by revealing the values of missing information.
Gambles were played out by drawing chips from different bags each containing two colors of chips in
proportions specified to represent the different probability levels. In actually playing the gambles that
had missing values, middle levels of each factor were used: 0.10/0.90 probability of winning/ losing,
$1.50 payoff. Subjects had been told that each gamble appeared twice in their booklet, once for practice,
and so only the last 11 gambles were actually played.
Because of the procedure for playing gambles, subjects were tested individually in a session that lasted
about 30 min. Total winnings were rounded up to the nearest e25 and ranged from 625 to $5.00,
averaging about $2.00.
zyxwvuts
Subjects
Different subjects were used in Experiment 2 than in Experiment 1 but they came from the same pool of
psychology students. As in Experiment 1, 20 subjects were assigned at random to each of the four
combinations of Inference condition and Framing condition.
Results and Discussion
Likelihood of taking gambles
Exhibit 3 shows how the ratings of likelihood of taking gambles vary across framing and inference
conditions in Experiment 2. A comparison with Exhibit 1 reveals some similarities and some differences
between results for the two experiments. Ratings, on the average, are less negative in Experiment 2 than
in Experiment 1. This is probably the result of allowing subjects to play with money given to them by
the experimenter. As in Experiment 1, the relative likelihood of taking gambles with complete and
incomplete information in Experiment 2 was different in the Implicit and Explicit Inference conditions
(e.g. note the relative elevation of the solid and dotted lines in the upper panels compared to the lower
panels). Finally, the effect of information frame (comparison of left and right panels) is less pronounced
in Experiment 2 than in Experiment 1.
The reduced framing effect is reflected in the statistical reliability of the results. For probability
payoff combinations (solid lines), the predicted framing effect was significant at the 0.05 level using a
one-tailed test, t(76) = 1.82; for probability-only trials (dotted lines), the effect did not approach
statistical significance.
Inspection of the points to the right of each of the upper panels of Exhibit 3 provides a possible
explanation for the particularly small framing effect in the Implicit Inference condition. Ratings for
payoff amount-only trials were significantly lower in the Positive Framing condition than in the
zyxwvuts
I. P. Levin et al.
IMPLICIT COND: NEGATIVE FRAME
IMPLICIT COND: POSITIVE FRAME
zy
zy
37
Confidence in Judgments
zyxwvuts
zyx
I
5
10
96
16
Arntonly ( $ 1
Chance of Winning ( % )
85
Arntonly (6)
Chance of Losing ( % )
EXPLICIT COND: NEGATIVE FRAME
EXPLICIT COND: POSITIVE FRAME
2.00
90
1-
zyxwvutsrq
Prob
2.00
1.00
1.00
0-
zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQP
-1-
-2-
zyxwvutsrqp
I
5
10
15
Arnt-
only
Chance of Winning (%)
($1
95
90
85
Arntonly ( $ )
Chance of Losing ( % )
Exhibit 3. Mean likelihood of taking gambles as a function of probability and payoff information for each
combination of inference condition and framing condition in Experiment 2.
Negative Framing condition. This is an unexpected result because these trials did not manipulate levels
of the framing variable. Furthermore, this result was not found in Experiment 1 or in the Explicit
Inference condition of Experiment 2 or in any of our earlier research with similar materials (Levin et al.,
1985, 1986). We therefore attribute it to sampling error; subjects initially more negative toward
gambling happened to be assigned to the Implicit Inference-Positive Framing group. This worked
against the framing effect on those trials involving the framing variable (probability-payoff combination trials and probability-only trials).
In the Explicit Inference condition ratings were significantly higher on payoff amount-only trials than
on complete-information trials; t(39)= 4.69, p <0.01. Ratings were also higher on probability-only trials
than on complete-information trials, but this difference did not reach statistical significance (0.10 > p
X.05).In the Implicit Inference condition ratings were significantly lower on probability-only trials
than on complete-information trials, t(39) = 2.94; and ratings were not significantly different on payoff
amount-only trials and complete-information trials. As was the case in Experiment 1, imputations of
missing values tended to be slightly on the optimistic side but were relatively constant across
38
zyx
zy
zyx
zyxwvutsrqpo
zyxwvuts
zyxwvutsr
zyxwvu
Vol. 1, Iss. No. 1
Journal of Behavioral Decision Making
manipulations of the presented values (e.g. the median imputed chance of winning was 15% for both
$1.OO and $2.00 payoffs).
Confidence ratings
Exhibit 4 summarizes the confidence ratings for Explicit and Implicit Inference conditions in
Experiment 2. Comparison with Exhibit 2 reveals similarities and differences between confidence
results for Experiments 1 and 2. Unlike Experiment 1 confidence ratings are higher in the Explicit
Inference condition than in the Implicit Inference condition, both for complete-information trials and
for incomplete-information trials. Again, the relative levels of confidence for complete- and incompleteinformation trials are of primary interest. Here the results are similar for the two experiments.
Confidence in judgments based on incomplete-information trials tends to be less than confidence in
judgments based on complete-information trials in the Implicit condition but the opposite trend can be
observed in the Explicit condition where confidence is higher on incomplete-information trials. Two of
these trends are statistically significant; t(39) = 2.64 for comparing confidence on probability-amount
trials with payoff amount-only trials in the Implicit condition, and t(39) = 2.19 for comparing
confidence on probability-amount trials with probability-only trials in the Explicit condition.
Analysis of covariance using each subject's mean absolute likelihood ratings on complete- and
incomplete-information trials as a covariate produced comparable results. Adjusted confidence ratings
were significantly higher in the Explicit Inference condition than in the Implicit condition. The relative
EXPLICIT INFERENCE CONDITION
IMPLICIT INFERENCE CONDITION
19 18 4-
17 -
"I
13
Prob.
I
10/90
Amt1
only ($1
Chance of Winning/Losing (%)
6/96
16/86
l2
T
Amtonly ($1
Chance of Winning/Losing ( % I
6/96
10/90
16/86
Exhibit 4. Mean confidence ratings as a function of probability and payoff information for each inference
condition in Experiment 2
zyxwvutsrqpo
zy
zyxwvutsrq
zyxw
zy
I. P. b v i n et al.
Confidence in Judgments
39
degree of confidence in judgments based on complete and incomplete information was different for
Explicit and Implicit Inference conditions, resulting in a significant interaction between amount of
information presented and inference condition when comparing complete information trials with
probability-only trials ( F (1,75) 5.59, p <0.05) and an interaction that approached significance when
comparing complete-information trials with payoff amount-only trials ( F (1,75) = 3.29,O. 10 > p X.05).
In each case adjusted confidence ratings were higher for incomplete- than for complete-information
trials in the Explicit Inference condition but not in the Implicit condition.
SUMMARY AND CONCLUSIONS
The effect of information frame on judgments and decisions has been demonstrated many times
(Kahneman and Tversky, 1979; Levin, 1987; Levin et al., 1985, 1986; Tversky and Kahneman, 1981) and
was replicated in the present study. Levin (1987) explains the effect in terms of differential associations
produced by different stimulus labels. In Levin’s study, more favorable associations were produced
when beef was described in terms of ‘percent lean’ rather than ‘percent fat’ and these associations were
assumed to serve as mediators of consumerjudgments and decisions. Analogous differences are likely to
occur in response to the labels, ‘chance of winning’ and ‘chance of losing’.
The reduced effect of information frame on the likelihood of taking real, as opposed to hypothetical,
gambles is consistent with Levin, Schnittjer and Thee’s (1988) recent extension of the associative model.
According to this ‘anchoring and adjustment’ model, the phrasing of stimulus labels determines initial
judgments or anchors. Associations produced by describing a gamble in terms of ‘chance of winning’
lead to a favorable anchor while associations produced by describing a gamble in terms of ‘chance of
losing’ lead to an unfavorable anchor. Degree of adjustment from the initial anchoring position depends
upon the extent to which subjects go beyond the labels to process the objective information. High levels
of personal involvement, such as created by providing real monetary consequences to gambles, can
serve to enhance the adjustment process and reduce the information framing effect.
Major interest in the present study was examining confidence in judgments based on incomplete
information. Results of the Levin and Johnson (1987) study led us to predict that confidence on missing
information trials would not decrease when subjects are explicitly asked to estimate missing values. This
prediction was upheld for both hypothetical and real gambles. While confidence ratings tended to be
somewhat lower on incomplete-information trials than on complete-information trials in the Implicit
Inference condition, some significant elevation of confidence with missing information was found in the
Explicit Inference condition of each experiment.* The effects of explicit inferences on confidence in the
‘real-play’ setting of Experiment 2 are especially important. Marketers should be aware of the
possibility that the greater the extent to which consumers are encouraged to make inferences about
missing information, the more confident will they be in their decisions.
This uncritical acceptance of one’s own inferences may help explain the overconfidence often found
in studies of human judgment and decision making. For example, Slovic, Fischhoff and Lichtenstein
(1977) reported overconfidence on tasks where people judged the odds that their responses to generalknowledge questions were correct. This could be because their answers were based on fallible inferences
which, as the present study suggests, were acted upon as if they were correct.
The present results also relate to research on heuristics. Kahneman and Tversky (1979) suggest that
human judges are overreliant on simplifying heuristics and are relatively insensitive to the potential
unreliability of their sources of information. The present research extends this principle to the use of
oneself as a source. We tend to trust our own inferences. Mehle, Gettys, Manning, Baca and Fisher
(198 1) point to the role of the availability heuristic in overconfidence. In the present case, one’s own
zyxwv
40
zy
zy
zyxwvutsrqpo
zyxwvutsrq
Vol. I , Iss. No. 1
Journal of Behavioral Decision Making
inference - especially if it is made explicit - is readily available in memory and thus increases
confidence.
Finally, these results may help explain the relatively low levels of information seeking commonly
observed in studies of judgment and decision making. For example, Howard and Sheth (1969) observed
that as confidence increases, a consumer’s tendency to take in information declines. The present
research suggests one reason why people refrain from seeking more information. Confidence in their
own inferences leads them to take insufficient account of the uncertainty produced by incomplete
disclosure of information.
zyxwv
zyxwvut
REFERENCES
Anderson, N. H. (1981). Foundations of Information Integration Theory. Academic Press, New York.
Fischhoff, B., Slovic, P. and Lichenstein, S. (1977). ‘Knowing with certainty: The appropriateness of extreme
confidence.’ Journal of Experimental Psychology: Human Perception and Performance, 3,552-564.
Howard, J. A. and Sheth, J . N. (1969). The Theory of Buyer Behavior. Wiley, New York.
Huber, J. and McCann, J. (1982). ‘The impact of inferential beliefs on product evaluations,’Journal of Marketing
Research, 19,324-333.
Johnson, R. D. and Levin, I. P. (1985). ‘More than meets the eye: The effect of missing information on purchase
evaluations,’ Journal of Consumer Research, 12, 169-1 77.
Kahneman, D. and Tversky, A. (1979). ‘Prospect theory: An analysis of decision under risk,’ Econometrica, 47,
263-291.
Koriat, A., Lichtenstein, S. and Fischhoff, B. (1980) ‘Reasons for confidence.’ Journal of Experimental Psychology: Human Learning and Memory, 6, 107- 1 18.
Levin, I. P. (1987). ‘An associative model of the effects of information frame on consumer behavior,’ Bulletin ofthe
Psychonomic Society, 25,85-86.
Levin, 1. P. and Johnson, R. D. (1987). ‘Price-quality inferences and framing effects in consumer judgments,’ in S.
Maital (Ed.), Applied Behavioral Economics, Wheatsheaf, Sussex.
Levin, I. P., Johnson, R. D., Deldin, P. J., Carsten, L. M., Cressey, L. J. and Davis, C. D. (1986). ‘Framing effects
in decisions with completely and incompletely described alternatives,’ Organizational Behavior and Human
Decision Processes, 38,48-64.
Levin, I. P., Johnson, R. D., Russo, C. P. and Deldin, P. J. (1985). ‘Framing effects in judgment tasks with varying
amounts of information,’ Organizational Behavior and Human Decision Processes, 36, 362-377.
Levin, I. P., Johnson, R. D. and Faraone, S. V. (1984). ‘Information integration in price-quality tradeoffs: The
effects of missing information,’ Memory & Cognition, 12,96-102.
Levin, I. P., Schnittjer, S. K. and Thee, S. L. ‘Information framing effects in social and personal decisions,’
Journal of Experimental Social Psychology (in press).
Mehle, T., Gettys, C. F., Manning, C., Baca, S. and Fisher, S. (1981). ‘The availability explanation of excessive
plausibility assessments’, Acta Psychologica, 49, 127- 140.
Singh, R. (1984). ‘Two problems in cognitive algebra: Imputations and averaging-versus-multiplying.’ Working
paper No. 490, Indian Institute of Management, Ahmedabad, India.
Slovic, P., Fischhoff, B. and Lichtenstein, S. (1977). ‘Behavioral decision theory.’ Annual Review of Psychology,
28, 1-39.
Tversky, A. and Kahneman, D. (1981). ‘The framing of decisions and the psychology of choice,’ Science, 211,
453-458.
Yamagishi, T. and Hill, C. T. (1981) ‘Adding versus averaging models revisited: A test of a path-analytic integration
model,’ Journal of Personality and Social Psychology, 41, 13-25.
NOTES
I. We recognize the uniqueness of asking subjects to rate their confidence in responses to a 7-point category scale.
This was designed to allow separate assessments of the polarity of a judgment and the subject’s confidence in that
I. P. Levin et al.
Confidence in Judgments
zy
zy
41
judgment. To be sure, the response scales are not completely independent. For example, low confidence was almost
never assigned to likelihood judgments that were at one extreme or the other of the scale. Analyses of confidence
ratings to be reported later take into account this covariation.
2. One could argue that imputations of missing values did not in fact occur in the Implicit condition. However,
there is reason to believe otherwise. Both probability and payoff information are crucial in evaluating gambles.
Furthermore, averaging theory (Anderson, 1981) suggests that the impact of a single source of information,
evaluated in isolation, would be greater than its effect when it was combined with another source of information.
This would be seen in Exhibit 1 as a steeper slope for the dotted line than the solid lines in each panel. This was
clearly not the case, suggesting that some imputing of missing values occurred even in the Implicit condition. The
fact that the dotted lines are lower in elevation than the solid lines in the Implicit condition conforms to the usual
‘negativity bias’ in responding to incomplete information (Johnson and Levin, 1985). Asking subjects to explicitly
fill in missing values wipes out this effect.
zyxwvutsrq
Authors’ Biographies:
Irwin P. Levin is Professor of Psychology and Director of Honors at The University of Iowa. He has been at The
University of Iowa since receiving his Ph.D from U.C.L.A. in 1965. His recent research has tested models of
information integration, inference processes and information framing effects.
Richard D. Johnson is Assistant Professor of Marketing at The University of Alberta. He received his Ph.D. in
Psychology at The University of Iowa in 1985 and worked as a postdoctoral scholar at The University of Chicago
for 1985-86. His research examines inference processes in consumer behavior.
Daniel P. Chapman is a doctoral student in the Department of Psychology at The University of Iowa. His research
interests include learning and memory processes in decision making.
Authors’ Addresses:
Irwin P. Levin, Department of Psychology, The University of Iowa, Iowa City, Iowa 52242, U.S.A.
Richard D . Johnson, Faculty of Business, The University of Alberta, Edmonton, Alberta, Canada T6G 2R6
Daniel P. Chapman, Department of Psychology, The University of Iowa, Iowa City, Iowa, 52242, U.S.A.