The role of personal interaction in the assessment of risk
attitudes
Benjamin Roth1, Stefan T. Trautmann1,* & Andrea Voskort2
1
2
University of Heidelberg
BaFin (German Financial Services Authority)
Feb 24, 2016
Abstract:
Many decisions under uncertainty are delegated to professionals, such as financial advisors or
medical doctors, requiring them to assess the risk attitudes of their clients or patients. To
gain a better understanding of the potential factors influencing risk attitude assessments, the
current study investigates the role of personal interaction in these assessments. Controlling for
information transmitted, we find that personal interaction leads to more risk-averse
assessments, but does neither harm nor benefit assessments in terms of precision. We replicate
previous findings of stereotypes in risk preference predictions, and discuss the influence of
the assessor’s own risk attitude on her assessments.
Highlights

Personal interaction affects the assessment of risk attitudes

No effects of personal interaction on the precision of the assessment

Stereotypes in risk assessment are observed

Assessments are affected by the assessors own risk attitude
Keywords: risk attitude, advice
JEL-Codes: C91, D81, G02
_____________________________
*Corresponding author: Alfred-Weber-Institute for Economics, University of Heidelberg, Bergheimer Str. 58
(Room 01.029), 69115 Heidelberg, Germany, Phone: +49 6221 54 2952, Fax: +49 6221 54 3592; email:
trautmann@uni-hd.de. We are grateful to the associate editor Cary Deck, the two reviewers, Christian KönigKersting, Rosemarie Nagel, Jörg Oechssler, Andreas Roider, Alex Roomets, as well as participants at the ESA
World Meeting for helpful comments.
1
1. Introduction
Advice is important in many domains of decision making under uncertainty. Prominent
examples are medical decisions supported by the advice of a doctor (Gigerenzer et al. 2007),
or investment decisions supported by the advice of a financial professional. An advice
relationship typically requires the advisor to carefully assess the advisee’s willingness to take
risks. For example, the MiFID (2006) guidelines specify that financial professionals need to
know their customers’ preferences; medical doctors need to discuss risks with their patients
and obtain informed consent.
We are interested in the aspects that influence such assessments of other person’s
risk attitudes. In particular, we investigate whether personal interaction conveys relevant
information about an advisee’s risk attitudes, over and beyond a set of demographical
background variables. We use experimental risk preference measurement and belief elicitation
methods to study this question. Importantly, in our study, the personal interaction does not
allow the advisor to formally assess or elicit the risk attitudes of the advisee. Rather, we are
interested in the implicit information that the advisor may obtain through an informal
interaction with a client.
Our investigation starts from the assumption that having more information, albeit
implicit one deriving from the unstructured interaction during a consultation, is beneficial to
the goal of assessing the advisee’s preferences. However, it is also conceivable that personal
interaction only adds noise and possibly biases to the assessment by the advisor. Our design
aims to detect both positive effects and negative effects of personal interaction. We also test
(i) if stereotypical assessments exist and whether they are more or less prevalent in the
personal interaction condition; and (ii) if the advisor’s risk attitudes influence her assessment
of the advisee’s preferences.
In the next section we present the details of our experimental design. The following
section discusses the results. It will be shown that participants predict systematically lower
risk aversion when they make their assessments based on merely a set of demographics,
compared to a situation where they additionally have the possibility to obtain information in a
personal interaction. However, the precision of the risk attitude assessments does not differ
systematically between conditions with and without personal interactions.1 These findings
replicate across two different methods of risk preference assessment. We also replicate
previously shown stereotypes in risk attitudes assessment, and find that the assessors own risk
1
Precision is measured as the absolute deviation of the prediction from the true attitude, or alternatively as the
number of correct predictions.
2
attitudes influence her assessment. The final section discusses our results in the context of the
existing literature.
2. Experimental Design
2.1. Methods
To assess whether personal interaction affects judgments of other people’s risk attitude, we
use a within-person design as follows. In Stage 1 of the experiment, participants make risky
choices as described in more detail below. In Stage 2, participants judge the risk attitude of
other people on the basis of a condensed demographic summary shown in Table 1. The
variables were selected based on previous research showing a correlation of these attributes
with risk attitudes (e.g., Dohmen et al. 2011; Noussair et al. 2014; Roth and Voskort, 2014).
Subjects were presented with 20 different profiles, all coming from real experimental
participants. 8 of these profiles came from other participants in the current experiment, 4 of
whom the assessor would encounter in Stage 3 of the experiment. The other profiles came
from a related study where risk attitudes were measured in a non-student sample with larger
variation in demographics. Risk attitude judgments were incentivized as described below.
Table 1: Demographic Information Provided in Stage 2
Variable
Possible realizations and aggregation
Gender
Age
Height
Income (net/p.m.)
male, female
25 or younger; 26-40; 41-65; 65 or older
cm
€1000 or less; €1001 to €3000; €3001 to €6000; €6000 or more
single; divorced; in a relationship; living separately; married;
widowed
children; no children
Family Status
Children
In Stage 3 of the experiment, participants had four one-to-one conversations with other
participants. All conversations took place at separated tables and were audio-recorded. In each
conversation of 2 minutes (in mode D) or 4 minutes (in mode F, defined below), initially one
person interviewed the other person. After a signal by the experimenters, the groups switched
roles and the other person took the role of the interviewer. That is, in each conversation, both
participants assessed the other person (and thus were assessed by another person). The
participants had to take notes during the interviews. After the last conversation, they returned
to their computers and filled in the risk-preference assessment for each of the four people they
3
talked to. Importantly, during the personal interview participants were not allowed to ask for
the other person’s behavior in the risk tasks (participants knew in advance that they would
assess the other person on these dimensions). With three experimenters present, and audiorecordings available, we did not detect any violations of this restriction.
The four interviews in Stage 3 were split in 2 groups of 2 interviews that differed
according to the questions the participants could ask during the interview. In type D
(“demographics only”) interviews, subjects had to ask for exactly the information that was
given in the short descriptions of people in Stage 2. In type F (“free conversation”) interviews,
subjects had to collect the information given in the Stage 2 descriptions, and were also
allowed to additionally collect any other information about the other person (with the
exception described above). This setup ensured that in Stage 3 assessments, assessors had at
least as much information about the other person as in Stage 2. Note that participants were not
aware of the fact that they would encounter people in Stage 3 whose profiles they assessed in
Stage 2 of the experiment.
Stage 1 risk preference elicitation. The Stage 1 risk elicitation employed two different
preference elicitation tasks: a financial investment task modeled upon a survey question in
Dohmen et al. (2005); and an abstract binary-choice based risk preference measure based on
Holt and Laury (2002). In the following we refer to the first measure as Investment, and to the
second measure as Choice. The measures are normalized in the analyses such that higher
values always refer to more risk aversion.
The investment decision elicited the share of a windfall gain that a participant was
willing to put at stake in a risky investment vs. the share she would prefer to keep uninvested.
There were six possible amounts. The question was stated as follows (cf. Dohmen et al., 2005,
p.8):
“Please consider what you would do in the following situation: Imagine that
you had won €100,000 in a lottery. Almost immediately after you collect the
winnings, you receive the following financial offer, the conditions of which
are as follows: The amount invested either gets doubled, or you lose half of
it, with equal probability. You have the opportunity to invest the full amount,
a part of it, or nothing and thus reject the offer. What share of your lottery
winnings would you be prepared to invest in this financially risky, yet
lucrative investment?
4
Your Decision: €100,000 - €80,000 - €60,000 - €40,000 - €20,000 - Nothing,
I would decline the offer.”
The investment decision was incentivized such that each €400 in experimental payoff yields 1
cent in real payoffs. That is, the risk-free option of not investing at all yields a payoff of
€2.50. In the analyses below we normalize the amount uninvested on a scale from 0 (no risk
aversion; full investment) to 10 (kept the full amount uninvested).
The second measure of risk preference was elicited by an adapted choice list task
using the payoffs and probabilities shown in Table 2. Because we needed unambiguous
choices in this choice list tasks to incentivize risk preference predictions, we enforced a
unique switching point by asking participants in which row they wanted to switch to the
riskier option B. As shown in Table 2, initially option A is much more attractive than option
B. When going down the list of choices, option B becomes relatively more attractive, and
dominates option A in choice no. 10. The choice item in which the decision maker switches
from the safer option A to the riskier option B indicates the decision maker’s risk attitude,
normalized as the degree of risk aversion on a scale from 0 (immediately) to 10 (never).
Table 2: Choice list risk measure
Choice no.
Option A
Option B
1
2
3
4
5
6
7
8
9
10
10%: €2; 90%: €1.60
20%: €2; 80%: €1.60
30%: €2; 70%: €1.60
40%: €2; 60%: €1.60
50%: €2; 50%: €1.60
60%: €2; 40%: €1.60
70%: €2; 30%: €1.60
80%: €2; 20%: €1.60
90%: €2; 10%: €1.60
100%: €2; 0%: €1.60
10%: €3.85; 90%: €0.10
20%: €3.85; 80%: €0.10
30%: €3.85; 70%: €0.10
40%: €3.85; 60%: €0.10
50%: €3.85; 50%: €0.10
60%: €3.85; 40%: €0.10
70%: €3.85; 30%: €0.10
80%: €3.85; 20%: €0.10
90%: €3.85; 10%: €0.10
100%: €3.85; 0%: €0.10
Stage 2 and 3 belief assessment. On the basis of stylized demographic information (stage 2) or
on the basis of the information collected during the personal conversation (stage 3), subjects
had to predict the decision maker’s exact decision in the two risky choice tasks. Showing the
demographics on-screen, subjects were asked "How do you think the advisee whose
characteristics are displayed in the box above decided in this risky choice task?" in the
5
respective risk measure. If the advisor correctly predicts the choice of the advisee we pay
€0.50 per risk elicitation mechanism. As we reward only for an exactly correct prediction, this
payoff structure is incentive compatible. Profiles were presented in a random order for belief
assessment; the assessment of the investment question always came before the assessment of
the choice list measure.
In Stage 3, subjects keep notes during their personal meeting. When returning to
their computers after the conversations, they answer the belief assessments for each of their
conversation partners in the order in which they met them.
2.2. Discussion of Design Choices
In this subsection we want to discuss some of the design choices we made in the current
study, and how they may affect the robustness and external validity of the results. We discuss
three aspects: the choice of the risk elicitation method; the belief elicitation method; and the
choice of a within-person design.
Various studies have shown a low correlation among different risk elicitation
methods (cf., Dave et al., 2010; Deck et al., 2013; Crosetto and Filippin 2015; Friedman et al.,
2014). Using two different methods (binary choice vs. investment decision), we can observe
the robustness of our results with respect to differences across elicitation tools. Results show
that behavior was largely consistent across the two tasks. We report that in our study the rank
ordering of participants in terms of their risk aversion as measured in the two tasks is
positively correlated with =0.34 (Spearman Rank correlation, p<0.01).
To elicit participants’ assessment of other participants’ risk attitude we had to make
a trade-off between simplicity/tractability and complexity (Trautmann and van de Kuilen,
2015). We note there that the simple incentivized measure we use here may not capture more
complex patterns of beliefs about others, and that a measure of prediction precision in terms
of absolute deviation from the true risk attitude may therefore not reflect the assessors’ goals
in the current task (see discussion below and robustness checks in terms of number of correct
predictions). Future research may fruitfully investigate more complex patterns of beliefs in
risk assessment.
The current study employs a within-person design with a fixed order of tasks. The
ordering with mere written information on group of potential clients followed by a more
personal interaction with a subgroup is clearly relevant in many situations (where some
information is submitted before a meeting). However, we acknowledge that the fixed order
may lead to order effects that we cannot detect in the current design.
6
2.3. Lab Details
The experiment was conducted at the facilities of the University of Heidelberg’s AWI Lab.
The computerized parts of the experiments were programmed in zTree software (Fischbacher
2007), and subjects were recruited via ORSEE system (Greiner 2015). In total 104 student
subjects participated in 13 sessions. The experiment took approximately 55 minutes, and
subjects earned on average €13.46.
3. Results
Predictions: Effects of Personal Interaction on Risk Assessment. We present results on
predictions in Figure 1 and Table 3. Figure 1 shows that there is a lot a variation in
predictions, and that predictions after a conversation seem to be more risk averse.2 Table 3
gives more detail, additionally separating between the two conversation modes. The table
shows averages of the revealed risk attitudes (columns 1 and 2) and the predictions under
different conditions. In particular, predictions in Stage 2 are shown in Columns 3 and 4;
predictions in Stage 3 for conversation mode D are shown in Columns 5 and 6; and
predictions in Stage 3 for conversation mode F are shown in Columns 7 and 8. We split the
data along four demographic dimensions: gender; age (median split at age 25); height (median
split at 175cm); and partnership status. For each of the dimensions we thus observe whether
there is a difference in risk behavior, and whether there is a belief that such a difference exists
in the difference prediction situations.
0
0
.05
.1
.1
Density
Density
.15
.2
.2
.3
.25
Figure 1: Distribution of Predictions
0
2
4
Investment
Demographics
6
8
Conversation
10
0
2
4
Choice
Demographics
6
8
10
Conversation
Notes: Left panel = investment task; right panel = choice task; distribution for conversation treatments combines
both modes
2
Repeat predictions averaged for each assessor on the basis of the normalized risk measures on scale 0 to 10.
7
The table confirms that there are differences between predictions in Stage 2 with
only a description of demographics, and Stage 3 with personal interaction. Predictions in
Stage 3 that are significantly larger than the corresponding predictions in Stage 2 are indicated
by shaded cells (see Table A1 in the appendix for the respective test statistics concerning
these across-column tests in Table 3). People are systematically assessed more risk averse
with personal interaction. Differences are more consistently significant over categories for
Stage 3 mode D than for mode F, but the general pattern obtains in both conditions.
Differences are of similar size between Stage 2 and Stage 3 mode D and mode F, respectively;
however, the free discussion in mode F seems to add more noise, leading to more variance in
assessments and making it more difficult to identify effects.
Looking in more detail at the different demographics, we firstly observe that in our
sample there are no significant differences in risk attitudes along any of the four dimensions,
and for neither of the two risk attitude measures.3 In contrast, Table 3 shows stereotypes about
risk attitudes. We find that women are predicted more risk averse than men in both Stage 2
predictions and Stage 3 predictions. We also find that taller people are predicted to be less risk
averse in Stage 2 and Stage 3 mode D; the effect points in the same direction for Stage 3
mode F, but is insignificant. There is some indication that younger people are perceived as
more risk averse, but predictions for the young and old differ significantly in only two out of
the six comparisons. No stereotypes for people with partner vs. singles show up in the
prediction data. Note that the observed stereotypes are not supported by the behavior of the
different groups. However, many studies have obviously found that woman are more risk
averse, and the predicted differences may in fact be reasonable when made on the basis of a
larger sample (although the size of the behavioral difference may in fact be much smaller than
perceived by the assessors; see Filippin and Crosetto, 2014). Most relevant to the current
paper, there is no evidence that there is a systematic bias induced by the personal interaction
in terms of stereotypes: there is no increase in stereotypical predictions in Stage 3 compared
to Stage 2.
3
Most significant effects of demographics are found in large population studies. The absence of effects in our
small student sample should therefore not be generalized.
8
Demographic
attribute
Table 3: Choices and Predictions of Risk Attitudes
Actual Choice
Prediction based on
Prediction based on
(Stage 1)
demographics (Stage 2) b
conversation (Stage 3):
mode D
Investment
Choice
Investment
Choice
Investment Choice
(1)
(2)
(3)
(4)
(5)
(6)
Prediction based on
conversation (Stage 3):
mode F
Investment
Choice
(7)
(8)
Male
4.63
4.80
4.13***
4.67***
4.73***
5.17***
4.7**
5.18
(3.29)
(1.99)
(2.57)
(1.85)
(2.59)
(1.72)
(2.87)
(2.21)
4.60
5.30
5.93
5.65
6.45
5.87
6.14
5.81
(2.86)
(1.67)
(2.33)
(1.6)
(2.23)
(1.76)
(2.63)
(2.06)
4.67
5.13
5.22***
5.16
5.58
5.50
5.46
5.47**
(3.11)
(1.9)
(2.27)
(1.52)
(2.15)
(1.63)
(2.46)
(1.83)
4.38
4.67
4.04
4.82
4.94
4.93
5.06
4.99
(3.01)
(1.65)
(2.59)
(2.03)
(3.35)
(1.95)
(2.7)
(2.11)
4.59
4.72
4.45***
4.84***
4.96***
5.28***
4.80
5.19
(3.22)
(1.96)
(2.63)
(1.86)
(2.5)
(1.73)
(2.86)
(2.18)
4.64
5.38
5.64
5.49
6.09
5.90
5.85
5.54
(2.95)
(1.69)
(2.39)
(1.59)
(2.25)
(1.68)
(2.49)
(2.2)
4.61
5.36
4.87
5.11
5.38
5.64***
5.49
5.39
(3.34)
(1.91)
(2.51)
(1.66)
(2.51)
(1.65)
(2.64)
(2.21)
a
Female
Young
Old
Tall
Short
Single
Partner
Overall average
4.62
4.87
5.26
5.09
5.79
5.15
5.19
5.55
(2.96)
(1.81)
(2.41)
(1.79)
(2.51)
(1.86)
(3.03)
(2.17)
4.62
5.04
5.05
5.08
5.53
5.49
5.34
5.37
(3.08)
(1.85)
(2.19)
(1.66)
(2.15)
(1.59)
(2.26)
(1.74)
Notes: Larger numbers indicate higher (predicted) risk aversion; risk measures and predictions normalized on scale 0 (least risk averse) to 10 (most risk averse); a:
Average measure of risk aversion in the shown category. b: values for all observations where only demographics have been available to the subjects, irrespective of
their Stage 3 condition. Median split for height and age: medians are at a body height of 175cm and at an age of 25; *, **, *** indicate a within-attribute difference in
attitudes or predictions between the two categories in each attribute, at the 10%, 5%, 1% significance level of a Mann-Whitney-U test (e.g., male v. female). For
predictions we aggregated observations for the same subject per category for testing purposes. Standard deviations in parentheses. Shading indicates that Stage 3
prediction is significantly larger than the corresponding Stage 2 prediction. Details are given in Table A1.
9
We probe the above-described pattern in a multivariate analysis. Table 4 shows results for
linear regression analyses of predictions across all conditions, clustered at the level of the
assessor (each assessor made multiple assessments and across conditions in the within-subject
design). Columns 1 and 2 confirm the gender stereotype, but not the effect of height. The
latter seems to be affected by a strong correlation with gender. The effect of Stage-3- personal
interaction emerges significantly. The effects of mode D is individually significant; moreover,
an F-test rejects the hypothesis that both the mode D and the mode F dummies are equal to
zero (Investment F=4.39 p=0.02, Choice F=4.50 p=0.01). Columns 3 and 4 show that the
latter effects are robust if we do not include demographics (Investment F=4.39, p=0.01,
Choice F=4.39, p=0.01).
Table 4: Regression analyses  target’s attributes and treatment effect
Dependent variable: Predicted degree of risk aversion
Investment Choice
Investment
Choice
Attribute
(1)
(2)
(3)
(4)
Male
Young
Tall
Single
Stage3D
Stage3F
a
a
Constant
-1.83***
-0.98***
(0.68)
(0.46)
0.50
0.15
(0.62)
(0.6)
0.33
0.31
(0.60)
(0.48)
-0.11
0.05
(0.56)
(0.4)
0.48***
0.34***
0.48***
0.34***
(0.34)
(0.24)
(0.17)
(0.12)
0.29
0.22
0.29
0.22
(0.36)
(0.28)
(0.19)
(0.14)
5.50***
5.34***
5.05***
5.14***
(0.70)
(0.68)
(0.22)
(0.15)
N
832
832
832
832
Clusters
104
104
104
104
Notes: a: Omitted category for condition is Stage2: Demographics only. Standard errors in
parentheses and clustered on subjects. Demographic attributes refer to targets.
Precision of predictions. We next consider the precision of the risk assessments. We define
for each prediction the absolute accuracy as |predicted choice – actual choice|. Table 5 gives
these values for each category, and tests for differences on the whole sample (independently
of category). We observe few differences and no systematic improvement in the precision of
10
the predictions in the case of conversation over mere demographics. That is, while predictions
are more risk averse with interaction (as shown in tables 3 and 4), this does not lead to
predictions being more accurate in terms of absolute deviation. In contrast, predictions seem
become too risk averse (relative to the true choices) with interaction, keeping accuracy at a
similar level.
As discussed in section 2.2, assessors were not rewarded for minimizing absolute
deviation. As a robustness test, we study precision in terms of the share of correct predictions
(which were rewarded). Results are given in Table A2 in the appendix. We observe few
significant differences. Over all predictions we only find Stage 3 mode D investment
predictions marginally less accurate than Stage 2 predictions at the 10% significance level. No
differences emerge for Stage 3 mode F or the mode D choice task predictions. Thus, we
conclude that there is no evidence that personal interaction improved accuracy of risk attitude
assessments in our study.
The role of the assessor’s risk attitude. An important aspect in the assessment of other
people’s risk attitudes is the question whether the assessor can avoid an influence of her own
risk attitude on her assessment, and whether such an influence may depend on the condition.
Figure 2 provides a graphical illustration of the relationship between an assessor’s own risk
attitude and her average prediction (averaged over all conditions). There is a clear and
significant positive correlation between the assessors own risk attitude and her predictions
(=0.59 and =0.41, with p<0.01 for investment and choice tasks). We also looked at the
relation between own and the predicted risk attitude for the different treatment conditions
separately. The same pattern emerges in all treatments as in the pooled data shown in Figure
2. Personal interaction does neither eliminate nor increase reliance of the assessor on her own
attitude as an indicator for what the attitude of the other person may be.
11
Demographic
attribute
Table 5: Accuracy of Predictions (Absolute Deviations from Actual Choice)
Prediction based on
Prediction based on
Prediction based
conversation
conversation
on demographics
a
(Stage 3): mode F
(Stage 3): mode D
(Stage 2)
Investment Choice
Investment
Choice
Investment
Choice
(1)
(2)
(3)
(4)
(5)
(6)
Male
Female
3.58 (2.03)
3.16 (1.83)
2.38 (1.37)
1.80 (1.05)
3.66 (2.22)
3.35 (2.1)
2.07 (1.33)*
1.99 (1.81)
3.70 (2.64)
2.70 (1.97) **
2.39 (1.75)
1.93 (1.39)
Young
Old
3.26 (1.39)
3.22 (2.17)
2.01 (0.95)
2.01 (1.53)
3.28 (1.81)
3.22 (2.45)
1.95 (1.34)
1.76 (1.23)
2.95 (1.91)
3.11 (2.21)
2.07 (1.28)
2.15 (1.78)
Tall
Short
3.65 (1.98)
3.07 (1.83)
2.34 (1.34)
1.80 (1.1)
3.81 (2.06)
3.20 (2.13)
2.08 (1.21)
1.98 (1.74)
3.51 (2.47)
2.76 (1.93)
2.49 (1.73)
1.96 (1.45)
Single
Partner
3.46 (1.54)
3.02(2.08)
2.10 (1.04)
2.01 (1.45)
3.29 (2.08)
3.76 (2.19)**
1.91 (1.5)
2.08 (1.27)
3.20 (2.13)
3.09 (2.66)
2.22 (1.57)
2.07 (1.57)
2.04 (0.84)
3.39 (1.65)
1.98 (1.15)
3.09 (1.7)
2.11 (1.1)
Overall average
3.32 (1.29)
Notes: Entries are average absolute deviations from true risk attitude; a: values for all observations where only demographics have been available to the
subjects, irrespective of their Stage 3 condition; Median split for height and age: medians are at a body height of 175cm and at an age of 25; *, **, ***
indicate a difference between absolute deviations in stage 2 and stage 3, at the 10%, 5%, 1% significance level, Wilcoxon test: we test whether (1) is equal
to (3) and (1) is equal to (5), and whether (2) is equal to (4) and (2) is equal to (6). Standard deviations in parentheses.
12
0
0
Own degree of risk aversion (Choice)
2
4
6
8
Own degree of risk aversion (Investment)
2
4
6
8
10
10
Figure 2: Assessor’s risk attitude and predictions
0
2
4
6
8
Predicted degree of risk aversion (Investment)
95% CI
Linear prediction
10
2
4
6
8
Predicted degree of risk aversion (Choice)
95% CI
10
Linear prediction
Notes: Left panel = investment task; right panel = choice task; predicted degree: averaged over all predictions by an assessor
13
4. Discussion and Conclusion
The current study investigates the effects of personal interaction on the assessment of other
people’s risk attitudes, compared to a situation where only basic demographic information is
available to the assessor. We make the interesting observation that interaction leads to more
risk averse assessments. This is an important insight that suggests that personal interaction
does matter. However, there is no evidence that it leads to more accurate predictions; on the
good side, there seem to be no systematic biases induced by personal interaction either. For
example, assessors do not become more stereotyping under personal interaction.
We do observe clear stereotyping in the assessment of risk attitudes though. We
replicate the previously observed pattern of men and taller people being predicted to be more
risk tolerant, and that this prediction is not reflected in actual choices (e.g., Ball et al. 2010;
Eckel and Grossman 2002). These patterns suggest that financial advice relationships suffer
from systematic judgment errors (irrespective of the personal or impersonal format). One
factor that seems to become important here is the assessor’s reliance on her own risk attitude
as a predictor of the client’s risk attitude.
However, we also have to stress the limitations of our approach, and the question
regarding the external validity of our results. As in any experiment, we had to make various
design decisions that could have potentially influenced the results. Probing the robustness
with respect to variation in risk task, belief measurement, and the within-person design in
future research can speak to this question. External validity should also be judged with care.
Since we worked with convenience samples of assessors, it is not clear whether professional
advisors would suffer from similar stereotype biases. Moreover, they may have been selected
for the job exactly because they are more skilled in using personal interaction to identify
clients’ preferences. Very likely they are less prone to rely on their own views to predict
others’ attitudes. Future research could use the current research paradigm to examine the
question whether the current results replicate for financial professionals (e.g., Kirchler et al.
2015).
14
Appendix
A Additional analyses and statistics
Table A1: Differences between stage 2 and stage 3 predictions: Test statistics
Prediction based on
Prediction based on
conversation (Stage 3):
conversation (Stage 3):
mode F
mode D
Demographic
Investment
Choice
Investment
Choice
attribute
(3) vs (5)
(4) vs (6)
(3) vs (7)
(4) vs (8)
Male
-1.988**
-2.462**
-2.031**
-1.994**
Female
-2.547**
-1.981**
-0.416
-0.891
Young
Old
-2.388**
-1.072
-2.732***
-0.537
-1.077
-1.314
-1.695*
-0.307
Tall
Short
-1.541
-2.337**
-1.233
-2.949***
-1.311
-0.272
-1.842*
-1.066
Single
Partner
-1.550
-1.600
-3.086***
-0.631
-2.279**
0.587
-1.906*
-0.369
Overall average
-2.30**
-2.30**
-1.312
-1.268
Notes: Entries are z-values of Wilcoxon sign rank tests. We test for the equality of the predictions in Stage 2 (columns (3) and (4) of Table 3) vs. the
prediction in Stage 3 (columns (5/7) and (6/8) of table 3). *, **, *** indicate the 10%, 5%, 1% significance level.
15
Demographic
attribute
Male
Female
Table A2: Accuracy of Predictions (% of correct predictions)
Prediction based on
Prediction based on conversation
demographics (Stage 2)a
(Stage 3): mode D
Investment
Choice
Investment
Choice
(1)
(2)
(3)
(4)
0.18 (0.3)
0.12 (0.24)
0.12 (0.27)*
0.1 (0.27)
0.18 (0.28)
0.2 (0.29)
0.14 (0.33)*
0.15 (0.32)
Prediction based on conversation
(Stage 3): mode F
Investment
Choice
(5)
(6)
0.15 (0.3)
0.18 (0.37)
0.27 (0.39)
0.11 (0.29) ***
Young
Old
0.19 (0.25)
0.19 (0.36)
0.16 (0.22)
0.15 (0.34)
0.16 (0.33)
0.19 (0.38)
0.15 (0.31)
0.14 (0.33)
0.24 (0.33)
0.21 (0.38)
0.13 (0.28) **
0.24 (0.42)
Tall
Short
0.17 (0.28)
0.19 (0.3)
0.13 (0.23)
0.2 (0.3)
0.11 (0.25) **
0.16 (0.33)
0.07 (0.21) **
0.17 (0.32)
0.18 (0.34)
0.25 (0.37)
0.16 (0.34)
0.14 (0.32) *
Single
Partner
0.16 (0.26)
0.22 (0.35)
0.15 (0.24)
0.17 (0.31)
0.16 (0.31)
0.11 (0.31) **
0.18 (0.34)
0.06 (0.21) ***
0.21 (0.33)
0.27 (0.42)
0.16 (0.33)
0.15 (0.34)
Overall average
0.19 (0.21)
0.16 (0.19)
0.15 (0.26) *
0.14 (0.25)
0.23 (0.27)
0.15 (0.26)
Notes: Entries are the percentage of exactly correct predictions of the true risk attitude; a: values for all observations where only demographics have been
available to the subjects, irrespective of their Stage 3 condition; Median split for height and age: medians are at a body height of 175cm and at an age of
25; *, **, *** indicate a difference between stage 2 and stage 3 success rates, at the 10%, 5%, 1% significance level, Wilcoxon test: we test whether (1) is
equal to (3) and (1) is equal to (5), and whether (2) is equal to (4) and (2) is equal to (6).
16
B Translation of Experiment
E
tal Instructtions
Basic Inforrmation
Please enterr your gender, yeear of birth, mariital status, if youu have children, your
y
highest leveel of education, your
y
monthly nett income and
your body hheight.
Decision I
Imagine thaat you have just won
w 100,000 Eurros in a lottery. Im
mmediately after receiving the 100,000 Euro you obtain
o
the follow
wing proposal
for a new loottery: On the onne hand you havee the chance of doubling
d
your mooney; on the other hand, you could lose half of thee money you
have wagereed. These events happen with the same
s
probabilityy.
What fractiion of you winniings do you wannt to wager on thhe risky but alsoo profit-promisingg lottery? The to
otal amount of 100,000€; the
amount of 880,000€; the amou
unt of 60,000€; thhe amount of 40,000€; the amountt of 20,000€; noth
hing, I would nott take part in the llottery.
Your Comppensation: In term
ms of your actual compensation, thhe 100,000 Euro are
a equivalent to 2.50 Euro (80,0000 Euro corresponnd to 2 Euro,
etc.).
Decision II
w of the Table yoou can choose eithher Option A or Option
O
B. Option A and Option B are two lotteries.. Your task is to choose
c
of the
In every row
two lotteries (either Option A or Option B).
Consider thhe first row for ex
xample: In Optionn A you receive a payment of 2 Euuro with a probabbility of 10% and
d a payment of 1.660 Euro with
a probabilityy of 90%. If you imagine a ten-sidded-dice this wouuld mean that youu receive 2 Euro if you rolled a 10 and 1.60 Euro foor rolling any
number betw
ween 1 and 9. Iff you choose Optiion B you will reeceive 3.85 Euro with a probabilitty of 10% and 0.10 Euro with a probability
p
of
90%. If youu again imagine the ten-sided-dicce, this would ind
dicate that you receive
r
3.85 Euroo if you roll a 100 and 0.10 Euro if you roll a
number betw
ween 1 and 9.
You can chooose Option A att the beginning annd then switch to Option B for thee rest of the rows or immediately choose
c
Option B for all of the
rows. Pleasee indicate the row
w in which you want
w to choose Opption B for the firsst time.
Your Comppensation: A randdom row will be chosen
c
for your acctual Euro-paymeent.
-Phase 1: Prrofiles
In this section you have to assess how otheer people decided
d in the game decisions I and III that you have already
a
faced. Thhe following
characteristics of those people are available: Level
L
of Educatiion, Age, Incomee, Marital Statuss, Gender, Child
dren, Body Heigh
ht.
-19 differentt demographic pro
ofiles are presentted
-Decision I
o his or her
How do yoou think the individual, whose chharacteristics aree displayed in thee box above deccided in Decisionn I? What part of
winnings diid the described in
ndividual wager on
o the risky but also
a profit-promissing lottery?
Your Comppensation: If you assess
a
the describbed individual’s choice
c
correctly, you
y will receive 0.50
0
Euro.
Decision II:
Please conssider how the inddividual whose characteristics aree displayed in thee box above deciided in Decisionn II. The individuual described
above first cchooses Option B in row: Enter a number between 1 and 11.
Your Comppensation: If you assess
a
the describbed individual’s choice
c
correctly, you
y will receive 0.50
0
Euro.
-Phase 2: Coonversation
In the folloowing section of the experiment you
y will interactt with other partiicipants. The parrticipant that youu interact with with
w has been
determined randomly selecteed form the otherr people in the session. In order too identify particippants, you will recceive a player nuumber for the
remainder of
o the experimentt. Your player nu
umber is XXX.
The sequencce of events for th
he following secttion is as follows:

Yoou will successiveely meet with fou
ur other experim
ment participantts. You will receiive the player num
mber of the partnner you are to
meeet with before thhe meeting. An addministrator will direct
d
you to yourr partner.
17



You will then ask your partner questions and your partner will ask you questions. You will receive a questionnaire before the
meeting.
Please return to your seat as soon as you have answered all the questions from the questionnaire. Wait until the administrators
distribute the next questionnaire and take you to meet your next partner.
Your task is to find out information about your partner during the interview. You will have to assess the risk preferences (in
Decisions I and II) of your partners later in the experiment.
There will be two different types of questionnaires. For the first two interview partners you will use questionnaire A. For the final two
interview partners you will employ questionnaire B.

Questionnaire A: Begin by writing down your own player number and your partner’s player number on the questionnaire. Then
ask the questions from the questionnaire and make sure to write down the answers on the questionnaire. The participant with the
lower player number begins asking the questions. Do not deviate from the questions. Only ask questions from the questionnaire.

Questionnaire B: You will have the opportunity to think of further questions before the start of the interview. Write these additional
questions on the back of the questionnaire. Begin by writing down your own player number and your partner’s player number on the
questionnaire. Ask about both the questions from the questionnaire and the questions that you have thought of yourself. Write down
the answers to all the questions on the questionnaire. The participant with the lower player number begins asking the questions. You
ask answer all questions that you deem to be important as long as you do not directly ask your partner how he or she answered
game decisions I and II. Asking for the choices in decisions I and II will lead to your exclusion from the experiment.

Please note that the conversations are recorded. After all four conversations you will go back to your spot and enter your assessment
to the computer in private.
-These instructions are repeated for all four partners before the conversations (X2, X3, X4, X5). After the conversations the subjects went
back to their computer spots.
-Please read the questionnaire from the conversation with player B. Recall player X2.
-Decision I
Evaluated individual: Assess your first partner with the player number X2. Consider the person’s decision in Decision I. What fraction of his
or her winnings did the described individual wager on the risky but also profit-promising lottery?
Your Compensation: If you assess the described individual’s choice correctly, you will receive 0.50 Euro.
Decision II:
Evaluated individual: Assess your first partner with the player number X2. Consider the person’s decision in Decision II. The person first
chooses Option B in row: Enter a number between 1 and 11.
Your Compensation: If you assess the described individual’s choice correctly, you will receive 0.50 Euro.
-Four repetitions (X2, X3, X4, X5).
References
Ball, S., Eckel, C. C., and Heracleous, M. (2010). Risk Aversion and Physical Prowess:
Prediction, Choice, and Bias. Journal of Risk and Uncertainty 41, 167193.
Crosetto, P. and Filippin, A. (2015). A theoretical and experimental appraisal of four risk
elicitation methods. Experimental Economics, forthcoming.
Dave, C., Eckel, C., Johnson, C. and Rojas, C. (2010). Eliciting risk preferences: When is
simple better? Journal of Risk and Uncertainty 41, 219243.
Deck, C., Lee, J., Reyes, J.A., and Rosen, C. C. (2013). A failed attempt to explain within
subject variation in risk taking behavior using domain specific risk attitudes. Journal of
Economic Behavior and Organization 87, 124.
18
Dohmen, T., Falk, A., Huffman, D., Sunde, U., Schupp, J. and Wagner, G. G. (2005).
Individual Risk Attitudes: New Evidence from a Large, Representative, ExperimentallyValidated Survey. IZA Working Paper 1730.
Dohmen, T., Falk, A., Huffman, D., Sunde, U., Schupp, J., and Wagner, G. (2011). Individual
Risk Attitudes: Measurement, Determinants, and Behavioral Consequences. Journal of
the European Economic Association 9, 522550.
Eckel, C. C. and Grossman, P. J. (2002). Sex differences and statistical stereotyping in
attitudes toward financial risk. Evolution and Human Behavior 23, 281–295.
Filippin, A. and Crosetto, P. (2014). A Reconsideration of Gender Differences in Risk
Attitudes. Working paper, Grenoble.
Fischbacher, U. (2007). Z-Tree: Zurich Toolbox for Ready-Made Economics Experiments.
Experimental Economics 10, 171178.
Friedman, D. R. M. Isaac, D. James, and S. Sunder (2014). Risky Curves: On the empirical
failure of expected utility. Routledge, New York.
Gigerenzer, G., Gaissmaier, W., Kurz-Milcke, E., Schwartz, L. M., and Woloshin, S. (2007).
Helping Doctors and Patients Make Sense of Health Statistics. Psychological Science in
the Public Interest 8, 5396.
Greiner, B. (2015). Subject Pool Recruitment Procedures: Organizing Experiments with
ORSEE. Journal of the Economic Science Association 1, 114125.
Holt, C. and S. Laury (2002). Risk Aversion and Incentive Effects. American Economic
Review 92, 16441655.
Kirchler, M., Lindner, F. and Weitzel, U. (2015). No Time For Losers! Rankings and RiskTaking in the Finance Industry. Working paper, Radboud University.
MiFID (2006). Market in Financial Instruments Directive 2006/31/EC of the European
Parliament and the Council.
Noussair, C.N., S.T. Trautmann, and G. van de Kuilen (2014). Higher Order Risk Attitudes,
Demographics, and Financial Decisions. Review of Economic Studies 81, 325−355.
Roth, B. and Voskort, A. (2014). Stereotypes and false consensus: How financial
professionals predict risk preferences. Journal of Economic Behavior and Organization
107, 553565.
Trautmann, S.T. and van de Kuilen, G. (2015). Belief Elicitation: A Horse Race among Truth
Serums. Economic Journal 125, 21162135.
19