Asking Better Questions How Presentation Formats Influence Information Search

Asking Better Questions: How Presentation Formats Influence
Information Search
Charley M. Wu and Björn Meder
Max Planck Institute for Human Development, Berlin, Germany
Flavia Filimon
Max Planck Institute for Human Development, Berlin, Germany,
and Humboldt University
Jonathan D. Nelson
Max Planck Institute for Human Development, Berlin, Germany
While the influence of presentation formats have been widely studied in Bayesian reasoning tasks, we
present the first systematic investigation of how presentation formats influence information search
decisions. Four experiments were conducted across different probabilistic environments, where subjects
(N ⫽ 2,858) chose between 2 possible search queries, each with binary probabilistic outcomes, with the
goal of maximizing classification accuracy. We studied 14 different numerical and visual formats for
presenting information about the search environment, constructed across 6 design features that have been
prominently related to improvements in Bayesian reasoning accuracy (natural frequencies, posteriors,
complement, spatial extent, countability, and part-to-whole information). The posterior variants of the
icon array and bar graph formats led to the highest proportion of correct responses, and were substantially
better than the standard probability format. Results suggest that presenting information in terms of
posterior probabilities and visualizing natural frequencies using spatial extent (a perceptual feature) were
especially helpful in guiding search decisions, although environments with a mixture of probabilistic and
certain outcomes were challenging across all formats. Subjects who made more accurate probability
judgments did not perform better on the search task, suggesting that simple decision heuristics may be
used to make search decisions without explicitly applying Bayesian inference to compute probabilities.
We propose a new take-the-difference (TTD) heuristic that identifies the accuracy-maximizing query
without explicit computation of posterior probabilities.
Keywords: information search, presentation formats, Bayesian reasoning, probability gain, optimal
experimental design
Supplemental materials: http://dx.doi.org/10.1037/xlm0000374.supp
Before it is possible to arrive at the correct answer, one must
first find the right question. The ability to ask good questions is
essential for cognition and decision-making, because the choice of
query determines what information is acquired, influencing all
subsequent inferences and decisions. Consider a doctor choosing a
test for diagnosing a patient, where a medical test is an example of
a search query with binary outcomes (e.g., positive or negative test
results). Not all tests are equally useful, and a doctor must consider
the probabilities and diagnostic implications of each test outcome
to identify the best test for diagnosing a patient. The term “infor-
mation search” applies to any decision-making task where the goal
is to actively acquire information, including directing eye move-
ments toward informative parts of a scene (Legge, Klitz, & Tjan,
1997; Najemnik & Geisler, 2005, 2008; Nelson & Cottrell, 2007;
Renninger, Verghese, & Coughlan, 2007) or conducting experi-
ments to differentiate between competing hypotheses (Lindley,
1956; Slowiaczek, Klayman, Sherman, & Skov, 1992).
Choosing the right search query requires evaluating the useful-
ness, or, more precisely, expected usefulness of each potential
query, because the outcome of a query is not known before it is
This article was published Online First March 20, 2017.
Charley M. Wu and Björn Meder, Center for Adaptive Behavior and
Cognition, Max Planck Institute for Human Development, Berlin, Ger-
many; Flavia Filimon, Center for Adaptive Behavior and Cognition, Max
Planck Institute for Human Development and Berlin School of Mind and
Brain, Humboldt University; Jonathan D. Nelson, Center for Adaptive
Behavior and Cognition, Max Planck Institute for Human Development.
This research was supported by Grant ME 3717/2-2 to BM and Grant
NE 1713/1-2 to JDN from the Deutsche Forschungsgemeinschaft (DFG) as
part of the priority program “New Frameworks of Rationality” (SPP 1516).
This article is based in part on CMW’s master’s thesis at the University of
Vienna, and early results were presented at the SPUDM 2015 conference
in Budapest, Hungary. We thank Matthias Gloel for designing and piloting
a preliminary version of the Turtle Island experimental scenario, and Gary
Brase, Michelle McDowell, Patrice Rusconi, Laura Martignon, Katya
Tentori, and Vincenzo Crupi for useful feedback.
Correspondence concerning this article should be addressed to Charley
M. Wu, Center for Adaptive Behavior and Cognition (ABC), Max Planck
Institute for Human Development, Lentzeallee 94, 14195 Berlin, Germany.
E-mail: cwu@mpib-berlin.mpg.de
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
Journal of Experimental Psychology:
Learning, Memory, and Cognition
© 2017 American Psychological Association
2017, Vol. 43, No. 8, 1274–1297
0278-7393/17/$12.00 http://dx.doi.org/10.1037/xlm0000374
1274

posed. Forming an expectation about the usefulness of a query
depends on the probabilities of each outcome and their implica-
tions for the hypotheses under consideration (a preposterior anal-
ysis; Raiffa & Schlaifer, 1961). Computationally, calculating the
expected usefulness of a query can be demanding, with part of the
complexity arising from the derivation of posterior probabilities
(e.g., probability that a patient has the disease given a positive test
result). There is a large literature on elementary Bayesian reason-
ing (for reviews see Barbey & Sloman, 2007; Brase & Hill, 2015;
Koehler, 1996), studying the influence of presentation formats on
the ability to derive a posterior probability; however, little is
known about how presentation formats affect human behavior in
other kinds of probabilistic reasoning tasks, such as information
search. Currently, the only literature on this topic has shown that
1–2 hr of experience-based learning (i.e., sequential presentation
of naturally sampled stimuli with immediate feedback) greatly
increased the proportion of accuracy-maximizing search queries
when compared with using numerical conditional probabilities
(Nelson, McKenzie, Cottrell, & Sejnowski, 2010). Our key goal
was to fill this gap in the literature by systematically investigating
how various design features of presentation formats influence
information search behavior. We specifically focused on descrip-
tive presentation formats, including both numerical and visual
formats, because they have the advantage of not requiring the
extensive training time of experience-based learning.
Presentation Formats and Bayesian Reasoning
Human performance in Bayesian reasoning tasks can be sub-
stantially improved by presenting information in terms of natural
frequencies (Gigerenzer & Hoffrage, 1995; Meder & Gigerenzer,
2014; for a meta-analysis see McDowell & Jacobs, 2016). Natural
frequencies represent the Bayesian reasoning problem in terms of
joint frequencies (e.g., number of patients who test positive and
have the disease), which resembles how an individual experiences
events in daily life (Cosmides & Tooby, 1996). Natural frequen-
cies also simplify the calculations necessary to apply Bayes’s rule
for deriving a posterior probability, because base rate information
is preserved (Hoffrage, Gigerenzer, Krauss, & Martignon, 2002;
Kleiter, 1994). This is in contrast to the standard probability
format, which provides information in terms of conditional prob-
abilities (e.g., probability of a positive test result given that the
individual has the disease), which requires introducing base rate
information via Bayes’s rule (Gigerenzer & Hoffrage, 2007). To
illustrate, consider a Bayesian reasoning task with binary hypoth-
eses (disease or no disease) and a single binary outcome (positive
or negative test result). The posterior probability that a patient has
the disease given a positive test result can be computed using
Bayes’s rule:
P(diseaseⱍpos) ⫽
P(disease)P(posⱍdisease)
P(disease)P(posⱍdisease) ⫹ P(no disease)P(posⱍno disease)
(1)
In Equation 1, the use of conditional probabilities requires
explicitly considering base rate information, P(disease) and P(no
disease). Substituting natural frequencies for conditional probabil-
ities, the same posterior probability can be computed without
reintroducing the base rate:
P(disease|pos) ⫽ N(pos ⵩ disease)
N(pos ⵩ disease) ⫹ N(pos ⵩ no disease)
(2)
where N denotes the number of cases for each combination of
disease and positive test result.
The set of presentation formats that have been shown to improve
Bayesian reasoning also includes formats that visualize natural
frequencies (e.g., bar graphs and icon arrays), which have system-
atically yielded superior estimation accuracy over numerical rep-
resentations (using only words and numbers) in Bayesian inference
tasks (Ancker, Senathirajah, Kukafka, & Starren, 2006; Brase,
2009; Galesic, Garcia-Retamero, & Gigerenzer, 2009; Garcia-
Retamero & Hoffrage, 2013; Sedlmeier & Gigerenzer, 2001; but
see Martire, Kemp, Sayle, & Newell, 2014). Additionally, studies
where subjects sequentially experienced single events naturally
sampled from the environment also yielded improvements over the
standard probability format (Lindeman, van den Brink, & Hoog-
straten, 1988; Medin & Edelson, 1988).
From Bayesian Reasoning to Information Search
While many Bayesian reasoning tasks in cognitive psychology
deal with probability judgments about a single binary outcome,
here we are concerned with information search in classification
problems (Skov & Sherman, 1986), which are more complex
probabilistic reasoning tasks. In the studies presented here, sub-
jects were asked to choose between two information search que-
ries, each with binary outcomes. The goal was to choose the query
that would be most useful for improving classification accuracy.
To determine which query is most useful, one must consider the
probability of each outcome (e.g., probability of a positive or
negative test result), as well as the usefulness of each outcome
(e.g., the informational value of a test result for diagnosing a
patient) for the purpose of making a classification decision. Thus,
this type of information search task is considerably more complex
than elementary Bayesian reasoning tasks, because it requires
reasoning about two binary outcomes instead of one, interpreting
the usefulness of each outcome, and forming an expectation to
derive the overall usefulness of the query.
There are many ways to define the usefulness of a query
outcome (e.g., the reduction of uncertainty or improvement of
classification accuracy), corresponding to different information
search goals (for reviews, see Nelson, 2005, 2008). Because it
is outside the scope of this study to determine which goal is
normatively appropriate or optimal, or which goal is the most
accurate description of human search behavior in general, we
used an information search task where the goal was explicitly
defined as the maximization of classification accuracy. This
goal corresponds to selecting queries according to their ex-
pected increase in classification accuracy (i.e., probability gain
or any equivalent metric; Baron, 1985). Nelson and colleagues
(2010) found probability gain to be the best account for how
humans select queries in probabilistic categorization tasks,
when information about the search environment was acquired
through experience-based learning.
Goals and Scope
How do presentation formats influence information search
decisions? We systematically examine both numerical and vi-
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1275
ASKING BETTER QUESTIONS

sual methods of presenting probabilistic information to assess
the important design features that could help facilitate better
information search decisions. We address two main questions.
First, how are search decisions influenced by presentation for-
mats and design features? Second, is search behavior mediated
by probabilistic reasoning, numeracy skill, or both? The first
question serves the practical purpose of identifying how infor-
mation should be presented to improve search behavior. Be-
cause natural frequencies and visualizations thereof have been
shown to improve people’s ability to calculate posterior prob-
abilities (Gaissmaier, Wegwarth, Skopec, Müller, & Broschin-
ski, 2012; Gigerenzer & Hoffrage, 1995; Hoffrage et al., 2002;
Micallef, Dragicevic, & Fekete, 2012), these formats may also
positively influence information search behavior. Our second
question examines the basis of how information search is re-
lated to probabilistic reasoning and numeracy skill. Because the
usefulness of a query depends on the statistical structure of the
environment, we elicited judgments about the relevant proba-
bilistic variables that made up the search environment, with the
hypothesis that more accurate probability judgments would be a
predictor for better search decisions. We also hypothesized that
probability judgment accuracy would be a function of both
presentation format and individual numeracy skill, as numeracy
has been found to be positively related with Bayesian reasoning
accuracy (Chapman & Liu, 2009; Hill & Brase, 2012; McNair
& Feeney, 2015) and other decision making tasks (Bodemer,
Meder, & Gigerenzer, 2014; Peters et al., 2006; Reyna, Nelson,
Han, & Dieckmann, 2009). Figure 1 summarizes our initial
hypotheses about the relationships between key variables.
In the following section, we present prominent models of
information search and describe how they can make different
predictions about which query is most useful. Subsequently, we
describe the theoretical motivations behind the presentation
formats we investigated and present them in detail. Finally, we
describe four experiments investigating the influence of these
presentation formats on human search behavior, each using the
same stimulus and procedure, but with different probabilistic
environments.
Models of Information Search
We considered both Bayesian statistical models and simple
heuristic strategies for modeling search behavior. Bayesian Opti-
mal Experimental Design (OED) models provide a means to
quantify the usefulness of a query, based on the probabilistic
structure of the environment. The OED framework has been
widely used to construct normative and descriptive models of
human information acquisition (Baron, 1985; Bramley, Lagnado,
& Speekenbrink, 2015; Klayman & Ha, 1987; Nelson, 2005;
Savage, 1954; Skov & Sherman, 1986; Slowiaczek et al., 1992).
OED models currently provide the best available computational-
level description (Marr, 1982) of many probabilistic information
search tasks (Gureckis & Markant, 2012; Nelson et al., 2010;
Ruggeri, Lombrozo, Griffiths, & Xu, 2016; Ruggeri & Lombrozo,
2015; for reviews see Markant & Gureckis, 2012; Nelson, 2005).
Yet, it has also been shown in some cases that simple heuristic
strategies can approximate or even exactly implement particular
OED models, thereby establishing a link to psychologically plau-
sible mechanisms (Navarro & Perfors, 2011; Nelson, 2005, 2008;
Nelson, Meder, & Jones, 2016). In the following sections, we
introduce prominent OED and heuristic models of information
search.
OED Models
OED models quantify the expected usefulness of possible que-
ries within a Bayesian decision-theoretic framework (Savage,
1954). For the purposes of our study, each OED model is a
candidate descriptive model of how people might intuitively eval-
uate the expected usefulness of a query. We describe several OED
models, each using a distinct informational utility function1
(also
known as sampling norm; Nelson, 2005). We use capital Q to
denote a query, where lowercase q1, . . . , qm are the m possible
outcomes of the query (e.g., medical test results, or forms of a
features). Because the search tasks presented in our experiments
are classification tasks, we use C ⫽ {c1, . . . , cn} to denote the
different hypotheses (i.e., categories). Equation 3 shows the gen-
eral framework used by all OED models to quantify the expected
usefulness of a query, eu(Q):
eu(Q) ⫽ 兺
j⫽1
m
P(qj)u(qj) (3)
The various OED models differ in how they calculate the
usefulness of each individual query outcome, u(qj), which may
lead to disagreements about which query is most useful, corre-
sponding to different information acquisition goals (Nelson, 2005;
see Figure 2). We describe several prominent OED models below.
Probability gain. Probability gain (PG; Baron, 1985; Nelson,
2005) values a query in terms of its expected improvement in
classification accuracy, assuming that the most probable category
will always be chosen. The model’s informational utility function
is shown in Equation 4, where the max operators choose the
leading (i.e., most likely) hypothesis given the outcome of a query
and the initially leading hypothesis before any query. The differ-
ence between the two terms is the probability gain of a query
outcome:
uPG(qj) ⫽ max
i
P(ci |qj) ⫺ max
i
P(ci) (4)
Probability gain corresponds to what Martignon and Krauss
(2003) have called average validity. The highest probability gain
1
Note that we only consider disinterested utility functions, which do not
factor in situation-specific payoffs or search costs (see Meder & Nelson,
2012). Our experiments use a cover story where payoffs correspond to
classification accuracy and there are no explicit search costs.
Figure 1. The hypothesized relationships between presentation format,
numeracy, probability judgment accuracy, and information search.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1276 WU, MEDER, FILIMON, AND NELSON

query is the query that results in the highest expected improvement
in classification accuracy over the initial best guess (i.e., relying
solely on base rate information). Probability gain has been used as
model for a variety of information acquisition tasks, such as
experience-based categorization (Nelson et al., 2010), the predic-
tion of eye movements where the goal is to find a target in a
cluttered environment (Najemnik & Geisler, 2008), and in medical
test selection (Baron, Beattie, & Hershey, 1988). In all of our
experiments in this article, we explicitly identify the goal of the
information search task as the maximization of classification ac-
curacy, which implies that probability gain should be used to
identify the most useful query.
Information gain. Information gain (IG) quantifies how
much a query outcome reduces the uncertainty about the hypoth-
eses, where uncertainty is measured using Shannon (1948) en-
tropy2
(Lindley, 1956):
uIG(qj) ⫽ 兺
i⫽1
n
P(ci) log2
1
P(ci)
⫺ 兺
i⫽1
n
P(ci |qj) log2
1
P(ci ⱍqj)
(5)
While the reduction of Shannon entropy can correspond to
improved classification accuracy, sometimes information gain and
probability gain strongly disagree about which query is more
useful (Figure 2; see Nelson, 2005; Nelson et al., 2010). In Ex-
periments 1 and 2, we specifically studied search behavior in
environments where the query with the highest expected reduction
of Shannon entropy does not increase expected classification ac-
curacy at all. It should also be noted that the expected Kullback-
Leibler (KL) divergence of a query (Kullback & Leibler, 1951) is
exactly equivalent to its expected information gain (Oaksford &
Chater, 1996), although the usefulness of individual outcomes may
differ. Therefore, when we describe information gain making a
prediction about query selection, it should be understood that KL
divergence always makes the same prediction.
Impact. Impact quantifies the usefulness of a query as the
absolute change in beliefs (Nelson, 2005, 2008; Wells & Lindsay,
1980), from the prior probability to the posterior probability of the
hypotheses conditional on a query outcome (Equation 6):
uImpact(qj) ⫽ 兺
i⫽1
n
|P(ci|qj) ⫺ P(ci)| (6)
In the case of a binary category classification task where the
base rates are equiprobable (Experiment 4), the highest impact
query also has the highest probability gain (Nelson, 2005). If the
base rates are not equiprobable, probability gain and impact can
disagree about which query is more useful, as illustrated by the
search environments used in Experiments 1 and 2.
Heuristic Models
In addition to several OED models, we considered the possibil-
ity that human search behavior might be best described with
simple heuristic strategies. In principle, all of the strategies we
consider, OED models and heuristics alike, can be applied to all
presentation formats and probabilistic environments in our exper-
iments. However, the heuristics operate directly on specific subsets
of the environmental probabilities, for example test likelihoods,
and may be easier to use given particular presentation formats.
More important, under certain conditions heuristic strategies im-
plement the same behavior as OED models (Klayman & Ha, 1987;
Navarro & Perfors, 2011; Nelson, 2005; Nelson, Divjak, Gud-
mundsdottir, Martignon, & Meder, 2014; Nelson et al., 2016).
Likelihood difference heuristic. The likelihood difference
heuristic (likDiff; Nelson, 2005; Slowiaczek et al., 1992) chooses
the query with the largest absolute difference in feature likelihoods
(i.e., test outcomes), for either query outcome. We can describe
this using an expected usefulness function, analogous to the OED
models, even though the likelihood difference heuristic does not
explicitly compute expectations:
eulikDiff(Q) ⫽ ⱍP(q1ⱍc1) ⫺ P(q1ⱍc2)ⱍ⫽ ⱍP(q2ⱍc1) ⫺ P(q2ⱍc2)ⱍ (7)
The likelihood difference heuristic has been proven to invariably
select the query with the highest impact3
(Nelson, 2005). In a medical
diagnosis scenario, one could apply the likelihood difference heuristic
by selecting the test with the largest absolute difference between the
2
We use Shannon entropy with log2 to measure information gain in bits,
although the choice of logarithm base is arbitrary.
3
One caveat is that the likelihood difference heuristic only applies for
binary classification tasks with binary features, whereas impact can apply
in situations with multivalued features and any number of categories
(Nelson, 2005).
Figure 2. An illustration of how models can make divergent predictions
about which query is more useful. In Experiment 1, probability gain
predicts selection of the DE test, while information gain predicts selection
of the LM test. Probability gain measures the usefulness of a query based
on how much the outcome is expected to increase the probability of making
a correct classification, P(correct)post – P(correct)prior, while information
gain measures usefulness in terms of the reduction in Shannon entropy (in
bits), Entropyprior – Entropypost. Although the LM test has a larger expected
reduction in entropy, there is no change in expected classification accuracy,
whereas both outcomes of the DE test result in an increase in classification
accuracy. Thus, probability gain predicts selection of the DE test, while
information gain predicts the LM test. See the online article for the color
version of this figure.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1277

true positive, P(positive|disease), and the false positive rate,
P(positive|no disease).
Probability of certainty heuristic. The probability of cer-
tainty heuristic (ProbCertainty) selects the query (if any) with the
highest probability of an outcome granting certainty about the true
hypothesis (Nelson et al., 2010). A query outcome has certainty
when the posterior probability of one of the hypotheses given an
outcome qj is 1. Analogous to the likelihood difference heuristic,
we can describe the probability of certainty heuristic using the
OED framework by assigning the usefulness of a query outcome to
1 if it implies certainty about a hypothesis, and zero otherwise:
uProbCertainty(qj) ⫽ 再1 if maxi⫽1
n
P(ci|qj) ⫽ 1
0 otherwise
(8)
Note that an informational utility function numerically very
similar to probability of certainty can be derived by substituting a
high order Tsallis (1988) entropy in place of Shannon entropy
within the information gain model (Crupi, Nelson, Meder,
Cevolani, & Tentori, 2016). Thus, probability of certainty could
also be viewed as a type of generalized information gain model.
The probability of certainty heuristic was tested in Experiments 2
and 3, where we introduced certain query outcomes, and explored
how the possibility of obtaining a certain outcome influenced
search behavior.
Information Search Scenario
We devised an information search task involving the classifica-
tion of artificial turtle stimuli to examine the influence of presen-
tation formats on search behavior.4
In all experiments, participants
were told that 100 turtles live on a remote island, with each turtle
belonging to either the “Freqosian” or “Bayosian” species. The
two species of turtles look identical, but differ in their genetic
makeup. Thus, identification of a turtle can be aided with the help
of one of two genetic tests, which constitute the available infor-
mation search queries. The DE test yields truthful information
about whether the D form or the E form of the DE gene is present
in the turtle, while the LM test yields truthful information about the
L form or the M form of the LM gene. While each turtle possesses
one form of the DE gene and one form of the LM gene, the extent
to which these gene forms are present varies between the two
species of turtles. Thus, each test outcome provides probabilistic
information about the true species of the turtle. This is a binary
information search task, where there are two possible classes, C ⫽
{Bayosian, Freqosian}, and two possible search queries, QDE and
QLM, with each query having two possible outcomes, QDE ⫽ {D,
E} and QLM ⫽ {L, M}. Participants were given the explicit goal of
choosing the test that maximizes the probability of correctly clas-
sifying a randomly encountered turtle from the island.
Presentation Formats
We presented information about the search scenario using 14
different numeric and visual formats. While many presentations
formats have been studied in the context of Bayesian reasoning,
including tree diagrams (Binder, Krauss, & Bruckmaier, 2015;
Sedlmeier & Gigerenzer, 2001), signal detection curves (Cole,
1989; Cole & Davidson, 1989), and Bayesian boxes (Burns, 2004),
our formats were chosen for the purpose of systematically studying
the influence of six design features (see Table 1) that have been
prominently related to improvements in elementary Bayesian rea-
soning. All design features are binary and are linked to theories
about how people perform probabilistic reasoning, such as pre-
senting natural frequencies instead of conditional probabilities, or
highlighting part-to-whole information. All formats provide com-
plete information about the probabilistic search environment, and
can be used to implement any of the OED or heuristic strategies.
However, the formats are not equivalent in the ease of extracting
or computing specific pieces of information about the search
environment, such as test likelihoods or posterior probabilities. In
Herbert Simon’s (1978) terms, the formats are informationally
equivalent, but not necessarily computationally equivalent. Thus,
people’s tendency to use particular heuristic or OED strategies
could depend on which format is used to present the probabilistic
task information.
Numerical formats. Four conditional probability formats and
four natural frequency formats collectively comprise the numerical
formats, which use words and numbers to express probabilistic
information (see Table 2). All conditional probabilities were
rounded to the nearest whole percentage point, while natural
frequencies were derived from a population of 100 turtles. Each set
of four formats was constructed using a 2 ⫻ 2 factorial design, (a)
varying the type of information presented (either likelihoods or
posteriors) and (b) varying the presence or absence of complement
information about each binary variable. These two factors are
described in detail in the subsequent design feature section.
Visual formats. We tested six different visual formats, com-
prised of two types of icon arrays, two types of bar graphs, and two
types of dot diagrams (see Table 3). Each visual format has a
likelihood variant and a posterior variant, providing a visual rep-
resentation of the corresponding natural frequency format (Freq-
Lik⫹ or FreqPost⫹). Stimuli were grouped by species (likeli-
hoods) or by test outcome (posteriors) in the same way as the
numerical natural frequency formats. The colors used in all visual
formats were defined using a tool called Paletton (http://paletton
.com), which allowed us to construct a symmetrically divergent
and colorblind-friendly palette.
Icon arrays and bar graphs are two of the most common
visualizations explored in the Bayesian reasoning literature (An-
cker et al., 2006; Brase, 2009; Gaissmaier et al., 2012; Galesic et
al., 2009), using the number of icons or the length of a bar to
represent the number of data points with each joint outcome (i.e.,
each relationship between species and test outcome). Icon arrays
and bar graphs are unique out of the 14 formats, because they
utilize spatial extent (i.e., the length of a bar or icon array) to
visualize natural frequency information. We used a variant of icon
arrays that are more comparable with bar graphs (i.e., a separate
array for each joint outcome), rather than stacked into a single unit
(in contrast to Brase, 2009; Galesic et al., 2009; Garcia-Retamero
& Hoffrage, 2013), allowing us to examine differences in count-
ability (see Table 1).
We introduce the dot diagram as a novel format inspired by
Euler diagrams (alternatively called “Venn diagrams”), which
have been found to be successful at improving Bayesian reasoning
4
For screenshots of the experiment see supplemental material Figures
1–3.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

(Brase, 2009; Micallef et al., 2012; Sloman, Over, Slovak, &
Stibel, 2003). One main difference is that Euler diagrams present
only one particular result of a binary query, whereas the dot
diagram is designed to show both of the possible test outcomes
equivalently, to fairly present an information search task, as op-
posed to a Bayesian reasoning task. The dot diagram uses Uniform
Poisson Disk Sampling (Lagae & Dutré, 2008) to place the indi-
vidual dots in an approximation of a uniform random distribution.
Dots (each representing a single item) are distributed within con-
tainers that highlight the part-to-whole relationships between spe-
cies and test outcomes. To avoid idiosyncrasies of a particular
random distribution of dots affecting search behavior, we gener-
ated 20 dot diagrams for each set of probabilities and selected one
at random for each participant assigned to the dot diagram condi-
tion.
Design Features
Natural frequencies. Conditional probabilities present differ-
ent quantitative information than natural frequencies (both nu-
meric and visual representations, where all visual formats in this
investigation are representations of natural frequencies). Condi-
tional probability formats normalize likelihoods and posterior
probabilities (i.e., to the interval [0, 1]) irrespective of the prior or
marginal probabilities (Gigerenzer & Hoffrage, 2007), whereas
natural frequencies express information about likelihoods and
posteriors without normalization, incorporating the base rate and
marginal probabilities. Natural frequencies are constructed as out-
comes of natural sampling (Kleiter, 1994) and are frequently
associated with higher Bayesian reasoning accuracy than condi-
tional probabilities (Gigerenzer & Hoffrage, 1995; Hoffrage et al.,
2002; Zhu & Gigerenzer, 2006; for reviews see Brase & Hill,
2015; McDowell & Jacobs, 2016). We hypothesized that the
advantage of natural frequencies would also carry over to our
information search task.
Posterior information. For all formats, we tested a variant
presenting information in terms of likelihoods and a variant pre-
senting posterior probabilities. The likelihood variants provide
information about the base rates (i.e., the distribution of the turtle
species before any tests being performed), along with the likeli-
hoods of each test outcome relative to a given species (e.g., the
likelihood of a D outcome of the DE test if the turtle is Bayosian).
The posterior variants provide information about the marginals
(i.e., the marginal probability of a test outcome independent of
species), as well as the posterior probability of a turtle belonging
to a species given a specific outcome (e.g., the probability that a
turtle is Bayosian given the D outcome of the DE test). It should
be understood that because natural frequencies do not renormalize
information, the same set of numerical quantities are presented in
both likelihood and posterior variants, but are grouped differently
(grouped by species or by outcome, respectively), whereas the
conditional probability formats present distinctly different numer-
ical quantities (because of normalization). We hypothesized that
posterior variations may be more helpful for information search
tasks, since all of the OED models make use of the posterior
probabilities in the calculation of usefulness. If the posteriors do
not need to be inferred, this simplifies the required computations.
However, the likelihood difference heuristic does not require pos-
terior probabilities, and it is possible that other heuristic strategies
Table
1
Design
Features
of
Presentation
Formats
Design
feature
Numerical
formats
Visual
formats
Conditional
probabilities
Natural
frequencies
Icon
arrays
Bar
graphs
Dot
diagrams
ProbLik
ProbLik⫹
ProbPost
ProbPost⫹
FreqLik
FreqLik⫹
FreqPost
FreqPost⫹
IconLik
IconPost
BarLik
BarPost
DotLik
DotPost
Natural
frequencies
0
0
0
0
1
1
1
1
1
1
1
1
1
1
Posteriors
0
0
1
1
0
0
1
1
0
1
0
1
0
1
Complement
0
1
0
1
0
1
0
1
1
1
1
1
1
1
Spatial
extent
0
0
0
0
0
0
0
0
1
1
1
1
0
0
Countability
0
0
0
0
1
1
1
1
1
1
0
0
1
1
Part-to-whole
information
0
0
0
0
1
1
1
1
0
0
0
0
1
1
Note.
1
denotes
the
presence
of
a
design
feature
and
0
denotes
the
absence.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1279

making the same predictions as OED models would not need to
compute posteriors either.
Complement. Each variable in the probabilistic search envi-
ronment is binary, that is, a turtle is either Bayosian or Freqosian.
We tested the differences between presenting information with
complement (e.g., 70% of turtles are Bayosian and 30% of turtles
are Freqosian) and presenting information without complement
(e.g., 70% of turtles are Bayosian). Rusconi and McKenzie (2013)
have shown that the inclusion of complement information with the
standard probability format improved sensitivity to the informa-
tiveness of an answer, which we hypothesized would also be
helpful for making search decisions. This design feature was not
manipulated in the visual formats, all of which intrinsically include
complement information.
Spatial extent. The icon arrays and bar graphs presented in
this investigation convey quantitative information using the
spatial extent of an array of icons or the length of a bar, where
there is a fixed ratio of area to length. In contrast, the numerical
formats use words and numbers to convey information, and the
dots in the dot diagrams are restricted to a container of fixed
size, conveying quantitative information using density rather
than spatial extent. Empirical work by Cleveland and McGill
(1985; Heer & Bostock, 2010) has shown that graphical repre-
sentations exploiting basic perceptual abilities, such as length
comparisons, can engage automatic visual perceptual mecha-
nisms, thus, reducing the required mental computation and
leading to higher quantitative reasoning capabilities. Perceptual
accuracy is highest when judging lengths against a common
scale (e.g., when comparing the lengths of bars in a bar graph
with a common axis), and progressively worse for comparing
area (e.g., circles) and volumes or densities (Ancker et al.,
2006). This would suggest that visualizations of natural fre-
quencies using spatial extent (icon arrays and bar graphs) would
reduce the computational complexity of the task and lead to
better search decisions than formats using numbers (conditional
probabilities and natural frequencies) or density (dot diagrams).
Countability. By countability we refer to the presentation
of quantitative information in discrete units. Conditional prob-
abilities are not countable because they represent statistical
information on a continuous scale (i.e., range from 0 to 1),
whereas the numerical natural frequencies use words and num-
bers to present discrete frequencies. However, when natural
frequencies are translated into visualizations, countability is not
necessarily preserved. With respect to the six design features
(see Table 1), the only difference between the bar graphs and
icon arrays in our experiments is that the discrete number of
icons in an array can easily be counted, whereas the length of a
bar represents information on a continuous scale rather than in
discrete units. If the countability of formats has a positive effect
on Bayesian reasoning or information search, we would expect
to see higher performance for icon arrays compared with bar
graphs. The dot diagrams are countable in principle, although
the random distribution makes it considerably more difficult to
arrive at the exact number. Brase (2009) found that dotted Venn
Table 2
Numerical Formats
Format Quantity Example
Standard probability
(ProbLik)
P(species) Consider a turtle picked at random from the 100 turtles on the island: The probability that it is a
Bayosian turtle is 70%.
P(outcome|species) If a turtle is a Bayosian, then the probability that it has the D form of the DE gene is 3%.
Standard probability with
complement
(ProbLik⫹)
P(species) Consider a turtle picked at random from the 100 turtles on the island: The probability that it is a
Bayosian turtle is 70%, and the probability that it is a Freqosian turtle is 30%.
P(outcome|species) If a turtle is a Bayosian, then the probability that it has the D form of the DE gene is 3%, and
the probability that it has the E form is 97%.
Posterior probability
(ProbPost)
P(outcome) Consider a turtle picked at random from the 100 turtles on the island: The probability that it has
the D form of the DE gene is 11%.
P(species|outcome) If a turtle has the D form of the DE gene, then the probability that it is a Bayosian turtle is 19%.
Posterior probability with
complement
(ProbPost⫹)
P(outcome) Consider a turtle picked at random from the 100 turtles on the island: The probability that it has
the D form of the DE gene is 11%, and the probability that it has the E form is 89%.
P(species|outcome) If a turtle has the D form of the DE gene, then the probability that it is a Bayosian turtle is 19%,
and the probability that it is a Freqosian turtle is 81%.
Natural frequency
(FreqLik)
N(species) Out of the 100 turtles on the island, 70 are Bayosian turtles.
N(outcome ⵩ species) Out of the 70 Bayosian turtles, 2 turtles have the D form of the DE gene.
Natural frequency with
complement
(FreqLik⫹)
N(species) Out of the 100 turtles on the island, 70 are Bayosian turtles and 30 are Freqosian turtles.
N(outcome ⵩ species) Out of the 70 Bayosian turtles, 2 turtles have the D form of the DE gene, and 68 turtles have the
E form of the gene.
Posterior frequency
(FreqPost)
N(outcome) Out of the 100 turtles on the island, 11 have the D form of the DE gene.
N(species ⵩ outcome) Out of the 11 turtles with the D form of the DE gene, 2 are Bayosian turtles.
Posterior frequency with
complement
(FreqPost⫹)
N(outcome) Out of the 100 turtles on the island, 11 have the D form of the DE gene, and 89 turtles have the
E form of the gene.
N(species ⵩ outcome) Out of the 11 turtles with the D form of the DE gene, 2 are Bayosian turtles and 9 are Freqosian
turtles.
Note. Examples of conditional probability and numerical natural frequencies values from Experiment 1. The corresponding values for all experiments are
available in supplemental material Tables 1–4. The short names for each format are provided in brackets below the full name. P(•) refers to the probability
of an item and N(•) refers to the natural frequency of an item.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

Table 3
Visual Formats in Experiment 1
Format Example
Icon array
(IconLik)
Posterior icon array
(IconPost)
Bar graph
(BarLik)
Posterior bar graph
(BarPost)
Dot diagram
(DotLik)
Posterior dot diagram
(DotPost)
Note. Full format examples for all experiments available in supplemental material Tables 1–4. Short names for each
format are provided in brackets below the full name. See the online article for the color version of this figure.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1281

diagrams sometimes resulted in better Bayesian reasoning per-
formance than Venn diagrams solely using the area of a con-
tainer to communicate quantitative information, while Stone et
al. (2003) independently proposed that the ability to derive
exact quantities from a visual format accounted for increased
risk perception (i.e., increased willingness to pay for a product
with a lower posterior probability of a negative outcome).
These hypotheses were further tested by Micallef et al. (2012),
who found that Euler diagrams with randomly distributed dots
(similar to dotted Venn diagrams) led to the highest Bayesian
reasoning performance, compared with other variations that
incorporated ordered icon arrays and better facilitated counting,
in contrast to Brase (2009). Thus, it is not yet clear whether
countability is important or helpful for Bayesian reasoning,
although it may be beneficial in a search task.
Part-to-whole information. By part-to-whole information
we refer to the ability to relate the proportion of a set of objects
(e.g., patients who test positive and have the disease) to any
larger set of objects (e.g., patients who test positive). A review
by Ancker and colleagues (2006) on the effectiveness of visual
formats in conveying health-related risk proposed that part-to-
whole information substantially improved risk perception. Sim-
ilar notions of “nested-set relations” (Neace, Michaud, & Bol-
ling, 2008; Sloman et al., 2003; Stone et al., 2003) or the
“subset principle” (Johnson-Laird, Legrenzi, Girotto, Legrenzi,
& Caverni, 1999) have also been proposed as explanation for
the effectiveness of natural frequencies in Bayesian reasoning
tasks, which by definition contain part-to-whole information,
whereas conditional probabilities do not (Gigerenzer & Hof-
frage, 2007). Because numeric formats are also differentiated
by other design features, we used differences between visual
formats to examine the influence of part-to-whole information
on search decisions. We designed the dot diagrams to provide
accessible information about the part-to-whole relationships
between all features of the environment, for example, that the
number of Bayosian turtles with the D form of the DE gene
represents a portion of the total number of Bayosian turtles, as
well as a portion of the total number of turtles with the D form.
In contrast, the design of our icon arrays and bar graphs do not
present part-to-whole information with the same accessibility as
the dot diagrams, because each combination of gene form and
species is expressed as a separate entity. Numeric natural frequen-
cies are always expressed in relation to the larger set of objects
(e.g., out of the 70 Bayosian turtles, 2 turtles have the D form of
the DE gene), and provide more accessible part-to-whole informa-
tion than icon arrays or bar graphs. If part-to-whole information
has a positive influence on search behavior, we would expect to
see more correct search choices for the dot diagrams and numeric
natural frequencies than for icon arrays or bar graphs.
Experiments
We conducted four different experiments using the same pro-
cedure, but each with a unique probabilistic environment. We used
optimal experimental design principles to generate statistical en-
vironments in which various subsets of OED and heuristic models
made divergent predictions about which query is more useful.
Accordingly, these environments are not necessarily representative
of the distribution of search environments in the real world, but are
meant to “stress test” the presentation formats in scientifically
interesting cases where competing models disagree. Table 4 pro-
vides the parameters for each experiment, while Table 5 provides
an overview of model predictions across the search environments.
Experiment 1
In Experiment 1 we used a search environment where probabil-
ity gain maximally disagreed with information gain, under the
constraint that no queries can lead to certain outcomes.5
The result
is a search environment where probability gain predicts selection
of the DE test, while information gain, impact, and the likelihood
difference heuristic make the opposite prediction (LM test); prob-
ability of certainty makes no prediction in this experiment because
no query outcome provides certainty about the true species. Here
we present the general methods used in all experiments.
Method
Participants and design. All experiments were conducted
using the Amazon Mechanical Turk (AMT) Platform. The Eth-
ics Committee of the Max Planck Institute for Human Devel-
opment approved the methodology and all participants con-
sented to participation through an online consent form at the
beginning of the survey. Human Intelligence Tasks (HITs) were
published exclusively to experienced AMT workers who had
completed at least 1,000 HITs and had at least a 95% accep-
tance rate on previous tasks. In total, 821 participants com-
pleted Experiment 1, with four excluded because of missing
data or because of a self-reported elementary or limited under-
standing of English. The final sample included 817 participants
(46% female, median age of 32 years, range 18–76). To make
sure there was no overlap of participants between experiments,
once a HIT was completed, the subject’s AMT worker ID was
added to the exclusion criteria for subsequent studies. Each
participant who completed the HIT was paid a fixed amount of
$1.50 USD, meeting the minimum hourly wage recommended
by Paolacci, Chandler, and Ipeirotis (2010) for experiments on
the AMT platform. The same exclusion criteria were applied to
all other experiments. Table 6 provides a full description of
demographic information for all experiments.
In Experiment 1, participants were randomly assigned to 1 of 14
different presentation formats upon accepting the HIT. In subse-
quent experiments, we adopted a prerandomized list, to better
equalize the number of participants within each condition. Within
each presentation format, participants were also randomly assigned
to 1 of 16 different randomizations of the probability values, to
5
A two feature, binary category search environment can be fully described
using a single prior and four likelihood probabilities, assuming class-
conditional independence (Jarecki, Meder, & Nelson, 2016). We randomly
generated 1 million random variations of these five probabilities under the
constraint that all posterior probabilities belonged in the range of .05 and .95.
The environment used in Experiment 1 had the largest pairwise disagreement
strength between probability gain and information gain, where we measured
the preference strength of a model as the normalized difference between
the expected usefulness of two queries, PStrm ⫽ 100共eum共Q1兲 ⫺ eum
共Q2兲兲 ⁄ maxPstrm, and disagreement strength as the geometric mean between
the preference strength of two opposing models m1 and m2, DStrm1m2⫽
(|PStrm1 | ⫻ |PStrm2 | )0.5, if PStrm1⫻PStrm2ⱕ0. See Nelson et al.
(2010) for a full description of this process.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

avoid possible confounds with aspects of the stimulus (e.g., the
naming of the turtle species, the naming of the genes, and choice
of colors in the visual formats).6
For the analyses, the 16 random-
izations were recoded to match a single canonical variation, de-
scribed in Table 4. For simplicity, we present each of the experi-
ments using a randomization where selecting the DE test is the
correct choice given the goal of maximizing classification accu-
racy.
Procedure and materials. Once participants gave their con-
sent to participate in the study, they were presented with the Turtle
6
For example, swapping the query outcomes for the DE and LM test
changes which query is most useful respective to each of the considered
models. We then recoded the test selection and the probability judgments
relative to the assigned randomization.
Table 4
Search Environments
Probabilities
Tree diagrams for Exp. 1
Parameters
Experiment
1
Experiment
2
Experiment
3
Experiment
4
Base rate
DE test LM test
70%
D E D E
3% 97% 30% 70%
L M L M
41% 59% 93% 7%
30% 70% 30%
Bayosian Freqosian Bayosian Freqosian
Conditional Probability
DE test LM test
70
D E D E
2 68 9 21
L M L M
29 41 28 2
30 70 30
Bayosian Freqosian Bayosian Freqosian
Natural Frequency
Likelihood Trees:
P(Bayosian) .7 .7 .72 .5
P(Freqosian) .3 .3 .28 .5
Likelihoods
P(D|Bayosian) .03 .04 .03 .1
P(E|Bayosian) .97 .96 .97 .9
P(D|Freqosian) .3 .37 .83 .8
P(E|Freqosian) .7 .63 .17 .2
P(L|Bayosian) .41 .43 .39 .1
P(M|Bayosian) .59 .57 .61 .9
P(L|Freqosian) .93 1 1 .3
P(M|Freqosian) .07 0 0 .7
Marginals
DE test LM test
11%
Bay. Freq. Bay. Freq.
19% 81% 76% 24%
51% 49% 95% 5%
89% 57% 43%
D E L M
Species Species Species Species
Conditional Probability
DE test LM test
11
2 9 68 21
29 28 41 2
89 57 43
D E L M
Species Species Species Species
Natural Frequency
Posterior Trees:
P(D) .11 .14 .25 .45
P(E) .89 .86 .75 .55
P(L) .57 .6 .56 .2
P(M) .43 .4 .44 .8
Posteriors
P(Bayosian|D) .19 .2 .09 .11
P(Freqosian|D) .81 .8 .91 .89
P(Bayosian|E) .76 .78 .94 .82
P(Freqosian|E) .24 .22 .06 .18
P(Bayosian|L) .51 .5 .5 .25
P(Freqosian|L) .49 .5 .5 .75
P(Bayosian|M) .95 1 1 .56
P(Freqosian|M) .05 0 0 .44
Note. Squares in the tree diagrams denote decision nodes and circles denote chance nodes. See the online article for the color version of this figure.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1283

Island story. Before proceeding, participants had to correctly an-
swer a comprehension question to ensure that they had understood
the nature of the task. Specifically, participants were asked one of
four randomly selected questions about the binary nature of the
DNA test, for example, “If a turtle does not have the D form of the
DE gene, then which form of the gene does it have?” Responses
were selected from one of the following options: “D form,” “E
form,” “L form,” or “M form.” Participants were required to
correctly answer this question before continuing the study. If an
incorrect answer was given, participants were asked to reread the
instructions and attempt a new randomly selected question.
Search task and probability judgments. After completing
the comprehension question, participants were given informa-
tion on the environmental probabilities in one of the 14 ran-
domly assigned formats, and asked to choose the test that would
yield the highest chance of correctly classifying the species of
a randomly selected turtle. Specifically, they were asked:
“Which test is better for having the highest chance of correctly
classifying the animal as either a Freqosian or a Bayosian
turtle?”
Once participants made a search decision, they were asked to
give their estimates for 11 different probabilistic variables in the
search environment. To avoid potential memory confounds, the
assigned presentation format was shown again for the probability
judgment task. The questions were arranged in blocks, with three
questions referring to the prior probability of the species and the
Table 5
Expected Utilities and Query Predictions for Each Search Environment
Experiment Model
Prediction
(test)
DE test LM test
P(D) u(D) P(E) u(E) eu(DE) P(L) u(L) P(M) u(M) eu(LM)
1 Probability gain DE 11% .11 89% .06 .07 57% ⫺.19 43% .25 0
Information gain LM .18 .09 .1 ⫺.12 .6 .19
Impact LM 1.02 .12 .22 .38 .5 .44
Prob. certainty — 0 0 0 0 0 0
Information gain LM .16 .12 .13 ⫺.12 .88 .28
Impact LM 1 .16 .28 .4 .6 .48
Prob. certainty LM 0 0 0 0 1 .4
Information gain DE .44 .51 .49 ⫺.14 .86 .29
Impact DE 1.26 .44 .65 .44 .56 .5
Prob. certainty LM 0 0 0 0 1 .44
4 Probability gain DE 45% .39 55% .32 .35 20% .25 80% .06 .1
Information gain DE .5 .32 .4 .19 .01 .05
Impact DE .78 .64 .7 .5 .12 .2
Prob. certainty — 0 0 0 0 0 0
Note. Predictions for each model correspond to the search environments described in Table 4. For each test, P(•) denotes the probability of the outcome,
u(•) denotes the utility of the outcome, and eu(•) denotes the expected utility of the test. All of the OED models compute eu(•) through a normalized mean,
weighting each individual outcome utility, u(•), by the probability of the outcome, P(•). The expected utility for the most useful query is shown in bold.
The likelihood difference heuristic invariably makes the same prediction as impact (proof in Nelson, 2005), so it is not listed separately. The probability
of certainty (prob. certainty) heuristic makes no predictions in Experiments 1 and 4, because neither of the queries contains a certain outcome.
Table 6
Participant Demographics
Variable
Experiment
1
Experiment
2
Experiment
3
Experiment
4
Final N 817 683 681 677
Completed HITs 821 690 684 686
Excluded 4 7 3 9
Average completion time in minutes (⫾SD) 13.8 (⫾7.7) 14 (⫾19.9) 13.1 (⫾8.8) 14.5 (⫾8.5)
Gender 46% female 45% female 48% female 50% female
Age
Median 32 31 31 32
Mean 34 34 34 35
Range 18–76 18–83 18–73 18–71
Education
High school 14% 15% 14% 12%
Some university 34% 33% 37% 39%
Bachelor’s degree or higher 51% 52% 48% 49%
Other 1% 0% 1% 1%
Note. Participants were excluded because of missing data or a self-reported elementary or limited understand-
ing of English. Percentages may not add up to 100% because of rounding.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

marginal probability of the outcomes of each test, four questions
concerning the likelihood of a test outcome given the species, and
four questions regarding the posterior probabilities of the species
given the test outcomes (supplemental material Figures 4–6). The
order of the three blocks was randomized across participants.
Responses were given using a visual analog scale slider. As the
slider was adjusted, text displayed the two complementary values
of the slider (e.g., “Bayosian turtles: 70%” and “Freqosian turtles:
30%”) corresponding to its position to the nearest percentage
point.
Numeracy test. Subsequently, participants completed a nu-
meracy test to assess their ability to comprehend and reason with
numerical information. Numeracy has been shown to correlate
positively with accuracy in Bayesian reasoning (Brown et al.,
2011; Hill & Brase, 2012) and other tasks (Peters et al., 2006). We
used a hybrid test consisting of the Schwartz, Woloshin, Black,
and Welch (1997) numeracy test and the adaptive Berlin Nu-
meracy Test (BNT; Cokely, Galesic, Schulz, Ghazal, & Garcia-
Retamero, 2012). The choice to combine the two tests follows the
recommendation of Cokely and colleagues (2012), who found that
this method offers the best discriminability specific to the broad
range of numeracy levels within an AMT sample population. The
BNT is well suited for discriminating among highly numerate
populations, whereas the Schwartz et al. test is better for discrim-
inating among relatively less numerate populations. The scores for
the two tests were added together, yielding a combined numeracy
score in the range of 1 to 7. Upon completion of the numeracy test,
participants provided demographic information and were de-
briefed.
Results
Information search decisions. Figure 3A shows the results of
the information search task, with the height of the bars representing
the proportion of correct (i.e., accuracy-maximizing) queries for each
format. Overall, 61% of participants chose the correct query, with the
A
B
C D E
Figure 3. Experiment 1 results. (A) Proportion of correct (accuracy-maximizing) search queries by format,
(B) probability judgment error by format, (C) correct search responses by judgment error split into quartiles,
(D) correct search responses by numeracy score, and (E) probability judgment error by numeracy. Bar
graphs displays Agresti-Coull 95% confidence interval (CI). Box plot whiskers indicate 1.5 IQR, with the
notch illustrating bootstrapped 95% CI of the median (10k replications). See the online article for the color
version of this figure.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1285

proportion varying between 27% and 86% across formats. A ␹2
test found that search decisions were strongly affected by
format, ␹2
(13, N ⫽ 817) ⫽ 78.57, p ⬍ .001. The standard
probability format (ProbLik) was the least helpful format for
identifying the correct query, with only 27% of participants
making this choice. This result was virtually identical to Nelson
et al. (2010) who used the standard probability format in a
similar search task. Participants given the posterior icon array
(82% correct choices) or posterior bar graph (86% correct
choices), both visual formats, achieved the highest proportion
of correct choices, and were able to meet the same levels of
performance as those who had undergone 1–2 hr of experience-
based learning in Nelson et al. (2010).
Consistent with the findings reported in the Bayesian reasoning
literature (e.g., Gigerenzer & Hoffrage, 1995), the natural fre-
quency format (FreqLik) yielded more correct responses (51%)
than the standard probability format (ProbLik; 27%). However,
different variants of numerical natural frequency formats did not
consistently outperform the different variations of conditional
probability formats. The posterior probability with complement
(ProbPost⫹) was most effective out of all numerical formats (77%
correct choices), suggesting that conditional probabilities can be
helpful in information search tasks if the complement is added and
if posterior probabilities are given, rather than in the form of priors
and likelihoods associated with the standard probability format.
We conducted a logistic regression to model search deci-
sions, using the design features of the assigned presentation
format, individual numeracy level, and probability judgment
error as predictors (Table 7). Design features were coded as a
binary vector relative to each presentation format (see Table 1),
while numeracy levels ranged from 1 to 7, and mean absolute
error was used as the measure of probability judgment error.
The regression analysis provides a comparison of search be-
havior at the level of design features, with the result that spatial
extent, posterior information, and including complements were
the strongest predictors for correct search decisions. For visual
formats, using spatial extent and posterior groupings (IconPost
and BarPost) were the most effective at soliciting correct search
decisions, while for numeric formats, presenting posterior and
complement information (ProbPost⫹ and FreqPost⫹) were also
relatively effective. Our regression model also indicates that the
use of natural frequencies was not a reliable predictor for search
behavior, in contrast to our initial hypothesis. Countability and
part-to-whole information also failed to consistently predict
correct search decisions.
Formats and probability judgments. In contrast to search
behavior (i.e., identification of the most useful test), probability
judgment accuracy did not vary substantially across formats
(Figure 3B; distribution of probability judgment error per for-
mat). Accordingly, a one-way between-participants analysis of
variance (ANOVA) found that the mean absolute error7
of
probability judgments was not significantly influenced by for-
mat, F(13, 804) ⫽ 1.15, p ⫽ .31. As a manipulation check, we
tested differences between formats on specific subsets of prob-
ability questions. Likelihood formats (numerical and visual)
had less error than their posterior counterparts on the base-rate
questions, t(816) ⫽ ⫺.73, p ⬍ .001 and on the likelihood
question, t(816) ⫽ ⫺.33, p ⬍ .001, whereas the posterior
variants had lower error on the posterior questions than the
likelihood variants, t(816) ⫽ ⫺.66, p ⬍ .001. Likelihood and
posterior formats did not systematically differ on the marginal
probability questions,8
t(816) ⫽ ⫺.8, p ⫽ .42. Consistent with
findings from Bayesian reasoning studies (Brase & Hill, 2015;
Gigerenzer & Hoffrage, 1995), natural frequency formats
(FreqLik, FreqLik⫹) had less error on posterior probability
judgments than their conditional probability counterparts (Prob-
Lik, ProbLik⫹), t(231) ⫽ ⫺3.45, p ⬍ .001. However, the
natural frequency formats were not better when aggregated over
all probability judgments questions.
Probability judgments and search decisions. In contrast to
our initial hypothesis there was virtually no correlation (rpb ⫽ .04)
between probability judgment error and the proportion of correct
search decisions (Figure 3C; participants split into quartiles based
on judgment error). This is supported by the regression analysis,
which controls for design features and individual numeracy level
(see Table 7). Similar results are obtained when judgments are
broken down by question type (i.e., base rate, marginals, likeli-
hoods, and posteriors). This suggests that contrary to our hypoth-
esis lower probability judgment error does not necessarily lead to
better search decisions.
Numeracy. There was no correlation between numeracy and
search decisions (rpb ⫽ .003; Figure 3D), with the regression
model also finding no significant relationship (see Table 7). How-
ever, higher numeracy was correlated with lower error on the
probability judgment task (Pearson r ⫽ ⫺.4; Figure 3E). Thus,
participants with higher numeracy performed better on the prob-
ability judgment task, but had no advantage on the information
search task.
Discussion
Experiment 1 tested a search environment where the correct
response (DE test) as predicted by probability gain was in
contradiction to the predictions made by information gain,
impact, and the likelihood difference heuristic (LM test).
Search decisions were strongly influenced by presentation for-
mat, with the posterior icon array and posterior bar graphs
providing the most effective means to help participants identify
the correct test. The key advantage of these graphical formats is
that they were able to elicit the same (correct) intuitions about
the relative usefulness of a query as experience-based learning
(Nelson et al., 2010), but without the same time requirements
(several hundred trials of training, over 1–2 hr). Spatial extent,
posteriors, and complements were the strongest predictors for
correct search decisions. Contrary to our initial hypotheses,
neither numeracy nor probability judgment accuracy were reli-
able predictors of search behavior. Most participants presented
with the posterior icon array or the posterior bar graph were
able to identify the correct query, even though they were not
7
Using mean squared error (MSE) to measure performance on the
probability judgment tasks yields equivalent results in all experiments. We
use MAE because it provides a more intuitive representation of the mag-
nitude of error.
8
The manipulation check results were replicated in all subsequent
experiments, with the exception of Experiment 3, where the posterior
formats had significantly lower error on the marginal questions than the
likelihood formats, t(680) ⫽ 2.40, p ⫽ .017. We do not separately report
the corresponding t tests for the subsequent experiments.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

any better at the probability estimation task than participants
assigned to other presentation formats.
Experiment 2
In Experiment 2 we introduced the possibility of certain
outcomes with respect to the hypotheses under consideration,
that is, one of the two search queries had a probabilistic out-
come that provided certainty about the true species of a turtle.
We used a search environment that was similar to Experiment
1, but adjusted so that the M result of the LM test gives 100%
certainty that the turtle belongs to the Bayosian species. How-
ever, the M result is only present in 40% of the turtles, and the
alternative L result gives maximal uncertainty about the species
(50% probability for each species). Even though the LM test has
a possibility of certainty, the DE test leads to higher expected
classification accuracy. In Experiment 2, the probability of
certainty heuristic along with information gain, impact, and the
likelihood difference heuristic predict selection of the LM test,
while only probability gain predicts selection of the DE test. We
expected Experiment 2 to be more difficult than Experiment 1,
because of the addition of the probability of certainty heuristic
as a model that disagrees with the correct probability gain
prediction. In a similar information search task, the probability
of certainty heuristic accounted for 42% of search decisions
when subjects were given the standard probability format, but
only 3% of search decisions for subjects assigned to experience-
based learning (Experiments 1 and 2, Condition 4; Nelson et al.,
2010). This finding suggests that when the probabilistic struc-
ture of the environment is not learned through experience, the
possibility of certainty is an important factor in how queries are
selected.
Results
Information search decisions. Overall, only 36% of partic-
ipants chose the correct query, compared with 61% in Experi-
ment 1, with all formats having a comparatively lower propor-
tion of correct responses (Figure 4A). As in Experiment 1, the
posterior icon array and posterior bar graph had the highest
proportion of correct responses (48% for both). However, there
were no reliable differences between formats in terms of search
behavior, ␹2
(13, N ⫽ 683) ⫽ 17.86, p ⫽ .16, which may be
because of floor effects. A logistic regression analysis found
that only spatial extent was a statistically reliable predictor for
correct search decisions (see Table 7). Remarkably, no other
design features, numeracy skill, or probability judgment error
were reliable predictors for search behavior, which is a strong
result given the large sample size of 683 subjects. It seems that
the introduction of a certain outcome made it substantially
harder for participants to identify the more useful query, and the
query with the possibility of certainty was valued beyond its
contribution to classification accuracy. We address this possi-
bility in Experiment 3, where we isolate the prediction of the
probability of certainty heuristic from all other models.
Probability judgments. Consistent with Experiment 1, prob-
ability judgment error was not significantly affected by presenta-
tion format, F(13, 670) ⫽ 1.26, p ⫽ .24 (Figure 4B), nor was
probability judgment error correlated with search decisions (rpb ⫽
.06; Figure 4C).
Numeracy. Again, there was no correlation between numeracy
and the proportion of probability gain search decisions (rpb ⫽ ⫺.06;
Figure 4D), although higher numeracy was correlated with lower
error on the probability judgment task (Pearson r ⫽ ⫺.45; Figure 4E).
Both of these results were consistent with the previous experiment.
Table 7
Logistic Regression Results
Dependent variable: Correct search decision
Model
(1) (2) (3) (4) (5) (6)
Variable Experiment 1 Experiment 2 Experiment 3 Experiment 4 All experiments
Design features
Natural frequencies ⫺.114 (.375) .028 (.354) ⫺.127 (.362) .333 (.454) .034 (.177) .042 (.187)
Posteriors .662ⴱⴱⴱ
(.150) .232 (.162) .223 (.163) ⫺.076 (.202) .287ⴱⴱⴱ
(.077) .306ⴱⴱⴱ
(.082)
Complement .483ⴱⴱ
(.179) ⫺.040 (.201) ⫺.114 (.198) .196 (.247) .120 (.094) .133 (.099)
Spatial extent .961ⴱⴱⴱ
(.271) .533ⴱ
(.250) .197 (.263) ⫺.265 (.296) .343ⴱⴱ
(.126) .378ⴱⴱ
(.133)
Countability .056 (.330) .127 (.291) .497 (.305) ⫺.744ⴱ
(.376) .007 (.150) .009 (.158)
Ind. differences
Numeracy .043 (.048) ⫺.035 (.053) .140ⴱⴱ
(.053) .361ⴱⴱⴱ
(.066) .072ⴱⴱ
(.025) .101ⴱⴱⴱ
(.026)
Probability judgments (MAE) .953 (.837) .994 (.880) ⫺.409 (.802) ⫺1.910 (1.183) ⫺.049 (.420) .209 (.441)
Environment
Certainty ⫺.973ⴱⴱⴱ
(.083)
OED model disagreement ⫺1.081ⴱⴱⴱ
(.083)
Constant ⫺.766ⴱ
(.349) ⫺.930ⴱ
(.397) ⫺.253 (.384) .437 (.446) ⫺.250 (.180) .621ⴱⴱ
(.198)
Observations 817 683 681 677 2,858 2,858
Classification accuracy .62 .64 .63 .80 .60 .67
Akaike Information Criterion 1,042.758 891.693 890.972 636.189 3,815.619 3,520.556
Note. Log odds are shown with SE in brackets. The part-to-whole design feature is not presented as an independent predictor, because it is entirely
accounted for by taking natural frequencies without spatial extent. MAE ⫽ mean absolute error; OED ⫽ Optimal Experimental Design. Classification
accuracy is 10-fold cross validation prediction accuracy.
ⴱ
p ⬍ .05. ⴱⴱ
p ⬍ .01. ⴱⴱⴱ
p ⬍ .001.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1287

Discussion
Experiment 2 produced results that resemble Experiment 1, but
with the proportion of probability gain choices reduced substan-
tially across all formats. No format led to consistent selection of
the accuracy-maximizing query, although formats using spatial
extent led to relatively better search performance. The lowered
performance across all formats in Experiment 2 indicates that
introducing certain outcomes may have contributed to a search
problem where the accuracy-maximizing query is substantially
less likely to be selected, irrespective of format. Why does the
possibility of certainty create such challenging environment when
the query that can lead to a certain outcome (LM test) is in
opposition to the accuracy-maximizing query (DE test)? To exam-
ine how certainty influences query selection by itself, we con-
ducted Experiment 3 to examine search behavior in an environ-
ment where the probability of certainty prediction contradicted all
other models.
Experiment 3
We conducted Experiment 3 to try to isolate the extent to which
the possibility of a certain outcome influences search behavior,
across the various presentation formats. As in Experiment 2, the
LM test had a chance of yielding an M result, with the implication
that the turtle is certainly Bayosian. But again, it was a risky
choice, because the alternative L outcome resulted in maximal
uncertainty. The main difference in the environmental probabili-
ties, compared with the previous study, was that in Experiment 3,
probability gain, information gain, impact, and the likelihood
difference heuristic predicted selection of the DE test, while only
the probability of certainty heuristic predicted selection of the LM
test.
Results
Information search decisions. Overall, 64% of participants
chose the correct query, while the remaining 36% of search
A
B
C D E
Figure 4. Experiment 2 results. (A) Across all format there was a lower proportion of correct (accuracy-
maximizing) search queries compared to Experiment 1. (B) Performance on the probability judgment task
replicated Experiment 1 results, with neither judgment error; (C) nor numeracy; and (D) being predictors for
search decisions. (E) As in Experiment 1, there was a negative correlation between judgment error and
numeracy. See the online article for the color version of this figure.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

decisions were consistent with only the probability of certainty
heuristic. The proportion of correct responses varied between
47 and 73% (Figure 5A), but there was no overall difference
between formats, ␹2
(13, N ⫽ 681) ⫽ 17.91, p ⫽ .16. The
logistic regression analysis found no statistically reliable pre-
dictors for search behavior from the set of design features (see
Table 7). Thus, none of the design features nor individual
presentation formats we tested yielded a systematic advantage
in this particular search environment.
Probability judgments. There were differences between for-
mats on the probability judgment task, F(13, 668) ⫽ 2.32, p ⫽
.005; however, the descriptive statistics indicate that the overall
effect is mainly because of the likelihood variants of the natural
frequency formats (FreqLik, FreqLik⫹) having lower error than
the other formats (Figure 5B). If these two conditions are ex-
cluded, an ANOVA shows no significant differences, F(11,
569) ⫽ 1.2, p ⫽ .28. No correlation was found between judgment
error and search decisions (rpb ⫽ ⫺.07; Figure 5C), which is
consistent with the regression analysis (see Table 7). Together with
the previous findings, this suggests that probability judgment error
had no bearing on search decisions.
Numeracy. In contrast to Experiments 1 and 2, we found a
small correlation between numeracy and search decisions
(rpb ⫽ .12; Figure 5D). This suggests that numeracy may be a
predictor for search behavior when the more useful query is
relatively obvious, which could be the case in Experiment 3
where probability gain, information gain, impact, and the like-
lihood difference heuristic all make the same prediction. In-
deed, numeracy skill was the only reliable predictor for search
decisions in the regression model (see Table 7). On the other
hand, as in all previous experiments, numeracy was strongly
correlated with performance on the probability estimation task
A
B
C D E
Figure 5. Experiment 3 results. (A) Sixty-four percent of subjects (aggregated across formats) chose the
correct accuracy-maximizing query, with the remaining choices consistent with only the probability of
certainty heuristic, of the models in consideration. There were no differences in search behavior across
formats, although we found that the natural frequency formats (FreqLik and FreqLik⫹) had lower
probability judgment error than the other formats (B). Judgment error was not a predictor for search
decisions (C), although higher numeracy was weakly related to better search choices (rpb ⫽ .12; D). The
negative correlation between judgment error and numeracy was replicated (E). See the online article for the
color version of this figure.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1289

(Pearson r ⫽ ⫺.47; Figure 5E). The correlation between nu-
meracy and probability judgment accuracy is a robust finding,
which appears to be independent of the task environment.
Discussion
Sixty-four percent of subjects correctly selected the DE test,
which was predicted by all of the OED and heuristic models with
the sole exception of the probability of certainty heuristic. This
suggests that by itself, the possibility of obtaining a certain out-
come influenced search decisions a great deal. To understand more
precisely how much the possibility of certainty influences behav-
ior, it is important to have an idea of the highest performance that
can be obtained. We address this in our final experiment.
Experiment 4
We conducted one further experiment to assess the upper range of
performance with respect to the experimental methods and subject
population, in which there was no disagreement among model pre-
dictions. Experiment 4 used an environment with similar model
predictions as Experiment 3, but without any certain outcomes. Prob-
ability gain, information gain, impact, and the likelihood difference
heuristic all made the same prediction (DE test), while the probability
of certainty heuristic made no prediction. This experiment was unique
in that no model disagreed with the probability gain prediction.
Results
Information search decisions. Participants in all formats reli-
ably chose the correct query (80% overall, range: 73–88%), which
was the highest proportion out of all experiments (Figure 6A). In
comparison with Experiment 3 (64% correct choices), which had a
similar relationship between models, but without the possibility of a
certain outcome, we can infer that a strategy corresponding to the
probability of certainty heuristic accounted for a difference of about
16 percentage points (95% confidence interval [12, 21], Wilson score
interval). There were no differences in search choice between formats,
A
B
C D E
Figure 6. Experiment 4 results. (A) With no divergence between model predictions, the majority of subjects,
regardless of format, chose the correct accuracy-maximizing query (80% aggregated across formats). (B) There
were no substantial differences in judgment error across formats. (C) In this experiment, there were correlation
between (C) search decisions and judgment error (rpb ⫽ ⫺.16) and (D) between search decisions and numeracy
(rpb ⫽ .25). (E) A negative correlation between judgment error and numeracy was found, as in all previous
experiments, suggesting that this is a robust result. See the online article for the color version of this figure.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

␹2
(13, N ⫽ 677) ⫽ 12.92, p ⫽ .45. The logistic regression analysis
found that from the set of design features, only countability was a
reliable predictor; however, countability had a negative effect on
performance (see Table 7).
Probability judgments. Format did not have an effect on prob-
ability judgments, F(13, 664) ⫽ 1.18, p ⫽ .29 (Figure 6B). Taken
together with the previous experiments, our results suggest that no
single format has an overall advantage on the probability judgment
task, although there were differences for specific subsets of questions.
For the first time, we found a weak negative correlation (rpb ⫽ ⫺.16)
between probability judgment error and correct search behavior (Fig-
ure 6C); however, when controlling for differences in design features
and numeracy skill, our regression model indicates that judgment
error is not a reliable predictor for search decisions (see Table 7). In
the context of all previous experiments, people who are able to make
more accurate probability judgments do not seem to perform any
better in the search task.
Numeracy. Higher numeracy scores were correlated with a
higher proportion of accuracy-maximizing search decisions (rpb ⫽
.25; Figure 6D). This was a stronger correlation than in Experi-
ment 3, whereas no correlations were found for Experiments 1 and
2. These results are consistent with the regression models, which
found significant effects of numeracy on search behavior in Ex-
periments 3 and 4 (see Table 7). One explanation for this result is
that with an easier task (i.e., less disagreement among the models
considered), numeracy may be a predictor for search decisions.
Numeracy was also correlated with lower error on the probability
judgment task (Pearson r ⫽ ⫺.42; Figure 6E), consistent with all
previous experiments.
Discussion
In Experiment 4, where all models predicted selection of the DE
test (with the exception of the probability of certainty heuristic that
made no prediction), we measured the highest proportion of correct
query selections out of all the experiments (80% overall). Countability
was the only statistically reliable predictor for search behavior among
the design features, but it had a negative contribution toward correct
choices. Numeracy was a reliable predictor for search choice, as in
Experiment 3. Therefore, one might conclude that high numeracy is
helpful in simple environments in which there is no disagreement
among the heuristic and OED models’ predictions. However, not even
the highest level of numeracy was adequate for more difficult prob-
abilistic environments.
General Discussion
Our aims in this study were both practical and theoretical: to
explore how search decisions are influenced by presentation formats
and design features, while addressing the theoretical question of
whether search behavior is mediated by probabilistic reasoning, nu-
meracy skill, or both. We studied how search behavior and probability
estimation varied across 14 different numerical and visual presenta-
tion formats, in four different search environments. Table 8 presents
an overview of statistical tests conducted in each experiment; Figure
7 shows the results aggregated over all experiments.
We extended the logistic regression analysis by aggregating
over all four experiments (see Table 7), with Model 5 using the
same predictors as in the individual experiments and with
Table
8
Overview
of
Experiments
Experiment
Correct
search
queries
Influence
of
Format
on
Probability
judgment
error
and
search
decisions
Relationship
between
numeracy
and
Search
decision
Probability
judgment
error
Search
decision
Probability
judgment
error
1
(n
⫽
817)
61%,
range:
[27–86%]
␹
2
(13,
N
⫽
817)
⫽
78.57,
p
⬍
.001
F(13,804)
⫽
1.15,
p
⫽
.31
r
pb
⫽
⫺.04
r
pb
⫽
.003
r
⫽
⫺.40
2
(n
⫽
683)
36%,
range:
[21–48%]
␹
2
(13,
N
⫽
683)
⫽
17.86,
p
⫽
.16
F(13,670)
⫽
1.26,
p
⫽
.24
r
pb
⫽
⫺.06
r
pb
⫽
⫺.06
r
⫽
⫺.45
3
(n
⫽
681)
64%,
range:
[47–73%]
␹
2
(13,
N
⫽
681)
⫽
17.91,
p
⫽
.16
F(13,668)
⫽
2.32,
p
⫽
.005
r
pb
⫽
⫺.07
r
pb
⫽
.12
r
⫽
⫺.47
4
(n
⫽
677)
80%,
range:
[71–88%]
␹
2
(13,
N
⫽
677)
⫽
12.92,
p
⫽
.45
F(13,664)
⫽
1.18,
p
⫽
.29
r
pb
⫽
⫺.16
r
pb
⫽
.25
r
⫽
⫺.42
Overall
(n
⫽
2,858)
60%,
range:
[45–71%]
␹
2
(13,
N
⫽
2,845)
⫽
52.66,
p
⬍
.001
F(13,2845)
⫽
2.87,
p
⫽
.001
r
pb
⫽
⫺.02
r
pb
⫽
.05
r
⫽
⫺.43
Note.
Performance
on
the
probability
judgment
task
is
measured
in
terms
of
mean
absolute
error
across
all
11
questions.
Thus,
a
negative
correlation
between
numeracy
and
probability
judgments
signifies
that
higher
numeracy
skill
is
a
predictor
for
lower
error.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1291

Model 6 adding two environmental factors (that vary across the
individual experiments) as binary predictors. Certainty denotes
environments where there is the possibility of a query outcome
leading to certainty about the true species of a turtle, which in
our case was always in opposition to the correct probability
gain query (Experiments 2 and 3). OED Disagreement denotes
environments where there is disagreement between OED mod-
els, specifically when the probability gain prediction is in
opposition to the prediction made by information gain, impact,
and the likelihood difference heuristic (Experiments 1 and 2).
These two environmental factors are able to describe all four
experiments as a 2 ⫻ 2 factorial design. Both aggregate models
yield similar estimates of the log odds for shared predictors,
although Model 6 achieves higher classification accuracy and a
lower AIC by including environmental factors.
The Role of the Environment
Environments where the probability gain prediction disagreed with
the probability of certainty heuristic or with other OED models were
more difficult, with similar reductions to the log odds of making a
correct search decision (see Table 7). Why did these factors contribute
toward a substantially more difficult search task? Hogarth and col-
leagues (Hogarth, Lejarraga, & Soyer, 2015) have introduced the
notion of kind and wicked environments, where wicked environments
represent a disjoint between learning and test environments. Are the
environments we tested wicked in this sense? We unfortunately do not
know the true distribution of search environments in the world. As a
first step to address this question, we conducted simulations over 10
million randomly generated environments,9
and considered the rela-
tionship of the predictions made by each of the OED and heuristic
models in a pairwise manner. In cases where both competing models
9
We randomly generated 10 million sets of the five probabilistic vari-
ables (prior probability and test likelihoods) that fully describe a search
environment, like the ones presented in this study. For each environment,
we conducted pairwise comparisons between model predictions. Simulation
code is available at https://github.com/charleywu/AskingBetterQuestions. For
simulations of a larger variety of OED models, see supplemental material
from Nelson (2005).
A
B
C D E
Figure 7. Summary of all experiments. (A) Aggregated over all experiments, the choice of format led to
noticeable differences in the proportion of correct search decisions, but not in judgment error (B). Probability
judgment error (C) was not correlated with search behavior, while each incremental change in numeracy had a
small influence on search decisions (D). High numeracy skill was a robust predictor for lower judgment error
(E). See the online article for the color version of this figure.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

made a prediction, information gain and probability gain made the
same prediction 92% of the time; the probability of certainty heuristic
and probability gain made the same prediction 77% of the time; and
information gain and the probability of certainty heuristic made the
same prediction 83% of the time.
This illustrates a potential tension between using OED princi-
ples to test environments where we can disentangle overlapping
model predictions, and environments that are representative of
search problems experienced in the wild. The use of heuristics,
such as the probability of certainty, may represent an ecologically
valid strategy because of a correspondence with the correct prob-
ability gain query in most environments, trading off computational
complexity for accuracy. However, this still poses a challenge to
the design of presentation formats for conveying statistical infor-
mation about search problems, because we are seeking formats that
are robust across different types of environment. Indeed, in Nelson
et al. (2010) the possibility of certainty did not influence partici-
pants assigned to the experience-based learning conditions, who
preferentially selected the higher probability gain query in every
statistical environment that was tested. We have not identified
numeric or graphical formats that consistently lead to as high of
rates of selection of the probability gain test as experience-based
learning. However, across all our experiments, we found that some
formats and design features were better than others in soliciting
correct search decisions.
Presentation Formats and Search Decisions
Aggregated over all experiments (Figure 7A), the posterior bar
graph and posterior icon array had the highest performance (69 and
71% correct choices, respectively), and the standard probability
format (ProbLik) had the lowest performance (45%). The logistic
regression analysis shows that from the set of design features,
spatial extent and posterior information were the only statistically
reliable predictors for search behavior (Model 6, Table 7). There-
fore, the present findings suggest that bar graphs or icon arrays are
the most helpful for information search problems, where the un-
derlying natural frequency information is conveyed using spatial
extent, and the information is grouped by primarily by query
outcome, as in our posterior formats.
In contrast to results from the Bayesian reasoning literature and
our own initial hypotheses, the other design features did not
consistently influence search behavior across environments. Given
the large (N ⫽ 2,858) sample size, we take the lack of statistically
significant results for natural frequencies, complement informa-
tion, countability, and part-to-whole information as strong evi-
dence against our hypotheses that these design features (by them-
selves) facilitate adaptive information search.
Among the classes of presentation formats (conditional proba-
bilities, natural frequencies, and graphical visualizations), the larg-
est variability in performance was found within the four condi-
tional probability formats. Presenting information in terms of
posteriors and including complement information led to enhanced
performance for the conditional probability formats, although
these same design features did not substantially influence the
natural frequency formats (Figure 7A). Aggregating across the
experiments, the posterior probability with complement format
(ProbPost⫹) led to a similar proportion of correct search choices
(68%) as the posterior icon array (69%) and posterior bar graph
formats (71%). An interesting finding was that the posterior prob-
ability format with complement (ProbPost⫹) led to a higher pro-
portion of correct choices (68%) than any of the numeric natural
frequency formats (that ranged from 57 to 61%). This, combined
with natural frequencies failing to be a predictor in the regression
model (see Table 7), suggests that numeric natural frequencies are
not necessarily more advantageous for information search prob-
lems than conditional probabilities—if they are presented in the
right way (i.e., not the standard probability format).
Posterior bar graphs, despite not being countable, led to as good
of performance as the posterior icon arrays, and to better perfor-
mance than any of the numeric natural frequency formats. This
suggests that countability is not necessary for communicating
probabilistic information for information search. Furthermore, out
of all the formats we studied, icon arrays and bar graphs were the
only formats that represented natural frequencies using spatial
extent, a perceptual attribute. Spatial extent was found to be one of
the best predictors for correct test selection in the regression
analysis, suggesting that it may be an important design feature for
communicating probabilistic information in search tasks.
Not all visual formats were successful, with the dot diagrams
leading to fewer accuracy-maximizing search decisions than icon
arrays or bar graphs. Thus, the increased accessibility of part-to-
whole information did not compensate for the lack of spatial
extent, which were the two design features differentiating the dot
diagrams from icon arrays and bar graphs. Future studies should
test whether providing part-to-whole information could be helpful
if combined with spatial extent (e.g., using stacked icon arrays).
Information Search and Probabilistic Reasoning
In contrast to our initial hypothesis, the ability to make accurate
probability judgments was not related to search behavior (Figure
7C), nor was it a predictor in the logistic regression model, when
controlling for differences in design features and numeracy skill.
This suggests that presentation formats that improve explicit
Bayesian reasoning accuracy do not necessarily transfer to more
complex decision-making tasks, such as information search, even
though information search can be modeled “as if” it is based on
Bayesian reasoning processes. Of course, improving explicit prob-
ability judgments may be useful in some situations (e.g., where
doctors need to explicitly explain the meaning of a test result in
terms of posterior probabilities to a patient). However, people can
sometimes make decisions that correspond to the correct applica-
tion of Bayes’s rule without calculating the correct probability
estimates (Domurat, Kowalczuk, Idzikowska, Borzymowska, &
Nowak-Przygodzka, 2015). In our studies, the ability to reason
about information search decisions seems to be largely indepen-
dent from the ability to explicitly estimate probabilities.
Numeracy was robustly correlated with probability judgment
accuracy in all experiments (Figure 7E), and may also be helpful
in information search tasks (Figure 7D). We found weak correla-
tions between numeracy and search decisions in Experiments 3
and 4; however, these relationships were not found in Experiments
1 and 2, which, judging by the disagreement among model pre-
dictions and the overall rates of probability gain queries, were
substantially more difficult. Aggregated across all experiments,
numeracy was a statistically reliable predictor for correct search
decisions (see Table 7). Using Model 6 from the regression anal-
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1293

ysis, which includes environmental factors as predictors, we see
that the difference between the lowest and the highest possible
numeracy scores (a difference of six levels) can be represented by
a sixfold increase in the log odds (0.101 ⴱ
6 ⫽ 0.606), which is
similar to the combined influence of spatial extent and posteriors
(0.378 ⫹ 0.306 ⫽ 0.684). Thus, an individual with the lowest
numeracy skill, given a format with both spatial extent and pos-
terior information, is about as likely to identify the correct search
query as an individual with the highest numeracy, but given a
format lacking these two design features. The right format may
help compensate for low numeracy in information search, though
this effect may be limited by individual differences in statistical
reasoning. Just as the performance-boosting benefits of numeric
presentation formats (i.e., natural frequencies) require a sufficient
level of numeracy (Chapman & Liu, 2009), so do visual formats
require a sufficient level of graph literacy to yield the correct
interpretation (Galesic et al., 2009; Okan & Garcia-Retamero,
2012).
Take-The-Difference Heuristic
Why were some presentation formats more helpful than others,
particularly the posterior icon arrays and posterior bar graphs? We
discovered a simple heuristic strategy that identifies the higher prob-
ability gain query in our tasks, without requiring explicit computation
of posterior probabilities.10
The take-the-difference (TTD) heuristic is
a relevant strategy for all natural frequency formats, although we
believe it is especially well suited for visualizations of natural fre-
quencies using spatial extent. This new heuristic can be executed by
choosing the query with the largest absolute difference of natural
frequencies, when comparing each query outcome (see Figure 8). We
can describe the TTD heuristic using the OED framework, where the
expected usefulness of a query is measured as the absolute difference
between the joint frequencies N共c1⵩qj兲 and N共c2 ⵩ qj兲 for the out-
come qj where the absolute difference is largest:
euTTD(Q) ⫽ max
j
ⱍN(c1 ⵩ qj) ⫺ N(c2 ⵩ qj)ⱍ (9)
This heuristic strategy is most salient for the posterior variations
of the natural frequency and visual formats, because N共c1 ⵩ qj兲
and N共c2 ⵩ qj兲 are grouped together and easy to compare; this may
explain why the posterior variations generally yielded more correct
responses. Formats using spatial extent to visualize natural fre-
quencies have the added advantage that a visual comparison of the
lengths of two bars or arrays of icons is equivalent to calculating
the absolute difference between two natural frequencies specified
by the TTD heuristic. TTD also offers a potential explanation for
10
Through simulations over 10 million randomly generated environ-
ments, we found that TTD and probability gain never made contradic-
tory predictions. Thus, we conjecture that in every environment where
both models make a prediction about query selection, TTD always
agrees with probability gain. Simulation code is available at https://
github.com/charleywu/AskingBetterQuestions
Figure 8. An example of the take-the-difference (TTD) heuristic, applied to the posterior bar graph format from
Experiment 1. TTD is a simple heuristic strategy that is capable of identifying the accuracy-maximizing (i.e.,
probability gain) query, without explicit computation of posterior probabilities. (A) The expected utility of each
test, eu共·兲, can be described as the absolute difference between natural frequencies for the outcome that has the
largest difference. In the case of the DE test, the E gene has the largest difference, which can be determined
visually by comparing lengths of the “D and Bayosian” bar to the “D and Freqosian” bar, and the “E and
Bayosian” bar to the “E and Freqosian” bar. The expected utility of the DE test can be described using a line
spanning the distance between the lengths of “E and Bayosian” bar and the “E and Freqosian” bar, since the
length of each bar is a visual representation of a natural frequency. In panel (B), the relative usefulness of both
tests are compared, and the accuracy-maximizing query can be identified by picking the test with the longest line.
See the online article for the color version of this figure.
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

the poor performance of the dot diagrams in the search task. The
presentation of the dots—placed in a random uniform distribution
within a container of fixed area—makes the task of comparing
absolute differences much less intuitive than comparing relative
differences (through a density estimate). Comparing relative dif-
ferences is undesirable because it renormalizes the natural frequen-
cies as absolute frequencies, and loses the additional information
about base rates and the marginals that are contained within the
natural frequency representation. Comparing relative differences
does not implement the probability gain model and does not
consistently yield the same search predictions.
Conclusion
We systematically investigated a wide range of presentation
formats across different search environments. Differences in
search environments played a large role in influencing search
behavior across experiments, where the possibility of certainty
and disagreements between OED models contributed toward
substantially more difficult tasks. Aggregated over all search
environments, formats using spatial extent (icon arrays and bar
graphs) and presenting information in terms of posteriors were
the most effective design features for improving search behav-
ior, while numeracy skill also made a positive contribution.
Because some environments still proved difficult for all for-
mats, future studies may want to examine didactic methods such
as story boarding (Ottley et al., 2016) or adding interactivity to
visualizations (Tsai, Miller, & Kirlik, 2011) as a manipulation
orthogonal to the choice of presentation format, to further
improve search performance.
Although many of the presentation formats we tested are
useful approximations of how humans might learn about a
search problem through experience, there is a notable incon-
gruence between description and experience (Hertwig & Erev,
2009). In Experiments 2 and 3, each format gave the descriptive
information that for the entire population of turtles, those with
the M gene are invariably Bayosian. This may be qualitatively
different from learning about a probabilistic environment
through direct experience, where single events are sampled
from the population. This may explain why certain outcomes
did not prove challenging for experience-based learners in a
similar information search task (Nelson et al., 2010). This
incongruence of representing certain outcomes through descrip-
tive versus experiential information has not played an important
role in Bayesian reasoning tasks, since certainty makes deriving
posterior probabilities easier, rather than harder. However, this
is not the case in information search, specifically when the
query with a certain outcome does not lead to the highest
expected classification accuracy. We recommend future re-
search to examine behavioral differences in judgment and de-
cision making tasks when conveying statistical information
about certain or near certain outcomes, across descriptive pre-
sentation formats and experience-based learning, and manipu-
lating whether the information is presented as a sample or as
representative of the whole population, to examine these issues.
We identified a new information search heuristic, TTD, which
chooses the accuracy-maximizing query (i.e., higher probability
gain query) without explicitly computing posterior probabilities.
The TTD heuristic offers a potential explanation for differences in
search behavior across formats, as well as the lack of correlation
between probability judgments and search decisions. Thus, we
recommend that future studies investigate the effectiveness of the
TTD heuristic by providing explicit instruction in its use, together
with appropriate visualizations of natural frequencies using spatial
extent.
The lack of correlation between probability judgment accu-
racy and search behavior has the important additional implica-
tion that improving explicit probability judgment accuracy is
not necessarily the same as improving performance on more
complex probabilistic decision-making tasks, such as informa-
tion search. Thus, we recommend future research on the use of
presentation formats for communicating risk to not only address
the ability to make explicit probability judgments, but also how
formats influence more complex decisions that people make
about the world.
References
Ancker, J. S., Senathirajah, Y., Kukafka, R., & Starren, J. B. (2006).
Design features of graphs in health risk communication: A systematic
review. Journal of the American Medical Informatics Association, 13,
608–618. http://dx.doi.org/10.1197/jamia.M2115
Barbey, A. K., & Sloman, S. a. (2007). Base-rate respect: From ecological
rationality to dual processes. The Behavioral and Brain Sciences, 30,
241–254. http://doi.org/10.1017/S0140525X07001653
Baron, J. (1985). Rationality and Intelligence. New York, NY: Cambridge
University Press. http://dx.doi.org/10.1017/CBO9780511571275
Baron, J., Beattie, J., & Hershey, J. C. (1988). Heuristics and biases in
diagnostic reasoning. Organizational Behavior and Human Decision Pro-
cesses, 42, 88–110. http://dx.doi.org/10.1016/0749-5978(88)90021-0
Binder, K., Krauss, S., & Bruckmaier, G. (2015). Effects of visualizing
statistical information: An empirical study on tree diagrams and 2 ⫻ 2
tables. Frontiers in Psychology, 6, 1186.
Bodemer, N., Meder, B., & Gigerenzer, G. (2014). Communicating relative
risk changes with baseline risk: Presentation format and numeracy
matter. Medical Decision Making, 34, 615–626. http://doi.org/10.1177/
0272989X14526305
Bramley, N. R., Lagnado, D. A., & Speekenbrink, M. (2015). Conservative
forgetful scholars: How people learn causal structure through sequences
of interventions. Journal of Experimental Psychology: Learning, Mem-
ory, and Cognition, 41, 708–731. http://dx.doi.org/10.1037/xlm0000061
Brase, G. L. (2009). Pictorial representations in statistical reasoning. Ap-
plied Cognitive Psychology, 23, 369–381. http://dx.doi.org/10.1002/acp
.1460
Brase, G. L., & Hill, W. T. (2015). Good fences make for good neighbors
but bad science: A review of what improves Bayesian reasoning and
why. Frontiers in Psychology, 6, 340.
Brown, S. M., Culver, J. O., Osann, K. E., MacDonald, D. J., Sand, S.,
Thornton, A. A., . . . Weitzel, J. N. (2011). Health literacy, numeracy,
and interpretation of graphical breast cancer risk estimates. Patient
Education and Counseling, 83, 92–98. http://dx.doi.org/10.1016/j.pec
.2010.04.027
Burns, K. (2004, May 25–28). Painting pictures to augment advice. In
Proceedings of AVI ’04, Gallipoli, Italy, 344–349. http://dx.doi.org/10
.1145/989863.989921
Chapman, G. B., & Liu, J. (2009). Numeracy, frequency, and Bayesian
reasoning. Judgment and Decision Making, 4, 34–40.
Cleveland, W. S., & McGill, R. (1985). Graphical perception and graphical
methods for analyzing scientific data. American Association for the
Advancement of Science, 229, 828–833. http://dx.doi.org/10.1126/
science.229.4716.828
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1295

Cokely, E. T., Galesic, M., Schulz, E., Ghazal, S., & Garcia-Retamero, R.
(2012). Measuring risk literacy: The Berlin numeracy test. Judgment and
Decision Making, 7, 25–47.
Cole, W. (1989). Understanding Bayesian reasoning via graphical displays.
Proceedings of the SIGCHI Conference on Human Factors in Comput-
ing Systems, 20, 381–386. http://dx.doi.org/10.1145/67449.67522
Cole, W., & Davidson, J. (1989). Graphic representation can lead to fast
and accurate Bayesian reasoning. Proceedings of the Annual Symposium
on Computer Application in Medical Care, 8, 227–231.
Cosmides, L., & Tooby, J. (1996). Are humans good intuitive statisticians
after all? Rethinking some conclusions from the literature on judgment
under uncertainty. Cognition, 58, 1–73. http://dx.doi.org/10.1016/0010-
0277(95)00664-8
Crupi, V., Nelson, J. D., Meder, B., Cevolani, G., & Tentori, K. (2016).
Generalized information theory meets human cognition: Introducing a
unified framework to model uncertainty and information search. [Man-
uscript in preparation].
Domurat, A., Kowalczuk, O., Idzikowska, K., Borzymowska, Z., &
Nowak-Przygodzka, M. (2015). Bayesian probability estimates are not
necessary to make choices satisfying Bayes’ rule in elementary situa-
tions. Frontiers in Psychology, 6, 1194.
Gaissmaier, W., Wegwarth, O., Skopec, D., Müller, A.-S., Broschinski, S.,
& Politi, M. C. (2012). Numbers can be worth a thousand pictures:
Individual differences in understanding graphical and numerical repre-
sentations of health-related information. Health Psychology, 31, 286–
296. http://dx.doi.org/10.1037/a0024850
Galesic, M., Garcia-Retamero, R., & Gigerenzer, G. (2009). Using icon
arrays to communicate medical risks: Overcoming low numeracy.
Health Psychology: Official Journal of the Division of Health Psychol-
ogy, American Psychological Association, 28, 210–216. http://dx.doi
.org/10.1037/a0014474
Garcia-Retamero, R., & Hoffrage, U. (2013). Visual representation of
statistical information improves diagnostic inferences in doctors and
their patients. Social Science & Medicine, 83, 27–33. http://dx.doi.org/
10.1016/j.socscimed.2013.01.034
Gigerenzer, G., & Hoffrage, U. (1995). How to improve Bayesian reason-
ing without instruction: Frequency formats. Psychological Review, 102,
684–704. http://dx.doi.org/10.1037/0033-295X.102.4.684
Gigerenzer, G., & Hoffrage, U. (2007). The role of representation in
Bayesian reasoning. Behavioral and Brain Sciences, 30, 264–267. http://
dx.doi.org/10.1017/S0140525X07001756
Gureckis, T. M., & Markant, D. B. (2012). Self-directed learning: A
cognitive and computational perspective. Perspectives on Psychological
Science, 7, 464–481. http://dx.doi.org/10.1177/1745691612454304
Heer, J., & Bostock, M. (2010). Crowdsourcing graphical perception:
Using Mechanical Turk to assess visualization design. In Proceedings of
the 28th annual chi conference on human factors in computing systems
(pp. 203–212). New York, NY: ACM. http://dx.doi.org/10.1145/
1753326.1753357
Hertwig, R., & Erev, I. (2009). The description-experience gap in risky
choice. Trends in Cognitive Sciences, 13, 517–523. http://dx.doi.org/10
.1016/j.tics.2009.09.004
Hill, W. T., & Brase, G. L. (2012). When and for whom do frequencies
facilitate performance? On the role of numerical literacy. The Quarterly
Journal of Experimental Psychology: Human Experimental Psychology,
65, 2343–2368. http://dx.doi.org/10.1080/17470218.2012.687004
Hoffrage, U., Gigerenzer, G., Krauss, S., & Martignon, L. (2002). Repre-
sentation facilitates reasoning: What natural frequencies are and what
they are not. Cognition, 84, 343–352. http://dx.doi.org/10.1016/S0010-
0277(02)00050-1
Hogarth, R. M., Lejarraga, T., & Soyer, E. (2015). The two settings of kind
and wicked learning environments. Current Directions in Psychological
Science, 24, 379–385. http://dx.doi.org/10.1177/0963721415591878
Jarecki, J. B., Meder, B., & Nelson, J. D. (2016). Naïve and robust:
Class-conditional independence in human classification learning. [Man-
uscript submitted for publication].
Johnson-Laird, P. N., Legrenzi, P., Girotto, V., Legrenzi, M. S., & Caverni,
J. P. (1999). Naive probability: A mental model theory of extensional
reasoning. Psychological Review, 106, 62–88. http://dx.doi.org/10.1037/
0033-295X.106.1.62
Klayman, J., & Ha, Y. (1987). Confirmation, disconfirmation, and infor-
mation in hypothesis testing. Psychological Review, 94, 211–228. http://
dx.doi.org/10.1037/0033-295X.94.2.211
Kleiter, G. (1994). Natural sampling: Rationality without base rates. In
G. H. Fischer & D. Laming (Eds.), Contributions to mathematical
psychology, psychometrics, and methodology (pp. 375–388). New York,
NY: Springer.
Koehler, J. J. (1996). The base rate fallacy reconsidered: Descriptive,
normative, and methodological challenges. Behavioral and Brain Sci-
ences, 19, 1. http://dx.doi.org/10.1017/S0140525X00041157
Kullback, S., & Leibler, R. A. (1951). On Information and Sufficiency. The
Annals of Mathematical Statistics. Advance online publication. http://dx
.doi.org/10.1214/aoms/1177729694
Lagae, A., & Dutré, P. (2008). A comparison of methods for generating
poisson disk distributions. Computer Graphics Forum, 27, 114–129.
http://dx.doi.org/10.1111/j.1467-8659.2007.01100.x
Legge, G. E., Klitz, T. S., & Tjan, B. S. (1997). Mr. Chips: An ideal-
observer model of reading. Psychological Review, 104, 524–553. http://
dx.doi.org/10.1037/0033-295X.104.3.524
Lindeman, S. T., van den Brink, W. P., & Hoogstraten, J. (1988). Effect of
feedback on base-rate utilization. Perceptual and Motor Skills, 67,
343–350. http://dx.doi.org/10.2466/pms.1988.67.2.343
Lindley, D. V. (1956). On a measure of the information provided by an
experiment. Annals of Mathematical Statistics, 27, 986–1005. http://dx
.doi.org/10.1214/aoms/1177728069
Markant, D., & Gureckis, T. M. (2012). Does the utility of information
influence sampling behavior? In N. Miyake, D. Peebles, and R. P.
Cooper (Eds.), Proceedings of the 34th annual conference of the cog-
nitive science society (pp. 719–724). Austin, TX: Cognitive Science
Society.
Marr, D. (1982). Vision: A computational investigation into the human
representation and processing of visual information. San Francisco:
W. H. Freeman.
Martignon, L., & Krauss, S. (2003). Can L’Homme Eclaire be fast and
frugal? Reconciling Bayesianism and bounded rationality. In S. Sch-
neider & J. Shanteau (Eds.), Emerging perspectives on judgment and
decision research (pp. 108–122). Cambridge, England: Cambridge Uni-
versity Press. http://dx.doi.org/10.1017/CBO9780511609978.006
Martire, K. A., Kemp, R. I., Sayle, M., & Newell, B. R. (2014). On the
interpretation of likelihood ratios in forensic science evidence: Presen-
tation formats and the weak evidence effect. Forensic Science Interna-
tional, 240, 61–68. http://dx.doi.org/10.1016/j.forsciint.2014.04.005
McDowell, M., & Jacobs, P. (2016). Meta-analysis of the effect of natural
frequencies on Bayesian reasoning. [Manuscript submitted for publica-
tion].
McNair, S., & Feeney, A. (2015). Whose statistical reasoning is facilitated
by a causal structure intervention? Psychonomic Bulletin & Review, 22,
258–264. http://dx.doi.org/10.3758/s13423-014-0645-y
Meder, B., & Gigerenzer, G. (2014). Statistical thinking: No one left
behind. In E. J. Chernoff & B. Sriraman (Eds.), Probabilistic thinking
(pp. 127–148). The Netherlands: Springer. http://dx.doi.org/10.1007/
978-94-007-7155-0_8
Meder, B., & Nelson, J. D. (2012). Information search with situation-
specific reward functions. Judgment and Decision Making, 7, 119–148.
Medin, D. L., & Edelson, S. M. (1988). Problem structure and the use of
base-rate information from experience. Journal of Experimental Psy-
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.

chology: General, 117, 68–85. http://dx.doi.org/10.1037/0096-3445
.117.1.68
Micallef, L., Dragicevic, P., & Fekete, J. (2012). Assessing the effect of
visualizations on Bayesian reasoning through crowdsourcing. IEEE
Transactions on Visualization and Computer Graphics, 18, 2536–2545.
http://dx.doi.org/10.1109/TVCG.2012.199
Najemnik, J., & Geisler, W. S. (2005). Optimal eye movement strategies in
visual search. Nature, 434, 387–391. http://dx.doi.org/10.1038/
nature03390
Najemnik, J., & Geisler, W. S. (2008). Eye movement statistics in humans
are consistent with an optimal search strategy. Journal of Vision, 8,
4.1–414. http://dx.doi.org/10.1167/8.3.4
Navarro, D. J., & Perfors, A. F. (2011). Hypothesis generation, sparse
categories, and the positive test strategy. Psychological Review, 118,
120–134. http://dx.doi.org/10.1037/a0021110
Neace, W., Michaud, S., & Bolling, L. (2008). Frequency formats, prob-
ability formats, or problem structure? A test of the nested-sets hypothesis
in an extensional reasoning task. Judgment and Decision Making, 3,
140–152.
Nelson, J. D. (2005). Finding useful questions: On Bayesian diagnosticity,
probability, impact, and information gain. Psychological Review, 112,
979–999. http://dx.doi.org/10.1037/0033-295X.112.4.979
Nelson, J. D. (2008). Towards a rational theory of human information
acquisition. In M. Oaksford & N. Chater (Eds.), The probabilistic mind:
Prospects for rational models of cognition (pp. 143–164). Oxford,
United Kingdom: Oxford University Press. http://dx.doi.org/10.1093/
acprof:oso/9780199216093.003.0007
Nelson, J. D., & Cottrell, G. W. (2007). A probabilistic model of eye
movements in concept formation. Neurocomputing, 70, 2256–2272.
http://dx.doi.org/10.1016/j.neucom.2006.02.026
Nelson, J. D., Divjak, B., Gudmundsdottir, G., Martignon, L. F., & Meder,
B. (2014). Children’s sequential information search is sensitive to en-
vironmental probabilities. Cognition, 130, 74–80. http://dx.doi.org/10
.1016/j.cognition.2013.09.007
Nelson, J. D., McKenzie, C. R., Cottrell, G. W. G., & Sejnowski, T. J. T.
(2010). Experience matters: Information acquisition optimizes probabil-
ity gain. Psychological Science, 21, 960–969. http://dx.doi.org/10.1177/
0956797610372637
Nelson, J. D., Meder, B., & Jones, M. (2016). Optimal experimental
design, heuristics, and sequential search. [Manuscript submitted for
publication].
Oaksford, M., & Chater, N. (1996). Rational explanation of the selection
task. Psychological Review, 103, 381–391. http://dx.doi.org/10.1037/
0033-295X.103.2.381
Okan, Y., & Garcia-Retamero, R. (2012). When higher bars are not larger
quantities: On individual differences in the use of spatial information in
graph comprehension. Spatial Cognition & Computation, 12, 195–218.
Ottley, A., Peck, E. M., Harrison, L. T., Afergan, D., Ziemkiewicz, C.,
Taylor, H. A., . . . Chang, R. (2016). Improving Bayesian reasoning: The
effects of phrasing, visualization, and spatial ability. IEEE Transactions
on Visualization and Computer Graphics, 22, 529–538. http://dx.doi
.org/10.1109/TVCG.2015.2467758
Paolacci, G., Chandler, J., & Ipeirotis, P. (2010). Running experiments on
amazon mechanical Turk. Judgment and Decision Making, 5, 411–419.
Peters, E., Västfjäll, D., Slovic, P., Mertz, C. K., Mazzocco, K., & Dickert,
S. (2006). Numeracy and decision making. Psychological Science, 17,
407–413. http://dx.doi.org/10.1111/j.1467-9280.2006.01720.x
Raiffa, H., & Schlaifer, R. (1961). Applied statistical decision theory.
Cambridge, MA: Division of Research, Graduate School of Business
Administration, Harvard University.
Renninger, L. W., Verghese, P., & Coughlan, J. (2007). Where to look
next? Eye movements reduce local uncertainty. Journal of Vision, 7, 6.
http://dx.doi.org/10.1167/7.3.6
Reyna, V. F., Nelson, W. L., Han, P. K., & Dieckmann, N. F. (2009). How
numeracy influences risk comprehension and medical decision making.
Psychological Bulletin, 135, 943–973. http://dx.doi.org/10.1037/
a0017327
Ruggeri, A., & Lombrozo, T. (2015). Children adapt their questions to
achieve efficient search. Cognition, 143, 203–216. http://dx.doi.org/10
.1016/j.cognition.2015.07.004
Ruggeri, A., Lombrozo, T., Griffiths, T. L., & Xu, F. (2016). Sources of
developmental change in the efficiency of information search. Develop-
mental Psychology, 52, 2159 –2173. http://dx.doi.org/10.1037/
dev0000240
Rusconi, P., & McKenzie, C. R. M. (2013). Insensitivity and oversensi-
tivity to answer diagnosticity in hypothesis testing. The Quarterly Jour-
nal of Experimental Psychology: Human Experimental Psychology, 66,
2443–2464. http://dx.doi.org/10.1080/17470218.2013.793732
Savage, L. (1954). The foundations of statistics. Journal of Consulting
Psychology. New York, NY: Wiley.
Schwartz, L. M., Woloshin, S., Black, W. C., & Welch, H. G. (1997). The
role of numeracy in understanding the benefit of screening mammogra-
phy. Annals of Internal Medicine, 127, 966–972. http://dx.doi.org/10
.7326/0003-4819-127-11-199712010-00003
Sedlmeier, P., & Gigerenzer, G. (2001). Teaching Bayesian reasoning in
less than two hours. Journal of Experimental Psychology: General, 130,
380–400. http://dx.doi.org/10.1037/0096-3445.130.3.380
Simon, H. A. (1978). On the forms of mental representation. Perception
and Cognition: Issues in the Foundations of Psychology, 9, 3–18.
Skov, R. B., & Sherman, S. J. (1986). Information-gathering processes:
Diagnosticity, hypothesis-confirmatory strategies, and perceived hy-
pothesis confirmation. Journal of Experimental Social Psychology, 22,
93–121. http://dx.doi.org/10.1016/0022-1031(86)90031-4
Sloman, S., Over, D., Slovak, L., & Stibel, J. M. (2003). Frequency
illusions and other fallacies. Organizational Behavior and Human De-
cision Processes, 91, 296–309. http://dx.doi.org/10.1016/S0749-
5978(03)00021-9
Slowiaczek, L. M., Klayman, J., Sherman, S. J., & Skov, R. B. (1992).
Information selection and use in hypothesis testing: What is a good
question, and what is a good answer? Memory & Cognition, 20, 392–
405. http://dx.doi.org/10.3758/BF03210923
Stone, E. R., Sieck, W. R., Bull, B. E., Frank Yates, J., Parks, S. C., &
Rush, C. J. (2003). Foreground:background salience: Explaining the
effects of graphical displays on risk avoidance. Organizational Behavior
and Human Decision Processes, 90, 19–36. http://dx.doi.org/10.1016/
S0749-5978(03)00003-7
Tsai, J., Miller, S., & Kirlik, A. (2011). Interactive visualizations to
improve Bayesian reasoning. Proceedings of the Human Factors and
Ergonomics Society Annual Meeting, 55, 385–389. http://dx.doi.org/10
.1177/1071181311551079
Tsallis, C. (1988). Possible generalization of Boltzmann-Gibbs statistics.
Journal of Statistical Physics, 52, 479–487. http://dx.doi.org/10.1007/
BF01016429
Wells, G. L., & Lindsay, R. C. (1980). On estimating the diagnosticity of
eyewitness nonidentifications. Psychological Bulletin, 88, 776–784.
Zhu, L., & Gigerenzer, G. (2006). Children can solve Bayesian problems:
The role of representation in mental computation. Cognition, 98, 287–
308. http://dx.doi.org/10.1016/j.cognition.2004.12.003
Received May 31, 2016
Revision received October 24, 2016
Accepted November 5, 2016 䡲
This
document
is
copyrighted
by
the
American
Psychological
Association
or
one
of
its
allied
publishers.
This
article
is
intended
solely
for
the
personal
use
of
the
individual
user
and
is
not
to
be
disseminated
broadly.
1297

Asking Better Questions How Presentation Formats Influence Information Search

More Related Content

Similar to Asking Better Questions How Presentation Formats Influence Information Search (20)

More from Scott Bou (20)

Recently uploaded (20)

Asking Better Questions How Presentation Formats Influence Information Search