Preprint of the Book Chapter: “Bayesian Versus

Frequentist Inference”

Eric-Jan Wagenmakers1 ∗ ,
Michael Lee2 , Tom Lodewyckx3 , Geoff Iverson2
1 University of Amsterdam, Department of Psychological Methods,
Roetersstraat 15, 1018 WB Amsterdam, The Netherlands
2 University of California Irvine,

Department of Cognitive Sciences, Irvine, USA

3 University of Leuven,

Department of Psychology, Leuven, Belgium

1 Goals and Outline

Throughout this book, the topic of order-restricted inference is dealt with almost
exclusively from a Bayesian perspective. Some readers may wonder why the other
main school for statistical inference – frequentist inference – has received so little
attention here. Isn’t it true that in the field of psychology, almost all inference is
frequentist inference?

The first goal of this chapter is to highlight why frequentist inference is a less-than-
ideal method for statistical inference. The most fundamental limitation of standard
frequentist inference is that it does not condition on the observed data. The resulting
paradoxes have sparked a philosophical debate that statistical practitioners have
conveniently ignored. What cannot be so easily ignored are the practical limitations
of frequentist inference, such as its restriction to nested model comparisons.

The second goal of this chapter is to highlight the theoretical and practical advan-
tages of a Bayesian analysis. From a theoretical perspective, Bayesian inference
is principled and prescriptive, and – in contrast to frequentist inference – a method
that does condition on the observed data. From a practical perspective, Bayesian in-
ference is becoming more and more attractive, mainly because of recent advances
∗ Corresponding author.
Email address: [email protected] (Eric-Jan Wagenmakers1 ).
in computational methodology (e.g., Markov chain Monte Carlo and the WinBUGS
program [1]). To illustrate, one of our frequentist colleagues had been working with
the WinBUGS program and commented “I don’t agree with the Bayesian philoso-
phy, but the WinBUGS program does allow me to implement complicated models
with surprisingly little effort”. This response to Bayesian inference is diametrically
opposed to the one that was in vogue until the 1980’s, when statisticians often
sympathized with the Bayesian philosophy but lacked the computational tools to
implement models with a moderate degree of complexity.

The outline of this chapter is as follows: Section 2 introduces the Fisherian and the
Neyman-Pearson flavors of frequentist inference, and goes on to list a number of
limitations associated with these procedures. Section 3 introduces Bayesian infer-
ence, and goes on to lists a number of its advantages. Section 4 briefly presents our
conclusions.

2 Frequentist Inference and Its Problems

Frequentist inference is based on the idea that probability is a limiting frequency.

This means that a frequentist feels comfortable assigning probability to a repeatable
event in which the uncertainty is due to randomness, such as getting a full house in
poker (i.e., aleatory uncertainty, [2]). When n hands are played, and a full house is
obtained in s cases, then, with n very large, the probability of a full house is just
s/n. But a frequentist must refuse to assign probability to an event where uncer-
tainty is also due to lack of knowledge, such as the event of Alexander Grischuk
ever winning a major poker championship (i.e., epistemic uncertainty, [2] and [3]).

Because uncertainty about parameters is epistemic, frequentist inference does not

allow probability statements about the parameters of a statistical process. For in-
stance, the fact that a frequentist 95% confidence interval for the normal mean µ
is [-0.5, 1.0] does not mean that there is a 95% probability that µ is in [-0.5, 1.0].
Instead, what it means is that if the same procedure to construct confidence inter-
vals was repeated very many times, for all kinds of different data sets, then in 95%
of the cases would the true µ lie in the 95% confidence interval (cf. the example
presented in Section 2.1 below).

Discussion of frequentist inference is complicated by the fact that current practice

has become an unacknowledged amalgamation of the p-value approach advocated
by Fisher [4] and the α-level approach advocated by Neyman and Pearson [5].
Hubbard and Bayarri [6] summarize and contrast the paradigms as follows:

“The level of significance shown by a p value in a Fisherian significance test

refers to the probability of observing data this extreme (or more so) under a null
hypothesis. This data-dependent p value plays an epistemic role by providing

2
 a measure of inductive evidence against H0 in single experiments. This is very
different from the significance level denoted by α in a Neyman-Pearson hypoth-
esis test. With Neyman-Pearson, the focus is on minimizing Type II, or β, errors
(i.e., false acceptance of a null hypothesis) subject to a bound on Type I, or α, er-
rors (i.e., false rejections of a null hypothesis). Moreover, this error minimization
applies only to long-run repeated sampling situations, not to individual experi-
ments, and is a prescription for behaviors, not a means of collecting evidence.”
[6, p. 176]

Clearly then, Fisher’s approach is very different from that of Neyman and Pearson.
Yet, most researchers believe the paradigms have somehow merged, and interpret
the p-value both as a measure of evidence and as a repetitive error rate. It appears
that the confusion between the two different procedures is now close to total, and
it has been argued that this mass confusion “has rendered applications of classi-
cal statistical testing all but meaningless among applied researchers.” [6, p. 171].
Additional references include [7], [8], [9], [10], [11], [12], [13], and [14].

We now discuss several general problems of both the Fisherian and the Neyman-
Pearson procedure (e.g., [15], [16], [17], [18], [14], and [19]). Although only the
Neyman-Pearson procedure is truly frequentist (i.e., it requires knowledge about
performance in long-run sampling situations), we will perpetuate the confusion and
refer to both the Fisherian and the Neyman-Pearson procedure as “frequentist”.

2.1 Frequentist Inference Generally Does not Condition on the Observed Data

As argued by Berger and Wolpert [15], frequentist evidence is often pre-experimental

or unconditional. 1 This means that “A particular procedure is decided upon for use,
and the accuracy of the evidence from an experiment is identified with the long
run behavior of the procedure, were the experiment repeatedly performed.” [15, p.
5]. We illustrate the problem with this unconditional approach by an example that
highlights the pathological properties of frequentist confidence intervals (e.g., [20,
p. 468]).

Consider a uniform distribution with mean µ and width 1. Draw two values ran-
domly from this distribution, label the smallest one s and the largest one l, and
check whether the mean µ lies in between s and l. If this procedure is repeated very
many times, the mean µ will lie in between s and l in half of the cases. Thus, (s, l)
gives a 50% frequentist confidence interval for µ. But suppose that for a particular
draw, s = 9.8 and l = 10.7. The difference between these values is 0.9, and this
covers 9/10th of the range of the distribution; Hence, for these particular values of
1 Frequentist’ procedures sometimes do condition on important aspects of the data. Con-
ditioning is always partial, however, and there exist situations in which it is unclear what
aspects of the data to condition on.

3
 Data y

Distribution y=1 y=2 y=3 y=4 y=5 y=6

f (y)|H0 .04 .30 .31 .31 .03 .01

g(y)|H0 .04 .30 .30 .30 .03 .03
Table 1
Two different sampling distributions, f (y) and g(y), lead to two different p-values for y =
5. Note that y = 5 is equally likely under f (y) and g(y). See text for details

s and l we can be 100% confident that s < µ < l, even though the frequentist
confidence interval would have you believe you should only be 50% confident.

This example shows why it is important to condition on the data that have actually
been observed. The key problem is that frequentist methods do not do this, so that
for data x, “(...) a procedure which looks great pre-experimentally could be terrible
for particular x(...).” [15, p. 9]. Other examples of pathological behavior of frequen-
tist confidence intervals can be found in [20, pp. 466-469], [15], and in particular
[21].

2.2 Frequentist Inference Depends on Data that Were Never Observed

The p-value is the probability under the null hypothesis of observing data at least
as extreme as the data that were actually observed. This means that the p-value is
partly determined by data that were never observed, as is illustrated in the following
example (see also [22], [15], [23], and [19]).

Assume the data y can take on six integer values, y ∈ {1, 2, ..., 6}, according to one
of the sampling distributions f (y) or g(y). Further assume that what is observed is
y = 5. As can be seen from Table 1, the observed datum is equally likely under f (y)
and g(y). Yet, a one-sided p-value is .03+.01 = .04 under f (y), and .03+.03 = .06
under g(y). This is solely due to the fact that the more extreme observation y = 6,
which was never observed, is less likely under f (y) than it is under g(y). Jeffreys
famously summarized the situation: “What the use of P implies, therefore, is that a
hypothesis that may be true may be rejected because it has not predicted observable
results that have not occurred. This seems a remarkable procedure.” [24, p. 385,
italics in original].

4
2.3 Frequentist Inference Depends on the Intention With Which the Data Were
Collected

Because p-values are calculated over the sample space, changes in the sample space
can greatly affect the p-value. For instance, assume that a participant answers a se-
ries of 17 test questions of equal difficulty; 13 answers are correct, 4 are incorrect,
and the last question was answered incorrectly. Under the standard binomial sam-
pling plan (i.e., “ask 17 questions”), the two-sided p-value is .049. The data are,
however, also consistent with a negative binomial sampling plan (i.e., “keep on ask-
ing questions until the fourth error occurs”). Under this alternative sampling plan,
the experiment could have been finished after four questions, or after a million. For
this sampling plan, the p-value is .021.

What this simple example shows is that the intention of the researcher affects sta-
tistical inference – the data are consistent with both sampling plans, yet the p-value
differs. Berger and Wolpert ([15, page 30-33]) discuss the resulting counterintuitive
consequences through a story involving a naive scientist and a frequentist statisti-
cian.

In the story, a naive scientist has obtained 100 independent observations that are
assumed to originate from a normal distribution with mean θ and standard deviation
1. In order to test the null hypothesis that θ = 0, the scientist consults a frequentist
statistician. The mean of the observations is 0.2, and hence the p-value is a little
smaller than .05, which leads to a rejection of the null hypothesis. However, the
statistician decides to probe deeper into the problem and asks the scientist what
he would have done in the fictional case that the experiment had not yielded a
significant result after 100 observations. The scientist replies that he would have
collected another 100 observations. Thus, it may be hypothesized that the implicit
sampling plan was not to collect 100 observation and stop; instead, the implicit
sampling plan was to first take 100 observations and check whether p < .05. When
the check is successful, the experiment stops, but when the check fails, another 100
observations are collected and added to the first 100, after which the experiment
stops.

The statistician then succeeds in convincing the scientist that use of the implicit
sampling plan requires a correction in order to keep the Type I error rate at α =
.05 [25]. Unfortunately, this correction for planning multiple tests now leads to a
p-value that is no longer significant. Therefore, the puzzled scientist is forced to
continue the experiment and collect an additional 100 observations. Note that the
interpretation of the data (i.e., significant or not significant), depends on what the
scientist was planning to do in a situation that did not actually occur. If the very
same data had been collected by a scientist who had answered the statistician’s
question by saying, whether truthfully or not, “I would not have collected any more
observations”, then the data would have been judged to be significant. Same data,

5
different inference.

But the story becomes even more peculiar. Assume that the scientist collects the
next 100 observations, and sets up another meeting with the statistician. The data
are now significant. The statistician, however, persists and asks what the scientist
would have done in case the experiment had not yielded a significant result after 200
observations. Suppose that the scientist now answers “This would have depended
on the status of my grant renewal; If my grant is renewed, I would have had enough
funds to test another 100 observations. If my grant is not renewed, I would have
had to stop the experiment. Not that this matters, of course, because the data were
significant anyway”.

The frequentist statistician then explains that the inference depends on the grant
renewal; if the grant is not renewed, the sampling plan stands and no correction
is necessary. But if the grant is renewed, the scientist could have collected more
data, in the fictional case that the data would not have been significant after 200
observations. This calls for a correction for planning multiple tests, similar to the
first one. Berger and Wolpert ([15, p. 33]) end their story: “The up-to-now honest
scientist has had enough, and he sends in a request to have the grant renewal denied,
vowing never again to tell the statistician what he could have done under alternative
scenarios.”.

We believe that most researchers find it awkward that the conclusions from fre-
quentist statistics depend critically on events that have yet to happen – events that,
moreover, seem to be irrelevant with respect to the data that have actually been
obtained.

2.4 Frequentist Inference Does not Prescribe Which Estimator is Best

Frequentist inference is not derived from a set of simple axioms that describe ra-
tional behavior. This means that any statistical problem potentially affords more
than one frequentist solution, and it may be unclear which one is best. For instance,
many different estimators may be proposed for a particular parameter θ. Which es-
timator for θ should we prefer? The common strategy is to narrow down the set of
admissible estimators by considering only estimators that are unbiased. An estima-
tor t(·) based on data y is unbiased when

Z
t(y)p(y|θ)dy = θ, (1)
Y
for all θ, where Y indicates the sample space (cf. [26]). That is, the only estimators
taken into consideration are those that, averaged over the data that could arise, do
not systematically over- or underestimate θ.

6
Although the criterion of unbiasedness has intuitive appeal, it is in fact highly con-
tentious. First, the criterion is based on all possible data sets that could be observed
(i.e., the sample space Y ). This means that the intention of the researcher affects
which estimators are unbiased and which are not. For instance, for the binomial
sampling plan the unbiased estimator is s/n, where s is the number of correct re-
sponses out of a total of n questions, but for the negative binomial sampling plan
the unbiased estimator is (s − 1)/(n − 1). Second, an estimator that is√unbiased for
θ may well be biased for some nonlinear transformation of θ such as θ.

Finally, unbiased estimators may perform uniformly worse√ than biased estimators.
Consider, for instance, the datum y distributed as N ( θ, 1) with θ > 0. The unbi-
ased estimator for θ is t(y) = y 2 − 1. But when |y| < 1, t(y) is negative, which
conflicts with the knowledge that θ > 0. A new estimator tnew (y) may be proposed
that is given by tnew (y) = y 2 − 1 when |y| ≥ 1 and tnew (y) = 0 otherwise. The new
estimator tnew (y) is biased but does uniformly better than the unbiased estimator
t(y) – this means that t(y) is inadmissible (e.g., [27]).

It should also be noted that in the above example, t(y) is biased downwards when
|y| < 1, but biased upwards when |y| ≥ 1. Thus, an estimator may be unbiased for
all possible data sets taken together, but it may – at the same time – be biased for
every single data set considered in isolation [27, p. 122].

Frequentist statisticians are aware of this problem, in the sense that they acknowl-
edge that “(...) an overly rigid insistence upon unbiasedness may lead to difficul-
ties.” [28, p. 432]. This statement highlights an important problem: frequentist
inference does not specify a unique solution for every statistical problem. When
unfortunate consequences of, say, “an overly rigid insistence upon unbiasedness”
become apparent, ad-hoc estimators may be proposed that remedy the immediate
problem – but this clearly is not a satisfactory state of affairs.

2.5 Frequentist Inference Does not Quantify Statistical Evidence

According to the Fisherian tradition, p-values reflect the strength of evidence against
the null hypothesis. General guidelines associate specific ranges of p-values with
varying levels of evidence: a p-value greater than .1 yields “little or no real evidence
against the null hypothesis”, a p-value less than .1 but greater than .05 implies “sug-
gestive evidence against the null hypothesis”, a p-value less than .05 but greater
than .01 yields “moderate evidence against the null hypothesis” and a p-value less
than .01 constitutes “very strong evidence against the null hypothesis” [29, p. 9];
see also [30, p. 157].

If p-values truly reflect evidence, a minimum requirement is that equal p-values

provide equal evidence against the null hypothesis (i.e., the p-postulate [19]). Ac-
cording to the p-postulate, p = .05 with 10 observations constitutes just as much

7
evidence against the null hypothesis as does p = .05 after 50 observations.

It may not come as a surprise that Sir Ronald Fisher himself was of the opinion that
the p-postulate is correct: “It is not true...that valid conclusions cannot be drawn
from small samples; if accurate methods are used in calculating the probability [the
p-value], we thereby make full allowance for the size of the sample, and should
be influenced in our judgement only by the value of probability indicated.” [31, p.
182], as cited in [14, p. 70].

Nevertheless, some researchers believe that the p-postulate is false, and that p =
.05 after 50 observations is more reliable than p = .05 after 10 observations. For
instance, Rosenthal and Gaito found that the confidence with which a group of
psychologists were willing to reject the null hypothesis increased with sample size
([32]; see also [33]). Consistent with the psychologists’ intuition, an article co-
authored by ten reputable statisticians stated that “A given p-value in a large trial is
usually stronger evidence that the treatments really differ than the same p-value in
a small trial of the same treatments would be.” [34, p. 593], as cited in [14, p. 71].

Finally, several researchers have argued that when the p-values are the same, studies
with small sample size actually provide more evidence against the null hypothesis
than studies with large sample size (e.g., [35], [36], [37], and [33]; for a summary
of the debate see [38]). Abelson considers the very question of whether a researcher
would be happier with a p = .05 after testing 10 cases per group or after testing
50 cases per group, and then firmly concludes “Undergraduates inevitably give the
wrong reply: “Fifty cases per group, because a bigger sample is more reliable” The
appropriate answer is “Ten cases per group, because if the p values are the same,
the observed effect size has to be bigger with a smaller n.”” [35, p. 12].

In order to draw a firm conclusion about the veracity of the p-postulate, we first
need to define what “evidence” is. The details of a rational (i.e., coherent or Bayesian)
definition of evidence are presented in Section 3. Here it suffices to say that such an
analysis must always reject the p-postulate (e.g., [16], [39], and [40]): from a ratio-
nal perspective, p = .05 after only 10 observations is more impressive than p = .05
after 1000 observations. In fact, it may happen that that for a large data set, a fre-
quentist analysis will suggest that the null hypothesis should be rejected, whereas
a rational analysis will suggest that the null hypothesis is strongly supported.

2.6 Frequentist Inference Does not Apply to Non-Nested Models

Consider the study on Dissociative Identity Disorder (DID, [41]), introduced in

Chapter 2 and discussed throughout this book. In this study, Huntjens et al. set out to
study memory processes in DID patients. These patients often report inter-identity
amnesia, that is, impaired memory for events experienced by personalities that are
not currently present. For instance, the personality “lonely girl” may have limited

8
or no knowledge of the events experienced by the personality “femme fatale”. To
test whether DID-patients were really affected by inter-identity amnesia, or whether
they were simulating their affliction, the authors assessed the performance of four
groups of subjects on a multiple-choice recognition test. The dependent measure
was the number of correct responses. The first group were the DID-patients, the
second group were normal controls, the third group were normal controls instructed
to simulate inter-identity amnesia, and the fourth group were normal controls who
had never seen the study list and were therefore truly amnesic.

From the psychological theorizing that guided the design of the experiment, one
can extract several hypotheses concerning the relative performance of the different
groups. One hypothesis, H1a , states that the mean recognition scores µ for DID
patients and True Amnesiacs are the same, and that their scores are higher than
those of the Simulators: µcon > {µamn = µpat } > µsim . Another hypothesis,
H1b , states that the mean recognition scores µ for DID patients and Simulators
are the same, and that their scores are lower than those of the True Amnesiacs:
µcon > µamn > {µpat = µsim }.

The hypotheses H1a and H1b are non-nested, and frequentist inference is not well
suited for the comparison of such models (e.g., [42]). The main problem is that
it is not clear whether H1a or H1b should be considered the null hypothesis. One
might try both possibilities, but this runs the danger of simultaneously rejecting
(or accepting) both H1a and H1b . Moreover, it is not clear how to interpret the
hypothetical result of p = .04 when H1a serves as the null hypothesis, and p = .06
when H1b serves as the null hypothesis – even though H1a is rejected and H1b is
not, this does not mean that H1b is much better than H1a .

2.7 Interim Conclusion

In the preceding analyses we have argued that frequentist procedures suffer from
fundamental philosophical and practical problems. These problems are not some
kind of well-kept secret, as statisticians have written about these frequentist flaws
for many decades – the website http://biology.uark.edu/coop/Courses/
thompson5.html documents some of their efforts by listing 402 articles and
books that criticize the use of frequentist null hypothesis testing.

Indeed, the selection of problems mentioned in Sections 2.1 to 2.6 was certainly not
exhaustive. Other problems include the fact that α levels are arbitrary (“surely, God
loves the .06 nearly as much as the .05”, [43, p. 1277]), the fact that inference in se-
quential designs is overly complicated and conservative (““Sequential analysis” is
a hoax.”, [44, p. 381]), the fact that p-values are often misinterpreted, even by those
teaching statistics [45], the fact that Fisherian frequentist inference does not allow
one to obtain evidence in support of the null hypothesis, and the fact that frequen-

9
tist inference is internally inconsistent or incoherent. The latter means that when
statistical conclusions need to be backed up by betting on them, the frequentist will
be a sure loser (for details see Section 3.1).

All of this makes one may wonder why – despite the harsh criticism – the flogged
horse of frequentist inference is still alive and well, at least in the field of psychol-
ogy [35]. We believe the reason for this is most likely an unfortunate combination
of several factors. Among the most important of these are ease of application, pre-
sumed lack of an appealing alternative, limited statistical knowledge among prac-
titioners, faulty and one-sided teaching of statistics at universities, historical prece-
dent, and – for a few special cases – exact numerical correspondence of frequentist
“flogged horse” inference with rational inference to be discussed below.

This concludes our summary of frequentist inference and its problems. We now
turn to a discussion of the other major statistical paradigm for statistical inference,
which differs from frequentist inference in a few key assumptions. We will argue
that, both philosophically and practically, this paradigm constitutes a superior al-
ternative to frequentist inference.

3 Bayesian Inference and Its Advantages

In Bayesian inference, parameters are random variables. Uncertainty or degree of

belief with respect to the parameters is quantified by probability distributions. For
a given model, say H1 , the prior distribution p(θ|H1 ) for a parameter θ is updated
after encountering data y to yield a posterior distribution p(θ|y, H1 ). The posterior
information contains all of the relevant information about θ. Note that the posterior
distribution is conditional on the data y that have been observed; data that could
have been observed, but were not, do not affect Bayesian inference.

Specifically, Bayes’ rule states that the posterior distribution p(θ|y, H1 ) is propor-
tional to the product of the prior p(θ|H1 ) and the likelihood f (y|θ, H1 ):

p(θ|y, H1 ) = p(θ|H1 )f (y|θ, H1 )/m(y|H1 ). (2)

In this equation, m(y|H1 ) is the marginal probability of the data – it is computed

by integrating out the model parameters using the law of total probability:
Z Z
m(y|H1 ) = p(y, θ|H1 )dθ = p(θ|H1 )f (y|θ, H1 )dθ. (3)

This shows that m(y|H1 ) can also be interpreted as a weighted average likelihood
where the weights are provided by the prior distribution p(θ|H1 ). Because m(y|H1 )
is a number that does not depend on θ, m(y|H1 ) can be conveniently ignored when
the goal is to estimate θ. However, when the goal is Bayesian hypothesis testing,

10
m(y|H1 ) becomes critically important.

For concreteness, consider the choice between two possibly non-nested models, H1
and H2 . The extension to more than two models is entirely possible and follows the
same recipe. Bayes’ rule dictates how the prior probability of H1 , p(H1 ), is updated
through the data to give the posterior probability of H1 , p(H1 |y):
X
p(H1 |y) = p(H1 )m(y|H1 )/ p(Ht )m(y|Ht ). (4)
t

In the same way, one can calculate the posterior probability of H2 , p(H2 |y). The
ratio of these posterior probabilities is given by

p(H1 |y) p(H1 ) m(y|H1 )

= , (5)
p(H2 |y) p(H2 ) m(y|H2 )

which shows that the posterior odds p(H1 |y)/p(H2 |y) is equal to the product of the
prior odds p(H1 )/p(H2 ) and the ratio of marginal probabilities m(y|H1 )/m(y|H2 ).
Thus, the ratio of marginal probabilities – henceforth referred to as the Bayes factor
[24] – quantifies the change from prior to posterior odds brought about by the data.
The Bayes factor, or the log of the Bayes factor, is often interpreted as the weight
of evidence coming from the data [46]. Thus, a Bayes factor hypothesis test prefers
the model under which the observed data are most likely (for details see [47], [20,
Chapter 6], [48, Chapter 7], [49], and [50]; for an introduction in Bayesian infer-
ence see Chapters 3 and 4 of the present book).

Jeffreys [24] proposed to label the evidence provided by the Bayes factor according
to a classification scheme that was subsequently revised by Raftery [51]. Table 2
shows the Raftery classification scheme. The first column shows the Bayes factor,
and the second column shows the associated posterior probability when it is as-
sumed that both H1 and H2 are a priori equally plausible. The third column shows
the verbal labels for the evidence at hand, in the case of a comparison between two
models. Note that these verbal labels are associated with the level of evidence that
is provided by the Bayes factor (i.e., a comparison between two models). These
verbal labels should not be associated with posterior model probabilities (PMPs)
when the set of candidate models is larger than two – for example, consider the
problem of finding the best set of predictors for a regression equation. By consider-
ing all possible combinations of predictors, the model space can easily comprise as
many as 100,000 candidate models. When a single model out of such a large set has
a posterior probability of, say, .50, this would constitute a dramatic increase over
its prior probability of 0.000001, and hence the data provide “very strong” rather
than “weak” evidence in its favor.

Bayesian procedures of parameter estimation and hypothesis testing have many ad-
vantages over their frequentist counterparts. Below is a selective list of ten specific
advantages that the Bayesian paradigm affords.

11
 Bayes factor p(H1 |y) Evidence
BF12

1–3 .50–.75 Weak

3–20 .75–.95 Positive
20–150 .95–.99 Strong
> 150 > .99 Very strong
Table 2
Interpretation of the Bayes factor in terms of evidence (cf. Raftery, 1995, Table 6). Note.
p(H1 |y) is the posterior probability for H1 , given that p(H1 ) = p(H2 ) = 1/2

3.1 Coherence

Bayesian inference is prescriptive; given the specification of a model, there exists

only one way to obtain the appropriate answer. Bayesian inference does not require
ad-hoc solutions to remedy procedures that yield internally inconsistent results.
Bayesian inference is immune from such inconsistencies because it is founded on a
small set of axioms for rational decision making. Several axiom systems have been
proposed, but they all lead to the same conclusion: reasoning under uncertainty
can only be coherent if it obeys the laws of probability theory (e.g., [20], [52], [53],
[54], [55], [18], [56], [57], [58], and [59]).

One of the famous methods to prove this far-reaching conclusion is due to Bruno de
Finetti and involves a betting scenario [53]. Assume there exists a legally binding
ticket that guarantees to pay 1 euro should a proposition turn out to be true. For
instance, the proposition could be “In 2010, the Dutch national soccer team will
win the world cup”. Now you have to determine the price you are willing to pay for
this ticket. This price is the “operational subjective probability” that you assign to
the proposition.

The complication is that this scenario also features an opponent. The opponent can
decide, based on the price that you determined, to either buy this ticket from you
or to make you buy the ticket from him. This is similar to the “I cut, you choose”
rule where the person who cuts a cake gets to choose the last piece; it is then in that
person’s own interest to make a fair judgement.

In the example of the ticket, it is obviously irrational to set the price higher than
1 euro, because the opponent will make you buy this ticket from him and he is
guaranteed to make a profit. It is also irrational to set the price lower than 0 euro,
because the opponent will “buy” the ticket from you at a negative price (i.e., gaining
money) and is again guaranteed to make a profit.

12
Now suppose you have to determine the price of three individual tickets. Ticket A
states “In 2010, the Dutch national soccer team will win the world cup”; ticket B
states “In 2010, the French national soccer team will win the world cup”; and ticket
C states “In 2010, either the Dutch or the French national soccer team will win the
world cup”. You can set the prices any way you want. In particular, there is nothing
to keep you from setting the prices such that price(ticket A) + price(ticket B) 6=
price(ticket C). However, when you set the prices this way you are guaranteed to
lose money compared to your opponent; for instance, suppose you set price(ticket
A) = 0.5 euro, price(ticket B) = 0.3 euro, and price(ticket C) = 0.6 euro. Then the
opponent will buy ticket C from you, sell you tickets A and B, and he is guaranteed
to come out ahead. A set of wagers that ensures that somebody will make a profit,
regardless of what happens, is called a Dutch book.

Using betting scenarios such as the above, de Finetti showed that the only way to
determine subjective values and avoid a certain loss is to make these values obey
the rules of probability theory (i.e., the rule that probabilities lie between 0 and
1, the rule that mutually exclusive events are additive, and the rule of conditional
probability). That is, the only way to avoid a Dutch book is to make your prices for
the separate tickets cohere according to the laws of probability calculus.

The concept of coherence refers not just to the betting scenario, but more generally
to the combination of information in a way that is internally consistent. For exam-
ple, consider Bayesian inference in the case that the data arrive in two batches, y1
and y2 [27, pp. 64-65]. Following the adage “today’s posterior is tomorrow’s prior”
[26, p. 2], we can update from the initial prior p(θ) to a posterior p(θ|y1 ), and then
update this posterior again, effectively treating p(θ|y1 ) as a prior, to obtain finally
p(θ|y1 , y2 ). The crucial aspect is that, when the data are conditionally independent,
it does not matter whether we observe the data set batch-by-batch, all at once, or in
reverse order.

As an additional example of coherence, consider a set of three models: H1 postu-

lates that µa = µb = µc , H2 postulates that {µa = µb } > µc , and H3 postulates that
µa > {µb = µc }. Then, denoting the Bayes factor for model Hi over model Hj by
BFij , we can deduce from the identity

m(y|H1 ) m(y|H1 ) m(y|H3 )

= , (6)
m(y|H2 ) m(y|H3 ) m(y|H2 )

that BF12 = BF32 × BF13 . This means that if the data are twice as likely under
H1 than under H3 , and thrice as likely under H3 than under H2 , we know that the
data are six times more likely under H1 than under H2 . If the Bayes factors would
not commute like this, one could construct a situation in which one could hold
intransitive beliefs – a situation that would violate the axioms of rational decision
making upon which Bayesian inference rests.

13
3.2 Automatic Parsimony

In statistical hypothesis testing, the ideal model captures all of the replicable struc-
ture and ignores all of the idiosyncratic noise. Such an ideal model yields the best
predictions for unseen data coming from the same source. When a model is too
complex, it is said to overfit the data; the model mistakenly treats idiosyncratic
noise as if it were replicable structure. When a model is too simple, it is said to
underfit the data, which means that the model fails to capture all of the replicable
structure in the data. Models that underfit or overfit the data provide suboptimal
predictions and are said to generalize poorly (e.g., [60] and [61]).

The main challenge of hypothesis testing or model selection is to identify the model
with the best predictive performance. However, it is not immediately obvious how
this should be done; complex models will generally provide a better fit to the ob-
served data than simple models, and therefore one cannot simply prefer the model
with the best “goodness-of-fit” – such a strategy would lead to massive overfit-
ting. Intuition suggest that this tendency for overfitting should be counteracted by
putting a premium on simplicity. This intuition is consistent with the law of par-
simony or “Ockham’s razor” (e.g., see http://en.wikipedia.org/wiki/
Occam’s_Razor) which states that, when everything else is equal, simple mod-
els are to be preferred over complex models [18, Chapter 20].

Formal model selection methods try to quantify the tradeoff between goodness-of-
fit and parsimony. Many of these methods measure a model’s overall performance
by the sum of two components, one that measures descriptive accuracy and one that
places a premium on parsimony. The latter component is also known as the Ock-
ham factor [62, Chapter 28]. For many model selection methods, the crucial issue
is how to determine the Ockham factor. One of the attractive features of Bayesian
hypothesis testing is that it automatically determines the model with the best pre-
dictive performance – Bayesian hypothesis testing therefore incorporates what is
known as an automatic Ockham’s razor. In order to see why this is the case, we
explore two lines of reasoning.

First, recall that Bayesian model selection is based on the marginal probability
of the data given model t, m(y|Ht ). Now denote a sequence of n data points by
y n = (y1 , ..., yn ); for example, yi−1 denotes the (i − 1)th individual data point,
whereas y i−1 denotes the entire sequence of observations ranging from y1 up to
and including yi−1 . Quantify predictive performance for a single data point by
the logarithmic loss function − ln p̂i (yi ): the larger the probability that p̂i (de-
termined based on the previous observations y i−1 ) assigns to the observed out-
come yi , the smaller the loss. From the definition of conditional probability, i.e.,
p(yi |y i−1 ) = p(y i )/p(y i−1 ), it then follows that the marginal probability of the
data may be decomposed as a series of sequential, “one-step-ahead” probabilistic
predictions (e.g., [63] and [64]):

14
 m(y n |Ht ) = p(y1 , ..., yn |Ht )
= p(yn |y n−1 , Ht )p(yn−1 |y n−2 , Ht )...p(y2 |y1 , Ht )p(y1 |Ht ). (7)

Thus, Equation 7 shows that the model with the highest marginal probability will
also have the smallest sum of one-step-ahead prediction errors, as − ln m(y n |Ht ) =
Pn i−1
i=1 − ln p(yi |y , Ht ).

According to the second line of reasoning, every statistical model makes a priori
predictions. Complex models have a relatively large parameter space, and are there-
fore able to make many more predictions and cover many more eventualities than
simple models. However, the drawback for complex models is that they need to
spread out their prior probability across their entire parameter space. In the limit,
a model that predicts almost everything has to spread out its prior probability so
thinly that the occurrence of any particular event will not greatly add to that model’s
credibility. Formally, the marginal probability of the data is calculated by averaging
the likelihood f (y|θ, Ht ) over the prior p(θ|Ht ). When the prior is very spread out,
it will occupy a relatively large part of the parameter space in which the likelihood
is almost zero, and this greatly decreases the average or marginal likelihood.

As a more concrete example, consider two people, Bart and Lisa, who each get 100
euros to bet on the winner of the 2010 world cup soccer. Bart decides to divide his
money evenly over 10 candidate teams, including those from Brazil and Germany.
Lisa divides her money over just two teams, betting 60 euros on the team from
Brazil and 40 euros on the team from Germany. Now if either Brazil or Germany
turn out to win the 2010 world cup, Lisa wins more money than Bart. By betting all
her money on just two teams, Lisa was willing to take a risk, whereas Bart was just
trying to keep his options open. For Bart, this means that even if his prediction of
Brazil winning turns out to be correct, he will still lose the 90 euros he betted on the
other countries to win. The point of the story is that, both at the betting office and
in Bayesian inference, hedging your bets is not necessarily the best option, because
this requires you to spread your resources – be it money or prior probability mass
– thinly over the alternative options.

3.3 Extension to Non-Nested Models

Bayesian hypothesis testing is based on the marginal probability of the data given
model t, m(y|Ht ), and therefore it does not make a fundamental distinction be-
tween nested and non-nested models. This means that Bayesian hypothesis testing
can be applied in many more situations than frequentist hypothesis testing. In cog-
nitive psychology, for instance, important substantive questions concern the extent
to which the law of practice follows a power function versus an exponential func-

15
tion, or the extent to which category learning is best described by an exemplar
model or a prototype model. For Bayesian inference, the substantive questions can
be statistically tested in exactly the same way, whether the competing models are
nested or not. For frequentist inference, however, the fact that the models are non-
nested causes grave complications.

Another class of possibly non-nested models that are of great relevance for psy-
chologists are those that incorporate order-restrictions. For instance, consider again
the case of the Huntjens et al. study on dissociative identity disorder discussed in
Section 2.6 and throughout this book. For the data from the Huntjens et al. study,
hypothesis H1a states that the mean recognition scores µ for DID patients and True
Amnesiacs are the same, and that their scores are higher than those of the Simula-
tors: µcon > {µamn = µpat } > µsim , whereas hypothesis H1b states that the mean
recognition scores µ for DID patients and Simulators are the same, and that their
scores are lower than those of the True Amnesiacs: µcon > µamn > {µpat = µsim }.
Within the frequentist paradigm, a comparison of these models is problematical.
Within the Bayesian paradigm, however, the comparison is natural and elegant
(e.g., [65], [66], [67], [42], [68], [69], and [70]).

The general recipe, outlined in O’Hagan and Forster [27, pp. 70-71] is to carry out
order-restricted inference by first considering the posterior distribution of the un-
constrained model, and then restricting one’s attention to the part of the posterior
distribution that obeys the parameter constraints. In a Markov chain Monte Carlo
(MCMC) simulation, for instance, this can be accomplished automatically by re-
taining only those samples that are in line with the constraints. The work reported
in this book attests to the ability of Bayesian inference to address substantive psy-
chological questions that involve order-restrictions in a manner that is unattainable
by frequentist means.

3.4 Flexibility

Bayesian inference allows for the flexible implementation of relatively complicated

statistical techniques such as those that involve hierarchical nonlinear models (e.g.,
[71], [72], [73], [74], [75], [76], and [77]). In hierarchical models, parameters for
individual people are assumed to be drawn from a group-level distribution. Such
multi-level structures naturally incorporate both the differences and the common-
alities between people, and therefore provide experimental psychology with the
means to settle the age-old problem of how to deal with individual differences.

Historically, the field of experimental psychology has tried to ignore individual dif-
ferences, pretending instead that each new participant is a replicate of the previous
one [78]. As Bill Estes and others have shown, however, individual differences that
are ignored can lead to averaging artifacts in which the data that are averaged over

16
participants are no longer representative for any of the participants (e.g., [79], [80],
and [81]). One way to address this issue, popular in psychophysics, is to measure
each individual participant extensively, and deal with the data on a participant-by-
participant basis.

In between the two extremes of assuming that participants are completely the same
and that they are completely different lies the compromise of hierarchical modeling
(see also [82]). The theoretical advantages and practical relevance of a Bayesian hi-
erarchical analysis for common experimental designs has been repeatedly demon-
strated by Jeff Rouder and colleagues (e.g., [76], [74], and [77]). Although hier-
archical analyses can be carried out using orthodox methodology (i.e., [83]), there
are strong philosophical and practical reasons to prefer the Bayesian methodology
(e.g., [57] and [84], respectively).

3.5 Marginalization

Bayesian statistics makes it easy to focus on the relevant variables by integrating

out so-called nuisance variables (e.g., [85] and [86]). Consider for instance the case
of the normal distribution, for which the likelihood function is given by
à !
1 (y − µ)2
f (y|µ, σ) = √ exp − . (8)
σ 2π 2σ 2

For this example, we follow [62, Chapter 24] and propose conjugate improper pri-
ors for µ and σ. A prior is said to be conjugate when it is in the same distributional
family as the posterior distribution. For instance, when the prior for µ is normal, the
posterior for µ is also normal. Conjugate priors are often the only ones that allow
analytical derivation of the posterior. A prior is said to be improper when it does
not integrate to a finite number. For instance, when the prior for µ is a normal distri-
bution with mean µ0 = 0 and standard deviation σµ → ∞, this yields a prior that is
flat across the entire real line. For the present example, we use conjugate improper
priors on µ and σ because they lead to elegant analytical results that correspond to
results from frequentist inference.

In particular, we assume here that the prior on µ is normal with mean µ0 = 0 and
standard deviation σµ → ∞. This flat prior simply states that all values of µ are
equally likely a priori. Because σ is always greater than 0, but log σ covers the entire
real line, a standard “uninformative” prior is flat on the log scale, which transforms
to the prior p(σ) = 1/σ. Using these priors, one can analytically derive the joint
posterior distribution of µ and σ given the data, that is, p(µ, σ|y) (e.g., [62, Chapter
24]).

Now that we have defined the priors and know the joint posterior distribution of µ

17
and σ, consider two scenarios in which one needs to eliminate a nuisance parameter.
In the first scenario, we want to learn about the mean µ of a normal distribution with
unknown standard deviation σ. Thus, µ is the parameter of interest, whereas σ is a
parameter that one would like to ignore (i.e., a nuisance parameter).

Using the law of total probability,

R
it is straightforward to marginalize over, or inte-
grate out, σ, asR p(µ|y) = p(µ, σ|y)dσ. The fact that this equation can be rewritten
as p(µ|y) = p(µ|σ, y)p(σ)dσ highlights the fact that the nuisance parameter σ
can only be integrated out once it has been assigned a prior distribution. After in-
tegrating out σ, the resulting posterior marginal distribution for p(µ|y) turns out to
be the Student-t distribution, the famous frequentist distribution for a test statistic
that involves the mean of a normal distribution with unknown variance [87].

In the second situation, we want to learn about the standard deviation σ of a normal
distribution with unknown mean µ. This means that σ is the parameter of interest,
whereas µ is now the nuisance parameter. From the joint posterior distribution of µ
and σ, we can again apply
R
the law of total
R
probability, this time to integrate out µ,
as follows: p(σ|y) = p(σ, µ|y)dµ = p(σ|µ, y)p(µ)dµ. As before, this equation
shows that the nuisance parameter µ can only be integrated out when it has been
assigned a prior distribution. After computing the marginal posterior distribution
p(σ|y), the Most Probable value for σ (given the data y) turns out to be σM P =
q P
S 2 /(n − 1), where n equals the number of observations and S 2 = ni=1 (yi − ȳ)2 .
The factor n−1 (instead of n) also occurs in frequentist inference, where S 2 /(n−1)
is the unbiased estimator for the variance of a normal distribution with unknown
mean.

In sum, Bayesian inference allows the user to focus on parameters of interest by

integrating out nuisance parameters according to the law of total probability. The
resulting marginal posterior distributions may have matching frequentist counter-
parts, but this only holds in a few special cases.

3.6 Validity

Bayesian inference yields results that connect closely to what researchers want to
know. To clarify this claim by analogy, Gerd Gigerenzer has suggested that for
many researchers statistical inference involves an internal Freudian struggle be-
tween the Superego, the Ego, and the Id (e.g., [9] and [12]). In Gigerenzer’s anal-
ogy, the Superego promotes Neyman-Pearson hypothesis testing, in which an α-
level is determined in advance of the experiment. The Ego promotes Fisherian hy-
pothesis testing, in which the precise value of p supposedly measures the strength
evidence against the null hypothesis. Finally, the Id desires that the hypotheses un-
der consideration are assigned probabilities, something that the Superego and Ego
are unable and unwilling to do. As a result of this unconscious internal conflict,

18
researchers often report results from frequentist procedures, but often believe – im-
plicitly or even explicitly – that they have learned something about the probability
of the hypotheses under consideration.

We agree with Gigerenzer that, deep down inside, what researchers really want is to
draw Bayesian conclusions. Or, in the words of Dennis Lindley, “Inside every Non-
Bayesian, there is a Bayesian struggling to get out” (as cited in [18]). This assertion
is supported by the fact that researchers often misinterpret frequentist concepts –
and misinterpret them in a manner that is decidedly Bayesian (i.e., the interpretation
would have been correct if the method of inference had been Bayesian) [45].

To illustrate the foregoing with a concrete example, consider a frequentist confi-

dence interval for the normal mean µ: µ ∈ [−0.5, 1.0]. As we have seen in Sec-
tion 2.1, the correct but counter-intuitive interpretation of this result is that when
the frequentist procedure is applied very many times to all kinds of possible data
sets, the different intervals cover the true value of µ in 95% of the cases. But why
would this be relevant for the researcher who wants to learn about µ for his or her
data? In contrast, consider the same [−0.5, 1.0] interval for µ, but now assume it
is a Bayesian 95% credible interval. Consistent with intuition, and consistent with
what researchers want to know, this Bayesian interval conveys that there is a .95
probability that µ lies in [-0.5, 1.0]. From the viewpoint of “operation subjective
probability” discussed in Section 3.1, this confidence interval means that when a
coherent researcher is asked to set a fair price for a ticket that promises to pay 1
euro should the assertion “µ is in [-0.5, 1.0]” turn out to be true, that researcher will
set the price of the ticket at exactly .95 euro.

3.7 Subjectivity that is Open to Inspection

A common objection to Bayesian inference is that it is subjective and therefore has

no place in scientific communication. For instance, in an article entitled “Why Isn’t
Everyone a Bayesian?”, Bradly Efron argued that “Strict objectivity is one of the
crucial factors separating scientific thinking from wishful thinking” and concluded
that “The high ground of scientific objectivity has been seized by the frequentists”
[88, p. 4].

Efron’s claims need to be amended for several reasons. First, from a subjective
Bayesian perspective, there is no such thing as “strict objectivity”, as reasoning
under uncertainty is always relative to some sort of background knowledge. In
this view, the search for “strict objectivity” is a quixotic ideal. Thus, subjective
Bayesians might want to change Efron’s claim to “The high ground of scientific
objectivity is a concept that cannot be seized by anyone, because it does not exist.”

Second, there exists a school of objective Bayesians, who specify priors according
to certain predetermined rules [89]. Given a specific rule, the outcome of statisti-

19
cal inference is independent of the person who performs the analysis. Examples of
objective priors include the unit information priors (i.e., priors that carry as much
information as a single observation, [90]), priors that are invariant under transfor-
mations [24], and priors that maximize entropy [91]. Objective priors are generally
vague or uninformative, that is, thinly spread out over the range for which they
are defined. Thus, objective Bayesians might want to change Efron’s claim to “Al-
though the high ground of scientific objectivity may appear to be seized by the
frequentists, objective Bayesians have a legitimate claim to scientific objectivity
also.”

Third, frequentist inference is not as objective as one may (wishfully) think. As

illustrated in Section 2.3, the intention with which an experiment is carried out
can have a profound impact on frequentist inference. The undisclosed ideas and
thoughts that guided experimentation are crucial for calculating frequentist mea-
sures of evidence. Berger and Berry conclude that the perceived objectivity of fre-
quentist inference is largely illusionary [92]. Thus, critics of frequentist inference
might want to change Efron’s claim to “Although the high ground of scientific ob-
jectivity may appear to be seized by the frequentists, upon closer inspection this
objectivity is only make-believe, as in reality frequentists have to rely on the hon-
esty and introspective ability of the researchers who collected the data.”

In contrast to frequentist inference, Bayesian inference generally does not depend

on subjective intentions (cf. Section 2.3), nor on data that were never observed (cf.
Section 2.2, [93]). The posterior distribution of parameters θ is written p(θ|y), and
the marginal probability of a model, say H0 , is given by m(y|H0 ) – in both cases, y
is the observed data, and it is irrelevant what other data could have been observed
but were not.

In Bayesian inference, the subjectivity that Efron alluded to comes in through

the specification of the prior distribution for the model parameters. Regardless of
whether this specification occurs automatically, as in the case of objective priors,
or whether it occurs through the incorporation of prior knowledge, as in the case of
subjective priors, the crucial point is that the prior distribution is formally specified
and available for all other researchers to inspect and criticize. This also means that
Bayesian subjectivity can be analyzed by formal methods that quantify robustness
to the prior (e.g., [94] and [95]). Note how different the notion of subjectivity is
for the two paradigms: Bayesian subjectivity is open to inspection, whereas fre-
quentist subjectivity is hidden from view, carefully locked up in the minds of the
researchers that collected the data. Therefore, a final adjustment of Efron’s state-
ment might read “Scientific objectivity is illusionary, and both Bayesian inference
and frequentist inference have subjective elements; the difference is that Bayesian
subjectivity is open to inspection, whereas frequentist subjectivity is not.”

20
3.8 Possibility of Collecting Evidence in Favor of the Null Hypothesis

Bayesian hypothesis testing allows one to obtain evidence in favor of the null hy-
pothesis. In the Fisherian paradigm, p-values can only be used to reject the null hy-
pothesis. The APA task force on statistical inference stressed this point by issuing
the warning “Never use the unfortunate expression “accept the null-hypothesis”.”
[96, p. 599]. Of course, what is unfortunate here is not so much the expression, but
rather the fact that Fisherian p-values are incapable of providing support for the null
hypothesis. This limitation hinders scientific progress, because theories and models
often predict the absence of a difference. In the field of visual word recognition, for
instance, the entry-opening theory [97] predicts that masked priming is absent for
items that do not have a lexical representation; Another example from that literature
concerns the work by Bowers et al. [98], who have argued that priming effects are
equally large for words that look the same in lower and upper case (e.g., kiss/KISS)
or that look different (e.g., edge/EDGE), a finding supportive of the hypothesis that
priming depends on abstract letter identities.

A final example comes from the field of recognition memory, where Dennis and
Humphreys’ bind cue decide model of episodic memory (BCDMEM) predicts the
absence of a list-length effect and the absence of a list-strength effect [99]. This
radical prediction of a null effect allows researchers to distinguish between context-
noise and item-noise theories of inference in memory. Within the Fisherian paradigm,
support for such informative predictions can only be indirect.

In contrast to the Fisherian hypothesis test, the Bayesian hypothesis test quantifies
evidence by comparing the marginal probability of the data given one hypothesis,
say m(y|HA ), to the the marginal probability of the data given another hypothesis,
say m(y|HB ). The null hypothesis has no special status in Bayesian inference,
and evidence for it is quantified just as it is for any other hypothesis, in a way
that automatically strikes a balance between goodness-of-fit and parsimony (cf.
Section 3.2).

3.9 Opportunity to Monitor Evidence as it Accumulates

Bayesian hypothesis testing allows one to monitor the evidence as the data come
in [100]. In contrast to frequentist inference, Bayesian inference does not require
special corrections for “optional stopping” [19].

Consider, for instance, a hypothetical experiment on the neural substrate of disso-

ciative identity disorder. In this experiment, the researcher Marge has decided in
advance to use functional magnetic resonance imaging (fMRI) to test 30 patients
and 90 normal controls in a total of four between-subjects conditions, using the
same design as Huntjens et al. [41]. Marge inspects the data after 15 participants

21
in each condition have been tested, and finds that the results quite convincingly
demonstrate the pattern she hoped to find. Unfortunately for Marge, she cannot
stop the experiment and claim a significant result, as she would be changing the
sampling plan halfway through and be guilty of “optional stopping”. She has to
continue the experiment, wasting not just her time and money, but also the time
and efforts of the people who undergo needless testing.

Within the frequentist paradigm, it is possible to adopt special sampling plans that
take into account the need or desire to monitor the data as they accumulate; how-
ever, these sampling plans yield conclusions that are much more conservative than
the one that assumes a fixed sample size. Thus, the very same data may lead to a
clearly significant result under a fixed sample size scheme, but to a clearly non-
significant result under a variable sample size scheme; the difference is due to the
fact that the variable sample size scheme incorporates a correction for the eventual-
ity that the experiment could have ended at a different point in time than it actually
did.

In contrast, for Bayesian hypothesis testing there is nothing wrong with gathering
more data, examining these data, and then deciding whether or not to stop collecting
new data – no special corrections are needed. As stated by Edwards et al. [16], “(...)
the rules governing when data collection stops are irrelevant to data interpretation.
It is entirely appropriate to collect data until a point has been proven or disproven,
or until the data collector runs out of time, money, or patience.” ([16, p. 193]).

3.10 Possibility of Incorporating Prior Knowledge

Bayesian inference allows prior knowledge to influence conclusions [93]. Priors are
not only tremendously useful for incorporating existing knowledge, they are also
a prerequisite for rational inference: “If one fails to specify the prior information,
a problem of inference is just as ill-posed as if one had failed to specify the data.”
[18, p. 373]. Another perspective on priors was put forward by Berger, who argued
that “(...) when different reasonable priors yield substantially different answers,
can it be right to state that there is a single answer? Would it not be better to admit
that there is scientific uncertainty, with the conclusion depending on prior beliefs?”
[101, p. 125]. Thus, rather than considering priors a nuisance, we believe they are
useful [93], necessary [18], and informative with respect to the robustness of one’s
conclusions [101]. Priors are an integral part of rational inference; one can only
enjoy the Bayesian omelet when one is prepared to break the Bayesian eggs [102,
p. 578].

22
4 Concluding Comments

In experimental psychology, the dominance of frequentist inference is almost com-

plete. The first goal of this chapter was to demonstrate that the frequentist frame-
work, despite its popularity, has several serious deficiencies. The second goal of
this chapter was to show how the Bayesian framework is both flexible and prin-
cipled. Our conclusion is that the field of psychology can gain a lot by moving
toward the Bayesian framework for statistical inference, and by moving away from
the frequentist framework.

Perhaps frequentist inference has survived for so long because researchers translate
the frequentist statistical outcomes to informal Bayesian conclusions. For instance,
most experienced experimental psychologists would take seriously a priming ef-
fect of 25 msec (p = .03, N = 30 subjects, k = 20 items per condition), whereas
they would be skeptical of a priming effect of 4 msec (p = .03, N = 257 sub-
jects, k = 20 items per condition). Such an informal Bayesian interpretation of fre-
quentist results is another indication of the internal conflict between the frequentist
Superego and Ego versus the Bayesian Id (see Section 3.6; [9]).

It is our hope that more and more psychologists will start to move away from fre-
quentist inference and turn instead to formal Bayesian inference. It may take ther-
apy, medication, or perhaps even surgery, but in the end researchers will be happier
people once they allow their inner Bayesian to come out.

23
References

[1] D. J. Spiegelhalter, A. Thomas, N. G. Best, WinBUGS Version 1.4 User Manual,

Medical Research Council Biostatistics Unit, Cambridge, UK, 2004.

[2] A. O’Hagan, Dicing with the unknown, Significance 1 (2004) 132–133.

[3] M. C. Galavotti, A Philosophical Introduction to Probability, CSLI Publications,

Stanford, 2005.

[4] R. A. Fisher, Statistical Methods For Research Workers (13th ed.), Hafner, New York,
1958.

[5] J. Neyman, E. S. Pearson, On the problem of the most efficient tests of statistical
hypotheses, Philosophical Transactions of the Royal Society A 231 (1933) 289–337.

[6] R. Hubbard, M. J. Bayarri, Confusion over measures of evidence (p’s) versus errors
(α’s) in classical statistical testing, The American Statistician 57 (2003) 171–182.

[7] J. O. Berger, Could Fisher, Jeffreys and Neyman have agreed on testing?, Statistical
Science 18 (2003) 1–32.

[8] R. Christensen, Testing Fisher, Neyman, Pearson, and Bayes, The American
Statistician 59 (2005) 121–126.

[9] G. Gigerenzer, The Superego, the Ego, and the Id in statistical reasoning, in:
G. Keren, C. Lewis (Eds.), A Handbook for Data Analysis in the Behavioral Sciences:
Methodological Issues, Erlbaum, Hillsdale (NJ), 1993, pp. 311–339.

[10] G. Gigerenzer, We need statistical thinking, not statistical rituals, Behavioral and
Brain Sciences 21 (1998) 199–200.

[11] G. Gigerenzer, S. Krauss, O. Vitouch, The null ritual: What you always wanted to
know about significance testing but were afraid to ask, in: D. Kaplan (Ed.), The Sage
Handbook of Quantitative Methodology for the Social Sciences, Sage, Thousand
Oaks, CA, 2004, pp. 391–408.

[12] G. Gigerenzer, Mindless statistics, The Journal of Socio–Economics 33 (2004) 587–

606.

[13] S. N. Goodman, P values, hypothesis tests, and likelihood: Implications for

epidemiology of a neglected historical debate, American Journal of Epidemiology
137 (1993) 485–496.

[14] R. M. Royall, Statistical Evidence: A Likelihood Paradigm, Chapman & Hall,

London, 1997.

[15] J. O. Berger, R. L. Wolpert, The Likelihood Principle (2nd ed.), Institute of

Mathematical Statistics, Hayward (CA), 1988.

[16] W. Edwards, H. Lindman, L. J. Savage, Bayesian statistical inference for

psychological research, Psychological Review 70 (1963) 193–242.

24
[17] C. Howson, P. Urbach, Scientific Reasoning: The Bayesian Approach (3rd. ed.),
Open Court, Chicago, 2006.

[18] E. T. Jaynes, Probability Theory: The Logic of Science, Cambridge University Press,
Cambridge, UK, 2003.

[19] E.-J. Wagenmakers, A practical solution to the pervasive problems of p values,

Psychonomic Bulletin & Review 14 (2007) 779–804.

[20] J. M. Bernardo, A. F. M. Smith, Bayesian Theory, Wiley, New York, 1994.

[21] E. T. Jaynes, Confidence intervals vs Bayesian intervals, in: W. L. Harper,

C. A. Hooker (Eds.), Foundations of Probability Theory, Statistical Inference, and
Statistical Theories of Science, Vol. II, D. Reidel Publishing Company, Dordrecht,
Holland, 1976, pp. 175–257.

[22] G. A. Barnard, The meaning of a significance level, Biometrika 34 (1947) 179–182.

[23] D. R. Cox, Some problems connected with statistical inference, The Annals of
Mathematical Statistics 29 (1958) 357–372.

[24] H. Jeffreys, Theory of Probability, Oxford University Press, Oxford, UK, 1961.

[25] S. J. Pocock, Group sequential methods in the design and analysis of clinical trials,
Biometrika 64 (1977) 191–199.

[26] D. V. Lindley, Bayesian Statistics, a Review, SIAM, Philadelphia (PA), 1972.

[27] A. O’Hagan, J. Forster, Kendall’s Advanced Theory of Statistics Vol. 2B: Bayesian
Inference (2nd ed.), Arnold, London, 2004.

[28] A. Stuart, J. K. Ord, S. Arnold, Kendall’s Advanced Theory of Statistics Vol. 2A:
Classical Inference & the Linear Model (6th ed.), Arnold, London, 1999.

[29] W. J. Burdette, E. A. Gehan, Planning and Analysis of Clinical Studies, Charles C.

Thomas, Springfield (IL), 1970.

[30] L. Wasserman, All of Statistics: A Concise Course in Statistical Inference, Springer,

New York, 2004.

[31] R. A. Fisher, Statistical Methods for Research Workers (5th ed.), Oliver and Boyd,
London, 1934.

[32] R. Rosenthal, J. Gaito, The interpretation of levels of significance by psychological

researchers, The Journal of Psychology 55 (1963) 33–38.

[33] N. Nelson, R. Rosenthal, R. L. Rosnow, Interpretation of significance levels and

effect sizes by psychological researchers, American Psychologist 41 (1986) 1299–
1301.

[34] R. Peto, M. C. Pike, P. Armitage, N. E. Breslow, D. R. Cox, S. V. Howard, N. Mantel,

K. McPherson, J. Peto, P. G. Smith, Design and analysis of randomized clinical trials
requiring prolonged observation of each patient, I: Introduction and design, British
Journal of Cancer 34 (1976) 585–612.

25
[35] R. P. Abelson, On the surprising longevity of flogged horses: Why there is a case for
the significance test, Psychological Science 8 (1997) 12–15.

[36] D. Bakan, The test of significance in psychological research, Psychological Bulletin

66 (1966) 423–437.

[37] D. V. Lindley, W. F. Scott, New Cambridge Elementary Statistical Tables, Cambridge

University Press, London, 1984.

[38] R. Royall, The effect of sample size on the meaning of significance tests, The
American Statistician 40 (1986) 313–315.

[39] D. V. Lindley, A statistical paradox, Biometrika 44 (1957) 187–192.

[40] E.-J. Wagenmakers, P. Grünwald, A Bayesian perspective on hypothesis testing,

Psychological Science 17 (2006) 641–642.

[41] R. J. C. Huntjens, M. L. Peters, L. Woertman, L. M. Bovenschen, R. C. Martin,

A. Postma, Inter–identity amnesia in dissociative identity disorder: A simulated
memory impairment?, Psychological Medicine 36 (2006) 857–863.

[42] I. Klugkist, O. Laudy, H. Hoijtink, Inequality constrained analysis of variance: A

Bayesian approach, Psychological Methods 10 (2005) 477–493.

[43] R. L. Rosnow, R. Rosenthal, Statistical procedures and the justification of knowledge

in psychological science, American Psychologist 44 (1989) 1276–1284.

[44] F. J. Anscombe, Sequential medical trials, Journal of the American Statistical

Association 58 (1963) 365–383.

[45] H. Haller, S. Krauss, Misinterpretations of significance: A problem students share

with their teachers?, Methods of Psychological Research 7 (2002) 1–20.

[46] I. J. Good, Weight of evidence: A brief survey, in: J. M. Bernardo, M. H. DeGroot,

D. V. Lindley, A. F. M. Smith (Eds.), Bayesian Statistics 2, Elsevier, New York, 1985,
pp. 249–269.

[47] J. O. Berger, L. R. Pericchi, The intrinsic Bayes factor for model selection and
prediction, Journal of the American Statistical Association 91 (1996) 109–122.

[48] J. Gill, Bayesian Methods: A Social and Behavioral Sciences Approach, CRC Press,
Boca Raton (FL), 2002.

[49] R. E. Kass, A. E. Raftery, Bayes factors, Journal of the American Statistical

Association 90 (1995) 377–395.

[50] A. O’Hagan, Fractional Bayes factors for model comparison, Journal of the Royal
Statistical Society B 57 (1997) 99–138.

[51] A. E. Raftery, Bayesian model selection in social research, in: P. V. Marsden (Ed.),
Sociological Methodology, Blackwells, Cambridge, 1995, pp. 111–196.

[52] R. T. Cox, Probability, frequency and reasonable expectation, The American Journal
of Physics 14 (1946) 1–13.

26
[53] B. de Finetti, Theory of Probability, Vol. 1 and 2, John Wiley & Sons, New York,
1974.
[54] M. H. DeGroot, Optimal Statistical Decisions, McGraw–Hill, New York, 1970.
[55] P. C. Fishburn, The axioms of subjective probability, Statistical Science 1 (1986)
335–345.
[56] D. V. Lindley, Scoring rules and the inevitability of probability, International
Statistical Review 50 (1982) 1–26.
[57] D. V. Lindley, The philosophy of statistics, The Statistician 49 (2000) 293–337.
[58] F. P. Ramsey, Truth and probability, in: R. B. Braithwaite (Ed.), The Foundations of
Mathematics and Other Logical Essays, Kegan Paul, London, 1926, pp. 156–198.
[59] L. J. Savage, The Foundations of Statistics, John Wiley & Sons, New York, 1954.
[60] I. J. Myung, M. R. Forster, M. W. Browne, Model selection [Special issue], Journal
of Mathematical Psychology 44 (1–2).
[61] E.-J. Wagenmakers, L. Waldorp, Model selection: Theoretical developments and
applications [Special issue], Journal of Mathematical Psychology 50 (2).
[62] D. J. C. MacKay, Information Theory, Inference, and Learning Algorithms,
Cambridge University Press, Cambridge, 2003.
[63] A. P. Dawid, Statistical theory: The prequential approach, Journal of the Royal
Statistical Society A 147 (1984) 278–292.
[64] E.-J. Wagenmakers, P. Grünwald, M. Steyvers, Accumulative prediction error and
the selection of time series models, Journal of Mathematical Psychology 50 (2006)
149–166.
[65] A. E. Gelfand, A. F. M. Smith, T.-M. Lee, Bayesian analysis of constrained parameter
and truncated data problems using Gibbs sampling, Journal of the American
Statistical Association 87 (1992) 523–532.
[66] H. Hoijtink, Confirmatory latent class analysis: Model selection using Bayes factors
and (pseudo) likelihood ratio statistics, Multivariate Behavioral Research 36 (2001)
563–588.
[67] I. Klugkist, B. Kato, H. Hoijtink, Bayesian model selection using encompassing
priors, Statistica Neerlandica 59 (2005) 57–69.
[68] I. Klugkist, O. Laudy, H. Hoijtink, Bayesian eggs and Bayesian omelettes: Reply to
Stern (2005), Psychological Methods 10 (2005) 500–503.
[69] O. Laudy, M. Zoccolillo, R. H. Baillargeon, J. Boom, R. E. Tremblay, H. Hoijtink,
Applications of confirmatory latent class analysis in developmental psychology,
European Journal of Developmental Psychology 2 (2005) 1–15.
[70] A. F. M. Smith, G. O. Roberts, Bayesian computation via the Gibbs sampler and
related Markov chain Monte Carlo methods, Journal of the Royal Statistical Society:
Series B 55 (1993) 3–23.

27
[71] R. D. Morey, M. S. Pratte, J. N. Rouder, Problematic effects of aggregation in zROC
analysis and a hierarchical modeling solution, Journal of Mathematical Psychology.

[72] R. D. Morey, J. N. Rouder, P. L. Speckman, A statistical model for discriminating

between subliminal and near–liminal performance, Journal of Mathematical
Psychology.

[73] D. J. Navarro, T. L. Griffiths, M. Steyvers, M. D. Lee, Modeling individual

differences using Dirichlet processes, Journal of Mathematical Psychology 50 (2006)
101–122.

[74] J. N. Rouder, J. Lu, An introduction to Bayesian hierarchical models with an

application in the theory of signal detection, Psychonomic Bulletin & Review 12
(2005) 573–604.

[75] J. N. Rouder, J. Lu, R. D. Morey, D. Sun, P. L. Speckman, A hierarchical process

dissociation model, Journal of Experimental Psychology: General.

[76] J. N. Rouder, J. Lu, P. L. Speckman, D. Sun, Y. Jiang, A hierarchical model for

estimating response time distributions, Psychonomic Bulletin & Review 12 (2005)
195–223.

[77] J. N. Rouder, J. Lu, D. Sun, P. Speckman, R. Morey, M. Naveh-Benjamin, Signal

detection models with random participant and item effects, Psychometrika.

[78] W. H. Batchelder, Cognitive psychometrics: Combining two psychological traditions,

CSCA Lecture, Amsterdam, The Netherlands, October 2007.

[79] W. K. Estes, The problem of inference from curves based on group data,
Psychological Bulletin 53 (1956) 134–140.

[80] W. K. Estes, Traps in the route to models of memory and decision, Psychonomic
Bulletin & Review 9 (2002) 3–25.

[81] A. Heathcote, S. Brown, D. J. K. Mewhort, The power law repealed: The case for an
exponential law of practice, Psychonomic Bulletin & Review 7 (2000) 185–207.

[82] M. D. Lee, M. R. Webb, Modeling individual differences in cognition, Psychonomic

Bulletin & Review 12 (2005) 605–621.

[83] L. Hoffman, M. J. Rovine, Multilevel models for the experimental psychologist:

Foundations and illustrative examples, Behavior Research Methods 39 (2007) 101–
117.

[84] A. Gelman, J. Hill, Data Analysis Using Regression and Multilevel/Hierarchical

Models, Cambridge University Press, Cambridge, 2007.

[85] D. Basu, On the elimination of nuisance parameters, Journal of the American

Statistical Association 72 (1977) 355–366.

[86] J. O. Berger, B. Liseo, R. L. Wolpert, Integrated likelihood methods for eliminating

nuisance parameters, Statistical Science 14 (1999) 1–28.

28
[87] H. Jeffreys, On the relation between direct and inverse methods in statistics,
Proceedings of the Royal Society of London. Series A, Mathematical and Physical
Sciences 160 (1937) 325–348.
[88] B. Efron, Why isn’t everyone a Bayesian?, The American Statistician 40 (1986) 1–5.
[89] R. E. Kass, L. Wasserman, The selection of prior distributions by formal rules,
Journal of the American Statistical Association 91 (1996) 1343–1370.
[90] R. E. Kass, L. Wasserman, A reference Bayesian test for nested hypotheses and its
relationship to the Schwarz criterion, Journal of the American Statistical Association
90 (1995) 928–934.
[91] E. T. Jaynes, Prior probabilities, IEEE Transactions on Systems Science and
Cybernetics 4 (1968) 227–241.
[92] J. O. Berger, D. A. Berry, Statistical analysis and the illusion of objectivity, American
Scientist 76 (1988) 159–165.
[93] D. V. Lindley, The analysis of experimental data: The appreciation of tea and wine,
Teaching Statistics 15 (1993) 22–25.
[94] J. O. Berger, Robust Bayesian analysis: Sensitivity to the prior, Journal of Statistical
Planning and Inference 25 (1990) 303–328.
[95] J. M. Dickey, Scientific reporting and personal probabilities: Student’s hypothesis,
Journal of the Royal Statistical Society B 35 (1973) 285–305.
[96] L. Wilkinson, the Task Force on Statistical Inference, Statistical methods in
psychology journals: Guidelines and explanations, American Psychologist 54 (1999)
594–604.
[97] K. I. Forster, K. Mohan, J. Hector, The mechanics of masked priming, in:
S. Kinoshita, S. J. Lupker (Eds.), Masked Priming: The State of the Art, Psychology
Press, New York, NY, 2003, pp. 3–38.
[98] J. S. Bowers, G. Vigliocco, R. Haan, Orthographic, phonological, and articulatory
contributions to masked letter and word priming, Journal of Experimental
Psychology: Human Perception and Performance 24 (1998) 1705–1719.
[99] S. Dennis, M. S. Humphreys, A context noise model of episodic word recognition,
Psychological Review 108 (2001) 452–477.

[100] J. O. Berger, D. A. Berry, The relevance of stopping rules in statistical inference, in:
S. S. Gupta, J. O. Berger (Eds.), Statistical Decision Theory and Related Topics: Vol.
1, Springer Verlag, New York, 1988, pp. 29–72.
[101] J. O. Berger, Statistical Decision Theory and Bayesian Analysis, 2nd Edition,
Springer, New York, 1985.

[102] L. J. Savage, The foundations of statistics reconsidered, in: J. Neyman (Ed.),

Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and
Probability, Vol. 1, University of California Press, Berkely, CA, 1961, pp. 575–585.

29