This article was downloaded by: [University of Alberta]

On: 7 January 2009

Access details: Access Details: [subscription number 713587337]
Publisher Informa Healthcare
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

Encyclopedia of Biopharmaceutical Statistics

Publication details, including instructions for authors and subscription information:
[Link]

Crossover Design
Stephen Senn a
a
Pharmaceutical and Health Statistics, University College London, London, United Kingdom

Online Publication Date: 23 April 2003

To cite this Section Senn, Stephen(2003)'Crossover Design',Encyclopedia of Biopharmaceutical Statistics,1:1,255 — 262

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: [Link]

This article may be used for research, teaching and private study purposes. Any substantial or
systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or
distribution in any form to anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses
should be independently verified with primary sources. The publisher shall not be liable for any loss,
actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly
or indirectly in connection with or arising out of the use of this material.
 Crossover Design
C
Stephen Senn
University College London, London, United Kingdom

INTRODUCTION first three attacks with the treatments provided in the

order indicated. The main outcome measure was a visual
‘‘A cross-over trial is one in which individual subjects are analogue pain score 2 hr after treatment.[1]
given sequences of treatments with the object of studying Crossover trials are of greater importance for the
differences between individual treatments (or subse- biopharmaceutical statistician than is generally the case
quences of treatments).’’[1] for fellow medical statisticians working outside of the
Crossover trials are potentially extremely efficient industry. Some typical applications in drug develop-
designs for studying the effects of treatment. They are, ment are:
Downloaded By: [University of Alberta] At: 07:19 7 January 2009

however, not suitable for all disease indications and they

can be affected by carryover, a phenomenon that may Phase I
make it impossible to produce unbiased estimates except Pharmacokinetic studies to establish concentration
by sacrificing the efficiency crossover designs are em- time profiles.
ployed to provide. This important issue will be dealt with Bioequivalence.[4]
extensively below. First, however, some general features Food interaction studies.
of such trials, and possible uses, will be reviewed. Dose proportionality.
Dose-escalation studies for investigating maximum
tolerated dose.
OVERVIEW Phase II
Studies of pharmacodynamic response.
The most studied form of crossover is the so-called AB/ Dose finding.
BA crossover in which patients are treated in two periods Parallel assay.
and are randomized to receive either treatment A followed Phase III
by treatment B or vice versa. Specialist scientific studies.
Example 1. A trial was carried out to compare the Determination of individual response.
effects in Swedish schoolchildren suffering from asthma Phase IV
of a single dose of 200 mg salbutamol and 12 mg for- Studies of patient preferences.
moterol, both given by metered-dose inhaler.[2] Fourteen
children were randomized in equal numbers to one of two Example 3. This is an example of a pharmacodynamic
sequences: either formoterol followed by salbutamol or parallel assay in asthma.[5] Three single doses (6, 12, and
salbutamol followed by formoterol. A washout period of 24 mg) for each of two formulations of formoterol to be
at least 2 days was observed between treatments. The given by inhalation in the form of a dry powder were
principal outcome measure was peak expiratory flow in 1 compared to placebo. It was not possible to treat patients
sec (PEF).[1] in more than five periods. Hence, an incomplete blocks
More complicated crossover designs are often used, design in five periods and 21 sequences was chosen, each
however, and designs in which patients receive three, four, patient receiving five of the seven treatments. The trial
or five treatments in as many periods are not uncommon. was carried out in 15 centers in four countries and 161
Example 2. A placebo-controlled trial in migraine was patients were recruited. The main outcome variable was
run in Sweden and Finland to estimate and compare the the area under the forced expiratory volume in 1 sec
effects of two doses (50 and 100 mg) of the potassium salt (FEV1) curve over 12 hr.
of diclofenac.[3] Each of 72 patients received each Crossover trials are also used in preclinical work
treatment once and once only. Patients were randomized involving animals. This topic is not covered here.
to receive one of the six possible sequences of the three Crossover trials are rarely used as pivotal Phase III
treatments. Patients were not to use the first treatment studies in support of a drug application. One reason is that
until they had observed a treatment-free and attack-free drug regulatory agencies are wary of crossover trials
period of at least 1 week. Thereafter they were to treat the because of the potential danger of carryover. This issue is

Encyclopedia of Biopharmaceutical Statistics 255

DOI: 10.1081/E-EBS 120007432
Copyright D 2003 by Marcel Dekker, Inc. All rights reserved.
 256 Crossover Design

dealt with in detail in the section ‘‘Carryover.’’ An even origin of all carryover, however indirectly, must ultimately
more important reason is that a principle advantage of be pharmacological.
crossover trials is to reduce the number of patients in a Where it occurs, the consequence of carryover is that
trial. However, for most drug developments, more the trialist will measure the combined (in some cases
patients are needed to satisfy the regulator as regards partial) effects of two or more treatments. (S)he may be
the tolerability of the product than are needed to de- unaware that this is happening and this may lead to biased
monstrate its efficacy, even if a parallel group trial is assessment and even where detected the disentangling
chosen. There is thus less incentive for the sponsor to use of individual effects may be difficult. The most ap-
crossover trials in Phase III. The exception as regards propriate way to deal with carryover is with washout
sample size requirements would be trials in which sur- periods. These may be ‘‘passive’’ in that no treatment is
vival is the outcome, but these are, of course, in any case given or ‘‘active.’’ The latter strategy occurs quite
not suitably run as crossovers nor, indeed, are any trials commonly in multidose therapeutic trials. Here patients
in which cure or death is the outcome. may be switched over almost immediately to the subse-
In fact, suitable indications for crossovers are chronic quent treatment. The previous treatment will, of course,
diseases in which the patients are relatively stable. Such be being eliminated by the patient’s body during the
indications include asthma, mild or moderate hyperten- time that the new treatment is being given. Provided
sion, rheumatism, migraine, sleep disturbances, angina, measurement of the effects of the new treatment is
Downloaded By: [University of Alberta] At: 07:19 7 January 2009

and epilepsy. delayed until this process is complete, the problem of

It is not the disease alone, however, that determines carryover is dealt with.
the suitability of a crossover. Also relevant is the nature If, however, as is sometimes the case in single-dose
of the treatment. Quickly reversible treatments are more pharmacodynamic studies, it is desired to study onset of
suitable than those for which the effect is more per- action, then a passive washout is necessary. The issue
sistent. Thus, for example, in asthma, beta-agonists are of modeling carryover is considered in due course in
easier to study using crossover trials than are steroids. the following.
Other considerations also apply. An important distinc-
tion in drug development is between single-dose studies,
which often involve pharmacodynamic measures and
may study onset and duration of action, and multidose ANALYSIS USING THE GENERAL
studies in which, more usually, the therapeutic steady- LINEAR MODEL
state effect of treatments is studied. Crossover trials
may be used for both but are generally more suited to In many cases the primary outcomes for a crossover
the former. trial will be continuous. For example, FEV1 is com-
monly used in asthma and diastolic or systolic blood
pressure in hypertension. As explained above, crossover
CARRYOVER trials are often more relevant to studying pharmacody-
namic rather than therapeutic outcomes. In particular,
The outstanding problem of crossover trials is regarded except in rare cases to be discussed below, survival is
by many as being carryover. Carryover has been defined not a relevant outcome for a crossover trial. A general
as, ‘‘the persistence (whether physically or in terms of linear model in which disturbance terms are assumed to
effect) of a treatment applied in one period in a sub- be normally distributed will often be a suitable frame-
sequent period of treatment’’ (Ref. [1]). The simplest work for analysis.
example of carryover is where the half-life of a treat- Crossover trials share a feature of fractional factorials
ment has been underestimated and some relevant con- that can make their representation in a linear model
centration of the previous treatment is still present at somewhat awkward. Consider an AB/BA crossover trial
a time when another is being measured. However, except in two patients only. There are two patients, two periods,
for bioequivalence studies in which drug concentra- and two treatments but only four and not eight observa-
tion itself is the outcome, it is the pharmacodynamic tions. A consequence of this is that and two of the three
time course rather than the pharmacokinetic course that factors patient, period, and treatment determine the third.
is relevant. Much play has also been made in the lit- Thus, if patient 1 is allocated to the AB sequence if the
erature of so-called psychological carryover, where per- current observation is the observation in period 2, it must
haps some memory of the therapeutic experience under be under treatment B.
the previous treatment affects the patient’s current judge- If we ignore the problem of carryover, then, following
ment. However, it should not be forgotten that in a double- Jones and Kenward,[6] a general model for observed
blind trial, to the extent that blinding is successful, the values yijk of a random variable Yijk from a crossover trial
 Crossover Design 257

in g sequences, p periods, N patients, and t treatments is least squares using proc glm1 of SAS1 and treating the
given as follows patient effect as fixed. Alternatively, the patient effects

Yijk ¼ m þ sik þ pj þ td½i;j þ eijk ð1Þ

are sometimes allowed to be random, and this may be
conveniently modeled using proc mixed1 of SAS1.
C
Sometimes, where this is done, a factor for the sequence
Here, m is a general mean, sik is the effect of subject k in itself is introduced to force the treatment estimate to be
sequence i, i = 1,2. . .g, k = 1,2,. . .ni, pj is the effect of a ‘‘within-patient’’ estimate. If this is done, however,
period j, j = 1,2. . .p, td[i,j] is the direct effect of the between-patient information cannot be recovered.
treatment administered in period j to subjects (either This issue can be understood simply by considering
patients or healthy volunteers) in group i, and eijk is a an AB/BA crossover in which some patients have
stochastic disturbance term. dropped out at random after the first period. If a fixed
A number of points may be noted about this model. effect is included for patients, only patients who have
completed both periods will contribute information.
. The eijk terms are often assumed independent. For However, direct comparison of patients who have
designs with three or more periods, however, it may be received A only with those who have received B only
appropriate to allow for autocorrelation within patients can be made if only the patient effects are allowed to be
over time. random (as they must be in any parallel group trial). This
Downloaded By: [University of Alberta] At: 07:19 7 January 2009

. The model is overparameterized, but it is identifiable will generally contribute a very small amount of
contrasts, in particular those of the treatment para- information but can in principle be combined with that
meters, which are of interest. from patients who have completed both periods using an
. The treatment parameter is indexed by a functional appropriate approach to analysis as provided, say, by
subscript d[i,j], which identifies it by position (se- proc mixed1 of SAS1.
quence and period). This is because, as discussed This is a simple example where so-called interblock
above, crossover designs are a form of fractional de- information is available to be recovered. There are more
sign: There are t treatments, N patients, and p periods complex cases. Usually, if carryover is introduced into
but only N  p observations. To complete the spe- the model, it means that interblock information can in
cification, the function d[i,j] must be separately de- principle be recovered. An exception is given by models
fined, say by a table that indicates which treatment in which simple carryover only is allowed for and the
is given to which sequence group in a given period. design chosen is such that simple carryover is orthogonal
. There is no term for carryover in the model. Jones and to the treatment effect. For example, such a design is the
Kenward[6] propose adding a term ld[i,j  1]. This form four-period, two-treatment design using the sequences
of carryover has been referred to as simple carryover[1] AABB, BBAA, ABBA, and BAAA. For incomplete block
and is appropriate if carryover lasts only for one period designs, such as illustrated by example 3 above, inter-
and depends on engendering treatment only (is not block information is available to be recovered whether or
modified by the perturbed treatment). not carryover is fitted.
. The term sk can be treated as fixed or random de- A general feature of conventional approaches to ana-
pending on approach. This may or may not make a lyzing crossover trials is that the residual degrees of
difference to the resulting inference regarding treat- freedom for error can exceed the number of patients.[8]
ment contrasts: Whether it does so depends on whe- Consider, for example, the degrees of freedom for the
ther interblock information is present and this in turn analysis of variance for a complete blocks design in N
depends on the balance of the design, whether auto- patients, p periods, and p treatments. (This sort of design
correlation is allowed or not and whether a carryover is very common. Examples 1 and 2 above are cases in
term is fitted.[7] point with N = 14 and p = 2 and N = 72 and p = 3,
. Whether or not the patient effect is treated as fixed or respectively.) The degrees of freedom for the analysis of
random, the model does not allow for a true random variance may be partitioned as follows:
effect of treatment: the possibility that the effect of
treatment can vary from patient to patient. The fixed Degrees of freedom
effect approach is extremely common but it can be of
interest on occasion to use a random effects approach, Source In general Example 1 Example 2
in particular in designs where the number of periods Patients N1 13 71
exceed the number of treatments. Periods p1 1 2
Treatments p1 1 2
Error Np  N  2p + 2 12 140
A popular approach to analyzing crossover trials
Total Np  1 27 215
within the pharmaceutical industry is to apply ordinary
 258 Crossover Design

Now, in the first example, the degrees of freedom In the above scheme, the two sequences AA and BB
for error are 12 and thus inferior to the number of pa- contribute nothing to the estimate of the treatment effect
tients. In the second, however, there are nearly twice as and would appear to be a complete waste of time and
many degrees of freedom for error as patients and this resources. If, however, carryover were present and that
signals that strong assumptions are involved.[8] An alter- from A into B were not the same as from B into A, the
native approach to using ordinary least squares under treatment estimate so formed would have a bias, â. If it
such cases can be to reduce the data for a given patient could be assumed that the carryover from A into A were
to a contrast of interest and then analyze these contrasts. the same as from B into B, this bias could be estimated
A similar problem arises with any crossover trial with using the weights 0, 0, 1/2,  1/2 on the four cell mean
more than two periods and where the same treatment, differences. Let us call this contrast C2. A little reflection
as may be the case in so-called, ‘‘n of 1 studies,’’ is shows that this second scheme of weights only estimates a
repeatedly given to the same patient, this summary carryover effect: patient, period, and treatment effects are
measures approach may be particularly valuable.[1,8] Al- eliminated from it. It thus follows that the difference
ternatively a true random effect (multilevel) model can between the first linear combination and the second,
be applied. C1  C2, provides an unbiased estimate in the presence of
‘‘simple’’ carryover.
This does no mean that such a design is useful. On
Downloaded By: [University of Alberta] At: 07:19 7 January 2009

the contrary it is not generally useful. The variance of

MODELING CARRYOVER the estimated treatment effect is 4 times what it would
be if all patients had been allocated to the first two se-
There is an enormous literature on adjusting the analysis quences only. This is a very high price to pay for un-
of crossover trials to allow for the effect of carryover and biasedness and modern approaches to data analysis re-
a corresponding interest in appropriate efficient designs. cognize the importance of variance bias trade-offs.[13]
Most authors have enthusiastically (and usually uncrit- There is, however, a more important criticism, namely,
ically) adopted the simple carryover model [6,9–11] that if carryover occurred, it is extremely unlikely that the
whereby carryover depends only on the engendering carryover from A into A would be the same as from A
treatment and last for one period only. Of course, for the into B. For example, A might be an active treatment with
AB/BA design, there are only two periods and the a response either at steady state (in a multiple trial) or
treatment that follows is completely determined by the near the top of a dose – response curve. On the other hand
treatment that precedes. Hence, the simple carryover B might be a placebo. The carryover from A into B would
model is the only one that needs to be considered. thus be more important than from A into A.[1,14–16] There
Approaches to dealing with carryover in the case of the are, in fact, occasions where correcting for simple carry-
AB/BA design will be considered below in due course, over can increase the bias of the resulting treatment
but in any case, those interested in designing supposedly estimate.[1,16]
more efficient crossover trials have usually considered When applying the simple carryover model one can
more complex designs. For these more complex designs, distinguish two rather different cases. The first occurs
simple carryover may not be realistic. where the design is such that adjusting for simple
Consider, for example, the design in two periods in carryover has no adverse effect on the variance. This is,
which patients are randomized in equal numbers to four for example, the case in the four-period design in two
sequences AB, BA, AA, and BB.[12] We can summarize treatments using the four sequences AABB, BBAA,
the responses using eight cell means defined by the ABBA, and BAAB. Note that in this design A follows
product of the four sequences and two periods. If we fit a A twice and B follows A twice and now consider the 16
fixed effect for each patient any estimate of the treatment cell means for this design defined by the cross-
effect (A –B) can be expressed as a linear combination of classification of sequence and period. A natural and fully
the four contrasts produced by taking the period 2 cell efficient estimate of the treatment effect A – B is obtained
mean from a given sequence from the corresponding by multiplying each cell mean corresponding to an A by
period 1 value. If carryover is ignored altogether, the re- 1/8, each cell mean corresponding to B by  1/8, and
levant weights for these four contrasts are 1/2,  1/2, 0, forming the weighted sum. But if simple carryover occurs
and 0. The patient effects are eliminated by virtue of two of the cell means under A are affected by carryover
having used within-cell differences, the period effects are from a preceding A and so are two of the cell means by B.
eliminated because the second weight is the negative of The weights associated with these cell means add to zero
the first and it can also be seen by inspection that the and thus simple carryover is eliminated automatically.
treatment difference is appropriately recovered. Let us Under such circumstances, fitting simple carryover in
call this contrast, C1. the model only reduces the error degrees of freedom and
 Crossover Design 259

because these will generally be many, there is little reason Table 1 Weights for estimating effects for an AB/BA crossover
not to do so. The danger here is rather in assuming that
because simple carryover has been fitted that all forms of
carryover have been eliminated.
Sequence AB BA
C
Treatment A B B A
The second case is where the design is not fully
PAR 1 0 1 0
efficient in the presence of simple carryover. Here, fitting
CROS 1/2  1/2  1/2 1/2
carryover will increase the variance of the treatment CARRY 1/2 1/2  1/2  1/2
estimate. In some cases, it may even increase the bias. It PAR  (CROS + CARRY) 0 0 0 0
is extremely doubtful whether the simple carryover
model is of any value in such cases. A useful review of
approaches to modeling carryover in multiperiod designs since each sequence mean reflects both treatments and
is given by Matthews.[17] both periods so that the difference between means can
only reflect either differences between patients, which
are assumed random, or the order in which treatments
TESTING FOR CARRYOVER were administered.
The general situation is given in Table 1. Note that
Until fairly recently an alternative to adjusting for carry- these three estimates satisfy the relationship PAR =
Downloaded By: [University of Alberta] At: 07:19 7 January 2009

over was popular with applied statisticians. This was CROS + CARRY.[20]
to perform a preliminary test of (simple) carryover, use In 1965, in an influential paper, Grizzle[18] proposed
an adjusted estimate (or test) of the treatment effect if a scheme whereby a preliminary test for CARRY was
this preliminary test was significant and an unadjusted performed at the 10% level. (This higher than usual no-
test if (as would more usually be the case) this was minal level was suggested in view of the low power of
not. This was a particularly popular approach to analy- this test.) If the result was not significant CROS would
zing the AB/BA design and will now be described in be used as the basis for tests of the treatment effect at
that connection.[18,19] the 5% level. If it was significant PAR would be used,
In the presence of carryover, the only unbiased also at the 5% level. This proposal was also clearly
estimate possible from the AB/BA design is that based described in a further influential paper by Hills and
on first-period data only. Clearly, if the second-period Armitage.[19] This scheme was very popular for many
data are discarded, the remaining data have the structure years, and some pharmaceutical companies had even
of a parallel group trial and can be analyzed as such. Note written SAS macros to perform this two-stage testing pro-
that as discussed above, the implicit assumption is that the cedure automatically. However, in an extremely import-
patient effect is random. An estimate produced under such ant paper published in 1989, Freeman[21] showed that
circumstances will be referred to in the discussion that the procedure as a whole does not have the correct im-
follows as PAR.[20] plicit nominal size of 5% but, in the case where there is
In the absence of carryover, an unbiased estimate of no carryover, has a Type I error rate that lies between
the treatment effect is provided by weighting the two 7% and 9.5%.
cell means from the AB sequence with weights 1/2 and The reason is the extremely high correlation between
 1/2 and the two cell means from the BA sequence PAR and CARRY. In fact, if anything, Freeman’s paper
with weights  1/2 and 1/2 and then summing. In the understates the problem with the two-stage approach
discussion that follows, this particular contrast will be because the pretest is either irrelevant or the conditional
referred to as CROS. In the presence of carryover, how- Type I error rate of PAR lies between 25% and 50%.[1,20]
ever, the bias in CROS is  l/2, where l is the diffe- In fact, although general medical statistics textbooks
rence between the carryover from A into B and from B continue to recommend this procedure, none of three
into A. monographs devoted to the crossover does so[1,6,9] and its
An estimate of the semi carryover effect, l/2, is given use in drug development and regulation appears to be
by using weights 1/2 for both cell means in the first being abandoned.
sequence and  1/2 for both cell means in the second It is possible to adjust the two-stage procedure in a
sequence. In other words, it corresponds to the difference number of ways, for example by carrying out the test
between the mean response over both periods in the AB using PAR at a lower nominal level: Performing this test
sequence and the corresponding mean response in the at the 0.5% rather than the 5% level deals with the
BA sequence. Like PAR, this estimate, which will be problem, for example.[22,23] However, under such circum-
referred to as CARRY below, is a between-patient esti- stances any power advantages in the face of carryover
mate, and so to base analyses upon it, a random patient disappear when compared to the simpler strategy of
effect must be assumed. It is unbiased for carryover always using CROS.
 260 Crossover Design

It seems plausible that the problems with the two-stage discussed above, the former corresponds roughly to the
procedure in the case of the AB/BA designs are likely to CROS analysis and the latter to the PAR analysis, but for
be present with strategies involving pretesting for more the fact that some variance information is recovered from
complex designs. Either the carryover effect will be or- the second period.[32] However, prior odds for these cases
thogonal to the treatment effect in which case the only are elicited and these can be updated to produce posterior
loss involved in adjusting for carryover would be residual odds via a Bayes factor. These can be used to mix over the
degrees of freedom, or the effects are not orthogonal in two models and hence produce an integrated posterior
which case similar problems to that with the AB/BA distribution for the treatment effect.[26]
design may occur. In an investigation of the pretesting
strategy, Abeyesekera and Curnow[24] came to the
conclusion that always adjusting was the best strategy. OTHER OUTCOMES
They implicitly assumed, however, that the simple
carryover model would be correct. In a more general The discussion above has been in terms of the general
investigation allowing for the possibility of an alterna- linear model. This has relatively more importance for the
tive, steady-state model, whereby carryover from a crossover trial than for parallel group trial as the former is
treatment into itself is zero[14,15] and also allowing for restricted in use to chronic diseases for which continuous
any arbitrary mixture of steady-state and simple carryover outcome measurements are common. It may be useful,
Downloaded By: [University of Alberta] At: 07:19 7 January 2009

model, Senn and Lambrou[25] came to the opposite however, to supplement such analyses with nonparametric
conclusion: that never adjusting was the best policy. techniques. A simple approach for the AB/BA design uses
The field remains controversial and the best advice to the the Wlicoxon– Mann – Whitney test.[33] A good review of
trialist seems to be not to rely on statistical modeling to approaches in more complex cases is given by Tudor and
eliminate carryover. Koch.[34]
Binary outcomes present particular problems in cross-
over trials because of the difficulty in handling depend-
BAYESIAN APPROACHES ence of observations on the same subject. A useful review
of various approaches is given by Kenward and Jones[35]
One interpretation of the difficulties that the above ap- who themselves have developed an approach based on log
proaches to carryover have encountered is that they are linear models. The matter is also treated extensively in
too extreme. For example, in the case of the AB/BA their book.[6] An alternative approach, particularly useful
design the choice is between an unbiased inefficient es- in the more general case where ordered categorical
timator PAR or an efficient but potentially biased es- variables are involved, is that of Ezzet and Whitehead.[36]
timator CROS. The former estimator would correspond They generalize the proportional odds model to the
roughly to a Bayesian analysis in which uninformative crossover context by allowing the true log-odds to have a
priors were used for both treatment and carryover effects. normal distribution over all patients. It is relatively simple
The latter would apply in a Bayesian analysis in which a to implement the binary version of this[37] using proc
completely informative prior was used for carryover, nlmixed1 of version 8 of SAS1.
assigning it the value 0 with probability 1. (The two-stage Although survival in the classic sense is not relevant
procedure is, of course, completely incoherent in a for crossover trials, certain outcome measures may be
Bayesian sense in that a choice is made between these appropriately analyzed using survival approaches. One
two extremes on the basis of an amount of information example is that of exercise tolerance tests in angina.
that is quite inadequate to the task in hand.[26]) A general France et al.[38] show how such data may be analyzed by
advantage of Bayesian approaches, however, is that establishing patient ‘‘preferences:’’ if a patient ‘‘sur-
compromise positions are possible or, to reinterpret in a vived’’ longer under A than B then A is said to be
frequentist framework, that bias-variance trade-offs can ‘‘preferred’’ for that patient. A tie occurs where both
be achieved naturally.[13] measurements are censored. An alternative approach is
A series of papers by Grieve starting 1985 has given by Feingold and Gillespie.[39]
developed a Bayesian approach to analyzing crossover
trials.[26–30] (An earlier paper by Selwyn et al. considered
Bayesian approaches to bioequivalence.[31]) The general FURTHER READING
difficulty is that of handling the prior for carryover
appropriately. Grieve’s approach establishes the appro- The books by Jones and Kenward,[6] Ratkowsky et al.,[9]
priate posterior distribution under the assumption that and Senn[1] all give very different perspectives on the
carryover is zero and also making no such assumption. As crossover trial: The last of these deals most closely with
 Crossover Design 261

the concerns of the biopharmaceutical statistician. The freedom, the carryover problem and its dual. Stat. Med.
paper by Hills and Armitage[19] is an excellent introduc- 1991, 10, 1361 – 1374.
tion to the AB/BA design although, for reasons explained
above, its advice as regards carryover is not sound. The
9. Ratkowsky, D.A.; Evans, M.A.; Alldredge, J.R. Cross-over
Experiments: Design, Analysis and Application. Statistics,
C
Textbooks, and Monographs; Marcel Dekker: New York,
encyclopedia articles by Kenward and Jones[40] and
1993; Vol. 135.
Senn[41] may also be useful, as may be a web-based 10. Kershner, R.P.; Federer, W.T. Two-treatment cross-over
tutorial.[42] A special issue of Statistical Methods in designs for estimating a variety of effects. J. Am. Stat.
Medical Research was devoted to the crossover trial and Assoc. 1981, 76, 612 – 619.
has articles on the AB/BA design,[16] multiperiod de- 11. Lasker, E.M.; Meisner, M.; Kushner, H.B. Optimal cross-
signs,[17] binary data,[35] nonparametric,[34] and Baye- over designs in the presence of carryover effects. Bio-
sian[30] approaches. metrics 1983, 39, 1089 – 1091.
12. Balaam, L.N. A two period design with t2 experimental
units. Biometrics 1968, 24, 61 – 73.
13. Carlin, B.P.; Louis, T.A. Bayes and Empirical Bayes
CONCLUSION Methods for Data Analysis; Chapman and Hall: London,
1996.
Crossover trials are not suitable for all indications 14. Fleiss, J.L. Letter to the editor. Biometrics 1986, 42, 449 –
and even in indications in which they can sometimes 450.
Downloaded By: [University of Alberta] At: 07:19 7 January 2009

safely be employed, will not be suitable for investigat- 15. Fleiss, J.L. A critique of recent research on the two-treat-
ing all treatments. Their use is best restricted to early drug ment cross-over design. Control. Clin. Trials 1989, 10,
development or for answering specific scientific ques- 237 – 243.
tions. They are almost never suitable for phase III thera- 16. Senn, S.J. Is the ‘simple carry-over’ model useful? [pub-
lished erratum appears in Statistics in Medicine 1992 Sep
peutic studies. Nevertheless, it would be a grave mistake
15;11(12):1619]. Stat. Med. 1992, 11, 715 – 726.
to conclude, therefore, that they are not useful. Their
17. Matthews, J.N. Multi-period crossover trials. Stat. Methods
extreme efficiency makes them very attractive for early Med. Res. 1994, 3, 383 – 405.
phase drug development including pharmacokinetics, 18. Grizzle, J.E. The two-period change over design and its use
bioequivalence, dose-finding, single-dose pharmacody- in clinical trials. Biometrics 1965, 21, 467 – 480. (Cor-
namic studies and proof of concept. They are also par- rigenda JE Grizzle Biometrics 30: 727, 1965 and AP
ticularly suited for investigating treatment by subject Grieve Biometrics 38: 517, 1982).
interaction and as such may have an important role to play 19. Hills, M.; Armitage, P. The two-period cross-over clinical
in the developing field of pharmacogenomics.[43,44] trial. Br. J. Clin. Pharm. 1979, 8, 7 – 20.
20. Senn, S.J. The AB/BA crossover: Past, present and future?
Stat. Methods Med. Res. 1994, 3, 303 – 324.
21. Freeman, P. The performance of the two-stage analysis
REFERENCES of two-treatment, two-period crossover trials. Stat. Med.
1989, 8, 1421 – 1432.
1. Senn, S.J. Cross-over Trials in Clinical Research; John 22. Senn, S.J. The case for cross-over trials in phase III [letter;
Wiley: Chichester, 1993, pp. 3, 8. (Second edition 2002). comment]. Stat. Med. 1997, 16, 2021 – 2022.
2. Graff-Lonnevig, V.; Browaldh, L. Twelve hours bron- 23. Wang, S.J.; Hung, H.M. Use of two-stage test statistic
chodilating effect of inhaled formoterol in children with in the two-period crossover trials. Biometrics 1997, 53,
asthma: A double-blind cross-over study versus salbuta- 1081 – 1091.
mol. Clin. Exp. Allergy 1990, 20, 429 – 432. 24. Abeyasekera, S.; Curnow, R.N. The desirability of ad-
3. Dahlof, C.; Bjorkman, R. Diclofenac-K (50 and 100 mg) justing for residual effects in a crossover design. Biomet-
and placebo in the acute treatment of migraine. Cepha- rics 1984, 40, 1071 – 1078.
lalgia 1993, 13, 117 – 123. 25. Senn, S.J.; Lambrou, D. Robust and realistic approaches to
4. Chow, S.C.; Liu, J.P. Design and Analysis of Bioavail- carry-over. Stat. Med. 1998, 17, 2849 – 2864.
ability and Bioequivalence Studies, 2nd Ed.; Marcel 26. Grieve, A.P.; Senn, S.J. Estimating treatment effects in
Dekker: New York, 2000. clinical crossover trials. J. Biopharm. Stat. 1998, 8, 191 –
5. Senn, S.J.; Lillienthal, J.; Patalano, F.; Till, D. Cross-over 233.
Trials; Vollmar, J., Hothorn, L.A., Eds.; Fischer: Stuttgart, 27. Grieve, A.P. A Bayesian analysis of the two-period cross-
1997; 3 – 26. over design for clinical trials. Biometrics 1985, 41, 979 –
6. Jones, B.; Kenward, M.G. Design and Analysis of Cross- 990.
Over Trials; Chapman and Hall: London, 1989. 28. Grieve, A.P. Statistical Methodology in the Pharmaceut-
7. Chi, E.M. Recovery of inter-block information in cross- ical Sciences; Berry, D.A., Ed.; Marcel Dekker: New York,
over trials. Stat. Med. 1991, 10, 1115 – 1122. 1989.
8. Senn, S.J.; Hildebrand, H. Crossover trials, degrees of 29. Grieve, A.P. Extending a Bayesian analysis of the
 262 Crossover Design

two-period crossover to allow for baseline measurements. binary data from cross-over trials. Appl. Stat. 1991, 41,
Stat. Med. 1994, 13, 905 – 929. 117 – 126.
30. Grieve, A.P. Bayesian analyses of two-treatment crossover 38. France, L.A.; Lewis, J.A.; Kay, R. The analysis of failure
studies. Stat. Methods Med. Res. 1994, 3, 407 – 429. time data in cross-over studies. Stat. Med. 1991, 10, 1099 –
31. Selwyn, M.R.; Dempster, A.R.; Hall, N.R. A Bayesian 1161.
approach to bioequivalence for the 2  2 changeover de- 39. Feingold, M.; Gillespie, B.W. Cross-over trials with cen-
sign. Biometrics 1981, 40, 1103 – 1108. sored data. Stat. Med. 1996, 15, 953 – 967.
32. Senn, S.J. Consensus and controversy in pharmaceutical 40. Kenward, M.G.; Jones, B. Encyclopedia of Statistics,
statistics (with discussion). The Statistician 2000, 49, 135 – Update; Kotz, S., Read, C.B., Banks, D.L., Eds.; Wiley:
176. New York, 1998; Vol. 2, 167 – 175.
33. Koch, G.G. The use of non-parametric methods in the 41. Senn, S.J. Cross-over Trials. In Encyclopedia in Biostatis-
statistical analysis of the two-period change-over design. tics; Armitage, P., Colton, T., Eds.; Wiley: New York,
Biometrics 1972, 28, 577 – 584. 1998; Vol. 2, 1033 – 1049.
34. Tudor, G.; Koch, G.G. Review of nonparametric methods 42. Senn, S.J. Symptom Research Methods and Opportunities;
for the analysis of crossover studies. Stat. Methods Med. Max, M.B., Lynn, J., Eds.; National Institute of Dental
Res. 1994, 3, 345 – 381. and Craniofacial Research: Bethesda, 2001. http://
35. Kenward, M.G.; Jones, B. The analysis of binary and [Link]/chapter_6/[Link].
categorical data from crossover trials. Stat. Methods Med. 43. Kalow, W.; Tang, B.K.; Endrenyi, L. Hypothesis: Com-
Res. 1994, 3, 325 – 344. parisons of inter- and intra-individual variations can sub-
Downloaded By: [University of Alberta] At: 07:19 7 January 2009

36. Ezzet, F.; Whitehead, J. A random effects model for stitute for twin studies in drug research. Pharmacogenetics
ordinal response from a cross-over trial. Stat. Med. 1991, 1998, 8, 283 – 289.
10, 901 – 907. 44. Senn, S.J. Individual therapy: New dawn of false dawn
37. Ezzet, F.; Whitehead, J. A random effects model for Drug. Inf. J. 2001, 35, 1479 – 1494.