Received: 23 May 2016 Revised: 8 December 2016 Accepted: 6 January 2017

DOI: 10.1002/jrsm.1236

ORIGINAL ARTICLE

Meta‐analysis combining parallel and cross‐over trials with

random effects

François Curtin1,2,3

1
Division of Clinical Pharmacology and
Toxicology, Geneva University Hospital, Meta‐analysis can necessitate the combination of parallel and cross‐over trial
Geneva, Switzerland designs. Because of the differences in the trial designs and potential biases notably
2
Research Center for Statistics, Geneva associated with the crossover trials, one often combines trials of the same designs
School of Economics and Management,
University of Geneva, Geneva, Switzerland
only, which decreases the power of the meta‐analysis. To combine results of clinical
3
GeNeuro SA, Geneva, Switzerland trials from parallel and cross‐over designs, we extend the method proposed in an
accompanying study to account for random effects. We propose here a hierarchical
Correspondence
mixed model allowing the combination of the 2 types of trial designs and accounting
François Curtin, Division of Clinical
Pharmacology and Toxicology, Geneva for additional covariates where random effects can be introduced to account for
University Hospital, Ave Foretaille 20, CH‐ heterogeneity in trial, treatment effect, and interactions. We introduce a multilevel
1292 Chambesy, Geneva, Switzerland.
Email: [email protected]
model and a Bayesian hierarchical model for combined trial design meta‐analysis.
The analysis of the models by restricted iterative generalised least square and Monte
Carlo Markov Chain is presented. Methods are compared in a combined design
meta‐analysis model on salt reduction. Both models and their respective advantages
in the perspective of meta‐analysis are discussed. However, the access to the trial
data, in particular sequence and period data in cross‐over trials, remains a major
limitation to the meta‐analytic combination of trial designs.

K EY WO R D S
Bayesian, cross‐over, parallel, random effects, trial design

1 | INTRODUCTION estimate as well as on the inference when one combines these

2 trial designs.1 Notably the metrics used for cross‐over trials,
Parallel and cross‐over design trials can be used in the devel- for example, the conditional odds ratio, cannot be directly
opment process of drugs and often meta‐analyses have to combined with parallel trial odds ratio.2,3 Another issue
combine results coming from trials of the 2 designs, on the related to including cross‐over trials into a meta‐analysis is
basis of the fundamental assumption that these trials evaluate the carry‐over problem, which is often not properly addressed
all the same treatment effect. in meta‐analysis.4 Therefore, the inclusion of cross‐over trials
Parallel trials provide outcomes at the end of the follow‐ is often performed in an unsatisfactory way in meta‐analyses,
up period, and their analysis seeks the estimation of the for example by using only the first‐period cross‐over data or
difference between the marginal mean responses obtained in as if results came from a parallel trial.5
the different treatment groups. Cross‐over trials on the con- Expanding a first approach to combine trials belonging to
trary allow the evaluation of change within subjects as well parallel and cross‐over designs according to the classical
between subgroups of subjects. The differences between par- weighted average model,1,2 a regression technique based on
allel and cross‐over trials in the statistical analysis of the the generalised estimating equation (GEE) has been proposed
treatment effect and its variance may have direct conse- in an accompanying study.6 This approach models the out-
quences on the calculation of a meta‐analysis treatment come in each treatment group, ie, arms of parallel trials or

Res Syn Meth. 2017;1–12. wileyonlinelibrary.com/journal/jrsm Copyright © 2017 John Wiley & Sons, Ltd. 1
2 CURTIN

periods with a given treatment within sequences of cross‐over indexed by j = 1 to mi and the treatment groups are indexed
trials. Generalised estimating equation permits the combina- by q = 1 to ni, ni being the length of the sequence for trial i
tion of different trial designs allowing for the introduction (one assumes that the sequences have the same length for a
of the between–cross‐over period correlation as well as of given trial). Each trial indexed i = 1 to k provides a matrix
covariates according to a fixed effects model, which can con- YTi = (YTi1,…, YTiji) and for sequence j of trial i, there is a vector
trol at least for part of the heterogeneity of effects,7 notably
T
Yij = yij1 ; :::; yijni , where yijq is the average outcome of
that resulting from the carry‐over bias observed in cross‐over
treatment group q. For a parallel trial, the number of
trials.4
sequences m is equal to the number of arms of the trial and
Among clinical trials included in a meta‐analysis, there
the sequence length n = 1 by definition; in cross‐over trials,
are variations in protocols, methods, assessment of endpoints,
one has for example in a 2‐period 2‐treatment AB/BA
patient populations, or trial designs. These variations may
cross‐over trial m = 2 and n = 2 (see also Curtin6). Sequences
yield differences in the treatment effect estimates.8,9 If these
are considered as independent clusters of observations even if
differences cannot be adjusted or controlled, the observed
they belong to the same trial.
heterogeneity of results suggests considering the treatment
Assuming a normally distributed Yi and an identity link,
effect not as the result of a fixed effect similar for all trials
Yi is related to p predictors as follows:
but rather as a random treatment variable with its own distri-
bution10, chapter 9.5.4 this corresponds to the random effects Yi ¼ Xi βi þ Εi; (1)
meta‐analysis.11 The use of different trial designs may well
be a reflection of differences in the clinical settings compati- where XTi is defined as (XTil … XTimi ) where the ni × p design
ble with the heterogeneity of treatment effect. Different matrix Xij is decomposed into ni treatment‐group‐specific
frequentist meta‐analysis approaches to random effects have row vectors xijq. Each xijq has elements corresponding to
been discussed using generalised least squares or restricted treatment group‐specific variables and trial‐specific vari-
maximum likelihood, nonparametric maximum likeli- ables. The p‐dimensional random vector βi can be of different
hood,12–18 or the Bayesian model (for example, in previous dimensions as the parameters contained in the vector βi
studies19–23), which requires the definition of prior distribu- include trial‐specific parameters such as the treatment effect,
tions for parameters,24 the choice of which in particular for a trial effect, and any trial‐specific parameters of interest
variance estimation is not always straightforward.25,26 (dose of the drug, characteristic of patients…); these param-
In this paper, one proposes to extend the meta‐analysis eters will be random in this model, although certain may be
model combining parallel and cross‐over trials6 by including considered as fixed as for example the trial effect parame-
random effects for continuous outcomes following a Normal ter.36 The matrix Εi is the sampling error caused by estimat-
distribution either by a frequentist method on the basis of iter- ing the outcome through a study with a finite size and each
ative generalised least square (IGLS)27,28 or by a Bayesian of its element eijq is distributed as follows:
hierarchical linear model (BHLM) (in previous studies29–33

eijq eN 0; υijq
for example), evaluated through Markov Chain Monte Carlo
(MCMC) methods using a Gibbs sampler.34 The random and, according to the models described in Section 2.2 below,
effects models are introduced in Section 2. Iterative general- either υijq = vijq, where vijq is the sample variance of treat-
ised least square and Bayesian methods are presented in ment group q in sequence j of trial i, usually considered as
Section 3. The 2 approaches, as well as the DerSimonian known and obtained from trial results (“semirandom effects”
and Laird method, are compared in a meta‐analysis on anti- model) or υijq = σ2 to be estimated by the IGLS/Bayesian
hypertensive treatment are introduced in Section 4. These methods described in Section 3 (“fully random effects”
results are discussed in Section 5. model).
A simple bivariable model is detailed where the outcome
2 | HIERARCHICAL AND of treatment group ijq is a function of the intercept and treat-
STATI ST I CAL MO D E L S ment effect xijq and where both the intercept (heterogeneity
due to trial effect) and the treatment‐by‐trial effect are
2.1 | The model random:

We define a sequence as a set of treatment groups, ie, treat- yijq ¼ β1i þ β2i xijq þ eijq : (2)
ments that a group of patients will receive in a successive
manner in the different periods of a (cross‐over) trial.35 To This corresponds to the meta‐analysis models that Senn37
parallel the notation presented in the accompanying paper,6 calls “full random model.” It must be noted that with this
in the meta‐analysis, each trial is indexed by i = 1 to k; each model, treatment groups of parallel trials or sequences of
trial has a given number of sequences of treatment groups cross‐over trials are not independent as the random variation
CURTIN 3

of the intercept (trial effect) is the same for all treatment coefficient within sequence j, (it is assumed to be known, if
groups/sequences of trial i. not, see for example38 for imputation) Vi is
β1i and β2i are decomposed into fixed parameters β1 and
β2 and random parameters γ1i and γ2i, thus Equation 2 can 0 1=2 1=2
1
vi11 ri1 vi11 vi12 0 0
be written: B C
B C
B ri1 v1=2 1=2
i11 vi12 vi12 0 0 C
yijq ¼ ðβ1 þ γ 1i Þ þ ðβ2 þ γ 2i Þxijq þ eijq : B C:
B C
B 0 0 vi21
1=2 1=2
ri2 vi21 vi22 C
@ A
0 0 1=2 1=2
ri2 vi21 vi22 vi22
Defining γTi = (γ1i, γ2i) as the vector of random compo-
nents, one has
For each trial, one defines a variance matrix Ui for the
E ðγ i Þ ¼ 0; random effects meta‐analysis model:
!
τ21 τ12
varðγ i Þ ¼ varðβi Þ ¼ Ω ¼ ;
τ12 τ22 Ui ¼ Vi þ Τ i : (3)

where τ21 and τ22 are the between‐study variances for the inter-
cept and treatment effect respectively, τ12 is their covariance Therefore, for each trial i, Yi follows the distribution:
and for the moment one assumes τ12 = 0. One defines a
matrix Τi for the interstudy variance components correspond- Yi eNðXi β; Ui Þ
ing to the trial design, Τi is given by

Ti ¼ Xi ΩXTi :
2.2 | Hierarchical linear model for combined
design meta‐analysis
For example, in an AB/BA cross‐over trial where A is a
placebo and B is an active treatment with the bivariable Three different meta‐analysis approaches can be defined: (1)
model (Equation 2), one has the design matrices for a fixed effects method assuming the variance of each trial is a
sequences 1 and 2: known quantity and regression coefficients are fixed, this
corresponds to the model developed in the accompanying
   
1 0 1 1 study,6 (2) a random effects method, referred as the
X i1 ¼ and X i2 ¼
1 1 1 0 “semirandom effects” method, which assumes the sample
variance of the trials are known and regression coefficients
corresponding to the design matrix X Ti : are random, (3) a random effects method, referred as the
“fully random effects“ method, which relaxes the assumption
! of known trial variance. These approaches are presented as
1 1 1 1
XTi ¼ hierarchical structures in Table 1: Level I corresponds to the
0 1 1 0 trial observations and parameters, level II to the meta‐analytic
parameters, and level III is the level of prior information for
and with Ω defined as above and τ12 = 0, Τi is the Bayesian model. Taking the same notation as in Section
2.1, for the fixed effects method, at level I, Yi follows a
0 1 Normal distribution whose expectation is the vector θi, and
τ21 τ21 τ21 τ21 whose covariance matrix is Vi. θi is a linear combination of
B C
B τ2 τ21 þ τ22 τ21 þ τ22 τ21 C Xi and, assuming p different predictors, the p‐dimensional
B 1 C
B 2 C: vector βi. At level II, βi follows a Normal distribution whose
B τ1 τ21 þ τ22 τ21 þ τ22 2 C
τ1 A
@ expectation and variance parameters are respectively μ and
τ21 τ21 τ21 τ21 the p × p null matrix 0. At level III, μ has a normal prior with
expectation a and variance B.
For the semirandom effects method, level I distribution is
One defines a (nimi) × (nimi) matrix Vi a block diagonal similar to that of the fixed effects method. At level II, βi is
matrix equal to diagðVi1 ; …Vimi Þ, where Vij is the sample var- normal with expectation and variance parameters μ and
iance/covariance matrix for trial i, sequence j. For an AB/BA p × p matrix Ω, respectively. At level III in the Bayesian
cross‐over trial with rij being the between‐period correlation model, the hyperprior for Ω follows a p‐dimensional Wishart
4 CURTIN

TABLE 1 Hierarchical diagram for fixed and random effects meta‐analysis

Fixed Effects Semirandom Effects Fully Random Effects

I. Observed data Yi Yi|θi ~ N(θi, Vi) Yi|θi ~ N(θi, Vi) Yi|θi, ΣP, ΣX ~ N(θi, (1‐IX)ΣP + IXΣX)

II. Unknown Param. θi θi = Xiβi θi = X i β i θi = Xiβi

βi|μ ~ N(μ, 0) βi|μ, Ω ~ N(μ, Ω) βi|μ, Ω ~ N(μ, Ω)

III. Hyperpriors* μ|a, B ~ N(a, B) μ|a, B ~ N(a, B) μ|a, B ~ N(a, B)

Ω−1 |C, p ~ Wp(C, p) Ω−1 |C, p ~ Wp(C, p)

*Level III of hyperprior exists only in the Bayesian model.

Yi is the observed average outcomes of treatment groups in trial i,
θi is the unknown results of treatment groups in trial i,
Vi is the matrix of sample variance/covariance observed in trial i,
βi is the vector of random coefficients,
μ is the expected value of the regression coefficients,
0 is the null matrix,
Ω is the variance/covariance matrix of the random coefficients, !
σ2 0
ΣP is the variance matrix of parallel trials, for example, in a 2‐arm trial: , !
0 σ2 σ2 ρσ 2
ΣX is the variance/covariance matrix of cross‐over trials, for example, in a 1‐sequence 2‐period trial: , ρ is the between period correlation,
ρσ 2 σ2
Ix is a scalar indicator function =1 if trial i is cross‐over, and =0 if it is parallel,
a, B are the expectation and variance parameters of the μ hyperprior, and
C, p are the parameter matrix and degrees of freedom for the Ω−1 hyperprior.

distribution parametrised by p degrees of freedom and a p × p are presented here. For the variance estimation, a block diag-
matrix C, estimating the prior variance of μ. onal matrix Τe of dimension (kmini) × (kmini) is obtained
In the fully random effects method, at level I, the bi
where, for each trial i, the mini x mini variance submatrix T
unknown treatment group variance and cross‐over covariance is estimated as the product of the residuals:
are represented by a matrix ΣX and a diagonal matrix ΣP for
cross‐over and parallel trials, respectively. The fully random bi ¼ ðYi −Xi βÞðYi −Xi βÞT :
T
effects method corresponds to formula 1, and the variance
of eijq is equal to σ2, which is unknown. Levels II and III As in practice, the estimate b
β is used instead of β, for each
are similar to those of the semirandom effects method. The trial, one obtains
fully random effects method is only computed with the   T
Bayesian approach. bi  ¼ Yi −Xi b
T β Yi −Xi b β :
It should be noted that cross‐over trial treatment groups
could be considered as a level I, which is nested in a level bi  is a biased estimate of T
T bi .39 To obtain an unbiased
II corresponding to sequence, itself nested in a level III variance estimate, it is necessary to use the following equation:
corresponding to trial. This would introduce random effects   T  
specific for periods and sequences. Generally, there is a lack bþ b Yi −Xi b
β þ Xi var b
Ti ¼ Yi −Xi β β Xi T ;
of information on sequences and periods in published cross‐
over trials and inclusion of these levels of information is
where var(bβ) is the covariance matrix of the fixed coefficient
hardly possible practically; and for simplicity, these addi-
estimates. This correction is called restricted IGLS (RIGLS).39
tional levels would not be considered any further here.
In parameter estimation, the sample variance Vi is already taken
into account in the estimation of bβ and must be omitted during
3 | RANDOM EFFECTS META‐ the estimation of the variance parameters. Vi is an offset
ANALYSIS PARAMETER subtracted from T bi ) at each variance parameter estima-
bþ (or T
i
ESTIMATION tion. For each trial, the variance matrix becomes:
 
3.1 | Evaluation of random parameters by T bi −Vi or T
bi  ¼ T bþþ ¼ T
b þ −Vi :
i i
IGLS
The IGLS estimation of random parameters follows the The next steps of the evaluation algorithm follow the
method by Goldstein.27 The steps relevant to meta‐analysis usual IGLS algorithm.27 Evaluation algorithms according to
CURTIN 5

Goldstein27 with the above modifications were written in The full conditional distribution of μ given e
β, Ω, a, and
S‐Plus. B is:
  

p μe β; Ω; a; B ∝ exp
3.2 | Evaluation of random parameters by  k


1 T −1 T −1
Bayesian estimation − ðμ−aÞ B ðμ−aÞ þ ∑ ðβi −μÞ Ω ðβi −μÞ :
2 i¼1
The estimation of the Bayesian parameters uses a Monte
Carlo Markov chain (MCMC) simulation on the basis of After algebraic development, μ is normally distributed as
Gibbs sampling. To estimate parameters, the full conditional follows:
distribution of the parameters is necessary.40 It is detailed
below for the semirandom effects method. For the full condi- 
−1

−1 
N B−1 þ kΩ−1 B−1 a þ kΩ−1 β ; B−1 þ kΩ−1 ;
tional distributions for βi and for the kp‐dimensional vector e
β

T T 
e
defined as β ¼ β1 ; β2 ; :::; βk to derived, it is necessary to
T T
where β ¼ 1k ∑ki¼1 βi .
redefine distributions of Table 1 for multivariate outcomes.
With Yi, Xi, Vi, and Ω defined as before, in a p‐covariate The full conditional distribution for Ω−1 is obtained by
model, the following variables are defined: YT = combination of the inverse Wishart prior distribution for Ω

T T  with the normal likelihood for β. This likelihood is propor-
Y1 ; Y2 ; :::; YTk , a (knimi) × (kp) matrix X = diag(X1,
tional to
X2,..., Xk), a (knimi) × (knimi) matrix V = diag(V1, V2,...,
Vk), a pk × p matrix A = Ipxp ⊗ 1k, where Ipxp is a p × p  
1 k
identity matrix and 1k a k‐dimensional vector of 1 and a jΩj−k=2 exp − ∑ ðβi −μÞT Ω−1 ðβi −μÞ : (4)
pk × pk covariance matrix Θ = Ω ⊗ Ikxk., where Ikxk is a 2 i¼1
k × k identity matrix and ⊗ is the Kronecker multiplicator.
The levels I and II distributions are rewritten,
respectively: By writing S(μ) = ∑ki¼1 ðβi −βÞT ðβi −βÞ , Equation 4 is
   rewritten:

Ye β; VeN Xe β; V ;
 
1
jΩj−k=2 exp − trΩ−1 SðμÞ ; (5)
2
e
βjμ; ΘeN ðAμ; ΘÞ:
where trM is the trace of M.
The full conditional parameter distributions are deter- The likelihood (Equation 5) is combined with the conju-
mined through the conjugacy property of the different level gated inverse‐Wishart prior given on Table 1 to give the full
conditional distribution for Ω−1:
distributions. The full conditional distribution of e
β given X,
V, Y, Θ, and μ is written:    
−12pþ1 1 −1 −k=2 1 −1
  jΩj exp − trΩ C xjΩj exp − trΩ SðμÞ
2  2
p e
βjX; V; Y;Θ; μ ∝ exp
 1

 ¼ jΩj−2ðpþkÞþ1 exp − trΩ−1 ðSðμÞ þ CÞ :
1

1  e
T
−1

e
 
e
T 
−1 e 2
− Y−Xβ V Y−Xβ þ β−Aμ Θ β−Aμ
2

Matrix Ω−1 is therefore distributed according to the

After developing and completing the quadratic forms in e
β,
following Wishart distribution:
e
one obtains the conditional distribution of β defined as
follows:  k
 
W ∑ ðβi −μÞT ðβi −μÞ þ C ; p þ k :

−1
T −1 
−1  i¼1
N XT V−1 X þ Θ−1 X V Y þ Θ−1 Aμ ; XT V−1 X þ Θ−1 :

Because of the symmetry of the matrices, a similar In the fully random effects method, the full conditional
distribution for each βi is easily obtained: distributions for ΣP and ΣX are obtained in a similar way
but from likelihoods determined from parallel or cross‐over


T −1 
  trials exclusively. Evaluation programs were written in bayes-
−1 −1 −1
N XTi V−1
i Xi þ Ω Xi Vi Yi þ Ω−1 μ ; XTi V−1
i Xi þ Ω :
ian inference using gibbs sampling (BUGS).41
6 CURTIN

4 | EXAMPLES 4.2 | Meta‐analysis models

This meta‐analysis will be analysed with models presented in
4.1 | Example of meta‐analysis on salt
Table 3. Fully random effects models (models 1′ and 3′) are
reduction
only calculated with the Bayesian method. Models 1 and 1′
The combined‐design meta‐analysis of trials studying the included only an intercept and the treatment effect. In models
effect of a reduction of sodium intake on blood pressure42 2, 3, and 3′, a dummy covariate reporting hypertension was
includes parallel and cross‐over trials. This meta‐analysis used: in a given trial, if the average diastolic blood pressure
included 23 studies among which 4 trials were separated into in placebo and treatment groups was ≥95 mmHg, the trial
2 phases. Details are in Table 2. population was considered as hypertensive. This was the case

TABLE 2 Randomized trials of sodium reduction

Design Sample Size DBP Control Grp DBP Treatmt Grp Treatmt Variance Variance Corr. Age

Contrl Trtmt Mean SD Mean SD Effect Effect (cor = 0.5)

Morgan43 Parallel 31 31 99 8.4* 92 8.4* ‐7.0 4.6 . . >50

Costa 44
Parallel 20 21 83.9 11.0 78.1 9.0 ‐5.8 9.9 . . 24

Morgan 45
male Parallel 6 6 94.0 7.1* 87.0 7.1* ‐7.0 16.8 . . 39

Morgan 45
fem Parallel 6 6 92.0 5.7* 89.0 5.7* ‐3.0 10.8 . . 39

Puska46 Parallel 38 34 86.9 9.2 86.5 10.5 ‐0.4 5.4 . . 40

Silman47 Parallel 10 15 86.5 10.8* 80.9 10.8* ‐6.6 19.5 . . 56

Erwteman 48
Parallel 44 50 94.4 12.0 92.9 10.4 ‐1.5 5.5 . . 46

Fagerberg 49
Parallel 15 15 94.6 7.4 90.5 9.3 ‐4.1 9.4 . . 51
50
Maxwell Parallel 18 12 78 7.6 80 7.6 2.0 8.0* . . 48

Chalmers 51
Parallel 52 48 93.3 4.5** 89.5 4.5* ‐3.8 0.81 . . 52

Logan 52
Parallel 37 38 91.7 5.2 91.5 5.2 ‐0.2 1.4 . . 47

ANHMRC a 53
Parallel 53 50 94.6 6.6 91.4 4.9 ‐3.2 1.3 . . 58

Dodson54 Parallel 17 17 90.4 5.7 87.6 10.5 ‐2.8 8.4 . . 61

55
HPTRG Parallel 174 177 80.0 7.5* 80.2 7.5* 0.2 0.64 . . 39
56
Parijs Crossvr 15 15 112.3 15.1 115.5 12.5 3.3 18.7 13.0 0.22 41

Skrabal 57
Crossvr 20 20 73.1 9.8 70.1 8.5 ‐3.0 4.13* 4.2 0.5** 23

McGregor 58
Crossvr 19 19 97.0 8.7 92.0 8.7 ‐5.0 3.2 4.0 0.60 49

Watt 59
Crossvr 18 18 82.6 3.4* 82.3 3.4* ‐0.3 0.64 0.64 0.5** 52

Cooper 60
Crossvr 113 113 60.7 10.0 59.3 11.0 ‐1.4 1.1 0.99 0.44 16

Richards 61
Crossvr 12 12 92.4 12.1 90.6 12.5 ‐1.8 12.6* 12.6 0.5** 36

Watt62 LL Crossvr 31 31 63.6 9.5 65.0 9.5 1.4 0.81 2.9 0.86 23
62
Watt HH Crossvr 35 35 63.3 11.2 64.5 11.2 1.2 0.86 3.6 0.88 22

Grobbee 63
Crossvr 40 40 73.3 9.5 72.5 9.5 ‐0.8 2.3* 2.3 0.5** 24

ANHMRC b 64
Crossvr 88 88 94 3.8* 92 3.8* ‐2.0 0.16 0.16 0.5** 58

Dodson 65
Crossvr 9 9 92.4 10.0 87.3 6.7 ‐5.1 8.6 8.7 0.47 62

McGregor 66
Crossvr 20 20 100.0 8.9 95.0 8.9 ‐5.0 2.6 4.0 0.67 57

Myers 67
Crossvr 172 172 77.2 11.8 75.3 11.8 ‐1.9 0.81* 0.81 0.5** 37

Abbreviations: DBP, diastolic blood pressure; corr, correlation; SD, standard deviation.
LL subjects with parents in the bottom third of age specific BP distribution, HH subjects with parent in the top of their age specific BP distribution.
*value imputed,
**correlation fixed at 0.5,
.not available.
CURTIN 7

TABLE 3 Random effects meta‐analysis, models for treatment 4.4 | Semirandom and fully random effects
groups models evaluated by Bayesian method
Model 1: yijq = β1ix1ijq + β2ix2ijq + eijq In models 1 or 3 evaluated by the semirandom effects
Model 1′: yijq = β1ix1ijq + β2ix2ijq + uijq method, the elements of μ had independent vague priors N
Model 2: yijq = β1ix1ijq + β2ix2ijq + β3x3ijq + β4x2ijqx3ijq + eijq (0, 10 000), matrix Ω was diag( ω21 ; ω22 ) or diag(
Model 3: yijq = β1ix1ijq + β2ix2ijq + β3ix3ijq + β4ix2ijqx3ijq + eijq ω21 ; ω22 ; ω23 ; ω24), and ω−2 −2
1 and ω2 followed each a vague prior

Model 3′: yijq = β1ix1ijq + β2ix2ijq + β3ix3ijq + β4ix2ijqx3ijq + uijq

Gamma(0.05, 0.05). Because of the small number of observa-
tions with covariate hypertension x3i = 1 and thus the diffi-
Fixed effects (coefficients): β3 = hypertension,
culty to estimate variance for hypertension and interaction,
β4 = hypertension*treatment
an informative prior was used for ω−2 −2
3 and ω4 . Following
Random effects (coefficients): β1i = intercept, β2i = treatment,
the approach originally proposed by Smith et, al68 for compu-
β3i = hypertension, β4i = hypertension*treatment
tation of variance prior, 2 assumptions were made. Firstly, a
ei ~ N(0,Vi), plausible variance estimate is obtained if one assumes that a
ui ~ N(0, ΣX) for cross‐over trials or ui ~ N(0, ΣP) for parallel trials. difference of 10.0 units (mmHg in the present example) can
be observed between the lowest and highest effects for the
hypertension covariate: this is equivalent to a prior estimate
for variance of (5/1.96)2 = 6.51. Secondly, it is unlikely to
for the trials by Morgan et al43, Parijs et al56, McGregor et observe a difference of more than 20.0 units, equivalent to
al58, and McGregor et al66. In model 2, covariates hyperten- a limit variance of (10/1.96)2 = 26.03. For precision, the cor-
sion and interaction between treatment and hypertension responding Gamma distribution has a mean of 1/6.51 = 0.15
were fixed. Models 1′ and 3′ were fully random effects and a low 5th percentile of 1/26 = 0.04. A distribution
models. Gamma(3.0, 20.0) has a mean of 0.15 and a 95% probability
of exceeding 0.04. For the interaction effect, the same
approach68 was taken with the prior belief of a plausible
4.3 | Random effects models evaluated by
observed difference of 5.0 units, equivalent to a variance of
DerSimonian and Laird method and RIGLS
1.63, and a highest difference of 10.0 units, equivalent to a
The frequentist results obtained with the RIGLS method limit variance of 6.5. A distribution Gamma(2.5, 4.0) has a
and the classical DerSimonian and Laird method11 are in mean of 1/1.63 and a 95% probability of exceeding 0.15.
the left part of Table 4. When comparing DerSimonian For the fully random effects method, the prior for matrix
and Laird method with RIGLS model 1, the estimates of ΣP−1 was a 2‐dimensional Wishart distribution W2(R, 2),
the between study variance for treatment were very similar with R = diag(4.0, 4.0) whose elements were the average
as were the estimates of the treatment effect and of its var- treatment group variances, the prior for matrix ΣX−1 was a
iance. For RIGLS, the high variance associated with the 2‐dimensional Wishart distribution W2(S, 2), with matrix S
intercept reflected the high heterogeneity of the average whose elements on the diagonal were equal to 4.0 and those
blood pressure across trials. The introduction of the covar- off diagonal were equal to 2.0 assuming a prior between‐
iates hypertension and hypertension*treatment interaction period cross‐over correlation of 0.5 as suggested by results
in models 2 and 3 reduced the between‐study variance for in Curtin et al.1 Run‐in series of 20 000 iterations were used.
the intercept and treatment compared to model 1. For Parameter evaluation was based on series of 10 000 iterations
example, the reduction of the between‐study variance of for all programs. The convergence of the parameters were
the intercept and treatment effect was about 26% and evaluated by Geweke and Raftery and Lewis diagnostic tests
40%, respectively, when passing from RIGLS model 1 to and diagrams using CODA software.69
RIGLS model 2, suggesting that hypertension and The results are in the right part of Table 4. For
hypertension*treatment were important factors explaining semirandom effects model 1, the treatment effect was very
the heterogeneity of model 1. There were only small differ- similar to those of RIGLS, and the intercept was slightly
ences between models 2 and 3: Indeed, when the coeffi- higher. For the fully random effects model 1′, the treatment
cient for hypertension was random in model 3, its effect estimates was higher in absolute value (−2.27) than
between‐study variance remained equal to 0. Similarly for that obtained with the other methods (−1.80). The between‐
the interaction only the RIGLS method estimated a small study variance for treatment (6.04) was also largely increased
between‐study variance of 0.56. The introduction of ran- compared with that of semirandom model 2 (2.50), variance
dom variation in hypertension and interaction covariates estimates of variance coefficients are also reported. Estimates
inflated somewhat the variances of the intercept and treat- for level I variance matrices (noted ΣP and ΣX) were
ment effect when passing from model 2 to model 3. obtained: Variances were estimated between 0.74 and 0.79
 8

TABLE 4 Random effects meta‐analysis according to DerSimonian and Laird, RIGLS, and Bayesian methods

Bayes Bayes Bayes Bayes

DerSimonian IGLS RIGLS IGLS RIGLS IGLS RIGLS SemiR FullyR SemiR FullyR

and Laird Model 1 Model 1 Model 2 Model 2 Model 3 Model 3 Model 1 Model 1′ Model 3 Model 3′

Intercept . 86.35 86.36 83.54 83.55 83.55 83.56 86.37 86.67 83.48 83.74

Variance . 5.27 5.49 4.41 4.78 4.73 4.94 5.90 5.81 5.11 4.75

95%confidence . 81.85 to 90.85 81.77 to 90.95 79.42 to 87.66 79.26 to 87.84 79.29 to 87.81 79.20 to 87.92 81.61 to 91.13 81.95 to 91.39 79.05 to 87.91 79.47 to 88.01
int.

Between‐study . 139.23 144.95 98.98 107.27 105.87 111.00 146.8 149.92 108.70 109.50
var.

Var of BS variance 2342.56 2320.35 1305.38 1228.50

Treatment effect ‐1.81 ‐1.79 ‐1.81 ‐1.32 ‐1.34 ‐1.33 ‐1.35 ‐1.80 ‐2.27 ‐1.34 ‐2.06

Variance 0.20 0.20 0.21 0.16 0.18 0.17 0.18 0.23 0.29 0.18 0.30

95%confidence ‐2.69 to −0.93 ‐2.67 to −0.91 ‐2.71 to −0.91 ‐2.10 to −0.54 ‐2.17 to −0.51 ‐2.14 to −0.52 ‐2.18 to −0.52 ‐2.74 to −0.86 ‐3.33 to −1.21 ‐2.18 to −0.50 ‐3.14 to −0.98
int.

Between‐study 2.41 2.42 2.65 1.34 1.58 1.48 1.65 2.50 6.04 1.44 4.87
var.

Var of BS variance 2.56 5.43 1.32 6.40

Hypertension effect . . . 19.31 19.30 19.31 19.20 . . 19.60 19.84

Variance . . . 30.43 32.82 32.45 33.96 . . 34.46 24.90

95%confidence . . . 8.50 to 30.12 8.07 to 30.53 8.14 to 30.48 7.78 to 30.62 . . 8.09 to 31.11 10.06 to 29.62
int.

Between‐study . . . . . 0.0 0.0 . . 7.18 7.27

var.

Var of BS variance 70.73 81.36

Interaction effect . . . ‐3.58 ‐3.54 ‐3.55 ‐3.37 . . ‐3.52 ‐1.23

Variance . . . 1.60 1.70 1.65 2.31 . . 2.31 2.50

95%confidence . . . ‐6.06 to −1.10 ‐6.10 to −0.98 ‐6.07 to −1.03 ‐6.35 to −0.39 . . ‐6.50 to −0.54 ‐4.33 to 1.87
int.

Between‐study . . . . . 0.0 0.56 . . 1.44 1.98

var.

Var of BS variance 2.50 8.70

Abbreviations: IGLS, iterative generalised least square; RIGLS, restricted iterative generalised least square; SemiR, semirandom; FullyR, fully random; BS, between‐study.
.not available.
CURTIN
CURTIN 9

for parallel and cross‐over treatment groups and the estimated of random effects is of particular interest in combined design
average between‐period cross‐over correlation was .46. For meta‐analysis. When different trial designs are used for a
model 3, there were very slight differences between similar clinical question, one can expect that there may be
semirandom Bayesian regression coefficients and those significant variation in the treatment effect that justifies the
obtained by RIGLS‐model. Differences were observed for use of a random effects model. With random effects meta‐
the between‐study variances: For intercept and treatment analysis, the weights of trials become more homogenous, so
covariates, Bayesian between‐study variances were relatively that the original trial variance and covariance have a lesser
similar to those of frequentist methods, but for covariates impact on the meta‐analytic estimates. It also decreases the
hypertension and interaction, the Bayesian method gave weights of the cross‐over design, which have a high relative
higher variance values (7.18 and 1.44 respectively) than weight in combined design meta‐analysis analysed by a
RIGLS which were close to the mean prior values. classical weighted average method.1 This is of importance
For the fully random effects model 3′, the absolute value if one considers the potential risk of carry‐over bias associ-
of the treatment estimate (−2.06) was larger than that ated with cross‐over trials that may impact on the meta‐ana-
observed with the semirandom effects (−1.34) method, the lytic estimate.5,35 The interest of the regression model
between‐study variance for treatment was more than 3 times proposed here and in the accompanying study6 is notably
higher (4.87) than that observed with the semirandom effects the possibility to introduce a carry‐over parameter to control
method (1.44). The estimate for hypertension was similar to for this bias.
that of model 3, but the estimate for interaction was lower. When random effects are estimated by the frequentist
The estimates for ΣP and ΣX were similar to those obtained method, between‐study variance components can differ
in model 1′. between IGLS and RIGLS. The restricted method RIGLS
In Data DS1, one discusses Bayesian methods allowing should be the rule in meta‐analysis where the number of
for correlation in the regression coefficients and for regres- observations (ie, trials) is limited. Frequentist method
sion coefficients following a t distribution. assumes that variance V is known, which is not the case actu-
ally,73 and this can potentially bias the estimates of standard
errors, in particular for cross‐over as the degrees of freedom
5 | DISCUSSI ON used to establish variance is reduced in this trial design.74
“Fully Bayesian” models are relatively rare; however, they
In an accompanying study,6 a meta‐analysis model based on present advantages: The Bayesian approach permits the cal-
GEE allows the combination of cross‐over trials of various culation of the exact variance from the posterior distributions
designs with parallel trials in a fixed effects meta‐analysis. of parameters,71 which can be of value to circumvent the lim-
Here, one extends this model by introducing random effects itations in the estimation of standard errors of cross‐over tri-
using either a frequentist method on the basis of RIGLS or als. Another advantage of the Bayesian method is the
a Bayesian method using MCMC technique. In random introduction of prior estimates at level III. In complex models
effects meta‐analysis, the treatment is often considered as like the present one, priors can be useful for certain parame-
fixed and the effects of trials as well as the treatment‐by‐trial ters that could be difficult to estimate by a classical approach:
interaction are random.70,71 This is conceptually similar to Prior information was used here for cross‐over correlation
model 1 presented above: The random intercept corresponds and between‐study variances. Also of interest is the possible
to a random variation by trial; the random treatment effect is introduction of random variation in parameters such as the
the random treatment‐by‐trial interaction. In the approach trial variance or the between‐period cross‐over correlation
developed here, the treatment effect in treatment groups is at level I, which are considered as fixed in the frequentist
the variable modelled by the regression, not the treatment model.
effect contrast, and the random effect allows to recover inter- It should be noted that the models we have used are based
trial information. It should be noted that this has been consid- on a random trial effect and a random trial by treatment inter-
ered as controversial, as this involves comparing patients action; this is not usual: Indeed generally, the linear models
across trials and to a certain extent using these patient data keep the trial parameter as fixed.41
as if there were historical controls37 and it may lead to bias, Variations on the model such as correlation between
for example, when an imbalance occurs in trials and the pla- regression coefficients or random coefficients with t distribu-
cebo effects are varying across trials,36,72 a point that has tions can also be introduced. In the present example, changes
already been highlighted for the fixed effects approach.6 in the estimates were minor, but in other examples this might
However unless imbalance is large, the difference between potentially matter.
random and fixed effects is likely to be small. We have assumed trial outcomes following a normal dis-
Our model allows parameters, other than the treatment tribution as it is usually the case in meta‐analysis.75,76 The
effect, to be modelled as random effects. The introduction normality assumption may not hold in particular when trial
10 CURTIN

sample size is small. Gibbs sampler as well as IGLS can be 3. Makubate B, Senn S. Planning and analysis of cross‐over trials in
used in models with outcomes following nonnormal distribu- infertility. Stat Med. 2010;29:3203‐3210.
tions27,29; combined design meta‐analysis with random 4. Curtin F, Elbourne D, Altman DG. Meta‐analysis combining paral-
effects can be generalised to outcomes following other distri- lel and cross‐over clinical trials. III: the issue of carry‐over. Stat
butions of the exponential family, paralleling the possibilities Med. 2002b;21:2161‐2173.
of the GEE approach for fixed effects meta‐analysis. 5. Elbourne DR, Altman DG, Higgins JP, Curtin F, Worthington HV,
Overall, the regression models taking into account the Vail A. Meta‐analyses involving cross‐over trials: methodological
correlation between trial periods either using fixed effects issues. Int J Epidemiol. 2002;31:140‐149.
GEE models or IGLS/Bayesian models allow the combina- 6. Curtin F. 2016 Combined design meta‐analysis using generalized
tion of different trial designs in meta‐analysis, in particular, estimating equation.
these approaches allow the combination of marginal estima- 7. Higgins JPT, Thompson SG, Spiegelhalter DJ. A re‐evaluation of
tions,6 which are suitable for cross‐over trials as well as for random‐effects meta‐analysis. Journal of the Royal Statistical
parallel trials.2 The interest of a regression approach for Society A. 2009;172:137‐159.
meta‐analysis is the possibility to explain at least some 8. Glasziou PP, Sanders SL. Investigating causes of heterogeneity in
sources of heterogeneity in the results of the included trials.7 systematic reviews. Stat Med. 2002;21:1503‐1511.
When combining different study designs (cross‐over and par- 9. Thompson SG. Why sources of heterogeneity in meta‐analysis
allel trials), adjusting for design specific covariates (see dis- should be investigated. BMJ. 1994;309:1351‐1355.
cussion of cross‐over covariates in Curtin6) is of interest;
10. Higgins JPT, Green S (Eds). Cochrane Handbook for Systematic
now in the choice between a fixed effects GEE meta‐analysis Reviews of Interventions Version 5.1.0 [updated March 2011]. The
or a random effects approach with RIGLS/Bayes, the same Cochrane Collaboration, Chichester, UK: John Wiley & Sons;
reflection as for simpler meta‐analytic models should apply 2011. Retrieved from http://handbook.cochrane.org
(see for example Borenstein et al77): When there is a series 11. DerSimonian R, Laird N. Meta‐analysis in clinical trial. Control
of trials assessing a very similar clinical question, a fixed Clin Trials. 1986;7:177‐188.
effects approach with a GEE regression could be proposed,
12. Aitkin M. Meta‐analysis by random effect modelling in generalized
using covariates to adjust for well‐identified sources of het- linear models. Stat Med. 1999;18:2343‐2351.
erogeneity. When the heterogeneity of the study designs goes
13. Begg CB, Pilote L. A model for incorporating historical controls
beyond the well‐identified sources and when there is a need
into a meta‐analysis. Biometrics. 1991;47:899‐906.
to generalise results, one should opt for random effects
14. Berkey CS, Hoaglin DC, Mosteller F, Colditz GA. A random‐effects
models such as those described in the present paper. Hierar-
regression model for meta‐analysis. Stat Med. 1995;14:395‐411.
chical Bayesian models are certainly the most flexible of
these models and allow to introduce easily several levels 15. Berkey CS, Hoaglin DC, Antczak‐Bouckoms A, Mosteller F,
Colditz GA. Meta‐analysis of multiple outcomes by regression with
and random parameters, which can be the most appropriate
random effects. Stat Med. 1998;17:2537‐2550.
for the complexity of the combination of these different trial
designs. 16. Li S, Begg CB. Random effects models for combining results from
controlled and uncontrolled studies in a meta‐analysis. J Am Stat
Assoc. 1994;89:1523‐1527.
AC KN OWL ED GE M EN TS 17. Platt RW, Leroux BG, Breslow N. Generalized linear mixed models
for meta‐analysis. Stat Med. 1999;18:643‐654.
I want to thank Diana Elbourne, Douglas Altman, and Ste-
phen Senn for fruitful discussions on the subject as well as 18. Stram DO. Meta‐analysis of published data using a linear mixed‐
helpful comments on methodological aspects. effects model. Biometrics. 1996;52:536‐544.
19. DuMouchel WH, Harris JE. Bayes methods for combining the
results of cancer studies in humans and other species. J Am Stat
C ON F L I C T O F I NT E RE ST Assoc. 1983;78:293‐315.
Francois Curtin is an employee of Geneuro SA, Geneva. 20. Laird N, Louis T. Empirical Bayes confidence intervals for a series
of related experiments. Biometrics. 1989;45:481‐495.
21. Morris CN, Normand SL. Hierarchical Models for Combining Infor-
R E F E RENC E S
mation and for Meta‐Analyses. In: Bernardo JM, Berger JO, Dawid
1. Curtin F, Altman DG, Elbourne D. Meta‐analysis combining paral- AP, Smith AFM, eds. Bayesian Statistics 4. Oxford: Oxford
lel and cross‐over clinical trials. I: continuous outcomes. Stat Med. University Press; 1992:321‐344.
2002a;21:2131‐2144. 22. Raudenbush SW, Bryk AS. Empirical Bayes meta‐analysis. Journal
of Educational Statistics. 1985;10:75‐98.
2. Curtin F, Elbourne D, Altman DG. Meta‐analysis combining paral-
lel and cross‐over clinical trials. II: binary outcomes. Stat Med. 23. Sutton AJ, Abrams KR. Bayesian methods in meta‐analysis and
2002a;21:2145‐2159. evidence synthesis. Stat Methods Med Res. 2001;10:277‐303.
CURTIN 11

24. DuMouchel WH. Bayesian Meta‐Analysis. In: Berry DA, ed. on blood pressure and intralymphocytic sodium concentration in
Statistical Methodology in the Pharmaceutical Sciences. New York, young subjects with borderline hypertension. Clin Sci.
NY: Marcel Dekker; 1990:509‐529. 1981;61:21s–23s.
25. Lambert PC, Sutton AJ, Burton PR, Abrams KR, Jones DR. How 45. Morgan TO, Myers JB. Hypertension treated by sodium restriction.
vague is vague? A simulation study of the impact of the use of vague Med J Aust. 1981;2:396–397.
prior distributions in MCMC using WinBUGS. Stat Med.
46. Puska P, Iacono JM, Nissinen A, et al. Controlled randomised trial
2005;24:2401‐2428.
of the effect of dietary fat on blood pressure. Lancet 1983;1:1–5.
26. Senn S. Trying to be precise about vagueness. Stat Med.
47. Silman AJ, Locke C, Mitchell P, Humpherson P. Evaluation of the
2007;26:1417‐1430.
effectiveness of a low sodium diet in the treatment of mild to mod-
27. Goldstein H. Multilevel statistical models. Second ed. London: erate hypertension. Lancet 1983;1:1179–1182.
Arnold; 1995.
48. Erwteman TM, Nagelkerke N, Lubsen J, Koster M, Dunning AJ.
28. Goldstein H, Browne W, Rasbash J. Multilevel modelling of medical Blockade, diuretics and salt restriction for the management of mild
data. Stat Med. 2002;21:3291‐3315. hypertension: A randomised double blind trial. Br Med J.
29. Clayton DG. Generalized linear mixed model. In: Gilks WR, 1984;289:406–409.
Richardson S, Spiegelhalter DJ, eds. Markov Chain Monte Carlo 49. Fagerberg B, Anderson OK, Isaksson B, Bjorntorp P. Blood pres-
in Practice. London: Chapman and Hall; 1996.
sure control during weight reduction in obese hypertensive men:
30. Daniels MJ, Gatsonis C. Hierarchical generalized linear models in Separate effects of sodium and energy restriction. Br Med J.
the analysis of variations in health care utilization. J Am Stat Assoc. 1984;288:11–14.
1999;94:29‐42.
50. Maxwell M, Kushiro TK, Dornfeld LP, Tuck ML, Waks AU. BP
31. Lindley DV, Smith FM. Bayes estimates for the linear model changes in obese hypertensive subjects during rapid weight loss.
(with discussion). Journal of the Royal Statistical Society: Series Arch Intern Med. 1984;144:1581–1584.
B. 1972;34:1‐41.
51. Chalmers J, Morgan T, Doyle A, et al. Australian National Health
32. Prevost TC, Abrams KR, Jones DR. Hierarchical models in general- and Medical Research Council Dietary Salt Study in Mild Hyperten-
ized synthesis of evidence: An example based on studies of breast sion. J Hypertens. 1986;4:s629–s637.
cancer screening. Stat Med. 2000;19:3359‐3376.
52. Logan A. Sodium manipulation in the management of hypertension.
33. Zeger SL, Karim MR. Generalized linear models with random The view against its general use. Can J Physiol Pharmacol.
effects: a Gibbs sampling approach. J Am Stat Assoc. 1986;64:793–801.
1991;86:79‐86.
53. Australian National Health and Medical Research Council Dietary
34. Gelman A, Rubin DB. Markov chain Monte Carlo methods in
Salt Study Management Committee (a). 1989. Effects of replacing
biostatistics. Stat Methods Med Res. 1996;5:339‐355.
sodium intake in subjects on a low sodium diet. A crossover study.
35. Senn S. Cross‐over trials in clinical research. 2nd ed. Chichester: Clinical and Experimental Hypertension (A) 11: 1011–1024.
Wiley; 2002.
54. Dodson PM, Beevers M, Allworth R, Webberley MJ, Fletcher RF,
36. Senn S. A note regarding “random effects”. Stat Med. Taylor KG. Sodium restriction and blood pressure in hypertensive
2014;33:2876‐2877. type II diabetics: Randomised blind controlled and crossover studies
37. Senn S. The many modes of meta. Drug Inf J. 2000;34:535‐549. of moderate sodium restriction and sodium supplementation. Br
Med J 1989;298:227–230.
38. Follmann D, Elliott P, Suh I, Cutler J. Variance imputation for over-
views of clinical trials with continuous response. J Clin Epidemiol. 55. Hypertension Prevention Trial Research Group. The Hypertension
1992;45:769‐773. Prevention Trial: Three‐year effects of dietary changes on blood
pressure. Arch Intern Med. 1990;150:153–162.
39. Goldstein H. Restricted unbiased iterative generalized least‐squares
estimation. Biometrika. 1989;76:622‐623. 56. Parijs J, Joossens JV, Van der Linden L, Verstreken G, Amery A.
40. Spiegelhalter DJ, Best NG, Gilks WR, Inskip H. Hepatitis B: a case Moderate sodium restriction and diuretics in the treatment of hyper-
study in MCMC methods. In: Gilks WR, Richardson S, tension. Am Heart J. 1973;85:22–34.
Spiegelhalter DJ, eds. Markov Chain Monte Carlo in Practice. 57. Skrabal F, Auböck J, Hörtnagl H. Low sodium/high potassium diet
London: Chapman et Hall; 1996. for prevention of hypertension: probable mechanisms of action.
41. Spiegelhalter D, Thomas A, Best N, Gilks W. BUGS 0.5. Bayesian Lancet 1981;1:895–900.
inference Using Gibbs Sampling Manual. Cambridge: MRC Biosta- 58. MacGregor GA, Markandu ND, Best FE, et al. Double‐blind
tistics Unit; 1995 .Retrieved from: http://www.mrc‐bsu.cam.ac.uk randomised cross‐over trial of moderate sodium restriction in essen-
42. Cutler JA, Follmann D, Elliott P, Suh I. An overview of randomized tial hypertension. Lancet 1982;1:351–354.
trials of sodium reduction and blood pressure. Hypertension. 59. Watt GCM, Edwards C, Hart JT, Hart M, Walton P, Foy CJW. Die-
1991;17:27‐33. tary sodium restriction for mild hypertension in general practice. Br
43. Morgan T, Adam W, Gillies A, Wilson M, Morgan G, Carney S. Med J. 1983;286:432–436.
Hypertension treated by salt restriction. Lancet 1978;1:227–230. 60. Cooper R, Van Horn L, Liu K, et al. A randomized trial on the effect
44. Costa FV, Ambrosioni E, Montebugnoli L, Paccaloni L, Vasconi of dietary sodium intake on blood pressure in adolescents. J
L, Magnani B. Effects of a low salt diet and of acute salt loading Hypertens. 1984;2:361–366.
12 CURTIN

61. Richards AM, Nicholls MG, Espiner EA, et al. Blood pressure 70. Brown H, Prescott R. Mixed Models Analysis of Clinical Trials
response to moderate sodium restriction and to potassium sup- Using SAS. Proc Mixed and Beyond. Edimburgh: University of
plementation in mild essential hypertension. Lancet 1984; Edimburgh; 1998.
1:757–761. 71. Brown H, Prescott R. Applied Mixed Models in Medicine.
62. Watt GC, Foy CJW, Hart JT, et al. Dietary sodium and arterial blood Chichester: John Wiley Sons Ltd.; 1999.
pressure evidence against genetic susceptibility. Br Med J. 72. Senn S. Hans van Houwelingen and the art of summing up. Biom J.
1985;291:1525–1528. 2010;52:1‐10.

63. Grobbee DE, Hofman A, Roelandt JT, Boomsma F, Schalekamp 73. Biggerstaff BJ, Tweedie RL. Incorporating variability in estimates
MA, Valkenburg HA. Sodium restriction and potassium supplemen- of heterogeneity in the random effects model in meta‐analysis. Stat
tation in young people with mildly elevated blood pressure. J. Med. 1997;16:753‐768.
Hypertens. 1987;5:115–119. 74. Senn S. Various varying variances: the challenge of nuisance param-
eters to the practising biostatistician. Stat Methods Med Res.
64. Australian National Health and Medical Research Council Dietary
2015;24:403‐419.
Salt Study Management Committee (b). Fall in blood pressure with
modest reduction in dietary salt intake in mild hypertension. Lancet 75. Greenland S. Quantitative methods in the review of epidemiologic
1989;1:399–402. literature. Epidemiol Rev. 1987;9:1‐31.
76. Lunn D, Barrett J, Sweeting M, Thompson S. Fully Bayesian hierar-
65. Dodson PM, Beevers M, Allworth R, Webberley MJ, Fletcher RF,
chical modelling in two‐stages with application to meta‐analysis.
Taylor KG. Sodium restriction and blood pressure in hypertensive
Appl. Stat. 2013;62:551‐572.
type II diabetics: Randomised blind controlled and crossover studies
of moderate sodium restriction and sodium supplementation. Br 77. Borenstein M, Hedges LV, Higgins JP, Rothstein HR. A basic intro-
Med J 1989;298:227–230. duction to fixed‐effect and random‐effects models for meta‐analysis.
Res Syn Meth. 2010;1:97‐111.
66. MacGregor GA, Markandu N, Sagnella GA, Singer DR, Cappucio
78. Curtin F. Meta‐analysis combining cross‐over and parallel group
FP. Double‐blind study of three sodium intakes and long‐term
trials. MPhil Thesis. University of London, London 1999.
effects of sodium restriction in essential hypertension. Lancet
1989;2:1244–1247.

67. Myers JB. Reduced sodium chloride intake normalises blood pres- SUPP ORT ING INFORM ATION
sure distribution. J Hum Hypertens.. 1989;3: 97–104. Additional Supporting Information may be found online in
68. Smith TC, Spiegelhalter DJ, Thomas A. Bayesian approaches to
the supporting information tab for this article.
random‐effects meta‐analysis: a comparative study. Stat Med.
1995;14:2685‐2699.
How to cite this article: Curtin F. Meta‐analysis
69. Best N, Cowles MK, Vines K. CODA, Convergence Diagnosis and combining parallel and cross‐over trials with random
Output Analysis Software for Gibbs sampling output, Version 0.30.
effects. Res Syn Meth. 2017. https://doi.org/10.1002/
Cambridge: MRC Biostatistics Unit; 1996 .Retrieved from: http://
jrsm.1236
www.mrc‐bsu.cam.ac.uk