Introduction to the

MethComp package

(compiled Friday 20th May, 2011, 02:09)

Bendix Carstensen Steno Diabetes Center, Gentofte, Denmark

& Department of Biostatistics, University of Copenhagen
[email protected]
www.biostat.ku.dk/~bxc
Contents

1 Overview of MethComp 3
1.1 Data structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Wide format data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Function overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 Graphical functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.2 Data manipulating functions . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3 Analysis functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.4 Reporting functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Worked examples 9
2.1 Fat measurements: Exchangeable replicates . . . . . . . . . . . . . . . . . 9
2.2 Cardiac output: Linked replicates? . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Systolic blood pressure: Linked replicates by two methods . . . . . . . . . 19
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2
Chapter 1

Overview of MethComp

The purpose of the MethComp package is to provide computational tools to manipulate,

display and analyze data from method comparison studies. A method comparison study
is a study where two methods of quantitative measurement are compared by measuring
the same set of items with both methods.
There may be more than two methods, and there may be replicate measurements on
each item by each method.

1.1 Data structures

In general we are concerned with measurements by different methods, on different items
(persons, samples), possibly replicated.
Often such data are represented by a row of measurements for each item, with
possible replicates listed either below or beside each other. This implicitly assumes that
some replicate measurements belong together, which is not necessarily the case in all
situations.
All functions in MethComp assume data to be represented in the “long” form, with one
measurement on each row, and columns to indicate method, item and replicate.
Specifically, we assume the following columns are available in a data frame:
• meth The measurement method. Numeric or factor.
• item Identification of item (person, sample). Numeric or factor.
• repl Replicate number. Numeric or factor.
• y The measurement by method meth on item item, replicate number repl.
There is a class, “Meth” for this kind of data frame. It is a data frame with the facors
meth, item and repl representing the classification, and the numerical variable y
representing the measurements.
A dataframe with method comparison data in the long format is converted to a Meth
object by using the Meth function on it:
> data( ox )
> str( ox )

3
4 1.1 Data structures

'data.frame': 354 obs. of 4 variables:

$ meth: Factor w/ 2 levels "CO","pulse": 1 1 1 1 1 1 1 1 1 1 ...
$ item: num 1 1 1 2 2 2 3 3 3 4 ...
$ repl: num 1 2 3 1 2 3 1 2 3 1 ...
$ y : num 78 76.4 77.2 68.7 67.6 68.3 82.9 80.1 80.7 62.3 ...

> ox <- Meth( ox )

The following variables from the dataframe

"ox" are used as the Meth variables:
meth: meth
item: item
repl: repl
y: y
#Replicates
Method 1 2 3 #Items #Obs: 354 Values: min med max
CO 1 4 56 61 177 22.2 78.6 93.5
pulse 1 4 56 61 177 24.0 75.0 94.0

> summary( ox )

#Replicates
Method 1 2 3 #Items #Obs: 354 Values: min med max
CO 1 4 56 61 177 22.2 78.6 93.5
pulse 1 4 56 61 177 24.0 75.0 94.0

If variables meth, item, repl and y are not availabe in the data frame we may create
them on the fly or give the variable positions as arguments to the Meth function:
> data( fat )
> str( fat )

'data.frame': 258 obs. of 5 variables:

$ Id : num 1 1 1 3 3 3 5 5 5 11 ...
$ Obs: Factor w/ 2 levels "KL","SL": 1 1 1 1 1 1 1 1 1 1 ...
$ Rep: num 1 2 3 1 2 3 1 2 3 1 ...
$ Sub: num 1.6 1.7 1.7 2.8 2.9 2.8 2.7 2.8 2.9 3.9 ...
$ Vic: num 4.5 4.4 4.7 6.4 6.2 6.5 3.6 3.9 4 4.3 ...

> sc <- Meth( fat, 2, 1, 3, 4 )

The following variables from the dataframe

"fat" are used as the Meth variables:
meth: Obs
item: Id
repl: Rep
y: Sub
#Replicates
Method 3 #Items #Obs: 258 Values: min med max
KL 43 43 129 0.39 1.7 4.2
SL 43 43 129 0.51 1.7 4.1

> str( sc )

Classes 'Meth' and 'data.frame': 258 obs. of 5 variables:

$ meth: Factor w/ 2 levels "KL","SL": 1 1 1 1 1 1 1 1 1 1 ...
$ item: Factor w/ 43 levels "1","2","3","4",..: 1 1 1 3 3 3 5 5 5 11 ...
$ repl: Factor w/ 3 levels "1","2","3": 1 2 3 1 2 3 1 2 3 1 ...
$ y : num 1.6 1.7 1.7 2.8 2.9 2.8 2.7 2.8 2.9 3.9 ...
$ Vic : num 4.5 4.4 4.7 6.4 6.2 6.5 3.6 3.9 4 4.3 ...
The MethComp package. 5

> summary( sc )
#Replicates
Method 3 #Items #Obs: 258 Values: min med max
KL 43 43 129 0.39 1.7 4.2
SL 43 43 129 0.51 1.7 4.1

We may even give some of them as names of the columns in the dataframe:
> vi <- Meth( fat, 2,1,"Rep","Vic" )
The following variables from the dataframe
"fat" are used as the Meth variables:
meth: Obs
item: Id
repl: Rep
y: Vic
#Replicates
Method 3 #Items #Obs: 258 Values: min med max
KL 43 43 129 2.0 3.9 6.5
SL 43 43 129 2.3 4.1 6.7

However, more complicated operations on the dataframe is best done on the fly using
the with function (from the base package):
> data( hba1c )
> str( hba1c )
'data.frame': 835 obs. of 6 variables:
$ dev : Factor w/ 3 levels "BR.V2","BR.VC",..: 2 2 2 2 2 2 2 2 1 1 ...
$ type : Factor w/ 2 levels "Cap","Ven": 2 2 2 2 1 1 1 1 2 2 ...
$ item : num 12 12 12 12 12 12 12 12 12 12 ...
$ d.samp: num 1 1 1 1 1 1 1 1 1 1 ...
$ d.ana : num 2 3 4 5 2 3 4 5 2 3 ...
$ y : num 8.7 8.7 8.7 8.7 9.2 9 8.8 8.7 9.4 9.3 ...
> hb1 <- with( hba1c,
+ Meth( meth = interaction(dev,type),
+ item = item,
+ repl = d.ana-d.samp,
+ y = y, print=TRUE ) )
#Replicates
Method 3 4 #Items #Obs: 835 Values: min med max
BR.V2.Cap 0 38 38 152 5.3 8.0 12.6
BR.VC.Cap 19 19 38 133 5.3 8.2 12.1
Tosoh.Cap 0 38 38 152 5.0 7.8 11.8
BR.V2.Ven 19 19 38 133 5.5 8.1 12.0
BR.VC.Ven 19 19 38 133 5.3 8.0 11.6
Tosoh.Ven 20 18 38 132 5.3 8.0 12.1
> str( hb1 )
Classes 'Meth' and 'data.frame': 835 obs. of 4 variables:
$ meth: Factor w/ 6 levels "BR.V2.Cap","BR.VC.Cap",..: 5 5 5 5 2 2 2 2 4 4 ...
$ item: Factor w/ 38 levels "1","2","3","4",..: 12 12 12 12 12 12 12 12 12 12 ...
$ repl: Factor w/ 5 levels "0","1","2","3",..: 2 3 4 5 2 3 4 5 2 3 ...
$ y : num 8.7 8.7 8.7 8.7 9.2 9 8.8 8.7 9.4 9.3 ...

Objects of class Meth (which inherits from data.frame) has methods such as summary,
plot, subset and transform. The functions mostly do not require the data to be in
Meth format — if a dataframe with the right columns is supplied, it is normally
converted internally to Meth format.
6 1.2 Function overview

1.1.1 Wide format data

Sometimes data frames comes in the wide format, that is with measurements by
different methods in different columns. In this case a Meth object is formed by giving
the variables containing measurements by different methods as a vector argument to y,
either as numbers of columns or names of columns:
> data( rainman )
> str( rainman )

'data.frame': 30 obs. of 6 variables:

$ SAND: int 120 48 88 32 24 100 52 80 72 96 ...
$ ME : int 175 50 150 45 25 125 70 145 85 110 ...
$ TM : int 120 50 75 22 22 80 50 75 90 110 ...
$ AJ : int 105 45 75 28 25 91 48 68 55 84 ...
$ BM : int 100 50 60 30 20 80 45 55 60 65 ...
$ LO : int 100 70 80 30 20 70 50 60 60 65 ...

> RM <- Meth( rainman, item=1, y=2:6 )

The following variables from the dataframe

"rainman" are used as the Meth variables:
item: SAND
y: ME TM AJ BM LO
#Replicates
Method 3 #Items #Obs: 150 Values: min med max
AJ 10 10 30 18 57 120
BM 10 10 30 15 62 120
LO 10 10 30 20 55 100
ME 10 10 30 24 90 200
TM 10 10 30 20 75 120

> head( RM )

meth item repl y

1 ME 120 1 175
2 ME 48 1 50
3 ME 88 1 150
4 ME 32 1 45
5 ME 24 1 25
6 ME 100 1 125

1.2 Function overview

The following is a brief overview of the functions in the MethComp package. The full
documentation is in the help pages for the functions, and an illustration of the way they
work can be obtained by referring to the examples in the help pages. The help page for
plot.meth is brought up by:
> ?plot.Meth

The example code from the manual page can be run directly by:
> example( plot.Meth )
The MethComp package. 7

1.2.1 Graphical functions

The graphical functions generally have a lot of arguments that can be used to fine-tune
the looks of the plots. Refer to the help page for each to see them all.
BA.plot Makes a Bland-Altman plot of two methods from a data frame with method
comparison data, and computes limits of agreement. The plotting is really done by
a call to the function BlandAltman. The default is to plot the two first methods
against each other.

BlandAltman draws a Bland-Altman plot and computes limits of agreement.

bothlines Adds regression lines of y on x and vice versa to a scatter plot. Optionally,
the Deming regression line can be added too.

plot.Meth Plots all methods against each other in a square matrix, both as a scatter
plot (below diagonal) and as a Bland-Altman plot (above diagonal).

plot.MethComp plots the estimated conversion between methods with a ±2 sd interval,

corresponding to approx. 95% prediction interval. Recognizes transformations
applied to data.

1.2.2 Data manipulating functions

make.repl Generates (or replaces) a repl column in a Meth object.

perm.repl Randomly permutes replicates within (method,item) and assigns new

replicate numbers.

to.wide Transforms a data frame in the long form to the wide form where separate
columns for each method are generated, with one row per (item,repl).

to.long Reverses the result of to.wide. The function can also generate a long form
dataset from a dataset with different methods beside each other.

summary.Meth Tabulates items by method and no. replicates for a Meth object.

Meth.sim Simulates a dataset from a method comparison experiment for given

parameters for bias, exchangeability and variance component sizes.

1.2.3 Analysis functions

DA.reg Regresses the differeneces between methods on the averages and derives
approximate linear conversion equations, based on [?].

Deming Performs Deming regression, i.e. regression with errors in both variables.

BA.est Estimates in the variance components models underlying the concept of limits
of agreement, and returns the bias and the variance components. The model used
assumes constant bias between methods.
8 1.2 Function overview

AltReg Estimates via alternating regressions in the general model. Returns estimates
of mean conversion parameters and variance components.

MCmcmc Estimates via BUGS in the general model with non-constant bias (and in the
future) possibly non-constant standard deviations of the variance components.
Produces a MCmcmc object, which is an mcmc.list object with some extra
attributes. mcmc.list objects are handeled by the coda package, so this is
required when calling MCmcmc.

1.2.4 Reporting functions

Some of these functions take an MCmcmc object as input, others will postprocess the
output of DA.reg, BA.est or AltReg.
The functions DA.reg, BA.est, AltReg return objetcs that have class MethComp,
whereas the result of MCmcmc can be converted to an object of this type by the MethComp
function. The reason for this is that the results of the MCmcmc function is output from an
MCMC-simulation which we may want to monitor by special functions. The MethComp
function only extracts the central summaries from the MCmcmc object assuming the
chains have reached convergence.

print.MethComp Prints a table of conversion equation between methods analyzed, with

prediction standard deviations.

print.MCmcmc Prints a table of conversion equation between methods analyzed, with

prediction standard deviations, but also gives summaries of the posteriors for the
parameters that constitute the conversion algorithms.

plot.MethComp, plot.MCmcmc Plots the conversion lines between methods with

prediction limits.

post.MCmcmc Plots smoothed posterior densities for the estimates. Primarily of interest
for the variance components, but it has aruments to produce the posterior of the
intercepts and the slopes of the conversion lines between methods too.

check.MCmcmc Makes diagnistic plots of the traces of the chains included in the MCmcmc
object.
Chapter 2

Worked examples

2.1 Fat measurements: Exchangeable replicates

The fat data from the MethComp package contains measurements of subcutaneous and
visceral fat on 43 persons, by two observers, KL and SL. Each measurement is replicated
3 times.
First we examine the names in the dataframe, and then use Meth to convert it to a
form that comply with that required by the functions in the MethComp package for
analyzing visceral fat — we convert it to a Meth object:
> data(fat)
> str(fat)

'data.frame': 258 obs. of 5 variables:

$ Id : num 1 1 1 3 3 3 5 5 5 11 ...
$ Obs: Factor w/ 2 levels "KL","SL": 1 1 1 1 1 1 1 1 1 1 ...
$ Rep: num 1 2 3 1 2 3 1 2 3 1 ...
$ Sub: num 1.6 1.7 1.7 2.8 2.9 2.8 2.7 2.8 2.9 3.9 ...
$ Vic: num 4.5 4.4 4.7 6.4 6.2 6.5 3.6 3.9 4 4.3 ...

> vis <- Meth( fat, 2,1,3,5 )

The following variables from the dataframe

"fat" are used as the Meth variables:
meth: Obs
item: Id
repl: Rep
y: Vic
#Replicates
Method 3 #Items #Obs: 258 Values: min med max
KL 43 43 129 2.0 3.9 6.5
SL 43 43 129 2.3 4.1 6.7

> str(vis)

Classes 'Meth' and 'data.frame': 258 obs. of 5 variables:

$ meth: Factor w/ 2 levels "KL","SL": 1 1 1 1 1 1 1 1 1 1 ...
$ item: Factor w/ 43 levels "1","2","3","4",..: 1 1 1 3 3 3 5 5 5 11 ...
$ repl: Factor w/ 3 levels "1","2","3": 1 2 3 1 2 3 1 2 3 1 ...
$ y : num 4.5 4.4 4.7 6.4 6.2 6.5 3.6 3.9 4 4.3 ...
$ Sub : num 1.6 1.7 1.7 2.8 2.9 2.8 2.7 2.8 2.9 3.9 ...

> summary(vis)

9
10 2.1 Fat measurements: Exchangeable replicates

●
●
●●
●

6 ●● ●
●● ●
●
● ● ●
●● ●

●
● ● ● ●
5

● ● ●
● ●● ●
● ● ● ●
● ● ● ● ●
● ●
●
SL

●● ●●●
● ●● ● ●
● ●● ●
●
4

● ●● ●● ●
● ● ● ● ●
● ● ●
● ● ● ●●
● ●
● ●

● ●● ●●●
● ● ● ●
● ●
3

● ●
●● ●●●● ● ●
● ●
●
●
●
●
2

2 3 4 5 6

Figure 2.1: Two observers measuring visceral fat.

#Replicates
Method 3 #Items #Obs: 258 Values: min med max
KL 43 43 129 2.0 3.9 6.5
SL 43 43 129 2.3 4.1 6.7

The two methods plotted against each other requires that we use the replicate
number for pairing the measurements; so we just keep the ordering among the replicates
when using to.wide:
> pw <- to.wide( vis )

Note:
Replicate measurements are taken as separate items!

> par( mar=c(3,3,1,1) )

> with(pw, plot( SL ~ KL, pch=16, xlim=range(vis$y), ylim=range(vis$y) ) )
> abline( 0,1 )

Since replicates are exchangeable witin (method, item) we should get the same sort of
overview of the data after a random permutation of the replicates. Plotting the data
using the original replicate numbers for pairing and then a random permutation is
shown in figure 2.2:
> plot( vis )
The MethComp package. 11

2 3 4 5 6 2 3 4 5 6

2 2

●
●● ●
●
● 1 ●
●
●
● 1
●●● ● ● ●● ● ● ●
● ●● ●● ● ● ● ● ●● ● ● ● ● ● ●
●● ●●● ● ● ● ● ● ● ● ● ●
● ● ● ● ●● ●● ● ● ● ●● ● ●● ● ●
● ● ● ● ● ● ●● ● ● ●● ●
● ● ● ● ● ●●● ●● ● ● ● ●● ● ● ● ● ● ● ●●● ●● ● ●● ●

KL ● ●
●
●●
●● ● ●
● ● ●● ●
● ●
● ●
●
● ●
●●
●
●●

● ●
● ● ●
●● ●●●
● ●●

●
●
●

●
●
●● ●

●
●
● ●
●
●
● 0 KL ● ● ●●

● ●
●
● ● ●
●
●
● ●
●●●●
●
● ●
●●
●● ●●●
●

●
●● ●●
● ●

●
● ●
● ●
●

●
●
●●
●
● 0
● ● ● ●●
●
● ●

−1 −1

−2 −2

● ●
● ●
●● ●
● ●

● ●
6 ●●
●● ●
●
6 ● ●
●● ●
●
● ●
● ● ● ● ● ●
●● ● ● ● ●
● ●
● ●
5 ● ● ●
● ●
● ●● ●
●
● 5 ● ●● ●
●● ● ●
● ●● ● ●
● ● ● ● ● ●●● ●
● ● ● ● ● ● ● ● ● ●

SL SL
● ● ● ●
● ●
●● ●●● ●● ●●
● ●● ● ● ● ● ● ●
● ●● ● ● ●● ●
4 ●
● ●
●● ●●
●
● ●
●
●
4 ●
● ●● ● ●● ●
● ● ● ● ●
● ● ● ● ● ● ●●●
● ● ● ●● ● ● ● ●
● ● ● ●
● ● ● ●
● ●● ●●● ● ● ●●
●
● ● ● ● ● ● ●
3 ●
●● ●●●●
● ●
●
● ● ●
3 ●
●● ● ●●
●●
●
●
● ● ●●
●
● ● ● ●
● ●
● ●
● ●
● ●

2 2

2 3 4 5 6 2 3 4 5 6

Figure 2.2: Plot of two methods of measuring visceral fat, using different pairings of the
replicates; the left panel is using the pairing in the original coding, the right panel is with
a random permutation of replicates.

Note:
Replicate measurements are taken as separate items!

> plot( perm.repl( vis ) )

Note:
Replicate measurements are taken as separate items!

These two plots are shown in figure 2.2 where it is pretty clar that the random
permutation of replicates has little effect.
BA.plot produces a Bland-Altman plot and computes the limits of agreement using
the pairing of replicates across methods based on the numbering of replicates. However
we do not want the replicates to be connected, so we must specify this explicitly:
> par( mar=c(3,3,3,3), mgp=c(3,1,0)/1.6 )
> BA.plot( vis, conn.repl=FALSE )

Limits of agreement:
SL - KL 2.5% limit 97.5% limit SD(diff)
0.1550388 -0.5612718 0.8713493 0.3581553

We see that using this approximation we get limits of agreement for KL−SL of
(−0.86, 0.55).
Moreover, there seems to be no indication that the difference between observers or
the variance varies with the level of measurement. This can be a bit more formally
tested using the DA.reg function (again using the existing pairing of replicates):
> DA.reg( vis )
12 2.1 Fat measurements: Exchangeable replicates

1.0
●

● ●
0.87
●● ●

●●● ● ●

● ●● ●● ● ● ●
0.5

●● ●●● ● ●

● ● ● ● ●● ●●

● ● ● ● ●
SL − KL

● ● ● ● ● ●●● ●● ● ● ● ●● ●
0.16
● ● ●● ● ● ●● ●●● ●● ● ●
0.0

● ● ●● ● ●

●● ● ● ●● ● ●● ●

● ● ●● ● ●

● ● ● ● ● ● ●

● ● ● ● ● ● ● ● ●
−0.5

●
−0.56

3 4 5 6
( SL + KL ) / 2

Figure 2.3: Bland-Altman plot of two observers measuring visceral fat.

Conversion between methods:

alpha beta sd.pred beta=1 sd.|A=4 slope(sd) sd.=K
To: From:
KL KL 0.000 1.000 NA NA NA NA NA
SL -0.340 1.044 0.365 0.158 0.366 -0.024 0.275
SL KL 0.326 0.957 0.349 0.158 0.366 -0.024 0.275
SL 0.000 1.000 NA NA NA NA NA

From the last two columns (p-values for tests of constant difference and constant sd.) it
is clear that there are no obvious violations of the assumptions about constant difference
or about constant variation across the range of measurements.
Setting up a proper variance component model we get only slightly different limits of
agreement (note that we must specify the replicates to be exchangeable):
> ( vis.est <- BA.est( vis, linked=FALSE ) )

Conversion between methods:

alpha beta sd LoA: lower upper
To: From:
KL KL 0.000 1.000 0.273 -0.545 0.545
SL -0.155 1.000 0.364 -0.883 0.573
SL KL 0.155 1.000 0.364 -0.573 0.883
SL 0.000 1.000 0.245 -0.490 0.490

Variance components (sd):

IxR MxI res
The MethComp package. 13

1.0
●

● ●
0.87
●● ●

●●● ● ●

● ●● ●● ● ● ●
0.5

●● ●●● ● ●

● ● ● ● ●● ●●

● ● ● ● ●
SL − KL

● ● ● ● ● ●●● ●● ● ● ● ●● ●
0.16
● ● ●● ● ● ●● ●●● ●● ● ●
0.0

● ● ●● ● ●

●● ● ● ●● ● ●● ●

● ● ●● ● ●

● ● ● ● ● ● ●

● ● ● ● ● ● ● ● ●
−0.5

−0.56
●

3 4 5 6
( SL + KL ) / 2

Figure 2.4: Bland-Altman-plot of two methods of measuring visceral fat, using different
pairings of the replicates. The blue lines are the LoA based on taking the paired replicates
as items, the red lines are based on the estimates from the proper variance component
model.

KL 0 0.181 0.193
SL 0 0.181 0.173

Moreover we get the coefficient of reproducibility for each of the methods; that is an
upper 95% confidence interval for the absolute difference between two measurements by
the same method on the same
We can visualize the difference between the ad-hoc-computed LoA and the model
based ones by plotting them in the same graph:
> par( mar=c(3,3,1,3), mgp=c(3,1,0)/1.6 )
> BA.plot( vis, conn.repl=FALSE )

Limits of agreement:
SL - KL 2.5% limit 97.5% limit SD(diff)
0.1550388 -0.5612718 0.8713493 0.3581553

> abline( h=vis.est$LoA[1:3], col="red" )

As predicted by the theory, the limits based on the ad-hoc paired replicates are roughly
equal to those derived from the proper variance component model — see figure 2.4.
14 2.2 Cardiac output: Linked replicates?

1.0
●

● ●
0.87
● ● ●

0.75
● ● ● ● ● ●
●
●
● ● ● ● ● ● ● ●
● ●
●
0.5
● ● ● ● ● ● ● ●

●
● ● ● ● ● ● ● ● ●
● ● ●
● ● ●
● ● ● ● ● ●
●
● ●● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
0.16
SL − KL

● ● ● ● ● ● ●●
● ● ● ● ● ● ● ●

●
0.0

● ● ● ●●
● ● ●
● ● ●
● ● ● ●
● ● ●
● ● ● ● ● ● ● ●
●
●
● ● ● ● ● ● ● ●
●
●
● ● ● ● ● ● ●

●
● ● ● ● ● ● ● ● ●
−0.44
−0.5

●
−0.56
●
−1.0

2 3 4 5 6 7
( SL + KL ) / 2

Figure 2.5: Bland-Altman-plot of two methods of measuring visceral fat, based on the
arbitrary pairing of the replicates (black) and on the mean over replicates (grey).

In order to illustrate the effect of basing the limits of agreement on the mean over the
replicates we use the argument mean.repl, and the trick of using par(new=T) to over
plot:
> par( mar=c(3,3,1,3), mgp=c(3,1,0)/1.6 )
> BA.plot(vis,mean.repl=T,limy=c(-1,1),limx=c(2,7),col=gray(0.7),col.lines=gray(0.5), conn.repl=FALS
Limits of agreement:
SL - KL 2.5% limit 97.5% limit SD(diff)
0.1550388 -0.4371295 0.7472070 0.2960841
> par(new=T)
> BA.plot(vis,mean.repl=F,limy=c(-1,1),limx=c(2,7),cex=0.7,conn.repl=FALSE)
Limits of agreement:
SL - KL 2.5% limit 97.5% limit SD(diff)
0.1550388 -0.5612718 0.8713493 0.3581553

The two superposed Bland-Altman plots are shown in figure ??.

2.2 Cardiac output: Linked replicates?

The dataset is adapted from table 4 in: JM Bland and DG Altman: Measuring
agreement in method comparison studies. Statistical Methods in Medical Research,
The MethComp package. 15

8:136-160, 1999. Originally supplied to Bland & Altman by Dr LS Bowling, see:

Bowling LS, Sageman WS, O’Connor SM, Cole R, Amundson DE. Lack of agreement
between measurement of ejection fraction by impedance cardiography versus
radionuclide ventriculography. Critical Care Medicine 1993; 21: 1523-27.
It consists of measurements of cardiac output on 12 persons. For each person the
cardiac output is measured repeatedly (three to six times) by impedance cardiography
(IC) and radionuclide ventriculography (RV).
The dataset is supplied with the MethComp package, and comes with the correct
variable names, so it can immediately be transformed into a Meth object:
> data( cardiac )
> cardiac <- Meth( cardiac )
The following variables from the dataframe
"cardiac" are used as the Meth variables:
meth: meth
item: item
repl: repl
y: y
#Replicates
Method 3 4 5 6 #Items #Obs: 120 Values: min med max
IC 1 3 3 5 12 60 2.32 4.610 7.40
RV 1 3 3 5 12 60 2.85 5.105 7.89

It is not clear from the description of the dataset whether replicates are linked across
methods or not, but a quick check can be made graphically by making a Bland-Altman
plot on the data as supplied and on the dat where replicates are randomly permuted,
and then compare them as in figure 2.2.
> par( mfrow=c(1,2), mar=c(3,3,1,3), mgp=c(3,1,0)/1.6 )
> BA.plot( cardiac , limy=c(-3,3) )
Limits of agreement:
RV - IC 2.5% limit 97.5% limit SD(diff)
0.6021667 -1.3199476 2.5242809 0.9610571
> BA.plot( perm.repl(cardiac), limy=c(-3,3) )
Limits of agreement:
RV - IC 2.5% limit 97.5% limit SD(diff)
0.6021667 -1.3471230 2.5514563 0.9746448

A slightly more formal handle can be obtained by fitting models assuming constant
difference between methods. The models are fitted, one with an item(=person) by
replicate effect, and one without:
> BA.est( cardiac, linked=TRUE )
Conversion between methods:
alpha beta sd LoA: lower upper
To: From:
IC IC 0.000 1.000 0.449 -0.898 0.898
RV -0.705 1.000 1.022 -2.748 1.339
RV IC 0.705 1.000 1.022 -1.339 2.748
RV 0.000 1.000 0.374 -0.749 0.749

Variance components (sd):

IxR MxI res
IC 0.193 0.661 0.317
RV 0.193 0.661 0.265
16 2.2 Cardiac output: Linked replicates?

●
3

3
●

● ● 2.52 ●
2.58
● ● ●
● ● ● ●
2

2
●
● ●
● ● ● ●
● ● ● ●
● ● ● ●
● ●
● ●● ● ●
● ● ● ● ●
1

1
● ● ● ● ●
● ● ●
●●
● ● ● ●●
● ●
●
●
● ●
● ● 0.60 ● ● ● ● ● 0.60
● ● ●
RV − IC

RV − IC
● ● ● ●
●● ● ● ●● ● ●● ● ●
● ● ●
0

0
●
● ● ●
●● ● ● ● ● ● ●
● ● ●
●
● ● ● ●
● ●
● ● ●
●
−1

−1
● ●
●
−1.32 ●
−1.37
●
−2

−2
−3

−3
3 4 5 6 7 3 4 5 6 7
( RV + IC ) / 2 ( RV + IC ) / 2

Figure 2.6: Bland-Altman plots of the cardiac data. The left panel is the orginal data
using replicate numbers to pair mesurements, the right is using a random permutation of
replicates for the pairing. Even if replicates are claimed to be linked, the replicates the
LoA in the right panel are not substantially wider.

> BA.est( cardiac, linked=FALSE )

Conversion between methods:

alpha beta sd LoA: lower upper
To: From:
IC IC 0.000 1.000 0.525 -1.050 1.050
RV -0.702 1.000 1.049 -2.801 1.396
RV IC 0.702 1.000 1.049 -1.396 2.801
RV 0.000 1.000 0.463 -0.926 0.926

Variance components (sd):

IxR MxI res
IC 0 0.654 0.371
RV 0 0.654 0.328

We see that there is a some variation between replicates, which we would not expect to
see if replicates were exchangeable. In the model where we (erroneously) assume
replicates to be exchangeable, we see that it is the residual variances that gets inflated.
We can check the assumptions about constant bias and constant variance across the
range of measurements by fitting a straight line to the differences as function of the
averages (using the given linking of replicates). Note that the argument reg.line=3 gives
printed output and graph annotation of the relationship between methods with three
digits after the decimal point:
> BA.card <- BA.plot( cardiac, limy=c(-2,4), reg.line=3 )

Limits of agreement:
RV - IC 2.5% limit 97.5% limit SD(diff)
0.6021667 -1.3199476 2.5242809 0.9610571
The MethComp package. 17

RV-IC = 0.422 + 0.036 (RV+IC)/2 (95% p.i.: +/-1.937)

res.sd = 0.968 se(beta) = 0.103 , P = 0.7300

IC = -0.415 + 0.965 RV (95% p.i.: +/-1.903)

RV = 0.430 + 1.036 IC (95% p.i.: +/-1.972)
4

IC = −0.415 + 0.965 RV (95% p.i.: +/−1.903)

RV = 0.430 + 1.036 IC (95% p.i.: +/−1.972)
3

● ●

● ●
2

●
●
● ● ●
● ●
RV − IC

● ●
●
● ●●
● ●
1

●
● ●
●●● ● ●
●
● ● ●
● ●
● ●
●● ● ● ●● ● ●●
0

● ● ●● ● ●
● ●
●
●
●●
−1

● ●
●

RV−IC = 0.422 + 0.036 (RV+IC)/2 (95% p.i.: +/−1.937)

−2

3 4 5 6 7

( RV + IC ) / 2

Figure 2.7: Bland-Altman plot of the cardiac data with a fitted regression line.

There is a some indication that the variance is not constant, but seen from the figure
it does not seem alarming, it presumbaly hinges on the 6 points to the far left of the
plot. An informal test of this can be obtained by using the function DA.reg, which
regresses the Differences between methods on the Averages, and additionally regresses the
18 2.2 Cardiac output: Linked replicates?

absolute values of the residuals from this analysis on the averages, so as to give an
indication as to whether the residual standard deviation depends linearly on the mean:
> DA.reg( cardiac )

Conversion between methods:

alpha beta sd.pred beta=1 sd.|A=4.8 slope(sd) sd.=K
To: From:
IC IC 0.000 1.000 NA NA NA NA NA
RV -0.415 0.965 0.951 0.730 0.943 0.168 0.021
RV IC 0.430 1.036 0.986 0.730 0.943 0.168 0.021
RV 0.000 1.000 NA NA NA NA NA

If we fit a variance component model using BA.est as before, we can explore what effect
it has on the repeatability (the prediction of a method from itself) if we include the
variation between replicates or not:
> BA.est( cardiac, linked=TRUE, IxR.pr=FALSE )

Conversion between methods:

alpha beta sd LoA: lower upper
To: From:
IC IC 0.000 1.000 0.449 -0.898 0.898
RV -0.705 1.000 1.022 -2.748 1.339
RV IC 0.705 1.000 1.022 -1.339 2.748
RV 0.000 1.000 0.374 -0.749 0.749

Variance components (sd):

IxR MxI res
IC 0.193 0.661 0.317
RV 0.193 0.661 0.265

> BA.est( cardiac, linked=TRUE, IxR.pr=TRUE )

Conversion between methods:

alpha beta sd LoA: lower upper
To: From:
IC IC 0.000 1.000 0.525 -1.050 1.050
RV -0.705 1.000 1.022 -2.748 1.339
RV IC 0.705 1.000 1.022 -1.339 2.748
RV 0.000 1.000 0.463 -0.926 0.926

Variance components (sd):

IxR MxI res
IC 0.193 0.661 0.317
RV 0.193 0.661 0.265

The former is for the situation where we consider the variation between replicate
measurements as a part of the repeatability conditions (even if the replicates are linked),
the latter where we consider the variation between replicates to be irrelevant to the
assessment of repeatability. However there is not much indication of linked estimates,
since the other two variance components are virtually unchanged between the two
analyses, and hence the predictions between methods based on the two approaches will
be the same.
The MethComp package. 19

2.3 Systolic blood pressure: Linked replicates by

two methods
We first load the systolic blood pressure data from the MethComp package.
> data( sbp )
> sbp <- Meth( sbp )

The following variables from the dataframe

"sbp" are used as the Meth variables:
meth: meth
item: item
repl: repl
y: y
#Replicates
Method 3 #Items #Obs: 765 Values: min med max
J 85 85 255 74 120 228
R 85 85 255 76 120 226
S 85 85 255 77 135 228

> str(sbp)

Classes 'Meth' and 'data.frame': 765 obs. of 4 variables:

$ meth: Factor w/ 3 levels "J","R","S": 1 1 1 1 1 1 1 1 1 1 ...
$ item: Factor w/ 85 levels "1","2","3","4",..: 1 2 3 4 5 6 7 8 9 10 ...
$ repl: Factor w/ 3 levels "1","2","3": 1 1 1 1 1 1 1 1 1 1 ...
$ y : num 100 108 76 108 124 122 116 114 100 108 ...

> plot( sbp )

Note:
Replicate measurements are taken as separate items!

The resulting plot is shown in figure 2.8, clearly shows that the two manual
measurements are in much closer agreement than any of them are with the automatic.
plot.Meth pairs replicates according to their numbering and treat them as separate
items, so the plots fail to take the dependence of observations nto account.
We want to restrict our attention to the comparison of the two manual methods, but
using the replicate measurements.
In this context it is important that we recognize whether the replicates are linked
across the two methods or not. In this case they are, i.e. replicates are not exchangeable
within methods and items.
> par( mar=c(3,3,3,3), mgp=c(3,1,0)/1.6 )
> sbp <- subset( sbp, meth %in% c("J","R") )
> str( sbp )

Classes 'Meth' and 'data.frame': 510 obs. of 4 variables:

$ meth: Factor w/ 2 levels "J","R": 1 1 1 1 1 1 1 1 1 1 ...
$ item: Factor w/ 85 levels "1","2","3","4",..: 1 2 3 4 5 6 7 8 9 10 ...
$ repl: Factor w/ 3 levels "1","2","3": 1 1 1 1 1 1 1 1 1 1 ...
$ y : num 100 108 76 108 124 122 116 114 100 108 ...

> BA.plot( sbp )

Limits of agreement:
R - J 2.5% limit 97.5% limit SD(diff)
-0.08627451 -4.60761840 4.43506938 2.26067194
20 2.3 Systolic blood pressure: Linked replicates by two methods

A slightly more informative plot can be obtained by explicitly regulating the

y-dimension of the plot by the argument ymax=:
> BA.plot( sbp, ymax=15 )

Limits of agreement:
R - J 2.5% limit 97.5% limit SD(diff)
-0.08627451 -4.60761840 4.43506938 2.26067194

The resulting plots are shown in figure 2.9.

In order to properly partition the variance and produce limits of agreement or a
translation between the two observers, we should fit the relevant variance component
model, assuming linked replicates:

ymir = αm + µi + air + cmi + emir , air ∼ N (0, ω 2 ), 2

cmi ∼ N (0, τm ), 2
emir ∼ N (0, σm )

Since we only have two methods, we cannot identify separate variance components τ1
and τ2 , so we are forced to assume that τ1 = τ2 , hence the use of pdIdent and not
pdDiag in the specification of the matrix effects (i.e. the method by item interactions).
The model above is fitted to the dataset by:

●● ●
●
100 150 200 100 150 ● 200
●

●●

●
●
● ●
●● ●●
50
● ●
●
●● ● ●
● ●●●● ● ●
● ● ● ●●
● ●● ● ●● ●● ●
● ●
●
●● ● ●●● ● ●
●●●
●
● ● ● ●●● ●
● ●●
●
●
●●
● ●
●● ●
●● ●
● ●
● ●● ●● ● ●● ●
●●● ● ●●●●●● ●● ● ●● ● ● ●
● ● ● ●●● ●●●
●● ●
●
●
●●
● ●● ●●●●● ●●● ● ●
● ● ● ●● ●●●●
●●
●●● ●● ● ● ●● ● ● ● ●
● ●●● ● ●●●● ●● ● ●●● ●● ●● ●
●

J ● ●
● ●●
●
●
●●
●●●
● ●●
● ●●●
●
●
●
●
●
●●
●●
●
●●
● ●
●
●●
●●●●●●●
●
●
●● ●●●
●
●●
●●
●●
●
●
●
●●
●
● ●●●●●
●
●

●
●
●●
●●
●●●
●●
●●
●
●
●●
●

●
●
●●
●●
●
●
●
●
●● ●● ●● ●●●
●●
●
● ●
●●● ●
●●●●●
●● ●
●● ●●
●●
●
● ●● ● ●
●
● ● ●

●
●●
●●●●●● ●● ● ● ●●●
●●
●● ●●●
●
● ●●
● ●● ● ● ●●

●
● ●● ●
●
● ●
● ●
●
●
● ●
●
● ● ●
●
● 0
●
●

●
●
−50
●
●
●
●

●●
●
●● ●
●● ●
●●
● ●
200 ●●●
● ●●
●
50
●
●● ●
●● ● ●
●● ● ● ●● ● ●
● ● ● ● ● ● ● ●
●● ●● ●
●● ● ● ●● ●●●●● ● ●● ● ●●
●●●● ● ●● ●● ● ●
● ●
● ●● ●● ●
● ●
●●
●● ●●●● ● ● ●● ●
● ●● ● ●
●● ●●● ●●
●●
●●
● ●
●
●●●●●●
● ●● ●● ●
●●●● ● ● ●● ●
●●●● ●●● ●
● ● ●●●
●●
●●●
●●
●●●●●
●
●● ● ●
●●●●
● ●●●
● ●● ● ● ●
● ● ● ●● ●●● ●
●
●● ● ●●● ● ●● ●
●

R
●●
● ●● ●
● ● ● ●● ● ● ●●●● ● ●● ● ● ●● ● ●● ● ● ● ●● ● ●
150 ●●
●●
●●
●●● ● ●
●
●● ● ●
●● ● ●
● ● 0
●●●●
●●●● ● ● ●● ● ● ●●●
●●
● ● ● ●
●●●
●●●
●●● ● ●
● ● ●
● ● ●●
●● ●
●
●●●●
●●
●●
●●●
●● ● ●●
● ●
●
●●
●●●●
●
●●
●●●●●
●●
●●
●●
100 ●●
●●
●●●●●●●● −50
● ●●
●
●●●
●●
●●●
●●

●● ● ●●
● ● ●● ● ●●
● ●
● ●
● ● ● ● ● ●
● ● ● ●
● ● ● ●
● ● ● ● ● ●
200 ● ●
●●
● ● ●
●●
●

●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ●
● ● ●●● ●● ● ●● ● ●● ●
● ● ● ● ● ●
●● ● ● ●
● ● ● ● ●● ● ● ● ● ● ●
● ● ● ●●● ● ● ● ● ● ●●● ● ●
● ● ● ●●● ●●●●● ●● ●●
●

S
● ●●●●●● ●●● ● ●
● ●
●
● ● ●● ● ●
150 ●● ●● ●● ● ● ● ●● ● ● ●●
●
●
●●●
●● ● ● ●
●● ●●● ●●●●● ● ●●
● ● ● ●● ●●
● ●● ● ●●
● ● ● ● ●●
●●
●
●●● ●● ●
●
● ●● ●● ●
● ● ●● ●● ● ● ●●● ●● ●
● ●● ● ●● ● ●
●●●●●● ●●●● ●● ●
●●●●
● ●● ●
●●
● ●●
●
●● ●● ● ● ●●●●
●
● ● ●● ●●●●
● ●●● ●● ● ●
● ●●
● ●●
● ●●●● ●●●● ● ●
●●●● ● ●
● ●●●
●
●
●●
●●
●
●
●● ● ● ● ● ●● ●●
● ●
●●● ●●●
●●● ●
●● ●
●●
● ●●
●●● ● ●
●●●●
●●● ●● ●
●
● ●● ● ● ●
●●●●
●● ●● ●●●●●●● ●
●● ● ●● ●●●● ●
●●● ● ● ●
●●●
● ●●●●
● ●● ● ●● ● ● ●● ● ● ● ●
100 ● ●●
●
●● ● ●
●● ●
●
●● ● ●
●●
● ● ● ● ● ●

● ●

100 150 200 100 150 200

Figure 2.8: Graphical overview of the sbp data. The methods J and R are two human
observers, whereas method S is an automatic device.
The MethComp package. 21

> m1 <- lme( y ~ meth + item,

+ random=list( item = pdIdent( ~ meth-1 ),
+ repl = ~ 1 ),
+ weights = varIdent( form = ~1 | meth ),
+ data = sbp )
> m1
Linear mixed-effects model fit by REML
Data: sbp
Log-restricted-likelihood: -1163.807
Fixed: y ~ meth + item
(Intercept) methR item2 item3 item4 item5
103.47872449 -0.08627451 5.82189382 -22.17810618 1.89313629 13.45293925
item6 item7 item8 item9 item10 item11
25.82189382 5.82189382 7.96437876 2.92875753 -2.54706075 0.78627258
item12 item13 item14 item15 item16 item17
10.85751506 8.19084839 1.89313629 1.29771210 15.29771210 -2.10686371
item18 item19 item20 item21 item22 item23
14.63104543 33.29771210 43.29771210 53.36895457 40.17810618 66.03562124
item24 item25 item26 item27 item28 item29
60.48856049 39.22646963 27.22646963 37.59542419 45.22646963 115.89313629
item30 item31 item32 item33 item34 item35
95.66666667 -15.14248494 14.85751506 18.63104543 22.03562124 15.89313629
item36 item37 item38 item39 item40 item41
-12.70228790 4.19084839 105.29771210 25.00000000 30.55980296 -9.07124247
item42 item43 item44 item45 item46 item47
-8.10686371 17.52418172 58.55980296 -2.17810618 24.26209086 11.22646963
item48 item49 item50 item51 item52 item53
31.08398468 49.22646963 -11.80915161 52.63104543 -1.44019704 1.15522715
item54 item55 item56 item57 item58 item59
-4.47581828 -24.17810618 1.59542419 5.45293925 75.45293925 52.92875753
item60 item61 item62 item63 item64 item65
35.96437876 93.52418172 -11.73790914 24.26209086 36.92875753 33.59542419
item66 item67 item68 item69 item70 item71
10

●
15

●
10

● ●
5

●
● ● ●● ● ● ●
4.44
●
5

● ●●● ● ●●●●●●●● ● ●●● ●●●●● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● 4.44

● ●●● ● ●●●●●●●● ● ●●● ●●●●● ● ● ●● ●● ● ● ●
R−J

● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ● ● ●● ●●
0

−0.09
R−J

● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ● ● ●● ●● −0.09

0

●●● ●● ●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●● ● ● ● ● ●●● ●● ●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●● ● ● ● ●

● ● ● ● ● ●
−4.61
−5

● ● ● ● ● ●
●
−4.61
−5

●
●
−10

● ● ●
−10

−15

● ●
100 150 200

100 150 200 (R+J)/2

(R+J)/2

Figure 2.9: Bland-Altman plot of the sbp data. Replicates are linked between methods,
so the single replicates in the data has been used as single measurements when doing the
Bland-Altman plot. Measurements from the same person are joined by thin lines. The
only difference between the two plots is the scaling of the y-axis.
22 2.3 Systolic blood pressure: Linked replicates by two methods

53.82189382 29.59542419 9.52418172 13.22646963 17.52418172 112.63104543

item72 item73 item74 item75 item76 item77
30.55980296 53.89313629 -19.44019704 70.48856049 75.59542419 13.22646963
item78 item79 item80 item81 item82 item83
15.29771210 4.55980296 6.26209086 36.78627258 4.78627258 6.92875753
item84 item85
-2.10686371 12.48856049

Random effects:
Formula: ~meth - 1 | item
Structure: Multiple of an Identity
methJ methR
StdDev: 0.2483701 0.2483701

Formula: ~1 | repl %in% item

(Intercept) Residual
StdDev: 5.932962 1.48587

Variance function:
Structure: Different standard deviations per stratum
Formula: ~1 | meth
Parameter estimates:
J R
1.000000 1.122211
Number of Observations: 510
Number of Groups:
item repl %in% item
85 255

Now, the output from lme is pretty difficult to read, but the residual standard deviations
are σJ = 1.485870 and σR = 1.485870 × 1.122211 = 1.6674599, whereas τ = 0.2483701
(largely negligible) and ω = 5.932962, by far the largest variance component. Also from
the output we get the difference between methods R and J to be −0.08627451.
An easier way to get the relevant estimates is to use the wrapper BA.est, where the
only necessary specification is the dataset (assuming that columns meth, item, repl and
y are present) and whether replicates are linked across methods:
> BA.est( sbp, linked=TRUE )

Conversion between methods:

alpha beta sd LoA: lower upper
To: From:
J J 0.000 1.000 2.101 -4.203 4.203
R 0.086 1.000 2.261 -4.435 4.608
R J -0.086 1.000 2.261 -4.608 4.435
R 0.000 1.000 2.358 -4.716 4.716

Variance components (sd):

IxR MxI res
J 5.933 0.248 1.486
R 5.933 0.248 1.667

Which is identical to the quantities we fished out of the lme output. Actually BA.est fits
exactly the model we fitted, and then extracts the quantities that we are interested in.
The limits of agreement
√ between the two manual observers is then for R−J
−0.0863 ± 1.96 × 2 × 0.2482 + 1.4862 + 1.6672 = (−4.51, 4.34), i.e. on average they
agree, but in order to be sure to enclose 95% of all differences we need an interval
approximately as 0 ± 4.5 mmHg.
The MethComp package. 23

One way of seeing the lack of exchangeability is to make the overview plot using a
random permuation of the replicates. If replicates were truely exchangeable within
methods the plot would look similar when permuting the replicates — and it does not!
For completeness we reload the data to get observations by all three methods
included, and then make overview plots after random permutation of replicates within
(method,item):
> data(sbp)
> sbp <- Meth( sbp )

The following variables from the dataframe

"sbp" are used as the Meth variables:
meth: meth
item: item
repl: repl
y: y
#Replicates
Method 3 #Items #Obs: 765 Values: min med max
J 85 85 255 74 120 228
R 85 85 255 76 120 226
S 85 85 255 77 135 228

> str(sbp)

Classes 'Meth' and 'data.frame': 765 obs. of 4 variables:

$ meth: Factor w/ 3 levels "J","R","S": 1 1 1 1 1 1 1 1 1 1 ...
$ item: Factor w/ 85 levels "1","2","3","4",..: 1 2 3 4 5 6 7 8 9 10 ...
$ repl: Factor w/ 3 levels "1","2","3": 1 1 1 1 1 1 1 1 1 1 ...
$ y : num 100 108 76 108 124 122 116 114 100 108 ...

> plot( perm.repl(sbp) )

Note:
Replicate measurements are taken as separate items!

The two resulting plots are shown in figure 2.10.

The analysis should be based on a model where a random item by replicate effect is
included to accomodate the linking of replicates:
> BA.est( sbp, linked=TRUE )

Conversion between methods:

alpha beta sd LoA: lower upper
To: From:
J J 0.000 1.000 2.305 -4.610 4.610
R 0.086 1.000 2.272 -4.459 4.631
S -15.620 1.000 20.326 -56.272 25.032
R J -0.086 1.000 2.272 -4.631 4.459
R 0.000 1.000 2.187 -4.375 4.375
S -15.706 1.000 20.317 -56.339 24.927
S J 15.620 1.000 20.326 -25.032 56.272
R 15.706 1.000 20.317 -24.927 56.339
S 0.000 1.000 12.930 -25.860 25.860

Variance components (sd):

IxR MxI res
J 5.887 0.338 1.630
R 5.887 0.001 1.547
S 5.887 18.077 9.143
24 2.3 Systolic blood pressure: Linked replicates by two methods

The substantial item by replicate interaction (IR) clearly indicates that replicates are
linked between methods.
The resulting estimates from this model gives limits of agreement for R−J based on
the method by item and the residual variances:
√
−0.0863±1.96× 0.33852 + 0.00112 + 1.63012 + 1.54672 = −0.0863±4.4540 = (−4.54, 4.37)

which is in agreement with the limits computed based on the simplistic way of taking
replicates as items — a procedure wich is actually close to correct if replicates are linked.
Alternatively this could be formulated as a 95% prediction interval for R given a
measurement by J, yJ , which would be

yR |yJ = yJ − 0.0863 ± 4.4540 = yJ + (−4.54; 4.37)

The above analysis is based on the correct analysis of the entire dataset, including
the information from the machine measurement S. If we fit the model on the restricted
dataset, we of course get a common method by item interaction term because we then
only have two methods:
> BA.est( subset( sbp, meth!="S" ), linked=TRUE )

Conversion between methods:

alpha beta sd LoA: lower upper
To: From:
J J 0.000 1.000 2.101 -4.203 4.203
R 0.086 1.000 2.261 -4.435 4.608
R J -0.086 1.000 2.261 -4.608 4.435
R 0.000 1.000 2.358 -4.716 4.716

Variance components (sd):

●● ●
● ● ●
●
100 150 200 100 150 ● 200 100 150 200 100 150 ● ● 200
● ●
●
●● ●
●
●
●
● ● ●
● ●●●
●● ●●
50 50
● ● ●
● ● ● ●
● ● ●
●● ● ● ●●
● ●
● ●●●● ● ● ●
●
● ●● ●●● ●
● ● ● ●● ● ● ●●● ●● ● ●
● ●● ● ●● ●● ● ● ● ● ●●● ● ● ● ●
●● ● ●●● ● ● ● ● ● ● ●●● ● ● ● ●● ●●
● ● ● ●●● ●
● ●●●
●●●●
● ● ●● ● ● ●● ●● ● ●● ● ● ● ●
● ●●●
● ●●
●
●●●●
●●
●
●●
●
●
●
●●● ● ●●● ●
● ●
● ● ● ●
●●● ● ● ● ●● ●●
●●●●● ●● ● ●● ● ● ● ● ● ●●● ● ● ● ●
● ● ● ●●● ●●
●● ●●
● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●
● ●●●●
●●●●●● ●
●
● ●●
●●●
●●● ●● ● ●
● ●●●
● ●● ●●●●● ●●● ●● ● ● ● ●●●●●● ● ●●● ● ●● ● ● ● ●
● ● ● ●● ●●
●●● ●●
●
●●
●
● ●●●
● ●● ● ● ●●
● ● ●●●●●● ●
●● ●
●
●
●
● ● ●●
●●●●● ●●● ●● ● ●
●● ● ● ● ● ●● ●●●●● ●
● ● ● ●● ●● ●
●
●● ● ● ●● ●● ● ●● ●

J ● ●
● ●●
●
●
●●
● ●●
●
●●●●●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●● ●●●
●
●
●●●●●●●
●
●●
●●
●●
●
●
●
●●
●
● ●●●●●
●
●

●
●
●●
●●
●●●
●●
●
●
●
●●●
●

●
●
●●
●●
●●
●
●● ●● ●● ●●●
●
●●
● ●●● ●
●
● ●
●●●●●
●● ●
●● ●●
●●
●
● ●● ●
●● ● ●
● ●
●
●
●●
●●●●●● ●● ● ● ●●●
●●
●● ●●●
●
● ●●
● ●● ● ● ●●
● ●● ●
●
● ●
● ●
●
●
●
● ●
●
● ● 0 J ●
● ●
●
●●

● ●
● ●

●
●●
● ●●
●
●● ●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●

● ●●
●●●
●
●
● ● ●● ●
●
●
●
●
●
●
●
●● ●
●
●●
●●● ●●●
●
●
●
●
●
●
●
●
●●●● ●
●●
●
●● ●
●
●● ●● ● ●
●●
●
●
●●● ● ●●
● ●● ●
● ● ● ● ●● ●
●
●● ●
● ●
●
●●
●
● ●
●
●● ●
● ●
●
● ● ●● ●●●
●
●
●● ●
●
●
●●● ●
●● ●● ● ●●
●● ●● ● ● ● ● ●●
● ●
● ●●
●
●● ●
● ●
●● ●

●
●

●●
●
●
0
● ● ● ● ●● ● ● ● ●
●
●
● ● ●
● ●
● ● ● ●
●

●
●
−50 ●
−50
● ●
● ● ●●
●
●
●

●● ●
●
● ● ●
●● ● ●● ●
●● ● ●
●●
● ● ●● ● ●
200 ●●●
● ●●
●
50 200 ●●● ● ● ● 50
● ● ● ●
●● ●
●● ● ● ● ● ●
● ●
●●● ●● ●
●● ● ● ●● ● ● ● ●
● ● ●
● ● ● ● ● ● ●
●
● ● ●● ● ● ●● ●
●● ●● ● ● ● ● ● ● ●● ●
● ● ● ●●
●● ● ● ●● ●●●●●● ● ●● ●● ● ●● ● ●● ●●● ●● ●● ●● ●
● ●● ●● ●
●●●● ● ●● ●● ● ●●●●●
●●
●● ●
●● ● ● ●● ● ● ● ● ● ● ●● ● ●
● ●● ●
● ● ●
●
●●
● ●●
● ●●● ●
●●●
●● ●●● ●●●●●●
● ●●● ●● ●●● ● ● ● ●● ● ● ● ●● ● ●● ● ●●
● ●●
● ●
● ●
●● ●●● ● ● ● ● ●
● ●●● ● ●● ●●● ● ● ●●
● ● ● ●●● ●
●●●●
●
●
● ●●●
●
●
●
●
●●●●
●
●●
●
●
●●●
●●
●
● ●●●
● ●
●
●
●●●●●
● ●● ● ● ● ●●● ●
● ● ● ●
●
●●
●●
●
●
●●
●●
●●
● ●● ●●● ●
● ●
●
●●
●● ●
●● ● ●
●●● ● ● ●● ●● ●
● ● ●● ●●●● ●● ●●● ● ●● ●
● ●
●●● ●
● ● ●
●●●●●● ● ●
● ●
● ● ● ● ● ● ● ●
●

R R
●● ● ●● ●●●
●● ●●● ● ●
●●●● ● ●● ● ● ●●●● ● ●
● ● ●● ● ●●
● ● ● ● ● ●● ● ● ● ●●● ● ●● ● ● ● ●●● ● ●
150 ●
●●
●● ●
●● ● ●
●
●●
●● ●
●
●
● ● 0 150 ●
●●
●●
●
● ●●
●
● ●
● ● ● ●●●
●
● ●
●●
●● ● ●
● ● ●● 0
●●
●● ● ● ●● ● ● ●● ● ●●●
●●
● ● ● ●●
●●●●
●● ●
●
● ● ●●● ● ● ● ●● ● ●●
●
●●
●●●●
● ● ● ● ● ● ●●● ●● ●● ● ●
●● ● ●
● ● ●
●
● ● ●●●
●● ● ● ●●● ●● ● ● ●● ● ●
● ●
●●●●
●● ●● ●
● ● ●●●
●●● ●●●● ●
●●
●●● ● ●●●
●●●
●●
●● ●
●● ●
●●●● ●
●●●● ●●
●●
●●●●●
●● ● ●●●● ● ●●●●
●●● ● ●
●
●● ●●
●
●●●
●●● ●●
●●●●
●●●● ● ●
●●● ●●
●●●●●●
100 ●●●●
●
●●
●●
●●●● −50 100 ●●
●●
● ●● ●●
●● ●●● ●
●● ●
●●●● −50
● ●●● ● ●● ● ● ●
●●●
● ● ●●●●
●●● ●
●● ●● ● ●

●● ● ●●
● ● ●● ● ●●
● ● ●● ● ●●
● ●● ● ●●
●
● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ●● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ●
200 ● ●
●●
● ● ●
●●
●
200 ● ●
●●
● ●
●
●
●●
●

●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●
● ● ● ● ● ●●
● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ●
● ● ● ● ● ● ● ● ● ● ●●
● ● ●●● ●● ● ●● ● ●● ● ●● ●
● ●
●● ● ● ● ●●
● ●
● ● ● ● ● ● ● ● ● ●●
●● ● ● ● ●●
● ● ● ●●
● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●
● ● ●●● ● ● ●●● ● ● ●● ● ● ● ●●●
● ● ●● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ●● ●

S S
● ●●● ●●● ● ● ●●● ●● ● ●●● ● ●●●●
●
●●●
● ● ● ● ●●●●
● ● ● ●● ● ● ●● ●
●●● ● ● ●● ●●
●●
● ●●
150 ●● ●● ●● ● ● ● ●● ● ● ●●
●● 150 ●
●● ●● ● ● ●●●
●
● ●
●● ● ● ● ● ● ●●
● ●
●●●
●● ● ● ●
●● ●●● ●●●●● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ●
●
● ● ● ● ● ● ●
●● ●● ● ●● ● ● ●● ● ●●●●● ●●● ●
●
●●
● ● ●
●●● ● ●● ● ●
●
● ● ● ●
●●●●● ●
●● ●
● ●● ●● ●
● ● ●● ●● ● ● ●●● ●● ●
● ●● ● ●● ● ● ● ●●●●● ● ●● ● ● ● ●
●● ● ●●●●●
●
● ●● ●
● ● ● ●
●●●●●● ●
●●
● ●●
●
●● ●
●●●●
●●
●●●
●●
●●
●
●
●● ● ● ●
●●● ● ●
●●● ● ●
● ● ●● ●
●●●●
● ● ●
● ● ● ●●
●
●
● ● ●●
●●
●●
●●●●
● ●●● ●● ● ●●
● ●●●
● ●●●● ●●●● ● ● ●●
●●
●●●●
● ●● ● ● ●
●●●●
●●
●●
●●●●
●● ● ● ●●
● ●●●
● ●
● ● ●● ●
●●●●●●●
● ● ● ● ● ●● ●●
●● ●● ●●● ● ● ●● ●● ●●●●●●●
●
●●
●●
●● ● ● ● ● ● ●●● ●●●
●●● ● ●●●●
●
● ● ●● ● ● ●●●
●● ● ●
●● ●●● ●
●
●● ●
●● ●●● ●
● ●●
●●● ●● ● ●●
● ●●●● ● ●●●
●
●●●●
●●
●●● ●●● ●
● ●● ●●● ● ● ● ●●●●
● ●● ●● ●●● ● ●
●● ● ● ●●●
●
●● ●● ●●●●●●● ●
● ● ●● ●●●●
●●●●●● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ●● ●●● ● ● ●
●●●
● ●●
● ● ● ●●● ●● ● ●●
● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ●● ● ●● ● ●
100 ● ●●
●
●● ● ●
●● ●
●
●● ● ●
●●
100 ● ● ●
●
●● ●● ● ● ●
● ●● ●
●●
●
● ● ● ● ● ● ● ● ● ● ● ●

● ● ● ●

100 150 200 100 150 200 100 150 200 100 150 200

Figure 2.10: Graphical overview of the sbp data; the left panel with the original replicate
numbers used for matching; the other with replicates permuted randomly within methods.
The MethComp package. 25

IxR MxI res

J 5.933 0.248 1.486
R 5.933 0.248 1.667

Based on these estimates we get the limits of agreement for R−J to be:
√
−0.0863 ± 1.96 × 2 × 0.24842 + 1.48592 + 1.66742 = 0.0863 ± 4.4313 = (−4.52, 4.35)

i.e. effectively the same as before, based on all three methods. Again these limits are
those computed by BA.est.
Bibliography

26