1

Detection of Outliers in Bioequivalence Studies Data

Analysis with Williams Design
A. Rasheed
1
*, S. Junaid
2
and T. Ahmad
3

1
Department of Research, Dow university of health sciences, Karachi,
2
Department of Statistics,
University of Karachi, Karachi
3
Center for Bioequivalence Studies at ICCBS, University of Karachi,
Karachi, Pakistan. *[email protected]
Background: Drug Regulatory agencies all over the world generally discourage exclusion of
outliers in a BE (BE) study; on the other hand Good Statistical Practices requires it. If the
decision rules for identifying the outliers are clearly mentioned before the start of the study
and laid down in protocol by the responsible biostatistician in collaboration with clinicians,
the problem of outliers can be dealt smartly without jeopardizing the whole study for
redoing. The purpose of this article is to introduce procedure for reliably detecting outlier
subject(s) with Williams design.
Experimental: Literature review reveals many different methods for the detection of outlier
values in BE studies; most of them are for BE of two treatments. For BE studies with more
than two treatments use of Williams design seems imperative; but inclusion and deletion of
outlying subjects may lead to profound effect on conclusion of BE which in turn may be
dangerous for the health. The suggested method is an adjustment to a previously
introduced method using exploratory data analysis technique such as principle component
analysis and Andrews curves.
KEY WORDS: Bioequivalence; outliers; Williams design; Exploratory data analysis; Principal
component; Andrews curves
1 INTRODUCTION
1.1 Back Ground and Rationale
John Tukey [7] proposed exploratory data analysis (EDA) a methodology for data analysis
which generally employs a variety of techniques most of them are graphical. These methods
are useful in detecting outliers, uncover underlying structure, maximize insight into a data
set, extract important variables and determine optimal factor settings.
Enachescu and Enachescu [2] discussed the EDA technique such as Andrews curves and
Principal Component Analysis (PCA). Using PCA Enachescu and Enachescu [2] mentioned
that first two principal axes in PCA span one such plane, providing a projection such that
the variation in the projected data is maximized over all possible 2-D projections and applied
these techniques to identify the outlying subjects in 2x2 crossover BE trial. In present work
we extended these techniques to Williams design; a special variety of Crossover/Latin
square designs. With the help of our proposed approach, gathering information regarding
outliers subjects and their identification in a BE data with more than two formulations
becomes easily possible.
2
1.2 WILLIAMS DESIGN

In crossover designs if each formulation appears in same number of times within each
sequence is called uniform within sequence and if each formulation appears the same
number of times within each period than is called uniform within period. A crossover
design is called uniform if it is uniform within sequences and within periods. A Latin square,
in which every treatment is occurred once and only once in each row and each column yields
uniform crossover designs. In a balanced design, each of the treatments occur the same
number of times in each period and the number of subjects who receive treatment i in one
period and treatment j in the next period is the same for all
i j
[5].
Williams [4] introduced a crossover design in which every treatment follows every other
treatment the same number of times called Williams design that is the treatments are
balanced over prefix and suffix. These Williams designs require fewer subjects than those
based on the complete sets of orthogonal Latin squares [3]. In Williams design when the
number of formulations are even than balance can be achieved by a single Latin square
design, but when the number of formulations are odd than two Latin square designs are
needed.
2 EXPERIMENTAL TECHNIQUES

In this present work we used two EDA techniques, Andrews curve and PCA to ease
the problem of detecting outliers in BE studies with more than two treatments.
Modified z-scores method commonly used method for outlier detection also used
here to insert the fictitious outliers in original data set. In Modified z-scores method
subject having absolute z-scores greater than 3.5 are labeled as potential outliers.
EXPLORATORY DATA ANALYSIS TECHNIQUES
2.1 Andrews curves
Andrews [9] introduced a method of visualization for multivariate data. Each
multidimensional data point is mapped into a periodic function
1 2 3
( ) / 2 sin cos ......
X
f t X X t X t = + + +
This graphical approach displays a point in multidimensional space by a two-dimensional
curve using the function ( )
x
f t given above in the interval < < . The advantage of
this method is that it allows the inclusion of many dimensions. A collection of
multidimensional points, that is, a multivariate data set, is displayed as a group of curves. In
this method numbers of variables are unlimited. These curves are dependent on the order of
the variables. Lower frequency terms (i.e., those that are first in the sum given in the above
equation) exert more influence on the shape of the curves we can get more information
about data by re-ordering the variables and viewing the resulting plot. Observations showing
quite apparent different curves are considered as outliers.

3
2.2 Principal component analysis

The objective of PCA is to discover or to reduce the dimensionality of the data set and
identify new meaningful underlying variables.
In PCA number of (possibly) correlated variables are transformed into (smaller) number of
variables which are uncorrelated called principal components. Large amount of variability is
accounted by the first PC and each succeeding PC accounts for as much of the remaining
variability as possible.
PCA can be performed either by using a covariance matrix or correlation matrix both
matrices are calculated from the data matrix, if one is using correlation matrix so firs
variables should be standardized.
2.2.1 Eigen analysis: Eigen analysis is a mathematical technique used in the PCA, in this
technique Eigen values and Eigen vectors of a square symmetric matrix with sums of squares
and cross products are calculated. The eigen-vector associated with the largest Eigen-value
has the same direction as the first principal component. The Eigen-vector associated with
the second largest Eigen-value determines the direction of the second principal component.
X is p n the data matrix where(
p
= number of variables and
n
= number of observation)
, is covariance matrix obtained from the data matrix X , and Z is the standardized data
matrix , R is correlation matrix obtained from the data matrix Z
.

i
is called Eigen value denotes the variance of the i-th PC (i.e., ( )
i
Var ith PC = ) that can
be calculated by setting 0 R I = Where I is the identity matrix.
( ) i
i
U Z

= is called the
i-th PC where
( ) i
is denote the i-th eigen vector that can be calculated by setting
( )
( )
0
i
i
R I = where
(1) (2) (3) ( )
...
p
=

and each of
( ) i
is defined as
( ) ( ) ( ) ( ) ( )
1 2 3
...
i i i i i
p

=

with

p

2 2 2 2
( ) ( ) ( ) ( ) ( ) ( )
1 2 3
= + + +... + =1
The sum of the Eigen-values equals the trace of the square matrix (i.e.,
( )
1
p
i
i
tr
=
=

) and
the maximum number of Eigen-vectors equals the number of rows (or columns) of this
matrix. Enachescu and Enachescu [2] has mentioned that For normally distributed
observation
i
i
U

are independent
2
1, j
variables. Consider
2
1,
1
p
i j
i

=

the weighted sum of

square distance to zero of the projected data into principal factorial plane, with
2
1,
1 1
p p
i j i
i i
E p
= =
| |
= =
|
\


and
2 2
1,
1 1
2
p p
i j i
i i
Var
= =
| |
=
|
\

. Now the Observations with a
square distance greater than m (the rule of 2 ) may be considered as outliers where
2
1
2 2
p
i
i
m p
=
= +

. [2]
4

3 PROPOSED NEW TECHNIQUE
EDA techniques, Andrews curve and PCA have been used to detect outliers in BE
studies with two treatments, here we advocated Andrews curve and PCA for
gathering information regarding outliers in Williams design with a reference (R) and
two treatments formulations (T1 and T2).
Andrews curve function is defined as
1 2
( ) / 2 sin cos 1, 2,3... 12 f t R T t T t t i = + + < < =

Each observation is projected onto a set of orthogonal basis functions represented
by sines and cosines and then plotted. Thus, each sample point is now represented
by a curve. Observations with identical curves show the possible outlying subjects.
In PCA, X is p n the data matrix where (
p
= 3, number of formulations and
n
=12, number
of observation for each formulation) , is covariance matrix obtained from the data matrix
X , and Z is the standardized data matrix , R is correlation matrix obtained from the data
matrix Z
.

1 2 3
, and are Eigen values denoting the variances of first, second and third PC
respectively.
3
2
1,
1
i j
i

=

the weighted sum of square distance to zero of the projected data

into principal factorial plane, with mean 3 p = and variance
3
2
1
2
i
i

. Now the Observations

with a square distance greater than m (the rule of 2 ) may be considered as outliers
where
3
2
1
3 2 2
i
i
m
=
= +

.
4 APPLICATION AND VALIDATION OF PROPOSED TECHNIQUE
In the present work we selected a data set of Areas Under the Curve; AUC from a
bioequivalence study reported by Purich [8]. In the study twelve healthy volunteer were
employed to investigate the bioequivalence of two test tablets formulations in comparison
to a reference solution. Chow and Liu [6] mentioned that no assignment of sequences and
periods was given by Purich [8]. Thus for the purpose of illustration Chow and Liu [6]
assigned subject 1 and 2 to sequence 1; 3 and 4 to sequence 2; 5 and 6 to sequence 3; 7 and
8 to sequence 4; 9 and 10 to sequence 5; 11 and 12 to sequence 6. Table-I gives this AUC
data set after rearrangement of reference and period according to Williams design for
comparing the three formulations.
The numerical results given bellow are obtained with popular software system SAS.
5
Table I: AUC data set with a reference and two test formulation (domestic and European
tablets)
Sequence Subject R T1 T2
(R, T
2
, T
1
) 1 5.68 4.21 6.83
2 3.6 5.01 5.78
(T1,R, T
2
) 3 3.55 5.07 4.49
4 7.31 7.42 7.86
( T
2
,T
1,
R) 5 6.59 7.72 7.26
6 9.68 8.91 9.04
(T
2
,T
1,
R) 7 4.63 7.23 5.06
8 8.75 7.59 4.82
( T
2
,R,T
1
) 9 7.25 7.88 9.02
10 5 7.84 7.79
(R,T
1
, T
2
) 11 4.63 6.77 5.72
12 3.87 7.62 6.74

To apply the above defined EDA techniques for determining the possible outlying
subjects in Table-II formulation-wise AUC data set is presented on both linear and
logarithmic scales.
Table:II AUC(0-) data set for a reference and two test
formulations on both linear and logarithmic scales
Subject Linear Scale Logarithmic Scale
R T1 T2 R T1 T2
1 5.680 6.830 4.210 1.737 1.921 1.437
2 3.600 5.780 5.010 1.281 1.754 1.611
3 5.070 3.550 4.490 1.623 1.267 1.502
4 7.420 7.310 7.860 2.004 1.989 2.062
5 7.260 7.720 6.590 1.982 2.044 1.886
6 9.040 8.910 9.680 2.202 2.187 2.270
7 5.060 4.630 7.230 1.621 1.533 1.978
8 4.820 8.750 7.590 1.573 2.169 2.027
9 7.880 9.020 7.250 2.064 2.199 1.981
10 7.840 7.790 5.000 2.059 2.053 1.609
11 4.630 6.770 5.720 1.533 1.913 1.744
12 3.870 7.620 6.740 1.353 2.031 1.908

4.1 Andrews curves
The Andrews curves for this data set are as follows for linear and logarithmic scales

6
111111111111111111111111111111111111111111111111111 11111111111111111111111111111111111111111111111111
222222222222222222222222222222222222222222222222222 22222222222222222222222222222222222222222222222222
333333333333333333333333333333333333333333333333333 33333333333333333333333333333333333333333333333333
444444444444444444444444444444444444444444444444444 44444444444444444444444444444444444444444444444444 555555555555555555555555555555555555555555555555555 55555555555555555555555555555555555555555555555555
666666666666666666666666666666666666666666666666666 66666666666666666666666666666666666666666666666666 777777777777777777777777777777777777777777777777777 77777777777777777777777777777777777777777777777777 888888888888888888888888888888888888888888888888888 88888888888888888888888888888888888888888888888888
999999999999999999999999999999999999999999999999999 99999999999999999999999999999999999999999999999999
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11
12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12
subject 1 10 11 12 2 3
4 5 6 7 8 9
-6
-3
0
3
6
t
-4 -3 -2 -1 0 1 2 3 4
For linear scale
1 2
( ) / 2 sin cos 1, 2,3... 12 f t R T t T t t i = + + < < =
For logarithmic scale
1 2
( ) ln( ) / 2 ln( )sin ln( )cos 1, 2,3... 12 f t R T t T t t i = + + < < =

From linear and logarithmic Andrew Curves given in Figures-1a and 1b it is very
evident that there is no curve revealing distinct or outlying behavior suggesting any
subject a possible outlier.






Figure 1: Andrews curves 1a: linear scale data; 1b: logarithmic scale data

4.2 Principal component analysis
We employed the same data set given in the Table-II by using the principal
components analysis we obtained the following results with correlation matrix R.
The Eigen-values of the correlation matrix; R for the linear and logarithmic scales are
presented in Table-III

Table III: Eigen-values of the correlation matrix; R for the linear and logarithmic scales

The threshold value and the square distance to the zero for each observation are
give in Table-IV for the linear and logarithmic scales
Linear Scale Logarithmic Scale

Eigen
values Differences Proportion Cumulative
Eigen
values Differences Proportion Cumulative
1 2.0421 1.5176 0.6807 0.6807 1.9098 1.2942 0.6366 0.6366
2 0.5245 0.0910 0.1748 0.8555 0.6156 0.1410 0.2052 0.8418
3 0.4335 0.1445 1.0000 0.4746 0.1582 1.0000
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 22 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 33 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 44 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4
5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 55 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 66 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 67 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 77 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7
8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 88 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8
9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 99 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11
12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12
subj ect 1 10 11 12 2 3
4 5 6 7 8 9
- 15
- 12
- 9
- 6
- 3
0
3
6
9
12
15
18
21
24
t
- 4 - 3 - 2 - 1 0 1 2 3 4
1b
1a
7

Table IV Threshold value and the squared distance on linear and logarithmic scales
Subject Linear Scale Logarithmic Scale
Squared distance Threshold Squared distance Threshold
1 1.9756731 9.0880308 2.4626745 8.8320355
2 3.1857674 3.5669672
3 6.0492623 7.3577339
4 1.4066709 1.5539635
5 0.6469274 0.8066935
6 8.0766893 6.0561900
7 2.5784951 2.4345028
8 1.9496287 1.7079329
9 2.6834404 2.3762979
10 2.0350777 2.0368382
11 0.8310584 0.6595056
12 1.5813093 1.9807000

Any observation with square distance greater than corresponding threshold value may be
considered as outlier. As we can see that on both scales there is no observation with square
distance greater threshold.

AS evident from the above analysis no observation is found as an outlier in the above data
set. In order to verify the proposed extended EDA techniques (Andrews curves and PCA) it
was felt imperative to introduce intentionally some outlying values in the original data.
Accordingly we made some changes in the original data set by replacing few values with
fictitious (obvious outlier) values.

We replaced some original values with few fictitious extreme (very high and very low) values
for each treatment (i.e., R, T1, and T2) which were identified as outliers by certainity by
confirming them as outliers through modified z-scores method. We carried this exercise two
times.

In first instance we randomly selected the subject 3 from original data set and replaced its
values for all three treatments (5.07, 3.55 and 4.49) by fictitious values (15.2, 13.2 and
12.56) previously identified as outliers.

In second instance we selected randomly two subjects; 3 and 9 from original data set and
replaced their values by fictitious values previously identified as outliers. Subject 3 values
(5.07, 3.55 and 4.49) replaced by (15.95, 13.56 and 16.12) and subject 9 values (7.88, 9.02
and 7.25) replaced by (15.98, 14.80 and 15.7).

On these changed two data sets we applied the proposed EDA technique to confirm the
validity of the technique that whether it identify the outliers in these two data sets.





8
We are glad to report that propose technique both by Andrews curves and PCA correctly
identified subject 3 for fictitious data set 1 and subject 3 and 9 for data set 2 on linear scale
as well as on logarithmic scale. In Figure 2 (2a and 2b) the outlying subjects can be visualized
very clearly and similarly in Table-V the Threshold values and the squared distance for
fictitious data sets 1 and 2 can be seen. Both the sets in Table-V reveals squared
distance for outlying subjects are greater than the threshold values.

Figure.2. (2a)Andrews curves for the linear scale fictitious data set 1; (2b)Andrews curves for the
linear scale fictitious data set 2
Principal component analysis
Table V Threshold value and the squared distance for fictitious data sets 1 and 2
Subject Data Set # 1 Data Set # 2
Squared distance Threshold Squared distance Threshold
1 2.0180746 10.4300960 1.4646404 11.0065720
2 2.8926605 2.2777500
3 19.0678020 11.1028000
4 0.2055743 0.1342879
5 0.0750083 0.2062721
6 1.9922433 0.3078588
7 2.6616014 1.9995872
8 0.6348092 0.4616749
9 0.4082725 12.2598750
10 0.9686317 0.6733027
11 1.1422837 1.1532807
12 0.9330383 0.9586664


Conclusion
Through this work we report and recommend an extended exploratory data analysis
techniques for identification of outliers in a Williams design data set generated during
bioequivalence evaluation. In present research for identification of outliers we successfully
applied the Andrews curves and principal component analysis and validated it through z-
scores method.
111111111111111111111111111111111111111111111111111 11111111111111111111111111111111111111111111111111
222222222222222222222222222222222222222222222222222 22222222222222222222222222222222222222222222222222
333333333333333333333333333333333333333333333333333 33333333333333333333333333333333333333333333333333
444444444444444444444444444444444444444444444444444 44444444444444444444444444444444444444444444444444
555555555555555555555555555555555555555555555555555 55555555555555555555555555555555555555555555555555
666666666666666666666666666666666666666666666666666 666666666666666666666666666666666666666666666666667 777777777777777777777777777777777777777777777777777 7777777777777777777777777777777777777777777777777
888888888888888888888888888888888888888888888888888 88888888888888888888888888888888888888888888888888
999999999999999999999999999999999999999999999999999 99999999999999999999999999999999999999999999999999
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11
12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12
subject 1 10 11 12 2 3
4 5 6 7 8 9
-15
-12
-9
-6
-3
0
3
6
9
12
15
18
21
24
27
30
33
36
t
-4 -3 -2 -1 0 1 2 3 4
111111111111111111111111111111111111111111111111111 11111111111111111111111111111111111111111111111111
222222222222222222222222222222222222222222222222222 22222222222222222222222222222222222222222222222222 333333333333333333333333333333333333333333333333333 33333333333333333333333333333333333333333333333333 444444444444444444444444444444444444444444444444444 44444444444444444444444444444444444444444444444444
555555555555555555555555555555555555555555555555555 55555555555555555555555555555555555555555555555555
666666666666666666666666666666666666666666666666666 66666666666666666666666666666666666666666666666666 777777777777777777777777777777777777777777777777777 77777777777777777777777777777777777777777777777777
888888888888888888888888888888888888888888888888888 88888888888888888888888888888888888888888888888888
999999999999999999999999999999999999999999999999999 99999999999999999999999999999999999999999999999999
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11
12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12
subject 1 10 11 12 2 3
4 5 6 7 8 9
-15
-12
-9
-6
-3
0
3
6
9
12
15
18
21
24
27
30
t
-4 -3 -2 -1 0 1 2 3 4
2a 2b
9
References
1 FDA Guidance. Statistical approaches to establishing bioequivalence, 2001.
2 Enachescu D, Enachescu C. A new approach for outlying records in bioequivalence trials.
Proceedings of the 13th International Conference on Applied Stochastic Models and Data
Analysis, Vilnius, Lithuania 2009; 250-257
3 Wang B-S, Wang X-J, Gong L-K.The construction of a Williams design and randomization in
crossover clinical trials in using SAS. Journal of statistical software 2009; 29
4 Williams EJ. Experimental Designs Balanced for the Estimation of Residual Effects of
Treatments. Australian Journal of Scientific Research 1949; 2(3):149-168.
5 Jones B, Kenward MG. Design and Analysis of Cross-Over Trials. 2nd edn. Chapman & Hall:
London, 2003.
6 Chow S-C, Liu J-P. Design and analysis of bioavailability and bioequivalence studies. 2nd
edn. Dekker: New York, 2000.
7 Tukey J-W. Exploratory Data Analysis. Addison-Wesley, Reading, MA.1977.
8 Purich E. Bioavailability/Bioequivalence Regulations: An FDA Perspective in Drug
Absorption and Disposition: Statistical considerations, K.S Albert, edn. American
Pharmaceutical Association, Academy of Pharmaceutical Sciences, Washington, DC, 1980;
115-137.
9 Andrews D. Plots of high-dimensional data. Biometrics 1972; 28: 125136.