View Article Online / Journal Homepage / Table of Contents for this issue

Experimental design and multivariate calibration in the

development, set-up and validation of a differential pulse
polarographic and UV spectrophotometric method for the
simultaneous plasmatic determination of the therapeutic
metronidazolepefloxacin combination
Paola Gratteri*a and Gabriele Crucianib
a

Department of Pharmaceutical Science, Florence University, Via G. Capponi 9, I-50121

Florence, Italy
b Department of Chemistry, Perugia University, Via Elce di Sotto 10, I-06100 Perugia, Italy

Published on 01 January 1999. Downloaded on 09/09/2016 02:24:30.

Received 23rd June 1999, Accepted 3rd September 1999

Partial least squares regression (PLS1 and PLS2) and GOLPE variable selection procedures were used for the
treatment of differential pulse polarographic and UV spectrophotometric data obtained from the analysis of the
therapeutic combination of metronidazole and pefloxacin. The analytical method used for the determination was
set up using experimental design strategies (Doehlerts design, full factorial design, fractional face center cube
design, etc.) and by involving the simultaneous optimization of several responses (desirability function). Method
validation was also performed, determining accuracy, precision, linearity and range, detection and quantification
limits and robustness. The quantitative prediction abilities in determining metronidazole and pefloxacin plasma
levels of the PLS1 and PLS2 models were tested on spiked plasma samples and good results were obtained
(metronidazole, 97.5%, RSD = 4.8%, n = 3; pefloxacin, 100.6%, RSD = 3.6%, n = 3 ). The use of multivariate
calibration was particularly useful for spectrophotometric quantification because of the highly overlapping spectra
of the binary mixture.

Introduction
Pefloxacin (P) and metronidazole (M) are well-known antimicrobial agents used in combination in therapy for the
treatment of intra-abdominal infections due to aerobic (P) and
anaerobic (M) bacteria.1 This circumstance makes their plasma
determination very useful and interesting, also considering that
there are no techniques in the literature which report the
simultaneous assay of both drugs in this medium.
Among many existing analytical techniques, the structural
characteristics of the two molecules make them suitable for
polarographic determination. Moreover, the plasma concentrations of P and M allow the differential pulse polarographic
technique (DPP) to be successful in quantification. The
presence of the nitro group in M and of the carbonyl group in P
(Fig. 1) give rise to well-defined polarographic signals,2,3 which
do not overlap, at 2450 and 21410 mV vs. Ag/AgCl,
respectively.
Quantification was performed by multivariate calibration
using suitable predictive variables in order to find a better

Fig. 1 Differential pulse polarogram and structures of metronidazole (M)

and pefloxacin (P).

regression model than that obtained with a single variable (peak

height corresponding to peak potential). The partial least
squares (PLS) approach and the GOLPE selection variable
procedure were used in analysis of the experimental data.
Furthermore, spectrophotometric determination of the therapeutic MP combination was considered as an alternative
assay method both to check the polarographic results and to
verify the predictive ability of a unique multivariate calibration
model obtained with predictive variables of a different nature
(namely polarographic and spectrophotometric variables). The
use of the multivariate calibration strategy becomes, in this
case, essential for the treatment of the highly overlapped
spectral bands obtained from the MP binary mixture. This
overlapping is so severe that it does not allow the simultaneous
quantification of the drugs by means of the univariate
approach.
This paper describes the development of a differential pulse
polarographic method of quantification, set up using experimental design strategies, in order to reduce the overall analysis
time requirements involved in a step-by-step (or OVAT)
approach. Moreover, as the OVAT approach might not reveal
possible interactions existing among parameters, a multivariate
approach (i.e., data obtained from a suitably designed experiment sequence) was preferred and used, giving the advantage of
an improved statistical evaluation of the data. Furthermore, in
the case under study a multivariate approach appears to be the
only tool which allows quantification exploiting simultaneously
information derived from both UV and polarographic data.
This paper also reports the determination of suitable
parameters involved in the polarographic method validation,
i.e., in that process which proves that the analytical method is
acceptable for its intended purpose. The guidelines followed for
performing such validation were those from the Third International Conference on Harmonization (ICH3)4 and from the
Analyst, 1999, 124, 16831688

This journal is The Royal Society of Chemistry 1999.

1683

View Article Online

United States Pharmacopeia.5 With regard to the methods

utilized in this study, the parameters were accuracy, precision,
linearity and range, detection limit, quantification limit and
robustness.

Experimental

Published on 01 January 1999. Downloaded on 09/09/2016 02:24:30.

Reagents
Metronidazole was obtained from Sigma (Milan, Italy) and
pefloxacin was kindly supplied by Formenti (Milan, Italy). Both
drugs were used with no further purification.
Stock standard solutions of each drug in Milli-Q grade water
(M, 0.5 mg mL21) and in 0.05 mol L21 NaOH (P, 1 mg mL21)
were prepared weekly by weighing; solubilization was assisted
by sonicating for a few minutes. All other chemicals used were
of analytical-reagent grade.
Plasma was obtained from the Banca Militare del Sangue
(Florence, Italy).
Apparatus
The electrochemical equipment consisted of a computercontrolled polarographic analyzer (Amel Model 433) comprised of a three-electrode cell with Ag/AgCl reference
electrode, a platinum wire as auxiliary electrode and a dropping
mercury electrode as working electrode. The system was
controlled via a VAAS PC Pentium processor.
UV spectra were acquired in duplicate with a computercontrolled Perkin-Elmer (Monza, Italy) Lambda 2 double-beam
UVvisible spectrophotometer with 1 cm quartz cells.
Nemrod software v. 3.0 from LPRAI (Universit de
Marseille III, Marseille Le Merlan, France) was used for the
generation and evaluation of the statistical experimental
design.
Both spectrophotometric and polarographic data were processed using the PLS algorithm and the procedure of variable
selection included in the software package GOLPE 3.0.
Polarography
Polarographic investigations were carried out by deareating 10
mL of pH 7.4 BrittonWelford buffer (supporting electrolyte)
with highly purified nitrogen for 10 min. For each solution
examined, scans were performed twice and measurements were
made at 9 mV intervals in the range from 2250 to 21650 mV
using the differential pulse polarographic technique with the
following optimized experimental conditions: potential amplitude, 253 mV; pulse increment, 9 mV; and dropping size, 34
a.u. (arbitrary units). Peak potentials occurred at 2450 and
21450 mV for M and P, respectively.

plasma blank was prepared by mixing 3 mL of unspiked plasma

and 6 mL of CH3CN. The mixture was vortex mixed for 10 min
and the phases were separated by centrifugation at 11 000 rpm
for 4 min. The organic phase was evaporated under a stream of
pure nitrogen and the residue was reconstituted with 100 mL of
water. An 80 mL volume of this solution was added to in-cell 10
mL of BrittonWelford buffer and additions of M and P stock
standard solutions were made according to a 22 FFD scheme
with three center point.
The in-cell concentration levels of both M and P are
presented in Table 1.
Plasma samples. Plasma samples containing M and P were
prepared as described above for the plasma blank using, in this
case, 3 mL of plasma spiked with suitable amounts (ranging
from 20 to 80 mL) of M and P stock standard solutions.
Spectrophotometry. Calibration model. A plasma blank
was prepared by mixing 1 mL of unspiked plasma with 2 mL of
CH3CN. The mixture was vortex mixed for a few minutes and
the phases were separated by centrifugation at 11 000 rpm for 4
min. The organic phase was quantitatively transfered into a 10
mL calibrated flask and diluted to volume with water. The
samples used for the construction of the calibration model were
prepared as described above, adding to the organic phase 20, 40
and 80 mL (2, 0 and + level, respectively) and 20, 50 and 80 mL
(2, 0 and + level, respectively) of M and P stock standard
solutions, respectively, prior to diluting to volume with water.
Plasma samples. The samples to be analyzed were prepared
by mixing 1 mL of CH3CN and aliquots ranging from 20 to 80
mL of M and P stock standard solutions. The solutions were
diluted to 10 mL with water.

Results and discussion

DPP method development
Many parameters have to be optimized when investigating a
system during the development of an analytical procedure.
Finding the best experimental sets for variables involved in the
setting up of an analytical method is a delicate process which
may be completely inefficient if performed by considering one
separate variable at a time (COST). Fisher, in 1925, and later
numerous others (Box, Hunter, Deming, etc.) have shown that
laying out the so-called statistical experimental design (SED),
i.e., series of experiments where all factors are changed
together, provides much more information and reaches the
optimum conditions in fewer experiments than does COST
experimentation.6 Only by changing all variables simultaneously can the experimental domain be scanned efficiently and
the real optimum conditions reached.
Table 1 In-cell concentration levels of M and P

Spectrophotometry
All measurements were made at 1 nm intervals in the 200400
nm region using a 60 nm min21 scan speed.

Level

Concentration

1 mg mL21 (corresponding to a 20 mL M stock standard

solution in-cell addition).
2 mg mL21 (corresponding to a 40 mL M stock standard
solution in-cell addition).
4 mg mL21 (corresponding to an 80 mL M stock standard
solution in-cell addition).
2 mg mL21 (corresponding to a 20 mL P stock standard
solution in-cell addition).
5 mg mL21 (corresponding to a 50 mL P stock standard
solution in-cell addition).
8 mg mL21 (corresponding to an 80 mL P stock standard
solution in-cell addition).

Plasma analysis
Polarography. Calibration model. In order to not exceed
the plasma level concentration of both M and P,1 and in order to
respect the linearity range found for both drugs [also considering the constraints given by the + and 2 levels of the full
factorial design (FFD) used for the construction of the
calibration model], the following procedure was carried out. A
1684

Analyst, 1999, 124, 16831688

+
P

2
0
+

View Article Online

In the case under study, once the key factors (xi), both
instrumental and otherwise, were established affecting the
differential pulse polarographic procedure of quantification,
their number was not so high (four) to suggest the use of a
screening design. On the contrary, with few variables the study
of the response surface could be undertaken directly. It allowed
a second-order model to be obtained. This model is useful for
predicting the behaviour of the studied system (the y-dependent
variable values of the model) for any values of the x-variables
(independent variables). The true function f relating y to the
design factors is unknown but it is assumed that in the region of
interest (the selected experimental domain) f is well approximated by a low-order polynomial (second-order):
n

y = b0 +

bi xi` +

Published on 01 January 1999. Downloaded on 09/09/2016 02:24:30.

i =1

i =1

j =1
i j

bij xi x j` +

bij x 2i

(1)

i =1

The x factors considered in the DPP method were drop

growth rate, ds (arbitrary units); pulse amplitude, DE (mV);
pulse increment, pi (mV) taken as a measure of scan rate
(mV s21; pi/dt, dt = dropping time); and pH of the background
electrolyte. The choice of BrittonWelford (BW) buffer as
supporting electrolyte was derived from previous screening
experiments which showed the presence of two well-defined
signals due to M and P at 2450 and 21410 mV vs. Ag/AgCl,
respectively (Fig. 1). Consequently, there are two ys, i.e., the
responses, in which one is interested: the M and the P signals.
The response surface can be obtained by employing a design
such as central composites (CCD), BoxBehnken (BBD) or
Doehlert, i.e., a design that involves at least three levels of each
design variable to estimate the pure quadratic terms of the
model (1). In this specific case, consideration concerning the
number of experiments to perform led to the choice of a
Doehlerts design.7 The concentration levels used to carry out
the design were 2 mg mL21 for M and 4 mg mL21 for P,
corresponding to an in-cell 40 mL addition of both P and M
stock standard solutions. The use of different stock solutions did
not affect the results.
According to the philosophy of a Doehlerts design, variables
assuming 5, 7, 7 and 3 levels were, respectively, ds (experimental domain: 2040 a.u.), DE (experimental domain: 2555
mV), pH (experimental domain: 69) and pi (experimental
domain: 515 mV). The total number N of the experiments in
the Doehlert design used was 24, of which q = 4 experiments
were replications at the center of the experimental domain, p =
15 was the number of parameters in the model (1) and f
represents the number of distinctly different factor combinations in the experiments set (f = 21). The experiments were run
in a random order, adding three other experiments at the center
of the experimental domain to the one already existing in the
Doehlert design (q = 4). Such replications allow the calculation
of a model-independent estimate (s12) of the experimental
variance s2 (SSpe, sum of square due to pure error). Another
estimation (s22) of this parameter is the sum of squares caused
by lack of fit, SSlof. These two sums of squares, SSpe and SSlof,
can be used, in the ANOVA table, to estimate the significance
of the lack of fit of the model to the data:
F(f 2 p, N 2 f) = [SSlof/(f 2 p)]/[SSpe/(N 2 f)]

response. The ANOVA tables both for M1 and P show that the
regression is significant (at least one parameter in the model,
other than b0, is significantly different from 0) and the two
different values s12 and s22 are actually estimates of the same
experimental variance s 2. In other words, the postulated models
are good, that is, they provide an adequate approximation of the
true systems in the investigated experimental domain. From
these models one can extract the best independent set of
conditions for both M1 and P. However, since the goal is the
simultaneous optimization of responses, that is, to obtain a
desirable combination of properties, P and M1 are transformed
into a single desirability function, D, which is then submitted to
optimization.
The desirability function involves the transformation of each
estimated response, P and M1, into a desirability value, di, which
increases as the desirability of the corresponding response
increases.9 The response transformations considered belong to
the class of one-sided transformations. In particular, dP
increases as P increases while dM1 decreases as M1 increases,
with M1 = 1/M, and aiming at the maximization of both the P
and M signals.
The individual desirabilities dP and dM1 thus obtained are then
combined into the D function giving a different importance to
each of them due to the lower polarographic signal of P as
compared with that of M at the working concentration levels
(plasma concentrations) and also considering the noise that the
hydrogen ion reduction wave causes in the P signal. In fact, this
reduction occurs at potential values not far from the P peak
potential. As a consequence, the combined function, D,
takes the form D = 3 dM dP 2 , where dP assumes a double
1
importance compared with dM1. Once the D function is defined,
the algorithm employed by the Nemrod software searches
through the levels of variables (x1x4) to find the optimum value
for D. The resulting optimum conditions are reported in Fig. 2,
which also shows the graphic representation of D.
Calibration model and quantification
In order to avoid interference effects produced by the plasma
matrix, a multivariate calibration method to correlate the
polarographic and UV spectrophotometric data with P and M
concentrations was used. In so doing, the response is described,
quantified and characterized by a multitude of variables
(multivariate characterization). The projection of this multitude
of measurements will provide a better response variable than a
single measurement.10
For both polarographic and spectrophotometric quantification, the model was developed using 11 samples whose
concentrations fit a two-level full factorial design (4 3 2 + 3
replications at the center point) as the training set in order to
allow this set of objects to span the experimental region
properly and obtain a model with good predictive properties in

(2)

Once the experiments were performed, the analysis of the

lambda plot (i.e., the plot of the log-likelihood function L(l)
against l]8 pointed out the need for a transformation for the
response M. The principal transformation response types used
fall within the family of power transformations and can be
represented by yl. The analysis of the L(l) vs. l plot for M
indicates that the simplest and best transformation for M is that
corresponding to l = 21, that is, M = M21. An analogous plot
for P indicates that a transformation is not necessary for this

Fig. 2 Graphical representation of the total desirability function D and

optimum experimental conditions of pH, dropping size ds, potential
amplitude DE and pulse increment pi.

Analyst, 1999, 124, 16831688

1685

Published on 01 January 1999. Downloaded on 09/09/2016 02:24:30.

View Article Online

all the experimental region. The concentration levels of the

polarographic calibration samples are reported in the Experimental section, Plasma analysis. PLS was then used to build the
regression models (PLS1 and PLS2) correlating the polarographic (peak heights in the potential range from 2250 to
21600 mV with steps of 9 mV) and spectrophotometric
(absorbance values in the wavelength range 250400 nm in
steps of 1 nm) data with the concentrations of P and M in the
plasma samples. The models were then improved upon applying
the GOLPE (Generating Optimal Linear PLS Estimations)
variable selection procedure,11 which selected the most informative areas from both spectra and polarograms. The areas
are those that improve the predictive ability of the models.
However, the most efficient way to use the data is to put
together both polarographic and spectrophotometric data to
build a mixed model.
The multivariate approach has the advantage of using the
information provided by these two methodologies in an
optimum way. As a result, it is less sensitive to interference and
errors because they appear only seldom in both analytical
methods.
The PLS model has been embedded in a program written in
FORTRAN which uses as input the files produced as an output
by the analytical instruments and gives the concentrations of
analytes directly. To select the number of components for PLS
models, the cross-validation method of the randomized cancellation group was used.12
Table 2 shows both the fitting and the predictive ability of the
obtained regression models. Fitting ability refers to the
adequacy of the model to describe the data points and predictive
ability refers to the predictive power of the model. The
measurements of fitting and predictive ability are expressed by
R2 and SDEC (standard deviation of error of calculations) for
the former and Q2 and SDEP (standard deviations of error of
predictions)13 for the latter. Plasma samples which were
prepared according to the instructions in the Experimental
section and tested at concentration levels of 2 mg mL21 (M) and
5 mg mL21 (P) were predicted using the PLS1 and PLS2
regression models. The results obtained are reported in Table 3.
No great differences exist between the two quantification
methods although the statistical parameters are better for the
PLS1 models (Table 2). Fig. 3 shows loading plots for the PLS2
and PLS1 polarographic models, taken as examples.
Fig. 3 Loading plots for (a) the PLS2 (M and P) and (b) PLS1 (P)
models.

Method validation4,5,14
The development and validation of a new analytical method is
an iterative process. In order to determine if the developed

Table 2 Polarographic and spectrophotometric PLS models: optimum

number of components and statistical parameters
Number of PC
Polarography
PLS1:
M
3
P
4
PLS2:
M
4
P
4
Spectrophotometry
PLS1:
M
4
P
3
PLS2:
M
4
P
3

1686

PLS1
Polarography
Spectrophotometry
PLS2
Polarography
Spectrophotometry

94.7, 94.8, 103.1

97.6, 94.7, 93.8

96.4, 102.8, 102.6

100.5, 103.2, 103.8

101.8, 98.1, 100.8

96.9, 100.1, 97.2

92.6, 94.2, 97.1

103.8, 100.4, 103.1

SDEC

SDEP

R2

0.0743
0.1407

0.095
0.202

0.997
0.996

0.994
0.993

0.1076
0.2502

0.1421
0.3713

0.993
0.989

0.988
0.976

0.058
0.070

0.082
0.095

0.988
0.999

0.996
0.999

Analyte

Equation of line
(y in nA, x in mg mL21)

R2

DL/mg mL

0.300
0.086

0.335
0.106

0.949
0.999

0.937
0.998

M
P

y = 1.16x + 0.095
y = 0.172x + 0.064

0.994
0.990

0.133
0.250

Analyst, 1999, 124, 16831688

Q2

Table 3 Polarographic and spectrophotometric percentage recoveries of

spiked M and P plasma sample

Table 4 Linearity range and detection limit (DL) validation parameters for
the simultaneous polarographic determination of M and P

View Article Online

method conditions are acceptable, validation studies have to be

carried out with these conditions and, when necessary, they
have to be modified until they satisfy the validation process.
The validation procedure considered for the developed polarographic method takes into account the following parameters:
linearity, accuracy, repeatability, detection limit (DL) and
robustness.

Published on 01 January 1999. Downloaded on 09/09/2016 02:24:30.

Linearity. Standard solutions at five concentration levels,

from 50 to 150% of the target or test analyte concentrations,
were used. The test concentrations selected were P 5 mg mL21
and M 2 mg mL21. Each standard solution was examined three
times and the acceptability of linearity data was judged by
examining the correlation coefficient R2 ( > 0.99 for each of
three curves) and the y-intercept (less than a few per cent of the
test concentration response).
Detection limit (DL). The DL values were calculated
according to the equation DL = 3.3s/S, where S is the slope of
the linearity curves and s is the standard deviation of the yintercepts of regression lines. The factor 3.3, which was
multiplied by the s value, gives a 5% probability of each of the
two following kinds of errors occurring: (1) to claim the
presence of the analyte when it is actually absent and, on the
contrary, (2) to claim the absence of the analyte when it is in fact
present.15
The equation lines and DL values for both M and P are
reported in Table 4. The DL values found were validated by
performing independent analyses of samples prepared at those
concentration levels.

Repeatability. The repeatability data were available from the

triplicate analyses of the spiked plasma samples to obtain the
accuracy data.
Robustness. The robustness, i.e., the ability of the developed
method to remain unaffected by small changes in conditions,
was tested in order to ensure that the method is equally effective
(robust) in many situations likely to be encountered in
practice.
The parameters evaluated during the robustness test, for both
the M and P response, included ds, pH, DE and pi. Small
changes in the method parameters were evaluated simultaneously as a part of a 24 full factorial design (FFD). The design
also included two experiments at the 0 factor level in order to
obtain a model-independent estimation of the experimental
variance. The model considered was a first-order polynomial
model with interactions y = b0 + Sbi + SSbij where i j. The
results of the FFD robustness were evaluated by analyzing the
ANOVA data reported in Table 6. These data need different
comments depending on which of the two responses is
considered. Starting from P, it may be seen that the regression
is significant and the postulated model is good. The graphical
representation of the effect according to Pareto for each
parameter investigated indicates that the pH of the BW buffer
and pi are the only statistically significant parameters. With
regard to the M response, the results from the ANOVA table
show that a higher order model was necessary or, better, the
Table 5 Accuracy of the polarographic method for M and P plasma
determination by PLS1
Analyst

Concentration level (%)

Recovery (%)

RSD (%) (n = 3)

Ma

Accuracy. Spiked plasma samples were tested in triplicate at

three levels over 50150% of the test concentration (5 mg mL21
P; 2 mg mL21 M). Each sample was prepared independently
according to the method procedure and analyzed in random
order in one laboratory on one day. The results obtained are
reported in Table 5.

Table 6

50
105.3
3.1
100
97.5
4.8
150
102.0
2.4
50
108.3
3.3
Pb
100
100.6
3.6
150
105.0
3.8
a Test concentration 2 mg mL21. b Test concentration 5 mg mL21.

Robustness test: FFD analysis of variance for the M and P responses

Analyst

Source of variation

Sum of squares

Degrees of freedom

Mean square

Regression
Residual
Lack of fit
Pure error
Total
Regression
Residual
Lack of fit
Pure error
Total

0.75643
0.10272
0.10227
0.00045
0.85915
0.10740
0.01399
0.01274
0.00125
0.12139

10
6
5
1
16
10
16
5
1
16

0.07564
0.01712
0.02045
0.00045

Table 7

0.01074
0.00233
0.00255
0.00125

F0

Significance (%)

4.42

<5

45.45

11.7

4.60

<5

2.04

48.1

Robustness test: analysis of variance for the M and P responses

Analyst

Source of variation

Sum of squares

Degrees of freedom

Mean square

F0

Significance (%)

Regression
Residual
Lack of fit
Pure error
Total
Regression
Residual
Lack of fit
Pure error
Total

0.29331
0.09754
0.05642
0.041113
0.3905
0.04224
0.02485
0.00928
0.01557
0.06709

11
8
5
3
19
11
8
5
3
19

0.02666
0.01219
0.01128
0.01371

2.19

13.8

0.82

60.5

0.00384
0.00311
0.00186
0.00519

1.24

39.0

0.36

85.2

Analyst, 1999, 124, 16831688

1687

Published on 01 January 1999. Downloaded on 09/09/2016 02:24:30.

View Article Online

postulated model was not satisfactory. Therefore, a response

surface study was carried out which also involved the P
response since considering this signal together with M was not
time consuming and allowed easy conclusions about P robustness to be drawn due to the graphical representation of the
response vs. variables. The response surface design used was a
face center design with four repetitions at the center of the
experimental domain. In order to reduce the number of
experiments, the factorial part of this design became a 24 2 1
fractional factorial design (I 1234). The model considered
included all the main effects bi but only those interactions not
involving the X1 variable (ds): y = b0 + Sbi + SSbij, where i
j and j 1. The new ANOVA table (Table 7) showed the
adequacy of the postulated model to the data. From the analysis
of the graphical representations of the response surfaces
obtained both for M and P, it can be deduced that the response
remains constant in the investigated domain. However, the
setting of the variable pH requires greater care than the other
variables because small deviations from its optimum value
cause greater variations in the response than the setting of the
other variables.

Acknowledgement
This work was supported by a grant from the Italian Ministry of
the University.

1688

Analyst, 1999, 124, 16831688

References
1 Swedish Study Group, J. Antimicrob. Chemother., 1990, 26, 173.
2 A. M. Y. Jaber and A. Lounici, Anal. Chim. Acta, 1994, 291, 53.
3 A. Z. Abu Zuhri, S. I. Al-Khalil and M. S. Suleiman, Anal. Lett.,
1986, 19, 453.
4 Third International Conference on Harmonization of Technical
Requirements for Registration of Pharmaceuticals for Human Use.
Draft Guidelines on Validation of Analytical Procedures: Methodology, Yokohama, Japan, 29 November1 December 1995.
5 United States Pharmacopeia, 23rd Revision, US Pharmaceutical
Convention, Rockville, MD, 1995, pp. 19821984.
6 S. Wold, J. Pharm. Biomed. Anal., 1991, 9, 589.
7 P. Gratteri, S. Furlanetto, E. La Porta, S. Pinzauti and R. Leardi,
Farmaco, 1996, 51, 231.
8 P. D. Haaland, Experimental Design in Biotechnology, Marcel
Dekker, New York, 1989.
9 G. Derringer and R. Suich, J. Qual. Technol., 1980, 12, 214.
10 S. Wold, M. Sjstrm, R. Carlson, T. Lundstedt, S. Hellberg, B.
Skagerberg, C. Wikstrm and J. hman, Anal. Chim. Acta, 1986, 191,
17.
11 M. Baroni, G. Costantino, G. Cruciani, D. Riganelli, R. Valigi and S.
Clementi, Quant. StructAct. Relat., 1993, 12, 9.
12 S. Wold, Technometrics, 1978, 20, 397.
13 G. Cruciani, M. Baroni, S. Clementi, G. Costantino, D. Riganelli and
B. Skagerberg, J. Chemom., 1992, 6, 335.
14 J. M. Green, Anal. Chem., 1996, 68, 305A.
15 J. C. Miller and J. N. Miller, Statistics for Analytical Chemistry,
Wiley, New York, 3rd edn., 1993.

Paper 9/05038H