G Model

TAL-9721; No. of Pages 5 ARTICLE IN PRESS

Talanta xxx (2008) xxx–xxx

Contents lists available at ScienceDirect

Talanta
journal homepage: www.elsevier.com/locate/talanta

Optimization of Pb(II) biosorption by Robinia tree leaves using

statistical design of experiments
Javad Zolgharnein ∗ , Ali Shahmoradi, Mohammad Reza Sangi
Department of Chemistry, Faculty of Science, Arak University, 38156-876 Arak, Iran

a r t i c l e i n f o a b s t r a c t

Article history: The present study introduces Robinia tree leaves as a novel and efficient biosorbent for removing Pb(II)
Received 17 December 2007 from aqueous solutions. In order to reduce the large number of experiments and find the highest removal
Received in revised form 19 March 2008 efficiency of Pb(II), a set of full 23 factorial design with two blocks were performed in duplicate (16 exper-
Accepted 22 March 2008
iments). In all experiments, the contact time was fixed at 25 min. The main interaction effects of the three
Available online xxx
factors including sorbent mass, pH and initial concentration of metal-ion were considered. By using Stu-
dent’s t-test and analysis of variances (ANOVA), the main factors, which had the highest effect on the
Keywords:
removal process, were identified. Twenty-six experiments were designed according to Doehlert response
Biosorption
Pb(II)
surface design to obtain a mathematical model describing functional relationship between response
Robinia tree leaves and main independent variables. The most suitable regression model, that fitted the experimental data
Doehlert design extremely well, was chosen according to the lack-of-fit-test and adjusted R2 value. Finally, after checking
for possible outliers, the optimum conditions for maximum removal of Pb(II) from aqueous solution were
obtained. The best conditions were calculated to be as: initial concentration of Pb(II) = 40 mg L−1 , pH 4.6
and concentration of sorbet equal to 27.3 g L−1 .
© 2008 Elsevier B.V. All rights reserved.

1. Introduction
Industrial wastes and fertilizers can add excessive amount of
heavy metals to the environment. Heavy metals even at very low
concentrations are highly toxic, and pose a serious threat to biota
and the environment [1]. Among the heavy metals lead has the
most damaging effects on human health. It can enter a human
body through uptake of food (65%), water (20%) and air (15%) [2].
Continuous absorption of lead may cause serious injuries to health
such as encephalopathy, kidney damage and several others [3,4].
Several methods for the removal of toxic metals from wastew-
aters have been introduced and tried out [5]. Techno-economic
considerations have restricted the wide scale application of these
methods [6,7]. However, application of biosorption has been proved
to be an effective way to heavy metal ions pollutions [1,3–6]. Suc-
cessful metal biosorption, with a variety of biological materials
including, microalgae, seaweeds, bacteria, fungi, crop residues, and
papaya wood have been reported [8–10]. This study verifies a new
biomass as metal sorbent, and demonstrates its potential for effi-
cient removal of Pb(II) from aqueous solution. Robinia tree leaves
are in great supply, inexpensive and easily available in countries
∗ Corresponding author. Tel.: +98 861 4173400; fax: +98 861 4173406.
E-mail address: J-zolgharnein@araku.ac.ir (J. Zolgharnein).
such as Iran. These leaves have no commercial usage and are not
eaten by livestock. Factors affecting metal ions uptake by a sor-
bent in a batch system include sorbent mass (s), acidity of medium
(pH), metal-ion and the sorbent contact time (t), initial metal-
ion concentration (m), speed of shaking, etc. [10]. Thus several
factors should be optimized. Here the major concern is the large
number of experimental runs. To resolve this problem, some vari-
ables can be arbitrarily fixed. The application of statistical design
especially, Dohelert design for optimization of biosorption pro-
cess has comparatively rarely been reported in literatures than
other approaches such as industrial process or analytical research
[10–13]. This study tries to develop an efficient method for achiev-
ing maximum removal of Pb(II) ion in aqueous solutions. For this
purpose, by using Dohelert design and response surface methodol-
ogy (RSM) an optimized response model was proposed [12–15].
2. Experimental
2.1. Materials
2.1.1. Biosorbent materials
The Robinia tree leaves were gathered from twigs into clean plas-
tic bags, washed with doubly distilled water and dried on a clean
table. The dried leaves were ground and sieved to 40–50 mesh, then
stored into plastic bag.

0039-9140/$ – see front matter © 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.talanta.2008.03.039

Please cite this article in press as: J. Zolgharnein, et al., Talanta (2008), doi:10.1016/j.talanta.2008.03.039
G Model
TAL-9721; No. of Pages 5 ARTICLE IN PRESS
2 J. Zolgharnein et al. / Talanta xxx (2008) xxx–xxx
Table 1
Factors and levels used in the factorial design
Factor Low
Biosorbent mass (g) 0.1
Initial pH 2 5
Pb(II) initial concentrationa (mg L−1 ) 40
a
A sample volume of 10 mL was used in either case.
2.1.2. Metal solution
The stock solution of Pb(II) was prepared by dissolving appropri-
ate quantities of Pb(NO3 )2 salt in 5% (v/v) HNO3 . All the chemicals
were of analytical grade and obtained from Merck. Doubly dis-
tilled water was used throughout. Ten milliliters of Pb(II) solution
was poured into 100 mL Erlenmeyer. A known amount of dried
biomass was added to each sample solution. The pH of solution
was adjusted with HNO3 or NaOH before addition of biosorbent.
Fresh dilutions were made for each study. Based on pervious expe-
riences, the biosorption process is very fast in its earlier stages and
reaches its equilibrium state at about 15–25 min and the follow-
ing passage of time has no or marginal effect on removal efficiency
improvement [16,17]. So, the contact time was chosen to be 25 min.
Samples were filtered through 0.45 ␮m Whatman filter paper. The
factors and levels used in factorial design are presented in Table 1.
2.2. Metal analysis
The concentration of Pb(II), remaining in solution after biosorp-
tion was determined by using PerkinElmer 2380 atomic absorption
spectrophotometer [7]. Instrumental conditions were adjusted as
recommended by manufacturer applying a current 10.0 mA. The
sensitive wavelength for lead at 283.3 nm was used with slit band-
width of 0.7 nm.
2.2.1. Removal efficiency
Percent of removal efficiency is defined as
(m − f )
R= × 100
m obtained depends critically on the validity of these assumptions.
m and f stand for the initial and final concentration of Pb(II) in the
solution, respectively.
2.3. Optimization
2.3.1. Optimization design
The optimization procedure of any analytical assay can be
performed in two different ways: univariate and multivariate. In
first approach, the optimum values for all of the variables are
found one at a time. This method of optimization requires greater
amounts of reagent and time to be accomplished. In addition, pos-
sible interactions among variables are not considered. In contrast,
the multivariate optimization methodology considers the possi-
ble interactions and is also faster, more effective and economical.
Simultaneous optimization of several variables is also possible by
this approach [11–15]. Full factorial design is one of the most fre-
quently employed chemometrics technique used for preliminary
evaluation of significant variables [14,15,18]. After identification
of significant factor(s), response surface methodology (RSM) is
used for closer inspection of the relationship between the iden-
tified significant factors and the response variable. RSM can be
done by several models including central composite design (CCD)
[12–15], Box-Behnken [19,20] and Doehlert design [12,13,21,22].
The Doehlert design is a less known but very useful type of design
offering advantages with respect to the other two ones [13,21]. It
describes a spherical experimental domain but with less points
Please cite this article in press as: J. Zolgharnein, et al., Talanta (2008), doi:10.1016/j.talanta.2008.03.039
than CCD and emphasizes uniformly space filling. Apart from its
uniformity, the most important advantage is its sequentially poten-
High tial. One can reuse experiments when the boundaries were not well
0.3
chosen at first, provided that the old and new considered bound-
aries are adjacent [15]. Thus, Doehlert matrix design was chosen as
200 the multilevel optimization strategy to have a forecast of high qual-
ity. According to it, the total number of treatment combinations
was 2k + 2k + n0 , where k is the number of independent variables
and n0 is the number of replications of central point [12,23,24]. In
Doehelrt design, the number of the levels is not the same for all
variables. Generally, the levels are assigned to the factors accord-
ing to their significance in such a way that the highest level is for the
variable with greatest effect and lowest level for the factor of least
effect [15]. In this way, the most valuable information of the sys-
tem under study is obtained. Values in Doehlert design are given in
coded units and should be decoded to give the real values. Decoding
was performed according to the following equation:
  
ximax − xi∗ zimax − zimin
zi∗ = zimax − (2)
2ximax
where xi and zi are the values of the factors of interest in coded
and uncoded (real) units with the maximum and minimum value
represented by ximax and ximin (ximax = −ximin ) for coded form and
zimax and zimin for real form, respectively [14].
2.3.2. Regression analysis
Determination of regression equation which is a mathematical
relation between dependent factor(s) and independent variables is
the most important aim of analysis. After performing experiments
required for optimization, one needs to do a regression analysis to
find the desired function. Here factorial optimization and algebraic
calculations were done by using MINITAB Version 14 for windows
and MAPLE Version 10, respectively.
2.3.3. Model criticism and selection
A statistical method such as regression analysis is based on
(1) certain assumptions. The accuracy of the analysis and the results
Regression analysis is viewed as an iterative process, in which the
outputs are used to diagnose, validate, criticize and possibly mod-
ify input data. The process is repeated until a satisfactory result has
been obtained. A good model satisfies the regression assumptions
well and its lack of fit is not significant at certain confidence level
[25].
3. Results and discussion
3.1. Factorial design for variable screening
To identify effective factors and interactions, a 23 full factorial
design was performed [18,26]. The results are listed in Table 2. All
the experiments were done in duplicate and random order to eval-
uate pure error of the method and to minimize the effect of possible
uncontrolled variable(s), respectively. The blocks were performed
on two different working days.
The obtained results through ANOVA showed that all the main
factors, pH, initial concentration of Pb(II) (m), amount of sorbent (s)
and their interactions were statistically significant (P < 0.05), while
blocks were not effective (P > 0.05). This is a very important find-
ing, since biosorption was not influenced by the day of performing
the measurement and this shows robustness of biosorption pro-
cess. The effects, regression coefficients, standard errors, T and P are
shown in Table 3. The increase of the removal of metal-ion as the
pH increases can be explained based on decreases in competition
G Model
TAL-9721; No. of Pages 5 ARTICLE IN PRESS
J. Zolgharnein et al. / Talanta xxx (2008) xxx–xxx 3

Table 2
23 full factorial design results

StdOrder RunOrder CenterPt Blocks pH m (mg L−1 ) s (gr) R

6 1 1 1 5 40 0.3 96.92
5 2 1 1 2 40 0.3 95.47
1 3 1 1 2 40 0.1 74.82
7 4 1 1 2 200 0.3 90
4 5 1 1 5 200 0.1 81.17
8 6 1 1 5 200 0.3 92.7
2 7 1 1 5 40 0.1 95
3 8 1 1 2 200 0.1 50.33
13 9 1 2 2 40 0.3 95.6
16 10 1 2 5 200 0.3 92.7
9 11 1 2 2 40 0.1 74.7
11 12 1 2 2 200 0.1 50.5
10 13 1 2 5 40 0.1 95
14 14 1 2 5 40 0.3 97
15 15 1 2 2 200 0.3 90.2
12 16 1 2 5 200 0.1 81.17 Fig. 2. Pareto chart of the main effects and related interactions. The vertical dashed
line indicates the 95% confidence interval.
Variables are in uncoded units.

Table 3 when sorbent mass is in its lower level; increasing the pH of solu-
Statistical parameters for 23 designs tion has greater effect. This is a reasonable result, because a lower
Term Effect Coef SE Coef T P weight of sorbent implies less number of biosorption active sites
and in this case proton–metal-ion competition for occupation of
Constant 84.58 0.01892 4471.36 0.000
Block −0.029 0.01892 −1.52 0.172 active sites becomes more critical. It can also be concluded when
pH 13.755 6.878 0.01892 363.58 0.000 initial metal-ion concentration is high, increasing sorbent mass has
m −11.968 −5.984 0.01892 −316.33 0.000 grater effect. This result is also justifiable, because in the case of
s 18.488 9.244 0.01892 488.68 0.000
low initial metal-ion concentration, even low mass of sorbent can
pH × m 2.923 1.461 0.01892 77.25 0.000
pH × s −11.743 −5.871 0.01892 −310.39 0.000
remove it sufficiently and therefore increasing the sorbent mass has
s×m 7.12 3.56 0.01892 188.2 0.000 marginal positive effect. The evaluation of the effect of each variable
pH × m × s −2.335 −1.168 0.01892 −61.72 0.000 on the removal efficiency may also be verified from the Pareto chart
(Fig. 2). For 95% confidence level and 8 d.f., the t-value for two-sided
test is equal to 2.4. The vertical dashed line indicates minimum sta-
between protons and metal ions for the same functional groups. It tistically significant effect magnitude for a 95% confidence level.
also decreases the positive surface charge resulting in a lower elec- Horizontal column lengths are proportional to Student’s t-test val-
trostatic repulsion between the surface of biomass and metal ions ues for each effect. Any effect or interaction that exceeds vertical
[27]. Fig. 1 illustrates the interactions between main factors and line is considered operative. The significance of the factors can also
reveals that there is an interaction between s and pH, s and m, pH be confirmed from the normal probability plot displayed in Fig. 3.
and m. The presence of interaction means, the factors may affect When the effects present a large deviation from the normal distri-
the response interactively and not in an independent way, i.e., their bution, they probably are significant and are obviously not related
combined effect is greater or less than that expected for the straight to experimental random errors. In this sense, all the points devi-
addition of the effects [28,29]. These synergistic effects would not ated from the straight line on the probability plot, are the same
be detected in a univariate optimization of the system. The unparal- as those introduced as significant factors by Pareto chart, confirm-
lel effect lines for pH × s and m × s interaction illustrates that there ing the results obtained from it. The normal probability method is
is rather a strong two-way interaction between pH × s and m × s. especially useful for analyzing the effects in experiment without
Meanwhile, the interaction between pH and m is relatively weak, replicates [14].
because the effect lines are nearly parallel. It also can be concluded

Fig. 1. Interaction effects plot. The unit of the parameters is as the main effect plot. Fig. 3. Normal probability plot for standardized effects.

Please cite this article in press as: J. Zolgharnein, et al., Talanta (2008), doi:10.1016/j.talanta.2008.03.039
G Model
TAL-9721; No. of Pages 5 ARTICLE IN PRESS
4 J. Zolgharnein et al. / Talanta xxx (2008) xxx–xxx

Table 4
Doehlert matrix response surface design and results obtained for Pb(II) biosorption using Robinia tree leaves

Experiment Coded values Real values %R Predicted R Residuals

−1 a
X1 X2 X3 pH s (gr) m (mg L ) Rep . 1 Rep. 2 Rep. 1 Rep. 2

1 0 0 0 3.5 0.2 120 90.67 91.05 90.86 −0.19 0.19

2 1 0 0 5 0.2 120 92.21 91.18 91.69 0.52 −0.52
3 0.5 0.866 0 4.25 0.3 120 93.1 92.33 92.73 0.37 −0.37
4 0.5 0.289 0.816 4.25 0.233 200 87.61 88.16 87.87 −0.26 0.29
5 −1 0 0 2 0.2 120 91.46 91.46 91.46 0 0
6 −0.5 −0.866 0 2.75 0.1 120 84.38 83.87 84.11 0.27 −0.24
7 −0.5 −0.289 −0.816 2.75 0.167 40 94.3 94.5 94.41 −0.11 0.09
8 0.5 −0.866 0 4.25 0.1 120 87.46 86.69 87.06 0.4 −0.37
9 0.5 −0.289 −0.816 4.25 0.167 40 95.08 94.69 94.9 0.18 −0.21
10 0 0.577 −0.816 3.5 0.267 40 95.86 95.47 95.64 0.22 −0.17
11 −0.5 0.866 0 2.75 0.3 120 91.46 91.59 91.54 −0.08 0.05
12 −0.5 0.289 0.816 2.75 0.233 200 85.89 86.13 86 −0.11 0.13
13 0 −0.577 0.816 3.5 0.133 200 81.59 81.13 81.38 0.21 −0.25
a
Replicate.

Table 5
Analysis of variance for suggested third-order model

Source d.f. SS MS F P

Regression 11 457.72 41.611 324.35 0.000

Residual error 14 1.796 0.128
Lack of fit 1 0.004 0.004 0.03 0.867
Pure error 13 1.792 0.138

Total 25 459.516

3.2. Final optimization using Doehlert design

pH, s and m all were determined as effective factors and should

be optimized. At this point, they were optimized according to
Doehlert matrix. Seven levels were used for s, since this factor
presented higher effect than the other two ones. pH and m were
given five and three levels, respectively according to their effects
Fig. 4. Normal probability plot for residuals.
importance (Table 4).

3.3. Finding the best model using regression analysis
An appropriate model can be expressed as
R = 135 − 33.9 pH − 0.271 m + 26.4 s + 20.8 pH s + 0.0527 pH m
+ 0.950 s m − 0.222 pH s m − 218 s2 + 8.42 pH2 − 0.771 pH3
− 2 × 10−7 m3
R represents percent of removal efficiency.
ANOVA test results are illustrated in Table 5. The estimated
effects and coefficients for this model are listed in Table 6. The
results showed that the lack of fit is not significant (P > 0.05) and
regression is meaningful (P < 0.05). Goodness-of-fit for this model
was also evaluated by the coefficient of determination. In this case,
in fact, fitting is very well (R2 = 0.993) and only 0.7% of total vari-
ance was not explained by the model. The high value of adjusted
regression coefficient (0.997) is also an indication for high signif-
icance of the proposed model [30]. The residuals also had to be
examined for normal distribution. Anderson–Darling test is a pow-
erful statistical means that generates normal probability plot and
(4)
performs a hypothesis test to examine whether the observations
follow a normal distribution.

Table 6
Estimated effects and coefficients for suggested third-order model

Predictor Coef SE Coef T P

Constant 134.97 11.04 12.22 0.000

pH −33.915 5.884 −5.76 0.000
m −0.27081 0.06352 −4.26 0.001
s 26.4 39.44 0.67 0.514
pH × s 20.8 10.86 1.91 0.076
pH × m 0.05269 0.01803 2.92 0.011
s×m 0.9501 0.3144 3.02 0.009
pH × s × m −0.22224 0.08944 −2.48 0.026
s2 −217.94 23.26 −9.37 0.000
pH2 8.418 1.294 6.51 0.000
pH3 −0.7714 0.1225 −6.29 0.000
m3 −2 × 10−7 9 × 10−8 −2.3 0.038
Fig. 5. Standardized residuals vs. plotted fitted values.

Please cite this article in press as: J. Zolgharnein, et al., Talanta (2008), doi:10.1016/j.talanta.2008.03.039
G Model
TAL-9721; No. of Pages 5 ARTICLE IN PRESS
J. Zolgharnein et al. / Talanta xxx (2008) xxx–xxx
The result obtained by Anderson–Darling test for residuals is
shown in Fig. 4. The P value for test (0.846 > 0.05) shows that the
null hypothesis could not be rejected and it concludes that residu-
als follow the normal distribution. Plot of standardized residuals vs.
predicted values indicates the possible existence of outliers (Fig. 5).
If a point lies far from the majority of points, it may be an outlier. It
is important to identify the outliers, because they can significantly
influence the model, providing potentially misleading or incorrect
results. Except for two points (marked with triangle) all other points
were found to fall in the range of +1.5 to −1.5. Points corresponding
to these runs were slightly out of this range (standardized resid-
ual = 2). However, elimination of these points did not reduce lack
of fit but increased it. This shows that they were not really out-
liers. Application of Cochran’s test for variances gives 0.296 with
n = 2 and k = 13, which is smaller than Cochran’s critical value of
0.515 and emphasizes the homoscedasticity numerically [15]. All
these verify the fact that the model does not violate regression
assumptions.
3.4. Model validation
A model is valid when it has good predictability. The low value
for PRESS (6.45832) and high value for predicted R2 (98.59%)
showed the high predictive ability of the model.
By using optimization toolbox in MAPLE, the best conditions
for suggested model was found to be 97.3% as maximum removal
efficiency in pH = 4.6, s = 0.27 gr and m = 40 mg L−1 . Replicating
the experiment twice under this condition led to an average of
96.8 ± 1.2 at 95% confidence level for removal percent of Pb(II),
which is in very good agreement with the predicted value. This
shows that the model has high prediction ability.
3.5. Capacity of sorbent and effect of other ions
The maximum capacity of sorbent was found to be 144.1 mg g−1 .
Interference effects due to Mg2+ , Ca2+ , Na+ and K+ at 10-fold
excess were also studied. It was found that under these condi-
tions, there is no significant adverse effect on removal efficiency.
This can be attributed to high affinity of this sorbent to Pb(II)
[16,17].
3.6. Application
The major interest of this study was to show the feasibility of
biosorption of Pb(II) by Robina tree leaves through a typical math-
ematical model [12]. Since the concentration of Pb(II) in various
industrial wastes varies drastically, thus for each case, it is recom-
mended that researcher treat each problem individually and find
its appropriate model.
4. Conclusions
Factorial experimental design is time saving and economic,
because only significant effects are optimized and unnecessary
experiments avoided. Interactions among the factors are also con-
Please cite this article in press as: J. Zolgharnein, et al., Talanta (2008), doi:10.1016/j.talanta.2008.03.039
5
sidered. It was clearly shown the feasibility of biosorptive process as
a low cost and effective method for removal of Pb(II) from aqueous
solutions by Robinia tree leaves. The calculated optimum condi-
tions for this process were in good agreement with experimental
results.
Acknowledgement
The authors would like to thank Dr. H. Yazdani for final editing
of the manuscript.
References
[1] B. Volesky, Hydrometallurgy 59 (2001) 203.
[2] http://www.lenntech.com/Periodic-chart-elements/Pb-en.htm.
[3] T.W. Clarkson, L. Friberg, G.F. Nordberg, P.R. Sager, Biological Monitoring of Toxic
Metals, Kluwer Academic Publishers, New York, 1998.
[4] W.N.L. dos Santos, C.M.M. dos Santos, J.L.O. Costa, H.M.C. Andrade, S.L.C. Fer-
reira, Microchem. J. 77 (2004) 123.
[5] B. Volesky, S. Schiewer, in: M.C. Flickinger, S.W. Drew (Eds.), Encyclopedia of
Bioprocess-Engineering, Wiley, New York, 1999.
[6] A. Saeed, M.W. Akhtar, M. Iqbal, Sep. Purif. Technol. 45 (2005) 25.
[7] M.D.A. Korn, J.B. de Andrade, D.S. de Jesus, V.A. Lemos, M.L.S.F. Bandeira, W.N.L.
dos Santos, M.A. Bezerra, F.A.C. Amorim, A.S. Souza, S.L.C. Ferreira, Talanta 69
(2006) 16.
[8] S. Schiewer, B. Volesky, in: D.R. Lovely (Ed.), Environmental Microbe–Metal
Interactions, ASM Press, Washington, DC, 2000 (Chapter 14).
[9] A. Saeed, M. Iqbal, M.W. Akhtar, Pak. J. Sci. Ind. Res. 45 (2002) 205.
[10] E.C. Lima, B. Royer, J.C.P. Vaghetti, J.L. Brasil, N.M. Simon, A.A. dos Santos Jr., F.A.
Pavan, S.L.P. Dias, E.V. Benvenutti, E.A. da Silva, J. Hazard. Mater. 140 (2007) 211.
[11] L.A. Portugal, H.S. Ferreira, W.N.L. dos Santos, S.L.C. Ferreira, Microchem. J. 8
(2007) 77.
[12] P. Vanloot, J.L. Boudenne, L. Vassalo, M. Sergent, B. Coulomb, Talanta 73 (2007)
237.
[13] S.L.C. Ferreira, W.N.L. dos Santos, C.M. Quintella, B.B. Neto, J.M. Bosque-Sendra,
Talanta 63 (2004) 1061.
[14] D.C. Montgomery, Design and Analysis of Experiments, fifth edn., Wiley, New
York, 2001.
[15] D.L. Massart, B.G.M. Vandeginste, L.M.C. Buydens, S. DE Jong, P.J. Lewi, J. Smey-
ers Verbeke, Handbook of Chemometrics and Qualimetrics, Part A, Elsevier,
Amsterdam, 1997.
[16] M.R. Sangi, A. Shahmoradi, J. Zolgharnein, G. Azimi, M. Ghorbandoost, J. Hazard.
Mater., in press.
[17] M.R. Sangi, A. Shahmoradi, M. Ghorbandoost, J. Zolgharnein, G. Azimi, Desali-
nation, in press.
[18] M. Soylak, I. Narin, M.D.A. Bezerra, S.L.C. Ferreira, Talanta 65 (2005) 895.
[19] G.E.P. Box, W.G. Hunterand, J.S. Hunter, Statistics for Experimenters, Wiley, New
York, 1978.
[20] S.L.C. Ferreira, R.E. Bruns, H.S. Ferreira, G.D. Matos, J.M. David, G.C. Brandao,
E.G.P. da Silva, L.A. Portugal, P.S. dos Reis, A.S. Souza, W.N.L. dos Santos, Anal.
Chim. Acta 597 (2007) 179.
[21] A.F. Barbosa, M.G. Segatelli, A.C. Pereira, A.D.S. Santos, L.T. Kubota, P.O. Luccas,
C.R.T. Tarley, Talanta 71 (2007) 1512.
[22] A. Gustavo González, D. González-Arjona, Talanta 49 (1999) 433.
[23] F. Hellal, M. Dachraoui, Talanta 63 (2004) 1089.
[24] R.E. Santelli, M. de Almeida Bezerra, O.D. de SantAna, R.J. Cassella, S.L.C. Ferreira,
Talanta 68 (2006) 1083.
[25] S. Chatterjee, A. Hadi, Regression Analysis by Example, fourth edn., John Wiley
& Sons, Inc., Hoboken, New Jersey, 2006.
[26] F.A. Pavan, Y. Gushikem, A.C. Mazzocato, S.L.P. Dias, E.C. Lima, Dyes Pigments
72 (2007) 256.
[27] Z. Reddad, C. Gerente, Y. Anders, P. LeCloires, Environ. Sci. Technol. 36 (2002)
2067.
[28] R.L. Anderson, Practical Statistics for Analytical Chemists, Van Nostrand Rein-
hold, New York, 1987.
[29] J.N. Miller, J.C. Miller, Statistics and Chemometrics for Analytical Chemistry,
fourth edn., Dorset Press, Dorchester, 2000.
[30] M. Elibol, Process Biochem. 38 (2002) 667.