Journal of Chromatography B, 774 (2002) 215–222

www.elsevier.com / locate / chromb

Linear regression for calibration lines revisited: weighting schemes

for bioanalytical methods
˜ *
A.M. Almeida, M.M. Castel-Branco, A.C. Falcao
Laboratory of Pharmacology, Faculty of Pharmacy, Coimbra University, 3000 -295 Coimbra, Portugal

Received 26 January 2002; received in revised form 11 April 2002; accepted 11 April 2002

Abstract

When the assumption of homoscedasticity is not met for analytical data, a simple and effective way to counteract the
greater influence of the greater concentrations on the fitted regression line is to use weighted least squares linear regression
(WLSLR). The purpose of the present paper is to stress the relevance of weighting schemes for linear regression analysis and
to show how this approach can be useful in the bioanalytical field. The steps to be taken in the study of the linear calibration
approach are described. The application of weighting schemes was shown by using a high-performance liquid chromatog-
raphy method for the determination of lamotrigine in biological fluids as a practical example. By using the WLSLR, the
accuracy of the analytical method was improved at the lower end of the calibration curve. Bioanalytical methods data
analysis was improved by using the WLSLR procedure.  2002 Published by Elsevier Science B.V.

Keywords: Linear regression; Heteroscedasticity; Weighting schemes; Bioanalytical methods

1. Introduction data. Obviously, when the range in x-values is

somewhat larger—usually a concentration range of
A well-designed and interpreted calibration curve more than one order of magnitude—it might be
is essential in any analytical methodology. In fact, expected that the variance of each data point might
the quality of bioanalytical data is highly dependent be quite different [1]. Larger deviations present at
on the quality of the standard curve used to generate larger concentrations tend to influence (weight) the
it. Analyte concentrations in unknown samples are regression line more than smaller deviations associ-
typically evaluated by using the regression results ated with smaller concentrations, and thus the ac-
obtained from calibration curves and although some curacy in the lower end of the range is impaired
analytical procedures may require a non-linear cali- [1–3]. A simple and effective way to counteract this
bration approach, linear regression is the most com- situation is to use weighted least squares linear
monly adopted model. regression (WLSLR) [1,2,4–7]. The aim of the
However, the condition of equal variances, termed present paper is to stress the relevance of weighting
homoscedasticity, is frequently not met for analytical schemes for linear regression analysis and to show
how this approach can be used and be useful in the
*Corresponding author. Tel.: 1351-239-820-510; fax: 1351- bioanalytical field. Although statistical considera-
239-837-731. tions are not new for mathematical experts, we
˜
E-mail address: acfalcao@ff.uc.pt (A.C. Falcao). believe the present paper may be of great utility for

1570-0232 / 02 / $ – see front matter  2002 Published by Elsevier Science B.V.

PII: S1570-0232( 02 )00244-1
216 A.M. Almeida et al. / J. Chromatogr. B 774 (2002) 215–222

practising bioanalysts. The steps to be taken are of variance as a function of concentration [1,2,4–
described and illustrated with a data set obtained 6,10].
during the validation process of a high-performance In order to counteract the greater influence of the
liquid chromatographic (HPLC) method [8]. greater concentrations on the fitted regression line,
the weighted least squares linear regression is used.
The expression to be minimised now takes the
following form [1]:
2. Simple and weighted linear regression
models—background ( y observed,i 2 y predicted,i )2
SS 5 O ]]]]]]]
s 2i
(2)
The objective of a regression analysis is to find a
deterministic model which allows a prediction of the
where s 2i is the variance of the standard point.
values assumed by the dependent variable ( y) when
Taking the objective of the WLSLR into consid-
independent variables (x n ) are known or fixed. That
eration, the most appropriate weighting factor, w i , is
model defines the kind of relationship between
the inverse of the variance of the standard point:
variables. Since the experimental values hardly fit
the mathematical model, the methodology of mini- 1
w i 5 ]2 (3)
mising the sum of squares (SS) of the deviations si
between the data and the assumed model should
yield the best estimate of the model parameters. This However, this weight is usually impractical, and
is called the ‘‘method of least squares’’ and the other empirical weights based on x-variable (con-
expression to be minimised is: centration) or y-variable (response) may provide a
simple approximation of variance [1,2].
SS 5 O( y observed,i 2 y predicted,i )2 (1)

In the simple linear regression model, the relation- 3. Performance

ship between variables is established by a straight
line, mathematically expressed as y 5 a 1 bx, where 3.1. Test of homoscedasticity
y is the dependent variable (measured with error), x
is the independent variable (known without error), a The homoscedasticity assumption should be tested
is the y-intercept of the regression equation and b is in any linear regression analysis. It can be performed
the slope of the regression equation. by plotting residuals versus concentration
Usually, the simple least squares method considers [1,2,4,7,11] and by applying an F-test in accordance
that, for each value of x, there is a subpopulation of with the following statistics [12–14]:
y-values normally distributed, that the means of all s 22
the subpopulations of y lie on the same straight line Fexp 5 ]2 (4)
and that all the subpopulations of y-values have s1
equal variances [9]. Ftab ( f1 , f2 ; 0.99)
However, it is very common for the standard
deviation (SD) of the measurement to alter with x where the experimental F-value is expressed as the
(heteroscedasticity). In many cases, SD rises pro- ratio between the variances obtained at the lowest
portionally to the concentration, leading to a constant (s 21 ) and at the highest (s 22 ) concentration level of the
coefficient of variation. Nevertheless, taking into working range, and the tabled F-value is obtained
account that random error is caused by noise and from the F-table at the confidence level of 99% for
noise sources may be a function of signal or con- f1 5 f2 5 (n 2 1) degrees of freedom.
centration or other factors, different behaviours may If variance is constant over the whole calibration
be observed. Despite observing this, the most com- range, residuals will fall more or less randomly
mon occurrence of heteroscedasticity is an increase around the x-axis and Fexp will be lower than Ftab .
 A.M. Almeida et al. / J. Chromatogr. B 774 (2002) 215–222 217

3.2. Choice of the weighting factor estimated values for the a and b parameters of the
weighted regression equation can be obtained by the
In the light of the evidence of the heteroscedastic following modified formulas:
situation, the following step should be the choice of
the weighting factor, w i . As it is not suitable to O O O O
wi ? wi xi yi 2 wi xi ? wi yi
calculate the inverse of variance in laboratory
routine, taking into account the fact that it requires
O O O
b 5 ]]]]]]]]]
w i ? w i x i2 2s w i x id
2 (8)

several determinations for each calibration point and Ow x ? Ow y 2 Ow x ? Ow x y

2
i i i i i i i i i
a fresh calibration line should be performed each
time the method is used, other empirical weights Ow ? Ow x 2sOw x d
a 5 ]]]]]]]]]]
i
2
i i i i
2 (9)

such as 1 /x 1 / 2 , 1 /x, 1 /x 2 , 1 /y 1 / 2 , 1 /y and 1 /y 2 where (x i , y i ) is the ith data pair of n total data pairs
should be studied. and w i is the weighting factor chosen.
The best weighting factor is chosen according to a The degree of dependence established between the
percentage relative error (%RE), which compares the two variables, expressed by the correlation coeffi-
regressed concentration (Cfound ) computed from the cient (r-value), can be obtained by the following
regression equation obtained for each w i , with the modified formula:
nominal standard concentration (Cnom ):
Cfound 2 Cnom
O O O O
wi ? wi xi yi 2 wi xi ? wi yi
r 5 ]]]]]]]]]]]]]
%RE 5 ]]]] 3 100
Cnom
(5) O O O
]]]]]] 2
O O
]]]]]]2
œ wi ? wi x i2 2s wi xid ?œ wi ? wi y i2 2s wi yid O
The %RE, evaluated by plotting %RE versus (10)
concentration as well as by calculating %RE sum,
defined as the sum of absolute %RE values, is a A generic diagram of the process described is
useful and sensitive indicator of goodness of fit in represented in Fig. 1.
the evaluation of the effectiveness of a weighting
factor for WLSLR [1].
The best w i will be that which gives rise to a 4. Practice
narrow horizontal band of randomly distributed %RE
around the concentration axis and presents the least To exemplify the procedure described, we selected
sum of the %RE across the whole concentration the intra-day assay data set generated during the
range. validation process of a recently developed HPLC
method for the determination of lamotrigine in
3.3. Weighted straight line equation plasma [8]. Regression parameters were obtained by
introducing the respective formulas on a Microsoft
The model parameters (a and b) of the weighted Excel  worksheet.
straight line equation can now be estimated. Conver-
sion of the relations for unweighted least squares into 4.1. Test of homoscedasticity
their weighted counterparts can easily be done by
adding a term w i to any sum and changing any term The plot of residuals versus concentration obtained
n into ow i [1,15]. Knowing that estimated values for for the intra-day assay chromatographic data is
the a and b parameters for unweighted least squares shown in Fig. 2. The residual plot clearly showed
may be obtained by the following formulas: that error was not randomly distributed around the

O O O
n ? xi yi 2 xi ? yi
concentration axis. The F-test (Table 1) also re-
vealed a significant difference between the variances,
O O
b 5 ]]]]]]]
n ? x i 2 s x id
2 2 (6)
when the experimental F-value (Fexp 51.45310 5 )
was compared to the tabled one (Ftab 515.98).
Ox ? O y 2 Ox ? Ox y
2
i i i i i There was evidence that variances were signifi-
n ? Ox 2sOx d
a 5 ]]]]]]]] 2 2 (7) cantly different, thus homoscedasticity was not met.
i i
218 A.M. Almeida et al. / J. Chromatogr. B 774 (2002) 215–222

Fig. 1. Generic diagram of weighting schemes for linear regression analysis.

Fig. 2. Residuals plotted against lamotrigine concentrations for the validation intra-day assay.
 A.M. Almeida et al. / J. Chromatogr. B 774 (2002) 215–222 219

Table 1
Test of homogeneity of variances (s 2 ): F-test
Standard A A I.S. Response s2
(mg / l) (A /A I.S. )
0.1 75381 3373583 0.022 5.25310 27
62834 2744443 0.023
68050 2863618 0.024
73212 3190782 0.023
75930 3144582 0.024
15.0 12209191 3244845 3.763 7.60310 22
12953689 2911392 4.449
11682576 2903904 4.023
10940196 2852877 3.835
11147470 2672128 4.172
2 22
s 15.0 7.60 3 10
Fexp 5 ] 5 ]]] 5 1.45 3 10 15
s 20.1 5.25 3 10 27
Ftab (4, 4, 0.99) 5 15.98
A, lamotrigine peak area; A I.S. , internal standard peak area.

4.2. Choice of the weighting factor 5. Discussion

The %RE plots for unweighted (model 1) and Unlike pharmaceutical analysis, the concentration
weighted (models 2–7) regressions of the lamot- range in bioanalytical methods is usually dynamic
rigine intra-day assay data across the whole con- and broad, presenting three or more orders of
centration range are shown in Fig. 3. Model 1 clearly magnitude, in order to monitor concentrations effec-
underestimated the concentrations in the lower range tively [16]. When the range in data values is large, it
of the calibration curve, near the limit of quantifica- might be expected that the variance of each data
tion (LOQ). Models 4 and 7 presented the best %RE point might be quite different. In the present case,
distribution scatter at the lower end of the calibration the concentration data ranged between 0.1 and 15.0
curve. mg / l. Therefore, the test of homoscedasticity was
The regression parameters of the calibration curve carried out.
generated for each weighting factor and the respec- Residual plots can be used to evaluate the need for
tive sums of the relative errors are summarised in weighting when unweighted LSLR is applied. If the
Table 2. The weighting factor 1 /x 2 (model 7) data adequately fit the linear model, then the re-
produced the least sum for this data set providing the siduals should be randomly distributed in a horizon-
most adequate approximation of variance. Thus, the tal band centred on the concentration axis [2]. In this
1 /x 2 weighting factor was chosen. study, the residual plots for unweighted LSLR
clearly showed that the residuals were not randomly
distributed around the concentration axis. Instead, an
4.3. Weighted straight line equation increase in variance as a function of concentration
was observed. This need of weighting LSLR was
For the intra-day assay data the calibration curve confirmed with the results of the F-test.
obtained with the weighting factor 1 /x 2 was y 5 The best weighting factor was chosen taking into
0.2804x 2 0.0049, with r50.999. The accuracy of account either the plots or the sums of the %RE
the data, expressed by bias value, was evaluated calculated for each weighting factor. In the present
across the whole concentration range using weighted study, the 1 /x 2 weighting factor produced the small-
(model 7) and unweighted (model 1) linear regres- est %RE sum, and the most random distribution
sion. The results are shown in Table 3. around the x-axis at the lower end of the calibration
220 A.M. Almeida et al. / J. Chromatogr. B 774 (2002) 215–222

Fig. 3. Percentage of relative error (%RE) versus concentration obtained for model 1 (w i 5 1), model 2 (w i 5 1 /y 1 / 2 ), model 3 (w i 5 1 /y),
model 4 (w i 5 1 /y 2 ), model 5 (w i 5 1 /x 1 / 2 ), model 6 (w i 5 1 /x) and model 7 (w i 5 1 /x 2 ).
 A.M. Almeida et al. / J. Chromatogr. B 774 (2002) 215–222 221

Table 2 especially at the lower end of the calibration range

Regression parameters of the calibration curve generated for each (Table 3). Consequently, weighted linear regression
weighting factor (w i ) and the respective sum of the relative errors
(o%RE) for the intra-day assay data; n530 analysis was able to show a lower limit of quantifica-
tion (LOQ), the lowest concentration on the standard
Model wi b a r o%RE
curve that can be measured with acceptable accuracy
1 1 0.2713 10.0281 0.997 788.90 and precision [20]. Szabo et al. also referred to a
2 1 /y 1 / 2 0.2743 10.0033 0.998 264.31 reduction of LOQ by using weighted linear analysis
3 1 /y 0.2762 20.0034 0.998 133.48
4 1 /y 2 0.2789 20.0048 0.998 116.47 [3]. A broad linear dynamic calibration range could
5 1 /x 1 / 2 0.2747 10.0037 0.998 271.34 be used by this means, with a higher degree of
6 1 /x 0.2770 20.0033 0.998 133.71 accuracy [1,2], which may contribute to improve
7 1 /x 2 0.2804 20.0049 0.999 112.38 precision in laboratorial routine analysis [4]. This is
of the utmost importance, taking into account that
curve (Table 2 and Fig. 3). Additionally, as happens the accurate quantification of low serum concen-
with the majority of bioanalytical methods [10], the trations versus time is particularly relevant in phar-
SD of the response was proportional to the con- macokinetic and pharmacodynamic studies.
centration (Fig. 2). According to all these circum- Although linear regression is the most frequently
stances, the chosen weighting factor was 1 /x 2 . used approach for determining a best-fit calibration
In this work, the calibration model was chosen line, providing a most efficient way to fit experimen-
during validation. Although the choice of the cali- tal data to an appropriate model, it should be assured
bration model is currently normally included in the that correctness of the mathematical model is as-
pre-validation phase [17,18], it has been more recent- sumed. A more complete analysis of the regression
ly recommended to include it during the validation approach (WLSLR) should be considered, taking
procedure, by using all validation samples and into account the error pattern of the data. Although
individual calibration curves in several batches in WLSLR is more complex and laborious than or-
order to simulate the real conditions of routine dinary linear regression, involving the use of statisti-
analysis [19]. cal tests and mathematical operations, it should be
The comparison study between the weighted least performed in order to obtain more realistic results.
squares procedure and the conventional least squares In addition, this paper brings out an issue referred
calibration revealed useful improvements in accura- to but not specified in the most recent FDA guide-
cy, which was particularly evident at the lower end lines for bioanalytical methods validation [21]. Ac-
of the range, as expected. Percentage bias was cording to these guidelines, the ‘‘selection of weight-
considerably greater than acceptable limits of 620% ing and use of a complex regression equation should
[20], when simple least squares regression was used, be justified’’. The present paper may contribute to

Table 3
Comparison of the accuracy obtained by using unweighted (w i 5 1) and weighted (w i 5 1 /x 2 ) linear regression in lamotrigine assay
validation a
Nominal % Bias
concentration
Model 1 Model 7
(mg / l)
Unweighted (w i 5 1) Weighted (w i 5 1 /x 2 )
y 5 0.2713x 2 0.0281 y 5 0.2804x 2 0.0049
0.1 2117.99 10.28
0.5 223.38 22.33
2.5 11.31 12.73
5.0 14.50 13.46
10.0 11.64 20.48
15.0 21.21 23.63
a
Values represent %bias obtained from five replicates; %bias 5 (Cmean 2 Cnominal ) /Cnominal 3 100.
222 A.M. Almeida et al. / J. Chromatogr. B 774 (2002) 215–222

the justification required by the regulators, as it [10] J. Pateman, Standard curve, in: H.H. Blume, K.K. Midha
(Eds.), Bioavailability, Bioequivalence and Pharmacokinetic
shows when, why and how to use weighting schemes
Studies, Medpharm Scientific, Stuttgart, 1995, p. 399.
for practising bioanalysts. [11] Analytical Methods Committee, Analyst 119 (1994) 2363.
[12] M.J. Cardone, S.A. Willavize, M.E. Lacy, Pharm. Res. 7
(1990) 154.
Acknowledgements [13] C. Hartmann, W. Penninckx, Y. Vander Heyden, P. Van-
keerberghen, D.L. Massart, R.D. Mcdowall, Experience with
chromatographic methods—Europe, in: H.H. Blume, K.K.
A.M. Almeida was supported by PRAXIS XXI Midha (Eds.), Bioavailability, Bioequivalence and Phar-
BD/ 16288 / 98 and M.M. Castel-Branco by PRAXIS macokinetic Studies, Medpharm Scientific, Stuttgart, 1995,
XXI BD/ 18351 / 98. p. 331.
[14] Water Quality—Calibration and Evaluation of Analytical
Methods and Estimation of Performance Characteristics—
Part 1: Statistical Evaluation of the Linear Calibration
References Function, International Standard—ISO 8466-1:1990 (E),
ISO, Geneve, 1990, p. 1.
[1] E.L. Johnson, D.L. Reynolds, D.S. Wright, L.A. Pachla, J. [15] R. Levie, J. Chem. Educ. 63 (1986) 10.
Chromatogr. Sci. 26 (1988) 372. [16] A.R. Buick, M.V. Doig, S.C. Jeal, G.S. Land, R.D. Mcdowall,
[2] H.T. Karnes, G. Shiu, V.P. Shah, Pharm. Res. 8 (1991) 421. J. Pharm. Biomed. Anal. 8 (1990) 629.
[3] G.K. Szabo, H.K. Browne, A. Ajami, E.G. Josephs, J. Clin. [17] J.R. Lang, S. Bolton, J. Pharm. Biomed. Anal. 9 (1991) 357.
Pharmacol. 34 (1994) 242. [18] J. Wieling, G. Hendriks, W.J. Tamminga, J. Hempenius, C.K.
[4] J.S. Garden, D.G. Mitchell, W.N. Mills, Anal. Chem. 52 Mensink, B. Oosterhuis, J.H.G. Jonkman, J. Chromatogr. A
(1980) 2310. 730 (1996) 381.
[5] M. Thompson, Analyst 113 (1988) 1579. [19] N.V. Nagaraja, J.K. Paliwal, R.C. Gupta, J. Pharm. Biomed.
[6] J.N. Miller, Analyst 116 (1991) 3. Anal. 20 (1999) 433.
[7] M. Mulholland, D.B. Hibbert, J. Chromatogr. A 762 (1997) [20] V.P. Shah, K.K. Midha, S. Dighe, I.Y. Mcgilveray, J.P.
73. Skelly, A. Yacobi, T. Layloff, C.T. Viswanathan, C.E. Cook,
[8] M.M. Castel-Branco, A.M. Almeida, A.C. Falcao, ˜ T.A. R.D. Mcdowall, K.A. Pittman, S. Spector, Eur. J. Drug
Macedo, M.M. Caramona, F.G. Lopez, J. Chromatogr. B 755 Metab. Pharmacokinet. 16 (1991) 249.
(2001) 119. [21] Guidance for Industry. Bioanalytical Methods Validation,
[9] W.W. Daniel (Ed.), Biostatistics: a Foundation for Analysis in United States Food and Drug Administration, Washington,
the Health Sciences, John Wiley, 1987, p. 367. DC, 2001, p. 6.