1.

05 Quality of Analytical Measurements: Univariate

Regression
M. C. Ortiz, M. S. Sánchez, and L. A. Sarabia, University of Burgos, Burgos, Spain
ª 2009 Elsevier B.V. All rights reserved.

1.05.1 Introduction 128

1.05.2 Linear Regression in Calibration: Elements and Procedure 129
1.05.2.1 Estimation by Least Squares 129
1.05.2.2 Estimation when the Distribution of " Follows a N(0, 2) 131
1.05.2.3 The Correlation between Concentration, x, and Signal, y 133
1.05.3 Statistical Validation of a Calibration Model 134
1.05.3.1 Validity of the Functional Model 134
1.05.3.1.1 F-test for significance of regression 134
1.05.3.1.2 Lack of fit test 135
1.05.3.2 Verification of the Hypotheses on the Residuals 136
1.05.3.2.1 Graphical methods 137
1.05.3.2.2 Inferential methods 138
1.05.4 Confidence Intervals and Hypothesis Testing 140
1.05.4.1 Confidence Intervals for the Intercept, the Slope, and the Residual Variance 140
1.05.4.2 Joint Confidence Region for  0 and  1 141
1.05.4.3 Comparison of Two Regression Lines 142
1.05.4.4 Confidence Interval for the Prediction 144
1.05.4.5 Confidence Interval for the Determination of the Concentration of a Problem Sample 145
1.05.4.6 The Intersection of Two Linear Regression Lines 147
1.05.5 The Design of a Calibration 148
1.05.6 The Capability of Detection, the Decision Limit, and the Capability of Discrimination
Computed from a Regression Model 151
1.05.7 Standard Addition Method 153
1.05.8 Weighted Least Squares and Generalized Least Squares 154
1.05.9 Robust Regression in Calibration 159
1.05.10 Errors in Both Variables 163
1.05.11 Final Remark 164
References 165

Symbols H0 null hypothesis

bj estimated coefficients N number of observations (objects)
CC capability of detection (European N(,2) normal distribution with mean  and
Decision 2002/657) variance 2
Cov(X,Y) covariance between random variables NID(,2) independent random variables equally
X and Y distributed as normal with mean  and
E(X) expectation of random variable X variance 2
ei residuals rxy sample correlation between x and y
F,v1,v2 upper -perecentage point of an F s2p pooled estimated common variance
distribution with v1 and v2 degrees of s2x variance (of x’s data)
freedom so that  ¼ pr{x > F,v1,v2} s2yx residual variance of the regression line
Fv1,v2 F distribution with v1 and v2 degrees of t,v upper -percentage point of a stu-
freedom dent’s t distribution with v degrees of
Ha alternative hypothesis freedom so that  = pr{x > t,v}

127
128 Quality of Analytical Measurements: Univariate Regression

tv student’s t distribution with v degrees  probability of type II error, probability

of freedom of false negative
Var(X) variance of random variable X j true coefficients
x independent (predictor) variable " random variable (experimental error)
xc decision limit (ISO 11843-1)  true mean
xd capability of detection (ISO 11843-1) xy correlation coefficient between x and y
x mean (of x’s data) 2 true variance
x̂i predicted value of xi 2v (chi-squared) distribution with v
y dependent (response) variable degrees of freedom
ŷi predicted value of yi 2,v upper -percentage point of a chi-
 significance level, probability of type I squared distribution with v degrees of
error, probability of false positive freedom so that  ¼ pr{x > 2,v}

1.05.1 Introduction

In Dux,1 a calibration model is defined as the verification of the response of an instrument to a material
of known properties and, maybe, the correction by a factor to take the instrument to the corresponding
mark, whereas standardization means to characterize the response of an instrument according to the
known properties of the material. The standardization of an instrument is usually done by the ‘calibration
curve’. These two terms are often mistakenly used interchangeably. Many of the present-day instruments
consist of different systems and their complexity makes it difficult to calibrate them. In fact, what is done
is standardizing the response by means of a series of samples of known concentration. These standards
of concentration should be of enough purity, or a reference material with certified quality characteristics
should be used.
Many different meanings have been attributed to the term calibration: for example, the International Union
of Pure and Applied Chemistry (IUPAC)2 recognizes two different versions:
1. Calibration function for species identification and qualitative analysis. That is, calibration of those analytical
parameters that characterize types of chemical species by establishing a model for the purpose of identifica-
tion and qualitative analysis (for instance, a model that identifies a compound on the basis of its
chromatographic retention time).
2. Calibration function for quantitative analysis, whose main purpose is to obtain a function that allows one
to calculate the concentration (or amount) of an analyte as a function of an instrumental signal. In most cases,
the calibration function has to take into account the response relations of all relevant constituents and
interferents.
In the document from the International Organization for Standardization (ISO)3 related to the capability
of detection, the practical relevance of the calibration is highlighted: ‘‘The quantitative calibration of a system
of measurement is central in many chemical measurements. It is relevant for the trazability of the measurement,
contributes to its uncertainty and many figures of merit of the analytical methods are directly related to the
calibration.’’
The calibration curve is fit, usually, by the statistical technique of linear regression, which is a particular case
of the more general framework of fitting models to experimental data.
To fit empirical models involves specific peculiarities caused by the type of data used for the fitting; for this
reason, a net chemometric approach has been adopted that emphasizes its utility for the calibration, the
determination of the sensitivity of a method, the linear range, the detection of possible matrix and interferent
effects, the comparison of two methods, the calculation of a detection limit, etc.
 Quality of Analytical Measurements: Univariate Regression 129

1.05.2 Linear Regression in Calibration: Elements and Procedure

We will establish the notation and the basic aspects of regression in the simplest case of univariate linear
regression. Let x denote the independent variable and y the response. When calibrating, x will be the
concentration of the analyte in each standard sample and y denotes the recorded signal. The random error
supposed for the response is denoted as ". The model for the observed experimental responses is supposed to be
the following:
y ¼ f ðx Þ þ " ð1Þ
where f(x) is the calibration curve to be estimated. To do this, N observations are recorded, yi, that correspond to
k calibration standards. If there are no replicate observations, then N ¼ k, otherwise N > k. As usual, for each yi,
the corresponding concentration of the standard is written as xi, without distinguishing whether they are equal
(replicates) or not. Therefore,
yi ¼ f ðxi Þ þ "i ; i ¼ 1; . . .; N ð2Þ
In many calibration problems, the following hypotheses are considered whose compatibility with the experi-
mental results will have to be carefully evaluated:
1. The true calibration curve f is linear in the studied range of concentrations. Hence
yi ¼ 0 þ 1 xi þ "i ; i ¼ 1; . . .; N ð3Þ
2. The values of the predictor variable (xi, i ¼ 1,2 , . . ., N) are controlled values and their experimental
uncertainty can be neglected compared with the uncertainty in the determination of the response.
3. The random errors in the observed response are uncorrelated, Covð"i ; "j Þ ¼ 0; if i 6¼ j , have zero mean,
E ð"i Þ ¼ 0; i ¼ 1; . . .; N , and the same variance, Var ð"i Þ ¼ 2 ; i ¼ 1; . . .; N , independent of 0, 1, and xi.
As a result of the hypotheses on the errors, each response yi is a random variable whose mean is a linear function
of the value of the predictor E ðyi Þ ¼ 0 þ 1 xi , its variance is Varðyi Þ ¼ 2 , and the observed values (y1, . . ., yN)
are uncorrelated, that is, Covðyi ; yj Þ ¼ 0; if i 6¼ j .
Therefore, the problem of estimating the model of Equation (1), or calibration line, is technically equivalent
to estimating the parameters  0, 1, and  from the experimental results of the calibration set
C ¼ fðxi ; yi Þ; i ¼ 1; . . .; N g.
To propose a calibration model for experimental data, as the one in Equation (3), is the consequence of the
theoretical knowledge of the problem, but its validity has to be verified on the basis of a critical examination of
the experimental results. The model, f(x), is a working hypothesis and must be modified if the experimental
data are against it. However, the model is not directly observable, and therefore the decision on its validity
depends on the success when distributing the experimental response in the two addends of Equation (1):
random part and functional part. In other words, the hypotheses on the random error, which are those described
in previous hypothesis (3), should also be made explicit and validated.
When we say that a model is linear or nonlinear, we are referring to linearity or nonlinearity in the
parameters. The value of the highest power of a predictor variable in the model is called the order of the
model. For example,
y ¼ 0 þ 1 x þ 2 x 2 þ " i ð4Þ
is a second-order (in x) linear (in the ’s) regression model. The calibration model of Equation (4) is quadratic;
it is not a linear calibration but a linear regression.

1.05.2.1 Estimation by Least Squares

The least squares (LS) method does not require any hypotheses under the distribution of the errors, ", in
Equation (1). To obtain the estimates b0 and b1 of  0 and  1, respectively, the straight line that best describes the
experimental data is determined. The argument is as follows: for each couple of values b0 and b1, a different line
130 Quality of Analytical Measurements: Univariate Regression

yˆ = b0 + b1x

eN
ei
e2
yˆ 1

e1
The linear calibration line fitted by
least squares is the one that makes
y1 minimum the sum of squares of all
these vertical distances

x
Figure 1 Illustration of least squares method to fit a calibration line.

is obtained, ŷ ¼ b0 þ b1 x , with the corresponding residuals ei ¼ yi – ŷi ; i ¼ 1; . . .; N . The LS estimators of  0

and 1 are the values that minimize the sum of squares (SSs) of residuals,
X
N X
N X
N
Rðb0 ; b1 Þ ¼ ei 2 ¼ ðyi – ŷi Þ2 ¼ ðyi – b0 – b1 xi Þ2 ð5Þ
i¼1 i¼1 i¼1

The line obtained when minimizing R as a function of b0 and b1 is the one that minimizes the sum of the
distances ‘in vertical’ between the experimental data and the estimated ones with the fitted line (see Figure 1).
It can be analytically proven that the function Rðb0 ; b1 Þ has a unique minimum (except for
x1 ¼ x2 ¼ . . . ¼ xN , a clearly nonsense situation when calibrating) that can be explicitly determined. This
solution provides the LS estimators as
PN
i¼1 ðyi – yÞðxi – xÞ Covðx; y Þ
b1 ¼ PN ¼ ð6Þ
ð
i¼1 i – x
x Þ2 sx2

b0 ¼ y – b 1 x ð7Þ
 PN   PN 2
with y ¼ ¼
i¼1 yi =N the mean of all yi and x i¼1 xi =N the mean of all xi , and sx denotes the variance
of the x’s.
The procedure for obtaining the minimum shows that the LS line verifies
X
N
ei ŷi ¼ 0 ð8Þ
i¼1

and, if there is a 0 term in the model, as in our case, it is also true that
X
N X
N
ðyi – ŷ i Þ ¼ ei ¼ 0 ð9Þ
i¼1 i¼1

The LS method gives no estimator of  2 , but an estimator of 2 based on the LS estimators of 0 and 1 is the
variance of the regression, defined as
PN PN PN 2
2 i¼1 ðyi – b0 – b1 xi Þ2 i¼1 ðyi – ŷ i Þ2 e
syx ¼ ¼ ¼ i¼1 i ð10Þ
N –2 N –2 N –2
Equations (6) and (7) imply that b0 and b1 are linear combinations of the random variables yi. Consequently, both
estimates themselves are also random variables. As E ðyi Þ ¼ 0 þ 1 xi , from Equation (6), the following
statement holds:
PN PN
i¼1 ðxi – x
ÞE ðyi – yÞ i¼1 ðxi – x
Þð1 ðxi – xÞÞ
E ðb1 Þ ¼ PN 2 ¼ PN ¼ 1 ð11Þ
i¼1 ðxi – x
Þ Þ2
i¼1 ðxi – x
 Quality of Analytical Measurements: Univariate Regression 131

and, analogously from Equation (7),

E ðb0 Þ ¼ Eð
y Þ –  1 x ¼ 0 ð12Þ
Equations (11) and (12) mean that both estimates are unbiased estimates.
The variance of the estimates is given by
PN
i¼1 ðxi – xÞ2 Varðyi Þ 1 2
Varðb1 Þ ¼PN 2 2
¼ PN 2 ð13Þ
ð
i¼1 i x – 
x Þ i¼1 ð x i – 
x Þ
PN 2 !
i¼1 xi 2 1 x2 2
Varðb0 Þ ¼ PN  ¼ þ PN 2  ð14Þ
N i¼1 ðxi – xÞ2 N i¼1 ð x i – xÞ

x 2
Covðb0 ; b1 Þ ¼ – PN 2 ð15Þ
i¼1 ð x i – 
x Þ

Equations (13) and (14) show that the precision attainable in the estimatesP of the coefficients of a calibration
P line
is the product of two factors; one of them depends on the magnitude, Ni¼1 ðxi Þ2 , and dispersion, Ni¼1 ðxi – xÞ2 ,
of the concentrations and the other one, 2, depends on the uncertainty of the observed response. The second
factor is estimated with Equation (10) after making the experimental determinations of the calibration, but the
other factor is previously known and implies the introduction of the idea of designing the calibration, analogous
to any other experiment. This subject, together with other aspects to be taken into account as a consequence of
the validation of the model, will be dealt with in Section 1.05.5.
When analyzing Equation (15) one should remember that, in a calibration, the values xi cannot be negatives;
thus, x > 0 and the covariance (correlation) between b0 and b1 is always negative.
We consider the class of the unbiased estimators of  0 and 1 that are linear functions of the random variables
yi. In this class, the LS estimators b0 and b1 of the linear model of Equation (3) with the three hypotheses detailed
in Section 1.05.2 have smaller variances than any other estimators. This important theorem is often referred to
as the Gauss–Markoff theorem and guarantees the coherence between the linear calibrations by LS and the
accuracy (trueness and precision) demanded by the methods of chemical analysis.

1.05.2.2 Estimation when the Distribution of " Follows a N(0, 2)

The evaluation and comparative analysis of the estimated parameters in a calibration (b0, b1, and syx) is a
problem of statistical inference because all of them are random variables that depend on the distribution of the
error ". The only supposition on the distribution of the errors in the LS method is that they have common
variance 2 independent of the concentration xi. However, it is necessary to dispose of additional information
on their distribution to statistically infer anything about the parameters.
When the distribution of the random error, " in Equation (1), is known, the estimation of 0, 1, and  can be
done by the maximum likelihood method, which, in essence, looks for the values of b0, b1, and s that maximize
the probability of obtaining the experimental results that were observed. The method is based on
the computation of the likelihood function L of the N random variables ðy1 ; y2 ; . . .; yN Þ, that is, the joint
density, g, of these random variables considered as a function of the parameters 0, 1, and . Therefore,
Lð0 ; 1 ; 2 Þ ¼ g ðy1 ; y2 ; . . .; yN ; 0 ; 1 ; 2 Þ. The maximum likelihood estimator is the value of 0, 1, and  that
maximizes L.
If, further to the hypotheses on the errors established in hypothesis in point 3 in Section 1.05.2, the data are
assumed to be normally distributed, the consequence is that for each x, the experimental response is a normal random
variable N ð0 þ 1 x; 2 Þ. Furthermore, as Covðyi ; yj Þ ¼ 0; if i 6¼ j , the random vector ðy1 ; y2 ; . . .; yN Þt follows a
multivariate normal distribution N ðh; 2 IÞ with vector of means h ¼ ð0 þ 1 x1 ; 0 þ 1 x2 ; . . .; 0 þ 1 xN Þt
and the variance–covariance matrix, 2I, which is a diagonal matrix with all the terms equal to 2 .
Figure 2 shows the meaning of the hypotheses assumed; for each concentration xi and due to the random
character of ", any value of the signal can be obtained but those values close to  0þ 1xi are more probable,
because this number is the expected mean value when using a probability density function of a normal
distribution, independent from one concentration to another and with common variance 2 .
132 Quality of Analytical Measurements: Univariate Regression

y
y = β0 + β1x

•
•
• • •
• •
•

Xi XN x

X1

Figure 2 Calibration model of Equation (3), assuming that the recorded signal comes from a normal distribution.

It can be proven that the likelihood function L is maximum when the sum of squares, R in Equation (5), is
minimum. Consequently, the estimates b0 (Equation (7)) and b1 (Equation (6)) have the property of being
maximum likelihood estimates. For 2 , the maximum likelihood estimate is
PN 2
i¼1 ei
s2 ¼ ð16Þ
N
which is biased, contrary to s2yx (Equation (10)). In any case, the relation ½N=ðN – 2Þs 2 ¼ syx2 holds.
As a consequence of supposing normal distribution for the random errors, the following properties are
obtained:
1. The estimates b0, b1, and syx2 (Equations (7), (6), and (10)) are of minimum variance in the class of all unbiased
estimators, which is a class wider than the class of unbiased linear estimators.
2. Ns 2 =2 follows a 2 distribution with N–2 degrees of freedom (d.f.), so ðN – 2Þsyx2
=2 also does.
3. b0 and b1 follow a bivariate normal distribution with covariance matrix whose elements are given by
Equations (13)–(15).
4. b0 and b1 are independent of s 2 and, hence, of syx 2
.
A last comment, if another distribution is supposed for the random term of the calibration line in Equation (3),
the maximum likelihood estimate of 0, 1, and  is obtained with other criteria. For example, instead of the
normal distribution, let us suppose that the errors were independent and followed the Laplace distribution with
null mean and variance 22, that is, its probability density function is
1 – j"i j=
f ð"i Þ ¼ e ð – 1  "i  1Þ ð17Þ
2
Then, application of the maximum likelihood principle for estimating  0 and 1 would involve the minimiza-
tion of the sum of absolute errors
X
N X
N
j ei j ¼ jyi – ŷ i j ð18Þ
i¼1 i¼1

Of course, the LS solution is still available but it is not the maximum likelihood solution; thus, the statistical
validity of the conclusions will be wrong.
Consequently, the success in the selection of the distribution of the error is essential for a correct
determination of the calibration line, including the criterion that should be followed to obtain the estimators.
In the rest of the chapter, unless otherwise indicated, we will assume normal and independent residuals with
zero mean and equal variance, that is, " is a NID(0,2) distribution.
 Quality of Analytical Measurements: Univariate Regression 133

1.05.2.3 The Correlation between Concentration, x, and Signal, y

Usually, the correlation coefficient is used as an index of the linearity of a regression model. The validity of its
application to a calibration is a subject of controversy because it seems to be valid only when it is applied with
two random variables and the concentration of the standard samples is not random. In the following, this subject
is analyzed.
We consider two random variables, v and w, which follow some continuous joint bivariate probability
distribution. The correlation coefficient between v and w is defined by
Covðv; wÞ
vw ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffipffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð19Þ
VarðvÞ VarðwÞ

For example, from Equations (13)–(15), the correlation coefficient between the two random variables b0 and
b1 is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N
b0 b1 ¼ – x PN 2 ð20Þ
i¼1 xi

which depends only on the values of the concentration of the calibration samples and which is completely
independent of the recorded signals, yi.
In general, – 1  vw  1. The quantity vw is a measure of the linear association between the random
variables v and w. If vw ¼ 1, v and w are perfectly positively correlated and their values all lie on a straight line
with a positive slope in the (v, w) plane. If vw ¼ 0, the variables are said to be uncorrelated, that is, linearly
unassociated with each other; note that this does not mean that v and w are statistically independent. If
vw ¼ – 1, v and w are perfectly negatively correlated and their values again lie on a straight line, but this
time with a negative slope.
If a sample of size Nfðvi ; wi Þ; i ¼ 1; :::; N g is available from the joint distribution, the quantity
PN
i¼1 ðvi – v Þðwi – w Þ
rvw ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN ffiqffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN ð21Þ
2
i¼1 ðvi – v Þ i¼1 ðwi – w  Þ2

is called the sample correlation coefficient between v and w. It is an estimate of vw and provides an empirical
measure of the linear association between v and w. Like vw , – 1  rvw  1.
When ðvi ; wi Þ; i ¼ 1; :::; N , are all nonrandom values, that is, they are not sample values from some
distribution, rvw can still be used as a measure of linear association, because the values ðvi ; wi Þ; i ¼ 1; :::; N ,
can be considered as a finite discrete bivariate distribution with uniform marginals. In this case, rvw is a
population rather than a sample estimation, that is, vw ¼ rvw .
In our case, xi, i ¼ 1, . . ., N, are nonrandom values that represent the values of a finite discrete distribution for
concentrations and the corresponding recorded signals yi, i ¼ 1, . . ., N, are observed random variables. As a
consequence, it is possible to define xy by Equation (19). Equation (21), with x and y replacing v and w, can still
be used to estimate yx by ryx.
There is another permanent discussion about the usefulness or uselessness of the correlation coefficient to
evaluate the linearity of a calibration.4–8 The paper by Asuero et al.9 is an updated review on the correlation
coefficient and its use either as a way to infer correlation between variables or to test linearity with chemical
applications.
Here, we will only show that between the coefficient of the regression (the slope of the calibration, b1) and the
correlation coefficient ryx, the following relation holds:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ffi
pffiffiffiffiffiffiffiffiffiffiffiffiffi
ðyi – yÞ2 =ðN – 1Þ Varðy Þ
b1 ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ryx ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi ryx ð22Þ
ðxi – xÞ2 =ðN – 1Þ Varðx Þ

Thus, b1 and ryx are closely related, although they provide different information about the calibration line. The
unit-free and scale-free correlation coefficient ryx measures the degree of linear association between
134 Quality of Analytical Measurements: Univariate Regression

concentration and signal, whereas b1 measures the size of the change in signal when a unit change is made in
concentration. Scale changes in the data will affect b1 but not ryx. It is easy to see that
 
ryx  ¼ ryŷ ð23Þ

That is, the correlation between the recorded signals yi and the estimated ones ŷi is the same (in absolute value)
as the correlation between the concentrations xi and the signals yi.

1.05.3 Statistical Validation of a Calibration Model

1.05.3.1 Validity of the Functional Model
1.05.3.1.1 F-test for significance of regression
Obviously, the dispersion of the recorded signal has two sources: the one due to the regression and the residual
one. If the dispersion attributable to the regression is higher, in statistical terms, than the residual one, the
conclusion is that the model is significant, that is to say, it explains the recorded signals.
The variation in the y’s about their mean can be written as
X
N X
N X
N X
N
ðyi – yÞ2 ¼ ððyi – ŷ i Þ þ ðŷ i – yÞÞ2 ¼ ðyi – ŷ i Þ2 þ ðŷ i – yÞ2 ð24Þ
i¼1 i¼1 i¼1 i¼1

because the crossproduct term is null. Therefore, the total variation of the y’s, SST, is decomposed
P into the
variation that can be ascribed to the regression line, SSR, and the residual variation, SSE ¼ Ni¼1 ðyi – ŷi Þ2 , due
to the fact that the experimental observations do not all lie on the regression line. In a compact notation,
Equation (24) is
SST ¼ SS E þ SSR ð25Þ
2
If SSR  SSE (equivalently, if R ¼ SSR =SST is near one), it is clear that the regression explains the recorded
signal as a function of the concentration.
The degrees of freedom (d.f.) associated with any sum of squares are the number of independent pieces of
information that we need for its calculation. To compute SST, there are only N1 independent data because it
is necessary to calculate y. The SSE has N2 d.f. because we need to previously compute b0 and b1. By
subtraction, SSR has 1 d.f. For each term in Equation (25), the mean of squares (MSs) is defined as the quotient
between the sum of squares and its d.f., MS ¼ SS=d:f : The result of these computations is presented as an
analysis of variance (ANOVA) table like Table 1.
MSR and s2yx ¼ MSE, which are shown in Table 1, are random variables because they are obtained from the
variables yi. It can be shown that their mean values are as follows:
X
N
E ðMSR Þ ¼ 2 þ 12 ðxi – xÞ2 ð26Þ
i¼1
 
2
E ðMSE Þ ¼ E syx ¼ 2 ð27Þ

In addition, if "i are NID(0,2) and 1 ¼ 0, the random variable

MSR
Fcalc ¼ ð28Þ
MSE

Table 1 Analysis of variance for significance of a linear regression

Source of variation Sum of squares Degrees of freedom Mean of squares Fcalc

Regression SSR 1 MSR MSR/MSE

Error or residual SSE N2 MSE
Total SST N1
 Quality of Analytical Measurements: Univariate Regression 135

follows an F-distribution with 1 and N2 degree(s) of freedom in the numerator and denominator, respectively.
If there is a linear relation between signal and concentration, the regression coefficient 1 must be different
from zero. The appropriate hypotheses are
H0 :1 ¼ 0
ð29Þ
Ha : 1 6¼ 0
and the null hypothesis would be rejected at significance level  if Fcalc ¼ MSR =MSE exceeds the critical value
at level , F,1,N2. In other words, the linear calibration is statistically significant at level . Instead of the
critical value, it is usual to compute the probability pr{F1,N2 > Fcalc} called p-value. In this case, the null
hypothesis is rejected (the model is significant) at level  if the p-value is smaller than .
In a calibration experiment, the null hypothesis of the test for significance of the regression is almost always
rejected; hence the model is significant because the calibration is, by definition, based on the fact that the
response of the instrument changes with the concentration of the standard solutions.
Two additional relationships should be noted. In the explanation of Equation (25), we have defined
R2 ¼ SSR =SST as an indicator of the variability of the signal that is explained by the variation of the
concentration. R2 is called the determination coefficient and verifies for a fitted straight line that
ryx ¼ ðsign of b1 ÞR ð30Þ

where R is the positive square root of R2. Then, from Equation (23) we also have that,
ryŷ ¼ R ð31Þ

1.05.3.1.2 Lack of fit test

More interesting in practice is the lack of fit (LOF) test that allows us to decide if the supposed linear model in
Equation (3) is compatible with the experimental data. The only information available on the discrepancy
between the model and the experimental results is provided by the residuals of the fitting yi – ŷ i , which can be
decomposed as
yi – ŷ i ¼ ðyi – ŷ i Þ – ðE ðyi Þ – E ðŷ i ÞÞ þ ðE ðyi Þ – E ðŷ i ÞÞ ¼ ððyi – ŷ i Þ – E ðyi – ŷ i ÞÞ þ ðE ðyi Þ – E ðŷ i ÞÞ ¼ qi þ Bi ð32Þ
In Equation (32), Bi is the bias in the estimation when the concentration is xi. If the model is correct, then the
bias is zero. Otherwise, Bi 6¼ 0 and its value depends on xi and on the postulated
PN 2  model. The2 other term, the
random variables qi, i ¼ 1, . . ., N, have null mean, are correlated, and E i¼1 qi ¼ ðN – 2Þ . Consequently,
from Equation (32), it is deduced that MSE defined as
PN
SSE – ŷ i Þ2
i¼1 ðyi
MSE ¼ ¼ ð33Þ
N –2 N –2
P
is a random variable with mean 2 if the model is correct and 2 þ Ni¼1 Bi2 =ðN – 2Þ if the model is biased. Note that
2
if the assumed model for the calibration is not correct, syx ¼ MSE is an overestimation of 2 ¼ Varð"i Þ ¼ Varðyi Þ
because it includes the bias of the model.
The question is to have another independent estimate of 2 , for example, by previous knowledge of the
system of measurement, and comparing it with MSE. If the latter is significantly greater, we say that there
is LOF and we would reconsider the model for the calibration line. If no prior estimate of 2 is available, a
usual procedure in calibration is to repeat measurements of the standard samples in such a way that we can
use these replicates to obtain an estimate of 2 . It is said that this estimation represents the pure error
because if xi is the same for several observations, then the variability among the observed responses is due
to random fluctuations of measurements alone. It is important to understand that the replicated runs must
be genuine replicates and not just repetitions of the same instrumental reading over the same analytical
sample. The number and distribution of replicates in the concentration range is an important practical
issue because it influences the variance of the estimates made from the calibration line and will be studied
in Section 1.05.5.
136 Quality of Analytical Measurements: Univariate Regression

Pk
Suppose that in each calibration standard xi (i ¼ 1,P 2 yij are recorded so that i¼1 mi ¼ N . The
. . . ,k) mi signals
mi
internal variability of these observations is given by j ¼1 yij – yi with mi1 d.f. Therefore, pooling over the k
levels xi, we have the so-called sum of squares of pure error:
k X
X mi  2
SSPE ¼ yij – yi ð34Þ
i¼1 j ¼1
P
with ki¼1 mi – 1 ¼ N – k d.f. Dividing SSPE by its d.f. we have MSPE that verifies E(MSPE) ¼ 2.
On the other side, we have
X
N  k X
2 X mi
 2
SSE ¼ yj – ŷ j ¼ yij – ŷ i
j ¼1 i¼1 j ¼1
ð35Þ
k X
X mi  2 X
k
¼ yij – yi þ mi ðyi – ŷ i Þ2 ¼ SSPE þSSLOF
i¼1 j ¼1 i¼1

In Equation (35), we can see that the sum of squares for LOF, SSLOF, is a weighted sum of squared deviations
between the mean response yi at each xi level and the corresponding fitted value ŷi .
Table 2 shows the decomposition of the sum of squares of Equation (35) with the corresponding d.f.
If Fcalc ¼ MSLOF =MSPE is greater than F,k-2,Nk, the null hypothesis is rejected at level , that is, we
conclude that the model proposed for the calibration line has LOF (is biased). Alternatively, if Fcalc does not
exceed F,k-2,Nk, then there is no experimental evidence for LOF, and MSPE and MSLOF are often combined to
estimate 2.

1.05.3.2 Verification of the Hypotheses on the Residuals

The residuals are the differences ei ¼ yi – ŷi ; i ¼ 1; . . .; N , where yi are the experimental values and ŷi the
corresponding fitted values. The random variables "i are not observable. The intuitive idea is that, if the calibration
model is correct, the residuals will show tendencies that make us think that "i are effectively NID(0,2).
The difficulty arises when expressing ei as a function of "i :10
XN
ei ¼ ð1 – hii Þ"i – j ¼1; j 6¼i
hij "j ð36Þ

with
 
1 ðxi – xÞ xj – x
hij ¼ þ PN ð37Þ
N Þ2
i¼1 ðxi – x

and the NID assumption regarding "i implies that the ei must be normally distributed with
Varðei Þ ¼ ð1 – hii Þ2 ð38Þ
Thus, unless the concentration values are only two, the variance of the residuals Varðei Þ will be changing and
will differ from 2 . To avoid this effect, which can be important if the calibration range is large, it is better to use
the studentized residuals defined as

Table 2 Analysis of variance for significance and lack of fit of a linear regression

Source of variation Sum of squares Degrees of freedom Mean of squares Fcalc

Regression SSR 1 MSR MSR/MSE

Error or residual SSE N2 MSE
Lack of fit SSLOF k2 MSLOF MSLOF/MSPE
Pure error SSPE Nk MSPE
Total SST N1
 Quality of Analytical Measurements: Univariate Regression 137

ei
ei 9 ¼ pffiffiffiffiffiffiffiffiffiffiffi ð39Þ
syx 1 – hii

As the studentized residuals come from a normal distribution with variance one, they have to vary between 3
and 3.

1.05.3.2.1 Graphical methods

Usual graphics are the plots of residuals (or studentized residuals) versus the following: (1) the order of
obtaining the results (possible bias due to the sequence of the experiments); (2) the responses ŷi computed
with the model; and (3) the concentrations.
In these plots, typical patterns like those of Figure 3 can be observed, which not only alert on the possible
failure of the hypotheses but also aid in deciding on the actions to be taken to restore them.
Figure 3(a) shows a suitable distribution of the residuals, whereas the rest display certain anomalies. For
example, a tendency like the one observed in Figure 3(b), the variance of the residuals increases when
increasing the variable represented in the abscissa, indicates a lack of homoscedasticity. Nevertheless, the
tendency of Figure 3(c) shows a bias attributable to an inadequate model; a model of order 2 should be chosen
as the calibration model. Figure 3(d) shows both anomalies, a failure of equality of variances and also a bias of
the model.
To evaluate normality, it is useful to draw the graph of normality of studentized residuals, which consists of
plotting the cumulative relative frequency of each residual according to a normal distribution versus the
residual itself. If the distribution is normal, the points have to appear aligned; additional details can be seen in
Draper and Smith.11 When, instead of using the studentized residuals, their absolute values are used, the
resulting graph is called a half-normal plot. The graphical possibilities of computers facilitate the drawing of
these plots because the vertical scaling should be determined from the inverse function of a normal distribution.
When " is not normally distributed, an additional problem can occur. Each hii is different for each estimated
residual and, when it is large related to the residual itself, the corresponding ei will be approximately normally
distributed because of the central limit theorem. When hii is small, the distribution of ei would be similar to that
of "i, although closer to a normal distribution. Thus, we have the disturbing result that the residuals can have a
distribution that is closer to normal than the distribution of the errors (even if they are not normally
distributed). This property is known as supernormality by many authors. To rectify this problem, the use of

(a) 2 (b) 2

1 1

0 0

–1 –1

–2 –2

(c) 2 (d) 2

1
1
0
0
–1

–1 –2
Figure 3 Schematic view of some possible tendencies on the residuals: (a) no tendency; (b) lack of homoscedasticity;
(c) wrong model; and (d) both heteroscedasticity and wrong selection of model.
138 Quality of Analytical Measurements: Univariate Regression

a ‘simulation envelope’ is proposed for a half-normal plot together with the use of the ‘externally studentized
pffiffiffiffiffiffiffiffiffiffiffi
residuals’ defined as ei ¼ ei =syxðiÞ 1 – hii , where syxðiÞ is the residual standard deviation when in the regression
model the pair (xi, yi) is not considered (is suppressed).
The envelope boundaries are made up by the estimators of the 5th and 95th percentiles of the distribution for
each ei obtained by bootstrap. These simulated residuals are normally distributed and the boundaries
constructed essentially provide a limit on how non-normal the residuals can appear when the errors have
normal distribution. This technique has been deeply studied and there are variations. A review of the subject
with examples of application can be seen in Ryan.12

1.05.3.2.2 Inferential methods

1.05.3.2.2(i) Serial correlation The Durbin–Watson test is usually used to verify if each residual is
independent of the previous one and therefore if some relation between the results exists related to the
concentration of the standards or the sequence in which they were measured. This is frequent in the laboratory
because of indirect causes, for example, degradation of reagents and instrumental drift.
From the Durbin–Watson test, some procedures have been developed to detect the lack of linearity in real
calibration experiments. A comparative analysis, including a recent proposal, is made in the paper by
Brüggemann et al.13

1.05.3.2.2(ii) Normality The usual tests are those of Kolmogoroff and 2. The Kolmogoroff test detects
modifications in the form of the empirical histogram made with n data with respect to the theoretical one, in our
case with respect to that of a normal distribution.
The 2 test, similar to the Kolmogoroff test, detects whether the histogram of some experimental data
adheres suitably to the one of a theoretical distribution (in our case a normal one). Unlike the previous one, the
data are grouped into k intervals so that it demands greater sample sizes.

1.05.3.2.2(iii) Homoscedasticity (equal variances) To be able to decide whether the variance of the
responses stays or is nonequal for the different values of the independent variable, it is necessary to have several
measured responses yi for the same value of the predicting variable xi, at least in two levels of concentration.
The Bartlett test and the Cochran test are both adequate and, to a great extent, complementary.
The Bartlett test detects the existence of groups of values of the independent variable with variances in the
response similar to each other but different from one group to another. It is very sensitive to the failure in the
hypothesis of normality. The test of Cochran detects the existence of one discrepant variance.

Example 1
To determine cadmium by anodic stripping voltammetry, the intensity at 630 mV has been recorded for five
different samples. The results obtained for five replicates of each sample are shown in Table 3.
The calibration line fitted by LS is
y ¼ – 0:497 þ 0:265x
with the standard deviation of slope and the independent term equal to 0.004 and 0.186 nA, respectively.
The residual standard deviation is syx ¼ 0.308, the correlation coefficient is 0.997, and the coefficient of variation
is R2 ¼ 0.994.

Table 3 Data for the calibration of cadmium

Concentration (nmol l–1) Current (nA) Mean Variance

20.18 5.00 4.85 5.13 5.02 4.50 4.90 0.060

30.27 7.92 7.05 7.48 7.81 7.49 7.55 0.116
40.36 9.89 9.77 10.00 10.72 10.34 10.14 0.149
50.24 12.49 12.67 12.62 12.83 12.51 12.62 0.019
60.08 15.27 15.49 15.90 15.26 15.98 15.58 0.117
 Quality of Analytical Measurements: Univariate Regression 139

Table 4 ANOVA for the determination of cadmium including the lack of fit test

Source of variation Sum of squares Degrees of freedom Mean square Fcalc p-value

Regression 349.30 1 349.30 3686.37 <104

Error or residual 2.18 23 0.09
Lack of fit 0.34 3 0.11 1.22 0.33
Pure error 1.84 20 0.09
Total 351.481 24

The ANOVA table, as defined in Table 2, is presented in Table 4. The p-value for the test on significance of
the regression, Equation (29), is less than 104; thus, the linear calibration model is significant at a significance
level of 0.05.
Furthermore, Table 4 shows that there is no significant LOF; the p-value of the corresponding test is 0.33
and hence is greater than the usual significance level of 0.05. Therefore, there is no experimental evidence to
reject the null hypothesis ‘the model is biased’.
We can consider that the residuals follow a normal distribution (again at the significance level 0.05)
according to the tests of 2, Kolmogoroff, Shapiro–Wilks, skewness, and kurtosis with p-values of 0.89, 0.85,
0.96, 0.57, and 0.28, respectively.
Autocorrelation at the significance level of 0.05 does not exist either, because the p-value of the Durbin–
Watson test is 0.07.
Regarding the equality of variances of the residuals by concentration levels, the Cochran and Bartlett tests
have p-values of 0.79 and 0.42, respectively. Hence, none of the five variances is significantly greater or smaller
than the others, nor there are groups of them that are significantly different.
Figure 4 shows the externally studentized residuals of the fitted calibration line. No anomalous tendency is
observed.
The normality graph of the same externally studentized residuals is shown in Figure 5 (the straight line is
just a guide), which shows that the residuals are reasonably aligned corresponding to a normal distribution.

2.9
Externally studentized

1.9
residuals

0.9

–0.1

–1.1

–2.1
20 30 40 50 60 70
Cd (nmol l –1)
Figure 4 Graph of the externally studentized residuals vs. the true concentration of cadmium.

99.9
99
95
Percentage

80
50
20
5
1
0.1
–1.7 –0.7 0.3 1.3 2.3
Externally studentized residuals
Figure 5 Normal probability plot of the externally studentized residuals.
140 Quality of Analytical Measurements: Univariate Regression

1.05.4 Confidence Intervals and Hypothesis Testing

Once the calibration model yi ¼ 0 þ 1 xi þ "i ; i ¼ 1; . . .; N , with NID(0,2) random errors "i is validated, it is
useful to know how precise the estimated b0 and b1 are, as well as any other parameters obtained from them, as
important as the concentration of a problem sample.

1.05.4.1 Confidence Intervals for the Intercept, the Slope, and the Residual Variance
Section 1.05.2.2 already explains that the estimates b0 and b1 follow a bivariate normal distribution with the
variance–covariance matrix defined in Equations (13)–(15). syx 2
is an unbiased estimate of 2 independent of b0
and b1. In these conditions, the confidence intervals at significance level  for the slope and the intercept are
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  1
dðb1 Þ ¼ syx
b1 – t=2;N – 2 sb1 ; b1 þ t=2;N – 2 sb1 with sb1 ¼ Var PN ð40Þ
ð
i¼1 xi – x Þ2

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  1 x2
dðb0 Þ ¼ syx
b0 – t=2;N – 2 sb0 ; b0 þ t=2;N – 2 sb0 with sb0 ¼ Var þ PN ð41Þ
N i¼1 ðxi – x Þ2

Moreover, we can test the null hypothesis i ¼ i0 , where i is equal to 0 or 1 and i0 is a nonrandom value
previously fixed. For example, for the intercept when each measurement was blank-corrected, 0 ¼ 0. The
hypothesis test for each one is
H0 : i ¼ i0
ð42Þ
Ha : i 6¼ i0
The statistic for this hypothesis test is
bi – i0
tcalc ¼ ð43Þ
sb i
which follows a Student’s t-distribution with N2 d.f. if H0 is true. Then the null hypothesis is rejected at
significance level  if jtcalc j > t=2;N – 2 .

Example 2
With the calibration data in Example 1, we want to decide at significance level  ¼ 0.05 whether the
independent term is equal to zero. Thus, the hypothesis test is now H0 : 0 ¼ 0 versus Ha : 0 6¼ 0 and the
statistic, Equation (43), is tcalc ¼ ð – 0:497 – 0Þ=0:186 ¼ – 2:672.
As jtcalc j > t0:025;23 ¼ 2:069, the null hypothesis is rejected and we conclude that the independent term is
significantly non-null.
Two tests have been shown for testing H0 :1 ¼ 0 versus Ha :1 6¼ 0, an F-test (Equations (28) and (29)) and
the t-test (Equations (42) and (43)). The two tests are equivalent because a random variable that follows an
F-distribution with one and d.f. is equal to the square of a t-distributed variable with d.f., and it is simple to
2
prove that the statistic Fcalc in Equation (28) is equal to tcalc in Equation (43).
2
The confidence interval, at significance level , for the residual variance syx :
!
2 2
syx ðN – 2Þ syx ðN – 2Þ
; ð44Þ
2=2;N – 2 21 – =2;N – 2
2
The null hypothesis that the residual variance is equal to a known value H0 :syx ¼ 20 against Ha :syx
2
> 20 is
rejected at level  if
2
syx ðN – 2Þ
2calc ¼ ð45Þ
20
is greater than the critical value of a 2 distribution with N2 d.f.
 Quality of Analytical Measurements: Univariate Regression 141

In Draper and Smith11 and Asuero et al.,9 the procedure to compute a confidence interval for the correlation
coefficient ryx can be found. Also, to test the null hypothesis H0 : ryx ¼ 0, where 0 is a specified value
(e.g., zero), versus any of the alternative hypotheses Ha : ryx 6¼ 0, or ryx > 0, or ryx < 0.
The relation between the correlation coefficient and the slope of Equation (22) makes the following relation
holds:

ðb1 =sb1 Þ2
R2 ¼ ð46Þ
N – 2 þ ðb1 =sb1 Þ2
Therefore, the significance of the slope of a calibration line, the significance of the regression (ANOVA), and
the correlation and determination coefficients are related to each other in such a way that any of them can be
obtained from the rest.  
As a measure of the degree of linearity of a calibration, the use of the index ILC ¼ 1 – sb1 =b1 has been
proposed.14,15 It is a hyperbolic transformation of the statistic of the t-test (for the significance of the slope) and
the F-test (for the significance of regression) instead of evaluating the corresponding p-value. In fact, Equation
(46) states that ILC is a transformation of the correlation coefficient that takes into account the size of the
calibration set, N. Table 5 shows the values of the statistic b1 =sb1 between 1 and 100 together with the
corresponding ILC, R2, ryx, and p-values supposing that the calibration size is N ¼ 10 or N ¼ 20. For example,
stating that a calibration is good when ILC ¼ 0.95 is exactly the same as stating that the calibration has
ryx ¼ 0.99015 (N ¼ 10) or ryx ¼ 0.97823 (N ¼ 20), which is equivalent, in terms of p-values, to 6.0  1010 and
2.7  1015, respectively. It is evident that to fix a threshold value for any of them is equivalent to fixing it in
any other; the reader can verify that the practical meaning of all of them is the same.

1.05.4.2 Joint Confidence Region for 0 and 1

Equations (15) and (20) show that the estimates of b0 and b1 are correlated. Therefore, the joint confidence
region at level 100(1–)% is not the Cartesian product of the corresponding individual confidence intervals,
but is defined as the set of pairs (0, 1) that verify the following inequality:
PN PN
N ð0 – b0 Þ2 þ2 i¼1 xi ð0 – b0 Þð1 – b1 Þ þ 2
i¼1 xi ð1 – b1 Þ2
2
 F;2;N – 2 ð47Þ
2syx

This region is useful to establish the accuracy of a method of analysis in a range of concentrations and to
compare different analytical methods and/or calibrations. After the calibration model is obtained, it is used to
estimate the concentrations x̂i corresponding to the calibration standards xi. Independent of the form of
calibration (univariate, linear or nonlinear, multivariate, multiway, etc.), if the method is accurate, then the
regression of values x̂i versus xi has to be a straight line with slope 1 and independent term zero. The regression
2
x̂ ¼ b 0 þ b 1 x and its corresponding syx are enough to calculate, Equation (47), the confidence region at
significance level  of ðb0 ; b1 Þ. If point (0, 1) belongs to that region, it can be concluded that the method is
accurate.

Table 5 Some values of x¼ b1 =sb1 and the corresponding index of linearity, ILC ¼ 1 – ð1=x Þ , coefficient of determination, R2,
coefficient of correlation, ryx, and p-level for the t-distribution with the corresponding degrees of freedom (d.f.)

R2 ryx p-value

x 1(1/x) N ¼ 10 N ¼ 20 N ¼ 10 N ¼ 20 t (8 d.f.) t (18 d.f.)

1 0 0.11111 0.05263 0.33333 0.22942 0.08522 0.08231

5 0.8 0.75758 0.58140 0.87039 0.76249 0.00013 0.00002
10 0.9 0.92593 0.84746 0.96225 0.92057 3.9  107 8.0  1010
20 0.95 0.98039 0.95694 0.99015 0.97823 6.0  1010 2.7  1015
50 0.98 0.99681 0.99285 0.99840 0.99642 6.0  1014 <1017
100 0.99 0.99920 0.99820 0.99960 0.99910 5.5  1017 <1017
142 Quality of Analytical Measurements: Univariate Regression

The joint representation of these regions for several calibration methods fitted with the same analytical
signals allows a comparison between them; the smaller the surface of the confidence region, the more precise
the calibration method.

Example 3
In the simultaneous determination16 of steroid hormones estrone (E1) and 17--ethinylestradiol (EE2) by gas
chromatography with mass spectrometry detection (GC/MS), three calibration methods have been used with
the same data that correspond to the signals recorded on three nonconsecutive days:
1. A three-way Parallel Factor Analysis (PARAFAC)-based calibration.
2. A univariate linear regression with standardized sum of six monitorized ions.
3. A univariate linear regression with standardized peak base area.

For the EE2, Figure 6 shows the three joint confidence regions for the slope and the intercept of the three
regression models x̂i versus xi . One can see that the three calibrates are unbiased and that the PARAFAC-based
calibration model is the most precise.

1.05.4.3 Comparison of Two Regression Lines

Suppose that we have two calibration sets of size N1 and N2, with independent residuals and equal variances,
21 ¼ 22 . The equality of the estimated slopes of the two calibration lines, b11 and b12, is tested by the hypothesis
H0 : 11 ¼ 12 versus Ha : 11 6¼ 12
The statistic for this hypothesis test is
b11 – b12
tcalc ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 ð48Þ
1 1
sp2 PN1 þ PN2
ðx – x1 Þ2
i¼1 i1
ðx – x2 Þ2
i¼1 i2

where
2 2
syx1 ðN1 – 2Þ þ syx2 ðN2 – 2Þ
sp2 ¼ ð49Þ
N1 þ N2 – 4
is the pooled estimated common variance.

(c)

(b)
(a)
Slope

–6 –4 –2 0 2 4 6
Intercept
Figure 6 Joint confidence region at 95% for the slope and intercept of the regressions ‘estimated concentration vs. true
concentration’ for 3,17-di-TMS-EE2 with several calibration models: (a) PARAFAC; (b) AREAS; and (c) univariate.
 Quality of Analytical Measurements: Univariate Regression 143

The null hypothesis is rejected at significance level  if jtcalc j > t=2;N1 þN2 – 4 .
For the case of significantly unequal residual variances at level , that is, if
2
syx1
2 > F;N1 – 2;N2 – 2 ð50Þ
syx2

the comparison is carried out by the statistic

b11 – b12
tcalc ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2
ffi ð51Þ
syx1 syx2
PN1 2
þ P N2 2
i¼1
ðxi1 – x1 Þ i¼1
ðxi2 – x2 Þ

The distribution of tcalc in Equation (51) can be approximated by a t-distribution with d.f., where is the
nearest integer to
1
ð52Þ
c 2 =ðN1 – 2Þ þ ð1 – c Þ2 =ðN2 – 2Þ
with
PN1
2
syx1 = ðxi1 – x1 Þ2
c¼ P i¼1  PN2 ; if N1  N2
2 = N1 2 2 = 2
syx1 i¼1 ð x i1 – 
x 1 Þ þ s yx2 i¼1 ð x i2 – 
x 2 Þ

This value always lies between N12 and N1 þ N2 – 4.

The adequate formulas for testing the equality of more than two regression lines can be found in Sachs.17
Another approximation to the same problem consists of considering a multivariate regression involving a
dummy variable,11 z, which is zero for the values (xi ,yi) of the first calibration set C1 and one for the data of the
second calibration set C2.
The joint model is then
y ¼ 0 þ 1 x þ 0 z þ 1 zx þ " ð53Þ
The separate functional models for C1 and C2 lines are given by setting z ¼ 0 and z ¼ 1, respectively
C1 calibration model: 0 þ 1 x ð54Þ
C2 calibration model: ð0 þ 0 Þ þ ð1 þ 1 Þx ¼ 0 þ 1 x ð55Þ

1. To test whether the two calibration lines are parallel (they have the same slope), we would fit model in
Equation (53) and then test H0 : 1 ¼ 0 versus Ha : 1 6¼ 0.
2. To test whether the intercept is equal for both calibration lines, the null hypothesis used would be H0 : 0 ¼ 0
versus Ha : 0 6¼ 0.
3. To decide whether the two lines are equal, the null hypothesis is H0 : 0 ¼ 1 ¼ 0.
In cases 1 and 2, the test is a t-test similar to those of Equations (42) and (43), that is, tcalc ¼ i =si but now si is
the square root of the appropriate diagonal term of the (XtX)1s2 matrix, where s2 is the residual variance of the
fitting of the model in Equation (53) and
0 1
1 x1 0 0
B C
B 1 x2 0 C0
B C
B C
B  C
B C
B C
B 1 xN1 0 0 C
X¼B
B1
C ð56Þ
B xN1 þ1 1 xN1 þ1 CC
B C
B. .. .. C
B .. . . C
B C
B C
@1 xN1 þN2 1 xN1 þN2 A
144 Quality of Analytical Measurements: Univariate Regression

The jtcalc j is compared with the percentage point t=2;N1 þN2 – 4 for a two-sided test on the null hypothesis
H0 : i ¼ 0 versus Ha : i 6¼ 0.
The joint test of case 3 is similar to that in Equation (47). The details of this test can be consulted in
Chapter 3.02 and also in Chapter 14 of Draper and Smith.11
In the model of Equation (53), it is assumed that the variance of the residuals is equal in all experimental data,
and this should be carefully verified before applying the hypothesis test to decide about the calibrations C1 and C2.
In chemical analysis, the decision on the equality of two calibration lines is important to decide if there is a
matrix effect and, if so, what type. If this is the case, calibration C1 is made with samples spiked on a matrix (e.g.,
sulfonamide in pig kidney) and calibration C2 would be formed by samples of similar concentrations in the
dissolvent. If there is no matrix effect, the null hypothesis of case 3 will not be rejected. It is also possible that the
effect is a constant or proportional bias, in these cases the null hypothesis of case 2 or the null hypothesis of case
1 will be rejected respectively. When 0 and 1 in Equation (55) are significantly non-null, then both biases
exist.
Another important application is the verification that the calibration stays day to day; in this case, C1 and C2
would be the calibration made on the first and second day, respectively.

1.05.4.4 Confidence Interval for the Prediction

For a concentration x0 where no experiments were carried out, the response is estimated by
ŷ0 ¼ b 0 þ b 1 x 0 ð57Þ
However, as b0 and b1 are random variables that follow a bivariate normal distribution, ŷ0 is also a random
variable that follows a normal distribution whose mean and variance as computed with Equations (11)–(15) are
E ðŷ 0 Þ ¼ 0 þ 1 x0 ð58Þ
!
1 ðx0 – xÞ2 2
Varðŷ 0 Þ ¼ þ PN 2  ð59Þ
N i¼1 ð xi – 
x Þ

Varðŷ 0 Þ has a minimum when x0 ¼ x and increases as x0 separates from x in either direction. In other words, the
most precise prediction is expected to be in the mean of the calibration range and precision is lost as we separate
from x.
The confidence interval at level (1)100% for the true mean value of the response for a given x0 is then
computed as
 
ŷ 0 – t =2;N – 2 s y0 ; ŷ0 þ t =2;N – 2 s y0
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 ðx0 – xÞ2 ð60Þ
d
sy0 ¼ Varðŷ 0 Þ ¼ syx þ PN
N i¼1 ðxi – x Þ2

If we join up all the lower endpoints and all the upper endpoints of intervals defined in Equation (60) as x0
changes, we would obtain the two dotted hyperbolas as shown in Figure 7 for the data in Example 1, Table 3.
The individual values of the random variable ŷ0 are distributed around its mean with variance 2 indepen-
dent of Varðŷ 0 Þ. Therefore, the variance of the prediction of an individual observation would be
!
2 1 ðx0 – xÞ2
 1þ þ PN ð61Þ
N Þ2
i¼1 ðxi – x

and the corresponding estimated value is obtained by substituting 2 by syx

2
in Equation (61).
The confidence interval for a new observation is then
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 ðx0 – xÞ2
ŷ0 t =2;N –2 s yx 1 þ þ PN ð62Þ
N i¼1 ðxi – x Þ2
 Quality of Analytical Measurements: Univariate Regression 145

16

14

12

Current (nA)
10

4
20 30 40 50 60 70
Cadmium (nmol l –1)
Figure 7 Confidence intervals at 95% for the calibration data of Example 1, Table 3. Dotted hyperbolas are for the true
mean. Continuous hyperbolas are for a new prediction with q ¼ 1 in Equation (63). The squares are the experimental points.

A confidence interval for the average of q new observations is obtained similarly as follows:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 1 ðx0 – xÞ2
ŷ0 t =2;N – 2 s yx þ þ PN ð63Þ
q N i¼1 ðxi – x Þ2

Again, by joining the corresponding endpoints of the confidence intervals for one new observation, we obtain
two hyperbolas (those drawn with a continuous line in Figure 7). The rest of the hyperbolas that would be
obtained for q > 1 would be located between the two ‘limiting’ hyperbolas (the dotted and the continuous lines
in Figure 7).

1.05.4.5 Confidence Interval for the Determination of the Concentration of a Problem

Sample
The most important application of a calibration line is the determination of the concentration of an analyte
in a problem sample. We have fitted the calibration line (y ¼ b0 þ b1 x with residual standard deviation syx)
and now for a specified value of y, say y0, we wish to obtain an estimated value x̂0, the value of sample
concentration, as well as some sort of confidence interval for it. Initially, we suppose that y0 is the true
mean value of the distribution of the signal for a sample of unknown concentration, then the solution will
be adapted to the case of having one or several determinations on the same sample whose concentration
we want to obtain.
By using the calibration line, the estimated concentration is
y0 – b0
x̂0 ¼ ð64Þ
b1
A confidence interval for x̂0 at level (1)100% will be made up by the values of the concentration x such that
the confidence interval at level (1)100% for the corresponding value y ¼ b0 þ b1 x contains y0 . According to
Equation (60), this is equivalent to considering all the x values that verify
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 ðx – xÞ2
jy0 – ðb0 þ b1 x Þj  t=2;N – 2 syx þ PN ð65Þ
N i¼1 ðxi – x Þ2

Besides being consequence of the distribution of the response, this interval is also a likelihood ratio confidence
region as shown in Brown18 and also corresponds to the inversion of a similar test of the hypothesis x ¼ x0 (see
Cox and Hinkley19). It is closely comparable to a Bayesian posterior distribution for x corresponding to vague
prior knowledge.
146 Quality of Analytical Measurements: Univariate Regression

After squaring and rationalizing Equation (65), we obtain a second degree equation. Its solutions are:11,18.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 1 syx 1 ðy0 – yÞ2
x þ ðx̂ 0 – xÞ t =2;N – 2 ð1 – g Þ þ 2 PN ð66Þ
1–g 1 – g b1 N b1 i¼1 ðxi – xÞ2

with


t2 –2 t=2;N – 2 2
g¼  =2;NP ¼ ð67Þ
b12 = syx
2 = ðxi – xÞ2 b1 =sb1

Note that the denominator of Equation (67) is the square of the statistic of the t-test to test H0 : 1 ¼ 0 versus
Ha : 1 6¼ 0 (Equation (43)), which is equal to the F-test for significance of regression (Equation (28)). Thus, the
more significant b1 is (e.g., smaller p-value in Table 1), the larger the denominator of g and the smaller g and
ð1 – g Þ > 1 will be. This is a quite usual situation; so, instead of using Equation (66), the following approximation
for the confidence interval20,21 is used: sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
syx 1 ðy0 – yÞ2
x̂0 t =2;N – 2 þ 2 PN ð68Þ
b1 N b1 i¼1 ðxi – xÞ2

The above calculations are true mean value calculations. Nonetheless, in calibration, y0 is regarded not as a true
mean value but as the mean of q observations. Thus, Equation (63) will be used for Varðy0 Þ and the confidence
interval at level (1)100% is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
syx 1 1 ðy0 – yÞ2
x̂0 t =2;N – 2 þ þ 2 PN ð69Þ
b1 q N b1 i¼1 ðxi – xÞ2

Then q ¼ 1 provides the confidence interval for the concentration when we have an individual observation, y0 ,
of the signal.
Equations (13), (14), and (61) prove that the variance of the slope, intercept, and ŷ0 are the multiplication of
2 (estimated by syx
2
) by a factor that depends on the following: (1) the proximity of x0 to x, (2) the number N of
calibration samples, and (3) the value xi of the calibration standards. As regards the estimated variance of x̂0 ,
from Equation (69), is obtains the estimated variance of x̂0 as
!
2 1
dðx̂ 0 Þ ¼ syx 1 1 ðy0 – yÞ2
Var þ þ 2 PN
b12 q N b1 i¼1 ðxi – xÞ2
! ð70Þ
2 1 1 1 ðx̂ 0 – xÞ2
¼ syx þ þ PN
b12 q N i¼1 ðxi – x Þ2

that depends on the calibration design (N and xi values), on the proximity between y0 and y , and also on the
estimated slope.
To obtain the most precise estimation of the concentration, the problem sample should provide a signal
(or mean value of q signals), y0, close to the central value of those of the calibration concentrations in such
a way that
P the numerator of Equation (70) is small. Moreover, one obtains better precision if the sum of
squares Ni¼1 ðxi – xÞ2 in the denominator is large, which is achieved if the concentration of the standards
is far from the mean value. The remaining addends in Equation (70) also indicate that precision is better
if the number of calibration standards N is larger and/or if the number of replicate determinations in the
problem sample, q, is larger too. But we also obtain better precision in the estimated concentration if the
slope b1 is larger, that is, as the sensitivity of the calibration increases, because b1 intervenes (squared) in
the denominator of Equation (70). As in all confidence intervals, the value of N also influences through
the value of t/2,N2, which can be important. For example, for  ¼ 0.05, if N ¼ 5, t0.025,3 is 3.182, whereas
for N ¼ 10, t0.025,8 is 2.306.
 Quality of Analytical Measurements: Univariate Regression 147

Example 4
With the data of the calibration of cadmium, Example 1, Table 6 shows the variation of the confidence interval
at 95% for the concentration of a problem sample as a function of its position in the calibration range and the
number of replicates used.
To clearly see the effect, the values of the signals y0 supposedly obtained were 4.75, 10.20, and 15.50 nA,
which are values near to the minimum, the mean and the maximum respectively of the recorded calibration
signals. Furthermore, in each case, the signal corresponds to one, three, or five replicate(s), that is, q ¼ 1, q ¼ 3,
or q ¼ 5 in Equation (69). In our case, t=2;23 ¼ 2:0687.
From the results in Table 6, one can observe that for the signal of 10.20 nA to consider 3 or 5 replicates
causes a reduction in the radius of the confidence interval of 40 and 52%, respectively. This reduction is
much higher than that caused by the position of the sample in the calibration range, which represents a
reduction of 4% when only one replicate is used and 16% when five replicates are used. Note the
interaction between the number of replicates and the position of the calibration range. In general, it is
advisable to determine at least three replicates to make the precision increase significantly. In any case, it
is quite simple using Equation (70) to explore the combined effect of the position of the signal of the
sample and the number of replicates, similar to Table 6, for any calibration.
The expressions of the intervals and confidence region, Equations (40), (41), (44), (60), and (68), show
that all of them become ‘wider’ if syx is larger, that is, the estimates are less precise. Furthermore, in the
statistics of the t-test, Equations (43), (48), and (51), the standard deviation of the estimate, for example,
appears in the denominator and increases when increasing syx. Thus, tcalc decreases and it becomes more
difficult to reject the null hypothesis. In summary, an overestimation of the variance of the residuals, 2, has
a negative effect on all the inferences, including those related to the calculated concentration. At this point,
the reader must remember the analysis of the LOF in Section 1.05.3.1.2. Concretely, Equations (32) and (33)
show that the bias of the model necessarily causes an increase in syx. Therefore, much attention has to be
paid to the problem of assuring the adequacy of the model because in case the model is biased, all the
inferential aspects are altered.

1.05.4.6 The Intersection of Two Linear Regression Lines

In some analytical techniques, it is necessary to determine the abscissa of the point where two straight line
segments of an analytical signal intersect (e.g., conductometric and photometric titrations). In this case, we have
two series of measurements, {(xi1,yi1), i ¼ 1, . . ., N1} and {(xi2, yi2), i ¼ 1, . . ., N2}, before and after the break
2 2
point, and fit linear regression lines y ¼ b01 þ b11x and y ¼ b02 þ b12x with syx1 and syx2 as estimated residual
variance, respectively. The estimated abscissa in the intersection point is

b01 – b02 b0

x̂0 ¼ ¼ ð71Þ
b12 – b11 b1

Table 6 Confidence interval for the determination of the concentration of a

problem sample with the calibration in Example 1

q¼1 q¼3 q¼5

y0 (nA) x̂0 dðx̂ 0 Þ

Var h (nmol l1) Var
dðx̂ 0 Þ h (nmol l1) dðx̂ 0 Þ
Var h (nmol l1)

4.75 19.80 1.518 2.549 0.618 1.626 0.437 1.368

10.20 40.37 1.405 2.452 0.504 1.469 0.324 1.178
15.50 60.37 1.515 2.546 0.614 1.622 0.434 1.363

y0, signal of the problem sample; q, number of replicates used in the

determination; h, radius of confidence interval ( X̂0h, x̂0þh ) at 95%
(Equation (69))
148 Quality of Analytical Measurements: Univariate Regression

The statistical problem of the intersection of two experimental lines has been tackled in depth.21,22–25 A
(1)100% confidence interval on the intersection can be obtained as the roots of the following quadratic
equation in x̂0 :
     
x̂0 2 b12 – t 2 sb
2
1
– 2x̂0 b0 b1 – t 2 Covðb0 ;b1 Þ þ b02 – t 2 sb
2
0
¼0 ð72Þ

where t denotes the critical value t=2;N1 þN2 – 4 .

2 2
Supposing that the variance of the error, syx1 and syx2 , is the same in both segments, the pooled estimated
common variance of Equation (49) is used to obtain the variances and covariances of b0 and b1. They are
computed by the propagation of errors method, which is exact in these cases because b0 and b1 are linear
functions of random variables.
!
2 1 1
sb ¼ sp2 PN1 2 þ P N2 ð73Þ
1
i¼1 ð xi1 – 
x 1 Þ ð x 2 Þ2
i¼1 i2 – x
!
2 1 1 x1 2 x2 2
sb ¼ sp2 þ þ PN1 2 þ PN2 ð74Þ
0
N1 N2 i¼1 ðxi1 – x
1 Þ 2 Þ2
i¼1 ðxi2 – x
!
x1 2 x2 2
Covðb0 ; b1 Þ ¼ sp2 PN1 2 þ P N2 ð75Þ
i¼1 ðxi1 – x
1 Þ 2 Þ2
i¼1 ðxi2 – x

The groups of symbols enclosed in brackets in Equations (73) and (74) have the form of a two-tailed t-test for
significant differences of slopes and significant differences of intercepts, respectively. When the hypothesis of
equal slopes is not rejected using the statistic in Equation (48), the coefficient of x̂0 2 in Equation (72) becomes
negative and no real roots are obtained; hence, the confidence interval embraces the entire x axis at the chosen
level of confidence.

1.05.5 The Design of a Calibration

In this section, the effect of the distribution of the calibration standards on the quality of the calibration line will
be analyzed. In principle, it is necessary to answer a series of questions based on the prior knowledge of the
analyst on the behavior of the system of measurement in terms of the model supposed for the calibration.
In relation to the calibration set, which is the range of concentrations of interest? Is the total number, N, of
measurements to be made limited? Is it possible to replicate in all the levels of concentration? Should the values
of the concentration of the calibration standards be selected among some fixed values or can they take any value
from the specified range?
Other questions are relative to the calibration model such as is it probable that the relation signal
concentration is not linear in the specified range? To decide this, one’s previous knowledge or knowledge of
other experts should be used. Does one believe that the curvature, if it exists, can be modeled with a polynomial
model of low order (for example, of order 2 or 3)? Because, if this is not so, a study with a model of higher order
should be initiated.
The answer to these questions will be reflected in the number of d.f. needed to estimate the LOF and even in
the possibility of fitting a calibration model of higher order to the experimental data.
Finally, questions about the variance of the error should be posed: Is there any a priori information about it?
Should it be estimated together with the slope and the intercept? If the answer is yes, then replicates are
mandatory, which are much more important if, in addition, the possibility of failure of the equality of variances
in the chosen range of concentrations exists.
Once the above questions are answered, the quality of a calibration model is related to the variances of b0 and
b1 and the correlation between them; thus, Equations (13), (14), and (20) should be used. Equations (13) and (14)
show that the variance of both parameters is equal to a factor, called variance inflation factor (VIF), that
depends only on the values of xi and not on the recorded signals. The second factor, 2, is the variance of the
 Quality of Analytical Measurements: Univariate Regression 149

signal, which is either known or will be obtained from the experimental results. In both cases, we must accept
that it is a characteristic of the measurement system and cannot be modified during the calibration stage. It is
evident that by designing the distribution of the calibration standards, the value of VIF can be decreased and,
therefore, the variance of the estimators decreases.
The previous argument is the same as that under the methodology for response surfaces (see Chapter 1.12 of
the book dedicated to design of experiments) but there is a difference; basically, a calibration model is not used to
predict the instrumental signal, y0, but the concentration of a problem sample. The structure of Equation (59) of
the variance of the estimated response is also the product of 2 by a variance function. This function is quadratic
and its minimum is 1/N,P which is reached in the mean value of the concentrations. The form of this function
depends on the factor 1= Ni¼1 ðxi – xÞ2 , which is also the VIF of b1 (Equation (13)). Consequently, the quality of
the design of a linear calibration will be evaluated with the VIF and with the correlation between the slope and
the intercept.
Let us now continue our discussion in terms of a specific example. Let us suppose that the analyst has decided
to do N ¼ 17 determinations in concentrations around 80 ppb (a hypothetical permitted limit (PL) for a
residue); the calibration range considered is between 0.5 and 1.5 PL, that is, between 40 and 120 ppb (as usual
in official norms). Furthermore, she or he wants the sample with 80 ppb to be always among the calibration
samples. Also, the analyst thinks that the relationship is probably first order but is not absolutely sure. Finally, it
can be considered that the variance 2 does not change in the calibration range, although it is unknown and
should be estimated with the calibration data.
Figure 8 shows some of the possibilities. Each dot represents one experimental determination at the
concentration specified in ppb.
For the same experiments and to avoid excessively small numbers, instead of the VIF, Table 7 shows the
square root of the VIF, VIF1/2, for the independent term and the slope at each of the possibilities in Figure 8.
VIF1/2 is in fact the factor that when multiplied with the residual standard deviation gives the standard
deviation of the parameter.
The number of different standards, k, and the d.f. of pure error and those for the LOF have also been shown in
Table 7. As can be seen, all the designs include concentration 80 ppb and cover the proposed range.
Design (1) is not acceptable because it does not allow considering the pure error. Designs (8) and (9) are not
acceptable either because the estimation of the LOF is made with only 1 d.f. (in this case, if the LOF was
significant, the calibration line of order 2 will pass obligatorily through the mean value of the signals obtained in
the three concentrations 40, 80, and 120 ppb). Nevertheless, the factors of design (9) can be taken as the
reference of the minimal VIF1/2, whereas those of design (1) as the maximal. When comparing these two
designs, it is clear that factor VIF1/2 for the slope can be improved by about 36%, and thus also its standard
deviation by the same percentage.
The remaining designs would fulfill the conditions imposed, allowing, in case of curvature, the fitting of a
polynomial of degree two or three.
Designs (4) and (5) are especially adequate because they have 5 d.f. to estimate the LOF, and with design (5) a
higher reduction (23%) of the standard deviation of b1 is obtained. It is interesting to note that when decreasing
the number k of different standards, the value of VIF1/2 is not necessarily reduced – compare designs (3) and (4),
1/2
(5) and (6), or (7) and (8). The values of VIF P 2 for the independent term follow a similar tendency, although
they are much higher because of the term xi that appears in the numerator. The strong linear dependency
between the slope and the independent term must be observed, which stays between 0.96 and 0.90.
To study the effect of the design on the variance of the computed concentration, Equation (70) will be used,
2
which is the product of syx =b12 and VIFðx̂ 0 Þ where
!
1 1 ðx̂ 0 – xÞ2
VIFðx̂ 0 Þ ¼ þ þ PN ð76Þ
q N i¼1 ðxi – x Þ2
2
The value syx =b12 will be obtained after fitting the calibration line; both the residual variance and the slope are
specific to the system and do not depend on the design. However, VIFðx̂0 Þ of Equation (76) depends on the design
and on the estimated value, x̂0 , of the concentration of the problem sample. The minimum is reached when x̂0 ¼ x
(and is ð1=q Þ þ ð1=N Þ), and the maximum is in one of the extremes of the calibration range. Table 7 also contains
150 Quality of Analytical Measurements: Univariate Regression

(1)
40 50 60 70 80 90 100 110 120 ppb

(2)

40 50 60 70 80 90 100 110 120 ppb

(3)

40 50 60 70 80 90 100 110 120 ppb

(4)

40 50 60 70 80 90 100 110 120 ppb

(5)

40 50 60 70 80 90 100 110 120 ppb

(6)

40 50 60 70 80 90 100 110 120 ppb

(7)

40 50 60 70 80 90 100 110 120 ppb

(8)

40 50 60 70 80 90 100 110 120 ppb

(9)

40 50 60 70 80 90 100 110 120 ppb

Figure 8 Nine different possible designs for the calibration. Each dot represents an experiment at the concentration (ppb) in
the horizontal axis.

the maximum values of VIFðx̂0 Þ1=2 , which are the maximum inflaction factors of the standard deviation of x̂0 for
the nine designs analyzed. The minimum value when q ¼ 1, that is, when only one signal is recorded for the
problem sample, is 1.0290 and the maximum varies from 1.1026 to 1.0589; nevertheless, when q ¼ 2 (two signals are
recorded for the problem sample and the mean is taken), the minimum is 0.7475 and the maximum varies,
depending on the design, from 0.8460 to 0.7882. In the case of q ¼ 5, the minimum is 0.5087 and the maximum
varies from 0.6447 to 0.5669. It is clear that the effect due to the number of replicates in the problem sample is much
greater than that due to the design. From Equation (76) it is clear that for q ¼ 1 is VIFðx̂ 0 Þ1=2 > 1, therefore it is
advisable to make at least two replicates of the problem sample. In the example, this action diminishes the factor at
least by 15%. Design (5) with q ¼ 2 is a suitable choice to obtain acceptable inflaction factors.
 Quality of Analytical Measurements: Univariate Regression 151

Table 7 Characteristics of the different designs in Figure 8

VIF(x)1/2

Design VIF(b0)1/2 VIF(b1)1/2 Corr(b0,b1) q¼1 q¼2 q¼5 k d.f. for PE d.f. for LOF

1 0.8284 0.0099 0.9562 1.1026 0.8460 0.6447 17 0 15

2 0.7695 0.0091 0.9490 1.0919 0.8320 0.6262 9 8 7
3 0.6849 0.0080 0.9352 1.0777 0.8133 0.6012 9 8 7
4 0.7375 0.0087 0.9444 1.0863 0.8247 0.6165 7 10 5
5 0.6514 0.0076 0.9281 1.0725 0.8063 0.5918 7 10 5
6 0.6927 0.0081 0.9367 1.0789 0.8149 0.6034 5 12 3
7 0.6054 0.0069 0.9162 1.0657 0.7973 0.5747 5 12 3
8 0.6262 0.0072 0.9220 1.0687 0.8013 0.5849 3 14 1
9 0.5557 0.0063 0.8997 1.0589 0.7882 0.5669 3 14 1

VIF(.) represents the maximum of the variance inflation factor of x̂0, VIF(x̂0); PE, pure error; LOF, lack of fit; q, the number of
replicates; k, the number of calibration standards; d.f., the degrees of freedom.

Introducing the individual cost of each determination and by varying N and q, the cost can be optimized in
addition to the quality of the design. In Section 2.2.8 of Meier and Zünd,21 it is shown from Equation (70) that
using the wrong combination of N and q enormously increases costs.

1.05.6 The Capability of Detection, the Decision Limit, and the Capability of
Discrimination Computed from a Regression Model

One of the main reasons for the development of new instrumental techniques is to increase their capability to
determine much smaller quantities. It is therefore necessary to establish a rule to ensure whether the analyte is
present or not in the sample. In general, a detection rule should be as follows: affirm that the analyte is present in
a sample if the response obtained is above a threshold value. When applying this decision rule, two mistakes
might happen: (1) affirming that the analyte is present in the sample when it is in fact not present, in other
words, giving a ‘false positive’, and (2) affirming that the analyte is not present in the sample when it is in fact
present, or giving a ‘false negative’.
The classic definition for the detection limit26 only evaluates the probability of a false positive but does not
explicitly quantify the probability of a false negative. A good revision of the evolution of the stance of the
IUPAC can be found in the paper by Currie.27 Chapters 5 and 6 of Gibbons’s book28 contain an interesting
discussion of the methods proposed for the calculation of the detection limit and include a good selection of
bibliography on the subject.
The capability of detection for a given probability of false positive, , as defined by the ISO,3 is ‘‘the true net
concentration of the analyte in the material to be analysed which will lead, with probability 1, to the correct
conclusion that the concentration in the analysed material is larger than that in the blank material’’. The need to
evaluate the probability of false positive, , and of false negative, , has also been recognized by the IUPAC,
and in the EU it is mandatory for the identification and quantification of toxic residues or residues that come
from veterinary treatments in products for human consumption (European Decision 2002/657/EC).
The concept of capability of detection is applicable to any level of concentration x0 and not only for the case
of concentration zero.
The following one-tail test30 is performed to estimate the capability of detection: H0 : x ¼ x0 (the concentra-
tion of the analyte in the sample is x0) versus Ha : x > x0 (the concentration of the analyte in the sample is greater
than x0). Thus, the capability of detection, named CC in the above-mentioned European decision and xd in the
ISO 11843-1 standard, can be estimated in the univariate case31–33 through a linear calibration model signal
versus concentration using the following expression:
ð;  Þwx0 syx
xd ¼ ð77Þ
b1
152 Quality of Analytical Measurements: Univariate Regression

where (, ) is the value of noncentrality parameter of a noncentral t-distribution related to the probabilities
 and , and wx20 is the VIF(x0) already defined in Equation (76).
When the analyte is banned, the minimum detectable net concentration is estimated with x0 ¼ 0 in Equations
(76) and (77) and  and  are the probabilities of false positive and false negative, respectively. However, if the
analyte is not banned and a PL has been fixed, the capability of detection must also be evaluated at those levels
of concentration. In this case, the capability of detection is estimated31,32,34 as in Equations (76) and (77) with
x0 ¼ PL and  and  are the probabilities of false noncompliance and false compliance, respectively.
Regarding the decision limit (Equation (78)), the ISO 11843-2 standard has defined the critical value (at zero)
of the net concentration as ‘‘the value of the net concentration the exceeding of which leads, for a given error
probability , to the decision that the concentration of the analyte in the analysed material is larger than that in
the blank material’’.
t;N – 2 wx0 syx
xc ¼ ð78Þ
b1

where wx0 is the same as in Equation (77).

For substances with a PL, the decision limit also has to be estimated35,29,36 as in Equation (78) but substituting
x̂0 by PL in Equation (76) to compute wx0.
As the decision of accepting or rejecting the presence of analyte in the sample is established based on the
signal, for an unknown sample the null hypothesis has to be rejected if the recorded signal is above a threshold
yd (signal of detection); hence, the critical region of the test is the set of signals beyond the signal of detection,
and a calibration function that relates signal with concentration is needed.
The program in Sarabia and Ortiz33 implements the previously described procedure when x0 ¼ 0 and
calculates the characteristic curves associated with the hypothesis test. In this way, the statistical method
establishes the capability of detection of a specific analytical procedure, given by the characteristic curves of the
decision rule expressed as a hypothesis test in terms of the concentration taking into account the calibration
line. This procedure (despite being published earlier than the year 2000) is the one proposed by ISO 11843 and
Commission Decision 2002/657/EC, estimating the minimum detectable net concentration xd, taking into
account the probabilities  and .

Example 5
In the determination of ascorbic acid in soft drinks by High Performance Liquid Chromatography – Diode
Array Detection (HPLC-DAD) (see Garcı́a et al.37 for more details), we want to compute the capability of
detection for the experimental data in Table 8.
The LS regression line computes an intercept, b0, equal to 221.5 with sb0 ¼ 41.6 and a slope, b1, 308124.0 with
sb1 ¼ 2688.2. The residual standard deviation of the regression is 89.6 a.u. and the correlation coefficient is
0.9994. Furthermore, it is checked that the regression model is significant (p-value < 104) and that there is no
significant LOF (p-value 0.24). There is neither failure in the equality of variances (the p-values for the Cochran
and Bartlett tests are 0.63 and 0.14, respectively) nor failure of the normality of residuals.
Using DETARCHI,33 Equations (76) and (77) with x̂0 ¼ 0, with one replicate (q ¼ 1) it is possible to detect
0.0011 mg l1 with a probability of false positive, , equal to 0.05 and a probability of false negative equal to 0.05.
As the computed capability of detection is less than the lowest concentration of the calibration standards, this
value cannot be considered and the only fact we can assert is that the proposed procedure has a capability of

Table 8 Calibration of ascorbic acid to determine the capability

of detection

Concentration (mg l1)

0.004 0.006 0.008 0.012 0.016 0.020 0.026

1454 2063 2597 3863 5259 6463 8280

Peak area (a.u.) 1420 2198 2586 4042 5241 6279 8117
1437 2078 6255 8305
 Quality of Analytical Measurements: Univariate Regression 153

detection equal to 0.004 mg l1 (the lowest concentration whose signal was recorded) with a probability of false
positive equal to 0.05 and a probability of false negative less than 0.05. Garcı́a et al.37 contains a discussion about
the capability of discrimination at several non-null nominal concentrations, x̂0 6¼ 0.
The decision limit and the capability of detection defined in Equations (77) and (78) have the property of
being invariant for linear transformations of the response36 (the signal). As a consequence, one obtains the same
capability of detection and decision limit using the regression estimated concentration versus true concentra-
tion. In practice, this is useful because even when using nonlinear univariate calibration methods, or calibrations
based on multiway and multivariate regression methods or based on neural networks, etc., it is always true that
the relation between the estimated concentration, x̂i , and that of the calibration standards, xi , must be linear.
This regression enables the evaluation of the accuracy of different calibration methods (see Section 1.05.4.2) and
also the capability of detection and the decision limit of any of them.
The capability of discrimination is another piece of information of interest provided by the calibration line. A
method is said to be sensitive (IUPAC,29 Chapter 10) if a small change in concentration, x, causes a large change
in the measurement, y ; that is, when the derivative dy=dx is large. In the case of a linear calibration line,
dy=dx ¼ b1 . As the slope is a random variable, it is not enough to consider only its mean but also its probability
distribution or at least its variance. The capability of discrimination, that is, the smallest difference jx – x0 j such
that the signal of x can be distinguished from that of x0, will depend on the distribution of both and also on the
assumed  and  (risk of, respectively, concluding that there is a difference when there is none and concluding
that there is no difference when in fact there is difference). Formally, it is about testing H0 :jx – x0 j ¼ 0 versus
Ha :jx – x0 j > 0. The following equation has been proposed for the sensitivity (capability of discrimination):36
ð=2;  Þwx0 syx
jx – x 0 j ¼ ð79Þ
b1
As for the capability of detection, the capability of discrimination can be computed using Equation (79) for the
regression model x̂ ¼ b 0 þ b 1 x and, therefore, it is of use with any kind of calibration method (multivariate or
multiway). In the same paper,36 several applications with partial least squares (PLS) calibration models are shown.

1.05.7 Standard Addition Method

A calibration line cannot be used to determine an analyte in a sample when it is known that the matrix of the
sample interferes with the determination and no blank is available for spiking at levels around the expected
(‘nominal’) level. A possible solution to this problem is to apply the method of standard addition in which the
sample is used as a kind of ‘blank’ to which further analyte is spiked. To do this, several quantities of the analyte
to be determined are added to aliquots of the unknown sample. These spiked samples as well as the unknown
sample are measured. Figure 9(a) is the usual representation of the data signal versus concentration added, that

(a) (b)
Difference y–y0
Signal y

y0 y0

0 Amount added 0 Amount added

Estimated
Estimated
concentration
concentration
Figure 9 Procedure for the standard addition method: (a) extrapolation and (b) interpolation.
154 Quality of Analytical Measurements: Univariate Regression

is, the pairs (xi, yi) i ¼ 1, . . ., N. If conditions about linearity and NID(0,2) errors are fulfilled, the model fitted
by LS is y ¼ b0 þ b1 x and the quantity of analyte in the sample, xs, is obtained by extrapolating the line to the
abscissa (y ¼ 0, xs ¼ – ðb0 =b1 Þ) if there is no systematic error.
The estimated value for the concentration of the problem sample is thus x̂s ¼ b0 =b1 and its variance can be
obtained from Equation (70) for value x̂s :
!
2 1
dðx̂ s Þ ¼ syx 1 1 ð – ðb0 =b1 Þ – xÞ2
Var þ þ PN ð80Þ
b12 q N i¼1 ðxi – xÞ2

This variance can be reduced if the regression line is made with the data (xi, yiy0), i ¼ 1, . . ., N, as indicated in
Figure 9(b). It is clear that the slope of the new line of standard addition is the same as before. The
concentration of the problem sample is the one corresponding to ordinate y0 in the graph, which is also the
value b0 of the previous regression model, so its variance is
!
2 1
dðx̂ s Þ ¼ syx 1 1 ððb0 =b1 Þ – xÞ2
Var þ þ PN ð81Þ
b12 q N i¼1 ðxi – xÞ2

less than the previous one, Equation (80), due to the positive sign of b0/b1.
In a concrete problem, to guarantee the absence of bias in the determination and therefore that the standard
addition method is valid, the following should be demonstrated: (1) that proportional error does not exist,
verifying that the slope of the standard addition line is the same as the calibration line obtained for the pure
analyte in the absence of the matrix, and (2) the absence of absolute systematic error that will be verified by
means of reference spiked samples. The second point is the weakness of the method of standard addition
because on many occasions it is not possible to have reference samples without an analyte to evaluate the
recovery.
The H-point standard addition method38 is a modification of the standard addition method that permits
correcting for both absolute and proportional systematic errors. In the simplest case, an interferent whose
spectrum is known, the method is based on obtaining a standard addition line at two wavelengths,
1 and
2,
where the interferent shows the same absorbance. The two standard addition lines intersect at the so-called H-
point (xH, yH) and the estimated concentration is x̂s ¼ – x H . This method has been generalized39,40 to obtain the
total concentration or concentrations of different chemical forms of an analyte when the matrix of the sample is
completely unknown.
Another solution comes from the calibration methods based on the decomposition of three-way data that identify
the multivariate signal of the analyte and separate it from the signal(s) of the remaining constituents of the sample.
The procedure was initially developed41 for the ‘direct trilinear decomposition’ method known as the ‘second-order
standard addition method’ (SOSAM). It has been applied with other decomposition methods such as multivariate
curve resolution,42 or PARAFAC,43 or direct trilinear decomposition (TLD);44 even an adaptation of this method to
handle the case in which the standard addition method causes a reduction of the rank has been proposed.45

1.05.8 Weighted Least Squares and Generalized Least Squares

Sometimes the signals used in a calibration have different variance. This can be deduced from a priori
knowledge, from graphics y versus x, or from the analysis of the residuals of an LS fit. With a transformation
of the signal, y, and possibly of the concentration, x, it might happen that a linear relation between the new
variables is valid and a constant-variance additive error is obtained. When this is not possible, the adequate
estimation is obtained with a weighted least squares fitting. The presence of correlation in the errors leads to a
generalized least squares fitting.
To make the analysis easier, the matrix notation will be introduced for the model in Equation (3).
!
0
The set of N equations yi ¼ ð 1 xi Þ þ "i i ¼ 1, . . ., N, may be written in the matrix form as
1
y ¼ Xb þ e ð82Þ
 Quality of Analytical Measurements: Univariate Regression 155

where y is the column vector comprising the values of the response, " is the column vector of residuals, b is the
column vector with the slope and the independent term, and finally X is the matrix
0 1
1 x1
B C
B. . C
X ¼ B .. .. C ð83Þ
@ A
1 xN

and
EðeÞ ¼ 0; CovðeÞ ¼ 2 W; and e follows a N ð0; 2 W Þ ð84Þ
where W, the variance–covariance matrix of the errors, is an N  N positive definite square symmetric matrix.
Because of the properties of W, it can be shown that it is possible to find a nonsingular symmetric matrix U
such that
U – 1 U ¼ Ut U ¼ UU ¼ W ð85Þ
 1
By writing e ¼ U e, it is easy to prove that
Eðe Þ ¼ 0 and Covðe Þ ¼ 2 I ð86Þ
In addition, as e is a linear combination of normal distributions, it also holds that e follows a N(0, 2I).
Now, if in Equation (82) both members are multiplied by U1, the following new model is obtained:
U – 1 y ¼ U – 1 Xb þ U – 1 e ð87Þ
 
y ¼ X b þ e ð88Þ
2
whose residuals are NID(0, ). The LS solution of Equations (82) and (88) is written in the matrix form as
!
b0  –1
b¼ ¼ ðXt X Þ – 1 Xt y ¼ Xt W – 1 X Xt W – 1 y ð89Þ
b1

and its variance–covariance matrix is given by

 –1
CovðbÞ ¼ Xt W – 1 X 2 ð90Þ

The residuals that must be checked are those corresponding to e that are obtained for e ¼ U1(Y Ŷ), where
Ŷ ¼ Xb and b is obtained from Equation (89). These weighted residuals are used by the Analytical Methods
Committee4 to evaluate the LOF.
An estimate of  is given by
rffiffiffiffiffiffiffiffiffiffiffi
et e
s¼ ð91Þ
N –2
Often the error covariance matrix W is not completely known and the information provided by the residuals
can suggest a parametric specification with just a few unknown parameters. The Gauss–Markov optimality of
Equation (89) is lost when W is substituted by an estimate Ŵ except for large sample sizes.
According to the hypotheses on the residuals, several procedures have been studied in detail. When matrix
W is not a diagonal matrix, that is, when the residuals of Equation (82) are correlated, it is known as generalized
least squares (GLS) and, when it is diagonal, it is known as weighted least squares (WLS). In the last case, the
recorded signals for the calibration samples, and hence the residuals, are independent but with different
variance from one level of concentration to another. This is the most frequent situation in chemical analysis.
Also, an iterative procedure to estimate the diagonal W has been proposed. By taking a trial Ŵi , bi is estimated,
then re-estimating Ŵiþ1 and biþ1 until convergence, it is named iterative weighted least squares (IWLS). See,
for example, Ryan’s book.12
The generalized linear model (GLIM) is about the case in which the distribution of the error is an
exponential one (the normal distribution is one of them) and/or additionally, h(y), a nonlinear, monotone,
156 Quality of Analytical Measurements: Univariate Regression

differentiable function of the vector of responses y, that verifies h(y) ¼ Xb þ e. The maximum likelihood
solution for estimating the parameters, b of GLIM, involves the use of generalized (weighted) LS in an iterative
way. Since the work by Nelder and Wedderburn,46 the GLIM procedure has become a technique of increasing
use although it needs special computation programs that are more and more accessible. The complexity of
GLIM prevents its in-depth discussion in the present chapter, and the interested reader can consult Chapter 16
of the book by Draper and Smith11 for an introduction and the references cited therein for further details.
In many practical situations, it is difficult to initially have specific information on W, Equation (84). For this
reason, usually it is necessary to make the hypothesis (mistaken) that W ¼ I and to try to discover something
about matrix W examining the residuals of the usual LS regression.
The LS method provides estimates of the coefficients as bLS ¼ ðXt XÞ – 1 Xt y and E ðbLS Þ ¼ ðXt XÞ – 1 Xt Xb
¼ b but CovðbLS Þ ¼ ðXt XÞ – 1 Xt WXðXt XÞ – 1 2 . If the correct analysis is done, according to Equation (90),
–1
CovðbGLS Þ ¼ ðXt W – 1 XÞ 2 holds. The consequence of this wrong estimation is that the variances of the
independent term and slope in bLS are greater than the correct ones in bGLS. However, as the estimate bLS is
unbiased, the calculated values of the estimates are rarely affected by the fact that the correct regression has not
been made. Therefore, the analysis of the residuals of the fitting by LS allows checking the lack of homogeneity of
variances and even estimating the weights to be used, that is, matrix W.
In an analytical calibration, it can be assumed that W is a diagonal matrix and, thus, the WLS approach
proves to be adequate.
Sometimes, the observation of the residuals allows fitting the variance as a function, f, of the predictor
variable.11 Davidian and Haaland47 describe a procedure to estimate the GLS and the variance function, f, by
means of an iterative procedure. It is also possible to use a GLIM to estimate the values of W.
A common alternative consists of estimating the values wi of the main diagonal of matrix W by using
replicates in each level of concentration, xi.
In a calibration experiment, the estimate of the concentration of the problem sample, xu, is similar when using
the LS or WLS regression, but its standard deviation will be different, and, further, depending on the position in
the range of concentration, the incorrect estimation VarLS(xu) may be lesser or greater than VarWLS(xu).20,21,48
Tellinghuisen49 has analyzed recently this subject by using Monte Carlo simulation and showed that the use of
replicates gives great improvement at small concentrations but can underperform LS regression in the mid-to-
large concentration region. Nevertheless, an estimation of the variance function approximates minimum
variance even though the true variance functions are not well reproduced.
The WLS regression for calibration has been recommended by several organizations;50,51 consequently, it is
being used with increasing frequency in analytical chemistry. A survey of the analytical applications of the use
of weighted linear regression is given in Table 10 of Asuero and González.52 Therefore, it is interesting to
describe the basic inferential aspects of WLS regression.
As much if the variance for each xi is estimated with replicates, si2, as if this is done by a function s2 ¼ f(x), it is
usual to estimate the elements, wi, of the diagonal of W by 1=si2 (or 1=f ðxi Þ).
The confidence intervals and the hypothesis tests for the slope and the independent term are computed as in the
LS regression but taking the values of variances of Equation (90) and the residual standard deviation in Equation (91).
Given a value of concentration, x0, the response is estimated by
!
b0
ŷ0 ¼ ð 1 x0 Þ ð92Þ
b1

and the estimate of its variance is

!
 1 1
dðŷ 0 Þ ¼ ð 1
Var x0 Þ X Ŵt 1
X s2 ð93Þ
x0

Thus, the confidence interval for the expected mean value at level (1)100% is

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
dðŷ 0 Þ; ŷ0 þ t =2;N – 2 Var
ŷ 0 – t =2;N – 2 Var dðŷ 0 Þ ð94Þ
 Quality of Analytical Measurements: Univariate Regression 157

The estimated variance of the prediction for the average of q new observations is
" !#
1  1 1
dðŷ 0 Þ ¼
Var þ ð 1 x0 Þ Xt c
W1 X s2 ð95Þ
w0 q x0

where w0 is the weight corresponding to x0. As a consequence, the confidence interval is

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!ffi
u  1 1
u 1
ŷ0 t =2;N – 2 s t þ ð 1 x0 Þ Xt W c1 X ð96Þ
w0 q x0

Finally, once the calibration is over, for a problem sample with value y0, the mean of q recorded signals in q
replicates, the estimated value of the concentration is

y0 – b0
x̂0 ¼ ð97Þ
b1

An argument similar to that used to obtain Equation (69) provides the corresponding approximate confidence
interval
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
s 1 1 ðy0 – yw Þ2
x̂0 t =2;N – 2 þ PN þ 2 PN ð98Þ
b1 w0 q i¼1 wi b1 i¼1 wi ðxi – xw Þ2
PN PN PN PN
where again w0 is the weight corresponding to x0 and yw ¼ i¼1 wi yi = i¼1 wi and xw ¼ i¼1 wi x i = i¼1 wi
are the weighted means.

Example 6
For the determination of benzaldehyde by differential pulse polarography,53 a calibration was made with nine
levels of concentration and four replicates in each level. Table 9 shows the intensity recorded together with the
means and variances in each level of concentration.
The LS regression for these N ¼ 36 data (xi, yi) is y ¼ 0.0021 þ 1.9401x
The estimated standard deviations are 0.004697 and 0.04147 for the independent term and slope,
respectively.
Figure 10(a) shows the standardized residuals and evidences that the dispersion of the residuals increases
with concentration. The Cochran test to decide if one of the nine variances is significantly different from the
rest has p-value equal to 0.18; thus, there is no evidence to reject the null hypothesis at level 0.05. This is a quite
general situation in analytical calibrations because even when the variances vary much, usually they vary
progressively in such a way that none of variances is significantly greater or smaller than the others. However,

Table 9 Experimental data (with means and variances) of the determinations of

benzaldehyde

Concentration (mmol l–1) Current (mA) Mean, xi Variance, si2

0.0199 0.0330 0.0371 0.0340 0.0330 0.0343 3.73  106

0.0450 0.0823 0.0914 0.0905 0.0845 0.0872 2.00  105
0.0593 0.1233 0.1149 0.1134 0.1155 0.1168 1.96  105
0.0884 0.1769 0.1573 0.1606 0.1680 0.1657 7.58  105
0.0980 0.2068 0.1874 0.1913 0.1777 0.1908 1.47  104
0.1270 0.2553 0.2243 0.2340 0.2456 0.2398 1.83  104
0.1370 0.2825 0.2417 0.2612 0.2689 0.2636 2.90  104
0.1640 0.3427 0.3175 0.2961 0.3078 0.3160 3.93  104
0.1740 0.3660 0.3311 0.3097 0.3447 0.3379 5.59  104
158 Quality of Analytical Measurements: Univariate Regression

(b)
(a) 2.5 2.5
2.0 2.0
1.5
Standardized residuals

1.5

Standardized residuals
1.0
1.0
0.5
0.5
0.0
–0.5 0.0

–1.0 –0.5
–1.5 –1.0
–2.0 –1.5
–2.5 –2.0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Benzaldehyde (mmol l –1) Benzaldehyde (mmol l –1)
Figure 10 Standardized residuals: (a) for the LS regression and (b) for the WLS regression.

the Bartlett test, which is adequate to detect groups of variances, has a p-value equal to 0.01, which means that,
at a significance level of 0.05, the null hypothesis of homogeneity of variances should be rejected.
It is interesting to note that the residuals in Figure 10(a) are compatible with the hypothesis of normal
distribution. The tests of 2, Kolmogoroff, Saphiro–Wilks, skewness, and kurtosis have p-values of 0.43, 0.78,
0.97, 0.65, and 0.33, respectively; thus, clearly, at a significance level of 0.05, the hypothesis of normal residuals
cannot be rejected.
The plot of si2 versus xi suggests a quadratic relationship, which we estimate by LS as
s 2 ¼ 0:00004 – 0:00195 x2
x þ 0:02670
which explains 96.27% of the variance of s2.
Each individual xi value of concentration can be substituted in this equation, then estimate si2 (i ¼ 1, . . ., 36)
and invert these values to give the estimated weights ŵi of the diagonal of matrix W. Then, from Equation (89),
the WLS regression is
y ¼ – 0:0021 þ 1:9565x
with estimated standard deviations 0.00196 and 0.03392 for the independent term and the slope, respectively. As
the theory says, the estimated values of b0 and b1 are similar as both regressions provide unbiased estimates.
However, the difference is in the standard deviation estimated for each one, which is reduced by 58% for the
independent term and by 18% for the slope.
Figure 10(b) shows the standardized residuals of WLS regression; unlike the LS case now they are
practically equal for all the values of the concentration. Strictly speaking, they should have been represented
pffiffiffiffiffi
versus wi xi , but to make the comparison with the LS residuals easier, the unweighted values xi are
maintained; this does not modify the clear pattern.
Figure 11 shows the length of the confidence intervals (x–h, x þ h) at level 95%. For each abscissa value, x,
the ordinate shows the two values þh and –h. The red ones are those obtained when using LS; the continuous
line represents values that correspond to a mean signal (Equation (60)) and the dashed line those obtained in
prediction of a future signal, q ¼ 1 in Equation (63). The length of the analogous intervals obtained with WLS is
in blue; the continuous line represents the mean signal (Equation (94)) and the dashed line for the prediction
(Equation (96)).
The remarkable asymmetry in the length of the intervals introduced by the weighting is evident. Up to
0.1 mmol l1 of benzaldehyde, the intervals are shorter with WLS than with the incorrect LS, especially
between 0.03 and 0.04 mmol l1.
Much more interesting from the analytical point of view is the confidence interval (x0h, x0 þ h) for the
concentration x0 determined in a problem sample, Equation (69) for LS and Equation (98) for WLS. In both
cases, we use q ¼ 1 and 95% confidence level. Figure 12 shows along abscissas the values of x0 and along
 Quality of Analytical Measurements: Univariate Regression 159

0.06

Length of the confidence interval

0.04

0.02

0.00

–0.02

–0.04

–0.06
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20
Benzaldehyde (mmol l –1)
Figure 11 Length of confidence intervals for LS (in red) and for WLS (in blue). Continuous lines mark values corresponding
to the mean signal and the dashed lines those obtained for prediction of a future sample.

0.03

0.02
Length of the confidence interval

0.01

0.00

–0.01

–0.02

–0.03
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20
Benzaldehyde (mmol l –1)
Figure 12 Length of confidence intervals for LS (in red) and for WLS (in blue).

ordinates the values of h: red for LS and blue for WLS. Up to the concentration of 0.11 mmol l1, the interval
determined by WLS is shorter than that determined when incorrectly using LS.
The analysis of Figure 12 shows that the most important impact of WLS is in the decision limit and the
capability of detection. When x0 ¼ 0, with the (incorrect) LS calibration model, the capability of detection is
0.0232 mmol l1 for  ¼  ¼ 0.05. However, with the WLS calibration model, the capability of detection is
0.0118 mmol l1 for the same probabilities of false positive and false negative.

1.05.9 Robust Regression in Calibration

In a linear univariate calibration, the visual inspection of the data can indicate the presence of outlier data, that
is, data that are different from the majority and do not follow the general linear tendency. The first observation
is that the outlier data are not necessarily incorrect or mistaken data. Sometimes, their presence indicates some
160 Quality of Analytical Measurements: Univariate Regression

phenomenon unexpected when the experimentation began but that has chemical or experimental explanation.
The problem raised by outlier data is regarding their detection, after which the analyst will be able to evaluate
their significance and act accordingly.
The outlier data give rise to highly asymmetric distributions or distributions with great accumulated
probability far from the central values. In these conditions, the hypothesis of normality is not acceptable. In
addition, it is necessary to consider that, in a calibration, the sample sizes are not so large so that the normality
tests behave conservatively and do not detect ‘lack of normality’.
The effect of outlier data on LS regression may be notable because it first affects the estimates of the
intercept, b0, the slope, b1, and the concentration, x̂0 , of a problem sample. It also affects the precision of these
2
estimates through syx .
On examination of the equations of the confidence intervals of the coefficients of the model (Equations (40),
(41), (44), and (47)), those of the predictions (Equations (60), (62), and (63)), and those for the concentration of
the problem sample (Equations (68) and (69)), and also considering all the hypothesis tests used on these
2
parameters, it is clear that they depend on a good estimate of the residual variance syx of the random error, ", of
the model. It suffices to observe how this estimate is made in Equation (10) to deduce that the presence of
outlier data inflates this value and, as a consequence, all the intervals become larger, that is, the estimates will be
2
less precise. On the contrary, in the hypothesis tests about the slope and intercept, the syx appears in the
denominators, therefore causing the statistic tcalc to take smaller values, leading to the more frequent acceptance
of null hypotheses when they should not be accepted.
Several robust regression methods insensitive (to a greater or lesser extent) to the presence of outlier data
have been developed.11,54 This is why they are able to detect the true linear calibration curve and allow the
analyst to decide, with objectivity, which data disagree and to re-establish the conditions under which the LS
regression is optimal.
Different robust regression methods have already been introduced to analytical chemistry (updated revisions
of these applications can be found in Ortiz et al.55 and Daszykowski et al.56). Also, the ISO has recommended its
use (ISO 5725-5).
Given the relevance of the subject, the present book devotes a complete chapter to the introduction of robust
methods (Chapter 3.07). The robust regression methods are usually evaluated according to two indices: (1) The
‘sampling breakdown’: the smallest fraction of contamination that can cause the estimator to take on values
arbitrarily far from the correct ones. (2) The ‘influence function’ that describes the effect on the estimator of an
infinitesimal contamination in the sample. In a robust procedure, this function must be bounded.
For LS regression, one outlier suffices to carry LS over all bounds. Therefore, for a sample of size n, its
breakdown is 1/n, which tends to 0 for increasing sample size. With regard to the influence function, the LS
estimator is not bounded on both axes; hence, it is very sensitive to any deviation of data as much in the
abscissas (‘leverage point’) as in the ordinates (‘outlier point’), and, of course, both deviations can come together
(‘outlier-leverage point’).
Rousseeuw57 developed the so-called least median squares (LMS) regression and Massart et al.58 showed the
advantages of its use in chemical analysis. The theory and versatility of its application can be seen in Rousseeuw and
Leroy54 and Rousseeuw.59 The least median squares estimators of 0 and 1 are the values that minimize, instead of
the sum of squares, the median of squares of residuals, that is, those that minimize the following function R:
 
Rðb0 ; b1 Þ ¼ median ei2 ; i ¼ 1; 2; :::; N ¼ median ðyi – ŷi Þ2 ; i ¼ 1; 2; :::; N
 ð99Þ
¼ median ðyi – b0 – b1 xi Þ2 ; i ¼ 1; 2; :::; N

The LMS method has a bounded influence function on both axes and 50% breakdown. The effect of outlier
data in an LS regression depends on their relative position in the cloud of points, which is measured by indices
as the ‘leverage effect’.10 In a calibration, there will not be any leverage points, that is, samples whose
concentration greatly differs from the rest. But if the experimental data include concentration values beyond
the linear range of the analytical method, it is clear that they will exert a similar influence on regression to that
of a leverage point.
Furthermore, the LMS estimator has the property of the ‘exact fit’, which guarantees that if at least 50% of
the data (x, y) follow a linear model, the LMS regression finds it. This property is of great interest when one
 Quality of Analytical Measurements: Univariate Regression 161

needs to determine the range of linearity at the same time as the sensitivity of the analytical method, which is
usually the case in calibration problems.
The relation of LMS with the least trimmed squares (LTS) and a comparison with other 10 procedures
(including fuzzy methods) can be seen in Ortiz et al.55 For more statistical background about LTS and other
references, see Rousseeuw and Leroy54 and Stromberg;60 for computational advances, see Watson.61 The recent
developments in PROGRESS62 and fast LTS63,64 have been included in a MATLAB library of robust statistical
methods.65
The procedure for the calibration with LMS regression consists of the following: (1) obtaining with the
calibration data the LMS regression line, (2) evaluating the standardized robust residuals to establish the data
that can be considered outlier (if the residual in absolute value is greater than 2.5, the data are rejected), and (3)
re-making the LS regression without the outlier data detected with the LMS regression.
Traditionally, the detection of nonaligned data, decisive for the establishment of the range of linearity, is
carried out by a detailed inspection of the residuals of the LS regression and many indices such as Cook
distance, DFFITS, and DFBETAS.10,11 This procedure requires great experience in the analysis of regressions,
which is perhaps not available in the routine working of a chemistry laboratory. The availability of a
computerized method such as the LMS regression, with sufficient mathematical properties to successfully
perform this task, is the key to the calibration proposed.

Example 7
The linear range in the determination of nickel by square-wave adsortive-stripping voltammetry (SWAdSV)
using dimethylglyoxime (DMG) was studied.66 The saturation of the electrode surface, caused by the
voluminous complex of Ni–DMG, leads to an increasing curvature. Table 10 shows the experimental data
corresponding to a calibration of this system. For more details, see Sanllorente et al.66
The LS regression estimates the slope as 193.12 and the independent term as 9.64 with standard deviations
14.05 and 3.37, respectively. The correlation coefficient is 0.96 and syx is 7.64 mA.
The externally studentized residuals are not large; all of them are less than 2.5 in absolute value. Further, the
p-values of the normality tests are 0.26, 0.77, 0.09, and 0.58 for the tests of 2, Kolmogoroff, Shapiro–Wilks, and
skewness, respectively. However, the residuals show a systematic tendency, nonrandom, as can be seen in

Table 10 Experimental data and residuals of the regression least

squares (LS) and least median squares regression (LMS) for the
calibration of nickel

Externally studentized LMS standardized

Nickel (mmol l–1) Current (mA) LS residuals residuals

0 2.50 1.04 0.54

0.024 7.22 1.00 0.17
0.047 12.71 0.84 0.48
0.070 19.13 0.56 0.30
0.093 25.15 0.34 0.29
0.116 30.79 0.17 0.42
0.139 39.02 0.33 0.66
0.161 46.14 0.72 1.33
0.183 52.83 1.06 1.78
0.205 57.13 1.07 1.09
0.227 60.76 0.98 0.07
0.249 66.50 1.20 0.07
0.270 70.47 1.19 0.66
0.291 73.12 0.99 2.02
0.312 73.46 0.48 4.51
0.333 76.27 0.31 5.79
0.374 72.13 1.42 12.96
0.395 76.87 1.33 13.32
0.415 75.01 2.47 16.73
162 Quality of Analytical Measurements: Univariate Regression

80

60

Current (μA)
40

20

0
0 0.1 0.2 0.3 0.4 0.5
Nickel (μmol l –1)
Figure 13 Regression line by least squares for data of Table 10.

Figure 13. At the beginning, they are negative, then positive, and finally negative again. The most important
conclusion is that the calibration line does not reproduce the experimental data in the region of the concentra-
tions in which the electrode has not still saturated and the response is theoretically linear.
The LMS regression, with independent term 1.389 and slope 261.91, points as outlier data those correspond-
ing to the greatest five concentrations (fourth column of Table 10) because their standardized robust residuals
are greater than 2.5 in absolute value. In addition, they show a systematic tendency for those five data; the
residuals ei ¼ yi – ŷi are negative and decreasing. That is, the experimental data of intensity do not lie on a
straight line, and deviate more and more from it with increasing concentration, which is concordant with the
effect of saturation of the electrode.
Once the outlier data are removed and the LS regression is repeated with the remaining 14 data, the intercept
is 2.05 and the slope is 257.66 with standard deviations of 1.05 and 6.04, respectively. The standard deviation of
the regression syx is now 2.04 and the correlation coefficient is 0.997. When comparing these results with those
of the first LS regression, the negative effect of the outlier data on the estimates and, thus, on the confidence
intervals and the capability of detection is evident.
Figure 14 shows the calibration line obtained with LS without the outlier data detected by LMS regression.
Still a systematic tendency is observed in the residuals, a tendency that is assumable given the correlation
coefficient and the residual standard deviation, and taking into account the fact that the LOF cannot be tested
because there are no replicates. As a consequence, the linear range of this calibration can be established as the
interval of concentrations up to 0.291 mM.
The property of the exact fitting of LMS regression allows removing the influence of the curvature in the
determination of the end point in titration analysis, as graphically shown in Figure 15. This approach has been
used67 to determine the equation of the linear segments in amperometric, conductimetric, spectrophotometric,
and potentiometric titrations.

80

60
Current (μA)

40

20

0
0 0.05 0.1 0.15 0.2 0.25 0.3
Nickel (μmol l –1)
Figure 14 Regression line by LS without the outlier data.
 Quality of Analytical Measurements: Univariate Regression 163

Corrected signal

V V′
Added volume
Figure 15 V9 is the end volume calculated when the first linear section has been fitted by LS (dashed line). V is the end
volume calculated when the first linear section has been fitted by LS without outlier data detected by means of LMS (solid line).

1.05.10 Errors in Both Variables

In the previous sections, the predictor variable (e.g., the concentration of the standard samples) is supposed to
be determined without random error, unlike the recorded signal that is affected by a random variation. In
practice, there is no quantity experimentally determined without variability (random error) but, in many
occasions, the error in the predictor variable is much less than that of the response; in this case, a model as the
one in Equation (3) is adequate. However, consider for instance the case of comparing two analytical methods
by linear regression; in this case, the abscissa contains the concentrations determined by one method and the
ordinates those by the second method; therefore, the model supposed in Equation (3) cannot be adequate
because both concentrations will have a similar experimental variability.
The problem is posed by accepting that both the predictor variable, x, and the response, y, are

yi ¼ i þ "i ; i ¼ 1; :::; N
ð100Þ
xi ¼ i þ  i ; i ¼ 1; :::; N

and there is a linear relation between the true unobserved variables  y 

 ¼ 0 þ 1  ð101Þ
The solution to the problem of estimating the parameters 0 and  1 and the variance of the errors " and 
depends on the hypothesis under the distribution of the errors. Adcock68 suggested 130 years ago, minimizing
the sum of squared perpendicular distances to the fitted line.
Let us assume that "i are NID(0,2") and that i are NID(0,2) with " and  uncorrelated. It can be proved11
that the solution by LS will provide
PN little
 or no bias in the following cases: (1) If 2 is small compared with the
2 2  2
spread  in the i values (  ¼ i¼1 i –  =N ). (2) If the x’s are fixed by the experimenter. (3) If we want to
fit yi ¼ i þ "i , where i ¼ 0 þ 1 xi (the observed xi) instead of Equation (101).
The maximum likelihood estimates of  0 and  1 have an identifiability problem; the estimation cannot be
carried out without some additional information being added, for example knowledge of the ratio
¼ 2" =2 .
This situation can be assumed when analysis methods are compared because the repeatability of both methods
can previously be known.
If
were known, maximum likelihood leads to estimates
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN 2 PN 2ffi
PN P P
Þ2 –
Ni¼1 ðxi – xÞ2
i¼1 ðyi – y i¼1 ðyi – y Þ2 –
Ni¼1 ðxi – xÞ2 þ 4
i¼1 ðxi – x Þðyi – yÞ
b1 ¼ P þ P ð102Þ
2 Ni¼1 ðxi – xÞðyi – yÞ 2 Ni¼1 ðxi – xÞðyi – yÞ
b0 ¼ y – b 1 x
164 Quality of Analytical Measurements: Univariate Regression

P P qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN P
Note that, if
is estimated as Ni¼1 ðyi – yÞ2 = Ni¼1 ðxi – xÞ2 , then b1 ¼ i¼1 ðyi – y Þ2 = Ni¼1 ðxi – xÞ2 , which is
the geometric mean functional relationship. In this case, b1 is the geometric mean of the slope in LS fit of y versus x
and the inverse of slope of x versus y, so b1 is a compromising value lying in between the two equation slopes. In
addition, the geometric mean functional relationship minimizes the sum of areas obtained by drawing horizontal and
vertical lines from each data point to the fitted line. A more detailed discussion can be seen in Draper and Smith.11
An important particular case of Equation (102) corresponds to
¼ 1, that is, when both variances 2" and 2
are equal. The solution, called orthogonal regression, is the line that minimizes the sum of squares of
perpendicular deviations from the line. This line is the same as the first principal component of the data
(xi, yi), i ¼ 1, . . ., N, and is the solution proposed by Adcock. In Section 15.2 of Jackson’s book,69 the inferential
procedures for the orthogonal regression line can be found.
A very detailed and rigorous statistical analysis of diverse general models for the errors in Equations (100)
and (101) can be found in the book by Fuller.70
Also, solutions based on LS have been proposed, that is, without supposing any distribution on the errors "
and . In the approach by Ripley and Thompson,71 it is considered that " and  are uncorrelated and
heteroscedastic. The solution is then obtained by an IWLS procedure. By adequately modifying the weights,
the previous iterative procedure allows considering correlated errors in both axes, the so-called bivariate LS
procedure.72,73
Finally, it is proved that the total least squares (TLS)74 procedure estimates the true parameters of Equations
(100) and (101) consistently, that is, the solution that provides TLS converges to the true values of 0 and  1
when N tends to infinity. This property of TLS estimates does not depend on any assumed distribution for the
errors " and . TLS is a very useful method for the problem of errors in variable estimation.

1.05.11 Final Remark

There are many regression procedures that can be of interest for calibration in analytical chemistry; however,
there is no room for them in a chapter of introduction to the subject. Even so, the user of a regression for
analytical calibration must draw her/his calibration data. The numerical indices that serve to evaluate the
quality of a regression and therefore guarantee the analytical result and its precision are a limited vision of the
calibration. The advantage of the univariate calibration is the possibility of the visual inspection of the data and
it should not renounce to it, as well as to a careful reflection on the type of errors that are expected to obtain.
The unthinking uses of a linear regression lead to seriously misleading conclusions. In this respect, the example
of Anscombe75 is didactic; Table 11 shows the results of four linear regressions, N ¼ 11, with a single replicate.
The equality of the estimates and their standard deviations is evident; thus, any other parameter deduced from
them (e.g., the linearity coefficient, value and precision of the estimated response) will also be equal. Even the
statistic for the significance of the regression indicates that they are ‘equally’ significant.
It suffices to show the data graphically, Figure 16, to see the error made in the last three cases when using a
linear regression because the data, except for case 1, were not compatible with the supposed model.

Table 11 Analysis by means of a linear least squares regression of the Anscombe’s data

Case 1 Case 2 Case 3 Case 4

Intercept, b0 3.0001 3.0009 3.0024 3.0017

Slope, b1 0.5001 0.5000 0.4997 0.4999
Standard deviation of b0 1.1248 1.1253 1.1245 1.1239
Standard deviation of b1 0.1179 0.1180 0.1179 0.1178
Standard deviation of regression, syx 1.2366 1.2372 1.2363 1.2357
Correlation coefficient, r 0.82 0.82 0.82 0.82
Fcalc (Equation (28))a 17.99 17.99 17.97 18.00
p-value of Durbin–Watson test 0.010 0.372 0.402 0.290
a
p ¼ 0.002.
 Quality of Analytical Measurements: Univariate Regression 165

(a) (b)
12.2 11.1
Response variable, y

Response variable, y
10.2 9.1

8.2 7.1

6.2 5.1

4.2 3.1
4 6 8 10 12 14 4 6 8 10 12 14
Predictor variable, x Predictor variable, x
(c) (d)
13.3 13.2
Response variable, y

Response variable, y
11.3 11.2

9.3 9.2

7.3 7.2

5.3 5.2
4 6 8 10 12 14 8 10 12 14 16 18 20
Predictor variable, x Predictor variable, x
Figure 16 Graphical representation of the data from Anscombe: (a) case 1; (b) case 2; (c) case 3; and (d) case 4 of Table 11.
The lines are those fitted by least squares.

References

1. Dux, J. P. Handbook of Quality Assurance for the Analytical Chemistry Laboratory, 2nd ed.; Van Nostrand Reinhold: New York, 1990.
2. IUPAC Recommendation, Guidelines for Calibration in Analytical Chemistry. Part 1 Fundamentals and Single Component
Calibration. Pure Appl. Chem. 1998, 70, 993–1014.
3. International Standard ISO 11843. Capability of Detection. Part 1. Terms and Definitions, 1997 and Part 2. Methodology in the
Linear Calibration Case, 2000; International Organisation for Standardisation: Geneva, Switzerland, 1997 and 2000.
4. Analytical Methods Committee. Is my Calibration Linear? Analyst 1994, 119, 2363–2366.
5. Mulholland, M.; Hibbert, D. B. Linearity and the Limitations of Least Squares Calibration. J. Chromatogr. A 1997, 762, 73–82.
6. Huber, W. On the Use of the Correlation Coefficient r for Testing the Linearity of Calibration Functions. Accred. Qual. Assur. 2004,
9, 726.
7. Hibbert, D. B. Further Comments on the (Miss-)Use of r for Testing the Linearity of Calibration Functions. Accred. Qual. Assur.
2005, 10, 300–301.
8. Ellison, S. L. R. In Defence of the Correlation Coefficient. Accred. Qual. Assur. 2006, 11, 146–152.
9. Asuero, A. G.; Sayago, A.; González, A. G. The Correlation Coefficient: An Overview. Crit. Rev. Anal. Chem. 2006, 36, 41–59.
10. Cook, R. D.; Weisberg, S. Residuals and Influence in Regression; Chapman and Hall: New York, 1982.
11. Draper, N. R.; Smith, H. Applied Regression Analysis, 3rd ed.; John Wiley and Sons: New York, 1998; p 43.
12. Ryan, T. P. Modern Regression Methods; John Wiley and Sons: New York 1997; p 64.
13. Brüggemann, L.; Quapp, W.; Wennrich, R. Test for Non-Linearity Concerning Linear Calibrated Chemical Measurements.
Accred. Qual. Assur. 2006, 11, 625–631.
14. Cuadros, L.; Garcı́a, A. M.; Jiménez, C.; Román, M. Estimation of Performance Characteristics of an Analytical Method Using the
Data Set of the Calibration Experiment. Anal. Lett. 1993, 26, 1243–1258.
15. de Beer, J. O.; de Beer, T. R.; Goeyens, L. Assessment of Quality Performance Parameters for Straight Line Calibration Curves
Related to the Spread of the Abscissa Values Around their Mean. Anal. Chim. Acta 2007, 584, 57–65.
16. Arroyo, D.; Ortiz, M. C.; Sarabia, L. Multiresponse Optimization and Parallel Factor Analysis, Useful Tools in the Determination of
Estrogens by Gas Chromatography–Mass Spectrometry. J. Chromatogr. A 2007, 1157, 358–368.
17. Sachs, L. Applied Statistics. A Handbook of Techniques; Springer-Verlag: New York, 1982; p 442.
18. Brown, P. J. Measurement, Regression and Calibration; Clarendon Press: Oxford, 1993; pp 26–29.
19. Cox, D. R.; Hinkley, D. V. Theoretical Statistics; Chapman and Hall: London, 1974.
20. Massart, D. L.; Vandeginste, B. G. M.; Buydens, L. M. C.; de Jong, S.; Lewi, P. J.; Smeyers-Verbeke, J., (eds.), Handbook of
Chemometrics and Qualimetrics, Part A, Data Handling in Science and Technology 20A; Elsevier: Amsterdam, 1998.
21. Meier, P. C.; Zünd, R. E. Statistical Methods in Analytical Chemistry, 2nd ed.; John Wiley and Sons: New York, 2000; pp 118–120.
22. Fisher, R. A. Statistical Methods for Research Workers, 13th ed.; Oliver and Boyd: Edinburg, 1963.
23. Lark, P. D.; Craven, B. R.; Bosworth, R. C. L. The Handling of Chemical Data; Pergamon: London, 1968.
24. Liteanu, C.; Rica, I. Statistical Theory and Methodology of Trace Analysis; Ellis Horwood: Chichester, 1980.
25. Carter, K. N. Jr.; Scott, D. M.; Salmon, J. K.; Zarcone, G. S. Confidence Limits for the Abscissa of Intersection of Two Least-
Squares Lines such as Linear Segmented Titration Curves. Anal. Chem. 1991, 63, 1270–1278.
166 Quality of Analytical Measurements: Univariate Regression

26. International Union of Pure and Applied Chemistry. Analytical Chemistry Division. Commission on Spectrochemical and Other
Optical Procedures for Analysis. Nomenclature, Symbols, Units and Their Usage in Spectrochemical Analysis-II. Data
Interpretation. Spectrochem. Acta B 1978, 33, 241–245.
27. Currie, L. A. Detection and Quantification Limits: Origins and Historical Overview. Anal. Chim. Acta, 1999, 391, 127–134.
28. Gibbons, R. D. Statistical Methods for Groundwater Monitoring. Wiley: New York, 1994.
29. Inczédy, J.; Lengyel, T.; Ure, A. M.; Gelencsér, A.; Hulanicki, A. Compendium of Analytical Nomenclature, 3rd ed.; International
Union of Pure and Applied Chemistry, Port City Press, Inc.: Baltimore 2000; p 74.
30. Clayton, C. A.; Hines, J. W.; Elkins, P. D. Detection Limits with Specified Assurance Probabilities. Anal. Chem. 1987, 59,
2506–2514.
31. Garcı́a, I.; Sarabia, L. A.; Ortiz, M. C. Detection Capability of Tetracyclines Analysed by a Fluorescence Technique: Comparison
between Bilinear and Trilinear Partial Least Squares Models. Anal. Chim. Acta 2004, 501, 193–203.
32. Ortiz, M. C.; Sarabia, L. A.; Garcı́a, I.; Giménez, D.; Meléndez, E. Capability of Detection and Three-Way Data. Anal. Chim. Acta
2006, 559, 124–136.
33. Sarabia, L. A.; Ortiz, M. C. DETARCHI: A Program for Detection Limits with Specified Assurance Probabilities and Characteristic
Curves of Detection. Trends Anal. Chem. 1994, 13, 1–6.
34. Verdon, E.; Hurtaud-Pessel, D.; Sanders, P. Evaluation of the Limit of Performance of an Analytical Method Based on a Statistical
Calculation of its Critical Concentrations According to ISO Standard 11843: Application to Routine Control of Banned Veterinary
Drug Residues in Food According to European Decision. Accred. Qual. Assur. 2006, 11, 58–62.
35. Ortiz, M. C.; Sarabia, L. A.; Herrero, A.; Sánchez, M. S.; Sanz, B.; Rueda, M. E.; Giménez, D.; Meléndez, M. E. Capability of
Detection of an Analytical Method Evaluating False Positive and False Negative (ISO 11843) with Partial Least Squares. Chemom.
Intell. Lab. Syst. 2003, 69, 21–33.
36. Sanz, M. B.; Sarabia, L. A.; Herrero, A.; Ortiz, M. C. Capability of Discrimination: Application to Soft Calibration Methods. Anal.
Chim. Acta 2001, 446, 297–311.
37. Garcı́a, I.; Ortiz, M. C.; Sarabia, L.; Vilches, C.; Gredilla, E. Advances in Methodology for the Validation of Methods According to
the International Organization for Standardization. Application to the Determination of Benzoic and Sorbic Acids in Soft Drinks by
High-Performance Liquid Chromatography. J. Chromatogr. A 2003, 992, 11–27.
38. Bosch-Reig, F.; Campins Falcó, P. Campins Falcó P., H-Point Standard Additions Method. Part 1 Fundamentals and Application
to Analytical Spectroscopy. Analyst 1988, 113, 1011–1016.
39. Campı́ns-Falcó, P.; Verdú-Andrés, J.; Bosch-Reig, F.; Molı́ns-Legua, C. Generalized H-Point Standard Additions Method for
Analyte Determinations in Unknown Samples. Anal. Chim. Acta 1995, 302, 323–333.
40. Tortajada-Genaro, L. A.; Campı́ns-Falcó, P.; Blasco-Gómez, F.; Bosch-Reig, F. The Generalized H-Point Standard-Additions
Method to Determine Analytes Present in Two Different Chemical Forms in Unknown Matrix Samples. Part I. General
Considerations. Analyst 2000, 125, 771–776.
41. Booksh, K.; Henshaw, J. M.; Burgess, L. W.; Kowalski, B. R. A Second-Order Standard Addition Method with Application to
Calibration of a Kinetics-Spectroscopic Sensor for Quantitation of Trichloroethylene. J. Chemom. 1995, 9, 263–282.
42. Saurina, J.; Tauler, R. Strategies for Solving Matrix Effects in the Analysis of Triphenyltin in Sea-Water Samples by Three-Way
Multivariate Curve Resolution. Analyst 2000, 125, 2038–2043.
43. Sena, M. M.; Trevisan, M. G.; Poppi, R. J. Combining Standard Addition Method and Second-Order Advantage for Direct
Determination of Salicylate in Undiluted Human Plasma by Spectrofluorimetry. Talanta 2006, 68, 1707–1712.
44. Herrero, A.; Zamponi, S.; Marassi, R.; Conti, P.; Ortiz, M. C.; Sarabia, L. A. Determination of the Capability of Detection of a
Hyphenated Method: Application to Spectroelectrochemistry. Chemom. Intell. Lab. Syst. 2002, 61, 63–74.
45. Bahram, M.; Bro, R. A Novel Strategy for Solving Matrix Effect in Three-Way Data Using Parallel Profiles with Linear
Dependencies. Anal. Chim. Acta 2007, 584, 397–402.
46. Nelder, J. A.; Wedderburn, R. W. M. Generalized Linear Models. J. R. Stat. Soc. Ser. A 1972, 135, 370.
47. Davidian, M.; Haaland, P. D. Regression and Calibration with Nonconstant Error Variance. Chemom. Intell Lab. Syst. 1990, 9,
231–248.
48. Miller, J. N. Basic Statistical Methods for Analytical Chemistry. Part 2. Calibration and Regression Methods. Analyst 1991, 116,
3–14.
49. Tellinghuisen, J. Weighted Least-squares in Calibration: What Difference Does it Make? Analyst 2007, 6, 536–543.
50. Analytical Methods Committee. Uses (Proper and Improper) of Correlation Coefficients. Analyst 1988, 113, 1469–1471.
51. EPA, Document # 815 R 05 006. Statistical Protocol for the Determination of the Single-Laboratory Lowest Concentration
Minimum Reporting Level and Validation of Laboratory Performance at 0 Below the Minimum Reporting Level, 2004.
52. Asuero, A. G.; González, G. Fitting Straight Lines with Replicated Observations by Linear Regression. III. Weighting Data. Crit.
Rev. Anal. Chem. 2007, 37, 143–172.
53. Sanz, B. Metodologı́a para determinar ı́ndices de calidad de técnicas analı́ticas que incorporan calibrados multivariantes.
Aplicación al Electroanálisis. Ph. D. thesis, University of Burgos, 2002.
54. Rousseeuw, P. J.; Leroy, A. M. Robust Regression and Outliers Detection. John Wiley and Sons Inc.: Hoboken, NJ, 2001.
55. Ortiz, M. C.; Sarabia, L.; Herrero, A. Robust Regression Techniques. A Useful Alternative for the Detection of Outlier Data in
Chemical Analysis. Talanta 2006, 70, 499–512.
56. Daszykowski, M.; Kaczmarek, K.; Vander Heyden, Y.; Walczak, B. Robust Statistics in Data Analysis – A Review. Basic
Concepts. Chemom. Intell. Lab. Syst. 2007, 85, 203–219.
57. Rousseeuw, P. Least Median of Squares Regression. J. Am. Stat. Assoc. 1984, 79, 871–880.
58. Massart, D. L.; Kaufman, L.; Rousseeuw, P. J.; Leroy, A. Least Median of Squares: A Robust Method for Outlier and Model Error
Detection in Regression and Calibration. Anal. Chim. Acta 1986, 187, 171–179.
59. Rousseeuw, P. J. Tutorial to Robust Statistics. J. Chemom. 1991, 5, 1–20.
60. Stromberg, A. J. Computing the Exact Least Median of Squares Estimate and Stability Diagnostics in Multiple Linear Regression.
SIAM J. Sci. Comput. 1993, 14, 1289–1299.
61. Watson, G. A. Computing the Least Quantile of Squares Estimate. SIAM J. Sci. Comput. 1998, 19, 1125–1138.
 Quality of Analytical Measurements: Univariate Regression 167

62. Rousseeuw, P. J.; Hubert, M. In L1-Statistical Procedures and Related Topics; Dodge, Y., Ed.; Institute of Mathematical
Statistics: Hayward, CA, 1997; p 201.
63. Pison, G.; Van Aelst, S.; Willems, G. Small Sample Corrections for LTS and MCD. Metrika 2002, 55, 111–123.
64. Rousseeuw, P. J.; Van Driessen, K. Computing LTS Regression for Large Data Sets. Estadı́stica 2002, 54, 163–190.
65. Verboven, S.; Hubert, M. LIBRA: A MATLAB Library for Robust Analysis. Chemom. Intell. Lab. Syst. 2005, 75, 127–136.
66. Sanllorente, S.; Ortiz, M. C.; Arcos, M. J.; López-Palacios, J. Application of an Optimization Procedure in Adsorptive Stripping
Voltammetry to the Determination of Nickel with DMG. Electroanalysis 1996, 8, 285–292.
67. Ortiz, M. C.; Herrero, A. Regression by Least Median Squares, A Methodological Contribution to Titration Analysis. Chemom.
Intell. Lab. Syst. 1995, 27, 231–243.
68. Adcock, R. J. A Problem in Least Squares. Analyst (des Moines, Iowa) 1878, 5, 53–54.
69. Jackson, J. E. A User’s Guide to Principal Components. John Wiley and Sons: New York, 1991.
70. Fuller, W. A. Measurement Error Models. John Wiley and Sons: New York, 1987.
71. Ripley, B. D.; Thompson, M. Regression Techniques for the Detection of Analytical Bias. Analyst 1987, 112, 377–383.
72. Lisý, J. M.; Cholvadová, A.; Kutej, J. Multiple Straight-Line Least-Squares Analysis with Uncertainties in All Variables. J. Comput.
Chem. 1990, 14, 189–192.
73. Riu, J.; Rius, F. X. Univariate Regression Models with Errors in Both Axes. J. Chemom. 1995, 9, 343–362.
74. Van Huffel, S.; Lemmerling, P.; Eds. Total Least Squares and Errors-In Variables Modeling. Kluwer Academic Publishers:
Dordrecht, 2002.
75. Anscombe, F. J. Graphs in Statistical Analysis. Am. Stat. 1973, 27, 17–21.
168 Quality of Analytical Measurements: Univariate Regression

Biographical Sketches

Marı́a Cruz Ortiz received her Ph.D. in chemistry from the University of Valladolid (Spain)
in 1988. She is a member of the staff of analytical chemistry in the University of Burgos
(Faculty of Sciences) since 1989. She does research and teaching on analytical chemistry and
chemometrics. Her research activity has focused on experimental design, optimization,
pattern recognition, quality assurance, and validation of analytical methods according to
official regulations, multivariate and/or multiway regression models. All of these applied to
problems in food chemistry, typification, and so on, which produce about 100 papers. She is at
present the head of an active research group, the chemometrics and qualimetrics group of the
University of Burgos.

M Sagrario Sánchez obtained her Ph.D. in mathematics from the University of Valladolid in
1997. She is working at the University of Burgos since 1991 and as a member of the
permanent staff since 2002. The teaching activities are mostly directed to graduate and
postgraduate courses in chemistry and science and food technology. She is also a permanent
member of the chemometrics and qualimetrics group since its foundation. Her main research
activities are developed into the areas of interest of the group that include modeling and
analysis of n-way data, modeling of categories, design of experiments, optimization, and so
on, by using classical methods as well as computationally intensive methods (such as neural
networks or evolutionary algorithms).
 Quality of Analytical Measurements: Univariate Regression 169

Luis A. Sarabia received his Ph.D. in Statistics from the University of Valladolid (Spain) in
1979. Since 1974, he has been teaching Statistics and Mathematics mostly to graduate and
postgraduate students of Chemistry. At present, his research is centred on Chemometrics as a
member of the Chemometrics and Qualimetrics group of the University of Burgos. His
research activities include development of software and implementation of nonparametric
and robust statistical methods, genetic algorithms, neural networks, etc. He is also involved in
multivariate/multiway regression methods, methodology of the experimental design, quality
assurance, and validation. He is the author of about a hundred papers on these matters.