Talanta 172 (2017) 221–229

Contents lists available at ScienceDirect

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

Some practical considerations for linearity assessment of calibration curves MARK

as function of concentration levels according to the fitness-for-purpose
approach
⁎ ⁎
J.M. Juradoa, , A. Alcázara, R. Muñiz-Valenciab, S.G. Ceballos-Magañac, F. Raposod,
a
Departament of Analytical Chemistry, Faculty of Chemistry, University of Seville, c/ Profesor García González 1, 41012 Seville, Spain
b
Facultad de Ciencias Químicas, Universidad de Colima, Carretera Colima-Coquimatlán km 9, 28400 Coquimatlán, Colima, México
c
Facultad de Ciencias, Universidad de Colima, c/ Bernal Díaz del Castillo 340, 28045 Colima, México
d
Instituto de la Grasa, Consejo Superior de Investigaciones Científicas (IG-CSIC), Campus Universitario Pablo de Olavide, Carretera de Utrera km 1, Edificio
46, 41013 Seville, Spain

A R T I C L E I N F O A BS T RAC T

Keywords: Since linear calibration is mostly preferred for analytical determinations, linearity in the calibration range is an
Calibration curves important performance characteristic of any instrumental analytical method. Linearity can be proved by
Fitness-for-purpose approach applying several graphical and numerical approaches. The principal graphical criteria are visual inspection of
Linearity the calibration plot, the residuals plot, and the response factors plot, also called sensitivity or linearity plot. All
Response factor
of them must include confidence limits in order to visualize linearity deviations. In this work, the graphical
Relative error
representation of percent relative errors of back-calculated concentrations against the concentration of the
calibration standards is proposed as linearity criterion. This graph considers a confidence interval based on the
expected recovery related to the concentration level according to AOAC approach. To illustrate it, four
calibration examples covering different analytical techniques and calibration situations have been studied. The
proposed %RE graph was useful in all examples, helping to highlight problems related to non-linear behavior
such as points with high leverage and deviations from linearity at the extremes of the calibration range. By this
way, a numerical decision limit which takes into account the concentration of calibration standards can be easily
included as linearity criterion in the form of %RETh=2·C−0.11. Accordingly, this %RE parameter is accurate for
the decision-making related to linearity assessment according to the fitness-for-purpose approach.

1. Introduction
Linearity is one of the most important performance characteristics
of an instrumental analytical method when linear response is assumed;
thus it is an interesting parameter to be evaluated in validation studies
[1]. After adjusting a linear function to a set of calibration data, points
are expected to be randomly distributed around the obtained line. But
reality goes its own way and points can show trends differing from
linear behavior, mainly at low or high concentration levels. The
accuracy of an analytical method is highly linked to linearity, which
must be assured over the whole calibration range [2]. There are
numerous approaches for linearity testing, including both graphical
and numerical methods [3]. Visual inspection of calibration and
residuals plot are very useful, but they must be accompanied by
significance tests or numerical decision parameters [4,5]. Linearity
plots, obtained by representing the response factors versus concentra-
tion level, are very useful to indicate the range in which sensitivity
remains practically constant within a confidence limit of 5% according
to Dorschel et al. [6] and Huber [7]. A number of statistical tests, such
as experimental Fratio (IUPAC) [8], ANOVA lack-of-fit (LOF) [9,10],
Mandel's test [11,12] and significance of the quadratic term [5] have
been proposed to check linearity. Literature on linearity also consider
several numerical parameters, such as correlation coefficient (not
advisable alone), standard deviation of regression, standard deviation
of the slope [13], quality coefficients [14,15] and back-calculated
concentrations expressed as percentage of relative error (%RE)
[3,16–19]. The main problem is that some of these procedures are
not equivalent and contradictory results can be obtained. For this
reason, common sense must be used to select adequate tools in order to
assure linearity according to the fitness for purpose principle [20].
There is no magical recipe for linearity testing, but two important
aspects should be always considered, the easy visualization of data and

⁎
Corresponding authors.
E-mail addresses: jmjurado@us.es (J.M. Jurado), fraposo@cica.es (F. Raposo).

http://dx.doi.org/10.1016/j.talanta.2017.05.049
Received 3 March 2017; Received in revised form 10 May 2017; Accepted 15 May 2017
Available online 18 May 2017
0039-9140/ © 2017 Elsevier B.V. All rights reserved.
J.M. Jurado et al. Talanta 172 (2017) 221–229

Table 1
Calibration examples used in this work.

Example variable Data

1 [Cu], µg L−1 2, 2, 2, 4, 4, 4, 6, 6, 6, 8, 8, 8, 10, 10, 10

Absorbance, a.u. 0.023, 0.022, 0.023, 0.046, 0.047, 0.045, 0.066, 0.066, 0.065, 0.089, 0.090, 0.088, 0.113, 0.112, 0.113

2 [Quinine], mg L−1 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 5, 6, 6, 6, 7, 7, 7, 8, 8, 8, 9, 9, 9, 9, 10, 10, 10

Intensity, RFU 138.455, 137.885, 137.535, 248.2, 247.956, 247.466, 354.044, 354.117, 353.668, 455.47, 454.756, 454.312, 558.784, 558.449, 558.006,
650.048, 650.142, 649.991, 723.440, 723.672, 723.068, 815.512, 815.315, 815.493, 895.258, 895.022, 894.987, 972.986, 971.926,
970.385

3 [Zn], µg L−1 10, 10, 10, 20, 20, 20, 50, 50, 50, 75, 75, 75, 100, 100, 100, 500, 500, 500
Intensity, counts 161, 148, 155, 233, 224, 217, 455, 459, 448, 629, 639, 621, 795, 814, 799, 4043, 4031, 3970

4 [Isoamil acetate], mg L−1 0.5, 1, 2, 3, 4, 5, 10, 15, 20

Area ratio, dimensionless 0.013, 0.028, 0.058, 0.087, 0.131, 0.178, 0.379, 0.599, 0.804

Fig. 1. Linearity study for the calibration curve of copper by ET-AAS in the calibration range 2–10 µg L−1 (Example 1). A, Absorbance; a.u., absorbance units; RF, response factor.

a clear acceptance criterion allowing the decision-making. Finally, a plot of %RE of back-calculated concentrations against
In this work, a combination of graphical and numerical criteria for concentration level offers additional information. By this way, a
linearity testing is proposed by considering the adequacy to the numerical decision limit which takes into account the concentration
purposes of the method in a specified concentration range. Four of calibration standards can be easily included.
calibration datasets covering usual situations that can occur in
chemical analysis laboratories were selected. Calibration and residuals
2. Theory
plots are mandatory to inspect the fitting quality [21]. They must
include confidence limits, allowing us to visualize the goodness of
Consider a set of n calibration points (xi, yi), with apparent linear
linear fitting and detect outliers according to statistical criteria.
relationship, where each xi-value corresponds to a known concentra-
Linearity plot, based on the variation of the response factors, is also
tion of an analyte and each yi-value correspond to an instrumental
considered because it can highlight the needed of non-linear fitting.
signal. Data can be fitted by a linear calibration model obtained by

222
J.M. Jurado et al. Talanta 172 (2017) 221–229

Fig. 2. Linearity study for the calibration curve of quinine by fluorescence spectroscopy in the calibration range 1–10 mg L−1 (Example 2a). I, intensity; RFU, relative fluorescence units;
RF, response factor.

ordinary least squares (OLS) method by minimizing the sum of squares mental data are inside the confidence limits of the fitted model. The
of residuals: confidence limits must consider the leverage effect of each calibration
n n point on the regression line. The leverage score, hi, varying from 0 to 1,
S= ∑ ei2 = ∑ (yi − yˆ)i 2 measures the amount by which the predicted value would change if the
i =1 i =1 (1) observation was shifted one unit in the y direction. Leverage score of
where ŷi is the predicted dependent variable obtained from each each point in the case of XY linear regression is obtained by the
independent value xi as expression [4].

yˆi = b1⋅xi + b0 (2) 1 (x − x ) 2

hi = + n i
n ∑i =1 (xi − x )2 (3)
being b0 the estimate of intercept and b1 the estimate of slope for the
considered linear model. This leverage score is used to obtain the standard error (sŷ) of each
Following, different graphical criteria for linearity testing are estimate ŷi at position xi:
explained.
syˆ = sy / x⋅ hi (4)
2.1. Calibration plot and coefficients of correlation & determination
where sy/x is the standard error of the estimates, also known as residual
standard deviation.
The inspection of calibration plot must be the first action to be
And the 95% confidence interval (CI) for each estimate can then be
considered in the evaluation of linearity. This practice can be useful to
obtained from:
find out possible outliers, points of influence and non-linear data
trends. Relating to the coefficient of correlation (r), although widely CI = yˆ ± tcrit ·sŷ (5)
used, it is important to note that must not be used to check regression
data [5]. With respect to the coefficient of determination (R2), it must where tcrit is the t-Student value for α=0.05 and n-2 degrees of
not be considered as a measure of linearity taking into account that R2 freedom.
is the ratio of the variability explained by the model with regard to the A representation of CI against the concentration allows us to obtain
total (observed) variability. Sometimes, R2 values of 0.999 can be the confidence region for the calibration line. If we are interested in the
obtained for linear regression but the distribution of data along the confidence limits for the points used to obtain the regression line, the
calibration line indicates non-linear behavior. On the contrary, lower prediction interval (PI) must be computed as [8]:
R2 can be assumed for linear relationship depending on the fitness for
1
purpose of the method [5]. There are lots of well-established scientists PI = yˆ ± tcrit ·sy / x · + hi
m (6)
recommending not using this parameter as synonymous of linearity
[3]. where m is the number of replicates used to obtain the mean signal for
A better approach to evaluate linearity is verifying that experi- each calibration point. Experimental data must lie inside this region as

223
J.M. Jurado et al.
Fig. 3. Linearity study for the calibration curve of quinine by fluorescence spectroscopy in the calibration range 1–6 mg L−1 (Example 2b). I, intensity; RFU, relative fluorescence units;
RF, response factor.
a first test to assure that the model really fit the data. Accordingly,
outliers and deviations from linearity are highlighted when calibration
points appear out of this region. For simplicity, only PI will be plotted
in the graphics presented in this work.
2.2. Residuals plot
Residuals plots are strongly recommended in several manuals on
chemometrics and quality control [5,21–23]. Plotting residuals can
help us to detect outliers, influence points, lack of fit or uneven
variance. If the selected function fits well the data, residuals must
appear normally distributed and centered around zero. The use of
scaled residuals, i.e., residuals (ei) divided by sy/x is strongly recom-
mended. They are very useful because a sample from the standard
normal distribution N (0, 1) is obtained and the computed residuals
can be easily compared with the corresponding z-value. But this
standardization does not consider the leverage effect of each point on
the regression line. The standard error of each residual (sei ) can be
obtained by considering the leverage score and then compute the so
called Studentized residual as:
ei ei
ti = = errors must lie between ± 15% for all the calibration range, and
sei sy / x⋅ 1 − hi
Studentized residuals must lie between ± 1.96 limits (for conve-
nience rounded to ± 2) over the whole calibration range. On the
contrary, the selected model is not acceptable.
2.3. Response factors plot
Linearity plots, which are very used in chromatographic method
development, are obtained plotting the signal-to-concentration ratio
(the so called response factor) against the concentration of calibration
Talanta 172 (2017) 221–229
standards [6,7]. The median of the computed response factors, defined
as Rc, is traced as a central line along the calibration range, and two
5%-confidence lines obtained as 0.95·Rc and 1.05·Rc, are also plotted.
If the computed response factors lie within the two limits, it clearly
indicates that linearity can be assured over the calibration range.
Positive and negative deviations from linearity are usually found at low
and high concentration levels, respectively.
2.4. Percent relative error of back-calculated concentrations (%RE)
When calibration data are fitted by a model and the calibration
function is obtained, it is possible to calculate a predicted concentration
value for each calibration standard by substituting their corresponding
measured signal into the function. The difference between this so called
back-calculated concentration and the nominal concentration is the
error associate to the concentration of each standard when it is
predicted by the model. This difference can be used to calculate a RE
dividing its value by the nominal content. This parameter was used
previously in the analytical literature to try obtaining some measure of
the error when using the proposed regression equation [16–19].
According to some documents, to verify linear range these relative
(7)
between ± 20% when limit of quantification (LOQ) is reached [3,24].
Although these literature limits are accepted, however the dependence
on the concentration of calibration standards is not considered. The
Guidelines for Standard Method Performance Requirements published
in the AOAC Official Methods of Analysis proposes an expected
recovery interval as a function of analyte concentration [25] (Table
S1 in the Supplementary electronic material). The differences between
the limits of this interval and 100% could be considered as the expected
relative error depending on the concentration level. For instance, at
analyte concentrations of 0.1%, the AOAC recovery lies between 95%
224
J.M. Jurado et al.
Fig. 4. Linearity study for the calibration curve of zinc by ICP-AES in the calibration range 10–500 µg L−1 (Example 3a). I, emission intensity; RF, response factor.
and 105%, and the expected relative error at this level could be
established as ± 5%. We propose as novelty the use of these concen-
tration-dependent relative errors as a limit for linearity testing over the
calibration range. The main problem is that the intervals proposed by
AOAC are not symmetric. For instance, at 100 ppm the recovery
interval is 90–107%. When lower concentration levels are considered
this asymmetry grows. With the aim of establishing a single criterion,
we propose the use of the upper level of the AOAC recovery intervals, as
it is the most restrictive situation, to obtain the %RE for each
concentration level. The computed relative errors are included in
Table S1. Our final objective is to obtain a plot of relative errors of
back calculated concentrations versus concentration levels, henceforth
called the %RE-plot. In order to facilitate the drawing of confidence
limits on the %RE-plot, the computed relative errors has been related
with the analyte ratio expressed in mass fraction (C) using a function in
the form:
%RE = a·C −b
As can be seen, this function is very similar to the Horwitz function
relating reproducibility and analyte content [25,26]. The fitting proce-
dure was carried out in an Excel Spreadsheet by using Solver comple-
ment to obtain parameters a and b. In order to ensure the adequate
value of %RE at analyte ratio of 1, the value of a was restricted to ≤2
(that is the expected %RE at this concentration level). The errors of the
fitted parameters were evaluated by means of numerical differentiation
[27]. The results obtained were a=2.0 ± 0.4 and b=0.11 ± 0.01; giving a
R2 value of 0.95. The supplementary electronic material includes the %
RE-plot (Fig. S1), calculated from AOAC, against analyte ratio in
logarithmic scale. This plot also includes the representation of the fitted
function. As can be seen, the function fits better at the extremes of the
analyte ratio that at the central points. It is logical because from 10−7 to
10−5 the %RE-value expected according AOAC is always the same, 10%.
Talanta 172 (2017) 221–229
The estimated %RE-value at C=10−7 is 11.9%. Although this value is
higher than 10%, we consider our approximation adequate because it is
lower than the 15% limit proposed by previous literature [3,24]. From
10−6 to 10−3, our proposed limit always is more restrictive than the
expected according to AOAC recovery model. By this way, Eq. (9) is
proposed to estimate the theoretical %RE-values from any mass
fraction over the calibration range.
%RETh = 2·C −0.11 (9)
By using this expression, the calculated confidence %RE-limits can
be easily included on the %RE-plot for the graphical evaluation of
linearity.
3. Examples of analytical calibration
Analytical calibration is an overall procedure to express the
(8) relationship between concentration of analytes and detector response
that can be applied to different analytical techniques (atomic absorp-
tion spectroscopy-AAS, fluorescence spectroscopy-FS, inductively
coupled plasma-ICP, gas chromatography-GC, and high performance
liquid chromatography-HPLC). On this manuscript, the examples have
been selected mainly to cover different situations for calibration curves,
independently of the technique selected to carry out the measurements.
The data sets are shown in Table 1.
Example 1. External calibration data used for the direct
determination of Cu in aniseed spirits by electrothermal atomic
absorption spectroscopy (ET-AAS) [28]. Each standard was measured
in triplicate from genuine replicates.
Example 2. Calibration curve for quinine measured by FS.
Fluorescence measurements were carried out by using a Shimadzu
225
J.M. Jurado et al.
Fig. 5. Linearity study for the calibration curve of zinc by ICP-AES in the calibration range 10–100 µg L−1 (Example 3b). I, emission intensity; RF, response factor.
RF-1501 spectrofluorophotometer. Quinine standards were prepared
in 0.05 M sulfuric acid and fluorescence intensity was recorded.
Excitation and emission wavelengths were set at 350 nm and
450 nm, respectively. Each standard was measured in triplicate from
genuine replicates.
Example 3. External calibration data used for the determination of Zn
in acid-digested aniseed spirits by inductively coupled plasma atomic
emission spectroscopy (ICP-AES) [29]. Each standard was measured in
triplicate from genuine replicates.
Example 4. Internal calibration data used for the determination of
isoamil acetate in wines by head-space solid-phase microextraction and
GC with flame ionization detector (HS-SPME-GC-FID) [30]. Ethyl
heptanoate (10 mg L−1) was used as internal standard. As the HS-
SPME extraction time was 60 min for each standard, the calibration
points were measured once.
Calibration plots, standardized residuals plots, response factors
plots and %RE-plots have been obtained in order to comparatively
study the linearity for each of the four selected examples. All these
methods have been applied by plotting their corresponding variables
and drawing the confidence limits. For each figure four graphs have
been plotted: A) Calibration plot including prediction interval; B) Plot
of Studentized residuals including ± z limits (z=2); C) Response factor
linearity plot including ± 5% limits (α=0.05); D) %RE-plot including
± 15% limits (dotted lines) and the proposed limits according to the
concentrations of calibration standards (dashed lines).
4. Results and discussion
4.1. Example 1. A linear calibration with evenly spaced data
Fig. 1 shows the plots obtained in the case of the Example 1. As can
226
Talanta 172 (2017) 221–229
be seen, all calibration points are found within the prediction interval
(Fig. 1A) and R2 present a value of 0.9991. Studentized residual plot
(Fig. 1B) also shows all the computed residuals within the ± 2 limits.
The same can be observed in the case of response factors (Fig. 1C),
whose values are included between ± 5%. In Fig. 1D, back-calculated
relative errors (%RE) are represented against concentration of stan-
dard solutions. This plot includes the ± 15% limits (dotted lines) and
the limits, proposed by authors, related to the concentration levels
(dashed lines). Note that the proposed limits are slightly greater than
± 15% at the lowest levels. This fact can soften the decision-making
when calibration is performed at very low concentrations. In this case,
all %RE values lie within both limits (theoretical and calculated) and
then linearity can be assured on the whole calibration range.
4.2. Example 2. A curvilinear calibration with evenly spaced data
Molecular fluorescence is a highly sensitive technique. At low
concentration levels, the dependence between signal and analyte
content can be considered linear. However, it is usual to find non-
linear relationships between signal and concentration at high levels
and, accordingly, non-linear calibration models can be fitted [31,32]. In
this example, the calibration range was intentionally selected from 1 to
10 mg L−1 to highlight this kind of situations. As can be seen in Fig. 2A
linear model does not fit well to data and the residual distribution,
although all calibration points appeared within the limits of the
prediction interval. The distribution of the residuals (Fig. 2B) also
shows a non-linear behavior. According to the profile that can be seen
on the response factor plot (Fig. 2C), sensitivity decreases when
concentration of calibration standards is raised. Note that in the %
RE-plot, showed in Fig. 2D, the proposed limits are lower than 15%,
due to the concentration of the standards. In this graph, a high
deviation of the first calibration point is observed, although the
J.M. Jurado et al.
Fig. 6. Linearity study for the calibration curve of isoamyl acetate by HS-SPME-GC-FID in the calibration range 0.5–20 mg L−1 (Example 4a). AR, area ratio (dimensionless); RRF,
relative response factor, obtained as AR divided by C.
problem becomes from high concentration levels. It is known that this
deviation is due to a bad selection of the calibration levels and,
accordingly, some reduction in the calibration range must be necessary.
A new shorter calibration line was obtained in the range 1–
6 mg L−1. As can be seen in Fig. 3A, all calibration points lie within
the prediction interval of the regression curve, but the distribution of
residuals still shows non-linear trends. In fact, a second grade
polynomial would fit also the data leading to a R2 value of 0.9999. In
addition, linearity plot based on response factors also indicates a non-
linear response. From our experience with this kind of plots, in some
situations it is usual to find strong variations in the response factors,
but adequate linearity for the fitness for purpose of the determination
method is achieved. According to the profile of the response factor
showed in Fig. 3C, sensitivity still decreases when concentration of
calibration standards is raised. This is a typical deviation that occurs in
a number of real situations, but analysts can accept it if interpolation
errors are kept within adequate limits. When Fig. 3D is checked, it can
be seen that the %RE of all back-calculated concentrations are
distributed within the calculated acceptance limits and, in fact, they
are lower than 6%. In view of such results, linearity can be assured in
this calibration range.
4.3. Example 3. A linear calibration using uneven data distribution
Uneven data distribution is one of the principal causes of linearity
failing due to the influence of points with high leverage. Ideally, the
concentrations should be evenly spaced across the calibration range to
avoid this problem. As can be seen in Fig. 4, in the case of Example 3,
not equidistant calibration points were selected and the distance
between concentration levels grow along the x-axis. In Fig. 4A, it can
be seen how the last calibration point presents a high influence on the
calibration line. The determination coefficient presents a value of
227
Talanta 172 (2017) 221–229
0.9997, however the calibration must not be considered adequate.
The calibration points lie inside the prediction interval lines (Fig. 4A),
but the residuals plot (Fig. 4B) shows a not random distribution of the
five residuals corresponding to lower levels. Linearity plot (Fig. 4C)
shows a high deviation of the two first response factors. Finally, paying
attention to Fig. 4D, the first point is out of any acceptance limit of %
RE.
Considering these results, a new shorter calibration range must be
selected to obtain reliable results. Obviously, the problematic concen-
tration level is 500 µg L−1 and, accordingly, the calibration line has
been optimized by using the range from 10 to 100 µg L−1. The results
can be seen in Fig. 5. In this case, the R2 value is 0.9996. Fig. 5B shows
that the Studentized residuals were not randomly distributed. In
addition, four of the response factors (Fig. 5C) appear out of the ±
5% limits. Actually, response factor behaves as if the model was not
linear. This is the same situation found in Example 2, although
polynomial model could fit better, the linear model is selected due to
easy interpretation and calculations because the fitness for purposes is
achieved. In fact, %RE of back-calculated contents are lower than 3%
and lie within the accepted calculated limits, as can be seen in Fig. 5D.
4.4. Example 4. A linear calibration with deviations at lowest
concentration levels
Deviations from linearity are also frequent at low concentration,
when the measurement is done close to the quantification limit. At that
concentration levels the signal usually tends at the background signal.
In this example, positive deviation is observed in Fig. 6A, although all
calibration points appear inside the prediction interval, and no random
distribution is observed for the four first residuals (Fig. 6B). As
expected, the relative response factors (RRFs) increases with the
concentration of the calibration standards (Fig. 6C). The two first %
J.M. Jurado et al.
Fig. 7. Linearity study for the calibration curve of isoamyl acetate by HS-SPME-GC-FID in the calibration range 3–20 mg L−1 (Example 4b). AR, area ratio (dimensionless); RRF,
relative response factor, obtained as AR divided by C.
RE values are higher than 25%, being both data out of the proposed
limits.
The calibration line was accordingly built by reducing the calibra-
tion range from the lowest concentration level (Fig. 7). These points
were sequentially removed and finally, the calibration plot was
obtained using 6 points in the range of 3–20 mg L−1. The R2 value
increased from 0.9988 to 0.9998. In addition, the residuals profile is
randomly distributed, as can be seen in Fig. 7B. The response factor
still increases with concentration and most of the points appear out of
± 5% limits. However, the %RE-plot presents all points within the
calculated confidence interval, with values lower than 2.5%. Taking
into account these results, linear response can be assumed within this
calibration range according to the fitness for purpose approach.
5. Conclusions
The evaluation of linearity of calibration functions is a very
important subject relating to the selection of a model for unknown
samples quantitation. Traditionally, three different plots were consid-
ered for the assessment of linearity. Firstly, the widely used calibration
plot by means of the lack of reliability R2 criterion. Secondly, the
valuable residuals plot or the more sophisticated Studentized version.
Thirdly, the very useful response factors plot. A fourth graphical
criteria based on the relative error of back-calculated concentrations
(%RE-plot) has been also considered. For the first time, decision limits
are included in the plot depending on the concentration levels of
calibration standards. These limits are derived from the AOAC
expected recovery at each analyte level.
For typical linear calibration curves, the four graphical plots show
agreement between theory and experimental data. However, for day-to
day analytical calibration procedures the trends including non-linear
behavior are very frequent, mainly due to the influence of high leverage
228
Talanta 172 (2017) 221–229
points and deviations at low concentration levels. In these cases, the
results provided by the different graphical plots could be controversial.
To a better understanding of linearity subject, the main objective of this
paper is to propose the different graphical plots in the form of a flow
chart to definitively making a decision about the goodness-of-fit for
calibration curves. First of all, the quality of the adjustment is observed
in the calibration and residual plots. Following, the response factors
plot allow us to know any deviation on sensitivity in the considered
calibration range. The evolution of the response factors could indicate
some deviations from linearity and the need of other calibration
function. Finally, it is important to note the goodness-of-fit provided
by the %RE-plot in all these examples. By this way, after selecting the
appropriate calibration range, the linearity of response was assured by
means of this graph. Therefore, we propose here that for linearity
assessment the fitness for purpose of the method should prevail over
any other criterion. For that reason, the final decision about linearity
should be made accurately considering either numerically or graphi-
cally representing the plot of %RE of back-calculated concentrations. In
any case, the limits should be established according to the analyte
concentration level.
Appendix A. Supplementary material
Supplementary data associated with this article can be found in the
online version at doi:10.1016/j.talanta.2017.05.049.
References
[1] E. Hsieh, J.P. Liu, On statistical evaluation of the linearity in assay validation, J.
Biopharm. Stat. 18 (2008) 677–690.
[2] International Conference on Harmonisation (ICH) of technical requirements for
registration of pharmaceuticals for human use. Harmonized Tripartite Guideline.
Validation of Analytical Procedures: text and Methodology Q2 (R1), 2005.
J.M. Jurado et al.
[3] F. Raposo, Evaluation of analytical calibration based on least squares linear
regression for instrumental techniques: a tutorial review, TRAC-Trend Anal. Chem.
77 (2016) 167–185.
[4] S.V.C. De Souza, R.G. Junqueira, A procedure to assess linearity by ordinary least
squares method, Anal. Chim. Acta 552 (2005) 23–35.
[5] S.L.R. Ellison, V.J. Barwick, T.J. Duguid, Practical Statistics for the Analytical
Scientist, The Royal Society of Chemistry, Cambridge, 2009.
[6] C.A. Dorschel, J.L. Ekmanis, J.E. Oberholtzer, F.V. Warren Jr., B.A. Bidlingmeyer,
LC detectors: evaluation and practical implications of linearity, Anal. Chem. 61
(1989) 951A–968A.
[7] L. Huber, Validation of analytical methods: review and strategy, LC-GC Eur. 11
(1998) 96–105.
[8] K. Danzer, L.A. Currie, Guidelines for calibration in analytical chemistry. Part 1:
fundamentals and single component calibration, Pure Appl. Chem. 70 (1998)
993–1014.
[9] Analytical Methods Committee, Is my calibration linear? Analyst 119, 1994, 2363–
2366.
[10] M.G. Akritas, N. Papadatos, Heteroscedastic one-way ANOVA and lack-of-fit tests,
J. Am. Stat. Assoc. 466 (2004) 368–382.
[11] J. Mandel, The statistical Analysis of Experimental Data, Dover Publications, New
York, 1964.
[12] J.M. Andrade, M.P. Gómez-Carracedo, Notes on the use of Mandel´s test to check
for nonlinearity in laboratory calibrations, Analyst 5 (2013) 1145–1149.
[13] A.M. Cuadros-Rodríguez, C. García Campaña, Jimenez Linares, M. Roman Ceba,
estimation of performance characteristics of an analytical method using the data set
of the calibration experiment, Anal. Lett. 26 (1993) 1243–1258.
[14] P. Vankeerberghen, J. Smeyers-Verbeke, The quality coefficient as a tool in
decisions about the quality of calibration in graphite furnace atomic absorption
spectrometry, Chemom. Intell. Lab. Syst. 15 (1992) 195–202.
[15] J.O. de Beer, C. Naert, E. Deconinck, The quality coefficients as performance
assessment parameter of straight line calibration curves in relationship with the
number of calibration points, Accredit. Qual. Assur. 17 (2012) 265–274.
[16] E.L. Jhonson, D.L. Reynolds, D.S. Wright, L.A. Pachla, Biological sample pre-
paration and data reduction concepts in pharmaceutical analysis, J. Chromatogr.
Sci. 26 (1988) 372–379.
[17] M. Mulholland, D.B. Hibbert, Linearity and the limitations of least squares
calibration, J. Chromatogr. A 762 (1997) 73–82.
229
Talanta 172 (2017) 221–229
[18] P. Hubert, P. Chiap, J. Crommen, B. Boulanger, E. Chapuzet, N. Mercier,
S. Bervoas-Martin, P. Chevalier, D. Grandjean, P. Lagorce, M. Lallier, M.C. Laparra,
M. Laurentie, J.C. Nivet, The SFSTP guide on the validation of chromatographic
methods for drug bioanalysis: from the Washington Conference to the laboratory,
Anal. Chim. Acta 391 (1998) 135–148.
[19] J.W. Dolan, Calibration curves, part 3: a different view, LC-GC Eur. 22 (2009)
304–308.
[20] Eurachem Guide. The fitness for purpose of analytical methods. A Laboratory
Guide to Method Validation and Related Topics, Second ed., 2014.
[21] M. Thompsom, P.J. Lowthian, Notes on Statistics and Data Quality for Analytical
Chemist, Imperial College Press, London, 2011.
[22] E. Mullins, Statistics for the Quality Control Chemistry Laboratory, The Royal
Society of Chemistry, Cambridge, 2003.
[23] J.M. Miller, J.C. Miller, Statistics and Chemometrics for Analytical Chemistry, 6th
ed., Pearson Education Ltd, Harlow, 2010.
[24] US EPA. Method 8000C-Determinative Chromatographic Separations, Revision 3,
March, 2003, p. 66.
[25] AOAC, Appendix F: guidelines for standard methods performance requirements,
in: AOAC Official Methods of Analysis, 2012.
[26] W. Horwitz, L.R. Kamps, K.W. Boyer, Quality assurance in the analysis of foods and
trace constituents, J. Assoc. Off. Anal. Chem. 63 (1980) 1344–1354.
[27] E.J. Billo, Excel for Scientist and Engineers: numerical methods, John Wiley &
Sons, Hoboken, EEUU, 2007.
[28] J.M. Jurado, M.J. Martín, F. Pablos, A. Moreda, P. Bermejo-Barrera, Direct
determination of copper, lead and cadmium in aniseed spirits by electrothermal
atomic absorption spectroscopy, Food Chem. 101 (2007) 1296–1304.
[29] J.M. Jurado, A. Alcázar, M.J. Martín, F. Pablos, A.G. González, Determination of
Zn, B, Fe, Mg, Ca, Na and Si in anisette samples by inductively coupled plasma
atomic emission spectrometry, Talanta 63 (2004) 297–302.
[30] J.M. Jurado, O. Ballesteros, A. Alcázar, F. Pablos, M.J. Martín, J.L. Vilchez,
A. Navalón, Differentiation of certified brands of origins of Spanish white wines by
HS-SPME-GC and chemometrics, Anal. Bioanal. Chem. 390 (2008) 961–970.
[31] D.A. Skoog, F.J. Holler, S.R. Crouch, Principles of Instrumental Analysis, 6th ed.,
Thomson Brook/Cole, Belmont, USA, 2007.
[32] J.M. Jurado, R. Muñiz-Valencia, A. Alcázar, S.G. Ceballos-Magaña, J. González,
Fitting chemical data with Excel: a practical tutorial (in Spanish), Educ. Quím. 27
(2016) 21–29.