You are on page 1of 5

Article

pubs.acs.org/ac

Coordinate Swapping in Standard Addition Graphs for Analytical


Chemistry: A Simplied Path for Uncertainty Calculation in Linear
and Nonlinear Plots
Juris Meija,* Enea Pagliano, and Zoltan Mester
National Research Council Canada, 1200 Montreal Road, Ottawa, Ontario K1A 0R6, Canada
S Supporting Information
*

ABSTRACT: Uncertainty of the result from the method of


standard addition is often underestimated due to neglect of the
covariance between the intercept and the slope. In order to
simplify the data analysis from standard addition experiments,
we propose xy coordinate swapping in conventional linear
regression. Unlike the ratio of the intercept and slope, which is
the result of the traditional method of standard addition, the
result of the inverse standard addition is obtained directly from
the intercept of the swapped calibration line. Consequently,
the uncertainty evaluation becomes markedly simpler. The
method is also applicable to nonlinear curves, such as the
quadratic model, without incurring any additional complexity.

INTRODUCTION
It is common knowledge in analytical chemistry that the
instrumental response can be aected by the sample matrix.
This behavior leads to biased results when the analytical signal
produced from the sample is compared against the calibration
graph prepared from matrix-free standard solutions. To
overcome this problem, the method of standard addition was
introduced in the 1930s.1,2 To date, the method of standard
addition remains an essential tool in quantitation, especially
when isotope dilution cannot be applied. Like any scientic
tool, the method standard addition has witnessed new
developments, most notable of which is the concomitant use
of the internal standard.3
To wit, when more than one addition of a standard is
performed, the method of standard addition seeks the point
with zero response (assuming all signals are corrected for blank
contribution), i.e., the analytical result is as follows:4,5
x 0 = xy = 0
(1)

estimates, uncertainty propagation of their ratio must involve


the covariance between these two estimates, which is not trivial
and is often neglected in practice:69
u 2(x0)
x02

a02

u 2(a1)
a12

cov(a0 , a1) = x u 2(a1)

cov(a0 , a1)
a0a1

(3)

(4)

In general, uncertainty calculations of standard addition results


are rather tedious; and publications on the subject are riddled
with mathematical equations.3 In addition, several recent
studies have focused on a more general problem of nonlinear
calibration curves in analytical chemistry.10,11 The metrological
aspects of nonlinear calibration curves, however, have not
witnessed proper attention, largely due to the complexity that
this problem poses.

THEORY
Variable Transformation for Standard Additions. To
the annoyance of many chemists, traditional standard addition
regression does not yield the analytical result directly. This
problem has been noted before. Notably, Tellinghuisen has
proposed an alternative approach whereby the variable-ofinterest for standard additions, x0, is cast as an explicit variable
of the tting model rather than calculated from the tting
parameters.12,13 The method involves variable transformation

(2)

which means that the result of the classical standard addition is


the ratio of the intercept and the slope of the linear t. A
drawback of this traditional approach is that the uncertainty of
the analytical result cannot be obtained readily from the tting
parameters a0 and a1. Rather, uncertainty propagation must be
performed by combining the uncertainties of a0 and a1.
Moreover, because the intercept and the slope are correlated
Published 2014 by the American Chemical
Society

u 2 (a 0 )

where,

The calculation of x0 is done by tting the best line through


the observed data {x, y}, where x is the amount of the added
analyte, and y is the corresponding analytical response. For a
commonly used straight line t, y = a0 + a1x,
x0 = a0 /a1

Received: January 2, 2014


Accepted: August 6, 2014
Published: August 6, 2014
8563

dx.doi.org/10.1021/ac5014749 | Anal. Chem. 2014, 86, 85638567

Analytical Chemistry

Article

from the classical y = a0 + a1x, with x0 = a0/a1, to y = a0(a1 + x),


which gives a direct estimate of the standard addition result, i.e.,
x0 = a1. Likewise, quadratic tting of y = a0 + a1x + a2x2 can be
done with a modied quadratic function, y = a0(a1 + x) + a2(a1
+ x)2, which yields x0 = a1.5 Albeit simple in its concept, the
variable change from y = a0 + a1x to y = a0(a1 + x) = a0a1 + a0x
leads to a nonlinear tting problem because of the product of
both variables, a0a1, which appears in the model equation. As a
result, no simple algebraic expressions can provide the results
and iterative minimization methods must be employed, such as
the Excel SOLVER macro.
Coordinate Swapping. Consider a typical standard
addition experiment where x is the vector of the masses of
the added standard and y is the vector of the corresponding
instrumental responses. Now, instead of tting y = a0 + a1x, one
performs tting of x = b0 + b1y, which is known as the inverse
calibration.14 Ordinary least-squares t is obtained by
minimizing the sum (a0 + a1xi yi)2 with a0 and a1 as
variables, whereas the inverse t is found by minimizing the
sum (b0 + b1yi xi)2 with b0 and b1 as variables. The advantage
of the seemingly trivial coordinate swapping becomes evident
after calculation of x0,inv = xy=0; x0,inv equals the (negative)
intercept of the inverse regression, i.e.,
x0,inv = b0

all practical applications of analytical chemistry, the numerical


dierence between the results of direct and inverse linear
standard addition plots is irrelevant.
Nonlinear Models. Not all analytical methods provide
linear response with respect to the amount of the analyte.
When nonlinear response is observed, it is common among
chemists to employ empirical polynomials, such as the
quadratic function, in order to represent the observed curvature
while disregarding the physical basis of the nonlinearity. The
diculty of employing polynomial tting in standard addition
models, however, is in large part due to the complexity of
obtaining the analytical results. For example, the analytical
result from a quadratic standard addition t, y = a0 + a1x + a2x2,
is given as follows:16
x0 =

u 2(x0) =

Consequently, the uncertainty of the analyte concentration


x0,inv is the uncertainty of the intercept:

qcov(a0 , a1)
qd cov(a1 , a 2)
d cov(a0 , a 2)
+

2
a 2z
a2 z
2a 23z

(6)

cov(a0 , a1) = u 2(a1)

(8)

x x 4 x2 x3

cov(a0 , a 2) = u 2(a 2)
cov(a1 , a 2) = u 2(a 2)

( x 2)2 N x 4

(9)

( x 2)2 x x 3
( x)2 N x 2

(10)

N x3 x x2
( x)2 N x 2

(11)

In stark contrast to eqs 811 for quadratic standard addition


model y = a0 + a1x + a2x2, the uncertainty calculation in the
inverse approach, x = b0 + b1y + b2y2, remains simple:
u(x0,inv) = u(b0)

(12)

The simplicity of the inverse tting is a particularly attractive


feature given the growing use of nonlinear calibration functions
in analytical chemistry (see Table 1).1618
The inverse of the linear model is also a linear model and the
same holds true for Pade[1,1] model. Unlike these two

This expression establishes practical equivalence between the


uncertainty estimates of direct and inverse standard addition
method results. Second, the dierence between the results of
inverse and direct standard addition methods is as follows:

Table 1. Comparison of the Direct and Inverse Calibration


Approaches for the Standard Addition Method

x0,inv x0 = b1y (1 r 2)/r 2

direct calibration

The combination of these two expressions can be empirically


approximated as follows:
u(x0,inv)

q2u 2(a1)
u 2 (a 0 )
d 2u 2(a 2)
+
+
z
4a 24z
4a 22z

where z = 4a0a2, q = a1 z, d = a1q 2a0a2, and the


regression parameter covariances are calculated as follows:

u(x0,inv)/u(x0) r 2

(7)

a21

No additional uncertainty propagation is necessary, which


obviates the need to evaluate the covariance between the
regression parameters which is necessary when direct regression
is employed (see eqs 3 and 4).
Equivalence between Direct and Inverse Regression.
The ordinary least-squares minimization is not symmetric with
respect to both coordinates. Consequently, many heated
discussions have ensued with regard to the numerical
dierences obtained from direct and inverse regressions.14,15
In the context of analytical chemistry, the following two
expressions are helpful to establish the metrological compatibility between the results of the direct and inverse linear
standard addition models. First, for a suciently large
coecient of determination (r2), commonly, if not universally,
encountered in analytical chemistry,

x0,inv x0

2a 2

and the uncertainty calculation is prohibitively complex:

(5)

u(x0,inv) = u(b0)

a12 4a0a 2

a1

y
y
y
y

1 r2
r2

At approximately r 2 > 0.7, the two models become


metrologically compatible in a sense that the uncertainty of
x0 exceeds the numerical dierence between the direct and the
inverse models, x0,inv x0. Hence, it is fair to conclude that for

=
=
=
=

a0 + a1x
a0 + a1x + a2x2
(a0 + a1x)/(1 + a2x)
(a0 + a1x + a2x2)/(1 + a3x)
inverse calibration
x
x
x
x

8564

=
=
=
=

b0 + b1y
b0 + b1y + b2y2
(b0 + b1y)/(1 + b2y)
(b0 + b1y + b2y2)/(1 + b3y)

result
x0
x0
x0
x0

=
=
=
=

a0/a1
(a1 (a21 4a0a2)1/2)/2a2
a0/a1
(a1 (a21 4a0a2)1/2)/2a2
result
x0,inv
x0,inv
x0,inv
x0,inv

=
=
=
=

b0
b0
b0
b0

dx.doi.org/10.1021/ac5014749 | Anal. Chem. 2014, 86, 85638567

Analytical Chemistry

Article

Table 2. LINEST Fitting of Functions Commonly Encountered in Analytical Chemistrya


type
direct
linear
quadratic
polynomial
Pade[1,1]
Pade[2,1]

Excel directive (in English)b

function
y
y
y
y
y
y

=
=
=
=
=
=

a0x
a0 + a1x
a0 + a1x + a2x2
a0 + a1x + ... + amxm
(a0 + a1x)/(1 + a2x)
(a0 + a1x + a2x2)/(1 + a3x)

=LINEST(y,
=LINEST(y,
=LINEST(y,
=LINEST(y,
=LINEST(y,
=LINEST(y,

x, FALSE, TRUE)
x, TRUE, TRUE)c
x^{0,1,2}, FALSE, TRUE)
x^{0,...,m}, FALSE, TRUE)
x^{0,1,0}*y^{0,0,1}*x^{0,0,1}, FALSE, TRUE)
x^{0,1,2,0}*y^{0,0,0,1}*x^{0,0,0,1}, FALSE, TRUE)

a
For inverse tting x = f(y), all instances of x and y in the LINEST directive have to be interchanged. For example, the inverse linear t is performed
using LINEST(x, y, TRUE, TRUE). bSymbols x and y stand for the data input range. For example, if the values of x and y are in cells A1 to A5 and
B1 to B5, respectively, then the linear t is obtained by entering =LINEST(B1:B5, A1:A5, TRUE, TRUE). Also note that unlike many other
functions which are entered in a singe cell by pressing enter, LINEST is an array function which is entered over several cells (an array) by pressing
ctrl+shift+enter. cAlternatively, one can use =LINEST(y, x^{0,1}, FALSE, TRUE).

functions, however, polynomial models are not invertible. The


mathematical form of the tting model equation, however, is
irrelevant for most applications since nonlinear models are used
empirically in order to describe the observed curvature between
x and y. Consequently, the tting parameters often carry little
physical meaning. In many cases, nonlinear responses can be
made linear by using variable transformation. For example,
when the method of standard addition is used in potentiometry,
one employs linear tting function y = a0 + a1lnx, which can be
inverted to lnx = b0 + b1y.
Regressions with Excel. Calculations using Excel spreadsheets have become commonplace in science.19,20 Many tting
algorithms involve the SOLVER macro which performs
minimization of a user-dened function against the userdened argument. The use of such iterative approaches cannot
be subjected to a seamless copy-and-paste for other data, and
they do not provide the uncertainty estimates for the
parameters. On the contrary, the built-in Excel LINEST
function addresses all of the above shortcomings, and it can
be tailored for both linear and nonlinear model equations.
Table 2 summarizes useful LINEST directives in order to
perform tting on linear or polynomial models, including the
rational Pade functions, which were recently introduced for
nonlinear standard addition tting.21

Table 3. Determination of Bromide Ions in Groundwater


(BCR-611)/Dataa
mA(AA*)/g

mA*(AA*)/g

mfinal/g

AEtBr

AEtI

2.005 99
2.008 17
2.180 35
2.001 74
2.004 35
2.009 60
1.999 61
2.005 36
2.009 01
1.999 22
2.002 76
2.006 66
1.997 26
2.004 13
2.006 93

0.000 00
0.000 00
0.000 00
0.049 49
0.050 05
0.049 89
0.099 24
0.099 49
0.099 91
0.148 14
0.148 72
0.148 83
0.197 03
0.198 03
0.200 21

2.353 82
2.358 67
2.355 67
2.350 62
2.353 49
2.358 09
2.348 73
2.354 30
2.358 00
2.346 72
2.351 28
2.354 87
2.344 29
2.352 16
2.357 14

14 610
14 633
13 422
21 863
22 161
22 155
29 786
30 806
30 427
35 088
37 794
34 243
42 221
46 070
38 640

8153
8176
7695
8190
8339
8199
8357
8652
8432
7983
8531
7731
7909
8682
7286

Mass fraction of Br in primary reference standard solution, wA* =


1925(2)k=1 ng g1.

In our calculations, we use a more elaborate formalism of


standard additions which accounts for the (small) dierences in
the amount of sample in each aliquot, and for the (small)
dierences in the dilution factors. We also employ an internal
standard in order to attain highest precision. More details on
this precision-formalism of the standard additions can be
consulted elsewhere,3,24 and here we give only a brief summary
of the variables which are involved in constructing the standard
addition plots. When no internal standard is used, the following
variables are considered:
mA * (AA ), i
mfinal, i
* w
AEtBr, i and xi =
yi =
A*
mA(AA ), i
mA(AA ), i
(12)
*
*
When endogenous internal standard is used, there is no need to
account for the sample dilution, and the following variables are
considered:
mA * (AA ), i
AEtBr, i
* w
and xi =
yi =
A*
AEtI, i
mA(AA ), i
(13)
*
The incorporation of the mass fraction of the analyte in the
primary standard, wA*, serves to simplify the calculations. Such
practice is acceptable if the uncertainty of the wA* is negligible.
However, when the uncertainty arising from the wA* needs to
be explicitly accounted for in the combined uncertainty budget,
one can remove the wA* from eqs 12 and 13. Then, the mass
fraction of the analyte in the sample is wA = x0wA*, and the

EXPERIMENTAL SECTION
Bromide was analyzed in certied groundwater reference
material by using headspace gas chromatography mass
spectrometry (GCMS) after the aqueous derivatization with
triethyloxonium tetrauoroborate at room temperature.22
Iodide ions which are naturally present in the groundwater
were used as the internal standard for bromide since both of
these anions undergo ethylation in a similar fashion.
Experimental details are summarized in the Supporting
Information (SI) and further details can be found elsewhere.23

RESULTS AND DISCUSSION


To demonstrate the validity of the coordinate swapping in
standard additions, and also to show the usefulness of this
approach when dealing with nonlinear standard addition
models, determination of bromide in groundwater certied
reference material was undertaken using a state-of-the-art
GCMS method.22 All relevant experimental data are given in
Table 3. Symbol mA(AA*) stands for the mass of the sample in
the mixture of the sample (A) and standard (A*), mA*(AA*) is
the mass of the standard in the corresponding mixture, mfinal is
the mass of the AA* mixtures after addition of all reagents, and
AEtBr and AEtI stand for the peak areas of EtBr and EtI.
8565

dx.doi.org/10.1021/ac5014749 | Anal. Chem. 2014, 86, 85638567

Analytical Chemistry

Article

Table 4. Determination of Bromide in BCR-611 with Standard Addition Method Using Various Direct and Inverse Calibration
Methodsa
yi = AEtBr,i/AEtI,i xi = (mA*(AA*),i/mA(AA*),i) wA*
direct calibration
y = a0 + a1x
y = a0 + a1x + a2x2
y = (a0 + a1x)/(1 + a2x)
y = (a0 + a1x + a2x2)/(1 + a3x)
inverse calibration
x = b0 + b1y
x = b0 + b1y + b2y2
x = (b0 + b1y)/(1 + b2y)
x = (b0 + b1y + b2y2)/(1 + b3y)

equations

result, w(Br)/(ng/g)

complexity of x0b

complexity of u(x0)b

eqs
eqs
eqs
eqs

24
7, 811
s1, s2s3
s4, s5s8

96.45(1.14)
93.15(3.12)
93.50(3.25)
94.93(1.17)

1
8
1
8

30
177
216
460

eqs
eqs
eqs
eqs

5,
5,
5,
5,

96.37(1.14)
93.09(3.34)
92.42(3.21)
94.94(1.12)

1
1
1
1

0
0
0
0

6
6
6
6

Mass fraction of bromide in BCR-611 as determined by the state-of-the-art quadruple isotope dilution GCMS, w(Br) = 96.28(42) ng/g.15 All
uncertainties here are combined standard uncertainties (expressed with coverage factor k = 1). bThe number of elementary arithmetic functions
(addition, multiplication, and exponentiation) needed to obtain x0 or u(x0) starting from the results given by the LINEST function (calculated for
this data set with N = 15). Although the operation count depends on variable rearrangement (e.g., three operations in a b + a b vs two
operations in 2 a b), it nevertheless provides a reasonable estimate of the mathematical complexity at-hand.
a

uncertainty is (u(wA)/wA)2 = (u(x0)/x0)2 + (u(wA*)/wA*)2.


This way one can ensure that the uncertainty associated with
the primary standard is properly incorporated in the
uncertainty of the result. For the sake of simplicity, this was
not done in this work, also because the uncertainty in the
primary reference standard, u(wA*), is insignicant in
comparison to the uncertainty of the t for data presented in
this work.
Using data from Table 3, standard addition calculations were
performed with four modelslinear, quadratic, Pade[1,1], and
Pade[2,1] whose equations are given in Table 1. Least squares
tting was performed using LINEST function as detailed in
Table 2. For direct calibration functions, the uncertainty of x0
was obtained either from eqs 3 and 4 (linear model) or eqs
811 (quadratic model). Expressions for the Pade[1,1] and
Pade[2,1] parameter covariances are given in the SI. For all
inverse calibration functions, the uncertainty of x0 was obtained
directly from LINEST as per eq 6. The results are given in
Table 4.
In general, one can observe that the numerical results of
inverse calibration approach dier negligibly from the
corresponding results obtained from the direct tting of
analytical signal against the mass of the added analyte. In
addition, all results agree with the value established from the
state-of-the-art isotope dilution measurements.23 The calculation of the x0 values and their uncertainty estimates, however,
is signicantly simpler from the inverse regressions. Table 4
summarizes the minimum number of elementary arithmetic
functions needed to obtain x0 or u(x0) starting from the results
given by the LINEST function. In all cases, the inverse
approach provides answers without any additional calculations
whereas direct tting of y over x requires considerable amount
of additional calculations in order to obtain x0 and u(x0). In the
well-cited publication on the standard addition method, Bader
reects that once mathematical complexity sets in, it is easy to
make mistakes.7 In this vein, it is hoped that the marked
simplicity of the inverse approach will result in a wider use of
this technique and spur the much-needed evaluation of the
metrological performance of nonlinear standard addition
models. With the increased nonlinearity in the standard
addition curves, one should expect more discrepancies between
the results of the various empirical nonlinear models and also
between the direct and inverse calibration approaches. In such

situations, the analysts will be reminded to re-evaluate the


chosen method of analysis or the adequacy of the measurement
model equation.

CONCLUSIONS
Although, strictly speaking, the classical and inverse regressions
are not mathematically equivalent procedures, the distinction
becomes relevant only for data sets that are very noisy or
extremely nonlinear.25 Since this is not a norm in analytical
chemistry, the inverse tting oers signicant advantages to
chemists. Given that the coordinate swapping in standard
addition experiments simplies the data analysis signicantly
over the traditional tting approach, the use of inverse
regression should be embraced in the spirit of Occams razor.

ASSOCIATED CONTENT

S Supporting Information
*

Experimental details; a typical GC/MS chromatogram of


bromide and iodide sampled from the headspace of the
groundwater reference material BCR-611 with w(Br) = 93 4
ng/g and w(I) = 9.0 ng/g (Figure S1); theoretical information
and results. This material is available free of charge via the
Internet at http://pubs.acs.org.

AUTHOR INFORMATION

Corresponding Author

*E-mail: juris.meija@nrc.ca.
Notes

The authors declare no competing nancial interest.

REFERENCES

(1) Hohn, H. Chemische Analysen mit dem Polarographen; Springer:


Berlin, 1937.
(2) Kelly, W. R.; Pratt, K. W.; Guthrie, W. F.; Martin, K. R. Anal.
Bioanal. Chem. 2011, 400, 18051812.
(3) Hauswaldt, A.-L.; Rienitz, O.; Jahrling, R.; Fischer, N.; Schiel, D.;
Labarraque, G.; Magnusson, B. Accred. Qual. Assur. 2012, 17, 129
138.
(4) Harris, D. C. Quantitative Chemical Analysis, 4th ed.; Freeman:
New York, 1995.
(5) Tellinghuisen, J. Analyst 2005, 130, 370378.
(6) Bruce, G. R.; Gill, P. S. J. Chem. Educ. 1999, 76, 805807.
(7) Bader, M. J. Chem. Educ. 1980, 57, 703706.
8566

dx.doi.org/10.1021/ac5014749 | Anal. Chem. 2014, 86, 85638567

Analytical Chemistry

Article

(8) Serapinas, P.; Labarraque, G.; Charlet, P.; Ezirinskis, Z.; Juzikiene,
V. J. Anal. Atom. Spectrom. 2010, 25, 624630.
(9) Christopher, S. J.; Day, R. D.; Bryan, C. E.; Turk, G. C. J. Anal.
Atom. Spectrom. 2005, 20, 10351043.
(10) Rozet, E.; Ziemons, E.; Marini, R. D.; Hubert, Ph. Anal. Chem.
2013, 85, 63276335.
(11) Peters, F. T.; Maurer, H. H. Anal. Chem. 2007, 79, 49674976.
(12) Tellinghuisen, J. Analyst 2000, 125, 10451048.
(13) Salter, C.; de Levie, R. J. Chem. Educ. 2002, 79, 268271.
(14) Tellinghuisen, J. Fres. J. Anal. Chem. 2000, 368, 585588.
(15) Shukla, G. K. Technometrics 1972, 14, 547553.
(16) Koscielniak, P. Chemometr. Intell. Lab. 1999, 47, 275287.
(17) Kirkup, L.; Mullholland, M. J. Chromatogr. A 2004, 1029, 111.
(18) Gorazda, K.; Michalowska-Kaczmarczyk, A. M.; Asuero, A. G.;
Michalowski, T. Talanta 2013, 116, 927930.
(19) Harris, D. C. J. Chem. Educ. 1998, 75, 119121.
(20) Kragten, J. Analyst 1994, 119, 21612165.
(21) Bartkovjak, J.; Karovicova, M. Meas. Sci. Rev. 2001, 1, 6365.
(22) DUlivo, A.; Pagliano, E.; Onor, M.; Pitzalis, E.; Zamboni, R.
Anal. Chem. 2009, 81, 63996406.
(23) Pagliano, E.; Mester, Z.; Meija, J. Anal. Bioanal. Chem. 2013,
405, 28792887.
(24) Rienitz, O.; Hauswaldt, A.-L.; Jahrling, R. Anal. Bioanal. Chem.
2012, 404, 21172118.
(25) Besalu, E. Talanta 2013, 116, 4549.

8567

dx.doi.org/10.1021/ac5014749 | Anal. Chem. 2014, 86, 85638567