Analytica Chimica Acta 762 (2013) 113

Contents lists available at SciVerse ScienceDirect

Analytica Chimica Acta

journal homepage: www.elsevier.com/locate/aca

Tutorial

A tutorial on the application of ion-selective electrode potentiometry:

An analytical method with unique qualities, unexplored opportunities
and potential pitfalls; Tutorial
Erno Lindner , Bradford D. Pendley
Department of Biomedical Engineering, The University of Memphis, Memphis, TN 38152, United States

h i g h l i g h t s

g r a p h i c a l

a b s t r a c t

 Electrochemical cells for potentiometric measurement.

 Characterization of potentiometric
electrodes.
 Measurement of ion activities and
concentrations with potentiometric
electrodes.
 Analysis of real samples: The role of
the selectivity coefcient.
 Methods to evaluate the agreement
between two methods.

a r t i c l e

i n f o

Article history:
Received 13 September 2012
Received in revised form 8 November 2012
Accepted 11 November 2012
Available online 23 November 2012
Keywords:
Ion-selective electrodes
Application of potentiometry
Sources of errors
Method comparison

a b s t r a c t
Ion-selective potentiometry enjoys practical utility as a simple analytical technique to measure ionic
constituents in complex samples. Advances in the eld have improved the selectivity and decreased the
detection limit of ion-selective electrodes (ISEs) by orders of magnitude such that trace analysis in micro
and nanomolar concentrations is now possible with potentiometric sensors. This tutorial reviews the
fundamental principles of ion-selective potentiometry, describes the practical considerations involved in
the use of these sensors to measure real samples, and discusses the statistical evaluation of experimental
results compared with alternative analytical techniques.
2012 Elsevier B.V. All rights reserved.

Contents
1.
2.

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Ion-selective potentiometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.
Electrochemical cells for potentiometric measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.
Characterization of potentiometric electrodes and measurement of ion activities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.
Measurement of ion concentrations with potentiometric electrodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.
Analysis of real samples: the role of the selectivity coefcient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.
Rate of response, drift of the measured voltage, reproducibility of the measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Corresponding author. Tel.: +1 901 678 5641; fax: +1 901 678 5281.
E-mail address: elindner@memphis.edu (E. Lindner).
0003-2670/$ see front matter 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.aca.2012.11.022

2
2
2
5
8
8
10

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

3.

4.

Method comparison studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.1.
Statistical methods to evaluate the agreement between two methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.
A typical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Acknowledgement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Erno Lindner, PhD. Erno Lindner is a R. Eugene Smith

Professor of biomedical engineering at the University Memphis. He graduated in 1971 and received his
Ph.D. from Technical University of Budapest in 1985.
During his scientic carrier he worked with the leading scientists of the eld of ion-selective electrodes
including E. Pungor and K. Toth (Budapest), W. Simon
and E. Pretsch (Zurich), and R.P. Buck (Chapel Hill), etc.
Dr. Lindners published 165 papers, which were cited
more than 4000 times. In 2010 he received the Willard
R. Sparks Eminent Faculty award the highest distinction given to a faculty member by the University of
Memphis.

1. Introduction
During the time we wrote this tutorial on ion-selective electrodes (ISEs), we received a letter from a colleague in which we
were asked about commercial instruments that utilize the recent
discoveries in the eld of ISEs for monitoring sub-nanomolar lead
ion concentrations in environmental samples. He approached us
because he came across our paper in which we reported picomolar
detection limits with lead ion-selective membranes [1]. We had to
inform him that, as far as we know, no such instrument exists. We
felt somewhat uncomfortable and embarrassed to give this information because the last 15 years of potentiometry was all about
the new wave of ion-selective electrodes [2] with spectacular
selectivities and previously unimaginable detection limits [28].
However, the focus of these papers was on the understanding of
the underlining chemical principles of the ISE responses and on
the elimination of experimental biases inherently related to their
response mechanism. Consequently, the majority of these papers,
including ours [1], were feasibility studies and only a very few
were aimed for documenting the unique benets of potentiometric ion-analysis through practical examples [9] with adequate error
analysis, statistical support or proper comparison to an existing
method. This is in sharp contrast to what occurred following the
discovery of the rst non-glass based ISEs in the late sixties [1013],
when the need for better methods to solve important practical
problems (e.g., the measurement of ion concentrations in whole
blood or uoride in drinking water) was the driving force of the
developments. At that time, the new ISEs were utilized to satisfy
real needs that would be very difcult or impossible to fulll with
the existing analytical methods, e.g., analysis of the blood samples
of astronauts during space ights. Consequently, simultaneous to
their appearance, the new ISE-based methods were analyzed in
detail for their precision and accuracy [14].
Compared to the meticulous examination of the analytical
results with the new ISEs in the early years, it appears that during the last 15 years of development, when the selectivities and
the detection limits of potentiometric sensors have been improved
by many orders of magnitudes, we have forgotten that to have
broad acceptance, new discoveries must satisfy real demands.
Although it is very exciting to measure ionic concentrations at the

10
10
11
13
13
13

Bradford D. Pendley, Ph.D., M.D. Dr. Pendley graduated from Cornell University with a Ph.D. in
analytical chemistry under the direction of Dr. Hector
Abruna. His research has involved ultramicroelectrode voltammetry and potentiometric sensors. He
began his academic career at Rhodes College but, after
becoming engaged in the applied eld of biomedical engineering, he resigned his tenured, endowed
position and attended medical school, completing
the M.D. degree and residency at The University of
Tennessee Health Science Center. He is currently an
afliate professor of biomedical engineering at The
University of Memphis as well as a practicing, Board
Certied, Internal Medicine Physician at Primary Care Specialists.

parts-per-trillion level (1010 M), the need exists to demonstrate

that these spectacular detection limits can be achieved in real samples and that the new method has obvious advantages over the old
ones, a task typically done in method comparison studies. At the
end of this tutorial we give a quick guide about the approaches for
evaluating the performance characteristics of a new method and
the appropriate statistical tools for assessing agreement between a
new and an existing method that measures the same quantity.
The feasibility of trace level measurements with potentiometric
ISEs has been demonstrated and the results have been summarized
in excellent reviews and book chapters [4,68,1519]. However
most of these reviews were written for experts without detailed
information on the methodology or validation of the novel ISEs.
Consequently, to nd practical information on ISE-based methods, one has to consult books [2025] and IUPAC documents [26]
from up to forty years ago. Therefore, in this tutorial, the emphasis will be on the sources of potential errors [21] and challenges
of direct potentiometry and ways to address them. We hope that,
by discussing these difculties, those who are new to the eld and
only consider using potentiometry as an analytical method can
understand how to translate these improvements in technology for
solving real problems. If we can reach this goal we will not hesitate how to answer the question of Karl Camman Ion-selective
potentiometry: an analysts dream or nightmare?[27].
2. Ion-selective potentiometry
2.1. Electrochemical cells for potentiometric measurement
The history of ion-selective potentiometry can be traced back
to the studies of Walter Nernst on the acidity of aqueous solutions
with a hydrogen gas sensor [28]. However, the discovery of the
pH sensitive glasses by Max Cremer [29] and recording acid base
titration curves with glass electrodes by Fritz Haber [30] about a
century ago are the real landmarks of a new era in potentiometry.
The rst papers describing the use of glass electrodes in combination with commercial pH meters were published about 20 years
later. The next revolutionary period in potentiometry started in
the sixties [1013] when new ISEs were reported at a rate that was
best characterized with titles like the Electrode of the Month [31].

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

The commercial realization of the new ISEs was also very fast. Only
four years after the rst publication describing the responses of the
precipitate-based iodide electrode it became commercially available in Hungary [32]. The rst calcium, uoride and potassium ISEs
became commercially available only a few years later. Reading the
history of ion-selective electrodes from different authors may give
the reader a valuable insight about this exciting period and show
how the different assumptions related to the importance of phase
boundary and membrane transport processes evolved in a unied
theory [3135].
The setup of a potentiometric measurement (a galvanic cell)
consists of an indicator and a reference electrode that are immersed
into the sample solution and connected to the two terminals
of a voltmeter as shown in Fig. 1. In this article we will focus
on the group of indicator electrodes that are classied as ionselective electrodes (ISEs) and will use the terms indicator and
ion-selective electrode interchangeably. In the most commonly
used ion-selective electrode arrangement, a unique membrane
made from (i) glass (glass electrodes), (ii) water insoluble precipitate (precipitate-based electrodes), or (iii) polymeric lm loaded
with water immiscible liquid, organic ion-exchanger and complexing agent (liquid membrane electrodes) is attached at the end of
a cylindrical electrode body as shown in Fig. 1. This membrane
separates the sample solution and the inner lling solution of the
ISE. The electrical contact of the ISE is made through a second
reference electrode immersed in the inner lling solution. Alternatively, the membrane can be sandwiched between the sample
solution and a solid contact (solid contact ion-selective electrodes)
[36]. From the large variety of solid contacts, inherently conductive

polymers (e.g., polypyrrole, poly(3-octylthiophene), and poly(3,4ethylenedioxythiophene)) emerged as materials of choice [17,37].
When an indicator electrode is immersed into a solution a
potential difference develops at its phase boundary. In a recent
article, Bhlmann [18] used the example of the extraction of tetrabutylammonium nitrate (TBA+ NO3 ) from an organic solvent to an
aqueous solution to explain the development of the phase boundary potentials. Extraction refers to the distribution of a solute from
one phase to another. In the most common case the two phases
are an organic solvent and an aqueous solution and the distribution between the two phases is characterized by the partition
coefcient:
Korg/aq =

aorg
corg

aaq
caq

(1)

where Korg/aq is the partition coefcient between the organic and

aqueous phases, a is the activity and c is the concentration of
the solute in the two phases as they are labeled with their corresponding subscripts. Using TBA+ NO3 as a solute in an extraction
experiment it is expected to partition between the two phases
according to its solvation energies in the two phases. Although
the concentration of TBA+ NO3 may be vastly different in the two
phases, the concentrations of TBA+ cation and NO3 anion in each
phase are equal to each other due to the electroneutrality requirement. However, at the utmost vicinity of the phase boundary the big
difference in the lipophilicity of the hydrophobic TBA+ cation and
hydrophilic NO3 anion generates a charge separation across the
interface. There will be a small excess of TBA+ cation on the organic
solvent side of the phase boundary and a small excess of and NO3

Fig. 1. Top: potentiometric cell assembly with a conventional, liquid inner contact ion-selective membrane electrode as indicator electrode and a double junction reference
electrode. Bottom: individual phase boundary potentials in the potentiometric cell assembly above: metalmetal (1 and 8 ), metalsalt (2 and 7 ), saltelectrolyte (3 and
6 ), electrolyteelectrolyte (4 and ED ), and electrolytesolid or liquid membrane (EPB and 5 ); Ecell is the sum of the phase boundary potentials that can be measured between
the ion-selective and reference electrode terminals of the cell, EPB and ED are the two potential terms that are inuenced by the sample composition.

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

the different species to the overall charging of the phase boundary

(QJ /QI = i0,J /i0,I ) determines the selectivity of the electrode.
The phase boundary potential is provided by the Nernst equation (Eq. (2)); it is a logarithmic function of the ionic activity in the
solution. The phase boundary potential of an ideal ISE is determined
by a single ion activity in a complex sample. The term ion-selective
electrode points to this unique property of this class of indicator
electrodes.
EI = EI0 +

Fig. 2. Currenttime proles at an ion-selective membrane (organic phase) aqueous

solution interface upon immersing the ISE into the sample or following a step concentration change in the sample composition. iorgaq and iorgaq indicate directed
ow of charge carriers across the interface, io,I and io,J are the exchange current densities for the primary ion I and the interfering ion J, QI and QJ are the contributions of
ion I and J to the charging of the phase boundary. Inset: a model of the ion distribution between a liquid membrane phase and a KCl solution. To facilitate the selective
transfer of K+ ions the liquid membrane is loaded with negatively charged cation
exchange sites (R ) and a selective complexing agent for K+ ().
From Cammann, K. Working with Ion-Selective Electrodes, Springer-Verlag, 1979
[20], Cammann K. A mixed potential ion-selective electrode theory, in Ion-Selective
Electrodes, Akademiai Kiado, 1978 [38] and Bhlmann P, Chen LD. Ion-Selective Electrodes with Ionophore-Doped Sensing Membranes, in Supramolecular Chemistry:
From Molecules to Nanomaterials, John Wiley & Sons, 2012 [18] with permission.

anion on the aqueous solution side of the phase boundary. This

interfacial charge separation is the origin of the phase boundary
potential.
To understand the process for the establishment of a charged
interface and the build-up of the phase boundary potential, Cammann [20,38] provided a very plausible electrokinetic model
(Fig. 2). Upon immersing an electrode into a solution, or in
Bhlmanns example upon merging two immiscible solutions,
charged species are crossing the phase boundary which contributes
to an increasing potential difference at the interface. In the case of
ion-selective electrodes, the transfer of one ionic species (e.g., K+ ) is
highly favored over all the others. In the example of Fig. 2, the selective transfer of K+ ions is achieved through the incorporation of a
selective complexing agent (e.g., valinomycin ()), and negatively
charged cation exchange sites, e.g., tetrakis(p-chlorophenylborate
(R )) into the organic phase (membrane). The potential buildup
related to the transfer of the preferred ion (K+ ) in one direction
facilitates its transfer to the opposite direction. The transport of
charge at the phase boundary is best represented by the current
density iorgaq and iorgaq .
At equilibrium, the current densities representing the charge
transfer across the interface in the two directions balance each
other. This means, that at equilibrium there is no net current ow
across the phase boundary, i.e., io,I = iorgaq = iorgaq . io,I is termed
as the exchange current density for the preferred ion I. The larger
the exchange current density the less sensitive is the electrode
potential to the presence of other ionic species (interfering ions)
in the sample or to external current (polarization). In Fig. 2, it is
assumed that the interfering ion J can also cross the phase boundary. The process to reach the equilibrium for the interfering ion J
is exactly the same as it is for the ion I, i.e., one can also dene
an exchange current density for the interfering ion J (io,J ). The ion
which contributes the most charge to the phase boundary potential
is called the potential determining ion. The relative contribution of

RT
2.303RT
ln aI (aq) = EI0 +
log aI (aq)
zI F
zI F

(2)

where EI is the phase boundary potential of the indicator electrode,

EI0 is a constant, aI is the activity of ion I in the sample solution,
zI is the charge of the ion I, R is the gas constant, T is the absolute
temperature, and F is the Faraday number.
According to Eq. (2), a 10 fold change in the activity of the ion I
with the charge zI results in 59.16 mV/zI change in the phase boundary potential at 25 C. For monovalent cations and anions (zI = 1)
1 mV error in the measurement of the indicator electrode potential introduces 4% error in the calculated single ion activity. For
di- and trivalent cations and anions this error is 8% or 12%, respectively. This relatively large error is a consequence of the logarithmic
response of ISEs and is one of the inherent disadvantages of potentiometry.
The reference electrode in the electrochemical cell is needed to
measure the indicator electrode potential. The voltmeter in Fig. 1
shows the potential difference between the indicator and the reference electrode, i.e., the cell voltage.
Ecell = EI Eref

(3)

In order to determine the indicator electrodes potential from

the cell voltage, the potential of the reference electrode is assumed
to be constant and independent of the sample composition. However, since the physical and chemical phenomenon on the phase
boundary of the reference electrode is the same as on the indicator electrode, i.e., the requirement for a reference electrode whose
potential is independent from the composition of the sample is not
trivial. If the potential of the reference electrode changes with the
sample composition, it introduces an error in the determination of
the indicator electrode potential and consequently in the ion activity that is calculated from the indicator electrode potential. Indeed,
the minimization of the errors related to unaccounted changes in
the reference electrode potential is one of the most challenging and
most neglected tasks of potentiometry.
To keep the reference electrode potential constant, one can set
the activity of the potential determining ion for the reference electrode constant in the samples through the addition of an inert
electrolyte in excess to each sample. A trivial example would be
when the measurements have to be performed in buffered samples of a given pH value. In such experiments a pH sensitive glass
electrode could be used as reference electrode. When applying this
approach however, one has to realize that by keeping the solution
pH within 0.01 pH unit (close to attainable precision with a high
quality combination glass electrode) corresponds to 0.6 mV error
in the potential measurement or 2.4% error in the ion activity of
a monovalent anion or cation. The error is proportionally larger for
multivalent cations and anions.
If this approach is not feasible, one can minimize the inuence of
the sample composition on the potential of the reference electrode
by separating the reference electrode from the sample solution
with the help of a salt bridge [39] as shown in Fig. 1. However, the
utilization of a salt bridge between the reference electrode and the
sample solution introduces a new potential source into the electrochemical cell that arises at the interface of the two solutions

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

with different compositions, namely the liquid junction or diffusion

potential (ED ). Consequently Eq. (3) has to be modied as:
Ecell = EI (Eref ED )

(4)

Although it is commonly assumed that ED is negligible (equal

to zero) or independent from the sample solution composition,
unaccounted changes in ED is one of the major sources of errors
in potentiometric analysis. To calculate the concentration dependence of the diffusion potential at the salt bridge electrolyte sample
solution interface different model assumptions were used [24] and
the different models provide slightly different ED values. In addition, experience shows that the liquid junction potential varies with
the construction of the merging zone between the salt bridge and
the sample. To estimate the diffusion potential, most commonly
the Henderson approximation is used:
ED =

 + +

)
|z |u (cm cm
|z |u (c c )
 m2 m
 2x x x x
+

zm um (cm cm ) +

zm um cm +

zx ux cx

By considering the electrochemical cell of Fig. 1 in more

detail, it is obvious that the cell voltage measured between the
indicator and the reference electrode terminals is the sum of several phase boundary potentials. These individual phase boundary
potentials, e.g., metalmetal (1 and 8 ), metalsalt (2 and 7 ),
saltelectrolyte (3 and 6 ), electrolyteelectrolyte (4 and ED ), and
electrolytesolid or liquid (EPB and 5 ) are shown schematically in
the lower panel of Fig. 1. In this panel, the potential difference which
can be measured between the connecting wires of the indicator and
reference electrode (Ecell ) as well as the two potential terms that
are inuenced by the sample composition (EPB and ED ) are printed
with capital letters.
2.2. Characterization of potentiometric electrodes and
measurement of ion activities
As it is shown in Eq. (2), the phase boundary potential of a potentiometric electrode is a function of the activity of the ion.

zx ux (cx cx )

 2 +  2
zm um cm +
z u c
RT
 x2 x x
ln  2 +

(5)

where
represent the sum over all charged species; zm and zx is
the electric charge in units of the proton charge for cations (m) and

anions (x); u+
m and ux are the electrochemical mobilities for cations
(m) and anions (x); c and c are the concentrations of cations (m)
and anions (x) in the sample and the salt bridge electrolyte solutions.
As seen from Eq. (5), the diffusion potential cancels if there is
no concentration difference between the two sides of the liquid
c = c c = 0) or when the ionic mobilities of the
junction (cm
m
x
x

cations (u+
m ) and anions (ux ) are equal with each other. The diffusion potential becomes practically constant, and dominated by
the composition of the salt bridge electrolyte, if its concentration
is much larger than that of the sample solution. Consequently the
best practical approach is to use highly concentrated salt bridge

electrolyte in which u+
equitransm ux or more precisely
 + to+ apply



ferent salt bridge electrolyte in which
|zm |um cm
|zx |u
=
x cx .
The most commonly used bridge electrolyte is a concentrated KCl
solution (e.g., saturated, or 3.5 M) with uK+ uCl . Unfortunately,
this traditional salt bridge electrolyte cannot always be applied or
is not always ideal for a specic task [39]. For example, upon the
direct potentiometric measurement of K+ or Cl ion activities in
dilute solutions the utilization of a reference electrode with a concentrated KCl solution as salt bridge is not recommended because
even the smallest leak of KCl from the salt bridge into the sample
may introduce a large systematic bias in the measurement. Since
the best potential reproducibility was reported for reference electrodes with free-owing, free-diffusion liquid junction [40], such
leaks (between 0.005 and 0.1 mL/24 h) are common with commercially available reference electrodes. Compared to other commonly
used reference electrodes with constrained liquid junctions (e.g.,
ceramic plug, ground-glass sleeve, etc.) the free owing liquid junction is also superior with respect of the time needed to reach a
steady state, time-independent reading.
The systematic error related to the contamination of the sample with ions from the salt bridge electrolyte can be minimized
through the addition of a second salt bridge (double junction reference electrode) as shown in Fig. 1. However, the errors related
to the concentration and time dependence of ED cannot be fully
eliminated, and its inuence on the quality of the analytical results
must always be considered as we will show later. KCl retained its
unique importance as a prime salt bridge electrolyte because of the
lack of better alternatives. The most promising recent approaches
apply slightly hydrophobic ionic liquids instead of KCl as salt bridge
electrolyte [39].

aI = I cI

(6)

where aI is the activity of ion I,  I is the individual activity coefcient and cI is the concentration of ion I in the standard solution.
Since the individual ion activity coefcients cannot be assessed on
exact thermodynamic bases, they are calculated from the measured
mean activity coefcients ( ) using different assumptions [24].
The different assumptions provide slightly different  I values.
During the calibration process the ion-selective and reference
electrodes are placed in standard solutions and the cell voltage
is measured. From the corresponding Ecell log aI data pairs one
can determine the parameters of an empirical function by linear
regression:
0
Ecell = Ecell
+ S log aI

(7)

0
is the potential difference between the indicator and
where Ecell
reference electrode in a solution in which aI = 1 and S is the experimentally determined slope of the Ecell vs. log aI linear function. Due
to uncertainties in the ion activity of the standards and in the diffusion potential between the indicator and the reference electrode,
0 and S is limited. In other
the accuracy of the determination of Ecell
words, inaccuracies in the assigned activity values of the standards
and in the calculated diffusion potentials, which are used for correcting the measured cell voltage, may be the source for deviations
from the Nernstian response. However, once the calibration curve
has been set, it can be used to measure the free ion activities even
in the most complex matrices, provided that the indicator electrode potential is solely controlled by the activity of ion I. Changes
in the measured cell voltage (Ecell ) in the presence of strong electrolytes, complexing agents or reagents that form precipitates with
the measured ion only reect changes in the free ion activities. The
only sources of error in the determination of the ionic activities
in unknown samples are related to the inaccuracies in  I and to
changes in the diffusion potential (ED ) between the standards and
the samples.
To demonstrate the magnitude of error that may be introduced
in the calibration of potentiometric electrodes, Table 1 shows single ion activities for CaCl2 solutions that were calculated with the
DebyeHckel equation (Eq. (8)) and diffusion potential values calculated with the Henderson equation (Eq. (5)) using different salt
bridge electrolytes in combination with potential data calculated
with Eqs. (2) and (4).

A |z+ z | I
(8)
log  =
+ CI
1+Ba I

where  is the mean activity coefcient, A = 0.509, B = 0.328

in water at 25 C, z+ and z are charges of the cation and anion of

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

Table 1
Activity coefcients (Ca2+ ) in CaCl2 solutions (calculated with Eqs. (8)(10)a ), diffusion potentials (ED ) at the interface of CaCl2 solutions and different salt bridge (SB)
electrolytes (calculated with Eq. (5)b ), and the expected cell voltage values (calculated with Eq. (4)) by using different salt bridge electrolytes.
1
c (M)
CaCl2

2
Ca2+

3
EI
Eq. (2)
E0 = 0

4
ED
KCl 3.5 M

5
ED
KCl 0.1 M

6
ED
NaCl 0.1 M

7
Ecell
SB-KCl (3.5 M)
Eq. (4), Eref = 0

8
Ecell
SB-KCl (0.1 M)
Eq. (4), Eref = 0

9
Ecell
SB-NaCl (0.1 M)
Eq. (4), Eref = 0

101
102
103
104
105

0.273
0.534
0.791
0.924
0.975

46.3
67.2
91.8
119.3
148.0

0.8
0.7
1.3
1.7
2.1

11.3
2.9
0.1
1.0
1.5

9.0
6.2
18.9
30.7
42.2

45.4
67.9
93.0
121.1
150.1

34.9
64.3
91.8
120.3
149.5

37.2
73.4
110.64
150.0
190.2

a
b

The parameters of Eq. (8) for Ca2+ ions were a = 5.00 and C = 0.04.
The ionic mobilities used for the calculations of ED in [cm2 s1 J1 mol] units were: uCa2+ 109 = 3.22; uK+ 109 = 8.00; uNa+ 109 = 5.47; uCl 109 = 8.11.

the electrolyte in consideration, respectively, a, and C are constants

determined by tting Eq. (8) to measured data [40], and I is the ionic
strength:
I=

1
cn zn2
2

(9)

where cn and zn are the concentration and charge of the nth ion in
the sample, respectively.
To calculate the individual ion activity coefcients from the
mean activity coefcient the convention by DebyeHckel was
used:

 
 
 z+ 
 z 
log  and log  =   log 

z
z

log + = 

(10)

From the data shown in Table 1 one must conclude that the
assumptions related to an ideal solution (Ca2+ = 1, i.e., aCa2+ =
cCa2+ ) and negligible or constant diffusion potential (ED = 0 or
ED = const) are not true. Consequently, using these assumptions
may introduce large errors in the determination of the response
function of ISEs or in the calculated ion activities of unknown samples (Examples are shown in Table 2).
In Table 2 we provide the expressions for the calibration curves
which were determined by linear regression using the data of
Table 1 and applying different assumptions (e.g., ED 0 or ED is
equal with the diffusion potential in the standard solution). In the
last column of Table 2 we provide an estimate of the errors in
the determination of calcium ion activity of 103 using the equations in Table 2 for the calculations (the common practice of direct
potentiometry).
In Fig. 3 we show examples for the consequences of the assumption of ED = 0 and  I = 1. Fig. 3a shows two calibration curves for
a calcium ion-selective electrode that one may expect by using
3.5 M KCl or 0.1 M NaCl as salt bridge electrolyte in the reference
electrode. The parameters of the Ecell log aCa calibration curves
determined by linear regression are provided in Table 2. The

regression analysis for both data sets provided excellent correlation coefcients (rKCl = 0.9999, rNaCl = 0.9992) but the response
slopes deviate from the Nernstian slope (SKCl = 30.3 mV/log aK ,
SNaCl = 42.7 mV/log aK ). However, by correcting the Ecell values with
the calculated diffusion potentials the slope of the line tted to the
(Ecell ED ) vs. log aCa data pairs becomes Nernstian. In our example, this perfect agreement is a consequence that ideal Nernstian
response data were modied and corrected with the same diffusion potential data (ED was calculated by Eq. (5)). Although one
cannot expect such perfect agreement in real experiments, the
example was aimed to show the importance of (i) selecting adequate salt bridge electrolyte; (ii) correcting the measured potential
data with the calculated ED ; (iii) recognizing that an excellent correlation coefcient does not necessarily mean that the results are
meaningful.
In Fig. 3b, the potential data in column 3 of Table 1 are plotted
as a function of the logarithm of the calcium ion activities (open
rectangles) and the logarithm of the concentrations (lled circles).
The separation between the two curves at higher concentrations is a
consequence that the solutions at high concentrations deviate from
the ideal behavior which is expressed in the individual ion activity
coefcients. Plotting the data as a function of the logarithm of the
calcium ion concentrations is nonlinear. However, recognizing this
nonlinear response in real experiments is not always trivial.
In Fig. 3c, the EI vs. log cCa data points from Fig. 3b are
plotted together with a line tted to the data points by linear
regression. The correlation coefcient of the tted line is close
to its limiting value (r = 0.998) which suggests that the electrode
potential changes linearly with the logarithm of the calcium ion
concentration and the slope of the tted line is close to the Nernstian slope (S = 25.6 mV/log aCa ). However, the regression analysis
0 = 17.9 mV) of the tted
beyond the slope and the intercept (Ecell
line also provides the variability of the data points around the tted line (RMSD = 2.9 mV). The acronym RMSD stands for residual
mean standard deviation which in statistical texts is also termed

Table 2
Parameters of calibration curves determined by linear regression using the data of Table 1 and the expected errors in the determination of a calcium ion activity of 103 .
Assumption

Data from Table 1 for LRa (column # in

Table 1 vs. column # in Table 2)

Equation of the
tted line

RMSDc (mV)

EI0 = 0, ED = 0d
ED 0 if SBb is 3.5 M KCl
ED 0 if SBb is 0.1 M KCl
ED 0 if SBb is 0.1 M NaCl
ED 0 if SBb is 0.1 M KCl and
aCa = cCa , i.e.,  Ca = 1

EI vs. log aCa (3 vs. 2 1)e

Ecell vs. log aCa (7 vs. 2 1)e
Ecell vs. log aCa (8 vs. 2 1)e
Ecell vs. log aCa (9 vs. 2 1)e
Ecell vs. log cCa (8 vs. 1)

Ecell = 0.0 + 29.6 log aCa

Ecell = 1.4 + 30.3 log aCa
Ecell = 12.7 + 32.8 log aCa
Ecell = 22.8 + 42.7 log aCa
Ecell = 6.6 + 28.5 log cCa

0
0.5
3.3
1.1
0.5

a
b
c
d
e
f

Error in aCa f (%)

0
+5
23
+143
+31

Linear regression.
Salt bridge.
Residual mean standard deviation.
The cell voltage data were generated by the equation EI = 0.0 + 29.6 log aCa with the assumption of ED = 0.
aCa it is calculated by multiplying data in columns #1 and #2 (aCa = cCa  Ca ).
It is assumed that the measured Ecell =88.8 mV (Ecell = 0.0 + 29.6 log aCa ) and the unknown activity is calculated from the corresponding calibration curve.

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

Fig. 3. (a) Theoretically expected calibration curves for a Ca2+ ion-selective electrode using 3.5 M KCl or 0.1 M NaCl salt bridge electrolyte in the reference electrode; (b)
theoretical dependence of the Ca2+ ion-selective electrode potential on the logarithm of the sample solution activity () or concentration (); (c) a calibration plot constructed
from the E vs. log c data points in (b) with a line tted to the data points by least square regression; (d) potential differences between the data points and the best t line in
(c).

as the standard error of the estimate. The RMSD value in linear

regression provides guidance about the average of expected errors
in the dependent variable. The 2.9 mV RMSD value in this example
corresponds to approximately 23% error in the calcium ion concentration (1 mV corresponds to 8% error). This large error is
the result of forcing a linear regression line through nonlinear data.
Consequently, characterizing the quality of an apparently linear
calibration curve only by the correlation coefcient of the tted line
can be misleading. By plotting the potential differences between
the measured points and the tted line (Residuals) as a function of
log c (Fig. 3d) it becomes immediately apparent that the data points
are not randomly scattered around the tted line and that the average deviation from the regression line is far too large for accurate
measurements. Due to the logarithmic response of ISEs relatively
large changes in the ionic activities (e.g., 10%) generate only
relatively small changes in the measured potential (2.4 mV for
monovalent and 1.2 mV for divalent ion response). Similar to Fig. 3c,
such small potential deviation from a tted response curve generally cannot be easily recognized on the calibration curves most
commonly provided in scientic papers with potential scales that
may span over 600 mV (z = 1) or 300 mV (z = 2) corresponding
to 10 orders of magnitude linear range (sub-nanomolar detection
limit). Indeed, these plots are generally not aimed for quantitative
analytical purposes but rather to show the wide dynamic range of
the electrodes, i.e., to show their response characteristics between
their upper and lower detection limits. For quantitative analysis
however, the reproducibility of the potential measurements in the
same solution following measurements in more dilute and more

concentrated solutions and the RMSD value around the tted line
of the calibration curve, are the most informative. To provide the
analytical results with their condence intervals the condence
intervals for the measured potentials and the tted line should be
considered. To get a quick quantitative overview about the magnitude and distribution errors in the potential measurements instead
of the conventional calibration curves (Ecell vs. log aI or log cI ) one
can plot the potential deviations from the tted line as a function
of log aI or log cI as shown in Fig. 3d.
The data in Table 2 and Fig. 3 show that the parameters of
the calibration curves determined by linear regression may be signicantly different from each other if they are calculated by the
neglecting of the diffusion potentials and/or assuming that the
standard solutions can be treated as ideal (innitely dilute) solutions with I = 1. Consequently, if these calibration curves were
used to determine the ionic activities in unknown samples the
errors may become unacceptable large. In our example, even with
3.5 M KCl as salt bridge electrolyte (which might be considered the
best choice) the error in the calcium ion activity would be around
5%. Since, these calculations did not consider any other sources of
errors, e.g., the reproducibility of the potential measurement, or the
inuence of other ions in the sample on the measured potential, the
error in unknown samples are expected considerably larger.
Another important message of Table 2 is that the characterization of newly developed indicator electrodes, to show that
its response corresponds to the expected Nernstian response, is
only meaningful if these calibration curves are determined from
cell voltage data corrected by the diffusion potentials and ionic

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

activities that were adequately calculated. In papers reporting the

performance characteristics of novel electrodes providing details
on procedures used for calibration and evaluating the response
parameters is essential for adequate assessment. As it is shown in
Table 2 inadequate assumptions still may lead to close to Nernstian
response slope (Rows 2, 3 and 5).
2.3. Measurement of ion concentrations with potentiometric
electrodes
The requirements for determining ion concentrations with
potentiometric electrodes with the help of calibration curves
are very different from the measurement of ionic activities. The
difculty in activity measurement is the uncertainty of activity
coefcients in the standard solutions. However, once the calibration curve has been established, it is adequate for the determination
of the ionic activities in any sample (the potential sources of errors
were discussed above). In clear contrast, when calibration curves
are used for determining ion concentrations, the composition of
the standard solutions should be as similar to the samples as possible. During concentration measurement it is always assumed that
the activity coefcients in the sample and the standard solutions
are the same. This can be achieved by preparing the standard solutions with a matrix (background) as similar to the sample matrix
as possible. Any difference between the ionic strength of the standards and the sample will introduce an error. The magnitude of
this error depends on the level of concentration. The concentration
of a 103 M solutions (z = 1, monovalent ions) can be determined with 2% accuracy as long the difference between the ionic
strength of the sample and the standard is less than a factor of
5. The same difference in the ionic strength introduces 5% error
in the concentration determination of divalent ions. If the samples
are more concentrated the deviations in the ionic strength between
samples and standards must be smaller to maintain the same accuracy. For samples between 103 M and 102 M the difference in the
ionic strength should not exceed 50%, while for samples between
102 M and 101 M the difference should not be larger than 20%
[20].
For matching the composition of the samples with the standards, one should know the approximate composition of the
samples and the composition should be fairly constant. This criterion is almost perfectly met when potentiometric sensors are used
for the analysis of blood electrolytes. The concentrations of H+ , K+ ,
Ca2+ of blood vary only in very narrow ranges (physiologic ranges)
and its high NaCl content controls the total ionic strength. When
the composition of the samples is not known with adequate accuracy or the composition of the samples uctuate in a wide range,
the total ionic strength of the samples and standards can be set
to a constant value through the addition of a large excess of inert
electrolyte to both the samples and the standards (ionic strength
adjustment buffers) [42]. This approach not only provides constant
activity coefcients but at the same time sets the diffusion potentials fairly constant. In Table 1 we showed the diffusion potentials
between CaCl2 sample solutions and 3.5 M KCl, 0.1 M KCl, and 0.1 M
NaCl as salt bridge electrolytes. As the CaCl2 concentration of the
standards changed from 102 M to 105 M the overall changes in
the diffusion potentials were 1.4 mV, 4.4 mV, and 36 mV, respectively. In Table 2 we showed how these changes distort the slope
of the calibration curves. Now, if the CaCl2 standards were prepared with a constant 0.14 M NaCl background (corresponding to
the mean NaCl concentration in whole blood), the overall changes
in the diffusion potentials between 102 M and 105 M CaCl2 standards would drop to 0.2 mV, 1.0 mV, and 1.3 mV by using 3.5 M KCl,
0.1 M KCl, and 0.1 M NaCl as salt bridge electrolyte, respectively. At
the same time the activity coefcients for Ca2+ ions would change
only from  Ca = 0.326 (102 M CaCl2 + 0.14 M NaCl) to  Ca = 0.346

Fig. 4. Theoretically expected calibration curves for a Ca2+ ion-selective electrode

in the presence of a 0.14 M NaCl using 3.5 M KCl bridge electrolyte in the reference
electrode. () E vs. log a or () E vs. log c. The curves were constructed with the
assumptions that EI0 = 0 and the interference introduced by the presence of NaCl is
pot

negligible (KCa,Na = 0).

(105 M CaCl2 + 0.14 M NaCl). Since at constant ionic strength the

activity coefcients are constant and the diffusion potentials hardly
change the Ecell vs. log c and Ecell vs. log a calibration curves run parallel to each other with a slope that approaches the Nernstian slope
even without the correction for the diffusion potentials (Fig. 4).
The separation between the two calibration curves (14 mV) corresponds to the difference in the activity coefcients.
In summary, if an ion-selective electrode is to be used for quantitative analysis based on a calibration curve, before preparing the
standard solutions one should decide whether the ionic activity
or the concentration will be determined from the potential data
measured in the unknown sample. At the same time, among other
criteria, the level of acceptable errors in the analysis must be known
to design the experiments adequately. For assessing the quality of
the calibration curve, use of the RMSD instead of the correlation
coefcient is recommended. It provides a quantitative estimate on
the expected errors in the concentration rage of the calibration.
However, to provide the concentration or activity in an unknown
sample with its condence intervals the condence interval of the
calibration curve as well as the measured data points has to be
known.
2.4. Analysis of real samples: the role of the selectivity coefcient
The success of ion-selective electrodes and their broad acceptance in the late sixties is related to their outstanding selectivity.
The phase boundary potential of an ideal ion-selective electrode is
determined by the activity (concentration) of a single ion even in
the presence of others of the same charge sign (Eq. (2)). Although
several ion-selective electrodes approach this ideal behavior, other
ions (interfering ions) may contribute also to the phase boundary
potential of real ion-selective electrodes. We have discussed this
contribution in terms of the individual exchange current densities
related to Fig. 2. However, to take into account the contribution
of interfering ions to the measured cell voltage quantitatively the
pot
selectivity coefcient (KI,J ) is used:
Ecell =

0
Ecell,I

RT
+
ln
zI F

aI +

pot
KI,J azJ I /ZJ

(11)

where aI and aJ are the activities of the primary and interfering ions in the sample. Eq. (11) was rst derived for H3 O+ -selective
glass electrodes and is known as the Nikolsky-equation. Due to the
importance of the selectivity coefcients, IUPAC formulated recommendations for their determination and the measured values were
tabulated in extensive reviews [26,4347]. More recently it has

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

is plotted as a function of the logarithm of the primary ion activity using the data in Fig. 5a. One could construct an identical plot
pot
to Fig. 5 by setting the selectivity coefcient constant (KI,J = 102 )
and plotting the Ecell vs. log aI response curve for interfering ion concentrations ranging between 102 M and 105 M. The latter would
be an example of an ISE with a known selectivity coefcient being
used for the analysis of real samples with different interfering ion
content. Looking at these pictures one must ask how to use the tabpot
ulated KI,J and detection limit data and how much systematic error
one could expect due to (i) the non perfect selectivity of ISEs; (ii)
the inaccuracies in the published selectivity coefcient data; and
(iii) our limited knowledge on the level of interferences in the sample. Due to these difculties, instead of taking the inuence of the
interfering ions on the measured cell voltage into consideration,
pot
generally a required selectivity coefcient is calculated (KI,J(max) )
based on the acceptable error (PI,J ) in the activity determination (Eq.
(12a)) or the standard deviation () of the potential measurements
(Eq. (12b)).
aI(min)

pot

KI,J(max) =

(aJ(max) )

aI(min)

pot

KI,J(max) =

Fig. 5. (a) Theoretical calibration curves for a monovalent cation-selective electrode

in the presence of a constant background of interfering ions (aJ = 102 M) and selectivity coefcients for the interfering ion ranging between 0 and 102 . In the gure
pot
the detection limits calculated according to the IUPAC denition for KI,J = 102 and
KI,J = 103 are also marked. (b) Potential deviations from the Nerstian response
of the same electrode as a consequence of non-perfect selectivity. Inset: the same
deviations on an expanded potential scale.
pot

been recognized that some of the published selectivity coefcients

are biased due contaminations of the solution in direct contact with
the ion-selective electrode surface and protocols were developed
for the determination of the unbiased selectivity coefcients [48].
In Fig. 5a we show the inuence of the selectivity coefcients
on a calibration curve. The Ecell vs. log aI calibration curves, in general have
 three regions: (i) Complete interference region where
pot

aI 

KI,J aJ , i.e., the measured cell voltage is independent from

the primary ion activity (ii) Mixed response region where aI


pot

KI,J aJ , i.e., where the contribution of the primary and inter-

fering ions to the measured cell voltage is comparable; and the (iii)
Nernstian response region for the primary
ion where the inuence

pot
KI,J aJ .

of interfering ions is negligible, i.e., aI

The IUPAC denition of the detection limit [26] is the activity or concentration at the cross section of the extrapolated lines
that were tted to complete interference and Nernstian response
regions of the calibration curve recorded at a given level of interference. In Fig. 5a we show the determination of the detection limit
pot
for a monovalent ISE in the presence of aJ = 102 M and KI,J = 102

or KI,J = 103 . At the detection limit aI = KI,J aJ , i.e., the contribution of primary and interfering ions to the measured voltage is
the same, consequently the separation between the ideal response
pot
curve (KI,J = 0) and the actual response curve of the electrode is
18/zI mV. In Fig. 5b, the deviation from the ideal electrode response
pot

pot

zI /zJ

(aJ(max) )zI /zJ

PI,J
100

(12a)

(102/SI 1)

(12b)

where aI(min) is the smallest expected activity for the measured

ion in the sample, while aJ(max) is the largest expected activity of
an interfering ion in the sample. The two equations provide the
same result if the acceptable error is calculated from the standard
deviation of the potential measurements as long the slope of the
pot
calibration curve (SI ) is Nernstian. If KI,J and aJ(max) are known Eqs.
(12a) and (12b) allow calculating the smallest primary ion concentration that can be determined with a predetermined (acceptable)
accuracy. The same information can be obtained from Fig. 5b.
For analyzing samples in the mixed ion range Eq. (11) has to be
tted to the Ecell vs. log a calibration data points. However, when the
charges of the interfering ions are different from the primary ion Eq.
(11) is inconsistent. Consequently its application for calculating the
ionic activity of an unknown sample in which the charges of the primary and interfering ions are different will be an additional source
of error. Moreover, the required selectivity coefcients calculated
with Eqs. (12a) and (12b) may be inaccurate. With somewhat more
complex equations, the response of an ion-selective electrode in the
mixed ion region can be described more accurately (Eqs. (13a) and
(13b)) which provides also a better estimate for the required selectivity coefcient (Eq. (14)). However, due to the uncertainty related
to our limited knowledge on the activity of the interfering ions in
the sample, as well as reduced sensitivity of the electrode in the
mixed ion region the uncertainty of the measurements based on a
calibration curve in the mixed ion region may be still unacceptably
high.
0
+
Ecell = Ecell,I

RT
ln
zI F

1/2
1
aI
pot 2
+ (a2I + 4aJ (KI,J ) )
2
2

for zI = 1 and zJ = 2

Ecell =

0
Ecell,I

RT
+
ln
zI F

(13a)

1 pot
aI + KI,J a2J
4

for zI = 2 and zJ = 1

pot

KI,J(max) =

aI(min)
(aJ(max) )zI /zJ

1/2

pot
K a2
4 I,J J

1/2 
(13b)

PI,J
100

zI /zJ
(14)

10

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

2.5. Rate of response, drift of the measured voltage,

reproducibility of the measurements
One of the most attractive applications of potentiometric sensors is continuous monitoring including in vivo monitoring. In
continuous monitoring, the indicator and reference electrodes are
generally incorporated into an electrochemical ow-cell which is
connected to a stream of owing solution, i.e., the measurements
are performed under a hydrodynamically controlled condition [49].
In the ow cell, the reference electrode is placed downstream from
the indicator electrode to avoid any potential interference from the
leaching salt bridge electrolyte. In continuous monitoring applications the dynamic characteristics of the ion-selective electrode is
the key parameter for recording rapid changes in the concentration
of the owing sample [50]. The dynamic characteristics of ionselective electrodes is commonly characterized with their response
time, the time required to attain a given percentage of a new steady
state value upon a stepwise change in the analyte concentration
[51]. In general, ISEs have short response times ranging from a few
milliseconds to seconds [52]. However, the response time can be
much longer [53], e.g., in highly diluted solutions where drifting
potentials are common. Drift is dened by the IUPAC [26] as a slow
non-random change in the measured cell voltage as function of
time when the ion-selective and reference electrodes are immersed
in a solution of constant composition and temperature. Electrode
potential drift can be a serious source of error. In the presence of
drifting potentials it is impossible to unambiguously correlate a
measured potential value to a sample concentration. Drift deteriorates the reproducibility of the measurements, bias the response
slope of the calibration curve, and lead to erroneous sample concentrations. In continuous monitoring situations it can be very difcult
to differentiate between electrode drift and changes in the cell
voltage related to concentration changes. When the cell voltage
recorded in a solution of constant composition and temperature
changes linearly with time the slope of the line tted to the Ecell vs.
time data points provides the drift. To differentiate between drift
and potential changes related to changes in the solution concentration the RMSD value of the tting provides guidance. When the
change in Ecell is larger than 3 RMSD it is considered statistically
signicant, which indicates a change in the sample concentration.
The drift usually can be categorized as (i) parallel drift, (ii) concentration dependent drift, and (iii) random drift [21]. The drift is
0 changes in Eq. (7) but the slope
termed as parallel when only Ecell
of the calibration curve (S) remains constant. If this change is slow
and unidirectional it is possible to make corrections in the sample
0
cell voltage readings by repeated evaluations of the Ecell
value of
the calibration curve (single point calibration) between the sample
measurements. This protocol is widely applied in blood electrolyte
analyzers to achieve the required precision and accuracy of the
measurements. The drift is termed concentration dependent when
its value is inuenced by the sample composition and inuences
the slope of the calibration curve, e.g., the drift can be different in
its sign and magnitude at high and low concentrations. Such drift
is often correlated by slow changes in the ion-selective membrane
composition due to interference or leaching of membrane components into the sample. If the rate of such changes is slow, repeated
0 and S) can signiftwo point calibrations (determination of both Ecell
icantly improve the results of the analysis. The drift is considered
random when it cannot be characterized by any regular trend. Consequently the negative consequences of random drift cannot be
compensated.
In vivo continuous monitoring of electrolytes would signicantly improve the treatment of critically ill patients. However
besides the issues related to the biocompatibility of the implanted
sensors (e.g., thrombus formation, protein and platelet deposition on the sensor surface, etc.) [54,55] the drift in the measured

potential is the most serious concern that can lead to inaccurate

analytical results with implanted sensors. Leaching of plasticizer
and/or ionophore from the membrane presents biocomparibility
concerns and leads to reduced sensor sensitivity and potential
drifts. The errors associated with such drift cannot be compensated
since under in vivo conditions there is no possibility for frequent
calibrations. To compensate the drift related errors in monitoring ionic uxes through plasma membranes in cell physiology the
method of self referencing has been introduced [56].
3. Method comparison studies
In the previous discussion we described the challenges and
approaches for performing high accuracy measurements with ISEs
related to the logarithmic response of potentiometric indicator
electrodes and the uncertainties in the diffusion potentials. In the
1960s, these challenges were overcome in clinical analysis of blood
electrolytes. Method comparison studies were the key in the process in which ion-selective electrode potentiometry completely
replaced ame photometry and atomic absorption spectroscopy
for the measurement of K+ , Na+ , and Ca2+ ion concentrations in
biological samples. Indeed, the ion analysis of whole blood made
the ISEs one of the biggest success stories in the history of chemical sensors. To achieve the same success with the new class of
ion selective electrodes in microanalysis, the performance characteristics of the new class of ISEs, with sub-nanomolar detection
limits, the unique advantages of the potentiometric method must
be demonstrated in method comparison studies. Method comparison studies seek to answer the question of how well one method of
measuring some quantity reproduces that same quantity measured
via another method.
Evaluation of a new or improved method involves several steps.
The rst is to establish the performance criteria the new method
is required to achieve. Next, one must design and execute experiments that allow the objective assessment of each of the specied
criteria. The nal step of the evaluation is the comparison of the
results with the established performance criteria to determine
whether the new method meets the desired criteria. Additionally,
but not as a substitute for the previous steps, one can measure the
agreement between the results provided by an established method
and the new method in a method comparison study. While experimentally this approach may be straightforward many times the
statistical analysis can be quite complicated.
3.1. Statistical methods to evaluate the agreement between two
methods
For some time, a scatter plot (x vs. y) was commonly constructed
from the data measured in identical samples with Method 1 and
Method 2 (e.g., Fig. 6a) and least-squares regression (LSR) analysis was used to assess agreement. In evaluating the agreement by
LSR analysis, the slope (B1 ) and intercept (B0 ) of the regression line
is compared to the line of equality with a slope of one and intercept of zero. A proportional bias exists between the two methods
when the slope of the regression line is different from unity. On
the other hand, a constant bias exists between the two methods
when the intercept of the regression line is different from zero.
Since there is an uncertainty in the estimates of the slope and intercept of the regression line one has to know the condence intervals
of these parameters to decide whether the recorded differences
are statistically signicant. Cornbleet and Gochman [57] cautioned
against the use of LSR analysis when the underlying assumptions
of the LSR method are not met, i.e., the independent variable contains errors comparable to the errors of the dependent variable.
In particular, their analysis showed that signicant error in the

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

11

Fig. 6. Graphical representation of two data sets (Method 1 vs. Method 2). The data sets were generated by the y = 0.9x + 10 function with the Method 1(x) values ranging
between 135 and 145. The individual data in both Method 1 and Method 2 (x and the y data) are associated with the same relative error (Syx = Sxy = 1.0%). (a) Scatter plot
(Inset: enlarged section of the same gure showing the individual data points), (b) B&A plot of the differences ( vs. the average values Av) (c) B&A plot of the logarithmic
differences (log  vs. the average of the log x and log y), (d) The differences () vs. the results of Method 1. The horizontal lines in (b)(d) represent the mean values of the
differences (red lines in the middle) and the 95% condence intervals of the mean (blue lines below and above the red horizontal lines). The slanted lines in (a), (b) and (d)
represent the best line t to the data points (Eqs. (15)(17)). (For interpretation of the references to color in this gure legend, the reader is referred to the web version of
the article.)

LSR slope estimation occurs when the ratio of the standard deviation of measurement of a single x value (Syx ) to the standard
deviation of the x-data set (Sx ) exceeds 0.2 (RCG = Syx /Sx 0.2).
This typically occurs when the range of values measured is small.
Under such circumstances the method of Deming [58] or others
[5961] was proposed for the determination of the best-t line,
which minimizes the sum of squares in both the x and y directions
simultaneously.
In two seminal articles, Bland and Altman (B&A) [62,63] criticized using the correlation coefcient and LSR for characterizing the
agreement between two methods. B&A argued that the correlation
coefcient measures the association between two methods and not
their agreement. Instead of the scatter plot and regression analysis, Bland and Altman proposed plotting the differences between
the two methods ( = y x) against the average (Av = [y + x]/2)
of the two methods (e.g., Fig. 6b) [62,63]. If the data in such
a BlandAltman plot are normally distributed around a horizontal line the mean value of the differences gives the constant
bias between the two methods. Alternatively, when the difference
between the methods varies with the average of the two, a logarithmic transformation of the data and plotting the logarithmic
differences against the average of the logarithmic values were suggested (e.g., Fig. 6c). To estimate the proportional bias between the

two methods (a dimensionless ratio of the quantities measured by

Methods 1 and 2) the antilog value of the mean of the logarithmic
differences is calculated. As an alternative, other authors [6466]
suggested plotting the ratio of the corresponding x and y values
against the mean of the x and y values to demonstrate proportional
bias.
3.2. A typical example
In many analyses of biological samples, the clinical range is
narrow while the variations in the measured data can be large.
Under such conditions, the Cornbleet and Gochman [57] criterion
is expected to approach or exceed its critical value. Consequently
the LSR analysis may not be suitable for assessing the agreement
between two methods based on the scatter plot. A typical example
would be the measurement of sodium ion concentration in serum.
In Fig. 6, we show the analysis of a data set with 180 data points
scattered between 135 and 145 (corresponding to the sodium concentration in mM in serum measured with two methods). Fig. 6a
shows the data in a conventional scatter plot. Fig. 6b and c are the
B&A plots of the differences and logarithmic differences, respectively. Finally, Fig. 6d is a plot of the differences vs. the results of
Method 1 (x).

12

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

Table 3
Statistical analysis of the data set shown in Fig. 6 with RCG = 0.29. The data set
with 180 data points was generated by the y = 0.9x + 10 function in the range of
135 and 145. Both the x and the y data are associated with the same relative error
(Syx = Sxy = 1.0%). The data in parenthesis are the differences in % between the values
of B1 = 0.9 and B0 = 10 and the calculated values of B1 and B0 . The results in shaded
cells are associated with unacceptably large errors and/or indicate the inappropriate
selection or application of the applied statistical method.

Calculated slope (B 1)

Method used
1
2
3
4
5
6

LSR
Deming [58]
B&A difference plot
(Difference vs. average)
B&A logarithmic analysis
(log. difference vs. log. average)
Eq. (18)
Eq. (19)

Calculated intercept (B 0)

0.78 0.03 (13.3%)

0.88 0.04 (2.2%)

27 5 (+170%)
13 6 (+30%)

1.00 (+11%)a

4 2 (140%)

0.97 0.01 (+8%)

0.0 (100%)b

0.89 0.43 (1.1%)

0.78 0.03 (13.3%)

11 6 (+10%)c
27 5 (+170%)

Assumed value by using the B&A method of the differences.

Assumed value by using the B&A method of the logarithmic differences.
The standard deviations in B1 and B0 has been calculated from the standard deviations in b1 and b0 using the rules of error propagation.
b
c

The correlation between the x and y data in Fig. 6a is represented

by the equation:
y = B0 + B1 x

where B0 = 27.0 and B1 = 0.78

(15)

The correlation between the differences  = y x and the x values in Fig. 6b is represented by the equation:
 = b0 + b1 Av where b0 = 12.0 and b1 = 0.116

(16)

Finally the correlation between the differences  = y x and the

x values in Fig. 6d can be described with the equation:
 = 0 + 1 x

where 0 = 27.0 and 1 = 0.22

(17)

The parameters of Eqs. (15)(17) were determined by the LSR

method and the tted lines are plotted as slanted lines in Fig. 6a, b,
and d. Since the data are scattered in a narrow range (between 135
and 145) the standard deviation of the data set (Sx ) is hardly larger
than the standard deviation of a single x value (Syx ), which resulted
in RCG = 0.29, a value larger than 0.2, the Cornbleet and Gochman
[57] criterion.
The narrow range of the measured data with a large distance
from the intercept of the regression line (Fig. 6a) should be a concern in using the LSR method of analysis. The large value of the
Cornbleet and Gochman criterion (RCG = 0.29) is an additional indication that LSR method may not be adequate for the assessment
of the agreement between the two methods. Under such conditions the Deming regression has been recommended instead of LSR
analysis.
Apparently the data points in the B&A plot of the differences
(Fig. 6b) are not normally distributed around the mean (red horizontal line). Under such circumstances B&A suggests to calculate
the proportional error between the compared methods from the
mean of the logarithmic differences. The results of the LSR, Deming
and B&A analysis of the data in Fig. 6 are summarized in Table 3. By
comparing the parameters of the equation that was used to generate the data in Fig. 6a (y = 0.9x + 10), and the parameters of the
linear correlation between Methods 1 and 2, it is obvious that neither the LSR nor the B&A analysis provides acceptable results in this
particular case. On the other hand, the Deming regression provided
acceptable results.
Of note, it can be shown that the slope and intercept of the tted
lines in Fig. 6b (b1 and b0 ) and Fig. 6d (1 and 0 ) are related to
the slope (B1 ) and intercept (B0 ) of the tted regression line in the
scatter plot (Fig. 6a):
B0 =

2b0
,
2 b1

B1 =

2 + b1
2 b1

(18)

B0 = 0 ,

B1 = 1 + 1

(19)

Consequently, Eqs. (18) and (19) provide alternative evaluation

protocols for the assessment of agreement between two methods.
As shown in Table 3, the estimates for the constant (B0 ) and proportional (B1 ) biases calculated with Eq. (18) are very close to the
parameters of the linear function used to generate the data set
although, the uncertainty of these estimates are quite large. The
B0 and B1 values calculated with Eq. (18) are also very similar to
the values provided by the Deming method. The importance of Eq.
(18) in the assessment of agreement between two methods is that
it (i) retains the advantages of the B&A representation of the data
in the difference plot, and (ii) allows the simultaneous assessment
of proportional and constant biases. On the other hand, Eq. (19)
provides exactly the same B0 and B1 values as the LSR analysis of
the data in the conventional scatter plot. It cannot be used to assess
agreement if RCG 0.2.
Apparently, the different statistical methods provide quite different parameters for the assumed linear correlation between the
results provided by methods 1 and 2. In analyzing real data this
can be quite confusing since it is not obvious which function represents best the true relationship between the compared methods.
By recognizing the differences in the results of the evaluation
methods one must ask the source of the differences. The B&A
method assumes, that correlation can be adequately represented
with either only constant or only proportional bias. If this assumption is not true, the B&A analysis provides incorrect estimates for
the constant and proportional biases. However, when the population of the differences is a linear function of the mean Eq. (18) can
be used to estimate the constant and proportional biases. On the
other hand, if RCG 0.2 Eq. (19) provides the same incorrect answer
as the LSR analysis of the scatter plot data. The comparison of the
B0 and B1 values calculated from Eqs. (18) and (19) has additional
relevance because there are conicting arguments in the literature
whether potting the y x differences against the mean (Fig. 6b) has
unique advantages [67,68] compared to plotting the same differences against the results of Methods 1 or 2 as in Fig. 6d. The very
good agreement between the parameters used to generate the data
set in Fig. 6a and the values calculated by Eq. (18) using the parameters of the tted line in Fig. 6b (b0 and b1 ) is a clear benet by
plotting the differences against the mean as in Fig. 6b.
In summary, interpreting the results of method comparison
studies may be very difcult. Therefore, before applying any statistical methods the trends, sign sequences, outliers, etc., in the
original data (e.g., the scatter plot) or their transformations (e.g.,
the B&A plot of the differences) should be carefully studied, i.e., it
has to be determined whether the assumptions of the selected statistical method are met. We advocate plotting the data as a scatter
plot rst and determining the Cornbleet and Gochman [57] criterion. If the Cornbleet and Gochman criterion is much smaller than
its critical value (RCG  0.2) the use of the least-squares regression analysis of the x and y data is adequate to ascertain the type of
bias that exists between the two methods from the intercept and
slope of a tted line, respectively. However, when the Cornbleet
and Gochman criterion approaches its critical value, the LSR cannot be used to evaluate the agreement between two methods with
condence. In such cases, the use of the Deming and Passing or
Bablock [5860] method is recommended. Alternatively, plotting
the data as a difference plot ( vs. Av) and evaluating the correlation between y and x by utilizing Eq. (18) will also provide accurate
results.
Similar to Figs. 3d and 5b the B&A plot of the differences (Fig. 6b)
is an excellent way to show the distribution of the differences
between two methods. However, one should be very cautious when
using the B&A plot of the differences to determine the constant
bias between two methods. It can be used only in the absence of

E. Lindner, B.D. Pendley / Analytica Chimica Acta 762 (2013) 113

proportional bias. Indeed, in the absence of proportional bias it is

superior to all the regression methods and can be used even when
the Cornbleet and Gochman criterion is not met. The B&A method
of the logarithmic differences can be used to determine the proportional bias only in the absence of constant bias. However, neither B&A
methods will give accurate results if both biases exist and are not
suggested if the possibility of the combined presence of constant
and proportional biases cannot be excluded unambiguously.
4. Conclusions
In this article the methodology of direct potentiometry is discussed including the validation of novel ISE responses as well as
the results of potentiometric analysis. First, the underlining theory of ion-selective electrodes is discussed. Next, the challenges
in the accurate determination of the ionic activities related to
the uncertainties in the activities of the standard solutions and in
the diffusion potentials between calibration and measurement are
highlighted. Then, protocols are provided to minimize the errors in
the determination of the ion concentrations in real samples. Subsequently, the inuence of the imperfect selectivity of the electrodes
and drifts in the measured potentials on the analytical results is
discussed. Finally, the statistical methods for comparing the experimental results provided by independent analytical techniques are
discussed and recommendations are provided for the methodology
of such comparisons. We hope that the concepts in this contribution will help researchers capitalize on the recent advances in
potentiometry and extend this analytical method to practical applications microanalysis.
Acknowledgement
The nancial support from the FedEx Institute of Technology
for the establishment of the Sensor Institute of the University of
Memphis SENSORIUM is gratefully acknowledged.
References
[1] . Pergel, R.E. Gyurcsnyi, K. Tth, E. Lindner, Anal. Chem. 73 (2001) 42494253.
[2] E. Bakker, E. Pretsch, Anal. Chem. 74 (2002) 420A426A.
[3] T. Sokalski, A. Ceresa, T. Zwickl, E. Pretsch, J. Am. Chem. Soc. 119 (1997)
1134711348.
[4] E. Bakker, P. Bhlmann, E. Pretsch, Electroanalysis 11 (1999) 915933.
[5] E. Bakker, D. Diamond, A. Lewenstam, E. Pretsch, Anal. Chim. Acta 393 (1999)
1118.
[6] E. Bakker, E. Pretsch, Trends Anal. Chem. 20 (2001) 1119.
[7] E. Bakker, E. Pretsch, Angew. Chem. Int. Ed. 46 (2007) 56605668.
[8] E. Bakker, E. Pretsch, Trends Anal. Chem. 27 (2008) 612618.
[9] A. Ceresa, E. Bakker, B. Hattendorf, D. Gunther, E. Pretsch, Anal. Chem. 73 (2001)
343351.
[10] E. Pungor, E. Hollos-Rokosinyi, Acta Chim. Acad. Sci. Hung. 27 (1961) 63.
[11] M.S. Frant, J.W. Ross, Anal. Chem. 40 (1968) 1169.
[12] J.W. Ross, Science 156 (1967) 1378.
[13] Z. Stefanac, W. Simon, Chimia 20 (1966).
[14] G. Horvai, E. Pungor, Anal. Chim. Acta 113 (1980) 287294.
[15] E. Bakker, P. Bhlmann, E. Pretsch, Chem. Rev. 97 (1997) 30833132.
[16] P. Bhlmann, E. Pretsch, E. Bakker, Chem. Rev. 98 (1998) 15931687.
[17] J. Bobacka, A. Ivaska, A. Lewenstam, Chem. Rev. 108 (2008) 329351.
[18] P. Bhlmann, L.D. Chen, in: A.W. Steed, P. Gale (Eds.), Supramolecular Chemistry: From Molecules to Nanomaterials, vol. 5, John Wiley & Sons, Ltd, New
York, 2012, pp. 25392579.

13

[19] E. Lindner, R.E. Gyurcsnyi, E. Pretsch, in: H.-J. Schneider (Ed.), Applications of
Supramolecular Chemistry, CRC Press, Taylor & Francis Group, Boca Raton/New
York/London, 2012, pp. 81104.
[20] K. Cammann, Working with Ion-Selective Electrodes, Springer Verlag,
Berlin/Heidelberg/New York, 1979.
[21] R.A. Durst, in: H. Freiser (Ed.), lon-selective Electrodes in Analytical Chemistry,
vol. 1, Plenum Press, NY, 1978 (Chapter 5).
[22] J. Koryta, Ions, Electrodes and Membranes, John Wiley & Sons, Chichester/New
York/Brisbane/Toronto/Singapore, 1982.
[23] G.J. Moody, J.D.R. Thomas, Selective Ion Sensitive Electrodes, Merrow Publ. Co.
Ltd, Watford, Herts, 1970.
[24] W.E. Morf, The Principles of Ion-selective Electrodes and of Membrane Transport, Elsevier Sci. Publ. Co., Amsterdam/Oxford/New York, 1981.
[25] Orion Research, Handbook of Electrode Technology, Orion Research,
Cambridge, MA, 1982.
[26] R.P. Buck, E. Lindner, Pure Appl. Chem. 66 (1994) 25272536.
[27] K. Cammann, Schweiz. Lab. Z. 44 (1987) 57.
[28] W. Nernst Ber, Dtsch. Chem. Ges. 30 (1987) 15471563.
[29] M. Cremer, Z. Biol. 47 (1906) 562.
[30] F. Haber, Z. Klemensiewicz, Z. Phys. Chem. 67 (1909) 385.
[31] M.S. Frant, J. Chem. Educ. 74 (1997) 159166.
[32] E. Lindner, K. Tth, Electroanalysis 21 (2009) 18871894.
[33] M.S. Frant, Analyst 119 (1994) 2293.
[34] J. Ruzicka, J. Chem. Eng. 74 (1997) 166170.
[35] R.P. Buck, E. Lindner, Anal. Chem. 73 (2001) 2A11A.
[36] E. Lindner, R.E. Gyurcsnyi, J. Solid State Electrochem. 13 (2009) 5168.
[37] A. Michalska, Anal. Bioanal. Chem. 384 (2006) 391406.
[38] K. Cammann, in: E. Pungor, I. Buzas (Eds.), Ion-selective Electrodes, Akademiai
Kiado, Budapest, 1978.
[39] T. Kakiuchi, J. Solid State Electrochem. 15 (2011) 16611671.
[40] R.E. Dohner, D. Wegmann, W.E. Morf, W. Simon, Anal. Chem. 58 (1986)
25852589.
[42] M.S. Frant, J.W. Ross, Anal. Chem. 40 (1968) 11691171.
[43] E. Lindner, Y. Umezawa, Pure Appl. Chem. 80 (2008) 85104.
[44] Y. Umezawa, Handbook of Ion-selective Electrodes: Selectivity Coefcients,
CRC Press, Boca Raton, FL, 1990.
[45] Y. Umezawa, P. Bhlmann, K. Umezawa, K. Tohda, S. Amemiya, Pure Appl. Chem.
72 (2000) 1851.
[46] Y. Umezawa, K. Umezawa, P. Bhlmann, N. Hamada, H. Aoki, J. Nakanishi, M.
Sato, K.P. Xiao, Y. Nishimura, Pure Appl. Chem. 74 (2002) 923994.
[47] Y. Umezawa, K. Umezawa, H. Sato, Pure Appl. Chem. 67 (1995) 507518.
[48] E. Bakker, Anal. Chem. 69 (1997) 10611069.
[49] K. Tth, K. Stulik, W. Kutner, Z. Fehr, E. Lindner, Pure Appl. Chem. 76 (2004)
11191138.
[50] E. Lindner, K. Tth, E. Pungor, Dynamic Characteristics of Ion-Selective Electrodes, CRC Press, Inc., Boca Raton, FL, 1988.
[51] E. Lindner, K. Tth, E. Pungor, Pure Appl. Chem. 58 (1986) 469479.
[52] E. Lindner, K. Tth, E. Pungor, T.R. Berube, R.P. Buck, Anal. Chem. 59 (1987)
22132216.
[53] M. Huser, P.M. Gehrig, W.E. Morf, W. Simon, E. Lindner, J. Jeney, K. Tth, E.
Pungor, Anal. Chem. 63 (1991) 13801386.
[54] K.A. Mowery, M.H. Schoensch, N. Baliga, J.A. Wahr, M.E. Meyerhoff, Electroanalysis 11 (1999) 681686.
[55] M.E. Meyerhoff, Trends Anal. Chem. 12 (1993) 257.
[56] P.J.S. Smith, R.S. Sanger, M.A. Messerli, in: A.C. Michael (Ed.), Methods and
New Frontiers in Neuroscience, CRC Press, Boca Raton, FL, 2007, pp. 373405
(Chapter 18).
[57] P.J. Cornbleet, N. Gochman, Clin. Chem. 25 (1979) 432438.
[58] W.E. Deming, Statistical Adjustment of Data, John Wiley & Sons, New York, NY,
1943.
[59] H. Passing, W. Bablok, J. Clin. Chem. Clin. Biochem. 21 (1983) 709720.
[60] H. Passing, W. Bablok, J. Clin. Chem. Clin. Biochem. 22 (1984) 431445.
[61] H.W. Zwanziger, C. Sarbu, Anal. Chem. 70 (1998) 12771280.
[62] D.G. Altman, J.M. Bland, Statistician 32 (1983) 307317.
[63] J.M. Bland, D.G. Altman, Lancet 30 (1986) 307310.
[64] K. Dewitte, C. Fierens, D. Stockl, L.M. Thienpont, Clin. Chem. 48 (2002) 799801.
[65] S. Eksborg, Clin. Chem. 27 (1981) 13111312.
[66] K. Linnet, I. Bruunshuus, Clin. Chem. 37 (1991) 16691675.
[67] J.M. Bland, D.M. Altman, Lancet 346 (1995) 10851087.
[68] J.S. Krouwer, Stat. Med. 27 (2007) 778780.