£ rgs #s ft * ft ft « © ft ft ft ft ft ft ft ft ft ft ft ft « ft

ft • ft % m # ft ft « « * ft ft ft » m 9 ft ft ft ft ft ft # ft ft ft ft ft ft
ft © ft ft ft ft ft ft ft ft ft ft ft ft • ft ft ft ft ft ft 9 ft ft ft 0
ft ft ft * ft ft ft « ft ft ft ft ft ft # • ft ft ft ft ft ft ft ft • ft ft ft ft ft «
ft ft ft ft # ft ft ft ft ft ft ft ft ft ft ft ft # ft ft ft ft ft # ft
<2 « ft ® # ft ft ft e? # ft ft ft ft * ft ft ft ft ft ft ft ft ft « ft ft ft ft ft
ft ft ft ft ft ft ft * ft ft ft » • ft ft ft ft ft ft ft jit ft ft ft ft ft ft ft
#v
>

ft
ft
ft
•
ft
•
$
m W * I*
T-
ft
ft
ft
ft
ft
# ft
ft
ft
•
ft
ft m
ftw
ft
ft
ft
ft
m
ft
ft
ft
$
»
ft
« # ft ft * ft
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.
Downloaded via UNIV INDUSTRIAL DE SANTANDER on August 30, 2021 at 18:37:01 (UTC).

ft ft ft ft ft ft ft ft ft ft ft ft ft ft • s ft • ft ft ft ft
ft ft ft ft ft ft ft ft ft • ft ft ft ft ft # • • ft ft ft ft
ft ft ft ft • ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft # ft *
ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft * « ft ft ft ft
.

ft ft
ft ft ft ft # ft ft ft ft ft ft ft ft ft ft ft ft ft O ft ft ft ft ft ft ft ft
ft ft ft # ft ft ft • ft ft ft ft • ft * ft ft ft ft ft « ft ft ft ft # ft ft #
ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft ft * ft # ft ft ft ft ft ft ft ft
* ft ft ft ft ft # ft ft ft ft # ft ft ft ft ft m « « ft # ft ft ft ft ft ft -

« ft ft ft ft ft ft ft ft ft ft ft ft ft * ft ft ft ft ft ft # ft ft ft ft

Figure 1. Analytical calibration curve

of signal, x, vs. concentration, c
The unknown signal, x^, is related to an unknown
concentration, cUt through the equation x me +- —

i, where m is the slope and i is the intercept of

the line

The ability to quantify a trace ele- approaches (1-13). In spite of the fact examine the statistical meaning of
ment molecule in chemical and bio-
or that the International Union of Pure limit of detection values in a format
logical matrices using specific analyti- and Applied Chemistry (IUPAC) consistent with the IUPAC definition.
cal methods is often viewed in terms adopted a model for the limit of detec- It is not intended to be a comprehen-
of the limit of detection. This limit of tion calculations in 1975 (7), and the sive review of the various methods of
detection is a number, expressed in ACS Subcommittee on Environmental calculating limit of detection values.
units of concentration (or amount), Analytical Chemistry reaffirmed this For a more complete review, the read
that describes the lowest concentra- standard in 1980 (2), acceptance of er is referred to the excellent articles
tion level (or amount) of the element this model by the general analytical by Kaiser (3-6), Boumans (7-9), Cur-
that an analyst can determine to be community has been slow. The result rie (10), Glaser et al. (11), as well as
statistically different from an analyti- of this slow acceptance has led to a available textbooks (12,13). Rather,
cal blank (1). Although this definition great deal of uncertainty when limits this REPORT is intended to be a sim-
seems rather straightforward, signifi- of detection are used as a basis for ple and general discussion on methods
cant problems have been encountered comparison between various analytical for calculating limits of detection; it is
in expressing these values because of procedures, methods, or analytical in- geared to the analyst who does not
the various approaches to the term struments. Unless the limits of detec- have a rigorous knowledge of statis-
“statistically different.” The calculat- tion are calculated in a consistent tics. In this examination, the signifi-
ed limit of detection for an element manner, the comparison may be cance of the limit of detection values
can easily vary an order of magnitude meaningless. will be emphasized, and the possible
through the use of different statistical It is the purpose of this REPORT to problems encountered when using

712 A • ANALYTICAL CHEMISTRY, VOL. 55, NO. 7, JUNE 1983 0003-2700/83/0351-712A501.50/0

1983 American Chemical Society
&copy;
 Report
Gary L. Long
J. D. Winefordner
Department of Chemistry
University ol Florida
Gainesville, Fla. 32611

ft#*#########*####**#*#*####
ft###*#############*####*## ft #
######## ^###**#*##»##«######»#####« ft 4
####«#«#<; • •ft**##*##***##########*## «##••*#•
####### It*############*#*****###### & * * # # « * # ft
ft#################**###### ft#######
ft ft ft * ft * ft » •##*•#•#**#*####*••#•#*•*# »########
########
####*#«# l#« ••#••••••### ##•«##•#•###
The limit of detection is the lowest concentration level that can be determined
to be statistically different from an analytical blank. Significant problems * ft *
have been encountered in expressing these values because of the various # * * *
approaches to the term “statistically different” ###••##«
######*#
3 * ft ft # * ft #**##* #
# # # * # # i
.##•#*
#
* # # « # ® # ##«#####
# # # # # # * # # * n £#######
m # * m # ft#######
# # # » # ######« p ##*#•« #
* # # * « E # # # * ft ft ft#######
# # # • # # ft ft # ft ft S

Figure 2. Normal distribution curve for

a measured x variable
The pink shaded area, a, represents the
ity that Xe s' 0* + kc) or xL 5s (xB + 3sB)

limit of detection values obtained tected with reasonable certainty for a generally plots of signal, x, vs. analyte
from non-IUPAC methods as a basis given analytical procedure” (J). This concentration, c, and are represented
for comparing methods and instru- concept is further clarified by the ACS as in Figure 1. The relationship be-
ments will be discussed. Also, two definition, which states “the limit of tween x and c can be obtained by per-
methods will be introduced that per- detection is the lowest concentration forming a linear regression analysis on
mit measurement errors in the analyt- of an analyte that an analytical pro- the data. This analytical calibration
ical sensitivity to be included in the cess cart reliably detect” (2). To un- relationship can be expressed as
limit of detection calculation. These derstand what a reasonably certain
two methods as well as the IUPAC measure or a reliable detection is, the x =
me + i (1)
method will be evaluated for their method of measurement as well as the
ability to incorporate these errors into errors (including noises) associated where m is the slope or analytical sen-
the numerical limit of detection. with the measurement must be well sitivity and i is the intercept. When an
understood. unknown sample containing the ana-
Definition lyte is subjected to the analytical pro-
cedure, a value, xy, can be measured.
The IUPAC definition, adopted in Measurements This value may then be inserted into
1975, states that “the limit of detec- Generally, most analytical methods Equation 1 to determine the concen-
tion, expressed as a concentration cl require the construction of analytical tration of the unknown, cu- However,
(or amount, qy), is derived from the calibration curves for the determina- the ability to solve accurately for cu is
smallest measure, xy, that can be de- tion of unknowns. These curves are dependent upon how well the m and i

ANALYTICAL CHEMISTRY, VOL, 66, NO. 7, JUNE 1983 • 713 A

values areknown. As long as the cali- as shown in Figure 3. The limit of de-
bration is obtained in the linear
curve tection is found by relating ksg to a
response region of the method, the concentration value by dividing by the
more points obtained in the construc- slope of the calibration curve line ob-
tion of the calibration curve the better tained from the linear regression anal-
defined the m value will be. Also, if ysis. However, the cl value obtained
points are sampled near the origin, the can only be a true reflection of the
i value may be better defined. How- limit of detection when m is well-de-
ever, if the m and i values are not well- fined and i is essentially 0.
defined because of nonlinearity in the The use of k 3 allows a confidence
=

calibration curve or a poor choice of level of 99.86% that xl & (xb + 3sb)
calibration curve ranges, the result of for a measurement based on the error
the unknown determination, cu, may of the blank signal following a normal
be subject to considerable error. distribution. It must be emphasized
that if xb does not follow a normal
Statistics distribution, then the probability
Figure 3. Analytical calibration curve that xl > (xg + 3sg) would be
The amount of error associated with of signal, x, vs. concentration, c, show- 100(1 1/k2), or 89% according to
—

a measurement of x can be statistical- ing the relationship of ksB to the limit of Tschebyscheff’s inequality (5). Hence,
ly estimated. Most measurements are detection, cL values of k < 3 should not be used for
subject to error that follows a normal limit of detection calculations.
distribution. If a sufficiently large
number of observations is made, plot- Other Approaches
ting the measured responses would The majority of the other approach-
nB
produce a curve similar to that shown es to calculating cl values are similar
in Figure 2. The mean value of the re- E (xgj -

xb) 2 to the IUPAC model in that sB and k

j=l__
sponses, yu, occurs at the center of the (4)
=
sB2 factors are involved. However, it is be-
The curve is symmetric around (nE 1) -

curve. cause of these terms that trouble may

yu and extends outward in units
of for ng observations. Because a finite be encountered when cl values are
standard deviation, a. Since this curve small number of blank readings is nor- used as a basis for the comparison be-
includes all x values that could be ob- mally taken, e.g., ng 20 or greater,=
tween procedures, methods, or instru-
tained from the procedure for the sb must be used instead of ctb- If a ments. The most widely debated of
sample, the area under the curve can sample of this size is used, sb can be the two factors has been the choice of
be expressed in terms of probability, P considered to be a reasonable indica- a value for k. Kaiser was perhaps the
(i.e., there is a 100% chance, P 1, =
tor of erg. If the random error follows a first to stress the use of k 3 for cl
=

that a measured x value would fall normal distribution, a plot of these re- values (3, 4). This value has also been
somewhere under the curve). sponses (frequency of occurrence vs.
agreed upon by other authors (9, 13),
The relationship between area and xb values) would resemble Figure 2. by IUPAC (1, 14), and by the ACS (2).
probability can be measured to esti- The probability that the smallest dis- A value of 2 for k had been initially
mate the chance that a newly mea- cernible analytical signal, xl, can be suggested (8) but this value corre-
sured x value, xg, would be a certain measured and not be a random fluctu- sponds to a 97.7% confidence level for
number of standard deviation units ation of the blank is dependent upon normal distribution and 75% for a
away from the mean response, yx. In how many standard deviation units xl nonnormal distribution of measure-
Figure 2, xE is shown to the right of yu is from xg. If xl is 3sb away from xB, ment error.
and can be measured to be k<r away the area to the right of xl is no less Although the use of k 3 instead of
=

from n. By dissecting the curve with a than 0.0013. Thus, there is a 0.13% k 2 slightly increases the cl value, it
=

line drawn at this x value, the area to chance that a signal measured at xg or is clear that cl values must differ by a
the right of the line, a, is the probabil- greater would be the result of a ran- factor of three for the values to be sig-
ity that xE > (m + kcr). This chance, dom fluctuation of the blank signal. nificantly different. Nevertheless, fac-
which is represented by the pink shad- This small chance of error can then tors of less than three have been com-
ed area, can be determined from fulfill the requirement of a reasonably monly used for comparison purposes.
certain signal. In order to minimize confusion,
area =
(2) In defining cl, IUPAC states that IUPAC suggested that xl values be re-
ibX~exp(::r)dk xl =
xB + ksB (5) ported in all literature with their k
where the xE value is kc away from n, value, xg(k—si* It would be extremely
(i.e., (xE n)/a k).
-
= where k is a numerical factor chosen in useful to go one step further and in-
This illustration can also be used to accordance with the confidence level clude the k values when cl values are
aid in the explanation of the smallest desired. The cl is a function of xl and reported, CL(k=3)- This change would
detectable signal, xl, in the IUPAC therefore be beneficial because cl values are
definition. When the determination of more commonly reported than xL
(xL -

xb)
a limit of detection is performed, cl =-
_

(6) values.
m
blank measurements, xg, are normally A problem encountered in the com-
taken. The question, however, is “how where m is the analytical sensitivity. parative use of cl values is the use of
well are these xg values known?’’ A Because the mean blank reading, xB, the standard deviation of the mean, sB
mean value of the blank responses, xg, is not always 0, the signal must be (11), the pooled standard deviation, sp
can be calculated as background corrected. By substituting (7, 12), or the relative standard devia-
"B
Equation 5 into Equation 6, Equation tion (RSD) (7, 9). Although each of
7 is obtained these standard deviation expressions
E xBj is important and has its place in ana-
xg = J—-- (3)
nB
(7) lytical chemistry, the use or misuse of
these expressions in cl calculations
and the standard deviation as This definition of cl can be illustrated may result in significant deviation

714 A • ANALYTICAL CHEMISTRY, VOL. 55, NO. 7, JUNE 1983

 cl =
k P — =
k(RSO) ^ (13)
3sb 3sb Ub) W m

In many ICP determinations, the

RSD has been measured to be 0.01 (as
determined by emission flicker noise).
Unfortunately, many ICP cl values in
papers and current manufacturers’
bulletins have been calculated with
k = 2. Inserting these values into
Equation 13 yields
Xb
cl —
0.02 (14)

A problem that may be encountered

with this approach is that the RSD for
Figure 4. Normal distribution curves for the blank observation, xa (pink line), the
some spectrometric detection systems
limit of detection, xu (blue line), and the limit of identification, X| (purple line)
Here all standard deviations follow a normal distribution and are equal may be >0.01. Before using this meth-
od, an analyst should reconfirm the
validity of RSD < 0.01. Failure to
from the IUPAC model. measure and correct for high RSD val-
RSD (11)
The standard deviation of the mean Xb ues could result in erroneous cl val-
ues. An attempt to compare an ICP cl
(or the standard error), sb, is ex-
This relationship is useful in ex- value obtained from this method and
pressed by
pressing the “concentrational limit of an IUPAC cB value for the same anal-

SB =
SB
(8) determination,” cb (7). Here, the mea- ysis procedure would be meaningless.
(nB)1/2 sure xd corresponding to cd is defined A well-based but seldom used con-
as cept in the calculation of detection
This value is calculated by dividing limits is the limit of identification, cB
the standard deviation of the blank ko sb xd =
(12) as described by Boumans (7) or the
measures, sB, by the square root of the where sB is the standard deviation of limit of guarantee for purity, eg, de-
number of blank observations, nB. the blank (not the RSD), and kp, is a scribed by Kaiser (3). These concepts
Usually when sp values are used, the k confidence factor that is defined by are essentially the same and are based
factor is replaced by a t distribution the reciprocal of the magnitude of the on the idea that the lowest statistical-
value. Even though the t values reflect
RSD value. For a maximum allowed ly discernible signal should be
the problem of defining a standard de- RSD of 5%, ko 20 and for a RSD of
=
viation for a finite number of observa- 10. xi =
xL + 3si (15)
10%, kD =

tions, 3sb > ta SB/(nB)1/2 for a = 0.005 Another limiting expression involv-
and five observations (four degrees of where si is the error associated with the
freedom). This inequality greatly in- ing RSD values involves the rearrang- measurements at the xi level, or
creases as nB increases. If a large num-
ing of Equation 7 (9). If the back-
ber of observations are taken on the ground signal is included in the nu- xi =
xl + 3sb (16)
merator and denominator of Equation
blank, say 30, the use of the t and sb 7,then if si =
sB, This concept is illustrated in
will reduce the value of cl by a factor
of six from the IUPAC model.
The use of the pooled standard de-
viation includes the number of blank
measurements as well as sample mea-
surements. Generally, the pooled stan- Table I. Statistical Expressions for the Slope and Intercept
dard deviation, sp, is expressed as

SS2 SB2' 1/2 ^xx bscx^ 1/2

Sp = | (9) n-2
ns nB
n \2
where ss is the standard deviation of a E cjl
"X, -

sample measurement and ny is the cj2) (Exit2 1=1 /

number of sample observations. If sg
=
sB, then Equation 9 reduces to
1!=!_±«) fEcA
1/2
—
m _

« n n 2
Sp -

SB (10)
ns nB \ l E Xj
/
For ng 1, sp is essentially the same
= Sxx ~

E x2
!j=i rr
as 3b- However, if ns and nB are both Sm
(Sec)1
large, Sp < sg. Again, the use of a large
number of observations as mentioned
above would cause a significant reduc-
tion in the cl value as compared to the
IUPAC model.
(Mfv
(ns^)1'2
Another common practice involves
the use of the RSD, which is defined
as

716 A * ANALYTICAL CHEMISTRY, VOL. 55, NO. 7, JUNE 1983

 Figure 4. When measurements are
made for a sample, the x values ob-

NEED tained should follow normal distribu-

tion around a mean value. If a sample
ANSWERS? were measured to have a mean at the
xl value, the distribution of these x
values around xL would resemble the
blue line in Figure 4. One-half of the
time the measurement would fall
Finding the right below the xL value and could not be
considered a true signal according to
chromatography product the IUPAC definition. To avoid this
can be tough, but we’ve high probability, the limit of xj is set cL cR
at 3si away from xl. This level is cho- c
made getting the sen so the areas a and 0 are equal. If
Si
=
sb then this level is 3sb away from
answers easier xl or 6sb away from xg. By using
Figure 5. Analytical calibration curve
of signal, x, vs. concentration, c, show-
. . .

k = 3, the area of the xi distribution

curve below xl, 0, is no less than ing graphical approach to limit of detec-
tion calculation
0.0013. Thus, there is a 0.13% chance
White dashed lines are the limit of error bars. cL
that an x value measured at xj would is obtained Irom xL if no error occurs In slope. cr,
fall below the xl limit and not be con- is obtained from xL if slope error is significant
sidered as a true signal.
This idea of further statistically
separating the blank measurements dard deviation of the slope, sm, in the
distributions and true signal distribu- Cl expression. The second method, a
tions has been proposed by the ACS
Subcommittee on Environmental An-
propagation of errors approach, con-
siders the standard deviation of the
alytical Chemistry and has been concentration, sc. This value is calcu-
termed the limit of quantitation (2).
APPL1EI Since the numerical significance of the
lated by including the standard devia-
tions of the blank, slope, and intercept
SC1ENC analyte concentration increases as the
analyte signal increases above xL, a
in the equation. The statistical expres-
sions for these values are listed in
minimum criterion, representing the Table I.
ability to quantify the sample, can he Although these models require ad-
established reasonably far way from ditional calculations, most linear re-
xp. This criterion, called the limit of gression analyses are performed using
quantification (LOQ), is lOer away calculators or microcomputers. With
from xb For limit of detection work, additional programming, these calcu-
a = sr. Samples that are measured as
lations can be easily performed, allow-
AND THEY4 having a signal, x, where x > 10sb are ing more accurate determinations of
.
..'.. termed to be in the region of quantita- cl values to be made. These values
tion while samples where 3sb € x Y may also be used for a truer compara-
10sb are termed to be in the region of tive look at the ability of an analytical
detection. method or instrument to quantify
By setting the quantitation level as trace elements (or compounds) in a
10sb or the identification limit as 6sb, sample.
a much higher probability is afforded
that the sample signal is not just a
random fluctuation of the blank.
However, when making comparisons Graphical Approach
using LOQ or ci to IUPAC cl values, To obtain a more reliable cl value,
the analyst must bear in mind the dif- the m value should be expressed as a
ference in the k factors for each limit. confidence interval m ± t„sm, where
s,„ is the standard deviation of the
Methods Involving slope and t„ is a t distribution value
chosen for the desired confidence
Analytical Sensitivity Error level, cv, and the degrees of freedom, u.
The previous models for calculating The insertion of this interval into
detection limits consider the error in Equation 7 produces
the blank measurements. These mod-
els also consider the analytical sensi-
well-defined value. In ksB
tivity, m, as a CL (17)
m ± t(rsm
practice, however, m may have signifi-
cant error due to nonlinearity in the
calibration curve, or measurement er- The effect of the inclusion of the con-
rors. fidence interval can best be seen by
The following proposed detection referring to Figure 5. The error bars
limit approaches include errors associ- (confidence interval) generated
ated with measurements of the analyt- around the regression line are indicat-
ical sensitivity. The first method, a ed as white dashed lines. Because of
CIRCLE 3 ON READER SERVICE CARD graphical approach, includes the stan- error in the slope, three concentration

718 A • ANALYTICAL CHEMISTRY, VOL. 55, NO. 7, JUNE 1983

 values are found for a given xl value.
When t„sm 0, (as in Equation 7), the
=

value of cl is obtained. However, the

reduced analytical sensitivity yields a
value, cr, to the right of Cl. If
HIGH PURITY ACIDS m » t(,sm, then the cr will not be sta-
tistically different from cl. For the sit-
uation when t„sm is not sufficiently
small as compared to m, a cr may be
ACS REAGENT ACIDS
substantially larger than cl- There-
fore, only the larger value should be
REDISTILLED ACIDS used in reporting a limit of detection.
An important consideration in the
Produced by taking ACS Reagent use of this model is the choice of a
Grade acids and redistilling them t value. The recommended k value of 3
in glass. This distallation yields involves a 99.87% confidence level.
high purity acids with very low or Therefore, the t should reflect a simi-
absent trace elements. lar level. An a = 0.0005 results in a
confidence level of 99.9% for a two-
tailed t distribution. The degrees of
DOUBLE-DISTILLED ACIDS freedom, v, are n 2 for a linear re-
—

Manufactured by taking the gression model. The n value used

already very pure redistilled acids should be the number of points used
and passing them through vycor / to prepare the calibration curve, i.e.,
Quartz stills twice. This process each point consists of a mean measure
removes virtually all volatile and and a concentration value.
non-volatile impurities.
For details on the full line of GFS Chemicals High Purity Acids, circle the Propagation of Errors Approach
reader service number below.
In the second approach used to
evaluate Cl, the error in the intercept
term, i, as well as the error in m are
considered. To include these errors,
P.O. BOX 23214 Columbus, Ohio 43223 (614) 881-5501
Equation 1 must be rearranged to
CIRCLE 99 ON READER SERVICE CARD x —
1
c =-
(18)
m

From Equation 18, any value of x can

be related to a concentration using the
m and i values obtained from the lin-
ear regression model. The i term is

Were big usually neglected in most cases and

assumed to be 0 because the analytical
measures are background subtracted.
But in most linear regression analyses,

in small weighs. i ^ 0. If a true reflection of the error

in the determined concentration is
sought, the error in m and i must be
Sensitivity lo O.l ug. Simple to operate. Auto calibration. BCD included in Equation 7.
The contribution of each term to
output. Three models. Portable or remote weighing.'Hie the total error may be found by taking
Cahn Series 20 Mien 'balances. Wriie lor our live eaialog u >da\: the first derivative of c with respect to
Cato Instalments, Inc., l62(F South IflJMUlil each term:
Carmcnita Koad. Ccrriios. (A uirnl.
WIkhv weigbins’ a little means a lot.

A B TARF

Taking the designated derivatives and

the square root gives

(20)

and finally combining like terms

yields

(21)
CIRCLE 33 ON READER SERVICE CARD

720 A « ANALYTICAL CHEMISTRY, VOL. 55, NO. 7, JUNE 1983

 poorly defined i value and a well-de-
fined m value; C, Cu(I) fluorescence
Table II. ICP-Excited ICP Fluorescence Data data where there is nonlinearity in the
calibration curve resulting in ill-de-
(A) Ca(ll) (B) Ca(ll) fined m and i values; and D. Co(II)
c (ppm) X c (ppm) X fluorescence data where extreme non-
0 0 0 0 linearity in the calibration curve re-
0.1 402 10 67000 sults in severe errors in both m and i
1 6750 100 718000 values. The limit of detection values,
10 67000 1000 7600000 cp(k-r..i, for the three methods are tab-
m =
6710 =
101 m = 7610 13700
=
ulated in Table III. Here the cm m =
Si Si
values are reported only to one signifi-
Sm
=
20 sB
=
5.4 sm =
27 SB 5.4 =

cant figure, as all cL values should be.

i =
-79 i =
16600
The subscript numbers representing
the second significant digit in cases A
(C)Cu(l) (D) Co(ll) and B are included only for compara-
c (ppm) X c (ppm) X tive purposes.
0 0 0 0 For case A, the IUPAC model and
1 55 1 796 the graphical approach agree well.
10 1810 10 2800 Only the propagation of errors model
100 34500 100 60700 shows a significantly higher Cpu, :t)=

m —
350 Si
=
545 m 614=
Si
=
1200 value; this deviation is the result of
Sm *“10 sB = 3.5 sm = 24 sB = 97.6
the error associated with the intercept
i = -628 i = -980
value.
The values obtained in case B em-
phasize the problem of an ill-defined,
nonzero intercept. This problem is the
direct result of constructing calibra-
tion curves for detection limits when
the lowest point of the calibration
Table III. curve data is considerably removed
cMk=3) Values (ppm) from the CL(k ?,) value. Only the
=

propagation of errors model accounts

3sb 3sb ,|*,**,+ for this error, while the other two
methods indicate an erroneously low
m m
CL(k -
o value. Although there are no
(A) Ca(ll) 0.0024 0.0027 0.05 set guidelines for constructing calibra-
(B) Ca(ll) 0.002, O.OO24 5 tion curves, the third approach clearly
0.03 0.3 5 illustrates the problem of sampling
(C) Cu(l)
6
too far away from the limit of detec-
(D) Co(ll) 0.5 -2
tion.
In some instances, calibration
curves may not be linear. Although
the linear regression procedure will fit
Equation 21 allows the determination In the event that no significant a line through the data, the resulting
of standard deviation in a c value cal- error occurred in the slope, Equation m value is by no means a “true” repre-
culated from any x value. 21 reduces to sentation of the analytical sensitivity
In the case of limits of detection, at all concentrations. Nonlinear cali-
the Cp value is actually a confidence cp =
(24) bration curves generally produce sig-
m
value expression of how well the blank nificant sm, Sj, and i values. Case G
is known. If Equation 7 is reconsid- If the error in the intercept, sp is suffi- represents such conditions. The
ered, it can be written as ciently small, Equation 23 reduces to IUPAC model results in a gp(p ;fl of
=

Equation 7, which is the IUPAC defi- 0.03 ppm. Repeating the cp calcula-
cL =
ksc (22) nition of the limit of detection. tion using the graphical approach re-
sults in a value of 0.3 ppm. Finally al-
where the sB/m term describes the Evaluation of Approaches
lowing m and i errors to be included
error in terms of c (if c Cp). Equa- ~
The IUPAC, graphical, and propa- results in a cp(k » of 5 ppm for the
tion 21 could now be used to evaluate gation of errors models will be applied propagation of errors method. The
sc where x is the blank signal, xb, and to four different sets of experimental propagation of errors Cp value is 170
sB is substituted for sx. By measuring data to show the effect of certain ex- times the IUPAC cp<k m value.
»

xB, Sb, and calculating m, i, sm, and s„ perimental conditions on the estima- The problem of nonlinearity can be
the value of sc can be determined. tion of cp values (Table II). The data further emphasized by considering
In most determinations, the data in Table 11 have been taken from a re- case D. Here, the errors associated
are background corrected, that is, cent paper on ICP-excited ICP fluo- with m and i are greater; however, the
0- Substituting the above mea- detection limits (15). The
xb large value of sm results in a special
=
rescence
sured and calculated values in Equa- four sets of typical experimental con- problem with the graphical approach.
tions 21 and 22, the expression for cp ditions are: A, Ca(II) fluorescence If t„sm > m, then the concentration
is further simplified to data which have well-defined m and i value for the limit of detection can
1/2 values; B, Ca(II) fluorescence data even be negative. Such negative values
k SB* + Sp +
—-

Sm* where sB is essentially the same as in are the direct result of the graphical
(mj A, but the calibration curve data are model not being statistically valid. Al-
taken far away from eg resulting in a though the graphical model is easier to

722 A * ANALYTICAL CHEMISTRY, VOL. 55, NO. 7, JUNE 1983

 use than the propagation of errors ap- References
MODERN THIN-LAYER proach, the former may give erroneous (1) “Nomenclature, symbols, units and
results as seen in Case D. Thus, the
CHROMATOGRAPHY graphical method could be used only
their usage in spectrochemical analysis—
II, ” Spectrochim. Acta B 1978,33B, 242.
is an efficient analy- for approximating cl values. The de- (2) “Guidelines for Data Acquisition and
Data Quality Evaluation in Environmen-
tection limit using the propagation of
tical method that in errors approach is calculated to be
tal Chemistry,” Anal. Chem. 1980,52,
2242.
many cases offers a 6 ppm. Correspondingly, the detection
limit for the IUPAC model is 0.5 ppm,
(3) Kaiser, H. “Two Papers on the Limit of
Detection of a Complete Analytical Pro-
most attractive per- differing by a factor of 12 between the cedure”; Hafner: New York, 1969.
(4) Kaiser, H. Anal. Chem. 1970,42, (2),
formance/operating two methods. 24 A.
(5) Kaiser, H. Anal. Chem. 1970, 42, (4),
cost ratio. Conclusions 26 A.
(6) Kaiser, H. Spectrochim. Acta B 1978,
Based on the above considerations, 33B, 551.
the graphical approach to cl should (7) Boumans, P.W.J.M. Spectrochim.
ActaB 1978, 33B, 625.
not be used. The IUPAC approach is (8) Boumans, P.W.J.M.; deBoer, F. J.
valid only if the major source of error Spectrochim. Acta B 1972,27B, 391.
is in the blank, i.e., sb2 » Si2 or sm2. (9) Boumans, P.W.J.M. “Lines Coinci-
Therefore, the IUPAC approach in dence Tables for Inductively Coupled
Plasma Atomic Emission Spectrometry”;
most cases gives artificially low values
Pergamon; Oxford, 1980.
of cl- The propagation of errors ap- (10) Currie, L. A. Nucl. Instrum. Methods
proach is certainly the most liberal ap- 1972, 100, 387.
proach and will give values of cl con- (11) Glaser, J. A.; Foerst, D. L.; McKee,
G. D.; Quave, S. A.; Budde, W. L. Envi-
sistent with the reliability of the blank ron. Sci. Technol. 1981,15, 1426.
measures and the signal measures of (12) Peters, D. G.; Hayes, J. M.; Hieftje,
the standards. G. M. “Chemical Separation and Mea-
We recommend that analysts report surements”; Saunders: Philadelphia,
limits of detection using the IUPAC 1974; Chapter 2.
(13) Winefordner, J. D. “Trace Analysis:
approach with k 3 (cyk 3))- The
=
=
Spectroscopic Methods for Elements”;
use of the propagation of errors ap- Wiley: New York, 1976; Chapter 1.
proach is also recommended because (14) “Nomenclature, symbols, units, and
their usage in spectrochemical analysis—
errors in the analyte measurements
III, Spectrochim. Acta B 1978,33B,
”

can be incorporated into the cl value. 248.

(15) Kosinski, M. A.; Uchida, H.; Wine-
Of course, reliability By adopting these approaches, mean-
fordner, J. D. Anal. Chem. 1983,55, 688.
ingful comparisons of analytical meth-
and accuracy depend ods and instruments based on cl val- This work was supported by AF-AFOSR-
on the appropriate ues can be made. F49620-80-C-0005.

system of compatible
instrumentation.
The catalogue TL-10 »lnstrumental
Thin-Layer Chromatography«
contains detailed information
on CAMAG's complete range of
TLC instrumentation.
Methodological explanations
together with an easy-to-use
guide of the instruments and
their compatibility will help you
to compose the COMPLETE
SYSTEM that suits your James D. Winefordner (/.) is a gradu- ment of new instrumental methods
requirements best. ate research professor in chemistry at based upon new optical and detection
the University of Florida. His re- approaches; and applications of ana-
Ask for this 56-page cataloque, interests include diagnostical mea- lytical techniques.
surements of flames and plasmas;
or even better, laser-excited luminescence, pho- Gary L. Long is a postdoctoral asso-
have the CAMAG product toacoustic and photoionization of ciate in the department of chemistry
specialist (of the CAMAG molecules in gases and liquids; laser- at the University of Florida. He re-
distributor in your territory) excited fluorescence of atoms in ceived his BS in chemistry from Wake
discuss your requirements with flames and plasmas; development of Forest University in 1978 and his
him. trace analytical methods for atoms PhD in analytical chemistry from
and molecules based upon lumines- North Carolina State University in
CAMAG has agents in almost all cence, chemiluminescence, pho- 1982. His research interests include
countries. toacoustic, and ionization methods; ICP-excited ICP fluorescence spec-
specialized methods for molecules trometry, laser-excited flame fluores-
based upon synchronous lumines- cence spectrometry, and the investi-
_£A/\A/\& cence, room-temperature phospho-
rimetry, energy transfer lumines-
gation of chemical and physical inter-
ferences in flame and plasma spectro-
Sonnenmattstr. 11 •
CH-4132 Muttenz/Swltzerland
Tel. 0 61-61 34 34 Telex 62 649
•
cence, and time resolution; develop- metric analyses.
CIRCLE 34 ON READER SERVICE CARD
724 A • ANALYTICAL CHEMISTRY, VOL. 55, NO. 7, JUNE 1983