Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: The work presented herein suggests that the limit of quantitation concept may be rendered substantially less
Received 24 November 2013 ambiguous and ultimately more useful as a figure of merit by basing it upon the significant figure and relative
Accepted 29 March 2014 measurement error ideas due to Coleman, Auses and Gram, coupled with the correct instantiation of Currie's
Available online 8 April 2014
detection limit methodology. Simple theoretical results are presented for a linear, univariate chemical measurement
system with homoscedastic Gaussian noise, and these are tested against both Monte Carlo computer simulations
Keywords:
Limit of quantitation
and laser-excited molecular fluorescence experimental results. Good agreement among experiment, theory and
Quantitation limit simulation is obtained and an easy extension to linearly heteroscedastic Gaussian noise is also outlined.
Detection limits © 2014 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.sab.2014.03.012
0584-8547/© 2014 Elsevier B.V. All rights reserved.
70 J. Carlson et al. / Spectrochimica Acta Part B 96 (2014) 69–73
and Vos detection limit. Subsequently, Voigtman demonstrated, in a where zp is the critical z value for probability p of false positives [10].
series of publications [10–16], the optimum implementation of Currie's Typically, p ≡ 0.05, so that zp = z0.05 ≅ 1.6448536. As a result,
program and it has since become clear (vide infra) that it is the relation-
ship between the LOQ and the Currie decision level, not detection limit, z0p 39:2
X Q ¼ 20 X ≅ X : ð6Þ
that is of primary utility. zp C zp C
where α is the true intercept, β is the true slope and Y is the true Values of zp are easily found using Microsoft Excel, i.e., zp =
response. Systematic error is assumed to be zero. The noise is additive, −NORMSINV(p), or via standard tables.
Gaussian white noise (AGWN) on Y and is homoscedastic, i.e., σ(βX) ≡ The experimental, content domain decision level, x C , is given
σ0, where σ0 is the population standard deviation of the noise on the by [10,11]
blank. As a consequence of the noise, a single measured response, yi,
t p η1=2 s0
at any arbitrary value of X, is simply a random sample from a Gaussian xC ≡ ð8Þ
(aka “Normal”) distribution centered at α + βX and having population b
standard deviation σ0. If ideal blank subtraction is performed, to obtain
where b is an experimentally determined, unbiased estimate of β,
the net response, then this simply means that the errorless value α is
s 0 is the sample standard deviation determined with ν degrees
subtracted from yi, so the net response, yi − α, is a random sample
of freedom (dof) and tp is the critical t value for probability p of
of a Gaussian distribution centered at βX, with population standard
false positives and ν dof. Then the corresponding experimental,
deviation σ0. The shorthand notation for this is (yi − α) ∼ N: βX,σ0,
content domain LOQ, denoted by xQ, is
where “~” means “is distributed as”.
If α is unavailable, as is typically the case, then blank subtraction may 0 1=2 0 1=2
t p η s0 t p η s0 t η1=2 s0
be performed by subtracting α ^ , an unbiased estimate of α. Then ðyi −α
^Þ xQ ≡ ¼ 20 ¼ 20 0:025 ð9Þ
0:05b b b
∼N : βX; σ d , where σd is the population standard deviation of the
difference. The relationship between σd and σ0 is σd = η1/2σ0, where where t p′ is the critical t value for 95% confidence (analogous to
the factor η1/2 is of the general form: z0.025 = 1.96) with v dof. Combining Eqs. (8) and (9) then yields
1=2
η
1=2
≡
1
^
þ error termðsÞ due to α ð2Þ t 0p
M xQ ¼ 20 x ð10Þ
tp C
with M being the number of future blank measurements. Optimally, M is with ν dof for both critical t values. In the experimental, net response
defined as unity, so that η1/2 = 1 only if ideal blank subtraction is domain, yC = bxC and yQ = bxQ. Therefore
performed and is otherwise greater than unity.
For maximum generality, it will be assumed that non-ideal blank t 0p
^ is used for blank subtraction even if β
subtraction is performed, i.e., α yQ ¼ 20 y : ð11Þ
tp C
and σ0 are known. Therefore, a single net response at βX will be a
random sample from N: βX,σd. As is well known, there is a 95% probabil-
Values of t p are easily found using Microsoft Excel, i.e., t p =
ity, over many repetitions of the process, that the net response will be in
TINV(2p,ν), or via standard tables.
the (central) 95% confidence interval (CI) defined by βX ± zp′σd, where zp
Fig. 1 collects together Eqs. (6), (7), (10) and (11), and also gives the
′ ≡ z0.025 ≅ 1.959964. Following the lead of Coleman et al. [8], the abso-
Currie decision level expressions (“C” subscripts) and detection limit
lute measurement error is arbitrarily defined as zp′σd, i.e., the half-width
expressions (“D” subscripts) in all four detection quadrants. The critical
of the 95% CI, and the maximum relative measurement error, RME(X), is
values zq and tq in the detection limit expressions are for probability q
of false negatives. These are correct and unbiased for all degrees of
z0p σ d 1:96σ d
RMEðX Þ ≡ ≅ ð3Þ freedom, including ν = 1. Note that all of the theoretical domain expres-
βX βX sions are errorless real numbers that require population parameters
(i.e., true values, since systematic error is assumed to be negligible),
while still remaining within the 95% CI. Note that the RME is simply
while the expressions in quadrant 2 are χ variates and those in quadrant
defined as the half-width of the confidence interval, divided by its center
4 are modified noncentral t variates [10,11].
value. As noted above [8], in order for the LOQ to guarantee at least
one significant figure, RME(X) must not exceed 0.05. Therefore, the
3. Computer simulation
theoretical, content domain LOQ, denoted by XQ, is defined as that
value of X such that RME(XQ) ≡ 0.05. Hence
The theory presented above is very simple and perfectly suited
0 0 0 1=2 to testing via computer simulation. Accordingly, the following ideal
zp σ d zp σ d zp η σ 0 η
1=2
σ0 model parameters were used, ignoring units: α ≡ − 0.05, β ≡ 3.85,
XQ ≡ ¼ 20 ¼ 20 ≅39:2 ð4Þ
0:05β β β β σ0 ≡ 0.03, p ≡ 0.05, zp ≅ 1.644854, p′ ≡ 0.025, zp′ ≡ z0.025 ≅ 1.959964,
M ≡ 1 future blank replicate, and α ^ the sample mean of N = 7
since σd = η1/2σ0. However, the theoretical, content domain Currie i.i.d. blank replicates. In this case η 1/2 ≡ [M − 1 + N− 1] 1/2 =[1
decision level is defined as + (1/7)] 1/2 ≡ 1.069045, so that, from Fig 1, XC = zp η 1/2σ 0 /β ≅
0.013702 and XQ = 20(zp′/zp)XC ≅ 0.3265386.
zp η1=2 σ 0 Fig. 2 shows two different simulation programs designed to test the
XC ≡ ð5Þ
β model and its theoretically predicted XQ. The upper half of Fig 2 shows a
J. Carlson et al. / Spectrochimica Acta Part B 96 (2014) 69–73 71
Monte Carlo program implemented using the author's free LightStone events above the 95% CI was 24 948.6 ± 159.2, for ten sets of 1 million
component libraries, which are an add-on for the commercial ExtendSim® steps each. The ±term is one standard deviation. For the Mathematica®
simulation software. This combination has been used extensively in prior program in Fig 2, likewise with X ≡ XQ ≅ 0.3265386 and ten runs of 1
work [10–16] and a free demo copy of ExtendSim® is available from million trials each, the results were 24 982.0 ± 140.3. Each program
Imagine That Inc. (www.extendsim.com). The lower half of Fig 2 shows in Fig 2 was easily modified to test for numbers of events falling below
a Monte Carlo Mathematica® (v. 9) program with exactly the same func- the 95% CI. The results were 25 033.3 ± 108.7 (LightStone/ExtendSim®)
tionality. It may be run on the fully functional 30 day free trial version of and 24 954.1 ± 165.4 (Mathematica®). In all cases, the obtained results
Mathematica® available from Wolfram Research (www.wolfram.com/). were statistically the same as the theoretical expectation value.
For each of 1 million steps (LightStone/ExtendSim®) or trials
(Mathematica®), a new blank-subtracted variate is generated at the 4. Experiment
user-specified value of X. The total number of events that fall above
the 95% CI centered on βX is recorded. Since the 95% CI is a central The laser-excited molecular fluorescence experimental system was
CI, the expectation is that 25 000, i.e., the upper-tail 2.5% of 1 million exactly the same one as used previously [14,15], with the same laser
total events, will fall above the 95% CI. For the LightStone/ExtendSim® dye analyte: rhodamine 6G tetrafluoroborate (R6G) in ethanol. The
program in Fig 2, with X ≡ XQ ≅ 0.3265386, the mean number of standards were either the exact same solutions previously prepared,
Fig. 2. Monte Carlo programs for XQ tests: LightStone and ExtendSim® (upper) and Mathematica® (lower).
72 J. Carlson et al. / Spectrochimica Acta Part B 96 (2014) 69–73
used [14] and then carefully stored for possible future use, or were
prepared from them by one-step dilutions using digital pipettes
and pure ethanol. With the exception of using more standards,
all other conditions were as previously reported [15], with noise
attenuator setting of 1.00, i.e., 10% of full scale added noise. Each
analyte was measured 12 001 times, so that data were collected at
a rate of 1000 samples per second for 12 s.
The analyte solution concentrations (Xi), in mg of R6G/100 mL
ethanol, are reported in Table 1, along with their mean responses (yi ),
standard deviations and approximate RMEs, i.e., RME/η1/2. The mean
of the noise standard deviations was 0.0329589 V, which was taken
as the estimate, σ ^ 0 , of the population standard deviation of the
homoscedastic noise. Then the approximate RMEs were computed,
ignoring the η 1/2 factor in Eq. (3), as 1:959964σ ^ 0 =ðyi −aÞ, where
a = − 0.037961 V (vide infra).
Fig. 3 shows a plot of the mean fluorescence responses (yi), versus R6G
concentrations (Xi). The ordinary least squares fit yields a = −0.037961 V,
b = 3.88777 V/(mg/100 mL), sa = 0.00835 V, sb = 0.0436 V/(mg/100 mL),
and adjusted R2 = 0.9981. The η1/2 blank-subtraction factor is estimated Fig. 3. Molecular fluorescence calibration curve for the data in Table 1. The error bars are
±one standard deviation.
by pooling σ ^ 0 and sa:
h i1=2
1=2
η ¼ ^ 20 þ s2a =σ
σ ^ 20 ≅1:0316 ð12Þ quantitation purposes. Second, if false negatives are of no concern,
then the decision levels in Fig. 1 also serve as detection limits, and
so the approximate RMEs in Table 1 are about 3% low. the expressions with subscripts “D” are not used. This is the pre-
Fig. 4 shows a plot of the RME, with the above η1/2 factor included, Currie detection limit methodology underlying obsolete expressions
versus X. Also shown is an inset plot of the inverse relationship such as LOD = ksblank /slope. Third, since the LOQ is related to the
curve-fit between RME and X: decision level, which depends on noise on the blank rather than
noise at X values above 0, extension to heteroscedastic noise is
RME ¼ 0:017179=X ð13Þ straightforward. For example, if the noise was linearly heteroscedastic,
i.e., σ(βX) ≡ σ0 + μβX, then
so that RME ≡ 0.05 implies XQ = 0.344 mg R6G/100 mL ethanol.
The predicted values of XC and XQ for the experiment were XC = z0:025
XQ ¼ XC ð14Þ
1.644854 × 1.0316 × 0.0329589/3.88777 ≅ 0.0144 mg/100 mL, zp 0:05−z0:025 η1=2 μ
from Eq. (5), and, from Eq. (6), X Q ≡ (20 × 1.959964/1.644854)
XC ≅ 0.343 mg/100 mL. Note that from Eq. (3), with β estimated by b, where μ is the heteroscedasticity, i.e., asymptotic RSD, and the related
^ 0 =b ¼ 0:017141 mg/100 mL, in reasonable agreement
z0:025 η1=2 σ expressions for YQ, xQ and yQ are trivially found. Note that if μ = 0, the
with the obtained curve-fit value of 0.017179 mg/100 mL. system is homoscedastic and Eq. (14) reduces to Eq. (6). Eq. (14) also
shows that, for XQ to exist, μ has an upper limit: μ b 0.05/z0.025η1/2.
5. Discussion and conclusions Therefore, μ must be less than about 2.5% for a linearly heteroscedastic
system if RME(XQ) is defined as 0.05.
The LOQ methodology suggested above has several advantages. Mermet was right that “any LOQ can be selected” [6], but the authors
First, it assures the user that there is 95% probability, over many repeti- believe that LOQ may never step out from under the shadow of the
tions of the process, that analyte present at the LOQ will be quantifiable detection limit unless it can be seen as fulfilling a fundamentally
with at least one significant figure. The detection limit, in contrast, important and, at present, largely unmet need.
is essentially a qualitative assurance that, at high probability, analyte
is present and has not escaped detection: it is not usable for
Table 1
Concentrations and experimental response data.
Acknowledgments [7] J.-M. Mermet, G. Granier, P. Fichet, A logical way through the limits of quantitation
in inductively coupled plasma spectrochemistry, Spectrochim. Acta Part B 76 (2012)
221–225.
This research is dedicated to the memory of Dr. Elena Nikolova [8] D. Coleman, J. Auses, N. Grams, Regulation — from an industry perspective or
Dodova. relationships between detection limits, quantitation limits, and significant digits,
Chemom. Intell. Lab. Syst. 37 (1997) 71–80.
[9] A. Hubaux, G. Vos, Decision and detection limits for linear calibration curves, Anal.
References Chem. 42 (1970) 849–855.
[10] E. Voigtman, Limits of detection and decision. Part 1, Spectrochim. Acta Part B
[1] L.A. Currie, Limits for qualitative and quantitative determination — application to 63 (2008) 115–128.
radiochemistry, Anal. Chem. 40 (1968) 586–593. [11] E. Voigtman, Limits of detection and decision. Part 2, Spectrochim. Acta Part B 63
[2] P.B. Adams, W.O. Passmore, D.E. Campbell, Paper 14, “Symposium on Trace (2008) 129–141.
Characterization – Chemical and Physical”, National Bureau of Standards, Oct., [12] E. Voigtman, Limits of detection and decision. Part 3, Spectrochim. Acta Part B
1966, 1966. 63 (2008) 142–153.
[3] ACS Committee on Environmental Improvement, Guidelines for data acquisition [13] E. Voigtman, Limits of detection and decision. Part 4, Spectrochim. Acta Part B 63
and data quality evaluation in environmental chemistry, Anal. Chem. 52 (1980) (2008) 154–165.
2242–2249. [14] E. Voigtman, K.T. Abraham, Statistical behavior of ten million experimental
[4] H. Kaiser, Part II quantitation in elemental analysis, Anal. Chem. 42 (1970) 26A–59A. detection limits, Spectrochim. Acta Part B 66 (2011) 105–113.
[5] C.G. Fraga, A.M. Melville, B.W. Wright, ROC-curve approach for determining the [15] E. Voigtman, K.T. Abraham, True detection limits in an experimental linearly
detection limit of a field chemical sensor, Analyst 132 (2007) 230–236. heteroscedastic system. Part 1, Spectrochim. Acta Part B 66 (2011) 822–827.
[6] J.-M. Mermet, Limit of quantitation in atomic spectrometry: an unambiguous [16] E. Voigtman, K.T. Abraham, True detection limits in an experimental linearly
concept? Spectrochim. Acta Part B 63 (2008) 166–182. heteroscedastic system. Part 2, Spectrochim. Acta Part B 66 (2011) 828–833.