QUANTIFYING MEASUREMENT UNCERTAINTY IN ANALYTICAL

CHEMISTRY – A SIMPLIFIED PRACTICAL APPROACH

Thomas W. Vetter

National Institute of Standards and Technology (NIST)

100 Bureau Drive Stop 8393

Gaithersburg, MD 20899-8393

301-975-4123

thomas.vetter@nist.gov

Abstract - Following a simple and practical uncertainty, since it is just an estimate anyway,
approach can minimize the headaches of isn’t it? However, the complete approach shown
calculating the components of measurement here is widely applicable and can minimize the
uncertainty. Basic requirements for planning headaches of calculating measurement
an analysis, realistically estimating the uncertainty. The complete approach can also
magnitude of uncertainty sources, and lead to a more accurate analytical determination.
combining uncertainty components will be A generalized scheme for planning the analysis
presented using examples of analytical and estimating and combining the uncertainty
chemistry methods. sources is followed by examples.

Keywords: cause and effect diagram; 2. Background

combined uncertainty; Kragten spreadsheet;
measurement; quantification; uncertainty Planning an analysis of the
composition of a sample
1. Introduction
The comments that follow are based on the
What follows is a simple and practical approach to assumption that the analyst is interested in
quantify measurement uncertainty, u, based on measuring one or more analytes in a sample that
information gathered from many helpful has been provided to the laboratory. This
resources. To the novice attempting to quantify planning can be applied to a specific one-time
the uncertainty in a chemical analysis, the task analysis or to the general analysis scheme for a
may seem daunting and far from simple because routine testing laboratory with minor modifications
of the multiple new terms, diagrams and made based on the sample, analyte, and
calculations required. It may be tempting to just experience.
hazard a guess as to the magnitude of the
• Select a general analytical method based on helpful, but not always necessary, to draw out a
the accuracy required. simple cause and effect diagram that includes
• Plan the analysis so that all the information of each component of the measurement equation
interest can be extracted from the data to and the major sources of uncertainty for each
reach a valid and repeatable conclusion with a component. For some complicated
minimal expenditure of time and cost. determinations it may be easier to write out first a
• Determine what constitutes a representative separate diagram for each of the measurement
test sample and the number of representative equation components and then combine them into
test samples required. a final simplified diagram. Appendix D of the
• Write out the measurement equation for the Eurachem Guide [3] and an article by Ellison [4]
result (i.e., the equation that explicitly shows provide guidance in the construction of a cause
how the various measured and theoretical and effect diagram. Uncertainty evaluated using
quantities from the experiment are combined statistical methods is referred to as Type A [5, 6].
to obtain the final result) and consider the All other ways of evaluating uncertainty are
factors that can influence variability and bias referred to as Type B [5, 6]. All uncertainties need
of the result. Factors that influence the to be expressed as standard deviations or
variability and bias of the analytical result standard uncertainties. Type B uncertainties are
should be controlled to the extent that a always expressed as probability distributions. The
subsequent repeat analysis of the original equations used to calculate the standard
sample likely will yield a result that agrees uncertainty for three common probability
within the limits of uncertainty. Apparent distribution models (normal, rectangular or
variability should not be controlled to the point uniform, and triangular) are shown in Figures 1 –
that repeated measurements are limited to a 3. After the magnitudes of the significant
single test sample in a short period of time, components of uncertainty are calculated they are
since that limitation likely will produce a result combined and multiplied by an coverage factor, k,
that cannot be repeated at a later date for to determine the expanded uncertainty.
other test samples. As stated by W. J.
Youden, a chemist and recognized expert in
statistical design of experiments, “Repeat
measurements cannot reveal the vicissitudes
of measurement making unless the operator
gives the vicissitudes a chance to occur” [1].
A book by Box, Hunter and Hunter [2] is an
excellent source of additional information on
experimental design.
• Make measurements.
• Estimate the magnitude of each uncertainty
source after all of the measurements in the
analysis plan are completed,

Estimating the magnitude of

uncertainty sources
Determine the major sources of uncertainty that
are associated with each component of the
measurement equation. Focus on the largest
uncertainty sources. Those sources that have a
magnitude of one-third or more of the largest
source should be examined in detail. Less
attention should be spent estimating smaller
sources. The Eurachem/CITAC Guide
“Quantifying Uncertainty in Analytical
Measurement” [3], hereafter referred to as the
“Eurachem Guide”, lists common sources and
values of uncertainty in its Appendix G. It is Figure 1. Normal distribution
 Combining uncertainty components
Although uncertainty components are not errors,
the individual components are combined using
law of propagation of error formulas in which the
combined uncertainty, uc, is the square root of the
sum of the squares (RSS) of the individual
uncertainty components, each scaled to reflect its
potential impact on the final result. Computing
partial derivatives of the measurement equation
with respect to each term in the equation using
calculus is a correct, but sometimes complicated,
way of obtaining the appropriate scaling factors
for combining components. Calculating the
relative uncertainty of the components and
combining them using a simple, unscaled RSS is
easier and sometimes avoids the need for explicit
calculation of the scaling factors from the
measurement equation. However, this method
may result in calculation errors, especially if the
measurement equation is not an exclusive mix of
either only additions and subtractions, or only
multiplications or divisions. In addition,
components of the measurement equation that
may have variables in common (mutually
dependent) may be calculated incorrectly using a
simple RSS. An alternate approach that avoids
the need to explicitly compute partial derivatives
Figure 2. Uniform distribution and minimizes the risk of calculation errors is the
calculation of the combined uncertainty using a
Kragten spreadsheet. A Kragten spreadsheet
(which is a tool for estimating uncertainty based
on calculus) uses a numeric approximation to the
scaled RSS [7] to combine uncertainty
components. Kragten [7] and the Eurachem
Guide [3] provide an alternative and more detailed
explanation than that provided here.

Although the explanation of a Kragten

spreadsheet may give the appearance that using
it is complicated, it is in fact, very straightforward.
In a Kragten spreadsheet the absolute magnitude
of each component of the measurement equation
and its absolute standard uncertainty are entered
into predetermined cells of a spreadsheet. Once
the template for the spreadsheet is set up, the
analyst only has to enter the measurement
equation, components of the measurement
equation, and the corresponding standard
uncertainties. With the exception of Type A
uncertainties, or uncertainties of zero bias
components that do not change the measurement
result (as explained in the examples that follow),
no relative values are added in the applications
Figure 3. Triangular distribution shown here.
 Table 1. Generalized form of Kragten spreadsheet with four components

1 2 3 4 5
1 Value A B C D
2 Uncertainty a b c d
3 A A+a A A A
4 B B B+b B B
5 C C C C+c C
6 D D D D D+d
7 R =1000*A*B*C/D Ra = 1000*[A+a]*B*C/D Rb = 1000*A*[B+b]*C/D Rc = 1000*A*B*[C+c]/D Rd = 1000*A*B*C/(D+d)
8 = R - Ra = R - Rb = R - Rc = R - Rd
9 DIFsum = DIFa+DIFb+DIFc+DIFd DIFa = (R - Ra)2 DIFb = (R - Rb)2 DIFc = (R - Rc)2 DIFd = (R - Rd)2

10 SQRT(DIFsum) = l R - Ra l/SQRT(DIFsum) = l R - Rb l/SQRT(DIFsum) = l R - Rc l/SQRT(DIFsum) = l R - Rd l/SQRT(DIFsum)

11 2 <-- k
12 k *SQRT(DIFsum) <-- k *u c
13 100*k*u c/R <-- relative unc. (%)

The following explanation applies to Table 1 as column 1, rows 3 - 6) are pre-formatted to equal
well as any other Kragten spreadsheet. A the contents of column 1 with the following
generalized form of the Kragten spreadsheet for a exception. Progressing diagonally from column 2,
four-component (A, B, C, D) measurement row 3 to column 5, row 6, and shaded in blue, the
equation with each component having a standard component of that row and the corresponding
uncertainty (a, b, c, d) and a coverage factor of k standard uncertainty of that column are added,
= 2 is shown in Table 1. The result, R, of the once for each component. The measurement
measurement equation is calculated. This result equation is entered in each column of the row
should agree with the determined value. The immediately below the entered components (row
result of the measurement equation is varied once 7) so that its result is determined by the contents
for each component (A, B, C, …) by adding the of the cells immediately above it (in this example,
magnitude of its uncertainty (a, b, c…) to the rows 3 – 6). In this example, once the
value of the component and calculating a new spreadsheet is set up, the analyst only has to
result (Ra, Rb, Rc,…). For each component the enter the measurement equation (row 7), the four
difference between the result, R, and the new components of the measurement equation
result (Ra, Rb, Rc, …). is squared. All of the (column 1, rows 3 – 6), the four corresponding
standard uncertainties (columns 2 – 5, row 2), and
squared differences (DIFa, DIFb, DIFc, …) are
the coverage factor, k (column 1, row 11). The
summed and the square root of the sum is original result and a new result for each
calculated. The value of RSS is the combined component are calculated in row 7, based on the
standard uncertainty. The assumptions and values in rows 3 - 6. Each of the differences in
calculations used to estimate the most significant the result is calculated in row 8. The sum of the
uncertainty components always should be re- squares of the differences is calculated in column
evaluated. 1 of row 9 and each square of the differences
corresponding to its entered component and
Though specific to Table 1, the following standard uncertainty is calculated in the remaining
explanation shows how any Kragten spreadsheet columns of row 9. The combined standard
is structured. The value of each of the four uncertainty, calculated as the RSS, is shown in
components is entered in the section of the first column 1, row 10.
column shaded in yellow (rows 3 – 6). The value
of each of the four standard uncertainties is This general spreadsheet has some additional
entered in the section of the second row shaded features that are not in the originally conceived
in green (columns 2 – 5). The spreadsheet is pre- Kragten spreadsheet [7]. In row 10, columns 2 -
formatted so that the entered component values 5, the magnitude of each component of the
also appear in row 1 above their corresponding standard uncertainty, relative to the combined
standard uncertainties entered in row 2. Each of standard uncertainty, is calculated. The coverage
the rows to the right of the entered values (in factor, k, is entered in column 1, row 11. The
expanded uncertainty is calculated in column 1, extra figures are carried through until the
row 12, and the percent relative expanded calculation of the total uncertainty to avoid
uncertainty is calculated in column 1, row 13. confusion that could be caused by rounding
Other modifications based on the analyst’s needs prematurely.
can be made. Note that it is easy to change any
re-evaluated components and to see the effect of
any changes on the total uncertainty. If additional Example 1. Titration of NaOH with
components were added, the columns and rows KHP
would be expanded once for each addition.
In this example sodium hydroxide (NaOH) is
The Kragten spreadsheet is based on the standardized by titration against standard
assumption that the result, R, changes linearly potassium hydrogen phthalate, KHP. The major
within the interval of R ± R(a, b, c, ...). Subtracting uncertainty sources and their magnitude are
the magnitude of the uncertainties instead of determined.
adding them serves as a check of this linearity. If
the assumption of linearity is correct, the
combined uncertainty does not change
Measurement equation
significantly whether the standard uncertainty is
First the measurement equation is written as
added or subtracted. Linearity should not be a
problem if the interval of R ± R(a, b, c, ...) is relatively
small and the measurement equation is not a (1) cNaOH = 1000 . ((mKHP) . (PKHP)/((MKHP) .
complicated function. Therefore, if the interval of (VT)) [mol L-1]
R ± R(a, b, c, ...) is relatively large (e.g., when the
signal-to-noise ratio of a measurement is small), where,
this approach to estimation of uncertainty may not cNaOH = concentration of the NaOH solution
be valid. [mol L-1]
1000 = conversion factor [mL] to [L]
Two examples of analytical chemistry methods mKHP = mass of the titrimetric standard KHP [g]
will be used to demonstrate the estimation and PKHP = purity of the titrimetric standard given as
quantification of measurement uncertainty. The a mass fraction
-1
first is a titration example taken directly from MKHP = molecular weight of KHP [g mol ]
Appendix A of the Eurachem Guide [3] and the VT = titration volume of NaOH solution [mL].
second is a gravimetry example based on a
recently completed determination at the National
Institute of Standards and Technology (NIST).
Keep in mind that different analysts are likely to
Cause and effect diagram
have different experience and judgement. When
Next, a simple cause and effect diagram is drawn
provided with the same information, they may
(Figure 4) showing the four components of the
model the uncertainty of the Type B components
measurement equation that may contribute to the
of the measurement equation differently, and
uncertainty of the result. The uncertainty of each
make different estimates of the total uncertainty.
component is evaluated and is added to the cause
and effect diagram as shown in Figure 5. Since
The magnitude of the some of the uncertainties
the repeatability terms for the volume, endpoint,
stated here, in the first example, are slightly
and mass of KHP are incorporated into the
different than some of those shown in the
repeatability of the result, they are combined into
Eurachem Guide [3] because some of the values
one contribution to the result as shown in Figure
in the Eurachem Guide are the result of rounding
6. Figure 6 shows that there are actually 5
prematurely. Rounding, especially for the most
uncertainty components: the original 4 and the
significant uncertainty sources, should be avoided
measurement replication. Figures 5 and 6 are
(especially when using spreadsheets) until the
simplifications of the diagrams in the original
final calculation of the total uncertainty. In both of
example [3].
the following examples some of the values may
appear to have too many significant figures. The
Figure 4. Simple cause and effect diagram for standardization of a NaOH solution with KHP

Figure 5. Detailed cause and effect diagram for standardization of a NaOH solution with KHP

Figure 6. Detailed cause and effect diagram with repeatabilities combined for standardization of a NaOH
solution with KHP
Quantification of Uncertainty Purity of the KHP

The magnitude of each of these components The certificate for the KHP states a purity of
needs to be evaluated. Each of these 1.0000 ± 0.0005 with no information on a
components, in turn, may be the product of coverage factor. By default the uniform
several uncertainty components. The amount of distribution model is used to calculate the Type B
effort expended to estimate the magnitude of the standard uncertainty as
uncertainty component should be commensurate 0.0005
with its relative contribution to the overall = 0.00029.
uncertainty. 3

Measurement replication Molecular weight of the KHP

The most obvious and easily calculated The uncertainty of the molecular weight of the
uncertainty component is the measurement KHP is calculated from the International Union of
replication, which is commonly calculated as the Pure and Applied Chemistry (IUPAC)
standard deviation of the mean. In this example uncertainties of the atomic weights [8]. A uniform
the replication uncertainty was calculated from a distribution is used, presumably by default, to
method validation study to be 0.05 %, relative. calculate the standard uncertainty in this example
The Type A measurement replication is entered in from the Eurachem Guide. It could be argued that
the Kragten spreadsheet as a relative uncertainty the uncertainties of the atomic weights have a
of 0.05 % with its value in the measurement central tendency so that a triangular or normal
equation replication set as 1. Since this distribution (See Figures 1 – 3) is more
uncertainty component implicitly combines the appropriate. However, since the magnitude of this
repeatability from all of the terms in the uncertainty component is relatively insignificant,
measurement equation, it is already in the units the model used is not important here. It could
relative to the result and does not need scaling. also be argued that because of its small
magnitude, this uncertainty source could be
Mass of the KHP ignored in the final calculation. Since this source
is included in the original example [3], it will be
The mass of the KHP was determined as the included here. However, the calculation of this
difference of the container with KHP, and the source will not be repeated here. The magnitude
container after the KHP was added to the sample. of the standard uncertainty is 0.0038 g mol-1.
As shown by Figure 6, the weighing repeatability
is considered separately. A Type B estimate is Volume of the NaOH solution
made for the linearity of the balance. The balance
calibration certificate states that the linearity is ± As shown by Figure 6, volume repeatability is
0.15 mg and recommends the use of a uniform considered separately. Type B estimates are
distribution model to calculate the standard made for the calibration of the buret and
uncertainty. The balance linearity component is temperature variability. The manufacturer states
calculated as the limits of accuracy as ± 0.03 mL. In this
example a triangular distribution is used because
0.15 mg
= 0.0866 mg. a central tendency is assumed. An argument also
3 could be made that ± 0.03 mL is a specification
that should be modeled as a uniform distribution
Since the mass of KHP is calculated as the because any buret within the specified range of ±
difference of two independent determinations this 0.03 mL has been deemed acceptable by the
uncertainty needs to be summed twice as the manufacturer. In addition, just as with the
RSS of the standard uncertainties. The standard uncertainty provided for purity, since the
uncertainty of the mass is calculated as manufacturer has not provided further information
on the uncertainty, the uniform distribution should
(0.0866 mg) 2 + (0.0866 mg) 2 = 2 ⋅ (0.0866 mg) 2 be considered by default. Since the buret
calibration is the major source of uncertainty, the
calculation of the magnitude of this uncertainty
= 2 ⋅ (0.0866 mg) = 0.122 mg. should be considered carefully (The calibration
bias and its uncertainty could be estimated more
accurately by measuring the mass of solution dividing that quantity by the factor for the normal
delivered by the buret, correcting for density to distribution model with 95 % confidence (1.96) as
determine volume, and comparing the received
volume value to the delivered volume value). ((18.64 mL) . (2.06 . 10-4 °C-1) . (3 °C))/1.96
Assuming a triangular distribution, the calibration = 0.0059 mL.
standard uncertainty is calculated as
The volume standard uncertainty is calculated by
0.03 mL the RSS of the standard uncertainties of the buret
= 0.0122 mL.
6 calibration and the temperature variability as

In this example the magnitude of the Type B (0.0122 mL)2 + (0.0059 mL)2 = 0.0136 mL.
estimate for temperature is calculated based on a
temperature variation of ± 3 °C and the
assumption of a normal distribution with 95 % Combined standard uncertainty
confidence (divide by 1.96). Since the NaOH
titrant is very dilute (mass fraction < 0.01 %), the The combined standard uncertainty is determined
expansion coefficient of water can be used. The to be 0.000101 mol L-1 in Table 2 (column 1, row
variation in the volume is calculated by multiplying 11). Page 48 of the Eurachem Guide [3] shows
the volume delivered (18.64 mL) by the expansion similar calculations. In Table 2 it is obvious that
coefficient for water (2.06.10-4 °C-1 at 20°C) [9] the uncertainty of the volume is the major source
and the temperature variation (± 3 °C) and of uncertainty. In this example the analyst should
consider a more thorough investigation of this
source of uncertainty.

Table 2. Kragten spreadsheet for NaOH titration

rep m KHP P KHP M KHP VT

1 2 3 4 5 6
1 Value 1.0 0.3888 1.0 204.2212 18.64
2 Uncertainty 0.0005 0.00013 0.00029 0.0038 0.0136
rep 3 1.0 1.0005 1.0 1.0 1.0 1.0
m KHP 4 0.3888 0.3888 0.38893 0.3888 0.3888 0.3888
P KHP 5 1.0 1.0 1.0 1.00029 1.0 1.0
M KHP 6 204.2212 204.2212 204.2212 204.2212 204.2250 204.2212
VT 7 18.64 18.64 18.64 18.64 18.64 18.654
c NaOH 8 0.102136 0.102187 0.102171 0.102166 0.102134 0.102062
u (y, xi) 9 0.000051 0.000035 0.000029 -0.000002 -0.000074
2 2
u (y), u (y, xi) 10 1.021E-08 2.61E-09 1.20E-09 9E-10 4E-12 5.53E-09
u c(c NaOH) 11 0.000101 51% 34% 29% 2% 74%

Example 2. Gravimetric determination determined gravimetrically and corrections are

of K mass fraction in KH2PO4 made for residual K, precipitate contaminants, and
estimated transfer losses. Since the Type A
In this example the mass fraction of potassium, K, measurement replication of the result of this
in potassium dihydrogen phosphate, KH2PO4, a procedure is less than 0.01 %, relative, the
fertilizer, is determined. In this method potassium uncertainties of these corrections have the
is separated on an ion exchange column and is potential to be a significant contribution to the
isolated as a precipitate of potassium sulfate, overall uncertainty. The major uncertainty
K2SO4. The bulk of the K (> 99.8 %) is sources and their magnitude are determined.
Measurement Equation mCont K2SO4 = mass K2SO4 contaminants [g].

First the measurement equation is written as

The complete measurement equation, a
(2) wK = combination of equations 2 and 3, is

100 . ((mTotal K2SO4) . (2AK/MK2SO4))/mSample [%] wK = 100 . ((mGrav K2SO4 + mRes K2SO4 +
(4)
where, mLoss K2SO4 - mCont K2SO4) . (2AK/MK2SO4))/mSample [%].
wK = mass fraction of K [%]
mTotal K2SO4 = total mass of K2SO4 [g]
AK = atomic weight of K [g mol-1]
Cause and effect diagram
MK2SO4 = molecular weight of K2SO4 [g
Next, a simple cause and effect diagram is drawn
mol-1] (Figure 7) showing the six components of the
2AK/MK2SO4 = 0.4487347 gravimetric factor [8] measurement equation that may contribute to the
mSample = sample mass [g]. uncertainty of the result. The uncertainty of each
component is evaluated and is added to the cause
The measurement equation is further expanded and effect diagram as shown in Figure 8. Note
by the equation for the total mass of K2SO4 as that, for ease of calculations, two significant
uncertainty components for the K2SO4 mass, mGrav
(3) mTotal K2SO4 = mGrav K2SO4 + mRes K2SO4 + and mTemp, are added separately. Since the
repeatability terms for the sample and precipitate
mLoss K2SO4 - mCont K2SO4 [g]
mass are incorporated into the repeatability of the
result, they are combined into one contribution to
where, the result as shown in Figure 9. Figure 9 shows
mTotal K2SO4 = total mass of K2SO4 [g] that there are actually eight uncertainty
mGrav K2SO4 = mass gravimetric K2SO4 [g] components: five of the original six from the
measurement equation (4), two components split
mRes K2SO4 = mass residual K2SO4 [g]
off from the remaining one of the original six, and
mLoss K2SO4 = mass calculated transfer loss one component of measurement replication.
[g]

Figure 7. Simple cause and effect diagram for gravimetric determination of K in KH2PO4
Figure 8. Detailed cause and effect diagram for gravimetric determination of K in KH2PO4

Figure 9. Detailed cause and effect diagram with repeatabilities combined for gravimetric determination of
K in KH2PO4

Quantification of Uncertainty
0.00647 %
= 0.00373 %.
The magnitude of each of these components is 3
evaluated. A total uncertainty based on the result
of an average sample is calculated. Details of the Mass of the sample
magnitude of each component, which are beyond
the scope of this explanation, can be found The mass of the sample was determined as the
elsewhere [10, 11]. difference of the container with the KH2PO4 and
the container after the KH2PO4 was added to the
Measurement replication sample. As shown by Figure 9, the weighing
repeatability is considered separately. A Type B
The result is based on duplicate samples of estimate, based on the range of masses obtained
KH2PO4 determined in three separate runs. The for a standard mass weighed along with the
mean result of each run is used to determine the samples, is made for the bias and resolution of
overall mean so that any Type A uncertainties for the balance. The uncertainty is estimated as ±
the measurement of the blank, sample, and 0.030 mg and is modeled as a triangular
precipitate masses, which are only applicable to distribution because it has a central tendency [5].
an individual run, are easily incorporated into the The balance uncertainty is calculated as
uncertainty for all three runs [10]. In this manner
they do not have to be calculated and added 0.030 mg
= 0.0122 mg.
separately to the total uncertainty. The standard 6
deviation of the three runs is 0.00647 %, relative.
The relative standard deviation of the mean of the Since the KH2PO4 mass is calculated as the
runs (N = 3) is calculated as difference of two independent determinations, this
uncertainty needs to be summed twice in the residual K correction of 0.0081 mg, as K2SO4, the
same manner as used in example 1. The standard uncertainty is calculated as
standard uncertainty of the mass is calculated as
(0.0081 mg) . (0.4) = 0.0032 mg.
2 ⋅ (0.0122 mg) = 0.0173 mg.
K2SO4 precipitate contaminants
Mass of the K2SO4 precipitate
Contaminants in the K2SO4 precipitate were
The uncertainty of the K2SO4 precipitate mass is determined by inductively coupled plasma mass
estimated in the same manner as that used for the spectrometry (ICPMS). The standard uncertainty
sample mass so that the standard uncertainty is of the ICPMS semi-quantitative determination is
also 0.0173 mg. estimated as 50 %, relative. Based on an
average correction of 0.0196 mg, as sulfates, the
standard uncertainty is calculated as
Temperature of heating the K2SO4 precipitate
(0.0196 mg) . (0.5) = 0.0098 mg.
The temperature required to obtain a
stoichiometric K2SO4 precipitate was determined
to be 825 °C in a preliminary study. A Transfer losses
temperature of 750 °C was used in previous work
Several sources of bias may be present from
[12]. A zero bias is estimated for the 825 °C losses in sample transfers and the ion-exchange
temperature selected for heating the precipitate. separation. These sources were addressed in
The relative difference of 0.0148 %, for the detail elsewhere [10, 13]. The maximum loss that
difference in mass at 750 °C and 825 °C could occur, estimated by linearly summing the
determined in the preliminary study, is used as an individual losses, is 0.0618 %, relative. It is
estimate of the uncertainty of the bias of the unlikely that a loss of this magnitude occurred.
heating temperature. This difference is modeled The magnitude of the correction is best modeled
as a uniform uncertainty. The standard using the assumption that it is most likely that
uncertainty is calculated as there is zero bias and that the probability of a bias
0.0148 % linearly decreases to zero at the maximum relative
= 0.00854 %. bias of 0.0618 %. Therefore, the mean bias can
3
be calculated [14] as having a magnitude of one-
third of the maximum interval, which equals
The result of the measurement equation is not 0.0206 %. The model for this bias is shown in
changed by the bias of the heating temperature, Figure 10. This bias has the effect of decreasing
however, the standard uncertainty is added to the the apparent K value. The negative of this bias is
total uncertainty. In effect, the measurement added as a correction in terms of the equivalent
equation is multiplied by a value of 1 and the mass of K2SO4 to each individual sample. The
heating temperature uncertainty is calculated average magnitude of this correction is 0.0435 mg
based on the relative standard uncertainty. based on the mean relative bias of 0.0206 % for
the K2SO4 mean mass of 0.21093 g. The
Residual K standard uncertainty of this correction is
The amount of K in solutions collected before and calculated by dividing the magnitude of the
after the fraction separated for precipitation as correction by the square root of 2 [14]. The
K2SO4, and in rinses of the transfer vessels, was relative standard uncertainty of the transfer loss
determined by flame atomic emission correction is calculated as
spectrometry (FAES). The standard uncertainty is 0.0435 mg
estimated to be 40 %, relative, based on the = 0.0307 mg.
2
range of replicate values for standards and
samples. This large relative uncertainty results
from measurements of K in the residual solutions Atomic weight of K and molecular weight of K2SO4
near the detection limit of the FAES
Since the uncertainties of the atomic and
determination. Since the magnitude of this
uncertainty is a minor component of the total molecular weights of the K and K2SO4 are
uncertainty, no additional effort was made to insignificant, relative to the combined uncertainty
improve this estimate. Based on an average
(< 5 %), the magnitude their uncertainties are not
estimated. Combined standard uncertainty and
expanded uncertainty
The combined standard uncertainty is shown in
Table 3 (column 1, row 13). The expanded
uncertainty (column 1, row 15) is calculated by
multiplying the combined standard uncertainty by
a coverage factor of 2. The relative expanded
uncertainty (column 1, row 16) is 0.041 %. Any
component of the measurement equation that
does not have a significant uncertainty (e.g.,
atomic weights) is not added as an extra row and
column. However, for all calculations of the result,
it is required that all of the components be part of
the measurement equation. Therefore, in this
example, each component of the measurement
equation, except the K atomic weight and the
K2SO4 molecular weight, are entered as a row. In
addition, the measurement replication and the
precipitate heating temperature are added as
components with a value equal to 1, since they do
not change the value of the result. In Table 3 it is
obvious that the uncertainty of the transfer loss is
Figure 10. Deming model [14] for likely zero bias the major source of uncertainty. It is appropriate
and probability of a bias linearly decreases to zero that this source of uncertainty has been studied
at maximum bias very thoroughly [10, 12].

Table 3. Kragten spreadsheet for wK in KH2PO4

rep m Sample m Grav m Res m Loss m Cont m Temp

1 2 3 4 5 6 7 8
1 Value 1.0 0.3294450 0.2109263 0.0000081 0.0000435 0.0000196 1.0
2 Uncertainty 0.0000373 0.0000173 0.0000173 0.0000032 0.0000307 0.0000098 0.0000854
rep 3 1.0 1.0000373 1.0 1.0 1.0 1.0 1.0 1.0
m Sample 4 0.3294450 0.3294450 0.3294623 0.3294450 0.3294450 0.3294450 0.3294450 0.3294450
m Grav 5 0.2109263 0.2109263 0.2109263 0.2109436 0.2109263 0.2109263 0.2109263 0.2109263
m Res 6 0.0000081 0.0000081 0.0000081 0.0000081 0.0000113 0.0000081 0.0000081 0.0000081
m Loss 7 0.0000435 0.0000435 0.0000435 0.0000435 0.0000435 0.0000742 0.0000435 0.0000435
m Cont 8 0.0000196 0.0000196 0.0000196 0.0000196 0.0000196 0.0000196 0.0000294 0.0000196
m Temp 9 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0000854
wK 10 28.7345 28.7355 28.7330 28.7368 28.7349 28.7387 28.7331 28.7369
u (y, xi) 11 0.001073 -0.001511 0.002359 0.000441 0.004188 -0.001337 0.002455
2 2
u (y), u (y, xi) 12 3.455E-05 1.15E-06 2.28E-06 5.57E-06 1.9E-07 1.754E-05 1.79E-06 6.03E-06
u c(w K) 13 0.005878 18% 26% 40% 8% 71% 23% 42%
k 14 2
M -1
k *u 15 0.0118 K2 SO4 174.2602 g mole [8]
-1
rel unc 16 0.041% AK 39.0983 g mole [8]
3. Conclusions results, 1994 ed., 1994; see also
http://physics.nist.gov/Document/tn1297.pdf
A generalized scheme for planning a
measurement and a simple and practical {7] Kragten, J., Calculating standard deviations
approach to estimating and combining the and confidence intervals with a universally
components of measurement uncertainty have applicable spreadsheet technique, Analyst, 119
been shown. Several resources have been noted (10), 1994, p. 2161 – 2166.
that should be consulted to help estimate and
quantify measurement uncertainty. A generalized [8] IUPAC Commission on Atomic Weights and
Kragten spreadsheet with useful modifications can Isotopic Abundances, Pure Appl. Chem. 71, 8,
be used as a tool to simplify the combination of 1999, p. 1593 – 1607.
uncertainty sources and minimize calculation
errors. Simple and practical approaches to [9] D. R. Lide, CRC Handbook of Chemistry and
estimating and combining uncertainties for two Physics, 80th edition, CRC Press: New York,
analytical chemistry methods have been 1999, p. 6-136.
demonstrated.
[10] Vetter, T. W., Yu, L. L., and Sieber, J. R.,
National Institute of Standards and Technology
Acknowledgements Report of Analysis 839.00-210, September 20,
2000.
The author acknowledges L. L. Yu of the NIST
Analytical Chemistry Division for the ICPMS [11] Vetter, T. W., Yu, L. L., Sieber, J. R. and
determinations and W. F. Guthrie of the NIST Guthrie, W. F., Integration of gravimetric and
Statistical Engineering Division for statistical instrumental techniques in the determination of K
consultations. in potassium dihydrogen phosphate (manuscript
in preparation)

[12] Bell, R. K., National Bureau of Standards

References Laboratory Notebook XIII RKB, April, 1973.
[1] Youden, W.J. Realistic estimates of error, [13] Moody, J. R. and Vetter, T. W. , Development
Instrumentation, Systems and Automatic of the ion exchange-gravimetric method for
Control Journal, October 1962, p. 57. sodium in serum as a definitive method, Journal
of Research of the National Institute of Standards
[2] Box, G. E. P., Hunter, W. G., and Hunter, J. S. and Technology, 101, 1996, p. 155 –163.
Statistics for experimenters, Wiley: New York,
1978. [14] Deming, W.E., Some theory of sampling,
Wiley: New York, 1950, p. 62.
[3] Eurachem/CITAC, Guide, Quantifying
uncertainty in analytical measurement, 2nd ed.,
2000; see also
http://www.vtt.fi/ket/eurachem/quam2000-p1.pdf

[4] Ellison, S. L. R. and Barwick, V. J., Estimating

measurement uncertainty: reconciliation using a
cause and effect approach, Accreditation Quality
Assurance, 3, 1998 p. 101 – 105.

[5] ISO, Guide to the Expression of Uncertainty in

Measurement, ISBN 92-67-10188-9, ISO:
Geneva, 1993.

[6] Taylor, B. N. and Kuyatt, C. E., NIST Technical

Note 1297, Guidelines for evaluating and
expressing the uncertainty of NIST measurement