You are on page 1of 0

FOOD COMPOSITION AND ADDITIVES

Comparison of Mercury and Copper Based Catalysts in the
Kjeldahl Determination of Nitrogen in Meat and Meat Products:
Collaborative Study
PRICE ET AL.: JOURNAL OF AOAC INTERNATIONAL VOL. 77, NO. 6, 1994
CINDY G. PRICE and NEIL B. WEBB
Webb Technical Group, Inc., 4320 Delta Lake Drive, Raleigh, NC 27612
WERTICE J. SMITH
U.S. Department of Agriculture, Agricultural Marketing Service, Science Division, Washington, DC 20090
HARRY M. MARKS and ARON M. YOFFE
U.S. Department of Agriculture, Food Safety and Inspection Service, Science and Technology, Washington, DC 20250
Collaborators: S. Arnold; E. Collins; N. Harton; C. Henry; J. Hess; B. Hoar; M. May; J. Michelson; D. Milner; G. Sakell; T.
Spratlin; S.F. Shepherd; J. Thompson; N.B. Webb; M. Wehr; W. Woods
A copper catalyst for use in AOAC Method 928.08,
Nitrogen in Meat, Kjeldahl Method, was subjected
to a collaborative study in which it was compared
to the standard mercury catalyst. Nine laboratories
(including one author’s laboratory) performed blind
duplicate determinations of the protein content of
4 samples from each of 6 products (ground beef,
canned ham, smoked ham, pork sausage, cooked
sausage, and dry cured ham). On the average, pro-
tein concentrations obtained with the copper cata-
lyst were lower than those obtained with the mer-
cury catalyst by approximately 1% of the protein
content of the meat. For example, with a sample
containing 15% protein, the results obtained with
mercury would be expected to be 0.15% higher
than those obtained with copper. This difference is
within the reproducibility standard deviations of
the methods. The repeatability standard deviation
with copper was on the average approximately 17%
larger than that obtained with mercury, but the re-
producibility standard deviations for the 2 proce-
dures were not, statistically, significantly different.
The Kjeldahl method for nitrogen determination in
meat and meat products using copper catalyst has
been adopted first action by AOAC INTERNA-
TIONAL.
T
he Kjeldahl Nitrogen method (1), which was published
in 1883 has, with modifications, been accepted as the
standard method for the determination of nitrogen (pro-
tein) for decades. After its publication, many researchers evalu-
ated different substances to aid in decreasing the total digestion
time. The most effective of the catalysts was mercury, as de-
scribed by Wilfarth (2). Although Arnold (3) is usually associ-
ated with the use of catalysts in the Kjeldahl digestion, his paper
was not published until later. Copper was mentioned as an al-
ternate catalyst for the determination of nitrogen in early edi-
tions of the Official Methods of Analysis (4–7). The use of mer-
cury has several disadvantages, including its high cost, disposal
problems arising from its toxicity, and the requirement that it
should be completely precipitated as a sulfide.
The re-evaluation of a copper catalyst as an alternative or
substitute for the mercury catalyst in the Kjeldahl method was
the result of concerns with the toxicity of mercury and its con-
tribution to environmental pollution. Kane (8) conducted a col-
laborative study for the determination of crude protein in feeds
with a copper catalyst. That method was adopted first action by
AOAC INTERNATIONAL. Florence (9, 10) conducted a col-
laborative study of a method using copper catalyst to determine
nitrogen content in cheese.
Collaborative Study
This article presents the results of a collaborative study to
compare and evaluate the equivalence of copper and mercury
catalysts in AOAC Official Method 928.08, Nitrogen in Meat,
Kjeldahl Method (11). The study was conducted by the Webb
Foodlab, Raleigh, NC, under contract with the Food Safety and
Inspection Service (FSIS). Six product types were chosen for
this study to provide a representative range of protein, fat and
water concentrations. A total of 24 samples (4 of each product
type) were split and analyzed by 9 laboratories (including one
Submitted for publication April 19, 1993.
The recommendation was approved by the Committee on Commodity
Foods and Products, and was adopted by the Official Methods Board of the
Association. See “Official Methods Board Actions (1993) J. AOAC Int. 76,
125A, and ”Official Methods Board Actions" (1993) The Referee 17,
September issue.
author’s laboratory) using both mercury and copper as cata-
lysts. As explained below, because of a procedural error in one
of the laboratories, only results from 8 laboratories were used
in the statistical analyses.
Sample Selection and Preparation
The 6 product types were: (1) ground beef (protein concen-
trations from 16.5% to 17.4%); (2) canned ham (protein con-
centrations from 18.2% to 18.8%); (3) smoked ham (protein
concentrations from 13.5% to 19.8%); (4) pork sausage (pro-
tein concentrations from 10.4% to 11.7%); (5) cooked sausage
(protein concentrations from 11.9% to 12.7%); (6) dry cured
ham (protein concentrations from 22.7% to 25.0%).
The samples were prepared according to instructions given
in the U.S. Department of Agriculture Chemistry Laboratory
Guidebook (12). The samples were tested for homogeneity be-
fore submission to the collaborators. Based on values obtained
from 4 different subsamples, the repeatability standard devia-
tion for each of the 24 samples was computed and compared to
the FSIS method performance standard for repeatability stand-
ard deviation of 0.24% (13). Only one sample exhibited a
standard deviation (0.26%) exceeding this performance stand-
ard. This difference is not considered of practical significance,
and thus all 24 samples were retained in the study.
After preparation, portions of approximately 100 g were
placed in bags and frozen. Samples were shipped frozen to ar-
rive at the laboratories the following day. Each laboratory re-
ceived randomly numbered blind duplicates of each of the
4 samples from each product group (48 subsamples per labora-
tory). The samples were stored under refrigeration after they
were received. Prior to analysis, the samples were tempered at
room temperature for 10–15 min. To improve homogeneity,
each tempered sample was kneaded in its plastic bag prior to
removal of the sample for analysis. Samples contained in each
of the 48 sample bags were further split at the laboratory, one
split analyzed with the copper catalyst, the other with the mer-
cury catalyst. Upon completion of the analysis the protein data
were returned to Webb Foodlab for decoding and compilation
of the results.
In addition to the test samples, the collaborators were also
supplied with prepackaged catalysts, Kelpak 5p (15.0 g K
2
SO
4
,
0.7 g HgO and 0.1 g pumice) and Propac 15p (15.0 g K
2
SO
4
,
0.45 g CuSO
4
and 0.1 g pumice). The collaborators were in-
structed to use 30 min digestion time after clearing for mercury
and 45 min after clearing for copper.
928.08 Nitrogen in Meat—Kjeldahl Method
Alternative I - (Mercury Based Catalyst)
Final Action 1974
A. Reagents
(a) Kjeldahl catalyst.—15 g K
2
SO
4
+ 0.7 g HgO (com-
mercially prepared catalysts are available containing pumice,
if desired).
(b) Sulfuric acid.
(c) NaOH solution.—Prepare 1200 mL NaOH (1 + 1). Let
stand until clear (about 10 days).
(d) Metallic zinc.—Powder, to be used if catalyst does not
contain pumice.
(e) Indicator solution.—Methyl purple, Fisher or equivalent.
(f) Acid potassium phthalate.—NIST Standard.
(g) Standard NaOH solution.—0.2000 ± 0.0004N. Add
108 mL NaOH (1 + 1) to CO
2
-free H
2
O and dilute to 10 L.
Standardize against potassium acid phthalate, using phenol-
phthalein indicator.
(h) Standard acid solution.—0.2000 ± 0.0004N. Prepare
either HCl or H
2
SO
4
solution.
(i) Hydrochloric acid.—Dilute 178 mL 35–37% HCl to
10 L. Standardize against standard NaOH solution, and adjust
strength accordingly.
(j) Sulfuric acid.—Dilute 55 mL 98% H
2
SO
4
to 10 L.
Standardize against standard NaOH solution, and adjust
strength accordingly.
(k) Sodium hydroxide–sodium thiosulfate solution.—Dis-
solve 460 g Na
2
S
2
O
3
⋅5H
2
O in H
2
O; dilute to 1 L with H
2
O, and
add this solution to 15 250 g NaOH dissolved in 14 250 mL
H
2
O. This will yield 20 L 50% (w/w) NaOH solution. If other
volumes are desired, adjust weights of NaOH and
Na
2
S
2
O
3
⋅5H
2
O accordingly. Specific gravity of final solution
should be 1.45.
B. Determination
Place weighed sample (2.0–2.2 g) in digestion flask. Add
0.7 g HgO or 0.65 g metallic Hg, 15 g powdered K
2
SO
4
or an-
hydrous Na
2
SO
4
, and 40 mL H
2
SO
4
. If sample >2.2 g is used,
increase H
2
SO
4
by 10 mL for each g sample. Place flask in in-
clined position and heat gently until frothing ceases (if neces-
sary, add small amount of paraffin to reduce frothing); boil
briskly until solution clears, and then ≥30 min longer (2 h for
samples containing organic material).
Cool, add ca 200 mL H
2
O, cool to <25°, add 25 mL of the
sulfide or thiosulfate solution, and mix to precipitate Hg. Add
few Zn granules to prevent bumping, tilt flask, and add layer of
NaOH without agitation. (For each 10 mL H
2
SO
4
used, or its
equivalent in diluted H
2
SO
4
, add 15 g solid NaOH or enough
solution to make contents strongly alkaline.) (Thiosulfate or
sulfide solution may be mixed with the NaOH solution before
addition to flask.) Immediately connect flask to distilling bulb
on condenser, and, with tip of condenser immersed in standard
acid and 5–7 drops indicator in receiver, rotate flask to mix con-
tents thoroughly; then heat until all NH
3
has distilled (≥150 mL
distillate). Remove receiver, wash tip of condenser, and titrate
excess standard acid in distillate with standard NaOH solution.
Correct for blank determination on reagents.
% N = [(mL std acid × normality acid) − (mL std NaOH ×
normality NaOH)] × 1.4007/g sample
Ref: J. Assoc. Off. Anal. Chem. 11, 408 (1928)
Alternative II - (Copper Based Catalyst)
First Action 1993
(Applicable to determination of nitrogen in meat and meat
products.)
Method Performance:
Ground beef (mean protein value = 17%)
s
r
= 0.22; s
R
= 0.26; RSD
r
= 1.30%; RSD
R
= 1.51%
Canned ham (mean protein value = 18%)
s
r
= 0.23; s
R
= 0.27; RSD
r
= 1.27%; RSD
R
= 1.48%
Smoked ham (mean protein value = 15%)
s
r
= 0.20; s
R
= 0.24; RSD
r
= 1.36%; RSD
R
= 1.58%
Pork sausage (mean protein value = 10.5%)
s
r
= 0.16; s
R
= 0.19; RSD
r
= 1.54%; RSD
R
= 1.80%
Cooked sausage (mean protein value = 12%)
s
r
= 0.18; s
R
= 0.21; RSD
r
= 1.47%; RSD
R
= 1.72%
Dry cured ham (mean protein value = 23%)
s
r
= 0.28; s
R
= 0.31; RSD
r
= 1.16%; RSD
R
= 1.36%
(Caution: Contact with concentrated boiling H
2
SO
4
, used
during digestion, or concentrated boiling NaOH, used during
distillation, will result in permanent damage; wear gloves,
safety goggles, and lab coat. In addition, H
2
SO
4
evolves SO
3
vapors, which are toxic and typically cause severe respiratory
irritation. Carry out digestion in Kjeldahl apparatus designed to
contain and remove fumes. After digestion, let Kjeldahl flasks
cool to room temperature before removal.)
C. Principle
Sample is digested in H
2
SO
4
, using CuSO
4
as catalyst, con-
verting nitrogen to NH
3
which is distilled and titrated.
D. Apparatus and Reagents
See A, except a copper catalyst is used in place of mercury
catalyst.
(a) Copper catalyst.—Prepare using 15.0 g K
2
SO
4
and
0.45 g CuSO
4
. As boiling aid, 0.1 g pumice may be added.
(Propac 15p, Alfie Packers, Inc., Omaha, NE, or other commer-
cially prepared catalysts are suitable.)
E. Preparation of Samples
See 983.18.
F. Determination
See B, except in para. 1, after solution clears, boil 45 min
longer (2 h for samples containing organic material).
G. Calculations
See B. Note: Alternative II yields results that are, on average,
99.0% of results from Alternative I.
Ref.: J. AOAC Int. 77, 1542 (1994)
Statistical Evaluation
To characterize bias between the mean results of the proce-
dures, regression analysis of the sample mean obtained using
copper vs that obtained using mercury, was used. The compari-
son of variability was expressed by computing the geometric
mean of the ratios of the sample standard deviations. In addition
to the above comparison, the standard deviations of the 2 meth-
ods were characterized through regression estimates.
An Initial Method Comparison (described in Results and
Discussion) was performed using all the data, except that from
Laboratory 9, and 2 other clearly aberrant data values, for rea-
sons that will be explained below. After this Initial Method
Comparison, a Further Data Analysis (see Results and Discus-
sion) was performed to identify results that would have an un-
due or disproportionate influence on the comparisons between
standard deviations for the use of the 2 catalysts, such as outlier
results, or results from samples exhibiting excessive heteroge-
neity. The Initial Method comparison was then repeated, ex-
cluding those results identified in the Further Data Analysis.
This Revised Method Comparison (see Results and Discussion)
provides a more precise comparison of the 2 methods, thereby
allowing any performance differences between the 2 methods
to be more accurately characterized. For example, if these iden-
tified results were included, they might inflate the variance as-
sociated with the method comparison, which could mask true
differences, or create artificial differences, between the meth-
ods.
The Further Data Analysis uses graphs in which outlier or
influential results can be readily identified. In addition, formal
outlier statistics and significance levels were computed to con-
firm that the graphically identified points should be considered
to be outlier values. Deleted results were from subsamples that
appeared to have either a larger degree of heterogeneity than the
other subsamples or a disproportionate amount of influence on
the comparison. Whenever a set of individual values obtained
with one catalyst was deleted, a second set of values, associated
with the other catalyst, was also eliminated. The second set of
values was chosen to reduce the bias that might possibly result
if only the first set of values were deleted.
Results and Discussion
Initial Method Comparison
Tables 1–6 present sample means (x
_
), and standard devia-
tions of repeatability and reproducibility, calculated using all
data under the heading “No data deleted.” The next set of num-
bers (under the heading “Initial method comparison”) are based
upon all the data except those from Laboratory 9 (for reasons
that will be explained in the next paragraph), and the clearly
aberrant results from both procedures on the second sample of
the pork sausage product in Laboratory 4, i.e., the 2 results
greater than 12%.
It was discovered, after completion of this study, that Labo-
ratory 9 did not follow instructions with respect to the sample
weight taken for analysis. Laboratory 9 used more than 2.7 g,
whereas the method specifies that about 2 g should be taken for
analysis. Also, as is clear from just a cursory glance in Tables 1–
6, the calculated repeatability and reproducibility standard de-
viations for many samples are noticeably larger when the re-
sults from Laboratory 9 are included. Refer to Tables 1–6,
where standard deviations are often more than doubled as a
result of the inclusion of data from Laboratory 9. The data from
this laboratory were, therefore, deleted from the statistical
analysis that was used in comparing the performance of the
2 catalysts.
To characterize bias, the differences between the subsample
mean results obtained with mercury and those obtained with
copper were calculated. Figure 1 presents a scatter plot of these
differences vs the mean results obtained using mercury (con-
sidered to be the standard in this study). There is a positive
correlation between the differences and the sample means ob-
tained with mercury. Also presented on this graph is the regres-
sion line obtained using an unweighted least-squares fit. A re-
gression analysis indicates that the expected value of results
using mercury will be higher, by approximately 1% of the pro-
tein concentration, than the expected value of results using cop-
per. In other words if, when using copper, a sample is deter-
mined to contain 15.00% protein, then a value of 15.15%
would be expected to be obtained with the use of mercury.
In addition, the geometric means of the ratios of the sample
repeatability and reproducibility standard deviations for the
2 catalysts were computed. For each sample, this ratio is equal
to the standard deviation obtained with the mercury catalyst
divided by that obtained with the copper catalyst. The geomet-
ric mean of the sample ratios for the repeatability standard de-
viation was 0.85, implying that the repeatability standard de-
viation with mercury is estimated to be approximately 15%
lower than with copper. The geometric ratio for the reproduci-
bility standard deviation was 0.97, implying that the reproduci-
bility standard deviation with mercury is estimated to be ap-
proximately 3% lower than obtained with copper. This latter
ratio is not, statistically, significantly different from 1.
However, even after deleting the data from Laboratory 9 and the
2 clearly aberrant results, there remain a few values which are dis-
proportionately influential on the estimated ratios. As is described
below, elimination of these values will lead to a more reliable com-
parison, though the general conclusions are not changed.
Further Data Analysis
Two ideas underlie the Further Data Analysis: (1) the ob-
served variability for a sample can be considered to have
2 sources: analytical variability and sample heterogeneity, i.e.,
subsamples that do not have the same amount of protein. The
greater the latter component, the more difficult it is to compare
the difference in analytical variability between the 2 proce-
dures; (2) estimates of variances can be highly influenced by
only a few results, some of which may be outliers.
The purpose of this Further Data Analysis is to identify re-
sults from subsamples that appear to be excessively hetero-
genous, or results that have a disproportionate influence on the
estimates. After these results are identified, the analysis is re-
peated without them.
To identify such subsamples, it will be necessary to employ,
as a criterion, the standard deviation that would be predicted in
the absence of excessive heterogeneity. This quantity will here-
after be referred to as the “predicted standard deviation.” Using
these predicted standard deviations, it is possible to associate,
with each result, a “normalized value.” These normalized val-
ues are not dependent upon the protein level of the sample or
the analytical variability of the method. The normalized values
are then combined and used in the identification of subsamples
with excessive levels of heterogeneity or which have dispro-
portionate influence on the comparison of the method catalysts.
To calculate the predicted standard deviations, a regression
analysis was performed. From Tables 1–6, it can be seen that
variabilities of results from samples generally increase with
levels of protein found in the samples, suggesting that a sam-
ple’s measured variability is expected to be an increasing func-
tion of protein level. Therefore, a level of heterogeneity that
would be acceptable in a high-protein sample might be unac-
ceptable in a sample with a lower percentage of protein. Thus,
it is necessary to take this expected dependence of subsample
variability on protein level into account when calculating the
predicted standard deviation. The predicted standard deviation
for results from a sample was derived through regression analy-
sis, using a function whose value increases with protein level.
The predicted standard deviation was calculated by a simple
model which related, through a linear regression, the natural
logarithmic transformation of the sample reproducibility stand-
ard deviations and the natural logarithmic transformation of the
sample means. This model is presently used in the FSIS Labo-
ratory Accreditation Program for predicting standard devia-
tions (14). The authors examined alternative models, such as
considering the standard deviation to be a linear function of the
mean, and found that they did not provide improved predict-
ability. The validity of the logarithmic model, in this applica-
tion, was confirmed when it was found that the logarithm of the
ratio of the standard deviations of the 2 methods was not a func-
tion of the sample protein content. In addition, analyzing the
logarithm of the standard deviation provides particular statisti-
cal advantages in making comparisons of standard deviations,
and is, therefore, a common statistical practice (15).
Figure 2 represents a scatter plot of the reproducibility of
mercury vs that of copper. The linear regression line and a 45°
line (i.e., one with a slope of 1 which intersects the origin) are
superimposed. This graph indicates the existence of 2 samples
with reproducibility standard deviations greater than 0.6% for
both catalysts. After eliminating these 2 samples (second sam-
ple canned ham, fourth sample smoked ham) and performing
the regression analysis, the resulting equations were derived:
Hg: ln (STD) = −2.95 + 0.576 ln (mean)
Cu: ln (STD) = −3.05 + 0.600 ln (mean)
Using the above equations, the predicted standard deviation
can be approximated as 0.05 times the mean raised to the 0.60
power for both catalysts.
The predicted standard deviations were in turn used to cal-
culate the normalized value for each result. This was accom-
plished by subtracting the sample mean for the catalyst from
each result, and then dividing the resulting difference by the
predicted standard deviation. Specifically, the normalized
value for a result from a subsample was set equal to the differ-
ence between the result and the sample mean for the catalyst,
divided by the product of 0.05 times the sample mean for the
catalyst raised to the 0.6 power. These normalized values are
not considered to be dependent upon protein level, and thus can
be used to develop protein-independent criteria by which to
identify subsamples with excessive heterogeneity.
Each normalized value can be thought of as a variable with
an expected value of 0 and a standard deviation of slightly less
than 1. The reason this standard deviation would not be exactly
equal to 1 is due to the correlation that exists between each re-
sult from a sample and the mean of the results from that sample.
However, for the purpose here, it is sufficient to assume it is
equal to 1, because the factors that would be needed to make it
in fact equal to 1 are small and would be assumed constant for
all the samples. Furthermore, the error in assuming it is equal
to 1 when it is actually less than 1, makes it slightly less likely
that data values would be identified as outliers.
Graphical analysis of the normalized values was the primary
data analysis tool used in the search for subsamples that ap-
peared to be outliers, or to otherwise exert disproportionate in-
fluence upon the results. Statistical outlier tests were then used,
to confirm that those points identified through the graphical
analysis warranted exclusion from the repeat analysis.
Two graphs were used to assist in the analysis. Figure 3
shows a scatter plot of the within-subsample standard devia-
tions of the normalized values obtained with mercury vs those
obtained with copper. The 3 points closest to the right corner of
this graph represent subsamples for which the copper and mer-
cury catalysts standard deviations are both large in comparison
to those of the other subsamples. This indicates that these sub-
samples exhibit an excessive amount of heterogeneity, and they
were, therefore, eliminated from the repeat statistical analysis.
These subsamples are: cooked sausage, Laboratory 8, sam-
ple 2; canned ham, Laboratory 8, sample 3; and cooked sau-
sage, Laboratory 7, sample 1.
When one of the 2 standard deviations was large and the
other one was small, data points can potentially exert a dispro-
portionate influence upon the comparison of the variabilities of
the results obtained with the 2 catalysts. These data points
would lie either in the upper left or lower right corners of Fig-
ure 3. An equal number of points were deleted from these 2 cor-
ners. Specifically, 4 data points were deleted in the repeat
analysis. The 2 deletions in the upper left corner had the effect
of decreasing the calculated variability of the mercury method
relative to that of the copper method. This was balanced by the
deletion of the 2 points in the lower right corner, which reduced
the calculated variability of the copper method relative to the
mercury method. This elimination is equivalent to the proce-
dure of “trimming” when estimating an unknown mean of a
population, where an equal number of high and low values are
deleted before estimating the mean (16). These 4 subsamples
were eliminated when the larger standard deviation ex-
ceeded 2.5 and the smaller one was less than 0.5. These sub-
samples were: canned ham, Laboratory 8, sample 2; pork sau-
sage, Laboratory 1, sample 3; smoked ham, Laboratory 7,
sample 4; and canned ham, Laboratory 7, sample 2.
A calculation was also performed to confirm that these
7 points would be considered as outliers based on a statistical
criterion. For each subsample, the sum of the normalized vari-
ances within a laboratory and the ratio of this sum to the sum
of all the 191 within-laboratory normalized variances was com-
puted. Under the assumptions that the 2 catalysts would yield
results with normal distributions and with the same expected
variances, such a ratio has a beta distribution (17) with 1 and
190 degrees of freedom. Using a Bonferroni approximation, as
presented by Hawkins (18), a ratio with a value greater than
0.0389 could be classified as an outlier at the 0.10 significance
level. Six of the 7 points identified above had the largest ratios,
ranging from 0.0342 to 0.0417. The next smallest ratio was
0.0262, which is 0.080 less than 0.0342. By contrast, the next
smallest ratios decrease by small amounts, i.e., 0.013, 0.002,
and 0.017. Because 5 of these 6 largest ratios are greater than
0.0389, each could be classified as an outlier at close to or less
than the 0.10 significance level; the sixth, with a ratio of
0.0342, has a significance level, calculated by the Bonferroni
approximation, of 0.25. The seventh sample, a smoked ham
sample from Laboratory 7, has a ratio of 0.25 and also a large
negative difference of normalized means (Figure 4, the lowest
plotted point). This raises sufficient suspicions regarding its
heterogeneity to justify its deletion.
These 7 deleted points can be seen in Figure 3. The 6 points
mentioned in the previous paragraph as having the largest ratios
are in the peripheral “circumference” of the scatterplot. The
point labeled “7,” representing the smoked ham sample from
Laboratory 7, is in the upper left corner of the scatterplot.
The second graph (Figure 4) represents an additional means
by which to identify outliers and other unduly influential data
points. It is based on the fact that the difference of variances can
be expressed as a covariance of a sum and a difference, i.e.,
cov(x+ y, x − y) = var(x) − var(y). Hence, a scatterplot of the
difference (x − y) between the means of normalized values on
a subsample within a laboratory vs the sum (x + y), can be used
to look for highly influential, or outlier, values (Figure 4). In
our application, for each subsample, x and y are the means of
the replicate normalized values obtained with mercury and cop-
per, respectively. Therefore, the sum and the difference of x and
y are variables with an expected value of 0 and standard devia-
tion of 1. Under the null hypothesis that the variances are the
same, and assuming normality, the sum and difference are sta-
tistically independent. If these assumptions were true, the scat-
terplot of (x − y) vs (x + y) would take on the shape of a circle
because the probability density under the normal distribution
depends only on the sum of the squares of (x − y) and (x + y).
Any data value at a distance from the origin greater than 4
would be a candidate point as an outlier and influential on the
comparison of variances between the 2 procedures. The value
of 4 approximates a P-value less than 0.10, based on a chi-
square distribution with 2 degrees of freedom and upon the
Bonferroni approximation for determining the significance
level. The Bonferroni approximation estimates the significance
level by equating it to the product of: (1) the number of sam-
ples and (2) the probability that a random variable from a chi-
square distribution with 2 degrees of freedom would have a
value exceeding 4. As is evident from Figure 4, there are 5 data
points that appear distinct from the main grouping of data val-
ues along the (x + y) axis and which are clearly not within the
depicted circle. Each of these 5 subsamples was thus desig-
nated for deletion in the repeat analysis. Four were from Labo-
ratory 8 and one was from Laboratory 6. However, 2 of the 4
(those from Laboratory 8) are among the 7 outlier points al-
ready designated for elimination by the previous analysis (rep-
resented in Figure 3). Therefore, the net result of this analysis
(represented in Figure 4) is the designation of 3 additional sub-
samples for elimination. These are: canned ham, Laboratory 6,
sample 2; canned ham, Laboratory 8, sample 1; and smoked
ham, Laboratory 8, sample 4. In sum, the analyses represented
in Figures 3 and 4 have resulted in the designation of a total of
10 subsamples for elimination from the repeat analysis, where
each subsample represents 4 data points (one pair of replicates
for each catalyst).
To gauge the effect of these data deletions on the compari-
son, analyses of variance for each laboratory were performed
on the normalized values before and after deletion. Following
deletion, the average variances computed for the different labo-
ratories became considerably more similar, and the reproduci-
bility and repeatability standard deviations for each of the
methods decreased by an average of approximately 14%.
Therefore, it is believed that a more precise and valid compari-
son of the procedures will be obtained using the results calcu-
lated following deletion of the aberrant data values.
In summary, there were initially a total of 864 data points.
For reasons given above, all 96 data points from Laboratory 9
were deleted, as well as 42 data points (the 2 mentioned in In-
itial Method Comparison plus the 40 identified in this section)
from the remaining 8 laboratories. Therefore, the following sta-
tistical analysis was performed on 726 data points from 8 labo-
ratories.
Revised Method Comparison
In Tables 1–6, the results of the Revised Method Compari-
son are presented under this heading. These are the sample
means and standard deviations of repeatability and reproduci-
bility that were calculated following deletion of the data iden-
tified in Further Data Analysis. Linear regression was per-
formed upon the subsample means obtained using the copper
catalyst vs those obtained using the mercury catalyst. The inter-
cept was 0.05%, with a standard error of approximately 0.05%.
The slope was 0.987, with a standard error of 0.003. Therefore,
because the intercept and slope are not significantly different
from 0 and 0.99, respectively, it would appear reasonable to
estimate that the average results of protein content using mer-
cury are 1% higher than average results using copper. This bias
is well within one reproducibility standard deviation unit.
Variability is compared by computing the geometric mean
of the ratio of the standard deviation obtained with mercury to
that obtained with copper. The summary of these results is pre-
sented in Table 7.
The geometric mean for repeatability, 0.83, is, statistically,
significantly different from 1, at approximately the 3% signifi-
cance level using either a Student’s t-test or Wilcoxin test on the
natural logarithm of the ratios. For reproducibility, the differ-
ence is not statistically significant.
Standard Deviations
To characterize the predicted repeatability standard devia-
tions (PSD
r
) and predicted reproducibility standard deviations
(PSD
R
) of the 2 methods, a linear regression was performed
using the natural logarithmic transformed value of the standard
deviation as the dependent variable and the natural logarithmic
mean as the independent variable, as described above. Of the
10 subsamples identified from Figures 3 and 4 for elimination
in the repeat analysis, those that were eliminated only because
of high influence on the comparison were not eliminated in this
analysis. These are the 4 subsamples identified in the lower
right and upper left corners of Figure 3. Therefore, only the
remaining 6 subsamples were deleted from the subsequent
analysis. An analysis of variance indicates that no statistically
significant difference exists among the slopes of the regressions
of the logarithms of the copper and mercury repeatability and
reproducibility standard deviations. Hence a common slope
can be used. PSD
r
and PSD
R
were derived from the exponential
of the predicted natural logarithmic standard deviations. For
mercury, it is estimated that PSD
r
= 0.030 × x
0.64
, where x is the
percentage protein content of the sample. The exponent 0.64
was the slope common to all 4 regression equations, as men-
tioned above, with a pooled standard error of 0.097. For copper,
the PSD
r
is estimated to be approximately 23% higher than that
for mercury. This difference is statistically significant (P
<0.05). For mercury, it is estimated that PSD
R
= 0.040 × x
0.64
,
which is approximately 5% lower than the PSD
R
when using
copper. This difference is not statistically significant (P = 0.50).
For example, for a sample containing 15% protein, PSD
r
values
with copper and mercury would be 0.21% and 0.17%, respec-
tively, and PSD
R
values would be approximately 0.23% for
both procedures. These equations were used to determine the
method performance values for the copper catalyst method.
Our conclusions correspond to the results presented in Ta-
ble 7, namely that the geometric mean of the PSD
r
using copper
is approximately 20% higher than that for mercury. The geo-
metric means of the PSD
R
values for the 2 catalysts were not,
statistically, significantly different. For both copper and mer-
cury, there were aberrant results. Finally, while there appears to
be a bias between the methods, it is small, averaging approxi-
mately 1% of the protein content of the sample.
Conclusions
In this study, the copper catalyst exhibited a negative bias
when compared to the mercury catalyst. The magnitude of the
bias is approximately 1% of the measured protein content of the
sample using mercury, and is less than the reproducibility
standard deviations of the methods. The use of the copper cata-
lyst yields results with a slightly larger PSD
r
, but there was no
statistically significant difference in PSD
R
.
Acknowledgments
We wish to acknowledge the participation of the following:
Theresa Spratlin and Wayne Woods, U.S. Department of
Agriculture, Athens, GA
Nancy Harton and Ernest Collins, Kentucky State Depart-
ment of Agriculture, Frankfort, KY
Brenda Hoar and Michael May, Kunzler and Co., Inc., Lan-
caster, PA
George Sakell, Meat Industry Laboratories, Inc., Chicago,
IL
Steve Arnold and Jack Michelson, Michelson Laboratories,
Inc., Los Angeles, CA
David Milner and Carolyn Henry, U.S. Department of Ag-
riculture, St. Louis, MO
Signe F. Shepherd and Michael Wehr, Oregon State Depart-
ment of Agriculture, Salem, OR
John Thompson and James Hess, U.S. Department of Agri-
culture, Alameda, CA
We would also like to thank George R. Heavner, AOAC
INTERNATIONAL, who, while employed at FSIS, U.S. De-
partment of Agriculture, helped in writing sections of the paper,
and Lily Chounlamountry, U.S. Department of Agriculture, for
the difficult task of typing this manuscript.
References
(1) Kjeldahl, J. G. C. (1883) J. Anal. Chem. 22, 336–382
(2) Wilfarth, H. (1885) Chem. Zentr. 17, 113
(3) Arnold, C. (1886) Arch. Phars. 24, 785–794
(4) Official and Tentative Methods of Analysis (1935) 4th Ed.,
AOAC, Arlington, VA, secs 19–25
(5) Official and Tentative Methods of Analysis (1940) 5th Ed.,
AOAC, Arlington, VA, secs 19–25
(6) Official and Tentative Methods of Analysis (1945) 6th Ed.,
AOAC, Arlington, VA, secs 2.22–2.26
(7) Official Methods of Analysis (1950) 7th Ed., AOAC, Ar-
lington, VA, secs 2.20–2.25
(8) Kane, P.F. (1984) J. Assoc. Off. Anal. Chem. 67, 869–877
(9) Florence, E., Harris, W.M., & Milner, D.F. (1985) Analyst
110, 971–973
(10) Florence, E., & Harris, W.M. (1987) Analyst 112, 317–320
(11) Official Methods of Analysis (1990) 15th Ed., AOAC, Ar-
lington, VA, sec. 928.08
(12) U.S. Department of Agriculture Food Safety and Inspection
Service (1984) Chemistry Laboratory Guidebook, Washing-
ton, DC, sec. 1.002B
(13) U.S. Department of Agriculture Food Safety and Inspection
Service (1984) Chemistry Quality Assurance Handbook,
Washington, DC
(14) “Food Safety and Inspection Service (Meat, Poultry),
USDA—Accreditation of Federal Laboratories” in Code of
Federal Regulations: Animals and Animal Products, 9, Parts
318.21 and 381.153, revised January 1, 1992
(15) Scheffe, H. (1959) Analysis of Variance, John Wiley and
Sons, New York, NY, p. 84
(16) Johnson, N. I. (1970) Continuous Univariate Distributions-I,
Houghton Mifflin Co., Boston, MA, p. 60
(17) Johnson, N. I. (1970) Continuous Univariate Distributions-
II, Houghton Mifflin Co., Boston, MA, p. 38
(18) Hawkins, D. M. (1980) Identification Of Outliers, Chapman
and Hall, New York, NY, Appendix 9
Table 1. Collaborative study results for protein determination (%) in ground beef (blind duplicates) by Kjeldahl
method: raw data and statistical summaries
Lab.
Sample
1 2 3 4
Cu Hg Cu Hg Cu Hg Cu Hg
1 17.36 17.11 16.94 16.44 17.06 17.38 17.18 17.15
18.19 17.33 16.17 16.65 16.60 17.37 16.98 17.38
2 17.13 17.17 16.45 16.74 17.15 17.06 16.92 17.44
16.98 17.40 16.55 16.99 16.62 17.18 16.91 17.32
3 17.24 17.16 16.79 16.33 17.21 17.33 17.91 17.39
17.18 17.52 16.03 16.23 17.64 17.67 17.21 17.30
4 17.09 17.70 15.89 15.97 17.27 16.91 17.04 16.70
17.15 17.28 16.02 16.12 17.55 17.12 17.39 17.71
5 16.98 17.35 16.17 16.44 16.86 17.08 17.09 17.21
17.40 17.34 16.56 16.59 17.32 17.34 17.00 17.22
6 17.04 17.08 16.31 16.60 17.35 17.48 16.62 16.77
16.80 17.22 15.94 16.47 17.24 17.35 16.61 17.22
7 17.12 17.07 15.92 16.11 16.87 17.37 17.28 17.39
16.94 17.01 16.20 16.75 17.07 17.30 17.05 17.42
8 16.99 17.26 16.72 16.33 17.68 17.40 17.51 17.36
16.97 17.16 16.45 16.79 17.59 17.25 17.17 16.96
9
a
16.78 17.12 16.27 16.20 17.00 16.79 17.13 17.27
16.72 17.18 16.09 15.84 17.11 16.60 17.14 17.25
No data deleted
No. of samples 18 18 18 18 18 18 18 18
x
_
b
17.114 17.249 16.304 16.422 17.177 17.221 17.119 17.248
s
r
c
0.2344 0.1570 0.3046 0.2273 0.2403 0.1338 0.2115 0.2847
s
R
d
0.3298 0.1736 0.3153 0.3127 0.3243 0.2655 0.3045 0.2405
Initial method comparison (Laboratory 9 data deleted)
No. of samples 16 16 16 16 16 16 16 16
x
_
17.160 17.260 16.319 16.472 17.193 17.287 17.117 17.246
s
r
0.2482 0.1659 0.3200 0.2236 0.2534 0.1337 0.2277 0.3020
s
R
0.3200 0.1805 0.3306 0.2852 0.3419 0.1866 0.3248 0.2523
Revised method comparison (Laboratory 9 and outlier data deleted)
e
No. of samples 16 16 16 16 16 16 16 16
x
_
17.160 17.260 16.319 16.472 17.193 17.287 17.117 17.246
s
r
0.2482 0.1659 0.3200 0.2236 0.2534 0.1337 0.2277 0.3020
s
R
0.3200 0.1805 0.3306 0.2852 0.3419 0.1866 0.3248 0.2523
a
Data deleted from the initial and revised method comparisons.
b
Sample mean.
c
Repeatability standard deviation.
d
Reproducibility standard deviation.
e
The further data analysis did not identify any outliers. Therefore, the initial and revised method comparisons gave identical results.
Table 2. Collaborative study results for protein determination (%) in canned ham (blind duplicates) by Kjeldahl
method: raw data and statistical summaries
Sample
Lab.
1 2 3 4
Cu Hg Cu Hg Cu Hg Cu Hg
1 18.16 18.26 19.13 18.58 18.23 18.36 18.41 18.89
17.81 18.11 18.93 18.67 18.34 18.49 18.82 18.59
2 18.11 18.33 18.41 19.12 18.25 18.44 18.69 18.97
17.93 18.08 18.86 19.02 18.49 18.78 18.76 19.01
3 17.50 17.44 18.89 18.96 18.22 18.06 18.94 18.61
17.76 17.81 18.50 18.54 18.24 18.29 18.88 19.02
4 17.78 17.78 18.73 19.05 18.02 18.16 18.56 18.74
17.69 17.74 18.53 18.61 18.32 18.27 18.62 18.96
5 18.11 18.22 18.89 19.08 18.22 18.39 18.96 19.11
17.91 18.02 18.67 19.15 18.35 18.53 19.01 19.03
6 18.21 18.03
b
19.49
b b
19.48
b
18.40 18.88 18.76 19.18
17.90 18.25
b
19.12
b b
19.72
b
18.51 18.58 18.66 19.47
7 17.25 17.85
b
19.34
b b
18.78
b
18.29 18.19 18.20 18.67
18.28 18.10
b
18.02
b b
18.80
b
17.79 18.26 18.69 18.60
8
b
18.68
b b
18.86
b b
17.32
b b
16.91
b b
17.93
b b
18.05
b
18.77 18.45
b
18.79
b b
19.05
b b
17.43
b b
18.31
b b
19.07
b b
18.91
b
19.15 19.55
9
a
18.45 17.93 17.79 18.21 17.90 17.81 18.44 18.72
17.49 17.92 19.07 18.82 17.51 17.77 18.09 18.39
No data deleted
No. of samples 18 18 18 18 18 18 18 18
x
_
c
17.989 18.099 18.618 18.767 18.227 18.346 18.689 18.887
s
r
d
0.3622 0.1510 0.4722 0.3932 0.3325 0.2419 0.1972 0.3093
s
R
e
0.4079 0.3921 0.6275 0.6075 0.3241 0.3276 0.2788 0.3228
Initial method comparison (Laboratory 9 data deleted)
No. of samples 16 16 16 16 16 16 16 16
x
_
17.992 18.121 18.641 18.799 18.292 18.415 18.743 18.928
s
r
0.3000 0.1602 0.3852 0.3882 0.3296 0.2564 0.1899 0.3176
s
R
0.4001 0.4131 0.6254 0.6313 0.2774 0.2692 0.2360 0.3116
Revised method comparison (Laboratory 9 and outlier data deleted)
No. of samples 14 14 10 10 14 14 16 16
x
_
17.886 18.001 18.754 18.878 18.262 18.406 18.743 18.928
s
r
0.3193 0.1635 0.2198 0.1982 0.1771 0.1493 0.1899 0.3176
s
R
0.2676 0.2503 0.1830 0.2210 0.1766 0.2339 0.2360 0.3116
a
Data deleted from the initial and revised method comparisons.
b
Data deleted from the revised method comparison only.
c
Sample mean.
d
Repeatability standard deviation.
e
Reproducibility standard deviation.
Table 3. Collaborative study results for protein determination (%) in smoked ham (blind duplicates) by Kjeldahl
method: raw data and statistical summaries
Lab.
Sample
1 2 3 4
Cu Hg Cu Hg Cu Hg Cu Hg
1 19.50 20.02 13.44 13.45 17.50 17.73 14.58 14.53
19.83 19.64 13.24 13.38 17.22 18.05 15.20 14.67
2 19.80 19.85 13.12 13.11 17.71 17.69 14.76 14.50
19.64 19.58 13.44 13.41 17.44 17.76 14.04 14.62
3 19.17 19.46 13.13 13.25 17.15 17.23 14.26 14.47
19.40 19.28 13.43 13.18 17.12 17.22 14.36 14.16
4 19.49 19.72 13.21 13.39 17.21 17.48 14.28 14.45
19.53 19.90 13.01 13.34 17.27 17.54 14.34 14.37
5 19.79 20.10 13.55 13.72 17.34 17.68 14.63 14.83
19.70 19.96 13.51 13.63 17.40 17.71 14.75 14.80
6 19.57 20.37 13.74 13.75 17.71 18.11 14.53 15.24
19.67 19.79 12.99 13.94 17.27 17.92 14.61 15.11
7 19.51 19.95 13.04 13.87 17.86 17.63
b
15.37
b b
14.14
b
18.80 19.26 13.12 13.64 17.48 17.60
b
15.26
b b
15.11
b
8 20.00 20.01 13.12 13.18 17.68 17.96
b
15.00
b b
15.25
b
19.61 20.11 13.25 13.33 17.29 17.68
b
15.83
b b
15.60
b
9
a
19.30 20.12 14.26 14.10 17.16 16.97 15.09 15.29
19.50 19.59 14.47 14.66 17.36 17.23 15.35 15.52
No data deleted
No. of samples 18 18 18 18 18 18 18 18
x
_
c
19.545 19.817 13.393 13.574 17.398 17.622 14.792 14.814
s
r
d
0.3416 0.3871 0.3273 0.3980 0.2915 0.3799 0.5424 0.5237
s
R
e
0.2666 0.2972 0.4163 0.3903 0.2141 0.3012 0.4795 0.4461
Initial method comparison (Laboratory 9 data deleted)
No. of samples 16 16 16 16 16 16 16 16
x
_
19.563 19.812 13.271 13.473 17.416 17.687 14.738 14.741
s
r
0.2332 0.2651 0.2318 0.1178 0.2015 0.1192 0.3200 0.2759
s
R
0.2842 0.3105 0.2200 0.2623 0.2259 0.2585 0.4905 0.4285
Revised method comparison (Laboratory 9 and outlier data deleted)
No. of samples 16 16 16 16 16 16 12 12
x
_
19.563 19.812 13.271 13.473 17.416 17.687 14.528 14.646
s
r
0.2323 0.2651 0.2318 0.1178 0.2015 0.1192 0.3408 0.1406
s
R
0.2842 0.3105 0.2200 0.2623 0.2259 0.2585 0.2784 0.3170
a
Data deleted from the initial and revised method comparisons.
b
Data deleted from the revised method comparison only.
c
Sample mean.
d
Repeatability standard deviation.
e
Reproducibility standard deviation.
Table 4. Collaborative study results for protein determination (%) in pork sausage (blind duplicates) by Kjeldahl
method: raw data and statistical summaries
Lab.
Sample
1 2 3 4
Cu Hg Cu Hg Cu Hg Cu Hg
1 11.50 11.71 10.34 10.51
b
10.41
b b
11.17
b
10.46 10.35
11.70 11.68 10.17 10.37
b
11.46
b b
11.29
b
10.32 10.50
2 11.57 11.71 10.15 10.27 10.80 11.01 10.18 10.41
11.63 11.74 10.31 10.47 10.94 10.92 10.15 10.38
3 11.30 11.40 9.79 10.14 10.99 11.04 9.96 9.96
11.59 11.58 10.30 10.26 11.14 10.65 10.27 10.06
4 11.68 11.64
a
12.34
a a
12.23
a
10.99 11.02 10.03 10.22
11.48 11.68 9.94 10.13 10.66 11.02 10.17 10.29
5 11.53 11.84 10.26 10.50 11.10 11.26 10.20 10.38
11.48 11.62 10.49 10.39 11.22 11.40 10.14 10.22
6 11.69 11.61 10.40 10.45 11.00 11.03 10.07 10.40
11.57 11.67 10.12 10.34 11.00 11.07 9.97 10.28
7 11.45 11.34 10.01 10.21 10.63 10.93 10.10 9.85
11.88 11.57 10.16 10.09 10.71 10.92 10.08 10.27
8 11.67 11.34 10.47 10.49 10.69 11.12 10.48 10.45
11.88 12.04 10.64 10.24 11.17 11.59 10.35 10.29
9
a
10.94 10.84 10.18 10.78 9.64 10.32 9.61 9.98
12.00 12.23 10.33 10.46 11.03 10.57 10.47 10.24
No data deleted
No. of samples 18 18 18 18 18 18 18 18
x
_
c
11.586 11.624 10.357 10.463 10.866 11.018 10.167 10.252
s
r
d
0.2287 0.3222 0.5907 0.5103 0.3813 0.3596 0.2747 0.2417
s
R
e
0.2310 0.2906 0.5322 0.4705 0.3971 0.2919 0.2094 0.1740
Initial method comparison (2 aberrant points and Laboratory 9 data deleted)
No. of samples 16 16 15 15 16 16 16 16
x
_
11.600 11.636 10.237 10.324 10.932 11.090 10.183 10.269
s
r
0.1608 0.1985 0.1885 0.1118 0.3067 0.1614 0.1022 0.1324
s
R
0.1510 0.1756 0.2257 0.1460 0.2639 0.2230 0.1606 0.1807
Revised method comparison (Laboratory 9 and outlier data deleted)
No. of samples 16 16 15 15 14 14 16 16
x
_
11.600 11.636 10.237 10.324 10.931 11.070 10.183 10.269
s
r
0.1608 0.1985 0.1885 0.1118 0.1695 0.1695 0.1022 0.1324
s
R
0.1510 0.1756 0.2257 0.1460 0.1958 0.2259 0.1606 0.1807
a
Data deleted from the initial and revised method comparisons.
b
Data deleted from the revised method comparison only.
c
Sample mean.
d
Repeatability standard deviation.
e
Reproducibility standard deviation.
Table 5. Collaborative study results for protein determination (%) in cooked sausage (blind duplicates) by Kjeldahl
method: raw data and statistical summaries
Lab.
Sample
1 2 3 4
Cu Hg Cu Hg Cu Hg Cu Hg
1 11.88 11.93 12.16 12.30 12.57 12.78 12.25 12.33
11.61 11.70 12.09 12.46 12.23 12.59 11.87 12.28
2 11.70 11.82 12.08 12.26 12.46 12.52 11.95 12.35
11.67 11.93 12.02 12.21 12.61 12.69 11.79 12.22
3 11.69 11.52 11.84 11.89 12.60 12.36 11.90 11.65
11.50 11.61 12.17 12.23 12.29 12.31 12.19 12.19
4 11.59 11.80 11.82 11.84 12.38 12.38 12.01 12.15
11.53 11.82 12.00 12.12 12.48 12.55 11.86 12.00
5 11.70 11.93 12.02 12.23 12.51 12.55 12.09 12.21
11.82 11.65 12.06 12.07 12.55 12.69 12.24 12.32
6 11.75 11.99 12.31 11.92 12.71 12.55 11.85 12.31
11.94 12.09 12.23 12.32 12.40 12.73 11.95 12.20
7
b
12.11
b b
12.01
b
12.22 11.74 12.29 12.65 11.96 12.12
b
11.67
b b
11.00
b
12.44 11.80 12.56 12.63 12.42 12.31
8 12.04 11.95
b
12.37
b b
12.31
b
12.54 12.69 12.44 12.65
12.08 12.14
b
12.94
b b
13.23
b
12.90 12.58 12.27 12.20
9
a
12.09 12.02 13.27 13.10 11.91 12.22 11.22 11.51
12.15 12.18 11.41 11.69 12.34 12.37 10.71 11.47
No data deleted
No. of samples 18 18 18 18 18 18 18 18
x
_
c
11.807 11.838 12.192 12.207 12.463 12.547 11.943 12.137
s
r
d
0.2841 0.3343 0.3447 0.4367 0.2784 0.2136 0.2893 0.3584
s
R
e
0.2084 0.2733 0.4129 0.4144 0.2088 0.1545 0.4256 0.3009
Initial method comparison (Laboratory 9 data deleted)
No. of samples 16 16 16 16 16 16 16 16
x
_
11.768 11.806 12.173 12.183 12.505 12.578 12.065 12.218
s
r
0.1498 0.2759 0.1823 0.2803 0.1845 0.1003 0.1853 0.1930
s
R
0.1947 0.2757 0.2716 0.3572 0.1673 0.1373 0.2113 0.2064
Revised method comparison (Laboratory 9 and outlier data deleted)
No. of samples 14 14 14 14 16 16 16 16
x
_
11.750 11.849 12.104 12.099 12.505 12.578 12.065 12.218
s
r
0.1088 0.1189 0.1214 0.1714 0.1845 0.1003 0.1853 0.1930
s
R
0.1822 0.1815 0.1685 0.2228 0.1673 0.1373 0.2113 0.2064
a
Data deleted from the initial and revised method comparisons.
b
Data deleted from the revised method comparison only.
c
Sample mean.
d
Repeatability standard deviation.
e
Reproducibility standard deviation.
Table 6. Collaborative study results for protein determination (%) in dry cured ham (blind duplicates) by Kjeldahl
method: raw data and statistical summaries
Lab.
Sample
1 2 3 4
Cu Hg Cu Hg Cu Hg Cu Hg
1 23.68 23.83 24.59 24.66 22.74 22.80 24.07 25.24
23.38 24.22 24.36 24.82 22.26 22.68 24.08 25.07
2 23.46 23.90 24.33 24.65 22.32 22.65 24.22 24.77
24.10 23.98 24.69 25.12 22.36 22.62 24.26 24.41
3 22.89 23.30 24.58 24.59 22.41 22.56 24.12 24.48
23.32 23.47 23.77 24.23 21.95 21.93 24.31 23.84
4 23.58 23.71 24.65 24.70 22.00 22.01 24.26 24.55
23.73 23.95 24.17 24.48 22.26 22.15 23.92 24.14
5 23.82 23.92 24.52 24.31 22.49 22.59 24.55 24.58
23.84 24.08 24.38 24.50 22.64 22.93 24.26 24.50
6 23.52 24.00 24.38 24.67 22.30 22.72 23.56 24.47
23.66 24.12 24.79 24.53 22.32 22.72 24.32 24.47
7 23.36 23.72 24.02 24.18 21.75 22.63 23.59 23.80
23.13 24.00 24.30 24.99 22.52 23.04 24.24 24.06
8 23.88 24.30 24.61 24.52 22.76 22.76 24.01 23.77
24.11 24.13 24.51 24.55 22.92 22.60 24.34 24.29
9
a
23.66 23.69 22.75 24.10 21.48 21.64 24.27 24.70
23.54 23.83 22.39 22.70 22.15 22.38 22.43 22.93
No data deleted
No. of samples 18 18 18 18 18 18 18 18
x
_
b
23.584 23.894 24.211 24.472 22.313 23.523 24.045 24.324
s
r
c
0.3886 0.3184 0.2862 0.4078 0.2983 0.4402 0.1293 0.4012
s
R
d
0.3074 0.2489 0.6654 0.5135 0.3646 0.3646 0.4876 0.5363
Initial method comparison (Laboratory 9 data deleted)
No. of samples 16 16 16 16 16 16 16 16
x
_
23.591 23.914 24.416 24.594 22.375 22.587 24.132 24.402
s
r
0.2281 0.1566 0.2899 0.2660 0.2684 0.2152 0.2901 0.2599
s
R
0.3385 0.2694 0.2620 0.2473 0.3119 0.3116 0.2614 0.4285
Revised method comparison (Laboratory 9 and outliers data deleted)
e
No. of samples 16 16 16 16 16 16 16 16
x
_
23.591 23.914 24.416 24.594 22.375 22.587 24.132 24.402
s
r
0.2281 0.1566 0.2899 0.2660 0.2684 0.2152 0.2901 0.2599
s
R
0.3385 0.2694 0.2620 0.2473 0.3119 0.3116 0.2614 0.4285
a
Data deleted from the initial and revised method comparisons.
b
Sample mean.
c
Repeatability standard deviation.
d
Reproducibility standard deviation.
e
The further data analysis did not identify any outliers. Therefore, the initial and revised method comparisons gave identical results.
Table 7. Geometric mean (GM) of the ratio of the
mercury and copper standard deviations
Statistic GM CV, %
a
RSD
r
0.995 5.6
RSD
R
0.829 7.8
a
Coefficient of variation of the GM.
Figure 1. Difference (Hg–Cu) vs mean (Hg).
Figure 2. Plot of reproducibility: Hg vs Cu.
Figure 3. Normalized standard deviations: Hg vs Cu. Each numeral identifies the laboratory that performed the
analysis. Note that 52 observations are coincident and therefore hidden.
Figure 4. Difference of normalized means (Hg–Cu) vs sum of normalized means. The asterisks describe a circle with
a radius of 4 (its shape is distorted into an ellipse because of the difference in scale between the x- and y- axes). Any
point outside of this circle represents an outlier with a level of significance less than 0.10. Each numeral identifies the
laboratory that performed the analysis.