Vous êtes sur la page 1sur 14

Exploration and Mining Geology, Vol. 17, Nos. 34, p.

131144, 2008
2009 Canadian Institute of Mining, Metallurgy and Petroleum.
All rights reserved. Printed in Canada.
0964-1823/00 $17.00 + .00

Quality Control of Assay Data: A Review of Procedures for


Measuring and Monitoring Precision and Accuracy
M. Abzalov1,*,
(Received August 30, 2007; accepted September 9, 2008)

Abstract Control of analytical data quality is usually referred to in the mining industry as Quality As-
surance and Quality Control (QAQC), and involves the monitoring of sample quality and quantification
of analytical accuracy and precision. QAQC procedures normally involve using sample duplicates and
specially prepared standards whose grade is known. Numerous case studies indicate that reliable control
of sample precision is achieved by using approximately 5% to 10% of field duplicates and 3% to 5% of
pulp duplicates. These duplicate samples should be prepared and analyzed in the primary laboratory.
Bias in the analytical results can be identified by inclusion of 3% to 5% of the standard in each sample
batch. Several different standards are used, with values spanning the practical range of grades in the actual
samples. A blank (a sample in which the concentration of metal of interest is below detection limit) should
also be included. Standard samples alone cannot identify biases introduced during sample preparation, and
therefore approximately 5% of the duplicate samples (coarse rejects and pulp) should be processed and
assayed at another, external, reputable laboratory.
This paper discusses techniques used for estimation of errors in precision and accuracy, and over-
views diagnostic tools. It is shown that one of the most commonly used methods, the Thompson-Howarth
technique, produces consistently lower results than other methods. These results reflect the nature of this
method, which relies on the assumption of a normally distributed error, and thus produces biased results
when errors have a skewed distribution. This study concurs with the suggestion of Stanley and Lawie
(2007: Exploration and Mining Geology, v. 16, p. 265274) to use the average coefficient of variation
(CVAVR(%)) as the universal measure of relative precision error in mine geology applications:

2 N (a - b )2
i
CVAVR (%) = 100
N
(a + b )
i
2
i =1 i i

Based on case studies, an acceptable level of sample precision is proposed for several different deposit
types. 2009 Canadian Institute of Mining, Metallurgy and Petroleum. All rights reserved.

Key Words: Mining geology, QAQC, duplicates, standards.

Sommaire Le contrle de la qualit des donnes analytiques est habituellement dsign sous lap-
pellation dAssurance et Contrle de la Qualit (ACQ), elle implique le suivi de la qualit des chan-
tillons et la quantification de la reproductibilit et la prcision analytiques. Les procdures des pro-
grammes dACQ impliquent habituellement lusage de duplicata et de standards analytiques dont
la teneur est connue. Plusieurs tudes de cas indiquent quune bonne fiabilit du contrle de la prci-
sion des chantillons est atteinte par lusage de 5% 10% de doublons sur le terrain et de 3% 5%
de doublons de pulpe. Ces duplicata devraient tre prpars et analyss au laboratoire principal.
Un biais des rsultats analytiques peut tre identifi par linclusion de 3% 5% dchantillons standard
dans chaque envoi dchantillons. Plusieurs standards sont utiliss, avec des valeurs couvrant ltendue
des teneurs habituellement rencontres dans des chantillons rels. Un blanc (un chantillon dans lequel la
concentration dun mtal donn est sous la limite de dtection) devrait aussi tre inclus. Lemploi exclusif
dchantillons standards ne permet pas lidentification de biais rsultant de la prparation des chantillons,
de telle sorte quenviron 5% des duplicata (rejets grossiers et pulpes) devrait tre traits a un autre labo-
ratoire indpendant et fiable.
Cet article passe en revue les techniques utilises pour valuer lerreur lie la prcision et a la repro-
ductibilit, et effectue aussi un survol de divers outils diagnostiques. Il ressort de cet examen quune des

1
MASSA Geoservices, Mt. Claremont, Western Australia, Australia WA6010.
* Present address: Rio Tinto Exploration, PO Box 175, Western Australia, Australia WA6984.

Corresponding Author: E-mail: marat.abzalov@riotinto.com


132 Exploration and Mining Geology, Vol. 17, Nos. 34, p. 131144, 2008

techniques les plus frquemment utilises, la mthode Thompson-Howarth, fournit des estims systmati-
quement infrieurs ceux provenant des autres mthodes. Ces rsultats sont un reflet de la nature de cette
mthode, laquelle assume une distribution normale de lerreur, biaisant ainsi les rsultats lorsque lerreur
prsente une distribution asymtrique. La prsente tude recommande de privilgier lusage du coefficient
moyen de variation (CVAVR(%)) en tant que mesure universelle de la prcision relative de lerreur dans les
applications minires tel que suggr par Stanley and Lawie (2007: Exploration and Mining Geology, v.
16, p. 265274):
2 N (a - b )2
i
CVAVR (%) = 100
N
(a + b )
i
2
i =1 i i

Appalaches du nord. Les gtes aux environs de Clarence Stream dans le sud-ouest du Nouveau
Brunswick sont un exemple de ce type de gisement, mettant en valeur ce qui pourrait devenir un district
aurifre majeur dans la rgion. 2009 Canadian Institute of Mining, Metallurgy and Petroleum. All
rights reserved.

Introduction ple can be a second blast-hole sample collected from the


same blast-hole cone, or another split (e.g., quarter) of the
The quality of geochemical data used in mineral ex- drill core. It also can be a duplicate sample collected at a
ploration and development is determined and reported in particular stage of the sampling protocol. Where the sam-
terms of precision and accuracy (JORC, 2004). Precision, pling protocol includes several stages of comminution and
or repeatability, is a measure of how close replicate sample subsampling, duplicate samples should be taken at every
values are to each other, whereas accuracy is a measure of subsampling stage.
how close the values are to the true grade. The procedures Duplicate samples are prepared and assayed using the
involved in monitoring precision and accuracy are referred same procedures that were used for the original samples
to as Quality Assurance and Quality Control (QAQC). and in the same laboratory. Variations of the results be-
Despite unanimous recognition of the importance of tween matching pairs of sample assays, when assayed in
QAQC in the minerals industry, no uniform, industry-wide the same laboratory, allow estimation of precision errors
procedures have been established, and the methods used associated with sample preparation and analysis.
vary significantly in complexity and effectiveness. Given
the growing need in the mining and exploration industry Accuracy
for development of a standardized approach to quantify- Accuracy of analyses is normally monitored by includ-
ing the quality of assay values, it is important to compare ing samples with known concentrations of elements (stan-
the currently used QAQC techniques. Such a comparative dards) with each assay batch. Standards are available com-
analysis will assist in developing standardized QAQC pro- mercially that have been prepared and assayed rigorously
cedures, which, in turn, will facilitate assessment of the following appropriate procedures (Kane, 1992), and the
technical merits of a mining project. results statistically certified. Best practice is to use stan-
This paper discusses techniques used in the mining in- dards from material mineralogically similar to that of the
dustry for monitoring precision and accuracy of geochem- studied mineralization; these are called the matrix-matched
ical analyses of samples, with a particular emphasis on standards.
the statistical methods used for QAQC data analysis. The Possible contamination during assaying is controlled by
most common estimators of precision errors are compared including blanks with the sample batch, which are materi-
by applying them to the same data sets, which have been als that have very low grades of the metal of interest, usu-
collected from operating mines and mining projects and ally below the detection limit of the analytical method to
which represent different commodities and mineraliza- be used. It is common practice to use barren quartz as the
tion styles. In order to facilitate statistical analysis of the blank material.
QAQC data, Visual Basic codes have been developed in Standards and blanks are called reference materials, and
an Excel spreadsheet (AMZ_QAQC.xls), which can be ob- are used for detecting possible biases in reported analytical
tained from the author upon request. values, and for quantifying the accuracy errors. Leaver et
al. (1997, p. 3) defined a reference material as a substance
Principles of Sample Quality Control for which one or more properties are established suffi-
ciently well to calibrate a chemical analyser or to validate a
Precision measurement process. A reference material is classified as
Analytical precision is normally monitored by using certified if it was issued and certified by an agency such as
paired analyses. The first sample is referred to as the ori- the Canadian Centre for Mineral and Energy Technology
ginal sample, and the second is called the duplicate. A (CANMET), the National Institute of Standards and Tech-
duplicate sample is merely another sample collected from nology (NIST), or other government agencies whose tech-
the same place following the same rules as used for col- nical competency is internationally recognized (Leaver et
lecting the original sample. In practice, the duplicate sam- al., 1997).
Quality Control of Assay Data: A Review of Procedures for Measuring and Monitoring Precision and Accuracy M. Abzalov 133

An alternative approach to measuring accuracy is to re- ever, the position of the duplicates within a given analytical
peat the assay of sample duplicates in an external laboratory batch should not be immediately sequential to the original
that has the authority to act as an independent auditor. This samples, and should also not be systematically related to
technique has its strengths and weaknesses. One strength is them. Numbering of the duplicate samples should disguise
that it allows detection of errors introduced at the sample them in the analytical batch.
preparation stage, which commonly cannot be recognized
by use of reference materials in a single laboratory. Un- Frequency of Inserting QAQC Material in Assay Batches
like standards, which check for bias only in the analytical The number of quality control samples and the fre-
protocol, the independent laboratory check permits the quency of their insertion in analytical batches should be
recognition of bias at various levels of the entire sampling sufficient for systematic monitoring of assay quality. Rec-
protocol, depending on the nature of the duplicate material ommended quality control materials vary from 5% to 20%
submitted to the second lab. Of course, if bias is detected, of the total analyses (Garrett, 1969; Taylor, 1987; Valle
further analyses must be undertaken (perhaps at a third lab- et al., 1992; Leaver et al., 1997; Long, 1998; Sketchley,
oratory) to determine the source of the errors. 1998) depending on mineralization type, location of the
mining project, and stage of the project evaluation. A brief
QAQC Procedures and Practices overview of the different recommendations on frequency
of insertion of QAQC materials is given below:
Planning and Implementation of QAQC Programs Garrett (1969) recommended that approximately 10% of
Practical recommendations regarding optimal distribu- geochemical samples should be controlled by collection
tion of QAQC material in analytical batches have been dis- of duplicate samples.
cussed by many researchers (e.g., Thompson and Howarth, Taylor (1987) recommended that from 5% to 10% of
1978; Valle et al., 1992; Long, 1998; Sketchley, 1998). samples analysed by a laboratory should be reference
Generally accepted rules are summarized below. materials.
Standards (with a range of compositions representative Leaver et al. (1997) suggested analyzing one in-house
of the practical range of grades in the actual samples) and reference material with every twenty assayed samples,
blanks should be inserted in every sample batch in such a including at least one certified standard in the same
way that they are anonymous to the analyst, and distributed batch.
to allow detection of possible bias in the results. Valle et al. (1992) recommended that at least 10% of
Duplicate samples should be chosen in such a manner the determinations in exploration or mining projects
that significant stages of a sampling protocol are properly should be QAQC samples, including standards, blanks,
addressed, and the precision errors associated with various and duplicates.
subsampling stages are estimated adequately. Some dupli- Long (1998) suggested that at least 5% of pulps (crushed
cates should be chosen at random, and others should cover and pulverized sample material) should be checked by
the entire range of expected grades in the deposit. The an independent laboratory; 5% of field and/or coarse
disadvantage of randomly selecting all duplicates is that rejects should have a second pulp prepared and ana-
such an approach does not lead to adequate representation lyzed by the primary laboratory; and every sample batch
of all different grade classes or mineralization types, and should include from 1% to 5% of standard reference ma-
might not assure adequate spatial coverage of the deposit. terials, such as certified standards and blanks.
Commonly, an excessively large proportion of randomly Sketchley (1998) recommended that every sample batch
collected duplicate samples will be taken from barren or submitted to an analytical laboratory should include
low-grade material, leaving the higher grade mineraliza- from 10% to 15% of quality control samples. In particu-
tion poorly tested. lar, every batch of 20 samples should include at least one
It is important to ensure that duplicate samples are rep- standard, one blank, and one duplicate sample.
resentative for the given deposit, but that at the same time
their distribution maintains, at least partially, the charac- It is a good practice, in addition to using matrix-matched
teristics of being randomly chosen. To meet these two dif- standard samples, to analyze approximately 5% of the pulp
ferent criteria, the author, based on personal experience, duplicates and coarse rejects at a reputable external labora-
recommends using the following approach in selecting tory that is authorized to act as independent auditor. Use
duplicate samples. Approximately half of the collected of an external laboratory is important because standard
duplicates should be located by a project geologist with in- samples alone cannot identify biases introduced at differ-
structions to ensure that these duplicates properly represent ent stages of sample preparation within a given labora-
the various mineralization styles, grade ranges, and spatial tory. Anonymity of the standards is another issue, because
coverage of the entire deposit. The other half of the dupli- standards can be easily recognized in the sample batches
cates should be chosen at random in each assay batch. and treated more thoroughly by laboratory personnel. The
Duplicate samples should be taken at the same time, author believes that assaying duplicate samples (pulp and
and using the same analytical procedures, as the original coarse rejects) in an external, reputable laboratory should
samples. It is important that the original samples and their be routine practice at the feasibility stage of a mining pro-
duplicates are included in the same analytical batch to per- ject, where accurate data are critical in economic evalua-
mit using them for within-batch precision studies. How- tion.
134 Exploration and Mining Geology, Vol. 17, Nos. 34, p. 131144, 2008

Statistical Estimation of Precision Errors level:


PTH = SC C . (2)
Techniques Based on Analysis of Duplicates
The intuitive approach in assessing paired analytical data Substituting Sc as defined in Equation 2 into Equation
is to plot duplicate sample assays (ai) versus original assays 1 and multiplying by 1001 yields an expression for rela-
(bi). Ideally, the paired data representing the independent tive error of precision (PTH(%)), estimated at one standard
assays of the same material should be similar, and should deviation confidence level and expressed as a percentage
therefore plot close to line y = x on a binary diagram. In of grade (C):
practice, the values of the paired data are rarely identical PTH (%) = 100 ( S0 C ) + K , (3)
due to sampling and analytical errors, and the data points
will be scattered to a varying degree about the y = x line. If where the parameters Sc and K have to be deduced empiric-
systematic differences (bias) exist between the original and ally from available data pairs. The relative precision error
duplicate assays, the data might define a more complicated (PTH(%)) calculated at the mean grade ( C ) of the studied
functional relationships (e.g., y = mx + c). The precision samples is given by Equation 4:
(error) can be quantified from the paired data through as-
sessing the scatter of the data points, with corrections for N N
a + b
their deviation from the y = x line.
i

i
i
i

Various methods are available for estimating precision C= . (4)
2N
from the paired data (Garrett, 1969; Thompson and How-
arth, 1973, 1976, 1978; Howarth and Thompson, 1976; Thompson and Howarth (1973, 1976, 1978) have pro-
Bumstead, 1984; Shaw, 1997; Francois-Bongarcon, 1998; posed two approaches for the experimental calculation of
Sinclair and Blackwell, 2002; Stanley and Lawie, 2007a). these parameters. The first approach is used when 50 or
The most common methods are based on the assumption more duplicated results are available (Thompson and How-
of a linear relationship between sampling and analytical arth, 1976, 1978). Using N (>50) of the data pairs ai and bi
errors, and concentration of the assayed metals (Thompson (i = 1,2,...N), the means of the data pairs (ai + bi)/2 and their
and Howarth, 1973, 1976, 1978; Howarth and Thompson, absolute differences |ai bi| are calculated. The data pairs
1976; Bumstead, 1984; Shaw, 1997). Francois-Bongarcon are sorted in increasing order of their mean grade, and are
(1998) has extended this principle to the more complex re- subdivided into groups of 11 results. The last group, ac-
lationships described by a quadratic model. A special type cording to Thompson and Howarth (1978, p. 25), should
of linear model is the Reduced Major Axis model (RMA; be ignored if it contains less than 11 records. The average
Sinclair and Bentzen, 1998; Davis, 2002; Sinclair and value of the means for each sample pair, and the median
Blackwell, 2002). This technique is applicable to linear value of the absolute differences are calculated for each
cases when different sets of paired data exhibit systematic group of data. Median values are then plotted as a function
differences (bias). There have also been attempts (e.g., of the group means, and a regression line is fitted to the
Garrett, 1969) to use geostatistical estimates (variograms) experimental points (Fig. 1). The slope of the regression
as a measure of precision. Pitard (1998) suggested a special line in Figure 1 represents K, and S0 is the intercept of the
diagnostic diagram, the Relative Difference Plot (RDP), regression line with the y-axis (i.e., at a mean of zero). In
which can be used for detailed analysis of the causes of practice, unless N >> 50, fitting regression lines to so few
poor data precision or sample bias. These methods are re- points commonly leads to peculiar, or even impossible, re-
viewed below. sults (e.g., a negative intercept with the y-axis).
Method 1: Thompson-Howarth: The Thomson-Howarth The Thompson-Howarth method has three constraints:
method was developed in the early 1970s (Thompson (1) all measurements must be positive, which is an intrinsic
and Howarth, 1973, 1976, 1978; Howarth and Thomp- characteristic of geochemical assays; (2) the slope of the
son, 1976) and is now widely used in the mining industry. regression line (K) must be positive; and (3) the intercept
The method assumes that error is normally distributed, of the regression line with y-axis (S0) must also be positive.
and that a systematic linear variation in error exists as a If one or more of these criteria are not met, the method is
function of sample grade: not applicable as described.
When the number of duplicate samples is less than 50,
Sc = S0 + KC , (1) parameters S0 and K cannot be reliably estimated, and
therefore the above method cannot be used to estimate pre-
where Sc (plotted on the y-axis) is the error measured as
the standard deviation of the duplicates, S0 is the intercept
with the y-axis representing the standard deviation at zero 1
Thompson and Howarth (1973) have used a coefficient of 200 in Equa-
concentration, K is the slope of the line, and C is the grade tion 3, which corresponds to precision at 2 standard deviations. In the
(x-axis). This method assumes no bias (i.e., that the data current paper, all precision errors are calculated and discussed at 1 stan-
follow a normal distribution) and, therefore, must be pre- dard deviation confidence level; therefore, in Equation 3, a constant of
100 was used. This change was made in order to avoid confusion when
ceded by a test for bias. PTH(%) is compared to other conventional estimates, such as coefficient
Relative error (PTH ) at a given grade (C) can be deter- of variation (CV(%)) representing relative error at 1 standard devia-
mined as the precision at one standard deviation confidence tion.
Quality Control of Assay Data: A Review of Procedures for Measuring and Monitoring Precision and Accuracy M. Abzalov 135

1 N
si2 2 N (ai - bi ) 2
CVAVR (%) = 100
N
m
i =1
2
= 100
N
(a + b )
i =1
2 . (6)


i i i

Hence, statistics such as AMPD (Equation 7) and HARD


(Equation 8) simply represent the CV(%) multiplied by
constants 2 and 2 , respectively (Stanley and Lawie,
2007a): 2

ai - bi
AMPD = 100 = 2 CV (%) , (7)
(ai + bi ) 2

ai - bi 2
HARD = 100 = CV (%) . (8)
(ai - bi ) 2

Fig. 1. Absolute difference vs. mean of data pairs (for N > 50 data pairs).
Vertical lines separate groups of 11 records. Open squares denote the Stanley and Lawie (2007a) correctly noted that using
experimental points calculated for each group of 11 records. The x-axis such statistics, which are directly proportional to the co-
value of a square corresponds to the average value of the pair means, and efficient of variation, offers no more information than the
the y-axis value corresponds to the median of the absolute differences of CV itself.
the data pairs in a group. The solid bold line is a regression line fitted to
the experimental points. Method 3: Geostatistical Approach: An alternative approach
to estimating the precision of analytical data involves calcu-
cision. To overcome this problem, Thompson and Howarth lation of various geostatistical measures of the spatial cor-
(1978) suggested a second graphical method that infers relation, such as variograms. Garrett (1969) suggested cal-
the approximate value of the relative precision error. This culating precision error between original samples and their
method has limited application in practice because of dif- duplicates by lognormally transforming their assay values,
ficulties in reliably estimating sample precision from such and then applying the formula of the lognormal variogram.
a limited number of data pairs. Consequently, this method This approach has not found wide application in practice.
is not discussed further here. F. Pitard has suggested (pers. commun., 2005) calculat-
2
Method 2: Relative Precision Error (CV, AMPD, HARD): ing relative variance ( s FP ) of the matching pairs of data:
This group of methods includes various estimators based 2
on calculation of the absolute differences between origin-
2
al and duplicate samples, normalized to (divided by) the 1 N ai - bi 2 N a -b
mean grades of corresponding data pairs (Bumstead, 1984; s 2
FP =

2 N i ai + bi
=
N
i ai + bi . (9)
i i
Shaw, 1997; Roden and Smith, 2001; Stanley and Lawrie, 2
2007a). Two of the measurements, Absolute Mean Per-
centage Difference (AMPD) and Half Absolute Relative
Difference (HARD), both described in Table 1, are widely This equation is known in geostatistics as a pair-wise
used in the mining industry. relative variogram (Goovaerts, 1997), and is particularly
useful for variables characterized by strongly skewed sta-
Stanley (2006) has shown that the abso- tistical distributions, and where outliers are present3.
lute difference between paired data is dir- The average precision error can be obtained from Equa-
ectly proportional to the standard deviation tion 9 by taking the square root of the relative variance
( | ai bi = 2 Standard Deviation ), and, consequently, and multiplying by 100 to represent the precision error as a
that the various measurements representing the ratio of the percentage at one standard deviation (Equation 10)4:
absolute differences between paired data to the means of
the corresponding pairs (Table 1) are directly proportional
to the coefficient of variation (CV(%))2:
3
Conventional formula of the pair-wise relative variogram (Goovaerts,
Standard Deviation a - bi 2 . (5) 1997):
CV (%) = 100 = 100 i 1 N
2
Z ( xi ) - Z ( xi + h)
Mean ai + bi

2
g PWR (h) =
Z ( xi ) + Z ( xi + h) ,
2
2N i =1

2
The average coefficient of variation (CVAVR(%)) can be
calculated from N pairs of duplicate samples: where Z(x) is a value of variable (Z) at the location x, and h is a vector
separating Z(x) from Z(xi + h) points.

2
For consistency with other measurements, the coefficient of variation, 4
When the formula of the pair-wise relative variogram (Equation 9) is
CV (one standard deviation divided by mean), is expressed as a percent- applied to duplicated pairs of samples whose spatial separation is zero
age (CV(%)). (h = 0), it calculates the PWR(0) value.
136 Exploration and Mining Geology, Vol. 17, Nos. 34, p. 131144, 2008

2 .
PFP (%) = 100 s FP (10) Table 1. Measures of Relative Error Based on Absolute Difference Between Duplicate Data Pairs
Single Duplicate
Error Estimator Pair Formula Reference Comments
Method 4: Partitioning of Total Vari-
ance of Analytical Data: Francois- Also known as Mean Percentage
Absolute Mean ai - bi Difference (MPD; Roden and Smith,
Bongarcon (1998) analysed sampling Percentage Dif- AMPD = 200
and analytical precision by subdividing (ai + bi ) Bumstead 2001) or Absolute Relative Difference
ference (AMPD) (1984) (ARD; Stanley and Lawie, 2007a).
the total precision error variance into
components. The practical recommen- Half Absolute ai - bi
Relative Differ- HARD = 100
dation of that study was to use Equation (ai + bi ) Shaw
ence (HARD) (1997) Half of the corresponding AMPD.
11 for estimating precision variance
if the data represent different grade Notes
classes. Error at one standard devia- ai = original sample; bi = duplicate.
tion is estimated using Equation 12:
(ST (%)), which were recommended by the State Ore Re-
(a - bi ) serves Committee as acceptable precision errors for given
2
s DFB = 2Var i , (11) grade classes (Eremeev et al., 1982). To obtain a reliable
(ai + bi ) estimation of sample precision, it was recommended to
analyze at least 30 data pairs for each grade class (Eremeev
2
PDFB (%) = 100 s DFB . (12) et al., 1982). Where samples have been assayed in the dif-
ferent laboratories, the data are grouped and analysed sep-
Calculation of other components of the global vari- arately for each laboratory.
ance is beyond the scope of the current review and can be
found in the abovementioned paper of Francois-Bongarcon Method 6: Reduced Major Axis (RMA): The Reduced
(1998). Major Axis (RMA) method is a linear regression tech-
nique that takes into account errors in two variables,
Method 5: Precision Estimation Technique Used by Rus- the original samples and their duplicates (Sinclair and
sian GeologistsInternal Quality Control: An approach Bentzen, 1998; Davis, 2002; Sinclair and Blackwell,
developed in the former USSR (Eremeev et al., 1982) 2002). This technique minimizes the product of the devi-
groups data by grade class, and calculates precision er- ations in both the x and y directions, which, in effect,
rors (SR(%)) separately for each class. The precision error minimizes the sum of the areas of the triangles formed
is calculated as a ratio of the standard deviation (S) of the by the observations and a fitted linear function (Fig. 2).
duplicated data to the mean of the data pairs ( C ; Equation The general form of the RMA is
4) in a given grade class (Equation 13):
b = W0 + W1a S RMA , (14)
N

i (ai - bi ) 2 N
2
where ai and bi are matching pairs of data plotted on the
S
S R (%) = 100 = 100 , (13) x- and y-axes, respectively, W0 is the y-axis intercept of the
C N N RMA linear model, W1 is the slope of the model relative
a + b 2 N

i
i i i to the x-axis, and SRMA is the standard deviation of the data
points around the RMA line. The slope of the RMA line
where N is number of sample pairs, and ai is the original (W1) is estimated as the ratio of standard deviations (SD) of
and bi the duplicate of the ith sample pair. It is noteworthy the values ai and bi:
that Equation 13 is the square root of the relative variogram
SD (bi )
applied to the duplicated samples, expressed as percent- W1 = . (15)
age5. SD (ai )
The approach used in Russia and some CIS countries
includes comparison of the precision errors (SR (%)) The intercept of the RMA model with the y-axis (W0) is
calculated in Equation 13 with the theoretical values estimated from

W0 = Mean (bi ) -W1Mean (ai ) . (16)


5
The conventional formula for the relative variogram is as follows (Goo-
vaerts, 1997): The RMA model allows quantification of the errors be-
N
2
Z ( xi ) - Z ( xi + h) tween matching data pairs (Equations 1720). Dispersion
1 ,
g R (h) =
2N
m2
of the data points about the RMA line (SRMA) is estimated
i=1
from
where Z(x) is a value of variable Z at the location x, h is a vector sepa-
rating Z(x) from Z(xi+h) points, and m is the mean of the variable Z(x). S RMA = 2 (1- r )(Var (ai ) + Var (bi )), (17)
When this formula is applied to duplicated pairs of samples whose spa-
tial separation is zero (h = 0), it calculates the value. where r is the correlation coefficient between the values ai
Quality Control of Assay Data: A Review of Procedures for Measuring and Monitoring Precision and Accuracy M. Abzalov 137

1989; Kane, 1992; CANMET, 1998). The purpose of these


tests is to demonstrate that analytical accuracy is compar-
able to the certified value of a standard and, hence, is ac-
ceptable with 95% confidence.

Statistical Test 1: Estimation Accuracy by Repeat Analy-


ses of Certified Standards, Single Laboratory Case
The most common situation is when one or several cer-
tified standard samples have been included in batches of
samples analyzed at the same laboratory. In such a case,
when several repeat assays of a given standard sample are
available, the analytical accuracy is assessed using statis-
tical tests. A common, and valid, approach is the statistical
t-test. An alternative method is presented in Equation 21
Fig. 2. Scatter diagram and the RMA model (thick line) fitted to paired Fe
grades (wt.%) of blast hole samples. A line with 1:1 ratio (thin) is shown (ISO Guide, 1989; CANMET, 1998):
for reference. The gray triangle shows an area formed by projection of the
data points to the RMA line.
SW2
m - m 2 s L2 + , (21)
and bi, and Var are variances of the values. n
The error on the y-axis intercept (S0) is estimated using where is the certified mean of a given standard sample
Equation 18, and the error on the slope (SSLOPE) is estimated (provided by the maker of the standard); L is the certi-
using Equation 19 for N data pairs: fied inter-laboratory standard deviation of a given standard
sample; m is the arithmetic mean of the replicate analyses
of this certified standard sample in assay batches; SW is the
(1- r ) Mean (ai )
2 estimated intra-laboratory standard deviation of the repli-

S0 = SD (bi ) 2 +
(1 + r ) , (18) cate analysis of the standard sample included in the assay
N
SD ( a )
i
batch; and n is the number of replicate assays of a given
certified standard in the analytical batches. If this condition
is satisfied, then the analytical results are considered ac-
SD (bi ) (1- r 2 ) . ceptable with regard to accuracy.
S SLOPE = (19) The base formula (Equation 21) can be simplified be-
SD (ai ) N cause the usual empirical relationships is that L 2SW
(CANMET, 1998). Consequently, for a large number of
The relative precision error (PRMA (%)) can be estimated replications (n > 10), the above condition can be written
from the RMA model6: as (ISO Guide 33, 1989)
2
S RMA
m - m 2s L. (22)
PRMA (%) = 100 2 , (20)
C If the certified value of the inter-laboratory standard
where C is the mean grade of the paired data (Equation deviation (L) of the reference material is not available,
4). then Equation 22 can be further simplified as (CANMET,
The RMA model is a useful tool for testing duplicated 1998)
data for the presence of bias. For example, the errors on
the slope estimate and/or intercept can be used to test if the m - m 4 SW . (23)
slope of the estimated RMA function is a sampling possi-
bility of one, and if the RMA function intercepts the y-axis Note, however, that the assumption L 2SW might not
at zero. If at least one of these criteria is rejected, then bias be valid in all cases (CANMET, 1998).
is indicated.
Statistical Test 2: Intra-Laboratory Precision, Single
Statistical Tests for Estimation of Performance Laboratory Case
of Certified Standards Comparison of the arithmetic mean of the replicate
analyses of a certified standard against its certified mean
The accuracy of laboratory analyses can be assessed by should be supported by estimation of the standard devia-
statistical tests on certified standard samples (ISO Guide, tion of the replicate analyses of the standard samples. This
is important because poor precision of the standard sample
assays (i.e., if SW is excessively large) can prohibit reliable
6
For consistency with other estimates discussed in this paper, PRMA(%)
detection of analytical biases.
is estimated at 1 standard deviation (i.e., multiplied by 100 in Equation The most common such situation occurs when multiple
20). analyses of a certified standard have been made in the same
138 Exploration and Mining Geology, Vol. 17, Nos. 34, p. 131144, 2008

laboratory. These data are used for calculating the standard against sequential numbers of the pairs (x-axis); calculated
deviation (SW) of replicate analyses of the certified stan- average grades of the data pairs are shown on the second-
dard, which should be compared with the corresponding ary y-axis (Fig. 3). A more traditional approach would be
certified value (C). The analytical precision of a laboratory to plot RD(%) values against the average grades of the
is considered acceptable if results satisfy the following sta- data pairs. However, in that case, the diagnostic ability of
tistical test (ISO Guide 33, 1989): this diagram is marred by uneven distribution of the data
between grade intervals. Instead, plotting RD(%) values
2 X (2n-1);0.95 against sequential numbers of the data pairs overcomes
( SW sC ) , (24) this problem. The link between average grades of the data
n -1 pairs and their sequential numbers is established by adding
a calibration curve to the RDP diagram (Fig. 3). RD(%)
where SW is the standard deviation of replicate analyses values usually exhibit a large range of variation; therefore,
of the standard; C is the certified value of the intra-lab- interpretation of this diagram can be facilitated by apply-
oratory standard deviation of the standard; and X (2n-1);0.95 is ing the moving window technique to smooth the initially
the critical value of the 0.95 quartile ( = 0.05) of the (X2) calculated RD(%) of the data pairs (Fig. 3).
distribution at (n 1) degrees of freedom, where n is the The example in Figure 3 is based on data collected from
number of replicate analysis of the standard. a massive Cu-sulfide project in Russia. Approximately 140
In practice, at least three repeat analyses of the certi- core duplicate samples were analysed in an external reput-
fied standard should be available for this test (CANMET, able laboratory as part of the project due diligence. The
1998). results, when plotted on an RDP diagram (Fig. 3), show
that copper assays of low-grade samples (Cu < 1.1%) are
Diagnosis of Errors biased. Assayed values have significantly underestimated
the true copper grade of those samples. Another fea-
When poor-quality analytical data have been recognized ture revealed by this diagram is that low-grade samples
from excessive variance of the paired data, or sample bias- (Cu < 1.1%) exhibit excessive precision error that is not
es have been detected, it is necessary to identify the cause shown by higher grade samples (Cu > 1.1%). These find-
of the error. In this section, two types of diagram are intro- ings triggered a special investigation of the laboratory pro-
duced that are particularly useful for preliminary diagnosis cedures, with a particular emphasis on differences between
by geological personnel, prior to calling for help from a high-grade and low-grade samples.
specialized sampling consultant. The RDP diagram can be used for testing the impact
The first diagram, introduced by Pitard (1998), is known of data precision against different factors such as sample
as the Relative Difference Plot (RDP),
and is used for processing paired
analytical data. The second diagram,
which is widely used in the mining
industry for diagnosing sample biases
and identifying their possible causes
(Leaver et al., 1997; Sketchley, 1998),
is a simple xy plot where assayed val-
ues of the certified standards are plot-
ted on a batch/time basis.

Relative Difference Plot (RDP)


The relative difference plot (Pitard,
1998) represents a graphic tool for
diagnosis of the factors controlling
precision error (Fig. 3).
The relative difference is estimated
for each pair of data, expressed as a
percentage:
Fig. 3. Relative Difference Plot (RDP) showing Cu(%) grades of duplicated drill core samples
assayed in different labs. The samples are from a Cu-sulfide deposit in Russia. These data are not
(ai - bi ) included in Table 2 because the duplicate samples have been processed and assayed in an external
RD(%) = 100 . (25) laboratory. Open symbols (diamonds) connected by fine tie-lines are RD(%) values calculated from
(ai + bi ) 2 matching pairs of data (i.e., original sample and duplicate). Average RD(%) value (thick dashed
line) and 2SD values (fine dashed lines) are shown for reference. The solid line is a smoothed line
Calculated RD(%) values are ar- of the RD(%) values calculated using a moving windows approach. The calibration curve sets the
relationship between RD(%) values on the primary y-axis, the sequential number of the data pairs
ranged in increasing order of the aver- plotted along the x-axis, and the average grades of the data pairs plotted on the secondary y-axis.
age grades of the data pairs, and then For example, point A (sequential number of the data pairs is 100) is characterized by an RD(%)
RD(%) values (y-axis) are plotted value equal to 16%, and the average grade of the data pair is 1.1% Cu.
Quality Control of Assay Data: A Review of Procedures for Measuring and Monitoring Precision and Accuracy M. Abzalov 139

size, core recovery, and depth of sample collection. In such errors in the database.
cases, the RD(%) values are arranged according to the fac- Data bias is easily recognizable by a consistent shift of
tor under investigation, which is plotted on the secondary standard sample assays from certified values (Fig. 4C).
y-axis (instead of the average grades of the data pairs), and Usually this occurs because of failed equipment calibra-
linked by a calibration curve to the sequential number of tion; it can also be caused by changed analytical proced-
the data points shown along the x-axis. ures in the lab.
A less common distribution pattern is when dispersion
Diagnosis of Accuracy Errors of the standard sample grades suddenly decreases over time
Diagnosis of sample biases and their possible causes (Fig. 4D). Such a decrease in variability can indicate pos-
can be made using a pattern recognition approach (Leaver sible data tampering (Sketchley, 1998), or at least suggests
et al., 1997; Sketchley, 1998) by plotting assayed values that standard samples have been recognized by laboratory
of certified standards against batch number or time (Fig. personnel and treated more carefully than other samples. In
4). This method is based on the fact that specific types of such cases, assayed values of the standard samples cannot
analytical problems have recognizable patterns on certain be used as reliable confirmation that the assayed samples
types of diagram. The different distribution patterns of the lack bias.
analytical results are indicative of the error sources and Accurate analyses are also characterized by a lack of
types. systematic data trends on grade versus order of process-
This technique is most effective when applied to certi- ing diagrams. Trends can be recognized by a systematic
fied standards (Leaver et al., 1997; Sketchley, 1998). Good increase or decrease of the assayed values of the standards
quality analyses will be characterized by a random distri- over time (Fig. 4E). Another commonly used criterion for
bution of data points around the certified mean value (Fig. identification of possible trends is when two successive
4A), with 95% of the data points lying within two stan- points lie outside two standard deviations, or when four
dard deviations of the mean. The same number of analyses successive points lie outside one standard deviation (Leav-
should fall above and below the mean. er et al., 1997). Such systematic drifts of assayed standard
In some cases, the analyzed grades of the standard values usually indicate instrumental drift. Alternatively,
samples randomly and significantly differ from the cer- they can also be caused by degradation of the standard
tified values (Fig. 4B). Presence of such outliers most samples. The author is familiar with cases when the char-
likely indicates data transcription errors. This feature does acteristics of standard samples have degraded in compari-
not imply data bias, but nevertheless indicates poor data son with their certified values because of inappropriate
management, and indicates the likely presence of random storage conditions; for example, storage in large jars that

Fig. 4. Schematic diagrams showing quality control pattern recogni-


tion methods. A. Accurate data, with statistically valid distribution
of the standard values; B. Presence of outliers suggesting tran-
scription errors; C. Biased assays; D. Rapid decrease in data vari-
ability indicating possible data tampering; E. Drift of the assayed
standard values.
140 Exploration and Mining Geology, Vol. 17, Nos. 34, p. 131144, 2008

have not been protected from vibrations from operating whereas field duplicates representing the first subsampling
equipment, leading to segregation of heavy minerals to the stage were left unattended.
bottom of the jar. Hygroscopic samples or standards can Duplicate samples should be representative for the
also be affected by storage in unsealed containers. These given deposit, covering the entire range of grade values
problems can be partially overcome if standard materials and mineralization styles; at the same time, their distribu-
are split into small aliquots, preferably using a rotary split- tion should maintain, at least partially, the characteristics
ter, and stored in sealed containers. Aliquots are then used of being randomly chosen. To meet these two conditions,
as needed for only a small number of analyses. a dual approach is recommended: approximately half of
the duplicates should be collected by a project geologist
Discussion and Conclusions with instructions to ensure that they properly represent all
mineralization styles and grade ranges, and that they spa-
QAQC Practice: Planning and Implementation tially cover the entire deposit; the other half of the dupli-
This review of recently published QAQC procedures cates should be selected randomly. All duplicates should be
used in the mining industry (Taylor, 1987; Vallee et al., placed at random in the assay batch.
1992; Leaver et al., 1997; Long, 1998; Sketchley, 1998)
and numerous case studies by the author (Abzalov and Precision Estimates
Both, 1997; Abzalov, 1999, 2007; Abzalov and Humphreys, Various statistical methods for estimating precision have
2002; Abzalov and Mazzoni, 2004; Abzalov and Pickers, been assessed by applying them to the same set of dupli-
2005; Abzalov et al., 2007) shows that reliable control of cate sample analyses (Table 2). Data for this study were
sample precision is achieved by using approximately 5% to obtained from operating mines and mining projects, in-
10% of field duplicates, and 3% to 5% of pulp duplicates. cluding published data (Abzalov and Both, 1997; Abzalov,
Duplicate samples should be prepared and analyzed in the 1999, 2007; Dominy et al., 2000; Roden and Smith, 2001;
primary laboratory. Abzalov and Humphreys, 2002; Sinclair and Blackwell,
To detect bias in analytical results, it is necessary to in- 2002; Abzalov and Mazzoni, 2004; Abzalov and Pickers,
clude 3% to 5% of standard reference materials with every 2005; Abzalov et al., 2007) and unpublished reports.
sample batch. A good practice is to use more than one stan- All errors are estimated as relative standard deviations
dard, so that their values span the practical range of grades and expressed as percentages (Table 2). The Thompson-
in the actual samples. However, standard samples alone Howarth method (Equation 3; Thompson and Howarth,
cannot identify biases introduced at different stages of 1973, 1976, 1978; Howarth and Thompson, 1976) has been
sample preparation. Anonymity of the standards is another applied to the average grades of the data sets to make the
issue, because standards can be easily recognized in the estimates by this method (Equation 3) comparable to other
sample batches and be treated more thoroughly by labora- estimates.
tory personnel. Duplicate samples (i.e., coarse rejects and Comparison of the results in Table 2 shows that precision
pulp) should be analyzed in an external, reputable labora- errors produced by the Thompson-Howarth technique are
tory as part of the accuracy control. It is suggested that at consistently the lowest among all the reviewed methods.
least 5% of the total analyzed duplicates, including pulp This result is not surprising because it reflects the assump-
duplicates and coarse rejects, should be analyzed in the ex- tion in the Thompson-Howarth method that measurement
ternal laboratory. errors are normally distributed (Thompson and Howarth,
Duplicate samples should be chosen in such a manner 1973, 1976, 1978; Howarth and Thompson, 1976). Based
that all significant stages of the sampling protocol are prop- on this assumption, the method uses the median value of
erly assessed. However, the main focus of routine duplicate absolute differences of the data pairs for calculation of
analyses should be on the very first split (i.e., field dupli- precision error, and this inevitably produces biased results
cates or coarse rejects), because precision error estimated when errors are not normally distributed.
from these data is a cumulative error, which will include all This study (Table 2) concurs with Stanley and Lawie
subsequent sample preparation and analytical stages. Good (2007b) that conventional Thompson-Howarth error analy-
QAQC practice is to collect the bulk of duplicate materials sis produces significantly biased results when applied to
at the first subsampling stage, whereas subsequent stages, skewed distributions, and is particularly inappropriate for
including pulp duplicates, play auxiliary roles7. Unfortu- estimating analytical precision of samples exhibiting high
nately, the author has witnessed cases when this simple nugget effects such as gold and silver ores (Stanley, 2006).
rule has been forgotten, and project personnel were col- To overcome this problem, Stanley (2006) suggested a
lecting unnecessarily large amounts of pulp duplicates, modification to the Thompson-Howarth method that calcu-
lates the root mean squares of the absolute pair differences
instead of their medians. This alternative approach (Stanley,
7
This is the first stage at which the size of the initial sample is reduced 2006; Stanley and Lawie, 2007b) is a more robust estima-
(subsampled) for further treatment. Usually these are duplicates of non- tor in comparison to the conventional Thompson-Howarth
core drilling samples (e.g., reverse-circulation drilling chips) collected method because it does not require a normality assumption.
from splitters at the drill sites, and are therefore conventionally refer-
enced as field or rig duplicates (Long, 1998). Alternatively, for drill However, this modification is likely to be more sensitive to
core or grab samples, these can be splits of the coarsely crushed mate- outliers, and should therefore be rigorously tested in order
rial. to better understand its possible limitations when applied
Quality Control of Assay Data: A Review of Procedures for Measuring and Monitoring Precision and Accuracy M. Abzalov 141

Table 2. Comparison of Precision Error Estimated by Different Techniques


Mineralization Metal # Sam-
Type/Deposit (Grade) Grade Range Avg. ples Corr. PTH%*,1 CV(%)2 AMPD3 SR%4 PFP%5 PDFB%6 PRMA%7 Sample Type Data Source
Orogenic gold, Au (g/t) 0.11206 89.600 36 0.92 74.5 105.4 78.7 74.5 75.5 105.2 Repeat channel Dominy et
very coarse and samples al., 2000
nuggety, USA
Archean gold, Au (g/t) 0.0350 4.300 201 0.71 34.4 47.2 66.7 82.9 47.2 46.6 110.1 1st (coarse) split Current study
medium to coarse Au (g/t) 0.01355 1.300 4209 0.98 15.3 24.5 34.6 94.1 24.5 24.5 132.2 Pulp duplicates
grained, Canada
Copper-silver Ag (g/t) 2393 72.310 42 0.98 15.3 21.7 15.0 15.3 15.5 21.1 Channel Sinclair and
veins, Canada samples vs. blast Blackwell, 2002
hole cones
Cu (%) 0.010.97 0.160 42 0.99 13.5 19.2 11.0 13.5 13.5 12.0
Native Cu- Cu (%) 1.171.51 1.370 10 0.55 9.6 13.6 9.8 9.6 10.1 14.5 Pulp duplicates Roden and
mineralization Smith, 2001
Cu-Au porphyry, Cu (%) 0.00014.83 0.100 4707 0.99 5.7 14.1 20.0 21.7 14.1 14.1 30.3 Field duplicates Current study
Australia Au (g/t) 0.00057.01 0.050 4784 0.95 18.6 41.7 59.0 113.9 41.7 41.7 157.7
Cu (%) 0.00025.3 0.400 6346 0.98 1.9 10.5 14.8 17.9 10.5 10.5 25.3 Pulp duplicates
Au (g/t) 0.02276 0.800 14 346 1.00 4.4 17.1 24.2 18.8 17.1 17.0 26.6
Cu-Mo por- Cu (%) 0.00314.4 1.450 398 1.00 1.4 7.1 10.0 2.8 7.1 7.1 3.9 1st (coarse) split Current study
phyry, USA Mo (%) 0.0051.14 0.030 398 1.00 4.0 14.2 20.0 16.0 14.2 14.2 14.2
Cu (%) 0.0159.6 1.430 346 1.00 0.9 2.5 3.6 1.3 2.5 2.5 1.8 Pulp duplicates
Mo (%) 0.0050.315 0.030 346 1.00 2.7 11.5 16.3 5.4 11.5 11.5 7.3
Iron Ore, Deposit Fe (%) 50.6367.37 62.200 228 0.94 0.6 1.4 2.0 1.3 1.4 1.4 1.9 Field duplicates Current study
1, Pilbara, Australia Al O (%) 0.197.38 2.060 228 0.92 5.8 12.4 17.5 19.1 12.4 12.3 26.3
2 3

SiO2 (%) 0.726.0 3.450 228 0.95 7.1 13.7 19.3 21.4 13.7 13.7 30.1
LOI (%) 0.8310.95 4.890 228 0.98 2.2 4.9 6.9 5.1 4.9 4.9 7.2
Iron Ore, Deposit Fe (%) 1.8467.3 51.270 8088 1.00 0.6 2.2 3.0 1.3 2.2 2.1 1.8 Field duplicates Current study
2, Pilbara, Australia Al O (%) 0.1150.66 5.700 8088 1.00 2.3 6.9 9.8 5.6 6.9 6.9 7.7
2 3

SiO2 (%) 0.6895.96 12.560 8088 1.00 2.3 7.0 9.9 4.9 7.0 6.9 6.8
LOI (%) 0.3426.03 7.380 8088 1.00 1.4 2.5 3.5 2.3 2.5 2.5 3.2
Iron oxide Cu-Au Au (g/t) 0.000541.3 0.740 1522 0.93 12.5 25.3 35.7 118.1 25.3 25.2 163.6 1st (coarse) split Current study
mineralization,
Australia
Cu-Au-Fe skarn Cu (%) 0.00312.3 0.910 806 1.00 2.6 8.7 12.4 4.5 8.7 8.5 6.3 Field duplicates Current study
Au (g/t) 0.0535.24 1.390 616 0.99 10.9 20.7 29.2 18.7 20.7 20.6 25.7
Ni-Cu-PGE- Ni (%) 0.00016.82 0.345 587 0.97 12.1 22.1 31.3 43.6 22.1 22.2 59.9 Field duplicates Current study
sulfides, Australia Cu (%) 0.00014.86 0.220 586 0.95 14.6 21.8 30.8 51.7 21.8 21.8 70.3
Co (%) 0.00010.16 0.013 586 0.96 9.4 14.5 20.5 37.2 14.5 14.5 52.2
Pd (g/t) 0.0052.46 0.080 323 0.96 12.3 29.4 41.5 41.1 29.4 29.1 57.8
Ni (%) 0.00015.4 0.170 961 1.00 1.2 11.0 15.6 4.2 11.0 11.0 5.3 Pulp duplicates
Cu (%) 0.0001-15.8 0.180 961 1.00 1.4 4.2 5.9 3.4 4.2 4.2 4.5
Co (%) 0.0001-0.175 0.079 961 1.00 1.7 7.5 10.6 4.4 7.5 7.5 5.9
Pd (g/t) 0.001-1.26 0.109 836 0.97 3.9 17.7 25.1 18.8 17.7 17.8 26.3
Detrital ilmenite Total Heavy 0.7-26.2 7.480 539 0.89 5.4 17.7 25.0 17.5 17.7 17.7 24.7 Field duplicates Current study
sands, Deposit Minerals (%)
1, Africa
Detrital ilmenite
Total Heavy
sands, Deposit 2.4-19.3 8.960 27 0.96 8.1 11.5 10.1 8.1 8.1 13.9 Field duplicates Current study
Minerals (%)
2, Africa
Notes
*
PTH% and SR% errors have been estimated at average (arithmetic mean) grade of the paired data.
1
Equation 3.
2
Equation 6.
3
Equation 7.
4
Equation 13.
5
Equation 10.
6
Equation 12.
7
Equation 20.
CV% = Coefficient of variation (%). This estimator is recommended in the current paper for use as the universal measure of relative precision error in
mine geology applications.
LOI = loss on ignition.
142 Exploration and Mining Geology, Vol. 17, Nos. 34, p. 131144, 2008

to various geological materials. and the second that uses a geostatistical approach based on
A second group of estimators include AMPD and other the application of a pair-wise relative variogram (PFP(%);
similar statistical methods (Table 1), which are based on Equation 10) for measuring precision error, are identical
calculation of the absolute differences between the original because both are based on calculating the average relative
and duplicate samples normalized by the mean grades of variance of the paired data. Not surprisingly, both estimates
the data pairs (Bumstead, 1984; Shaw, 1997; Roden and produce the same results (Table 2).
Smith, 2001). Relative errors obtained from this approach Other geostatistical tools, such as the lognormal vario-
are proportional to coefficients of variation (Stanley, 2006; gram (Garrett, 1969) or the relative variogram (Equation
Stanley and Lawie, 2007a), and represent nothing more 13; Eremeev et al., 1982), are not suitable for estimation
than the product of the CV and a constant. Hence, the vari- of average precision error because of their sensitivity to
ous statistics included in this group (e.g., AMPD, HARD) outliers. In particular, when a relative variogram (Equation
offer no more information than the CV itself (Table 2). 13) is applied to strongly skewed data without subdivid-
Stanley and Lawie (2007a) have, therefore, suggested that ing them onto grade classes, the results significantly differ
the CV be used as a universal measure of the relative pre- from their calculated CVAVR(%) values (Table 2).
cision error in mine geology applications, instead of the The approach suggested by Francois-Bongarcon (1998;
more confusing AMPD, HARD, and similar measures. Equations 1112) calculates the ratio of the difference be-
(a - b )
It is important to remember that the average coefficient tween the paired data to their sum ( (a + b ) ) and then estimates
i

i
i

of variation (CVAVR(%)) is calculated as a square root of variance of this complex variable. Results of this method
the average relative variance (Equation 6). The same ap- are similar to the average coefficient of variation (Table
proach, which is consistent with the additivity principle, 2). However, Equation 11 seems to be unnecessarily com-
must be used if precision error is estimated using AMPD plicated in comparison to the conventional CVAVR(%) ap-
and other measures proportional to the coefficient of varia- proach (Equation 26).
tion (Table 1)8. In other words, the correct procedure for The RMA model (Fig. 2) is another QAQC technique
obtaining an average AMPD value should also be based on that is becoming increasingly popular among geoscientists
calculating average relative variance and then taking the (Sinclair and Bentzen, 1998; Davis, 2002; Sinclair and
square root of it. However, this principle was not followed Blackwell, 2002). This method consists of the calculation
by Bumstead (1984) and Roden and Smith (2001), who of the RMA line (Equations 1416) and related parameters
used an average precision error calculated as the arithmetic (Equations 1720), and is particularly useful for identify-
mean of the AMPD values of the duplicated samples. This ing bias between paired data (Fig. 2). This method can be
is another reason to avoid measurements like AMPD and used to calculate the precision error estimate from disper-
HARD, whose statistical meaning is less evident than the sion of the data points about the RMA function (Equation
CV, and which cause confusion and incorrect treatment of 20). However, it is shown here that the errors estimated
the data. from this model (PRMA(%)) can significantly differ from
An alternative geostatistical approach is based on meas- CVAVR(%) values (Table 2). These differences are likely due
uring data precision by applying formulae of the vario- to sensitivities of the RMA model to the presence of out-
grams to duplicated samples (ai and bi; e.g., Garrett, 1969). liers, which makes this technique inappropriate for strongly
The two most common methods used for QAQC purpose skewed distributions.
are the pair-wise relative variogram (Equation 9) and the In summary, the author concurs with the proposal of
relative variogram (Equation 13). However, it is easy to Stanley and Lawie (2007a) that CVAVR(%) be used as the
see that when a pair-wise relative variogram is applied to universal measure of relative precision error. Based on
duplicated samples, it simply measures the relative vari- numerous case studies, the appropriate levels of sample
ance of the paired data. The square root of this value is precisions are proposed for different types of deposits in
equal to average coefficient of variation: Table 3, and it is suggested that these levels be used as ap-
proximate guidelines for assessing analytical quality. It is
2 N (a - b )2 important to remember that the values in Table 3, although
i
PFP (%) = 100
N
(a + b ) = CV
i
2 AVR (%) . (26) based on case studies of the mining projects, might not be
i =1 i i appropriate always because mineral deposits can signifi-
cantly differ by the grade ranges, statistical distribution of
the studied values, mineralogy, textures, and grain sizes.
Hence, the two approaches, one that uses the coefficient
of variation (CVAVR(%); Equation 6) as a universal measure Accuracy Tests
of the relative precision error (Stanley and Lawie, 2007a), The accuracy of analytical data is quantified by statis-
tical comparison of the arithmetic mean of the replicate an-
alyses of a certified standard with its certified mean com-
8
Standard deviation is not an additive parameter, and therefore its aver- position (Equations 2123). The purpose of these tests is to
age value cannot be calculated as the arithmetic mean of the standard demonstrate that analytical accuracy is comparable to the
deviations of the individual data pairs. Conversely, variance is an addi-
tive parameter, and therefore the mean variance can be calculated. The certified value of a standard, and is acceptable with 95%
average standard deviation is then derived by taking the square root of confidence. These tests should be supported by estimation
the mean relative variance (Equation 6). of the standard deviation of the replicate analyses (Equa-
Quality Control of Assay Data: A Review of Procedures for Measuring and Monitoring Precision and Accuracy M. Abzalov 143

Table 3. Best and Acceptable Levels of Precision Errors (CVAVR(%)) suggests the use of RMA diagrams as
Mineralization Best Acceptable a routine supplement to statistical cal-
Type / Deposit Metal Practice Practice Sample Type culation of the precision error. When
Gold, very coarse grained Au (g/t) 20 (?) 40 Field duplicates the estimated CVAVR(%) exceeds the
and nuggety allowed tolerance level (Table 3), an
Gold, coarse to medium Au (g/t) 20 30 Field duplicates RDP diagram can be used for a more
grained Au (g/t) 10 20 Pulp duplicates
detailed analysis and diagnosis of er-
rors.
Cu-Mo-Au porphyry Cu (%) 5 10 Field duplicates
Mo (%) 10 15 Acknowledgments
Au (g/t) 10 15
Cu (%) 3 10 Pulp duplicates The author expresses his sincere
Mo (%) 5 10 gratitude to all his colleagues in the
Au (g/t) 5 10 mining industry, in particular from
WMC Resources, Rio Tinto, BHPB,
Iron ore Fe (%) 1 3 Field duplicates
and Norilsk Nickel, for many stimulat-
Al2O3 (%) 10 15 ing discussions and useful comments
SIO2 (%) 5 10 on an earlier version of this paper.
Loss on Ignition (%) 3 5 C. Stanley is thanked for assistance in
Cu-Au-Fe skarn and Iron Cu (%) 7.5 15 Field duplicates obtaining the most recent papers on
oxide associated Cu-Au Au (g/t) 15 25 sampling and measurements errors.
Review and useful comments by M.
Cu (%) 5 10 Pulp duplicates
Rayner, A. Sinclair, an anonymous
Au (g/t) 7.5 15
journal reviewer, and, in particular,
Ni-Cu-PGE-sulfides Ni (%) 10 15 Field duplicates very helpful editorial comments by J.
Cu (%) 10 15 Richards are gratefully acknowledged.
PGE 15 30
Ni (%) 5 10 Pulp duplicates References
Cu (%) 5 10
Abzalov, M.Z., 1999, Gold deposits of
PGE (g/t) 10 20
the Russian North East (the North-
Detrital ilmenite sands Total Heavy Minerals (%) 5 10 Field duplicates ern Circum Pacific): Metallogenic
overview, in AusIMM, eds., Pro-
ceedings of the PACRIM 99 sym-
tion 24), because poor precision of the standard sample posium: Australasian Institute of Mining and Metal-
assays can inhibit reliable detection of analytical biases. lurgy, p. 701714.
Standard samples are used exclusively for quantifica- Abzalov, M.Z., 2007, Zarmitan granitoid-hosted gold de-
tion of performance of the certified standards, and can- posit, Tian Shan belt, Uzbekistan: Economic Geology,
not identify sample preparation biases. In order to obtain v. 102, p. 519532.
a more complete and conclusive assessment of accuracy Abzalov, M.Z., and Both, R.A., 1997, The Pechenga Ni-Cu
of the assayed data, it is necessary to process and analyze deposits, Russia: Data on PGE and Au distribution and
approximately 5% of the duplicate samples (coarse rejects sulphur isotope compositions: Mineralogy and Petrol-
and pulp) in an external reputable laboratory that has the ogy, v. 61, p. 119143.
authority to act as an independent auditor. The presence of Abzalov, M.Z., and Humphreys, M., 2002, Resource esti-
bias in the duplicate samples can be detected using RMA mation of structurally complex and discontinuous min-
(Fig. 2) and RDP (Fig. 3) methods. eralisation using non-linear geostatistics: Case study of
a mesothermal gold deposit in northern Canada: Explor-
Diagnostic Tools ation and Mining Geology, v. 11, p. 1929.
Diagnostic diagrams can supplement statistical tests, Abzalov, M.Z., and Mazzoni, P., 2004, The use of condi-
which are ineffective in detection of certain types of sam- tional simulation to assess process risk associated with
pling errors and identifying their possible causes. In par- grade variability at the Corridor Sands detrital ilmenite
ticular, sample biases caused by instrumental drift, tran- deposit, in Dimitrakopoulus, R., and Ramazan, S., eds.,
scription errors, and data tampering might not be detected Ore body modelling and strategic mine planning: Un-
by statistical tests, but are easily recognized from data certainty and risk management: Australasian Institute of
distribution patterns of standard assays plotted versus the Mining and Metallurgy, p. 93101.
order of processing (Fig. 4). Abzalov, M.Z., and Pickers, N., 2005, Integrating different
Reduced Major Axis (RMA: Fig. 2) and Relative Dif- generations of assays using multivariate geostatistics:
ference Plots (RDP; Fig. 3) can be used for diagnosis of A case study: Transactions of Institute of Mining and
factors causing poor reproducibility of results. The author Metallurgy, Section B: Applied Earth Science, v. 114,
144 Exploration and Mining Geology, Vol. 17, Nos. 34, p. 131144, 2008

p. B23B32. mineral inventory estimation: Exploration and Mining


Abzalov, M.Z., Menzel, B., Wlasenko, M., and Phillips, Geology, v. 7, p. 117127.
J., 2007, Grade control at Yandi iron ore mine, Pilbara Pitard, F., 1998, A strategy to minimise ore grade reconcili-
region, Western AustraliaComparative study of the ation problems between the mine and the mill: Mine to
blastholes and Reverse Circulation holes sampling, in Mill 1998 Conference, Brisbane, Australia, October 11
AusIMM, eds., Proceedings of the iron ore conference, to 14, Australasian Institute of Mining and Metallurgy,
2022 August, 2007, Perth, Western Australia: Austral- p. 7782.
asian Institute of Mining and Metallurgy, p. 3743. Roden, S., and Smith, T., 2001, Sampling and analysis
Bumstead, E.D., 1984, Some comments on the precision protocols and their role in mineral exploration and new
and accuracy of gold analysis in exploration, in Aus- resource development, in Edwards, A., ed., Mineral
IMM, eds., Proceedings AusIMM: Australasian Institute resources and ore reserve estimationThe AusIMM
of Mining and Metallurgy, No. 289, p. 7178. guide to good practise: Australasian Institute of Mining
CANMET, 1998, Assessment of laboratory performance and Metallurgy , p. 7378.
with certified reference materials: CANMET Canadian Shaw, W.J., 1997, Validation of sampling and assaying
Certified Reference Materials Project Bulletin, 5 p. quality for bankable feasibility studies, in AusIMM,
Davis, J.C., 2002, Statistics and data analysis in geology, eds., The resource database towards 2000: Australasian
3rd ed.: New York, John Wiley and Sons, 638 p. Institute of Mining and Metallurgy, AusIMM Illawara
Dominy, S.C., Annels, A.E., Johansen, G.F., and Cuffley, branch, Wollongong, p. 6979.
B.W., 2000, General considerations of sampling and Sinclair, A.J., and Bentzen, A., 1998, Evaluation of errors
assaying in a coarse gold environment: Transactions of in paired analytical data by a linear model: Exploration
Institute of Mining and Metallurgy, Section B: Applied and Mining Geology, v. 7, p. 167173.
Earth Science, v. 109, p. B145B167. Sinclair, A.J., and Blackwell, G.H., 2002, Applied min-
Eremeev, A.N., Ostroumov, G.V., Anosov, V.V., Beren- eral inventory estimation: Cambridge University Press,
shtein, L.E., Korolev, V.P., and Samonov, I.Z., 1982, 381p.
Instruction on internal, external, and arbitrary quality Sketchley, D.A., 1998, Gold deposits: Establishing sam-
control of the exploration samples assayed in the labora- pling protocols and monitoring quality control: Explora-
tories of the Ministry of Geology of the USSR: VIMS, tion and Mining Geology, v. 7, p. 129138.
Moscow, 106 p. (in Russian). Stanley, C.R., 2006, On the special application of Thomp-
Francois-Bongarcon, D., 1998, Error variance information son-Howarth error analysis to geochemical variables
from paired data: Applications to sampling theory: Ex- exhibiting a nugget effect: Geochemistry: Exploration,
ploration and Mining Geology, v. 7, p. 161165. Environment, Analysis, v. 6, p. 357368.
Garrett, R.G., 1969, The determination of sampling and Stanley, C.R., and Lawie, D., 2007a, Average relative error
analytical errors in exploration geochemistry: Economic in geochemical determinations: Clarification, calcula-
Geology, v. 64, p. 568574. tion, and a plea for consistency: Exploration and Mining
Goovaerts, P., 1997, Geostatistics for natural resources Geology, v. 16, p. 265274.
evaluation: New York, Oxford University Press, 483 p. Stanley, C.R., and Lawie, D., 2007b, Thompson-Howarth
Howarth, R., and Thompson, M., 1976, Duplicate analysis error analysis: Unbiased alternatives to the large-sample
in geochemical practice, Part 2: Examination of pro- method for assessing non-normally distributed measure-
posed method and examples of its use: The Analyst, v. ment error in geochemical samples: Geochemistry: Ex-
101, p. 699709. ploration, Environment, Analysis, v. 7, p. 110.
ISO Guide 33, 1989, Uses of certified reference materials: Taylor, J.K., 1987, Quality assurance of chemical measure-
Standards Council of Canada, 12 p. ments: Michigan, Lewis Publishers, 135 p.
JORC Code, 2004, Australasian code for reporting of ex- Thompson, M., and Howarth, R., 1973, The rapid estima-
ploration results, mineral resources, and ore reserves: tion and control of precision by duplicate determina-
AusIMM, 20 p. tions: The Analyst, v. 98, p. 153160.
Kane, J.S., 1992, Reference samples for use in analytical Thompson, M., and Howarth, R., 1976: Duplicate analysis
geochemistry: Their availability, preparation, and ap- in geochemical practice, Part 1: Theoretical approach
propriate use: Journal of Geochemical Exploration, v. and estimation of analytical reproducibility: The Ana-
44, p. 3763. lyst, v. 101, p. 690698.
Leaver, M.E., Sketchley, D.A., and Bowman, W.S., 1997, Thompson, M., and Howarth, R., 1978, A new approach to
The benefits of the use of CCRMPs custom reference the estimation of analytical precision: Journal of Geo-
materials: Canadian certified reference materials pro- chemical Exploration, v. 9, p. 2330.
ject: Society of Mineral Analysts Conference, MSL No. Valle, M., David, M., Dagbert, M., and Desrochers, C.,
637, 16 p. 1992, Guide to the evaluation of gold deposits: Geo-
Long, S., 1998, Practical quality control procedures in logical Society of CIM, Special Volume 45, 299 p.

Vous aimerez peut-être aussi