See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/249545172

Thompson-Howarth error analysis: Unbiased alternatives to the large-

sample method for assessing non-normally distributed measurement error in
geochemical samples

Article  in  Geochemistry Exploration Environment Analysis · May 2008

DOI: 10.1144/1467-7873/07-145

CITATIONS READS

8 2,952

2 authors:

Clifford Stanley Dave Lawie

Acadia University REFLEX Geochemistry
69 PUBLICATIONS   1,174 CITATIONS    10 PUBLICATIONS   217 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Drill Core Data Acquisition Methods View project

Geochemical Data Transformation Methods View project

All content following this page was uploaded by Clifford Stanley on 12 June 2015.

The user has requested enhancement of the downloaded file.

 Thompson–Howarth error analysis: unbiased alternatives to the
large-sample method for assessing non-normally distributed
measurement error in geochemical samples
Clifford R. Stanley1 & David Lawie2
1
Department of Geology, Acadia University, Wolfville, Nova Scotia, B4P 2R6, Canada
(e-mail: cliff.stanley@acadiau.ca)
2
ioGlobal, Level 3, IBM Building, 1060 Hay Street, West Perth, WA 6005, Australia
(e-mail: dave.lawie@ioglobal.net)

ABSTRACT: The Thompson–Howarth error analysis procedures have become

common in geochemical applications for assessing the magnitude of measurement
error at any stage of determination (initial sampling, sample preparation, geochemical
analysis). However, the large-sample method, as defined by Thompson and Howarth,
which relies on an assumption that the measurement errors are normally distributed,
produces significantly biased results when the errors are not normally distributed.
Four examples of quality control data-sets from a variety of mineral deposit types
illustrate that normally distributed errors are probably the exception rather than the
rule in ore deposits, and non-normally distributed geochemical data-sets may exist in
other geological materials. As a result, using Thompson and Howarth’s large-sample
error analysis approach, geoscientists may obtain a significantly inaccurate estimate of
the quality of their geochemical concentration data.
Two new methods, which are modifications to the Thompson–Howarth large-
sample technique, eliminate this bias because they ensure that the results are
independent of a normally distributed error assumption. Regression of group root
mean square standard deviations produces accurate error estimates, at least provided
that the concentrations are distributed relatively evenly across their range. Similarly,
regression of duplicate variances against duplicate means using a quadratic model,
and then taking the square root of this model, also results in an unbiased estimate of
measurement error. Furthermore, if this quadratic model is a perfect square, then the
square root of the quadratic model will be linear on a mean versus standard deviation
scatterplot. This approach is further independent of the distribution of the error in
the data. Using these modified approaches, geoscientists can now obtain unbiased
Thompson–Howarth estimates of measurement error that is not normally distributed
in quality control/quality assessment programmes.
KEYWORDS: measurement error, precision, relative error, Thompson–Howarth, duplicate, replicate,
bias, standard deviation, variance

INTRODUCTION THOMPSON–HOWARTH ERROR ANALYSIS

Assessments of measurement error in geochemical applications, The second approach commonly used to assess measurement
including errors introduced during sampling, sample prep- error involves development of a linear equation that describes
aration or geochemical analysis, have historically been under- error as a function of concentration. This approach, referred to
taken using two approaches. Both involve estimation of error, as Thompson–Howarth error analysis (Thompson & Howarth
or relative error, using p replicate determinations from a set of 1973, 1976, 1978; Howarth & Thompson 1976; Thompson
n samples (note, when p=2 these are commonly referred to as 1973, 1982; Fletcher 1981; Stanley & Sinclair 1986; Stanley
duplicate or paired determinations). 2003a, 2003b; Garrett & Grunsky 2003; Stanley 2006b), also
The first approach involves determining a measure of applies across the range of concentrations under consideration.
average relative error that applies to the range of element However, the Thompson–Howarth error analysis technique is
concentrations under consideration. A review of this approach more complicated than the average relative error approach.
was undertaken by Stanley & Lawie (2007b and references Although both approaches can be represented by straight lines
therein; Francois-Bongarcon 1998), and so is not presented on a scatterplot of replicate mean versus replicate standard
here. deviation, the average relative error line is defined by a single

Geochemistry: Exploration, Environment, Analysis, Vol. 7 2007, pp. 1–10 1467-7873/07/$15.00  2007 AAG/ Geological Society of London
2 C. R. Stanley & D. Lawie

Fig. 1. Thompson–Howarth scatterplot for larger data-sets contain-

ing more than c. 50 duplicate measurements. Grey circles are the
means and absolute differences of 66 duplicate measurements, Fig. 2. Thompson–Howarth scatterplot for smaller data-sets con-
whereas black circles are the average means and median absolute taining less than c. 50 duplicate measurements. Grey circles are the
differences of six groups of 11 consecutive duplicate pairs. Ordinary means and absolute differences of 15 duplicate measurements. A
least square regression is typically undertaken on these average means postulated two-standard-deviation measurement–error model (solid
and median absolute differences to determine a linear median line) is tested against the data by plotting the 90th and 99th percentile
measurement–error model for the median. Multiplication of the confidence bounds for this model (dashed; 2.32617 and 3.64277
slope and intercept of this median measurement error model times the slope and intercept of the model line, respectively; equal to
by 1.04836 yields the corresponding one-standard-deviation the appropriate half normal distribution cumulative percentiles times
measurement–error model slope and intercept. the square root of two). Using the binomial probability tables
presented in Thompson & Howarth (1978), one can determine how
likely the observed number of absolute differences that exceed these
percentile confidence bounds would occur at random (2 and 0 out of
15 duplicates exceed these bounds, respectively).
parameter (the non-negative slope of a line that passes through
the origin) equal to the average relative error. In contrast, the
Thompson–Howarth error analysis line is defined by two Large-sample Thompson–Howarth error analysis
parameters (slope and intercept), both of which are constrained method
to be non-negative. Furthermore, these two measures of error If a large number of duplicate measurements are available (e.g.
are derived using different mathematical approaches, each n > 50), a linear measurement–error function may be obtained
of which has embedded within them different underlying via regression (Thompson & Howarth 1976, 1978; Fletcher
assumptions and philosophical foundations. 1981). This has historically been achieved by first assembling
The Thompson–Howarth error analysis approach the duplicates into consecutive groups of 11, based on their
(Thompson & Howarth 1973, 1976, 1978; Howarth & means (although any number of duplicates per group other than
Thompson 1976; Thompson 1973, 1982; Fletcher 1981; Stanley 11 can be used; Thompson 1988). Larger numbers per group
& Sinclair 1986; Stanley 2003a, 2003b; Garrett & Grunsky 2003; (>11) result in fewer, more accurately estimated means and
Stanley 2006b) is actually composed of two methods that are standard deviations (absolute differences) to regress with;
widely used in applied geochemistry to establish a linear smaller numbers per group result in more, less accurately
measurement–error function. These methods utilize duplicate estimated statistics. Obviously a trade-off exists, and the
measurements (k=2) made on a subset of samples from a number of samples per group should probably be a function of
geochemical survey (used here sensu lato to include any the total number of duplicates in the data-set.
sampling/analysis programme involving the measurement of After grouping, the average means and median (or root mean
geochemical concentrations, including drilling assays) to estab- square, RMS) absolute differences (or standard deviations;
lish a linear relationship between an element concentration and Stanley 2003a, 2003b) of the duplicates in each group are
its associated one standard deviation measurement error (or an calculated and plotted on a Thompson–Howarth scatterplot
appropriate proxy; Stanley 2006b). Duplicates collected at the (Fig.1; Thompson & Howarth 1976, 1978; Fletcher 1981). The
various stages of sampling, preparation and analysis can thus be use of the median absolute differences (or standard deviations;
used to determine the magnitude of measurement error at these Stanley 2003a, 2003b) provides a more stable estimate of the
various stages of sample treatment. variation than the mean absolute differences (or standard
The methods first require that the means and absolute deviations) because it is less influenced by outliers (Thompson
differences (or standard deviations; Stanley 2003a, 2003b) of the & Howarth 1976). However, if the measurement errors cannot
duplicate measurements be calculated and plotted on the be assumed to be normally distributed, such as for gold
abscissa and ordinate of a scatterplot, respectively. Then, two concentrations, the Thompson–Howarth method will produce
different approaches are employed to determine the linear biased results if applied using the median absolute differences
relationship between the measurements and their error, depend- (or standard deviations; see below). As a result, RMS absolute
ing on how many duplicate measurements are available (Figs. 1 differences should be used when measurement errors cannot be
and 2). This paper considers, and suggests modifications to, assumed to be normally distributed to avoid imposing bias
only the large-sample Thompson–Howarth error analysis (Stanley 2006b).
method. A companion paper (Stanley & Lawie 2007a) addresses Ordinary least squares (OLS) regression of the group
issues associated with the small-sample Thompson–Howarth median (or RMS) absolute differences (or standard deviations;
error analysis method. Nevertheless, an overview of both Thompson 1988; Stanley 2003a, 2003b, 2006b) against the
methods is presented below. group average means yields a linear measurement–error
 Thompson–Howarth error analysis 3

relationship (
=mµ+b). If RMS absolute differences were inconsistent with the data (note that the double negative here is
employed in the regression, these regression parameters (m and appropriate because we fail to reject the hypothesis that this
b) may be converted to those that would be obtained had the model fits the data). Thus, at a first approximation, this model
RMS standard deviations been used in the regression by estimates the unknown measurement–error relationship in
multiplication of the slope and intercept by 1/√2 (=0.70711). these duplicate data.
Alternatively, if median standard deviations were used, the
regression parameters may be converted to those that would be Error measurement model
obtained had the RMS standard deviations been used in the
The Thompson–Howarth error analysis linear model defining
regression by multiplication of the slope and intercept by
the measurement–error relationship derived by either of the
1/0.67449=1.48260 (Thompson & Howarth 1976, 1978;
above techniques has three constraints. The first is implicit
Fletcher 1981). Lastly, if median absolute differences were
in Thompson and Howarth’s goal of specifically evaluating
used, both correction factors need to be employed (i.e. the slope
geochemical concentration data. The measurements must be
and intercept must be multiplied by 1/(1.41421  0.67449)=
positive because they are geochemical concentrations. In
1.04836; Stanley 2003a, 2003b, 2006b) to obtain what would
addition, Thompson and Howarth characterize sampling error
have been obtained had the RMS standard deviations been used
in terms of a standard deviation, which also must be positive.
in the regression. Thus, the expected (RMS) measurement error
Thus, a Thompson–Howarth error analysis is undertaken using
for samples at any concentration within the range of duplicate
a scatterplot of the positive quadrant.
concentrations may be estimated using the functional relation-
The second constraint is an assumption (Thompson &
ship determined by this simple regression (Thompson &
Howarth 1973) that measurement error increases linearly with
Howarth 1976, 1978; Fletcher 1981).
the measurement magnitude. Thus, the slope of the line
describing this relationship, which is the ‘relative error term’ in
Small-sample Thompson–Howarth error analysis method the model (m, sometimes called ‘proportional error’ and com-
monly expressed as a percentage), must be non-negative. If the
If only a small number of duplicate measurements are available
line describing measurement error has a negative slope, the
(e.g. n c 50), an alternative method is employed, in which
formula for the effective lower detection limit (see Equation 2
geoscientists postulate a linear measurement–error model by
below) would produce an upper detection limit, rather than a
arbitrarily choosing a slope and intercept that they think applies
lower one.
to their data, or which represent a maximum acceptable error
The third constraint (Thompson & Howarth 1973) is that the
for their data, and then test whether this function is consistent
intercept of this line, which is the ‘absolute error term’ in the
with the data (Thompson & Howarth 1976, 1978; Fletcher
model (b, sometimes called ‘fixed error’ or ‘error at zero
1981). This is achieved by plotting on a Thompson–Howarth
concentration’, and expressed in the same units as the measure-
scatterplot the two lines corresponding to the 90th and 99th
ments), must also be non-negative. This, again, is required
percentile bounds for the half normal distribution on the
because standard deviations are non-negative.
postulated model, a procedure achieved by multiplying the
Clearly, special procedures are necessary if an unconstrained
postulated slope and intercept by 2.32617 and 3.64277, respect-
linear regression of the absolute differences (or standard
ively, to obtain the corresponding slope and intercept for these
deviations) against the means produces an optimal slope (m) or
critical lines (Fig. 2; note that any number of percentile
intercept (b) that is negative. In order to be consistent with
confidence bound lines can be used, provided the appropriate
Thompson & Howarth’s (1973) original assumptions, if the
multiplier is known for the confidence bounds employed).
calculated intercept is negative (b < 0), the regression may be
Then, the number of duplicate absolute differences that plot
redone using the constraint that the intercept equals zero. This
above each of these lines is determined. Using binomial
forces the regression line through the origin and guarantees a
statistics, the probability that this number of duplicate absolute
non-negative slope. It assumes that the model no longer
differences (or standard deviations) plots above each percentile
involves an absolute error term, and describes all measurement
critical value line is calculated. If these probabilities are relatively
error using only a relative error term. Alternatively, if the
high, then the postulated one standard deviation linear
calculated slope is negative (m < 0), the regression may be
measurement–error model can be considered consistent with
redone using the constraint that the slope equals zero. This is
the data. Thompson & Howarth (1976, 1978; Fletcher 1981)
equivalent to equating the intercept to the mean of the group
assembled binomial statistics tables for the 90th and 99th
median (mean) absolute differences (standard deviations), and
percentiles for up to 50 observations to allow these calculations,
produces a ‘regression’ line that is horizontal, guaranteeing a
although Stanley (2003a, 2003b) provided a MATLAB program
non-negative intercept. It assumes that the model involves no
capable of outputting a corresponding binomial probability
relative error term, and describes all measurement error using
table for a wide range of percentiles and n, along with the
only an absolute error term (Stanley 2003a, 2003b).
associated slope and intercept multipliers.
Note that because the purpose of Thompson and Howarth’s
Figure 2 presents an example of this small-sample procedure
methods is to determine the level of measurement error in
on a Thompson–Howarth scatterplot. Based on the 15 dupli-
geochemical analysis, both the chemical concentrations and
cate pairs under consideration, a measurement–error relation-
their errors (standard deviations) must be positive. Thus,
ship is postulated (in this case,
=mµ+b; solid line). No
regressions resulting in a negative slope and intercept are not
duplicates (out of 15) plot above the 99th percentile line, and
realistic. Nevertheless, many other forms of natural science data
only two duplicates plot above the 90th percentile line. The
can be evaluated using the Thompson and Howarth procedures
probabilities of these happening at random are 45.0957% and
provided the data are positive and a linear (with positive slope
86.0058%, respectively (Thompson & Howarth 1976, 1978;
and intercept) measurement–error model is applicable.
Fletcher 1981). These probabilities thus allow determination of
whether it is likely that these duplicate data were derived from
the postulated measurement–error relationship defined by the Effective detection limit
model slope and intercept. In this case, the probabilities are Once a measurement–error relationship has been determined,
relatively high, and this measurement–error model is not both the slope and intercept of the regression line can be used
4 C. R. Stanley & D. Lawie

ALTERNATIVE LARGE-SAMPLE
THOMPSON–HOWARTH ERROR ANALYSIS
A fundamental mathematical property that must be exhibited by
data before many forms of numerical manipulation can be
undertaken on those data is ‘additivity’. This property of a
variable essentially dictates that linear operations (i.e. addition,
subtraction, multiplication or division) on the variable have
physical meaning (Stanley 2006a).
Thompson and Howarth’s large-sample error analysis
approach involves the regression of duplicate standard devia-
tions (or their proxies) on duplicate means. However, variances,
not standard deviations, are additive (Shaw 1961; Zitter & God
1971; Pitard 1993). Thus, Thompson and Howarth’s large-
sample error analysis approach appears to be inconsistent with
Fig. 3. Thompson–Howarth scatterplot graphically illustrating the the concept of additivity, involving numerical operations on
relationship between an error model (
=0.25µ+1; 25% relative error variables that are not additive (i.e. the standard deviations). If
plus 1 ppm absolute error), and two precision lines (
=µ/2 and

=µ/3), and the effective detection limits defined by the correspond- inconsistent with additivity, Thompson and Howarth’s large-
ing intersections (equal to 4=2b/(1
2m)=2/(1
Y) and 12=3b/ sample error analysis approach would be invalid, providing
(1
3m)=3/(1
¾)). biased estimates of measurement error from duplicate data.
However, by considering exactly what Thompson and
Howarth’s large-sample error analysis approach involves, one
can demonstrate that the approach is valid and consistent with
to determine the ‘effective detection limit’ of the duplicate data. the concept of additivity. Specifically, Thompson and
The ‘effective detection limit’ (de) is defined as the magnitude of Howarth’s large-sample error analysis involves a linear OLS
a measurement at which the precision (twice the total error regression of median group duplicate standard deviations (or
divided by the measurement) equals 100% (p=100%=2
/µ; or their proxies) on average group duplicate means to describe

=µ/2; Thompson & Howarth 1973, 1976, 1978; Howarth measurement error as a function of concentration. The regres-
& Thompson 1976). The Analytical Methods Committee sion involves median group standard deviations, and these are
(1987) suggests a three-standard-deviation detection limit, merely the square roots of the median group variances. How-
where
=µ/3 (Fig. 3). A discussion of the merits of this ever, these median group variances are functionally related to
definition of the effective detection limit is presented in the (proportional to) the mean group variances by a multiplicative
Appendix. factor derived from the distributional form of a normal
The expected (average) measurement error (
) predicted by distribution (requiring the original assumption by Thompson
the error model originally defined by Thompson & Howarth and Howarth that the errors be normally distributed;
(1973) can be determined at any concentration (µ) by the Thompson & Howarth 1973, 1976, 1978; Howarth &
following formula: Thompson 1976). This multiplicative factor, 2.1981=1.48262, is
the square of the factor used by Thompson and Howarth to

= mµ + b convert the median standard deviations into mean standard
deviations (Thompson & Howarth 1973, 1976, 1978; Howarth
The ‘effective detection limit’ (de) for these concentrations, & Thompson 1976). Thus, Thompson and Howarth’s regres-
based on their observed level of precision, is thus: sion uses a median standard deviation derived from a median
variance that is proportional to the mean variance. Conse-
de = 2b quently, although Thompson and Howarth’s use of the median
1
2m standard deviation (or its proxy) was originally intended to
produce numerically stable results, and was ironically not
where m is expressed in slope units (i.e. not as a percentage). employed in an effort to be non-parametric (because normality
For example, if the relative error term for an error model is assumed), additivity is also coincidentally achieved by this
derived from a set of duplicates is 25%, and the absolute error procedure because the median standard deviations are derived
term is 1 ppm, then the effective detection limit is 4 ppm from the additive mean variances. As a result, OLS regression
[=2b/(1
2m)=(2  1)/(1
2  X)]. This effective detection of the median group duplicate standard deviations on the
limit is equivalent to the abscissa axis coordinate of the average group duplicate means produces an unbiased estimate
intersection point between the model line describing measure- of the measurement–error relationship, provided the errors are
ment error as a function of concentration, and a line through normally distributed.
the origin with a slope of Y (
=Y µ; the line where precision Stanley (2006b) presented an alternative Thompson–
equals 100%; Fig. 3) on a Thompson–Howarth scatterplot. Howarth large-sample error analysis approach that does not
Clearly, because the Thompson–Howarth error analysis involve an assumption of normally distributed errors. In this
approach requires that the measurements be always positive, a modified technique, the RMS group standard deviations:
realistic (non-negative) effective (lower) detection limit cannot
be calculated unless the slope and intercept are constrained to
be non-negative, as dictated by Thompson and Howarth’s
model (Thompson & Howarth 1973, 1976, 1978; Howarth &
Thompson 1976). Note also that an effective detection limit
¯ =
!& n

i=1

i2

n
cannot be calculated when the relative error term is greater than
50% because the resulting Thompson and Howarth instead of the median group standard deviations (or their
measurement–error model is everywhere greater than the proxies) are regressed against the average group means. This
precision line (
=µ/2) within the positive quadrant. avoids use of the multiplicative factor required to convert
 Thompson–Howarth error analysis 5

median standard deviations into mean standard deviations, a

factor that to be known requires the assumption of a distri-
bution form (in this case the normal distribution; note that a
different distribution would require a different factor because it
would have a different mean/median ratio). As a result, this
RMS approach is independent of the normality assumption.
Using the RMS group standard deviations in an OLS regression
is also consistent with the concept of additivity because the
RMS standard deviations are the square roots of the average
group variances. Consequently, this alternative method also
provides unbiased estimates of the measurement–error relation-
ship, and has the added advantage of not requiring a normality
assumption. Thus, this alternative approach can be more
generally applied to error assessment efforts in applied
geochemistry because it makes no distributional assumptions.
Both of the above Thompson–Howarth large-sample error
analysis approaches use grouped data to make their calculations.
Unless dedicated software is available, these calculations are
typically undertaken using commercially available computer
spreadsheets. Unfortunately, spreadsheet calculations made on
grouped data are cumbersome because they must be undertaken
on every th row in the spreadsheet. This is both tedious and
time-consuming, particularly when large replicate data-sets are
under consideration. As a result, it would be useful to be able
to obtain an unbiased estimate of the measurement–error
relationship directly from the individual replicate means and
standard deviations, instead of from group means and standard
deviations (or their proxies).
This objective can be undertaken, provided that additivity
issues are addressed in the attendant numerical calculations.
Obviously, regressing the individual standard deviations (or
their proxies) against the individual means will produce biased Fig. 4. Thompson–Howarth error analysis results for two syntheti-
estimates because variances (not standard deviations) are addi- cally derived data-sets of 5005 duplicates exhibiting 5% relative error
tive. To illustrate the significance of this bias, two synthetic (A) and 5% relative error plus 2 ppm absolute error (B). Results from
data-sets were generated using a normal random-number OLS regressions on the group median standard deviations (triangle,
top) and group RMS standard deviations (filled circles, bottom)
generator to produce 5005 duplicate determinations (divisible approximate the true, underlying error within these data-sets.
by 11) with uniformly distributed true concentrations ranging
from 10 to 100 ppm. Random normal scores were then added
to these true concentrations to produce two duplicate data-sets: As a result, an alternative approach has been developed
(i) one exhibiting relative errors only (
=5%  µ), and (ii) one whereby, instead of individual means and standard deviations,
exhibiting relative and absolute errors (
=5%  µ+2 ppm). individual means and variances are plotted against each other on
Thompson–Howarth error analysis results calculated from a scatterplot (Fig. 6). Then, a regression of the individual
groups of 11 replicate median standard deviations (the classic replicate variances on the individual replicate means is under-
large-sample Thompson–Howarth error analysis method) and taken. The square root of this regression function converts it
from groups of 11 RMS standard deviations (Stanley’s (2006b) into a form that relates replicate mean to replicate standard
modification of the large-sample Thompson–Howarth error deviation. If the functional form of this ultimate measurement–
analysis method) for these data-sets are presented in Figure 4. error equation is to be linear, then the regression involving the
These results are highly comparable because these data-sets individual means and variances must have a quadratic form.
have error that is normally distributed, although in both cases Furthermore, this quadratic equation must be a perfect square if
the method involving median standard deviations produced its square root is to define a linear equation on a mean versus
larger slopes and intercepts (but not significantly so) than the standard deviation scatterplot.
method involving RMS standard deviations. These results also Three different quadratic regression models can be envi-
approximate well the 5%, and 5% plus 2 ppm error magnitudes sioned and calculated for any replicate data-set. These models
introduced into these data-sets. involve different numbers of parameters (1, 2 and 3), which
Thompson–Howarth error analysis was also undertaken must be constrained to be non-negative to be consistent with
using linear OLS regression of the individual standard the fact that concentrations and standard deviations are posi-
deviations on the individual means for each of these data-sets. tive. The first quadratic model is associated with a Thompson–
Figure 5 presents the results of these error calculations. When Howarth error described by only relative error. It has the form:
calculated from the raw duplicate data statistics, these results
underestimate the slopes and intercept defining these two
2 = (mµ)2 = m2µ2
measurement–error relationships by c. 20%. This illustrates that
significant biases exist in the estimates of error when a (or
=mµ in mean versus standard deviation space). When the
‘Thompson–Howarth-like’ error analysis is employed using the square root of this one-parameter (m) quadratic regression
individual duplicate statistics. This bias is caused by the fact that model is calculated, the result is a straight line (
=mµ) with
the regression involves data (the standard deviations) that are non-negative slope (m) through the origin on a scatterplot of
not additive. replicate mean versus replicate standard deviation.
6 C. R. Stanley & D. Lawie

Fig. 5. Thompson–Howarth error analysis results for the two Fig. 6. Thompson–Howarth error analysis results for the two
synthetically derived data-sets presented in Figure 4 (top and bottom, synthetically derived data-sets presented in Figure 4 (top and bottom,
respectively). Results from OLS regressions on the individual respectively). Individual duplicate variances have been regressed
duplicate standard deviations on the duplicate means do not approxi- against duplicate means using one- and two-parameter perfect square
mate the true, underlying error introduced to these data-sets and quadratic models, respectively.
significantly underestimate the parameters of the lines.
with non-negative intercept (because to reside in the positive
The second quadratic regression model is associated with a quadrant, b0, b1 and b2 must be non-negative.
Thompson–Howarth error described by both relative and Note that Thompson & Howarth (1976) acknowledged the
absolute error terms. It has the form: possibility of using a quadratic or higher polynomial model to
describe measurement error as a function of concentration.

2 = (mµ + b)2 = m2µ2 + 2mbµ + b2 Unfortunately, the model they referred to described error in
terms of duplicate standard deviations, and so was inconsistent
(or
=mµ+b in mean versus standard deviation space). When with additivity and would have produced biased results. In
the square root of this two-parameter (m, b) quadratic regression contrast, Thompson (1988) introduced a quadratic model that
model is calculated, the result is a straight line (
=mµ+b) with regressed mean against variance, acknowledging the advantage
a non-negative slope (m) and non-negative intercept (b) on a this model has in terms of additivity. Unfortunately, this model:
Thompson–Howarth scatterplot. Both of these first two quad-
ratic regression models are constrained to be perfect squares,
2 = m2µ2 + b2
and so when the square roots of these quadratic equations are
taken, these models result in straight lines on scatterplots of (Equation 13 in Thompson 1988) when plotted in mean versus
replicate mean versus replicate standard deviation. standard deviation space describes a non-linear measurement–
The third quadratic regression model is not associated with error relationship. Furthermore, this model is not general,
any specific Thompson–Howarth error model. It has the form: because it is merely equivalent to the quadratic model of
Equation 6, but without its second parameter.

2 = b 22µ2 + b1µ + b0 Consequently, the three general quadratic regression
models (Equations 4, 5 and 6) derived using non-linear OLS
regression and calculated in Excel using the ‘SOLVER’ utility
(or
=√b 22µ2+b1µ+b0 in mean versus standard deviation space). are presented for the two synthetic data-sets in Figure 6
This three-parameter (b2, b1, b0) quadratic model is not con- (Equation 4 for the first data-set, Equation 5 for the second).
strained to be a perfect square, and so there is no guarantee that The corresponding square roots of the three regression models
the square root of this quadratic model (
=√b 22µ2+b1µ+bo) will applied to the two synthetic data-sets are plotted on
be a straight line. As a result, when plotted on a scatterplot Thompson–Howarth scatterplots in Figure 7. In both cases,
of replicate mean versus replicate standard deviation, this the regression models produce slopes and intercepts closely
model most likely occurs as a curved line (concave downward) approximating the 5%, and 5% plus 2 ppm error magnitudes
 Thompson–Howarth error analysis 7

Fig. 8. Thompson–Howarth error analysis results from 3488 Cu

assays of duplicate drillcore samples from an anonymous porphyry
Cu–Mo–Au deposit. The square roots of the one-parameter (linear,
through origin), two-parameter (linear, with intercept) and three-
parameter (non-linear) quadratic mean-variance regression models
are plotted on this Thompson–Howarth scatterplot.

model in quality control/quality assessment programmes, and

(ii) the same collection procedures for duplicate samples used to
assess measurement error at various stages of preparation and
analysis, are probably the most important factors influencing
the conclusions of any quality control assessment. These ensure
that measurement of the same errors in different data-sets or
against minimum standards are made in a consistent way.
Four example duplicate quality control data-sets from oper-
ating mines that illustrate the differences obtained through use
of the three different large-sample Thompson–Howarth error
Fig. 7. Thompson–Howarth error analysis results for the two analysis methods are presented in Figures 9 to 12. These span
synthetically derived data-sets presented in Figure 4 (top and bottom, a range of mineral deposit types and commodities (Cu, Pb, Ni,
respectively). The square roots of the one- and two-parameter perfect Au), and provide insight into the magnitude of variation that
square quadratic models from Figure 6 illustrate that regression can occur when assessments of measurement error are under-
models derived using individual replicate statistics approximate the taken using different calculation methods and model choices. In
true, underlying error within these data-sets. all cases, duplicate standard deviations have been plotted against
duplicate means, and a relative error model has been deter-
introduced to these data-sets (Table 1). Clearly, these correctly mined using the three large-sample Thompson–Howarth error
founded, non-linear OLS regressions of replicate variance analysis methods consistent with additivity (i.e. grouped RMS
on replicate mean produce unbiased estimates of the true, variances were regressed using the two-parameter quadratic
underlying measurement error in a data-set. model against grouped mean concentrations, and the square
In order to illustrate the different characteristics of the three roots of the model are then calculated and plotted on a replicate
quadratic regression models, optimal OLS regression results for mean versus replicate standard deviation diagram). In one case
these three models are presented in Figure 8. These reveal the (Fig. 10; the Pb assays from an anonymous Broken Hill-type
typical relationships that exist between these models, and Pb–Zn–Ag deposit), the results are highly comparable. How-
illustrate that from a practical perspective, each model estimates ever, in the other cases, significantly different results are
a similar magnitude of measurement error across the range of obtained. In Figures 11 and 12, the original large-sample
concentrations. As a result, quadratic model choice is probably Thompson–Howarth error analysis method produces a much
not critical. However, the non-linearity of the three-parameter smaller estimate of measurement error than the other two
model adds complexity that may limit its utility in error analysis. techniques. In Figure 9, the results from the original
Similarly, the one-parameter model has an effective detection large-sample Thompson–Howarth error analysis method are
limit of zero, and so lacks the ability to identify a minimum intermediate between the results of the other two approaches.
reliable concentration (detection limit). As a result, the two- Figures 9B, 10B, 11B and 12B present frequency diagrams
parameter equation (Equation 5) may represent the best quad- for the coefficient of variation that illustrate what the distri-
ratic model for use. However, regardless of which model is bution of the relative errors should exhibit if the underlying
chosen, consistent use of (i) the same measurement–error measurement errors are normally distributed at 1%, 2%, 5%,

Table 1. Comparison of measurement–error models derived from several Thompson–Howarth error analysis-type methods.

Regression dependent variable (scatterplot ordinate) Data-set no. 1 slope (5% error) Data-set no. 2 slope (5% error) Data-set no. 2 intercept (2 ppm error)
Group median standard deviation 0.0507192 0.0476848 2.2166554
Group RMS standard deviation 0.0493528 0.0449496 2.0998716
Individual standard deviation 0.0401696 0.0369281 1.6933478
Individual variance 0.0504482 0.0454358 2.1852808
8 C. R. Stanley & D. Lawie

Fig. 9. Thompson–Howarth error analysis results from 3488 Cu Fig. 10. Thompson–Howarth error analysis results from 5500 Pb
assays of duplicate drillcore samples from an anonymous porphyry assays of duplicate drillcore samples from an anonymous Broken-
Cu–Mo–Au deposit. In (A), the OLS relative measurement error Hill-type Pb–Zn–Ag deposit. In (A), the OLS relative measurement
estimates derived using Thompson and Howarth’s original approach error estimates derived using Thompson and Howarth’s original
(black circles; dotted line; y=0.072517x), using Stanley’s modified approach (black circles; dotted line; y=0.028964x), using Stanley’s
approach (grey triangles; dashed line; y=0.089676x), and using a modified approach (grey triangles; dashed line; y=0.032118x), and
quadratic regression model for the individual duplicate statistics using a quadratic regression model for the individual duplicate
(open circles; solid line; y=0.059864x) are presented. In (B), the statistics (open circles; solid line; y=0.031055x) are presented. In (B),
distribution of relative errors is compared with what would be the distribution of relative errors is compared with what would be
expected from normally distributed errors. expected from normally distributed errors.

10%, 20%, 30%, 40% and 50% relative error levels (the thin line the duplicate concentration data exhibit something other than a
curves). On Figures 9B, 10B, 11B and 12B, the thick lines uniform distribution; this further undermines the ability of the
describe the corresponding frequency curves in the real dupli- two grouped data Thompson–Howarth error analysis tech-
cate data. These thick line curves are clearly fundamentally niques to produce accurate estimates of error (note that the two
different from what would be expected from normal error (half synthetic duplicate data-sets presented above have uniform data
normal distributions), as all have much larger positive tails than distributions, so the results from all three Thompson–Howarth
expected. As a result, none of these data-sets appear to exhibit error analysis approaches are highly comparable; Table 1).
even near-normally distributed errors. Thus, the disparities in Although weighted OLS regression could be used to address
the regression results of Figures 9A, 10A, 11A and 12A are this data grouping problem (using weights proportional to the
likely at least partially functions of the fact that these duplicate square of the range of mean concentrations in each group;
data do not have errors that are normally distributed (Stanley & Thompson 1988; Stanley 2003a, 2003b), a simpler approach is
Smee 2007; Stanley & Lawie 2007a). available, and is described above. Because OLS regression of
However, a cursory examination of the distributions of the individual replicate variances against the individual replicate
duplicate means in Figures 9 to 12 reveals that each of these means does not involve grouping of the data, it does not suffer
data-sets exhibits very positively skewed concentration distri- from the above positive skewness problem. It also is consistent
butions. This causes the groups of 11 with lower concentrations with the concept of additivity, and thus provides unbiased
to span small ranges in concentration, enabling the median and estimates of measurement error. As a result, OLS regression of
RMS standard deviations (or their proxies) to accurately esti- individual replicate variances against the individual replicate
mate error at these low concentrations. However, it also causes means represents the preferred approach for simply, accurately
the groups of 11 with higher concentrations to span large ranges and reliably estimating measurement error.
in concentration. This prevents the median and RMS standard The above arguments involving the assumption of normally
deviations (or their proxies) from accurately estimating error at distributed errors and the variations in accuracy of group
these concentrations because the duplicate data used to make statistics should give pause to geoscientists who, using the
these estimates span a large range in concentration (Fig. 1). As original implementation of the Thompson–Howarth error
a result, the quality with which the group statistics estimate analysis method, believe that their estimate of the magnitude of
error may vary across the range of concentrations observed if measurement error is accurate, and that their geochemical data
 Thompson–Howarth error analysis 9

Fig. 11. Thompson–Howarth error analysis results from 2156 Ni Fig. 12. Thompson–Howarth error analysis results from 928 Au
assays of duplicate drillcore samples from an anonymous dissemi- assays of duplicate drillcore samples from an anonymous high SX
nated sulphide Ni deposit. In (A), the OLS relative measurement epithermal Au–Ag deposit. In (A), the OLS relative measurement
error estimates derived using Thompson and Howarth’s original error estimates derived using Thompson and Howarth’s original
approach (black circles; dotted line; y=0.048989x), using Stanley’s approach (black circles; dotted line; y=0.099096x), using Stanley’s
modified approach (grey triangles; dashed line; y=0.063934x), and modified approach (grey triangles; dashed line; y=0.350179x), and
using a quadratic regression model for the individual duplicate using a quadratic regression model for the individual duplicate
statistics (open circles; solid line; y=0.072909x) are presented. In (B), statistics (open circles; solid line; y=0.360222x) are presented. In (B),
the distribution of relative errors is compared with what would be the distribution of relative errors is compared with what would be
expected from normally distributed errors. expected from normally distributed errors.

are ‘fit for purpose’ (Bettaney & Stanley 2001). Clearly, if

measurement errors are not normally distributed, and the four quadratic model, and then taking the square root of this model
examples considered above illustrate that this may be the rule to transform it onto a scatterplot of replicate mean versus
rather than the exception, and if the duplicate data exhibit a replicate standard deviation. Both of these adaptations are
non-uniform distribution, and most element concentration independent of any normality assumption, and are consistent
distributions are positively skewed, conventional Thompson– with the concept of additivity. Furthermore, under the assump-
Howarth error analysis will provide significantly biased and tion of normally distributed errors, these two methods and
inaccurate estimates of measurement error. However, if Thompson and Howarth’s original large-sample method are
the alternative individual mean versus individual variance equivalent.
large-sample Thompson–Howarth error analysis approach is However, the second approach avoids the detrimental
employed, these factors do not impact the results, and accurate interaction between data grouping and a non-uniform data
estimates of measurement error can be obtained. distribution, and is easier to implement in computer spread-
sheets. As a result, quadratic OLS regression of the individual
replicate variances on the individual replicate means, and then
CONCLUSIONS taking the square root of the resulting regression equation,
Thompson and Howarth’s original use of an ordinary least represents the preferred method to provide accurate esti-
squares regression of duplicate grouped absolute differences (or mates of measurement error that can be used to assess the
their proxy) against duplicate grouped means in their large- quality of geochemical concentration data in mining and mineral
sample error analysis method produces poor estimates of exploration applications confidently.
measurement error when measurement errors are not normally
distributed and when the data exhibit a non-uniform distri- This research was supported by a Natural Sciences and Engineering
bution. Two simple adaptations to Thompson and Howarth’s Research Council of Canada (NSERC) Discovery Grant, and a
regression have been developed to address these problems. The financial stipend and logistical support from CRC-LEME (Perth,
Western Australia) to the first author. The authors would like to
first adaptation involves the use of the group RMS standard thank the several mining companies which anonymously donated
deviation in a linear OLS regression of standard deviation their mine assay quality control data to this research project. Helpful
against mean. The second adaptation involves OLS regression reviews by Michael Whitbread, Richard Howarth and an anonymous
of individual replicate variance against replicate mean using a reviewer helped to improve the manuscript.
10 C. R. Stanley & D. Lawie

APPENDIX: A BETTANEY, L. & STANLEY, C.R. 2001. Geochemical data quality: the “fit-for-
purpose” approach. Explore, Newsletter of the Association of Exploration
In contrast to the two-standard-deviation effective detection Geochemists, 111, 21–22.
limit described in this paper, an alternative, three-standard- FISHER, N.I., LEWIS, T. & EMBLETON, B.J.J. 1987. Statistical Analysis of Spherical
deviation effective detection limit has been proposed (Analytical Data. Cambridge University Press, New York.
Methods Committee 1987). This detection limit can be calcu- FLETCHER, W.K. 1981. Analytical Methods in Geochemical Prospecting. Handbook
lated by substituting ‘3b’ for ‘2b’ in the numerator and ‘1
3m’ of Exploration Geochemistry, 1. Elsevier Scientific Publishing,
Amsterdam.
for ‘1
2m’ in the denominator of Equation 2 (Fig. 3). This FRANCOIS-BONGARCON, D. 1998. Error variance information from paired data;
alternative definition results in a larger effective definition of applications to sampling theory. Exploration and Mining Geology, 7 (1-2),
the detection limit, and although originally recommended 161–165.
in 1978 by the International Union of Pure and Applied GARRETT, R.G. & GRUNSKY, E.C. 2003. S and R functions for the display of
Thompson-Howarth plots. Computers and Geosciences, 29 (2), 239–242.
Chemists (IUPAC), this detection limit definition has been HOWARTH, R.J. & THOMPSON, M. 1976. Duplicate analysis in practice – Part 1.
‘slow to be adopted’ amongst applied geochemists (R. Howarth, Examination of proposed methods and examples of its use. The Analyst,
pers. comm.). 101, 699–709.
This may be because applied geochemists operate under PITARD, F.F. 1993. Pierre Gy’s Sampling Theory and Sampling Practice 2nd edn. CRC
different constraining parameters from analytical chemists. Press, Boca Raton.
SHAW, D.M. 1961. Manipulative errors in geochemistry. Transactions of the Royal
For example, until the recent use of inductively coupled mass Society of Canada, 60, 41–55.
spectrometry, many elements analysed in geochemical surveys STANLEY, C.R. 2006b. On the special application of Thompson-Howarth error
exhibited concentrations close to the conventional two- analysis to geochemical variables exhibiting a nugget effect. Geochemistry:
standard-deviation effective detection limit. Use of a higher Exploration, Environment, Analysis, 6, 357–368.
detection limit (such as the three-standard-deviation effective STANLEY, C.R. 2006a. Numerical transformation of geochemical data, II:
Stabilizing measurement error to facilitate data interpretation. Geochemistry:
detection limit) would thus define a large proportion of Exploration, Environment, Analysis, 6, 79–96.
samples as undetected, and could prevent the extraction of STANLEY, C.R. 2003a. THPLOT.M: A MATLAB function to implement
valuable information inherent in low concentration data. generalized Thompson-Howarth error analysis using replicate data.
Furthermore, in geochemical exploration, geochemists seek Computers and Geosciences, 29 (2), 225–237.
STANLEY, C.R. 2003b. Corrigenda to “THPLOT.M: A MATLAB function to
element concentration patterns to interpret, and thus rigorous implement generalized Thompson-Howarth error analysis using replicate
precision in individual samples is not required because the data” [Computers & Geoscience 29:2, 225-237]. Computers and Geosciences,
concentrations of many samples are interpreted collectively. 29(8), 1069.
As a result, lower effective detection limits, corresponding to STANLEY, C.R. & LAWIE, D. 2007a. Thompson-Howarth error analysis: 2.
Adaptations to the small sample method for assessing measurement
95% (not 99%) confidence levels, are likely to be acceptable error in geochemical samples. (In preparation). Geochemistry: Exploration,
in exploration. In mining applications, the issue of how to Environment, Analysis, 2.
define the detection limit is essentially irrelevant, because only STANLEY, C.R. & LAWIE, D. 2007b. Average relative error in geochemical
the precision of concentrations close to the cut-off grade is of determinations: clarification, calculation and a plea for consistency. (In
press). Exploration and Mining Geology.
prime importance. Cut-off grades are usually much greater STANLEY, C.R. & SINCLAIR, A.J. 1986. Relative error analysis of replicate
than the detection limit, regardless of how the detection limit geochemical data: advantages and applications. in: Programs and Abstracts.
is defined, and so these grades typically exhibit sufficient GeoExpo –, 1986, 77–78.
precision. STANLEY, C.R. & SMEE, B.W. 2007. Strategies for reducing sampling errors in
In short, whereas three-standard-deviation effective detec- exploration and resource definition drilling programs for gold deposits. (In
press). Geochemistry: Exploration, Environment, Analysis.
tion limits may be favoured by IUPAC (analytical chemists), the THOMPSON, M. 1988. Variation of precision with concentration in an analytical
choice of ‘3’ is arbitrary, and there is no reason why ‘2’ cannot system. The Analyst, 113, 1579–1587.
be used. Because geochemistry does not have the same analyti- THOMPSON, M. 1982. Regression methods and the comparison of accuracy.
cal precision requirements as chemistry, lower detection limits The Analyst, 107, 1169–1180.
THOMPSON, M. 1973. DUPAN 3, A subroutine for the interpretation of
that allow low concentrations to provide at least some infor- duplicated data in geochemical analysis. Computers and Geosciences, 4,
mation (even though there may be much noise in these data) 333–340.
may be favoured. Perhaps this is why the IUPAC definition of THOMPSON, M. & HOWARTH, R.J. 1973. The rapid estimation and control of
effective detection limit has not been whole-heartedly embraced precision by duplicate determinations. The Analyst, 98, 153–160.
THOMPSON, M. & HOWARTH, R.J. 1976. Duplicate analysis in practice - part 1.
by the applied geochemistry community. theoretical approach and estimation of analytical reproducibility. The
Analyst, 101, 690–698.
THOMPSON, M. & HOWARTH, R.J. 1978. A new approach to the estimation of
analytical precision. Journal of Geochemical Exploration, 9, 23–30.
REFERENCES ZITTER, H. & GOD, C. 1971. Ermittlung, Auswertung und Ursachen von
ANALYTICAL METHODS COMMITTEE 1987. Recommendations for the definition, Fehlern bei Betriebsanalysen. Fresenius’ Journal of Analytical Chemistry, 255 (1),
estimation and use of the detection limit. The Analyst, 112, 199–204. 1–9.

View publication stats