Bias testing of cross-belt samplers

J
o
u
r
n
a
l

by G. Lyman*, M. Nel, F. Lombard, R. Steinhaus**,

and H. Bartlett

The cross-belt sampler is a device that is very widely used even

though it has not been ratified by the Standards associations. This
paper is concerned with the bias testing of two cross-belt sampler
designs against stopped belt samples and a cross-stream cutter at
the belt end. The bias analysis is carried out using size distribution
as the analyte and conventional t-testing and the Hotelling T
squared test, which are compared for their effectiveness. The paper
considers the issue of detectable bias and the statistical planning of
bias tests.

Introduction
The cross-belt sampler has been a contentious
machine for many years. The first testing of
the device was carried out by Siebtechnik in
Germany in 1989 (Reiners and Mohrhauer1).
This test was carried out on coal which was
such that the ash content of the coal was
almost constant as a function of particle size,
and the bias test was carried out using ash
content as the test analyte. Consequently, the
machine was given a clean bill of health and
the cross-belt or hammer sampler became a
well-known piece of equipment. The original
machine had a cutter head with parallel side
plates which were perpendicular to the motion
of the belt. The design concept for the sampler
was that it should cut a swath of solids off the
belt, of a width determined by the aperture of
the cutter, and propel these solids into a chute
at the side of the belt from where they would
be collected as a sample increment. These
machines are still common in the mineral
industry. They have the advantage of being
relatively cheap to manufacture and can be
retrofitted to an existing conveyor at low cost.
No transfer point is needed for installation as
would be needed for a cross-stream sampler.
Subsequent to the development of the first
cross-belt sampler, a machine was designed in
South Africa which came to be known as a
go-belt sampler. This design was patented by
JCI (van der Merwe and English2) and had
parallel plates that were angled with respect to
The Journal of The Southern African Institute of Mining and Metallurgy

* Materials Sampling and Consulting Pty Ltd,

Australia.
Kumba Iron Ore, South Africa.
University of Johannesburg, South Africa.
** Multotec Process Equipment, Johannesburg, South
Africa.
Consultant, South Africa.
The Southern African Institute of Mining and
Metallurgy, 2010. SA ISSN 0038223X/3.00 +
0.00. This paper was first published at the SAIMM
Conference, Fourth World Conference on Sampling
& Blending, 2123 October 2009.

VOLUME 110

JUNE 2010

289

Synopsis

the belt. The idea was that if the speed of the

head was equal to the belt speed, the plates
would have an angle of 45 degrees. Such a
configuration would make the relative velocity
of the solids with respect to the head
nominally parallel to the side plates.
The authors are unaware of any detailed
sampler test data in the public domain other
than the German publication, referenced
above.
In line with this situation, the Standard
associations around the world have never
ratified the cross-belt or go-belt sampler as a
device suited to sampling for commercial
purposes, even though the machines are used
extensively for non-commercial sampling (and
some commercial sampling) in the coal
industry, largely due to the ease and low cost
with which they can be retrofitted to an
existing conveyor. The machines are also used
as a primary sampler for PGM and gold ores in
South Africa. While the samplers are
mentioned in sampling Standards, they are not
recommended, and it is usually stated that the
machine can be used after bias testing if the
parties involved in the contract which the
samples are used to control are in agreement.
Generally, only samplers taking samples from
a falling stream of material are endorsed by
Standards.
The greatest criticism levelled against the
cross-belt sampler is that it tends to leave
material behind on the belt. In an attempt to
rectify this situation, the van der Merwe patent
also introduced the idea of equipping the

P
a
p
e
r

Bias testing of cross-belt samplers

sampler head with a brush on the trailing edge of the head.
The brush was intended to ensure that the belt was left clean
and that a correct sample, corresponding to a clean diagonal
swath of material being removed from the belt, was collected.
A second criticism is that the cross-belt sampler can lead
to breakage of the material. It can be shown that the energy
imparted to the solids during sampling is significant at high
head speeds and that a 10 m/s head speed corresponds to a
10 m drop of the solids and that the energy dissipated in the
solids is a quadratic function of the head speed. Of course, if
the solids from any sampler are permitted to drop down a
chute from an elevated belt, substantial breakage may occur.
This is an issue, particularly with coal samples.
This paper discusses the bias testing of cross-belt
samplers of the original design and of the go-belt design.
The original design machine differs from the original
Siebtechnik machine in that a direct drive is used to power
the head rather than a clutched system where the sampler
head first falls by gravity and is then picked up and driven
through the load on the belt to a point at which is starts to
decelerate towards its park position. It was shown by internal
test work at Sasol, directed by the first author, that the direct
drive system for the head produced better samples than the
clutch drive system.
The machines were constructed by Multotec Process
Equipment with an aperture of 130 mm and installed on a
product belt at the Thabazimbi beneficiation plant of Kumba
Iron Ore. The product belt carries nominally 300 tph of
washed iron ore, but often less, if one of the washing
modules is down. Coarse material (approximately 28 mm,
+12.5 mm) is deposited on the belt first with finer material
(-12.5 mm, +6.3 mm) on top of the coarse. The belt load is,
therefore, strongly vertically segregated, which is quite
desirable from the point of view of providing a challenging
circumstance for bias testing.
The bias testing was carried out on the size distribution
of the material.

The analysis of the results was carried out in two ways;

individual t-tests, and the Hotelling T-squared test, which is
a direct generalization of the t-test that takes covariance of
the random variables into account.

Test system
The product belt at Thabazimbi has a run of 27.7 m where
the belt is accessible downstream of the second installed
cross-belt sampler (the go-belt type). There is a belt end
sampler of the Vezin type which was refurbished by Multotec.
This last machine is mechanically correct. The belt speed is
1.47 m/s and the belt width 0.75 m with 35 degree idlers.
Figure 1 shows the layout of the samples on the belt. The
sampling was initiated by taking three sets of increments
with the cross-belt samplers at 2 second intervals; a belt end
sample cut was initiated at the same time as the last crossbelt cut. The belt was then stopped as quickly as possible. A
stopped belt increment of nominally 0.4 m was taken, as
shown in Figure 1. The belt was then started and a second
belt end cut taken about 8.7 seconds after the start, giving
the sampling pattern as shown. Care was taken to reject any
sample sets where there was any overlap of the sample
increments; this could be verified visually.
The full set of samples was taken in a 19 m length of belt.
Close spacing of the samples was used to avoid, as far as
possible, the introduction of variance between the samples
due to distributional heterogeneity. Three cross-belt cuts
were taken for every stopped belt sample to match the sample
volumes and a stopped belt increment was about 3 times the
mass of each cross-belt increment. Two belt end samples
were taken to bracket the other samples and these samples
were split in half by rotary sample division, after drying, to
approximately match the sample mass of the other samples.
Samples from each stop of the belt were combined (3 to 6
stops) to make up a sample of nominally 100 kg, which was
then sized on 25, 20, 16, 12.5, 9.5, 8.0, 6.3 mm sieves. The
results of the sizings are provided in Appendix 1.

Figure 1Nominal locations of the samples in the test programme

290

JUNE 2010

VOLUME 110

The Journal of The Southern African Institute of Mining and Metallurgy

Bias testing of cross-belt samplers

A bias test has the intention of making comparisons that can
be used to decide whether one sampling device or method is
biased with respect to another. Implicit within this procedure
is the definition of what is an acceptable or unacceptable level
of bias. This issue is all too often sidestepped and a test is
done on the data at hand and the test system is deemed to be
biased if the t-tests done on say, ash content or iron content,
turns out to be significant. If bias is detected, the detected
bias may or may not be economically significant. If one is to
go to the expense of carrying out a bias test, it makes a lot of
sense to decide before any testing is done just how big a bias
must exist before the bias has an economic affect. The test
program should be designed to detect the predetermined
economically significant bias. The cost of the test program is
bound to be inversely proportional to the level of bias to be
detected.
Testing a sampling system for bias with respect to particle
size, is generally a more powerful method than testing for
bias with respect to a chemical constituent, especially when
the chosen analyte is present in small concentration. The
sample preparation and analysis uncertainties may be
sufficiently large under those circumstances that a very large
number of paired samples may be needed to arrive at the
detectable level of bias that has been targeted.
Samplers commonly show a bias against size if they are
in fact biased. A sampler that is biased against particle size
is, without doubt, biased against the analytes of interest.
Furthermore, if a bias against size shows up in a test when
the size distribution is close to the mean distribution
expected over time, the sampler will almost certainly be
biased when other size distributions are presented to it.
Francis Pitard has made the perceptive observation that bias
tests can be misleading as biases are by nature not constant
in time due to the changing nature of the material sampled
and the mass flow rate in the system. However, if a size bias
is picked up in one test, it is almost sure to be there, to some
extent, in a subsequent test. Actually, as segregation on the
belt increases, bias is almost sure to increase. An exception
to this is the case of a material having a very wide particle
density distribution and a narrow size distribution. Such
cases are very rare.
Bias tests that are not planned to detect an economically
relevant level of bias are a waste of time and money. Given a
knowledge of the size by size distribution of the analytes of
interest, the bias induced in the analytes as a consequence of
a particular degree of size bias can be estimated.
Consider a case in which a sampler is missing the coarse
particles. Such a defect will cause the size distribution to
become narrower; a sampler that is missing fines will have
the same effect. Therefore, if one is going to design a bias
test using sizing as the critical factor, the test should be
designed to detect a particular level or extent of narrowing of
the size distribution. It is a simple matter to simulate progressively, more narrow size distributions. This narrowing can be
readily translated into analyte biases. If the suspected change
due to bias is a shift in mean particle size, this effect can also
be dealt with by simulation.

Statistical methodologies for data analysis

The comparison to be made in the data analysis is among the
The Journal of The Southern African Institute of Mining and Metallurgy

stopped belt samples, the belt end samples, and the two
samples from the cross-belt samplers. These samples are all
paired and so should all have the same expected value of the
size distribution. The sample masses are all similar so the
variance of the size analysis should be similar for all
samples. There should be only two sources of variance
attached to a single sample: the difference between the
100 kg samples due to the intrinsic heterogeneity with
respect to particle size and the uncertainty associated with
the sizing procedure itself. The careful pairing of the samples
should eliminate the uncertainty due to distributional heterogeneity along the belt as far as possible.
In making a statistical comparison between two measured
size distributions, it must be recognized that the mass
fraction reporting to a given size fraction is statistically
correlated with the mass fraction reporting to any other size
fraction. The correlation comes about as a result of having
added up the masses in each size fraction and then dividing
each mass by the total to arrive at the mass fraction in the
size fraction. Therefore, it is incorrect to treat the amounts in
each size fraction (or the amounts passing or retained on a
given sieve) as independent random variables.
Treating them as such is, however, a common practice.
This practice is justified only when a single measure of bias
is involved (e.g a single analyte assay). When more than one
analyte and particularly when sizings are involved in making
paired comparisons, the covariance structure of the data must
be taken into account. The temptation to carry out individual
t-tests on the mean differences derived from n independent
measures of the difference should be avoided.
There are three problems associated with such a
procedure. First, working only with mean differences ignores
that fact that the bias in the sampler may be a bias of scale
which is not discovered using mean differences (see
Lyman3). Secondly, covariance in the data set is ignored, and
thirdly, one has to decide how many of the t-tests must fail
in order for the sampler to be deemed biased.
Does it suffice that one t test fail at the 95% level? Or
must a certain fraction of all possible t-tests fail at the 95%
level. With p size fractions defined, it is possible to carry out
p 1 t-tests (the pth test is not independent due to the
correlations of the mass fractions). Appendix 3 deals with
this problem comparing t-testing to the Hotelling T-squared
test, which fully accounts for the correlation between the size
fractions and provides a single test statistic for the entire
group of size fractions. It is shown there that determination
of the number of t-tests that should fail before deciding is not
trivial and that deciding, for example, that three tests must
fail for the sizing to be different may be too strong a
criterion, whereas deciding that two failures are sufficient
may be too weak a criterion. Use of the Hotelling statistic
entirely circumvents this problem.

Use of the t-statistic

Consider first a simple means of testing the difference
between two size distributions. Assume that there are n pairs
of observations on the size distribution which can be denoted
as observations xi, yi. These vectors contain p elements
giving, say, the mass fraction within p of the p + 1 size
fractions. The difference vectors, using p size fractions are
VOLUME 110

JUNE 2010

291

The detectable level of bias

J
o
u
r
n
a
l

P
a
p
e
r

Bias testing of cross-belt samplers

[1]
These can be put into a matrix D (n columns x p rows),
where the form the columns of D.
The mean difference vector is
[2]
The standard deviation of the mean difference of the
mass fractions in each size fraction can be calculated as
[3]
where ij is the j th element of the vector i and i is the j th
element of the vector .
If one were to focus interest on a single size fraction, one
would use a t-test and form the statistic
[4]
and this has a Students t distribution with n 1 degrees of
freedom. This test could be applied to each size fraction of
interest in turn.
The difficulty with carrying out such a test is that while
the mass fraction in one size may appear to be biased (that
is, it fails to pass the t-test), the material that is apparently
missing (or in excess) from that size fraction must consistently appear in some other size fraction or set of size
fractions for there to be a consistent bias. Therefore, t-testing
based on single size fractions fails to take account of the fact
that the missing or excess material must go into or come from
some other size fraction(s). When one considers that it is
necessary to test more than one size fraction at a time, it
should be immediately recognized that one is dealing with
two correlated, random variables and that simple t-tests that
assume independence of the two random variables are
inappropriate. The Hotelling T-squared test intrinsically
recognizes the correlation or covariance between mass
fractions in the various size fractions as is appropriate to the
technical realities of the situation.

Use of the Hotelling T squared statistic

Rather than simply estimating the variance of the amounts in
the size fractions, one estimates the covariance matrix of the
differences in the amounts in the p size fractions. This matrix
is estimated in an unbiased manner as
[5]
The T-squared statistic is then
[6]
and this is distributed as F(p,n p).
Note that one must have at least p + 1 pairs of
observations if the test is to be possible. With p + 1 pairs, the
F distribution has (p,1) degrees of freedom. This distribution
is broad and the power of the test will be poor unless the
number of observations exceeds p + 1 by some margin.
Note also that the T-squared statistic is proportional to
the square of the Mahalanobis distance of the mean

292

JUNE 2010

VOLUME 110

difference from the origin. The virtue of the T-squared

statistic is that it can be simply tested for significance using
the F distribution, if the data distribution is normal.
One must not make the mistake of using all p + 1 size
fractions in the test as the estimated covariance matrix is
singular in such a case as a result of the fact that the mass
fraction in any size fraction can be found by subtracting the
sum of the amounts in all other size fractions from one. The
singular matrix cannot be inverted.

Setting the detectable level of bias in the bias test

plan
The planning of a bias test must begin by estimating the size
distribution of the material in the process stream sampled by
the sampler to be tested. Then, it is necessary to determine
the precision of the sizing procedure to be used for the
samples taken from the process stream. This must be done by
collecting a large sample of material from the process stream,
homogenizing and dividing it by rotary, sample division into
a substantial number of nominally identical subsamples (10
to 15 is good and there must be more subsamples than there
are size fractions defined). The mass of the subsamples to be
sized must be large enough that the sizing is reasonably
accurate. Smaller samples will carry larger uncertainties due
to Poisson error and will lead to a requirement for a larger
number of pairs of samples to attain the detectable bias
target.
For the test work that is reported in this paper, the
covariance matrix derived from the sizings of 15 nominally
identical 100 kg samples was compared with the covariance
matrix derived from the simulation of sampling of an iron ore
with a similar size distribution. The simulation allowed for a
Poisson uncertainty in the number of particles in each size
fraction and a normally distributed random uncertainty in the
mass of material reporting to each size fraction. The former
component is what is described by Gys fundamental error as
applied to sampling for size distribution and the latter is
intended to account for misplacement of material from a size
fraction or loss of material in the sizing process.
The result of sizing of the 15 nominally identical samples
is reported in Appendix 1 (see Table A1-II). The size distribution used for simulation is shown in Figure 2. This distribution is slightly different from the observed mean distribution, but the initial distribution used for planning the tests
differs from that found in the test work: such a difference is a
hazard of the planning, but the differences are not gross.
From that Table, with the exception of the coarsest size
fraction, it is clear that the fundamental error is smaller than
the observed uncertainty. If one allows for an uncertainty of
2.4% relative (1 sigma) due to the analytical uncertainty,
the observed uncertainties are adequately modelled by the
simulated uncertainties.
Using the size distribution of Figure 2, which is linear on
the Rosin Rammler coordinates, it is easy to make the distribution narrower simply by pivoting the line about a point.
The 40% passing size was chosen as a pivot.
The determination of detectable bias is described in
Appendix 3. Basically, one changes the size distribution from
the base distribution and then carries out Monte Carlo
simulations of sampling from both the base and biased size
The Journal of The Southern African Institute of Mining and Metallurgy

Bias testing of cross-belt samplers

J
o
u
r
n
a
l

Rosin-Rammler coordinate

0.8
0.6
0.4
0.2
0
-0.2
-0.4

P
a
p
e
r

-0.6
-0.8
-1
1

10

100

Particle size (mm)

Figure 2Size distribution used for estimation of detectable bias

Results
It is instructive to note that based on the aperture of the
hammer samplers, the actual sample masses were 83 and
85% of the expected masses for samplers H1 and H2. This
suggests that both hammer samplers were leaving material
on the belt (or that the usual method of estimating the
hammer sample increment mass is in error).
The increment masses were checked after drying of the
samples to ensure that the samples had not been corrupted.
The expected increment masses were compared with the
actual increment masses. For a given sampler, these ratios
The Journal of The Southern African Institute of Mining and Metallurgy

were normally distributed over the 100 samples collected. For

the first hammer sampler (H1), it was observed that 10 of the
samples were substantially heavier than expected. During the
sampling, it happened that when the belt was restarted after
collection of a set of samples, the belt was overloaded and
there was spillage down the hammer sampler collection
chute. This spillage had to be dumped from the sample
bucket before collection of the next set of samples. Based on
the suspicion that some non sample material might have
been present in the bucket, increasing the sample weight,
those sets of increments where the hammer 1 sample was too
heavy were rejected and not used to make up the nominally
100 kg samples that were sized.
The data collected are provided in Appendix 1. The belt
end, hammer 1 (conventional hammer sampler) and the
hammer 2 (angled hammer sampler or go-belt sampler)
were tested by taking the differences between size distributions that they collected and the belt cut (stopped belt)
samples.
The results of making individual t-tests are provided in
Table I. The belt end sampler is apparently unbiased whereas
the other two samplers are significantly biased at the 95%
confidence level.
It is relevant to note that the mean differences for the belt
end sampler are under 1% absolute, with some about 0.5%
absolute, which is of the order of the detectable bias (see
Figure A2-2). The differences for the hammer samplers
exceed 0.5% in most cases and are, therefore, of a magnitude
which will cause the Hotelling test to indicate bias and to
cause a number of t-tests to fail.
The results of the Hotelling T squared tests are provided
in Table II.
The Hotelling T squared tests show all samplers to be
biased compared to the belt cuts. To find that the belt end
sampler is biased according to the Hotelling test is surprising
and one must seek a reason for the apparent assignment of
bias. To look for causes of failure of the Hotelling T squared
test, it is instructive to make box plots of the distribution of
the differences for the 14 samples for each of the
comparisons.
VOLUME 110

JUNE 2010

293

distributions, forming the desired numbers of pairs of

samples from such samples. For example, 10 pairs of samples
were selected for the work in Appendix 3. These ten pairs can
be used to calculate a T squared statistic and the value
compared to the appropriate F statistic. This sampling and
comparison procedure can be carried out thousands of times
to estimate the probability that the samples will fail the F
test. When the bias between the size distributions becomes
large enough, the probability of the failure of the F test will
rise to 95%. It can then be said that the bias between the two
distributions is detectable at the 95% level.
If the level of bias that can be detected at the 95% level is
deemed to be too large, the number of pairs of samples used
can be increased (from the 10 used as a trial value) and the
Monte Carlo simulation carried out again. The result will be
an improvement in the amount of detectable bias. Appendix 3
shows that with 10 paired samples, and with the exception of
the coarsest size fraction, a difference of less that 1.0%
(absolute) in the amount in a size fraction can be detected.
Including the coarsest material raises this figure to 2%. Note
also that such differences expressed as relative uncertainties
are of order 7 to14%, which exceeds the uncertainty
associated with the determination of the size distribution by
some margin.
The test was designed to consist of up to 20 paired
samples so that the detectable bias would be under 2%
(about 1.3%). In the final case, 14 pairs of samples were
collected each of, nominally, 100 kg.

Bias testing of cross-belt samplers

Table I

t-test results (significant results shown in red). Differences correspond to test reference sample. Size
fractions run from + 25 mm to 6.3 mm
Mean
difference
(% abs)

Belt end
SD mean
Rel
difference difference
(%)

0.649
-0.482
-0.086
-0.194
-0.616
-0.188
0.029
0.160

0.318
0.394
0.257
0.484
0.313
0.274
0.279
0.250

5.41
-3.82
-0.73
-1.27
-3.15
-1.64
0.26
2.68

t stat

Mean
difference
(% abs)

2.04
-1.22
-0.33
-0.40
-1.97
-0.69
0.10
0.64

2.481
1.819
-0.287
-1.058
-1.750
-0.811
-0.211
-0.050

Hammer 1
SD mean
Rel
difference difference
(%)
0.483
0.496
0.169
0.343
0.336
0.227
0.350
0.267

Table II

Results of Hotelling T squared test

Sampler
T squared stat
F probability

BE

H1

H2

8.71
0.0053

7.47
0.0083

9.60
0.0040

Figure 3 shows such plots. The box shows the median

(red line) and the 25 and 75% quartiles (limits of box). The
whiskers show the value of the last observation that does
not differ by more than 1.5 times the inter quartile range
(height of the box). Points shown as red crosses can be
considered to be outliers if the points are assumed to follow a
normal distribution.
It is evident that there are a number of suspected outliers
(7) in the comparison between the belt end sampler and the
belt cuts. It appears that these outliers have contributed
substantially to the result for the Hotelling statistic.
It is possible to devise a ranked version of the Hotelling
test in which the observed differences are replaced by their
signed rank. In this case, it is not correct to use the F statistic
to determine the significance of the outcome of this modified
Hotelling test. Instead, the distribution of the robustified
statistic must be found using a permutation test and the
significance determined in that manner. Note that
permutation tests underpin the original derivation of the t
statistic by Student. Using the ranked or robustified version
of the Hotelling test, the hammer samplers are found to have
significant bias and the belt end sampler to have no
significant bias. The further use of this test and especially a
means of connecting the outcome of the test to a detectable
level of bias is currently under study by Lyman and Lombard.
The box plots provide a picture of the nature of the bias
suffered by the hammer samplers. The +25 mm material is
overestimated. For the conventional hammer sampler
(hammer 1), the 20 x 25 mm is also overestimated with
compensation taking place by an undersampling of the 12 x 8
mm material. For the go-belt sampler, 20 x 12.5 mm material
is undersampled and 12.5 x 9.5 material oversampled.

294

JUNE 2010

VOLUME 110

20.71
14.42
-2.44
-6.93
-8.95
-7.10
-1.85
-0.84

t stat

Mean
difference
(% abs)

5.14
3.67
-1.70
-3.08
-5.21
-3.57
-0.60
-0.19

2.450
-1.241
-2.877
0.194
1.667
-0.012
-0.150
0.101

Hammer 2
SD mean
Rel
difference difference
(%)
0.490
0.391
0.313
0.496
0.536
0.303
0.319
0.310

20.45
-9.83
-24.45
1.27
8.53
-0.11
-1.31
1.68

t stat

5.00
-3.17
-9.18
0.39
3.11
-0.04
-0.47
0.33

It was observed that the hammer samplers did not cut a

clean swathe across the belt. There was a good deal of
material left at the side of the belt, despite the fact that the
machines were adjusted prior to the tests. It is also to be
noted that the material on the belt was heavily segregated
with coarse material on the bottom of the load and finer
material on top. There was evidence of ploughing of
material behind the conventional hammer sampler and loss of
material from the increment on the go-belt sampler. The
observed bias in these samplers was not unexpected, given
the observation of the belt. It appeared that the material on
the belt was difficult to remove as the coarse material at the
bottom of the load tended to lock up.

Conclusions
Bias testing of a sampler is a critical factor in the testing of a
machine. Beyond the collection of the magical 30 pairs of
samples suggested by some Standards, it is rare to see any
statistically rigorous discussion of the level of detectable bias
that will be achieved in the bias test. It is obvious that the
detectable bias is a figure of economic significance.
The application of individual t-tests to the results of
testing a sampler on size distribution is incorrect as such
testing ignores correlation between the amounts in size
fractions. t-testing when more than one analyte is used is
also incorrect when the analyte contents in samples are
correlated (as they usually are).
The Hotelling T squared test, correctly, takes correlation
in the data into account when tests on simple differences are
used. However, as apparently illustrated by the results in this
paper, the Hotelling test is sensitive to outliers and general
non normality of the data. A robustified version of the
Hotelling test can be devised based on the signed rank of the
differences observed. Having replaced the actual differences
by their signed ranks, the calculations for the robustified
method do not differ from the method described in here. The
critical value of the robustified statistic must be determined
by permutation tests.
The important point to note is that replicate analyses are
necessary to assess the accuracy of the size analysis or
analyte analyses carried out. In spite of working many years
The Journal of The Southern African Institute of Mining and Metallurgy

Bias testing of cross-belt samplers

J
o
u
r
n
a
l

P
a
p
e
r

Figure 3Box plots for samplers: Top left: belt end; top right: hammer 1; bottom, hammer 2. Size fraction 1 is +25 mm and fraction 8 is -6.3 mm. Note the
difference in vertical scales on the plots

Acknowledgements
Thabazimbi mineThe authors would like to
acknowledge Thabazimbi Mine, in particular Cornelia
Holtzhausen, for allowing the mine to be used as a test
facility.
The Journal of The Southern African Institute of Mining and Metallurgy

Kumba Iron Ore ManagementWe would like to

acknowledge the following from Kumba Iron Ore
Management for giving their approval to do this very
important test work:
Chris Griffith (CEO)
Christo van Loggerenberg (Head of Technology and
Business Development
OreProxWe would like to acknowledge OreProx,
Sample Preparation Laboratory based in Kimberley,
Northern Cape, South Africa for part sponsoring and
conducting the required sample preparation work to
generate the source data used in this paper. Their work
was done quickly by a very competent crew.

References
1. REINERS, E. and MOHRHAUER, P. Hammer Samplera proven sampling unit,
Aufbereitungs-Technik, vol. 30, no. 4. pp. 218225.
2. VAN DER MERWE, E. English; Ralph H, Device for sampling material from a
moving belt conveyor, US Patent No. 5,115,688, filed 1990.
3. LYMAN, G.J. Determination of Bias in Sampling Systems, Minerals
Engineering, vol. 6. pp. 99105.
VOLUME 110

JUNE 2010

295

in association with mineral processors, the first author has

rarely seen replicate size analyses carried out with the
objective of measuring the accuracy (or at least the precision)
of the sizing method. Without this information, the detectable
bias in a bias test based on size distribution cannot be
reliably estimated. The same is true of tests based on chosen
analytes. For the masses used for each pair of samples in the
bias test, the total sampling plus preparation plus analytical
variance and the covariance between the analytes must be
quantified before the detectable bias can be estimated.
The issue of using Hotelling tests in a robust manner is
under study by Lyman and Lombard. The issue of estimating
both absolute bias and bias of scale for individual size
fractions or a set of analytes is being studied by Lyman.

Bias testing of cross-belt samplers

Appendix 1
Data from bias test

Table A1-I

Data from bias test

Sample

H1 Sample 1
H2 Sample 1
BE Sample 1
BC Sample 1
H1 Sample 2
H2 Sample 2
BE Sample 2
BC Sample 2
H1 Sample 3
H2 Sample 3
BE Sample 3
BC Sample 3
H1 Sample 4
H2 Sample 4
BE Sample 4
BC Sample 4
H1 Sample 5
H2 Sample 5
BE Sample 5
BC Sample 5
H1 Sample 6
H2 Sample 6
BE Sample 6
BC Sample 6
H1 Sample 7
H2 Sample 7
BE Sample 7
BC Sample 7
H1 Sample 8
H2 Sample 8
BE Sample 8
BC Sample 8
H1 Sample 9
H2 Sample 9
BE Sample 9
BC Sample 9
H1 Sample 10
H2 Sample 10
BE Sample 10
BC Sample 10
H1 Sample 11
H2 Sample 11
BE Sample 11
BC Sample 11
H1 Sample 12
H2 Sample 12
BE Sample 12
BC Sample 12
H1 Sample 13
H2 Sample 13
BE Sample 13
BC Sample 13
H1 Sample 14
H2 Sample 14
BE Sample 14
BC Sample 14

296

Initial
mass
(g)

+25

25 x 20

16 x 20

12.5 x 16

9.5 x 12.5

8 x 9.5

6.3 x 8

-6.3

106890.7
96893.6
122211.4
131795.2
116468.7
107837.8
129272
151180.6
63533.6
92819.7
107154
119238.9
80191.7
96437.1
111628.5
105801.8
95208.5
97659.5
108957.2
118741
80443.7
93998.4
110978.6
113439.8
101970.7
110455
118381.2
120593.7
96406.8
106661.3
119804.4
122111.3
73585.5
98025.6
107699.7
112416.7
75343.8
103269.6
116373.3
117126.8
77859.4
104637
118804.5
129356.5
59259.5
95509.2
95468.5
113510.4
102442.6
112105.7
129651.8
132705.7
116895.2
107281.2
128267.2
139238.4

9.182651
9.620759
7.983216
7.007691
9.180664
9.418404
7.643186
6.729567
13.29722
13.56005
10.85802
11.41934
10.08758
8.774424
7.796127
6.677864
13.05923
13.64138
11.11813
10.55086
23.98808
23.9652
21.62372
22.05355
15.92889
18.22561
16.01175
15.74875
13.25166
16.47917
12.34562
11.88948
11.61003
12.82451
11.46224
10.00101
11.45482
12.07974
9.90442
9.956048
13.99934
13.42948
11.75275
9.756139
17.46555
15.06808
12.75363
13.67593
17.9318
14.84037
17.22876
17.54537
22.02879
20.10324
18.33087
14.719

13.80045
12.10317
11.48379
13.87979
14.11761
11.02025
11.15145
11.78438
18.56608
12.81862
13.96607
14.18354
14.99557
10.64217
12.11232
10.44103
12.94412
11.41589
12.7328
12.05321
18.1517
16.39539
18.09862
19.17325
16.97664
15.86447
14.28419
15.38463
17.58932
10.15176
13.09676
13.94416
14.42703
12.79635
11.65351
10.89509
12.9716
9.81189
10.73537
10.76568
10.84609
9.466154
10.44304
11.10729
11.21761
7.451324
11.01777
8.808444
13.98412
9.613963
9.531067
11.92616
11.51809
9.71736
9.580158
12.29029

13.01778
11.486
13.11072
13.16315
13.22973
10.07365
13.84646
12.92679
12.31726
7.972984
11.38558
12.75129
13.12243
9.684447
13.49593
13.24221
12.10092
7.838357
12.61569
12.64391
12.17671
10.05751
11.08313
12.30194
11.37954
9.741343
11.18328
12.29815
11.46071
8.901917
12.48819
12.65952
11.08112
8.645293
12.37153
11.21631
10.66936
9.004392
11.29898
10.04569
11.26184
9.431081
11.67464
10.45537
9.152963
7.680098
9.285052
10.47957
9.529336
6.479064
9.593388
9.562815
10.24045
7.490781
10.12792
11.01729

16.95077
19.34173
16.53864
15.93283
17.38922
18.29831
16.28419
16.70472
14.04312
14.46708
15.56097
13.96549
13.70616
15.13277
14.60765
16.93884
17.38847
18.96969
15.89101
19.90635
11.54447
11.8202
12.45051
11.78043
12.15388
12.88733
14.57715
14.7686
13.05582
14.80481
13.7818
14.65827
13.34652
13.9078
14.35909
14.66401
16.68034
16.50447
15.71314
18.0226
15.25031
13.44763
14.89211
14.90965
12.36392
17.36775
14.887
13.80587
14.25559
18.11139
19.02434
15.81711
10.64475
11.24148
12.30237
11.7069

20.31739
22.55536
20.79912
20.37146
19.04486
22.62574
20.89927
21.61607
16.99762
21.70035
19.11277
18.55217
18.13679
25.26165
19.97859
20.69747
19.87669
23.00903
20.23629
20.60527
14.27433
15.85101
14.45774
13.8707
16.4409
17.05835
15.40388
17.736
16.54406
19.91866
17.68975
18.92601
19.19563
20.03864
19.43432
21.12782
18.95591
22.1863
21.13045
23.45108
18.18997
20.29942
21.25551
20.70735
17.96134
23.43575
19.02586
19.30396
18.28985
23.55563
20.1391
19.10204
15.02962
19.60754
15.57935
17.69332

11.681
9.946787
12.404
12.72854
10.68244
12.23078
11.54225
12.36878
9.533381
11.95393
11.35123
11.23291
11.1529
11.78219
11.9723
12.45366
10.60231
11.29445
11.46312
10.75366
8.095103
8.94079
8.550117
8.328118
10.02965
9.69182
11.89902
9.74769
11.35646
12.13055
11.09183
10.75879
12.78078
12.23282
11.17905
13.15481
10.73134
11.89314
12.10742
12.19593
10.67591
13.21378
9.945415
11.76617
11.50432
11.8041
11.5349
12.0196
9.729741
11.18016
9.506308
9.997008
9.922991
11.36583
12.65904
12.32663

10.16693
9.824178
11.29723
11.51195
10.76032
10.80706
11.54643
11.77115
9.210717
10.79361
10.63544
10.45892
12.43346
12.68132
13.03941
13.59249
9.890609
9.522371
10.35875
9.530491
7.751135
8.263864
8.019384
8.226566
11.05602
10.76022
10.64358
9.456547
10.96614
11.77831
12.06049
11.2621
10.97716
12.80614
12.34005
12.5541
12.26471
12.17735
12.04108
9.884672
12.52283
13.2672
11.93246
13.9533
13.40241
11.48549
13.91328
14.01246
10.63991
10.74361
9.234812
10.69901
14.7912
12.77726
13.13368
12.87454

4.883025
5.12201
6.383283
5.404597
5.595151
5.525799
7.086763
6.098534
6.034602
6.733377
7.129925
7.436332
6.365123
6.041036
6.997675
5.95642
4.137656
4.308849
5.584211
3.956258
4.018463
4.706038
5.716778
4.265434
6.034479
5.770857
5.997152
4.859624
5.775837
5.834825
7.445553
5.901665
6.581731
6.748441
7.200206
6.386862
6.271916
6.342718
7.069147
5.678291
7.253716
7.445263
8.10407
7.344741
6.931884
5.707408
7.582501
7.894167
5.639646
5.475814
5.742226
5.350486
5.824106
7.696502
8.286608
7.372032

JUNE 2010

Mass % in fraction

VOLUME 110

The Journal of The Southern African Institute of Mining and Metallurgy

Bias testing of cross-belt samplers

J
o
u
r
n
a
l

Table A1-II

Replicate size distribution data

Fraction (mm)
Sample
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Average.
SD actual
RSD actual (%)
RSD fund. (%)*

+25
2.960
2.985
3.112
3.023
2.772
2.856
2.479
2.955
3.161
3.112
3.510
2.849
3.080
3.261
3.284
3.027
0.243
8.028
13.250

20 x 25

16 x 20

12.5 x 16

9.5 x 12.5

8 x 9.5

14.057
14.073
14.719
14.905
14.376
14.308
15.018
14.966
14.689
15.155
14.583
15.233
15.020
14.717
14.957
14.719
0.373
2.534
3.811

12.489
11.945
11.212
11.091
12.330
12.123
12.220
12.537
11.625
11.897
10.088
12.177
11.871
12.100
11.835
11.836
0.635
5.363
3.208

12.095
12.745
12.484
11.971
13.310
12.681
12.499
12.126
11.546
11.634
12.201
12.416
11.447
11.770
11.818
12.183
0.513
4.213
2.332

14.350
14.309
14.333
14.568
15.449
14.213
14.000
14.342
14.754
14.669
14.545
14.104
14.788
14.025
14.073
14.435
0.381
2.637
1.611

9.298
8.634
8.936
8.969
8.381
8.784
8.703
8.770
8.778
8.748
8.574
8.656
8.990
8.657
8.970
8.790
0.218
2.485
1.573

-8
34.751
35.307
35.204
35.473
33.382
35.035
35.081
34.303
35.447
34.785
36.499
34.565
34.805
35.471
35.063
35.011
0.680
1.943
0.992

P
a
p
e
r

*Based only on the Poisson distribution of particle numbers in the size fractions

The objective of this appendix is to compare the behaviour of

individual t-tests on the difference in mass fractions, within
size fractions and the Hotelling T-squared test, on the set of
differences in mass fraction between the paired samples and,
at the same time, estimate the bias in the size distribution
that can be reliably detected.
It is important to determine the amount of bias that can
be detected at a given level of confidence when the
covariance matrix for the difference in size analyses is
known or estimated from the data pairs.
The solution to the problem will be approached by using
both analytical results and simulation. The application of the
results is intended to be relevant to the bias testing of a
sampler which is likely to show an increased or decreased
probability of collecting coarse particles. In such a case the
proportion of fines in the samples from the test sampler, will
correspondingly decrease or increase.

Methodology
To illustrate the concepts behind the comparison of the
methods, a particular size distribution will be chosen and the
tendency to collect more coarse particles will be modelled by
choosing a size distribution that plots linearly on RosinRammler coordinates and pivoting the cumulative passing
curve at a particular point. It is convenient to choose the 40%
passing point. One sampler will sample from the true size
distribution and the second sampler will sample from the
distorted distribution.
The sampling and sizing of the samples from the two
samplers will be simulated by allowing for a Poisson error in
the numbers of particles in each size fraction and a random
The Journal of The Southern African Institute of Mining and Metallurgy

normally distributed error in the mass of material reporting to

each screen. The magnitude of this, normally distributed
error, can be changed to make the size analysis of the
samples more or less accurate.
In the simulations presented below, the true size distributions from which samples are realised will be taken to be
constant as a simplifying assumption.
The amount of bias detectable will be sought by
increasing the difference between the true size distributions
until a situation arises in which one is 95% certain of
detecting bias using a given number of paired samples of a
given mass and the Hotelling test. The distribution of the
number of failed t-tests under the same circumstances, will
be determined.

Results
Figure A2-1 shows the statistical results for the base size
distribution of A2-I.
The results were collected by simulation of 5 000 datasets and using 10 pairs of samples. There are 7 size fractions
that can be t-tested and 6 size fractions used in the Hotelling
test. The Hotelling test uses one less size fraction than is
present as the last size fraction is linearly dependent on the
other and provides no additional statistical information.
When there is no difference in size distribution between
the two samples (slope factor = 1.00), the critical F value for
the Hotelling test is chosen as the 95% point for 6 and 4
degrees of freedom corresponding to the test of 6 size
fractions and using 10 pairs of samples. The covariance
matrix in the test is estimated from the data. At this stage,
the probablility that there will be no t-test failures out of the
7 is 0.826, but the probability that one t-test will fail is
0.168. If the criterion for the detection of bias is a failed ttest, then there is a probability of 0.168 (16.8%) of failing
the sampler when it is perfect.
VOLUME 110

JUNE 2010

297

Appendix 2Comparison of the hotelling T-squared

test and individual t-tests for paired size distribution
observations: detectable bias

Bias testing of cross-belt samplers

Table A2-I

Base size distribution for analysis (slope factor 1.00)

Size fraction (mm)

Mass %

25
20
16
12.5
9.5
8
0.5

31.5
25
20
16
12.5
9.5
8

Figure A2-2Biased and unbiased distributions and difference in distributions at the critical point for detection of bias by the Hotelling T
squared test

8.86
8.77
11.22
13.50
14.21
7.80
35.65

Figure A2-1Analysis of single t-test failures and Hotelling T squared

failure probability as a function of difference in test size distributions.
The function fitted to the Hotelling probabilities is a logistic function

Table A2-II

Results of Monte Carlo simulation for statistical evaluation of the individual t-test method and the Hotelling T
squared test method
Rosin-Rammler slope

Probability of Hotelling
test failing

1.00
1.02
1.04
1.05
1.06
1.07

0.0494
0.1950
0.6210
0.8162
0.9280
0.9766

0.8258
0.4326
0.0178
0.002
0
0

Increasing the slope factor to 1.02 to provides some

difference in size distribution from the two samplers raises
the probability of a failing Hotelling test from 0.05 to 0.195
but raises the probability of a single t-test failure to 0.528
and the probability of 2 tests failing to 0.04.
Changing the slope factor to 1.04 raises the Hotelling
failure probability to 0.621 but the probability of one t-test
failure is now 0.803 and the probability of 2 failures is 0.176;
the probability of no t-test failures is practically zero.
Using the Hotelling test with a critical value place at 95%
confidence, the slope factor can increase to 1.065 before there
is a probability of 0.95 of the test failing. This point can be
regarded as the level of bias that will be detected with a
probability of 0.95. The size distributions are shown in
Figure A2-2.

298

JUNE 2010

Probability that j t-tests fail

1
2

VOLUME 110

0.1678
0.5278
0.803
0.7504
0.6622
0.5764

0.0064
0.0396
0.1762
0.243
0.3196
0.3926

3
0
0
0.003
0.0046
0.0182
0.031

It is clear from these results that the discovery of a single

t-test failure is a poor indicator of size distribution bias. To
require the failure of two t-tests before concluding the
presence of bias is a too-harsh requirement.
When the Hotelling failure rate is 95%, the probability of
two t-test failures is 0.4, meaning that the probability of
making an incorrect decision based on the failure of two tests
is 0.6 and this probability of incorrect decision is
unacceptably high. The probability of incorrect decision for
the Hotelling test is only 0.05.
It would seem that the Hotelling test is a much superior
decision making tool when compared to the application of
single t-tests.
The Journal of The Southern African Institute of Mining and Metallurgy