Applied statistics

J.W. MERKS
Degrees of freedom also occur in the
The variance is the fundamental
basic formula for the covariance, but
J.W. Merksis Dresident olMalrix
Consulrants
limited,
measure for random variations in stothey cancel in a simplified formula.
r . Columbia, Canada. nflnmeelin0 lla!Jer 95- However, covariances cannot be
chastic variables and stochastic sys- V a ~ o u ~ eBritish
319. Ori0inal mafluscri~lSe~lember1995; revised manuscri~l tested for statistical significance withterns. Variances are amenable to
r e asproved
v i e w e dPaller out taking degrees of freedom into
mathematical analysis. The proper- A P ~ I I ~ u~i ~ ~I .~ ~ ~ i f l n f l l l h i ~ ~ e e ~and
is
invited
and
must
be
submilled,
in
duplicare,
prior
to May 31,
ties of variances are derived from the
account,
variance of a general function. The
1991.
Associative dependence. Associaarithmetic mean and the weighted avtive dependence exists if the degree
erage are both functionally depenof association between two or more
dent variables of sets of measured values. The central
variables is statistically significant. An example in minlimit theorem defines the relationship between the varieral exploration is associative dependence between
ance of a set of measured values and the variance of the
grades and densities in massive sulfides. A significant
arithmetic mean of the set.
degree of associative dependence permits regression paThe variance of the weighted average, despite its sigrameters to be used to calculate densities from grades
nificance in mining, metallurgy and exploration, is not as
and the variance of density from the variance of grade.
well documented as the central limit theorem. As a reMeasured values in n-dimensional sample spaces
sult, kriging variances and covariances of sets of kriged
may display a significant degree of associative depenestimates (distance weighted averages) could evolve into
dence, which is frequently referred to as spatial depenthe essence of geostatistics.
dence. Ordered sets of on-stream data at mineralprocessing plants almost invariably display a significant
Dependenceand independence
degree of associative dependence.
Functional or mathematical dependence. The arithmetic mean is a function of a set of measured values with
The variance of the arithmetic mean
identical weight factors, and the weighted average is a
The arithmetic mean is a functionally dependent
function of a set of measured values with different
variable of a set of measured values with equal weight
weight factors. The requirement of
factors. It can be formulated as folmathematical or functional indepenlows
dence invalidates the variance and Abstract
covariance of a set of distance
Applied statistics provides % = (l/n)*x, + ... + ( l / n ) * x , + (1)
weighted averages. The concept of scores of powerful tests and tech- ... + (l/n)*x,
degrees of freedom in applied statis- niques that are widely used and
tics is the corollary of the require- abused in science and engineering. where
ment of functional independence in The variance is the basic measure k = arithmetic mean of a set of n meaprobability theory. Degrees of free- for variability, precision and risk.
surements,
dom are awarded when values are The formulas for the variances of xi = ith measured value and
measured but not when they are cal- the arithmetic mean and the n = number of measurements in the
culated from measured values. It is weighted average are derived from
set.
beyond the scope of this paper to the variance of a general function,
The variance of the arithmetic
prove that adding functionally depen- as defined in probability theory. mean can be derived from the fordent values to a set of measured val- The requirement of functional or mula for the variance of a general
ues does not add degrees of freedom mathematical independence in function. Intermediate steps in the
to the set.
probability theory translates into derivation are based on the premises
Because degrees of freedom are the concept of degrees of freedom in that the squared partial derivative for
fundamental in applied statistics, they applied statistics. Ceostatistics is each term in Eq. (1) is (l/n)2and that
occur not only in the formulas for the only variant of statistics in the variance of the arithmetic mean is
variances but also in many statistical which functional dependence and the sum of n terms as follows
tables. For example, the denomina- degrees of freedom are irrelevant.
tors in the formulas for the variances This paper examines whether such vat-(%)= (l/n)2*var(x,) +
of randomly distributed and ordered basic elements of probability and ... + (l/n)2*var(xi)+ ... +
(2)
sets of measured values are degrees statistics can be ignored with impu- ( l / n ) 2 * var(x,)
of freedom (Gy, 1979; Merks 1993). nity.
Introduction

78

FEBRUARY 1997 H M I N I N G ENGINEERING

When a set of measured values of a variable is obtained by applying the same measurement procedure to
the same sample space, the variance of the set is an unbiased estimate for the variance of each term, and the for-'\,
mula for the variance of the arithmetic mean becomes

TABLE 1

1 Different macrodiamond counts and sample I

Hole
number
1
2
3
4
SUM

which implies that

Equation (4) is the central limit theorem. It can be

used to show that the central tendency of samples selected of multinomial, binomial or Poisson distributions
tends toward the normal distribution. For example, if n
values are measured in randomly selected samples of different sample spaces, the variance of each term in Eq. (2)
should be weighted with the factor wi = niln, the ratio
between n,, the number of measured values in ith sample
space and n, the number of measured values in all
sample spaces.

The variance of the weighted average

If wi, the weight factor for each term, is defined such
that Cwi = 1 for the function, then the sum of all weight
factors Zln for the arithmetic mean meets this prerequisite. The weighted average is identical to the arithmetic
mean if each wi is equal to lln. Substituting the weight
factors in Eq. (1) gives the following formula for the
weighted average:

where
x = weighted average of a set of n measurements,
w, = ith weight factor,
x,= ith measured value and
n = number of measurements in the set.
Because the squared partial derivative for the ith
term is wi2,the variance of the weighted average of a set
of n measured values with weight factors wi is

Macro
count

Mass,
Kg

Weight
factor

4
2
15
19
40

16
12
64
72
164

0.098
0.073
0.390
0.439
1.OOO

Table 1 lists a set of macrodiamond counts determined in test samples of different mass.
The macrodiamond counts and sample masses in
Table 1 have one, and only one, mass weighted average,
namely, 0.098*4 + 0.073*2 + 0.390*15 + 0.439*19 = 14.7
or approximately 15 macrodiamonds. The variance of
this mass weighted average is 0.09S2*4 + 0.0732*2 +
0.3902*15+ 0.4392*19 = 6.0, for a standard deviation of
d6.0 = 2.45, a 95% confidence interval of 2*2.45 = k4.9 or
4.9*100114.7 f 33% and a 95% confidence range with a
lower limit of 14.7 - 4.9 =: 10 and an upper limit of 14.7 +
4.9 = 20 macrodiamonds.
The previous example shows how to calculate the
variance of the mass weighted average of a set of
macrodiamonds counts determined in test samples of
different mass. The next example is based on a hypothetical set of uranium data (Clark, 1979) and shows how
to calculate the variance of a distance weighted average
of a set of values measured at different positions within
a two-dimensional sample space.
Table 2 lists the measured values and their positions,
the position for which the weighted average is calculated,
the distances between the measured values, the inverse
distance for each position and its weight factor.
The arithmetic mean of the set of uranium values in
Table 2 is 366 ppm, and the variance of the randomly distributed set is 4,480 ppm2. The central limit theorem
teaches that the variance of the arithmetic mean is 4,4801
5 = 896 ppm2, for a standard deviation of 6896 = 29.9
ppm, a 95% confidence interval of t0.95;4*29.9 =
2.776*29.9 = +83 ppm or 83*1001366 = f23% and a 95%
confidence range with a lower limit of 366 - 83 = 283
ppm and an upper limit of 366 + 83 = 449 ppm. The tabulated t-value of 2.776 at 95% probability reflects that a
set of five randomly distributed measurements has 5 - 1
= 4 degrees of freedom.

Numerical examples show how to apply the formula

for the variance of the weighted average in sampling
practice. One deals with a mass
weighted average of macrodiamond
counts in test samples of different TABLE 2
mass, and the other deals with a disPositions and values for hypothetical uranium data - part one.
tance weighted average of uranium
values in geochemical samples seDatum Easting, Northing, Value, Distance Inverse Weight
lected at different positions in a twom
number
m
PPm
m
distance factor
dimensional sample space.
1
1,27 1
71 1
400
6.6
0.1523
0.319
In practical applications, the set
2
1,280
713
380
15.2
0.0656
0.137
of' diamonds detected in a test sample
3
1,268
722
450
9.6
0.1038
0.217
is often grouped into subsets of
4
1,265
704
280
9.1
0.1094
0.229
macro- and microdiamond counts, but
5
1,244
713
320
21.3
0.0469
0.098
they may be grouped into more than
Selected
two classes. In this example, only Position
1,265
713
macrodiamonds counts are used t o Weighted
calculate the mass weighted average Average
and its variance.
source: Practical Geostatistics (Clark, 1979).

MINING ENGINEERING W FEBRUARY 1997

79

TABLE 3

of 4496 = 22.3 ppnl, a 95% confidence

t0.95:Se22.3
=
interval
of
Positions and values for hypothetical uranium data - Part two. 2.306*22.3*2.306 = +51 ppm or
51*1001373 = +14% and a 95% confiDatum Easting, Northing, Grade, Distance, Weight
Aa
dence range with a lower limit of 373
number
m
PPm
m
m
factor
ppm
- 5 1 = 322 ppm and an upper limit of
5
1,244
7 13
320
373 + 51 = 124 ppm. The value of
I
4
1,265
704
280
23.2
1.630
40
t0.95;8 = 2.306 is listed in the t-distri1
1,271
400
71 1
9.1
0.636
120
bution at p = 0.95 and df = 8 and re2
1,280
7 13
380
9.4
0.664
20
flects that an ordered set of five
3
1,268
722
450
15.2
1.070
70
measured values has 2(n - I ) = S deSource: Practical Geostatistics (Clark, 19791.
grees of freedom.
The question of whether variSubstituting in Eq. (5) the uranium values and the
ances of 4,480 pprn' for the randomly
weight factors of Table 2 gives a distance weighted averdistributed set and 2.160 ppm' for the ordered set are
age of 0.319*400 + 0.137*380 + 0.217*450 + 0.229'k280+
statistically identical is solved by comparing their ratio of
0.098*320 = 373 ppni. Substituting in Eq (6) the weight
4,48012.160 = 2.07 with tabulated F-values. The F-ratio
factors of Table 2 gives the following terms for the variof 2.07 exceeds neither F0.99:4;8 = 7.01 (tabulated in the
ance of the distance weighted average o f 373 pprn
F-distribution for p = 0.99, dfr = 4 and dfo = 8) nor
F0.95;4:8 = 3.84 (tabulated at p = 0.95 with the same degrees of freedom). Therefore, the degree of spatial dependence in this two-dimensional sample space is not
statistically significant.
Because the variance for cach term is unknown. eiEven though these variances d o not differ signifither the variance of the randomly distributed set or the
cantly, the lowest may be used to calculate precision esvariance of the ordered set should be used in all terms.
timates. but it would be erroneous to claim that spatial
Substituting the variance of 4.480 ppm' for the randomly
dependence results in a significantly higher degree of
distributed set of uranium values gives
precision for the distance weighted average of 373 ppm.
The expression "marginally more precise" reflects in an
appropriate manner that the lowest variance gives a numerically more desirable degree of precision than the
highest variance does.
Mathematical analysis should not be applied to the
This variance is equivalent to a standard deviation of
difference between statistically identical variances, be41,029 = 32.1 ppm, a 95% confidence interval of
cause the difference of 4,480 - 2,160 = 2,320 ppm bet0.95;4*32.1 = 2.776*32.1 = f 89 ppm or 89*100/373 =
tween the variances of the randomly distributed and
*24% and a 95% confidence range with a lower limit of
ordered sets is only a random number. In geostatistics,
373 - 89 = 284 ppnl and an upper limit of 373 + 89 = 462
such
differences are entered into "smoothing relationPPm.
ships" to predict tonnages and grades (David, 1989).
The position of the distance weighted average of 373
Clark (1979) refers to inverse distance and extension
ppm is within the sample space defined by this set of uraestimates and reports the precision of the distance
nium values. Therefore, it makes sense to check whether
weighted average in terms of 95% confidence intervals,
the degree of associative dependence between ordered
and lower and upper limits of 95% confidence ranges.
values is statistically significant and, thus, whether the
Table 4 lists confidence limits calculated from statistical
probability to encounter continued mineralization
and geostatistical variances.
within this sample space is high. The question of whether
Because geostatistics is unencumbered with degrees
the ordered set exhibits spatial dependence within its
of freedom, it would be imprudent to apply ANOVA to
two-dimensional sample space can be assessed by applythe geostatistical variances in Table 4, but it can be aping analysis of variance (ANOVA) to the variances of
plied to obtain asymmetric lower limits of confidence
randomly distributed and ordered sets of measured valranges for variances. Dividing the variance of 496 ppm2
ues.
Table 3 lists the same data sets as Table 2 but ordered in such a manner that a systematic walk, which
TABLE 4
calls on each of the five positions only once, covers the
shortest possible distance.
1 Precision estimates for hypothetical uranium (
The variance of 2,160 ppm2 is a function of the differences between ordered measurements and the dis95%
95%
95%
tances between positions. Substituting in Eq. (6) the
Methodology
Variance
CI
CRL
CRU
variance of 2.160 ppm' and the weight factors of Table 2
Applied Statistics:
gives
284
462
+89
1,029
Random
322
424
f51
496
Order
Geostatistics:
349
396
Inverse Distance 138.6 f 23.6
340
392
163.8 f25.6
Extension
The variance of 496 ppm2 for the distance weighted
Source:
Practical
Geostatistics
(Clark
1979).
average of 373 ppm is equivalent to a standard deviation

Qtl

ccaallnav

iaa7

MINING ENGINEERING

for the ordered set of measured values by F0.95;8:.. =

1.94 gives a lower limit of 49611.94 = 256 ppm2 at 95%
probability, and dividing it by F 0 . 9 9 ; 8 ; ~= 2.51 gives a
lower limit of 49612.51 = 198 ppm2 at 99% probability.
Hence, the probability is much less than 1 % that geostatistical variances of 138.6 and 163.8 ppm2 are compatible
with the variance of 496 ppm2 for the ordered set.
Because the cost for data acquisition in mineral exploration is high and the geostatistical variances are low,
it is understandable that geostatistics is more popular
than applied statistics. How the geostatistical theorist
contrived to reduce geostatistical variances below confidence limits that probability theory and applied statistics
impose merits close examination and scrutiny. After all,
systematically low variances would underestimate the
risk associated with the measurement of grades and contents of ore deposits.
David (1977) describes how to krige 16 distance
weighted averages of the same set of nine measured values, and submits that "writing all the necessary covariances for that system of equations might be a good test
to find out whether one really understands geostatistics."
A powerful test to find out whether one understands statistics is to calculate the degrees of freedom for David's
system of equations. In applied statistics, a set of nine
randomly distributed measured values has n - 1 = 8 degrees of freedom, and the ordered set has 2(n - 1) = 16
degrees of freedom. In geostatistics, however, degrees of
freedom are deemed irrelevant.
It is tempting to dismiss degrees of freedom as restrictive, even redundant, if it were not for the fact that
the requirement of functional independence in probability theory translates into the concept of degrees of freedom in applied statistics. Because kriging variances and
covariances of sets of kriged estimates are mathematical
aberrations in applied statistics, the question is why
kriging became the quintessence of geostatistics.
Rendu (1978) reports that "statistical theory proves
that the geostatistical methods of ore reserves estimation
are greatly superior t o any other method." Given that
Clark's geostatistical variances of 138.6 and 163.8 ppm2
are significantly lower than the variance of 496 ppm2 for
the ordered set, Rendu's praise appears legitimate.
Yet, without the need for functional independence
and without the concept of degrees of freedom,
geostatistics enjoys an unfair advantage. So much so that
Armstrong and Champigny (1989), who noted that
kriging variances of sets of kriged estimates become
smaller as kriged blocks do, caution that oversmoothed
estimates should not be used for calculating recoverable
reserves. The authors did not reveal how much smoothing constitutes oversmoothing nor did they report that
kriged estimates approach the arithmetic mean of a set
of measures values as the distance between the selected
position and the measured set increases.
It is beyond dispute that the weighted average is a
functionally dependent variable of a set of measured values with different weight factors. It is also beyond dispute that the macrodiamond counts and test masses in
Table 1 define one, and only one, mass weighted average
but that the uranium values and positions in Table 2 define an infinite set of distance weighted averages.
Computing large sets of kriged estimates (distance
weighted averages) of small sets of measure values is
tantamount to perpetual motion in data acquisition. Se-

lecting the perfectly smoothed subset of kriged estimates

from the infinite set is already a daunting task, but calculating the covariance or the variance of that subset is an
exercise in infinite futility indeed.

Summary
The variance is the essence of probability theory and
applied statistics. The properties of variances are derived
from the variance of the general function as defined in
probability theory. The formulas for the variances of the
arithmetic mean and the weighted average are derived
from the formula for the variance of a general function.
Arithmetic means and weighted averages are functionally dependent variables of sets of measured values, but
arithmetic means are based on measured values with
identical weight factors and weighted averages on measured values with different weight factors.
The distance weighted average and its variance are
of crucial importance in mineral exploration. Set theory
teaches that two or more measured values of a variable
in an n-dimensional sample space define an infinite set
of positions and, thus, an infinite set of distance weighted
averages. In applied statistics, every distance weighted
average of the infinite set has its own variance.
In geostatistics, the distance weighted average metamorphosed into the kriged estimate (Clark, 1979; David,
1977). a simple genesis that evolved into a bizarre variant
of Byzantine complexity, a synthesis of probability without functional independence, statistics without degrees
of freedom and geomathematics with abandon. This
geostatistical variant of probability theory and applied
statistics is implemented with dogmatic determination to
emulate static stochastic systems of immense intricacy
and variability.
David (1977) alludes to the variance of the arithmetic mean as "the famous central limit theorem" and
mentions that it gives the variance of the mean of "a
group of n independent samples" but omits the variance
of the weighted average, which is, in fact, a homologue of
the central limit theorem. H e predicts correctly that
"statisticians will find many unqualified statements" in
the "few pages of theory" in which he "firmly grounds
the model" and that "the terminology may seem barbarous to statisticians" but appears unaware that the methodology, too, is barbarous.
David refers to "Student's t" but does not mention
that tabulated values of the t-distribution are a function
of degrees of freedom. H e refers to "independent
samples" but does not differentiate between spatial dependence and functional dependence. Not surprisingly,
his chapter on kriging does not refer to the requirement
of functional or mathematical independence. On the
contrary, David intimates that the ability to calculate the
variance and covariance of a set of functionally dependent variables (kriged estimates) is a test for geostatistical perspicacity.
Armstrong and Champigny (1989) caution against
oversmoothing, which seems to imply that the requirement of functional independence can be violated a little,
but they fail to define the perfect degree of smoothing.
The authors observe that "the kriging variance rises up
to a maximum and then drops off.'' and submit that this
extraordinary behavior is described by Brooker (1992)
who praises its robustness.
Rendu (1994) worries about "an endless list of
MINING ENGINEERING

FEBRUARY 1997

81

kriging methods" but wonders why skepticism remains.

Perhaps the skeptics wonder which kriging method is the
least biased, why kriging variances give the most precise
estimates, or whether kriging variances of small sets of
kriged estimates violate the requirement of mathematical independence as much as those of large sets do.
IVonetheless, Rendu pronounces that "geostatistics has
come of age," and predicts that "it is here to stay with all
its strengths and weaknesses."
In the section on sampling variograms, Gy (1979)
points out that the denominators in the formulas for the
variances of randomly distributed and ordered sets of
measured values are degrees of freedom. Even though
Gy's works are invariably quoted whenever geostatisticians require a reference to sampling theory or practice. the significance of degrees of freedom continues to
elude the geostatistical theorist. In fact. a geostatistical
reviewer recently postulated that "degrees of freedom is
an older terminology that is not relevant to the modern
development of statistics".
Matheron (1963) hails geostatistics as "a new science" but Philip and Watson (1986) dismiss it as "a
sham." Armstrong (1992) editorializes that Philip and
Watson's paper is "high on journalistic style and low on
scientific content" and is not surprised that "neither set
of authors succeeded in getting their ideas published in
the usual way in a technical journal." Armstrong commiserates with authors who "disagree violently with
geostatistics" and "express opinions that run against the
popular view" and whose papers are rejected because
they "do not back up their opinions with scientific fact".
The scientific facts are that kriging variances and
covariances of sets of kriged estimates violate the fundamental requirement of functional independence of probability theory and that the concept of degrees of freedm
is and will always be relevant in applied statistics.
Other scientific facts are that geostatistical "variances" are mathematical aberrations that cannot possibly give unbiased precision estimates as a measure for

e c o m l I A ~ V1007

r MINING ENGINEERING

the risk associated with the measurement of grades and

contents of ore deposits. Applied statistics, by contrast,
gives unbiased precision estimates in terms of confidence limits for grades and contents as intuitive and reliable measures for the risk that the progression from
prospect to producing mine entails.
The geostatistical theorist ought to explain how it is
possible to design a geostatistical model on the basis of
widely spaced exploration drill-hole grades when closely
spaced blasthole grades d o not even display a significant
degree of spatial dependence..

References
Armstrong, M.. and Champigny, N., 1989, "A study on kriging
small blocks," CIM Bulletin,Vol. 82, No. 923, March.
Armstrong, M., 1992, "Freedom of speech?." De Geo,statisticr.,
Julius MCMXCII, Numero 14.
Clark, I.. 1979, Practical Ceostatistics Applied Science Publishers.
Barking. England.
Brooker. PI.. 1986, "A parametric study of robustness of the
kriging variance as a function of the range and the relative nugget effect for a spherical semivariogram." Jourr~ulof Mathematical Geology,
Vol. 18, No. 5 .
David, M., 1977, Geostarislical Ore Reserve Esritrration, Elsevier
Scientific Publishing Company, Amsterdam.
Gy, P., 1979, Sampling of Particulate Materials, Elsevier Scientific
Publishing Company, Amsterdam.
r~.~
GrolMatheron, G., 1963. Prir~ciplesof C e o s ~ u t i s ~Economic
ogy, Vol. 58.
Merks, J.W., 1993,"Abuse of statistics." CIM Bulletit~.Vol..86. No.
966, January.
Philip. G.M., and Watson, D.F.. 1986, "Matheronian geostatistics
Q u o vadis'?," Journal of Malhemarical Geology. Vol. 18. No. 1 .
Rendu. J-M., 1978, "Kriging, logarithmic kriging and conditional
eupectation: Comparison of theory with actual results." 16th
APSCOM Svmposilrnz, Chapter 19.
Rcndu, J-M., 1994. "Mining geostatistics
Forty years passed. What Lies Ahead'?." Mining Engitrc,ering.
Vol. 46, No, 6. June.