EconomicGeology Vol. 64, 1969, pp.

538-550

A SimplifiedStatistical Treatmentof Geochemical Data by GraphicalRepresentation

CLAUDE LEPELTIER Abstract

Programme two selected in zones Guatemala, streamsediment of a reconnaissance was carriedout, and graphical methods interpretation of were attempted the search a in for simplified statistical treatmentof about 25,000 geochemical results. The data were grouped drainage lithological by and units,andthe frequency distributions the abunof

Inthe course mineral ofa explo-ation bythe sponsored United Nations Development

dance Cu, Pb, Zn andMo werestudied the form of cumulative of in frequency curves. The four elements appear be approximately to lognormally distributed. Background, coefficients deviation threshold of and levels weregraphically estimated. Examples are
givenof simple andcomplex populations. Mineral associations studied correlawere by
tion diagrams.
Contents
PAGE

ous targets for follow-up operationswould not be

encountered but rather more subtle features not so

easyto pinpointand interpret. Difficultyof the statistical approach the caseof 539 in The interpretation phaseof the surveywas charstream sedimentsurvey ..................... 539 acterized two essential by features:the great amount Adjustmentto a lognormaldistribution......... 539 of data to be analyzedand the lack of precisionof Definitions ................................. 539
Construction the cumulative of frequency curve 542 Comparisonwith histograms .................. 543 Informationgiven by cumulative frequency curves 544
Background ................................ Deviation .................................. Threshold .................................. Examples .................................. 544 544 544 545
these data.

Introduction

.................................

538

Advantagesof cumulativefrequencycurves ....

The coefficients of deviation ...................

546
546

Correlation diagrams .........................

Conclusion .................................. References ...................................
Introduction

548
550 550

Sampling and analytical methodsmust sacrifice precision speeddue to the nature of geochemical for prospecting,and the first consequence .this fact of is that an isolatedresult has little meaning in geochemistry. It must be part of a population as numerous and homogeneous possible. Indeed in as all kindsof phenomena, individualinaccuracies shade off progressively when observation extendedto is larger and larger populations. The first phaseof geochemical interpretationis to condenselarge massesof numerical data and extract from them the essential information. The most

objectiveand reliableway to do it (and sometimes Ti, United Nations Mineral Exploration Pro- the only one) is statistically. Large sets of numgrammein Guatemala relied heavilyon geochemical bers, cumbersome and difficult to interpret, may be prospecting. During one year (1967) 60 percent reducedto a useful form by the use of descriptive of the total Projectareawas covered systematicallystatistics.This is bestdoneby the graphical repreby a geochemical reconnaissance carried out in the sentationof the frequencydistributionof a given drainagesystems. Nine thousandstream sediment set of data; then the averagevalue, an expression samples were collected over about 12,000 km2 of the degreeof variation around the average,and

(roundedfigures). All the samples were analyzed

the limit

above which

the anomalies

start are im-

for copper and zinc, and the total number thinned mediatelyand preciselydetermined well as the as out to approximately4,000 before being run for existence one or severalpopulations the surof in lead and molybder/um. Finally about25,000 geo- veyed area.

chemical resultswere available compilation for and This treatment the data alsosimplifies comof the interpretation.As they accumulated, became it ap- parisonof the geochemical behaviorof an element parent that high-contrastanomalieswhich are obvi- in various geological surroundings of several or in unit. This article is publishedwith the authorization of the elements the samelithological

I am grateful Mr. Henry H. Meyer,Project to United Nations. The opinions expressed not necessarily are endorsedby this Organization. Manager of the Guatemalaand E1 Salvador Mineral
538

SIMPLIFIED STATISTICAL TREATMENT OF GEOCHEMICAL DATA

539

Surveys, and to Mr. Stephen S. Steinhauserfor technical criticismand much helpful discussion.
Difficulty of Statistical Approach in Stream Sediment Surveys

of sedimentary rocks,and others(Coulomb, 1959;

Cousins, 1956). In all theseexamples, character the studiedfollows the lognormallaw, which is probablymore
common than the normal one.

It is interestingto note here that the lognormal A reliable statisticalinterpretationrequires that law fits very well in the caseof low-gradedeposits a great quantity of data be treated and that these like gold but for high-gradedeposits, iron for indata be homogeneous. stance,the experimental distributions generally are In drainagereconnaissance surveys, the first con- negatively skexved because the limitationtowards of dition is easily filled but not the second. As a the high values. G. Matheron gives a thermomatter of fact, the importance samplingtechnique of dynamicinterpretation the proportional of effectin is sometimes overlooked this type of prospecting. in the caseof ore depositsand relatesit to the Mass But even if given the appropriate attention, too Action Law (Matheron, 1962). To the extent in many typesof rivers and too many lithological units which geochemical anomaliesare extrapolationsof are generally sampledto result in a homogeneous ore deposits theoryshould this applyto geochemical collectionof samples. The best way to limit the prospecting. inconvenience the heterogeneity the samples of of (particularly pH, organiccontentand grain size) is of to splitthe surveyareainto drainages lithological Construction the CumulativeFrequencyCurve and units, when possible,and to make the statistical A lognormaldistribution curve is definedby two interpretationfor each of them separately. How- parameters:one dependent the mean value, and on ever, even if this is done, the same degree of pre- the otherdependent the character value-distrion of cision cannot be achieved as in the case of a soil bution. This latter parameteris a measureof the survey where good homogeneity possible. is range of distributionof values,that is whether the
distributioncoversa wide or narrow range of values.

Adjustment to a Lognormal Distribution

Definitions

The two parameters can be determined graphically

as will be explainedon following pages. For practical purposes,we work on cumulative frequency curves,and their construction shall be explainedby meansof a concrete example. The various steps of this constructionare the following:

When dealing with a large mass of geochemical data, the first stepis to find what sort of distribution pattern best fits the various sets of observations. And, thus far, the lognormal distributionpattern appearsto be the one most applicable the results to (a) Selectionof a preciseset of data ("populaof most geochemical surveys (Ahrens, 1957). tion") as large and homogeneous possible. as In geochemical prospecting, study the content we (b) Groupingof the valuesinto an adequate numof trace elements in various natural materials, and
ber of classes.

to say that the values are lognormallydistributed (c) Calculating frequencyi occurrence the of in means that the logarithms of these values are distributed following a normal law (or Gauss' law) each class and plotting it against the class limits; well known as the bell-shapedcurve (Monjallon, this gives a diagram calledthe "histogram." (d) Smoothing the histogram to get the fre1963). quency curve. Many natural or economic phenomena can be (e) Plotting the cumulatedfrequencies ordias expressedby a value varying between zero and infinity, represented a skeweddistributioncurve. nates gives the cumulativefrequencycurve, which by If, instead of the actual value of the variable itself, is the integral of the frequencycurve. (f) By replacing the arithmetic ordinate scale we plot its logarithm in abscissae, the frequency with a probability scale the cumulative frequency curve takes a symmetrical, bell-shaped form, typical by of the normal distribution. This happenswhen a curve is represented .one or more straight lines. frequency curvesare shown phenomenon subjectto a proportionaleffect,that Examplesof lognormal is is to saywhenindependent initial causes variations in Figure 1. of

of the studiedvalue take effectin a multiplicative

way. It is the case,for instance, the distribution for of trace elementsin rocks, for the area of the dif(a) The larger the population be analyzed, to the ferent countries of the world, for the income of moreprecise and reliablethe results. If necessary,

Some briefcomments the different on steps follow:

individuals a country, the grainsizein samples as few as 50 values in for may be treated statistically but

54O

CL/I UDE

LEPELTIER

Figure 1. Lognorm. distribution i curves

Frequency ,%
Frequency,%

Value
Arithmetic scale

Value

Logarithmic scale

Cumulated frequency

Cumulated frequency .
99.99

,,

Value

'

'

, ,

Value

Logarithmic scale

Lolarithmic scale

of points (n) necessary construct correctline; to a the range of distributionof the values (R), ex(b) A correct grouping the values mandatory pressedas the ratio of the highest to the lowest of is if someprecisionis to be achievedin the statistical value of the population;and the width of the classes interpretation;too few classes will result in shading expressedlogarithmically(log. int.) which has to out important featuresof the curve; too many in be selected functionof the two first parameters. in losing significant details amidst a cloud of erratic Thesethreevariables are linked by the relation:
the analysisis meaningful.
ones. The results are distributed in classes, the

the confidence limits

must be calculated to see if

modulus whichshould proportional the preof be to cisionof the analyses: moreprecise analyses, the the
the smaller the modulus. The logarithmicinterval must be adapted to the variation amplitudeof the valuesand to the precision the analyticalmethods of (Miesh, 1967). In statistics, workingwith 15 to 25 intervals(or classes) is recommended. As a rule, the width of a class,expressed logarithmically, mustbe kept equal to or smallerthan half of standarddeviation(Shaw, 1964). For geochemical purposes, is convenient work it to with 10 to 20 points on the cumulative frequency line, that is to say with 9 to 19 intervals or classes.
There are three variables to consider: the number

log. int. -

log R

In mostof the cases variesfrom 6 to 300 (experiR mental average values), then, with (n) varying from 10 to 20, log R from 0.78 to 2.48, the extreme values for the logarithmic interval will be:
log. int. 0.78
20

- 0.039
O.25

2.48

log. int. -

10

The 0.10 was selected the best suitedlogarithmic as

interval for the classes because it suits most distri-

SIMPLIFIED

STATISTICAL

TREATMENT

OF GEOCHEMICAL DATA

541

bution, giving reasonable number of classes and a Practically,the histogram-frequency curve step is gooddefinition the curve. In case very reduced skippedand the cumulative of of frequencydirectly condispersion of the values around the mean, it may structed. However, note here an advantageof the be necessary use 0.05, and if the dispersionis histogram: it clearly illustrates the effect of the to specially large, 0.2 will be chosen. When the sensitivityof the analyticalmethodand more prelogarithmicinterval is selected, is easyto calculate ciselythe bias broughtto the low valuesby the use it a table giving the class limits in ppm. The only of colorimetric scales of standards. As a matter of precautionis to avoid starting with a round value fact, experience shows that there is an inevitable so that no analyticalresultswill fall on the limit of concentration the readings,whoeverthe analyst, of two classes. The most useful and commonlyem- on the values actuallyrepresented the colorimetric in ployedin geochemical work is the 0.1 log. int. classs scale. For instance, the caseof copper,the lower in part of the standardcolorimetricscale reads 0,2,4,7 table, a part of which is given below: . . ppm. Usually this results in an excessof 2, classlimit (log) .. 0.07, 0.17, 0.27, 0.37, 0.47, 0.57 4 and 7 values,and a conspicuous lack of 1, 3, 5 classlimit (ppm) . 1.17, 1.48, 1.86,2.34, 2.95, 3.72 ppm values. This is of importance for a correct It can be extended in both directions as far as construction the frequencycurve, and the raw of valuesmust often be corrected extrapolating by the necessary. (c-d) After selectingthe class table, the values general shapeof the curve. are groupedand the frequencycalculated for each (e-f) By plotting the cumulated frequenciesas one class (in percentage); then the frequencies are ordinatesinsteadof the frequencies, obtainsthe plotted against the class limits (the latter being integral curve of the preceding. It has the form of logarithmically calculated, ordinary arithmetic-arith- a straightline whenusingthe appropriate graphpaper metic paper must be used), giving a histogram (probability-log), and it is the one used in geowhich is smoothed a frequency to curve. But histo- chemicalpresentationand interpretation of the regrams are often misleading, being stronglyaffected sults. Then two questionshave to be answered: by slight changesin class intervals, and frequency where to start accumulatingthe frequencies,and curves are difficult to draw and handle: for instance, where to plot the cumulatedfrequencies ? it is necessary determinethe inflexion points of to As for the first point, the normal procedure folthe curve in order to evaluate the standard deviation. lowed by many authorsis to start cumulatingthe
Ftgure 2 - Cumulative lYequeuy Dtltributton for Zn and Cu

5o

2.5

0,5
O.X

0.3.

0.2

0.5

542

CLIUDE LEPELTIER

Figure3. Confidence limits(Pl, P2}=t 0.05probobllity levelCumulative

Frequency

inO/o

Number of

samples

against lower class the limits. Using the class center will entailan error of excess the centraltendency on parameters(backgroundand threshold) but not on the dispersion parameter(coefficient deviation). of This error, or difference, varies with the type of classes used and is easily calculated (6% for the

0.05 logarithmic class interval,12% for the 0.1 log. int. and 26% for the 0.2 log. int.). If the class limit is used,curvesconstructed from differentlog.
int. classes can be directly comparedwithout correction.

Let us take a concreteexample: the distribution of Zn in the quaternary alluvial deposits Block I of (Fig. 2). There are 989 resultsrangingfrom 10 to 230 ppm.
population:N-989 range:R230 10

- 23

The best class interval is selectedas explained

150

above'

log. log 1.36 0.097 int. n R 14 =

A 0.1 log. interval will give 14 intervals,which is acceptable. Usually, the histogram-frequency curve step is skippedand the cumulativefrequencydiagram directlyconstructed. 2 000 In Figure 2, the points fit fairly well along a 5OOO straight line, suggesting lognormaldistributionof a 10 000 zinc in the alluvial deposits. Actually, the points never fit the line exactly,but this doesnot matter provided they stay in a channeldelimited by the confidence limits usually taken at the 5% probSource: t iezou, A. Initiatiou pratique slatistiwe, i la Gauthief Villars, Parrs, 1961. ability level. This confidenceinterval has been drawn on Figure 2 by usinga graph (Fig. 3), which avoidsfastidious calculation and givesa fairly good frequencies from the lowest values toward the highprecisionfor the cumulativefrequencyvalues beest (Fig. 1) (Hubaux, 1961; Termant and White, 1959). However, one has to considera property tween 5% and 95%. The width of the confidence of the probabilityscaleused as ordinates:the values channelis inverselyproportionalto the importance considered: bigger the populathe zero and 100% are rejectedat the infinite; it does of the population tion, the narrower the confidence interval. To not matter for zero becausezero% never occurs, but in each case the last cumulatedfrequencyis check that a distribution fits a lognormal pattern, test (Rodionov,1965; 100%, and this value is impossibleto plot, lost one shoulduse the Pearson's Vistelius, 1960), but this longer operation is genfor the curve. Then consideringthe lack of precision in the low values and the importanceof the erally not warranted in this type of interpretation high ones for the determination of the threshold and, for practical purposes,the graphical control above is satisfactory. level, I considerit much better to cumu, the fre- described late
5OO 4OO

quencies from the hi#hestto the lowestvalues; thus, the 100% will correspond the lowest classand be to
eliminated.

Comparison with Histograms

For comparison purposes cumulative the frequency As for the secondpoint, the curve being an in- curve for Cu in the Motagua drainage (Fig. 2) tegral one, the ordinates must be plotted at class was also constructed, then, in Figure 4, the corlimits and not at class center; then, since one responding histograms and frequency curvesfor Cu cumulatesthe frequenciesfrom the highest values and Zn. Figures 2 and 4 presentthe samedata in to the lowest,cumulated frequencies to be plotted two different ways. Before enumerating are and com-

A SIMPLIFIED STATISTICAL TREATMENT OF GEOCHEMICAL DATA

543

Figure4. Histogrum frequency .nd curvefor Zn md Cu

I 1.5 2 3 4 5 6 7 8 O I0 15 30 40 50 GO70-80 100 Z. 150 200 300

x-'
_

I T-'I Illll

I
'.

..

tz.
I

:'"''

C.t

I.

': C.

q/

I.

....... ,,
19l 1. 1, 1,t l, l, ],1l 4.1 i LII 9.l

.....

11.1 14,8 18.f l,4 19, 1.l 1.81 1.9 14,1,

Information Given by Cumulative menting on the advantages the former presentaof tion over the latter, an interesting feature of the Frequency Curves histogram should be mentioned: in the case of The main purpose constructing cumulative in the colorimetricdeterminations made in the lower range frequency curvefor a given population to check is of sensitivity the analyticalmethod,the histogram of if it fits a lognormaldistribution, and if it does,to showsclearly the bias introduced the readingsby in estimate graphically basic its parameters: background the humanfactor and by the accuracy and sensitivity
limits of the method.
standard

This

effect is illustrated

for
of

(b), coefficients deviation s', s") and threshold of (s,

copper in Figure 4, where the classesincluding a

colorimetric are shaded and the value

level (t).

(b) gives an idea of the averageconcentration

level of the elements a given surrounding. in

the standarditself is given as a larger figure (1, 2, 4... ppm); the cumulation the frequency of reduces this effect,particularlyif it is startedfrom the high values,but it may be necessary bring somecoro to rections to the low value frequenciesin order to constructa precise distributioncurve. ComparingFigures 2 and 4, one seesimmediately

(s) expresses scatterof the valuesaround the (b): it corresponds the spread the values to of and
their range,from the lowestto the highest.

(t) is a complexnotionwhich might be termed

"conditional": statisticallyit dependson the prob-

ability level chosen; geologically, for practical and

fluctuations (b): it depends (b) and (s). of on The values equal or higherthan (t) are considered to
anomalous.

it to ihat it is easier compare straight to two linesthan purposes, is supposed be the upperlimit of the
two overlappingbell-shapedcurves; many more populationscan be presentedon the same diagram by using cumulativefrequencycurves than by using histograms. Cumulative frequency curves are of easier constructionand more precisethan ordinary frequency curves; it is simpler to draw a line that fits a set of pointsthan to draw a bell-shaped curve with inflexionpoints.

Adjustmentto the lognormal law is generallythe casewhensoil samples considered: the drainare in age reconnaissance surveyin Guatemala, found we
that trace element contents in stream sediments

appearalso to be lognormally distributed.

544

CLAUDE LEPELTtER

of the deviation: it is inverselyproportionalto the slopeof the line. We call it the geometricdeviation A straight line denotesa single populationlog(s'); it has no dimension: it is a factor obtained normally distributed. In this simplecase,the backby dividing the value read in .4 by the value read ground value (b) is given by the intersectionof inO: the line with the 50% ordinate. In the examples 21 given in Figure 2, we have:
s' -

Background

background value for copper .. b (Cu) --9.2 ppm background value for zinc .... b (Zn) = 48 ppm
Of course, these values must be rounded off; it

9.2

2.28

Then multiplyingor dividingthe background value by the geometricdeviationwill give the upper and will be illusoryto imply a precision out of reach lower limits of a range including68% of the populafar of the analyticalmethods. In the illustratedexample, tion (from b-s to b+s, or A'A on the figure). 10 and 50 ppm are taken as reasonably good ap- Multiplying or dividing by the square of the geometric deviationgivesa range includingabout95% proximations of the background levels. In the case of a perfect frequency distribution of the values( b -- 2s to b + 2s). curve, the backgroundthus calculatedcorresponds Becauseall the reasoningis made on logarithms, to to the mode (most frequent) and median (50% of it is also necessary expressthe deviationby a logarithm: the coefficientof deviantion (s) is the the valuesabove,50% below it) values,and is the deviation(s'). geometric meanof the results. This geometric mean logarithm(base10) of the geometric is a more significantvalue that the arithmetic mean. s' = 2.28
It is also a more stable statistic, less subject to change with the additionof new data and lessaffected by high values.
Deviation

s = logs = 0.36

It will be seen later that it might be interesting

to consider a third deviation index: the relative

deviation (s") sometimes called coefficient variof as Before explaininghow to determinegraphically ation. It is expressed a percentage: the deviationcoefficient, essential an property of the $ normal distribution(i.e., fitting the "bell-shaped" s"= 100 curve) mustbe recalled here: 0.36 (b) beingthe medianvalueand (s) the standard
deviation then:

s" = 100

9.2

- 3.9%

68.26% of the population falls betxveenb-s

and b + s

Threshold

This holds true in the case of the lognormaldistribution since the logarithms of the values are

After the background and the coefficient deviof ation,the third important parameter the threshold is 99.74% of the population falls betveen b- 3s level (t), whichis a functionof the two former. It and b + 3s has been seen that in the case of symmetricaldistribution (either normalor lognormal)95% of the 95.44% of the population falls betveen b- 2s
and b + 2s

individual values fall between b + 2s and b-

2s,

that is to say that only 2.5% of the population normallydistributed. Then, rounding the above- exceeds off the upper limit b + 2s. This upper limit mentioned percentages taking (b) as the back- is conventionally and taken as the thresholdlevel (t) ground,we can say that 68% of the population falls above which the values are considered as anomalies: between b-s and b-I-s or that 32% is outside log t = (log b) + 2s theselimits. The distribution curve beingsymetrical around an axis of abscissa (b) (Fig. 4), 16% of or to avoid usinglogarithms: the values will fall aboveb -I- s and 16% belowb -- s. t = b Xs '2 In Figure 2, the values b+s and b-s will be obtained projectingthe intersection the disby of t = 9.2 X 5.2 = 47.8 ppm
tribution line with the ordinates 16 and 84% on the abscissaaxis. Working with logarithms, one
has to consider the ratios and not the absolute values

Practically,(t) as well as (b), is read directly on the graph as the abscissa the intersection of of

thus established.Taking the sameexampleof Cu the distribution line with the 2.5% ordinate. In one reads47 ppm, and the slightdif(Fig. 2), onedetermines points (at the 16% this example the P ordinate)and.4. 0.4 is the geometrical expression ference is due to the rounding off of the exact

A SIMPLIFIED

STATISTICAL

TREATMENT

OF GEOCHEMICAL

DATA

545

b. a mixture of two populations a given set of in of the deviation in the estimationof the threshold; data; and two populations may havethe samebackground but, c. an excess low valuesin the considered of popunevertheless,different thresholds if their coefficients lation.

ordinate 2.28% to 2.5%. This shows importance the

of deviationare different. In Figure 2, the threshold

is five timesthe background Cu and only 2.7 for

times for Zn.

These three casesare represented graphicallyin Figures 5. They correspond real distributions to

encounetred the Guatemalandrainagesurvey and in appear as solid lines with slopebreaks on the diaplest case: a singlelognormal population, diathe grammatic expressionof which is a straight line. gram. Some indications are given below showing However, when constructing cumulative frequency how to interpret suchlines. CopperDistribution (in a lithologicalunit). The curves, a broken line is frequentlyobtainedsuggesting that the set of data consideredconsistsof a cumulativefrequencyline (Fig. 5) shows a break complexpopulationor of different ones. Whenever to a flatter slope at the 30% level. This is the possible practice,the interpretation made on case when there is an excessof high values in the in is population;the histogram will give a frequencycurve sets of data selected so as not to include more than two different distributions; for instance, a litho- skewed to the right, in the direction of the high logicalunit may include two typesof mineralization values (positive skewhess). If the populationwas showing in soil or sediment up samples; repre- lognormallydistributed,the main branch Oat should one sentativeof the normal or backgroundcontent of extend as a straight line in Oz whereas,in this case, Ox the materialsampled, the other,a superimposed is lifted to Oy whichmeansthat insteadof having and
In all the foregoing, I have consideredthe simmineralization related to ore.

2.5% of the values ppmor greater, 30 thereare 17% of them. The abscissa the breaking point, O, of Examples (in this case 18 ppm) indicatesthe limit above The three main casesof non-homogeneous dis- whichthereis a departure from the norm (i.e., from tribution that are the most likely to occur are, in the lognormaldistribution),an excess high values. of decreasing frequency order: In this case,background and coefficients deviation of
a. an excess of high values in the considered are calculated with the main branch Oat. The population; abscissa the breakingpoint may be conveniently of

546

CLAUDE LEPELTIER Advantages of Cumulative Frequency Curves

taken as threshold value if the break occurs above

instance) the thresholdshould be taken as usual

the normalthreshold level of 2.5%. If, however, Plotting the distributionof an elementin a selected the breakoccurs below2.5% level (at point p for unit as cumulativefrequencycurve on probability
graph paper is the easiest and most preciseway to

(abscissa point P). Positively of brokendistribu- presenta great amount of data (for instance,pretion linesare the moreinteresting because they in- sentingFigure 5 as histograms frequency and curves dicatean excess over the background mineralization. will result in an overloaded and illegiblediagram). Molybdenum Distribution a lithological (in unit). All the characteristic parameters the distribution of
The cumulative distribution line shows two breaks:
can be estimated without cumbersome calculations.

first a positive,then a negativeone. Such a graph Comparisonbetweenvarious populationsare easy is the expression a dual distribution, of suggesting andcomplex distributions clearlyidentified.Furare the existence two distinctpopulations the set thermore,the adjustmentto a lognormaldistribution of in of data considered.It givesa double-peaked histo- can be checked graphically. gram. We shall considerhere only the most freComparingthe geochemical featuresof the various quentcase a main "background" of population mixed units of a surveyarea is importantin assessing their with a smallerone of higher averagevalue, the two mineral potential. This is convenientlydone by of them beinglognormally distributed. On the dia- plotting the corresponding distributions the same on gram (Fig. 5), branchA corresponds the main diagram for instance Cu distribution in three or to or normal population, branch B to the anomalous four different drainagesin the case of a stream population and the central branch A q- B to a mix- sedimentreconnaissance. Distribution heterogeneiture of the two. By splitting the data at a value ties will be spotted and the correspondingunits taken aroundthe middleof A q- B (at 4 ppm for selectedfor further investigations. On a broader instance),it is possible separate total popula- scale, the geochemical to the behavior of trace elementsin tion into two elementary ones appearingas a and a given geologicalenvironmentfrom different counb on the diagram. The generalbackground will be tries or metallogenic provinces can be readily comtaken with branch A and the threshold as the abscissa pared. This is an approach a betterunderstanding to of the middleof branchA q- B, thoughthe threshold of the distribution laws of trace elements naturally in of population may alsobe considered, we have occurring a but materials. not enoughexamples suchcomplexdistributions of
to make definite recommendations, and we lacked computingfacilities to calculatetheoreticaldistributions. The coefficients of deviation must be calThe Coefficients of Deviation

culatedseparately distributions and b. for a Zinc Distribution (in a drainage unit). The negatively brokenline on Figure 5 is the expression of an excess low values an essentially of in lognormal distribution;in this case, the histogramis skewed to the left, toward theselow values (negativeskewness). Provided their proportionis not too high (20% or less or instance),they do not interfere in the interpretation,which is done on the main branchof the distribution line in the usual way. This excess low valuesmay be due to the incluof sionin the population a low-background of lithological unit or, more often, to poor sampling(for instance,collectingan important set of sedimentsamplesthat are too coarse). When the resultsdo not fit a lognormaldistribution, an explanation may generallybe found among these three factors: (1) lack of homogeneity in sampling,(2) complexgeology(imprecision the in lithological boundaries), and (3) analyticalerrors. It should alsobe kept in mind that someelements
tributed.

A lognormal distribution completely is determined by two parameters:the geometricmean (b) and the coefficient deviation (s). It has been seen of that the absolutedeviation can be expressedas a geometric factor s' or, more commonly, a logarithas mic coefficient The term "deviation"is preferred s. to "dispersion"which might be more expressive, because there is no geneticimplicationin the concept of statisticaldispersion whereasthere is one in the notion of geochemicaldispersion; however, many peopleuse the term "dispersion"in statisticalinterpretation geochemical of data. The coefficient deviationis a dispersion of index

specificfor the distribution of a given element in a given environment and expresses the degree of homogeneity this distribution. When rocks are of considered, similarityin the coefficient deviation, a of together with similar average values, may indicate similar geochemical processes their formation. in It is possible that a given value of s corresponds to each type of mineralizationin a lithologicalunit. Confirmingthis assumption would require very extensivegeological-statistical studiesencompassing all metallogenic cases. in somesurroundings not be lognormally may disThere is also a relationshipbetween the back-

ground (b) and the coefficient deviations(s) of

SIMPLIFIED STATISTICAL TREATMENT OF GEOCHEMICAL DATA

547

whichis the expression the geochemical which averageabsolute of law deviations (graphicallyestimated when the abundance the of statesthat the dispersionof an elementis inversely in Fig. 6) also decreases proportionalto its abundance. This is expressed element increases. The weightedmean valuesof b, s and s" for each very clearly by the relativd dispersion (or relas" tive deviation), a percentage related to b and s as elementhave been calculated separatelyfor Blocks
follows:
s

I and II:

s"= 100

Block I
Zn Cu

b
55. 8.

s
0.23 0.34

s"
0.42 4.2

Block II
Zn

b
70.

s
0.17

s"
0.24

Cu 8. 0.30 3.8 The higher the background,the lower the relative Pb 5.8 0.30 5.2 Pb 6.8 0.32 4.7 deviation. This is best shown on a log/log correlaMo 0.35 0.40 125. Mo 0.38 0.37 97.5 tion diagram by plotting s" as abscissaand b as ordinate. Figure 6, for instance,showsthe variation The fact that the absolute deviationfor Pb is equal of s"in functionof b in the different lithologicalunits of Blocks I and II, for Cu, Zn, Pb and Mo. The to or slightly lower than that for copper is due to

limit of the analytical diagram has been constructed taking, for each two factors: (1) the sensitivity by element, the extreme values for b and s" thus deter- methodfor lead, which entaileda numberof assumpin mining parallelograms includingall the individual tions and extrapolations the interpretation--devalues. One seesimmediatelythat there is an in- terminationof b and s, and (2) the existenceof verse linear relationshipbetween b and s" (which somePb mineralizedzonesin the surveyarea where is evident from the definition of s") and that the b was high and s low.
Figure 6. Correlationdiagramb/s"for blocksI andII
All lilholoicll units

6.2

0.3

0.4

O.i 0.8 !

,3

$ $ 78910

,=lsJ'1 o. 2o.2

548

CLAUDE LEPELTIER
Figure 7. CorrelationdiagramCu/Zn

cu (pp,.)
III III
\

Ill Ill

Ill

, /

/ '

"
Cu

x ., X ,

Ill

Ill

Ill

Ill III
,,i

III ,
.
....

- I l
/'
:-_ .

,,,
I II
tll

'.A......

Ill

III

Jill

III
III

...............III ',
;
I II

11

Ill

III

III

ill

5670910

4 5 67891

4 5 67091ffi0

N = n + n = 168 1 I 3

In Figure 6, it is also interestingto note the variationsof the dispersion the sameelementin of different lithological units which is particularly noticeable copper; the width of each parallelofor gram indicates the range of variation of s for each

is commonly high in Zn. This geologic also concept of a relationship between two-types mineralization of (only qualitative rathervague)may be substiand tutedby a precise factor,the coefficient correlaof tion p, whichgivesa rigorous measure their deof element. gree of dependency. the caseof geochemical In p the of of The coefficientof deviation is a very important prospecting,measures degree dependency character of the distribution of an element in a two lognormal variables namelythe tenorsof two in population (Matheron,1962). given surrounding; is probablyrelatedto the type elements a sample it The coefficient always p fallsbetween and + 1. -1 of geochemical dispersion, mechanical chemical, or a independence between the and consequently might give an indicationof the p--o means complete p indicates functional a relationtype of anomalyencountered: syngenetic epi- two elements, -- --+-1 or between them (it is a linear genetic. It appears that a highercoefficient devi- ship,director inverse, of between logarithms the tenors). the of ation indicates preponderantly a mechanical disper- relationship Simplified Calculation p.--There is a graphical of sion, but this has not been proved. Much remains to be done in this field. way to estimate slightlylessprecise p, but much faster than the complete statistical calculation: conCorrelation Diagrams structing correlation a cloudin full log. coordinates of under In the caseof a polymetallic mineralization, with (Fig. 7, 8). Each sample the population its two or more elements lognormally distributed, there study is plottedfollowingits tvo coordinates: is generally positive a correlation between them;for tenor in element / and its tenor in element B and appears a cloudof points. as instance between lead and zinc, a samplehigh in Pb the total population

SIMPLIFIED

STATISTICAL

TREATMENT

OF GEOCHEMICAL DATA

549

Practically, this presentation the data is very of convenientbecauseit gives a geometric image of the distribution laws. The axes passing by the gravity center (b, b)), that is to say by the point whosecoordinates the background are valuesfor the two considered elements, are then drawn. In Figure 7, the axes will passthrough the point (bc, = 5.3 ppm, bz, = 75 ppm). The points falling in each quadrantare summedup and countedas follows:

(1) eitherp is equalor near to zero: the elliptical

cloud has its axes parallel to the coordinate axes and the two variablesare independent, (2) or p is clearly different from zero and the cloud is an ellipsewhoseaxes are inclined relative to the coordinates. The slopeof the main axis has the samesign as p (if p > 0 the two elements vary in the same direction; if p < 0 the two elements vary inversely).

The correlation cloud is in fact a two dimensional N = numberof pointsin first and third quadrants histogram;it is the bestand simplest way to estabN. = numberof pointsin second fourth quadand rants. lish whethera population homogeneous heterois or geneous:in the first case,the points tend to group Then Ois givenby the formula: in a singleellipticalcloud; in the second, they split
into 2 or several attraction centers and form several

o= sin 'N+Ns 1 [r --N N

Practically,p is never equal to --1 (which would (or of the orderof +0.95) (Matheron,1962); then be the caseif all the pointswere on a straightline) it is likely that a geologically basedchemical equiand the points form an ellipticalcloud. Two cases librium exists between the two elements considered. may happen: In geochemical prospecting, correlationcoefficients
Figure8. Correletion diegram Pb/Zn
Pb (i)Pm)
1000

elliptical clouds more or less overlapping. G. Matheron pointsout that the relation expressed by p is an expression the MassAction Law if p = --+1 of

I I IIII

I II

...............

III

Ii lii-: i I111. [ IXI [ll../

'f' /

/, /

,,, Iii I I I ii

,,, Ill III ii

i ii

, /

/.

'

/I

,,,,,.

111

1111

'"'" Illl
I IIIII

III
Ill

" III
III

Ill

I llIi

III

n ]0 2
n 4

: n] n 83 3

n =45 3
6

=n +n =16 2 4

=sin

_N2

+N

550

CL,ZIUDE LEPELTIER

anomalous patterns may be usedto assess mineral associations ele- was useful in outlining subdued of mentsin natural samples. The correlation diagram in a complex geochemical surrounding,but much
showswhether two elementsare spatially associated more informationcan certainlybe extractedfrom the

and if onemay be usedas a pathfinder the other. analytical results by a more thorough, computerfor Let us consider two examples'the relationship of oriented, treatment. Cu/Zn in the drainageof the Suchiate River (Fig. The graphicalmethodsdescribed above have the 7) andthe relationship Pb/Zn in the Rio Grande great advantageof being quick, cheapand easy to of drainage(Fig. 8). use in the field without any specialmathematical The first example, Figure7, is intended in onlyto knowledge. It is a convenient and syntheticway illustrate the lack of relationshipbetween two types to present greatamount geochemical a of data,and of mineralization.The cloudof pointshasno definite I think it might be usefulto any geologist involved shape, but it can be dividedinto three zones' one in geochemical prospecting. aroundthe intersection point of the axes,including UNITED NATIONS MINERAL SURVEY, the majority of the pointswhich are spreadmore

or less equally among fourquadrants; elliptical the an one,markedCu, in the rangehigh-Cu/backgroundZn values; and a third one, includingonly a fev

GUATEMALA CITY, GUATEMALA,

January20; March 28,1969

REFERENCES

high-Zn/background-Cu points. Thisshows that,in

soever between the Cu and Zn mineralization, that

of the Suchiate drainage, thereis no relationship what- Ahrens, L. H., 1957, The lognormaldistribution the

the Cu anomaly moreimportant is than that for Zn

and that the two anomalies are well separated
correlation'

elements--a fundamental law of geochemistry: Geochim. et Cosmochim. Acta, v. 11, no. 4.

Coulomb, 1959,Contribution la Cochimie l'uranium R., 3, de dans les granitesintrusifs: Rapport C.E.A. 1173, Centre
minerals with special reference to the Witwatersrand
Gold Reefs: Geol. Soc. South Africa Trans. v. LIX.

d'Etudes Nucleires de Saclay, France. spatially.All thisis expressed the coefficient Cousins,C. A., 1956, The value distributionof economic by of

Hubaux,A., 1961,Representation graphique distributions des Ann. Soc.Gol. Belgique, LXXXIV-T. Its low absolute value indicates nearly complete d'oligo-lments: a Mars 1961.

p = -0.11

independence the two mineralizations, a Termant,C. B., and White, M. L., 1959,Study of the disof with data: ECON.GEOL., 54, V. tendency' inverse to relationship (negative value). tribution of somegeochemical p. 1281--1290. On the contrary, Figure8 shows example Matheron,G., 1962,Trait de gostatistique an of applique, tome directrelationship between types mineraliza- 1: Mmoire no. 14 du Bureau de RecherchesGologiques two of
et MiniSres, Paris. tion. In the Rio Grande drainage,Pb and Zn are for associated' the correlation cloud is an elongated Miesh, A. T., 1967,Methodsof computation estimating

ellipse whose main axis has a 45 slope and the

correlation coefficient +0.87. In this drainage, t--lead and zinc anomalies will have the same pattern

geochemical abundance--U. S. Geological Survey ProfessionalPaper 574-B. Monjallon, A., 1963, Introduction3. la mthode statistique:
Vuibert, Paris.

andwill be spatially related. In similar geological conditions, element one maybe usedas a pathfinder
for the other.

Rodionov, A., 1965,Distributionfunctions the elements D. of

and mineral contentsof igneousrocks: ConsultantBureau,
New York.

Shaw, D. M., 1964,Interpretation gochimique lments des

en trace dans les roches cristallines: Masson et Cie,
Paris.

Conclusion

In the Guatemalan geochemical reconnaissance, the statistical analysis the data,although of elementary,

Vistelius,A. B., 1960,The skew frequency distributions and fundamentallaw of the geochemical processes: Journal of
Geol. Jan. 1960.