]\gt ; INDEX PRARSON , Karl .
\mdash ; Mathematical Contributions to the Theory of Evolution.\mdash ; VII.\mdash ; On the Correlation of Charactel .
not Quantitatively Measurable .
Phil. Trans. , , pp. 1-47 .
oxin Treatment in Effectiveness Q Measured .
PEARSON , Karl .
Phil. Trans. , , vol. , 1900 , pp. 1-47 .
Aseociation of Attributes .
PEABSON , Karl . .
Phil. Trans. , , vol. 196 , , pp. 1-47 .
Correlation , Determination of , when Variables are not Quantitatively Measurable .
PBABSON , Karl .
Phil. , vol. 196 , .
1900 , pp. 1-47 .
Errors , Normal Iheory of\mdash ; Extension to Cases of Variables not Capable of Quantitative Measurementi .
PEARSON , Karl .
Phil. Trans. , , vol. 195 , 1900 , pp. 1-47 .
Heredity--Rate of Production of Exct.ptional Offspring by ExceptiontJ tock .
PeABSON , Karl .
Phil. Trans. , , vol. 195 , 1900 , pp. 1-47 .
Vaccination , Effectiveness of , Quantitatively Measured .
PEARSON , Kart .
Phil. Trans. , , vol. 195 , 1900 , pp. 1-47 .
PHILOSOPHICAL I. Mathematical Contributions to the Theory of .
On the Cor relation of Characters not tively KARL PEARSON , F. R. rtment Applied JIathematics , University College , Received February \mdash ; Read March 1 , 1900 .
NOTE .
In August , 1899 , I presented a memoir to the Royal Society on the inheritance of coat-colour in the horse and of eye-colour in man , which was read November , 1899 , and ultimately ordered } ) published in the ' Phil. Trans. ' Before that memoir was printed , valuable memoir on Association was read , and , further , Mr. LESLIE -MOORE showed me that the theory of my memoir as given in S6 of the present memoir led to somcwhat divergent results according to the methods of proportioning adopted .
We therefore undertook a new investigation of the theory of the whole subject , which is embodied in the present memoir .
The data involved in the paper on coat-colour in horses eye-colour in man all been recalculated , and that paper is nearly ready for presentation .
But it seemed best to the purely theoretical considerations from their application to special cases of inheritance , and accordingly the memoir now reappears in two sections .
The theory discussed in this paper was , further , the basis of a paper on the Law of Reversion with special reference to the Inheritance of Coat-colour in Basset recently communicated to the Society , and about to appear in the 'Proceedings .
' While I am responsible for the general of the present paper , the rough draft of it taken up and carried on in leisure moments by Mr. LESLIE .
L. N. G. and Miss ALICE LEE , D.Sc .
Mr. BRAMLEY-MOORE discovered the -fumctions ; Mr. FlLON proved most of their general properties and the conwergency of the series I alone am responsible for sections 4 , 5 , and 6 .
Mr. LESLIE BRAMLEY-MOORE sent me , without proof , on the eve of his departure for the Cape , the general expansion for on p. 26 .
I am responsible for the present proof and its applications .
To ALICE LEE we owe most of the illustrations and the table on p. 17 .
Thns the work is essentially a joint memoir in which we have equal part , and the use of the first personal pronoun is due to the fact that the material had to be put together and thrown into form by one of our number.\mdash ; K. P. CONTENTS .
PAGE S1 .
On a General Theorem in Normal Correlation for two Variables .
Series to Determine the Correlation . . . . . . . . . . . . . . . . . . .
2 S2 .
Other Series for the Determination of the Correlation . . . . . . . . . .
7 * Since ordered to be printed in the ' Phil. Trans. ' Read January 25 , 1900 .
'Boy .
Soc. Proc vol. 66 , p. 140 .
VOL. CXCV.\mdash ; A262 .
16.8.1900 .
S3 .
Proof of the General Convergency of the Series for the Correlation . . . . .
10 S4 .
On the Probable Error of the Correlation thus Determined . . . . .
10 S5 .
To Determine a Physical Meaning for the Sories and on Divers Measures of Association . . .
14 S6 .
On the " " Excess\ldquo ; and its Relation to Correlation and Relative Variability . . . . .
18 87 .
On a Generalisation of the Fundamental Theorem of the present Memoir .
Special Formulae for Triple and Quadruple Correlation . . . . .
23 S8 .
Illustrations of the Methods of the Memoir . . . . .
35 Illustration I. Inheritance of Coat-colour in Thoroughbred Horses .
Sire and Filly . . . . .
35 II .
Chance that an Exceptional bIan is born of an Exceptional Father . . . .
37 III .
Inheritance of Coat-colour in Dogs , Half- " " Siblings " " . . . .
38 IV .
Inheritance of Eye-colour between Maternal Grnndmother and Granddaughter 39 V. Inheritance of Stature between Father and Son for different groupings . . .
40 VI .
Correlation between Strengbh to resist Small-pox and Degree of Effective Vaccination 43 VII .
Effect of Antitoxin on Diphtheria Mortality . . . . . . . . .
44 VIII .
Chance of Stock above the Average giving Produce above the Average as compared with the chance of such Produce from Stock below the Average . . . .
45 , , IX .
Chance of an Exceptional Man being born of Exceptional Parents . . . .
46 S ( 1 .
) On General Theorem in Normal Correlation .
Let the frequency surface where total number of observations , standard deviations of organs and correlation of and be divided into four parts by two planes at right angles to the axes of and at distances and from the origin .
The total volumes or frequencies in these parts will be represented by , and in the manner indicated in the accompanymg plan : Then clearly , if and TO THE THEORY OF EVOLUTION .
Further , .
, and ( iii .
) , ( iv .
) , . .
Thus , when , and are known , and can be found by the ordinary table of the probability integral , say that of Mr. SHEPPARD ( ' Phil. Trans vol. .
p. 167 , Table VI .
* ) .
The limits accordingly of the integral for in are known .
Now consider the expression , say , ( vi .
) , and let us expand it in powers of .
Then , if the expansion be . .
, we shall have . . . . .
( viii .
) .
Taking logarithmic differentiak , we get at once Differentiating times by LEIBNITZ'S theorem , and putting , we have , after some reductions ( ix .
) .
Hence we find Sqe , however , foot-note , p. 5 .
Thus the foUowing laws are indicated:\mdash ; . . . . .
( xi .
) , where . . . . .
( xii .
) , . . . . .
( xiii .
) .
We shall now show that these laws hold good by induction .
Assume .
Thus But by ( ix .
) , substituting for from ( xi .
) and ( xiii .
) , .
, as we have seen above .
Thus , if the theorem holds for , it holds for .
Accordingly . . .
( xiv .
) , where the 's and 's are given by , ( xii .
) , and ( xiii .
) .
It is thus clear that consists of a series of which the general term is where It remains to find these integrals .
The general form of is given by \mdash ; \amp ; c. ( xv .
) .
TO THE OF EVOLUTION .
/ For this obviously gives .
Assume it true for and .
then . . .
\mdash ; . . . . . .
Thus the expression ( xv .
) is shown to hold by induction , the general terms being or the general term in We notice at once that . . . . .
( xvi .
) .
Thus , by ( xii .
) Multiply by and integrate Integrating the latter integral by parts , we have or Now can be found from any table of the ordinates of the normal curve , , Mr. SHEPPABD'S , ' Phil. Trans , vol. 192 , p. 153 , Table I. * We shall accordingly put . . . .
( xvii .
) , and look upon and as known quantities .
* For our present purposes the differsnces of Mr. SHEPPARD'S tables are occnsionnlly too large , but the following series givo very close results :\mdash ; Let by ( iv .
) , ' by .
Further , let us write as , and similarly as .
Thus .
( xviii .
) .
We have then from HK by ( ii .
) and ( iii .
) .
Or , remembering that , we can write this \amp ; c. .
( xix .
) .
Then . .
and . . . . .
These follow from the considerations that if whence it is easy to find the successive differentials of with regard to and with regard to , and then the above results by MACLAURlN'S theorem .
There is , of course , no difficulty in calculating and from ( xvii .
) dircctly .
That method was adopted in the numerical mustrations .
TO THE THEORY OF EyOLUTION .
Here the left-hand side is known , and since and are known , we can find the coefficients of any number of powers of so soon as the first two have been found , from ( xii .
) and ( xiii .
) .
Accordingly the correlation can be found if we have only made a grouping of our frequencies into the four divisions , , and If and be zero , we have from ( xvii .
) and The right-hand side of ( xix .
) is now or equal to Hence bc ) .
( xx .
) , which agrees with a result ofMr .
SHEPPARD'S , 'Phil .
Trans , vol. 192 , p. 141 .
We have accordingly reached a generalised form of his result for any class-index what Clearly , also , being known , we can at once calculate the frequency of pairs of organs with deviations as great as or greater than and S ( 2 .
) Other Series for the Determination of For many purposes the series ( xix .
) is sufficiently convergent to give for given and with but few approximations , but we will now turn to other We have by ( vii .
) . . . . . .
Put , and write for brevity .
( xxi .
) .
It follows at once from ( xix .
) that .
( xxii .
) , if Nou either of the quantities under the of integration in ( xxii .
) can be expanded in powers of by MACLAURIN'S theorem .
Thus let . . . . . .
Then . . . . .
, and it remains to find Now Hence \mdash ; Differentiating times by LEIBNITZ'S theorem , and putting . . . . . . . . . . .
( xxiii .
) .
Clearly , then we rapidly find ' ' Or , finally ( xxiv .
) , where more terms if required can be found by ( xxiii .
) .
If be fairly small , will be ligible .
Or if and be small , the lowest term in the next factor will be TO THE THEORY OF EVOLUTION .
and this into is generally quite insensible .
Very often two or three terms on the right-hand side of ( xxiv .
) give quite close enough values of , and accordingly of .
( xxiv .
) is clearly somewhat more convergent than ( xix .
) if and are , as usually happens , less than unity .
Returning now to , let us write it .
This is the equation that must be solved for .
Suppose a root of this when we retain only few terms on the right , say a root of the quadratic Then if Hence to a third approximation - nearly ( xxv .
) , which gives us a value of which , substituted in in the aboye equation , introduces only terms of the 6th order in Another integral expression for of Equation ( xxi .
) may here be noticed : Put , .
Hence Let , or , Therefore where and is It seems possible that interesting developments for might be deduced from this integral expression .
VOL. CXCV .
\mdash ; A. S ( 3 .
) To show that the Series for .
is Convergent be the Values of Write the series in the form of p. 6 , i.e. :\mdash ; Now by ( xii .
) and ( xiii .
) .
From these we deduce Now let Then we find Thus , when is large , we find the ratio of successive terms or is given by , where or , The ultimate ratio of to is accordingly given by , but this is the ratio of alternate terms of the original series .
The original series thus breaks up into two series , one of odd and one of even powers of .
Both these series are absolutely convergent whatever and be , having an ultimate convergence ratio of S ( 4 .
) To Jind the Error of the Correlation Coefficient Determined by the Method of this Memoir .
Given a division of the total frequency into groups , where , then the probable error of any one of them , say , is where* . . . . . . . .
( xxvi .
) .
Let , then * The standal.d deviation of an event which happens times and fails times in trials is well known to be .
The probable errors here dealt with are throughout , of course , those arising from different samples of the same general population TO THE THEORY OF EVOLUTION .
( xxvii .
) .
To obtain we have , if S denotes an error in any quantity . . . . .
( xxviii .
) , by squaring , summing for all possible variations in and , and dividing by the total number of variations .
Hence , substituting the values of the standard deyiations as found above , we deduce . . . . .
( xxix .
) .
In a similar manner Sn l@d . . . . .
( xxx .
) .
and . . . . .
( xxxi .
) , Now NHSh .
Thus NH . . . . . . . .
( xxxii .
) , and similarly . . . . .
( xxxiii .
) .
Hence the probable error of .
( xxxiv .
) , and of ( xxxv .
) .
They can be found at once , therefore , when and have been found from an ordinate table of the exponentia !
curve , and are given .
We have thus the probable error of the means as found from any double grouping of observations .
Next , noting that S S , we have PROFESSOR K. PEARSON ON IsIATHEMATICAL CONTRIBUTIONS But @n , . . .
( xxxvi .
) , therefore ( xxxvii .
) . .
( xxxviii .
) .
This is an important result ; it expresses the correlation between errors in the position of the means of the two characters under consideration .
But if the probabilities were independent there could be no such correlation .
Thus might be taken as a measure of divergence from independent variation .
We shall return to this point later .
Since HNSh , we have \mdash ; HNSdSh , whence we easily deduce . . . . .
( xxxix .
) .
jlarly . . . . .
( x1 .
) .
Now is a function of , and .
Hence if , @ S S Sk . . . . . . .
( xli .
) .
Whence transposing , squaring , summing , and dividing by the total number of observations , we find ( xlii .
) .
Substituting the values of the standard deviations and correlations as found above , we have ( xliii .
) .
TO THE THEORY OF EyOLUTION .
It remains now to determine 70 , , and By Equation ( xliv .
) , where Thus . .
, . . .
Similarly \mdash ; . . . . .
( xlvi .
) .
Here where ( xlvii .
) , .
, and thus and can be found at once from the tables when and are found from the known values of Lastly , we have from Equation ( xxi .
) or Udr .
Thus* NU , where ) ( xlix .
) a value which can again be found as soon as are known .
is clearly the ordinate of the frequency surface corresponding to Substituting in Equation ( xliii .
) we have , after some reductions , * By Equations ( ii .
) and ( iii .
) , and are independent of Probable error of . . . . .
( 1 .
) , where , and are readily f*ound from Equations ( xlix .
) , ( xlvii .
) , and ( xlviii .
) .
Thus the probable error of can be fairly readily found .
It must be noted in using this formula , that is the quadrant in which the mean falls , so that and are both positive ( see fig. , p. 2 ) .
In other words , we have supposed and .
Our lettering must always be arranged so as to suit this result before we apply the above formula .
S ( 5 .
) To Find Physical for the Series in , orfor the of ( xxi .
) .
Return to the original distribution of p. 2 .
If the probabilities of the two characters or organs were quite independent , we should expect the distribution Now re-arranging our actual data we may put it thus : Accordingly correlation denotes that has been transferred from each of the second and fourth compartments , and the same amount added to each of the first and third compartments .
If , then is the transfer per unit of the frequency .
The magnitude of this transfer is clearly a measure of the divergence of the statistics from independent variation .
It is physically quite as significant as the correlation coefficient itself , and of course much easier to determine .
It must vamsh with the correlation coefficient .
We see from ( xxi .
) that or we have an interpretation for the series in of ( xix .
) .
Now , obviously any function of , just like itself , would serve as a measure the divergence from perfectly independent variation .
It is convenient to choose a function which shall lie arithmetically between and 1 .
TO THE THEORy OF EyOLUTION .
Now consider what happens in the case of perfect correlation , i.e. , all the observations fall into a straight line .
Hence if , either or is zero , for a straight line cannot cut all four compartments , and and are obviously positive .
Thus and can only be zero if or .
In order that should be zero , it is needful that and , as given by ( iv .
) and , should be positive or , and the mean fall under the line through the vertical and horizontal lines dividing the table into four compartments , i.e. , .
These conditions would be satisfied if and .
Now suppose our four-compartment table arranged so that and consider the function , or .
This function vanishes if , and it further unity if .
Thus it agrees at the limits and 1 with the value of the correlation coefficient .
Again , when and are both zero , , and , is thus by ( xx .
) .
Hence we have found a function which vanishes with .
and equals unity with , while it is also equal to if the divisions of the table be taken through the medians .
Now , I take it that these are very good conditions to make for any function or which is to vanish with the transfer and to serve as a measure of the degree of dependent variability , or what Mr. YULE has termed the degree of ' ' association Mr. YULE has selected for his coefficient of association the expression ( liii .
) .
This vanishes with the transfer , equals unity if or be zero , and minus unity if or be zero .
The latter is , of course , unnecessary if we agree to arrange so that is always greater than .
Now it is clear that possesses a great advantage over in rapidity of calculation , but the coefficient of correlation is also a coefficient which measures the association , and it is a great advantage to select one which agrees to the closest extent with the correlation , for then it enables us to determine other important features of the system .
If we do not make all the above conditions , we easily obtain a number of coefficients which would vanish with the transfer .
Thus for example the correlation of Equation ( xxxviii .
) is such an expression .
* It has the advantage of a symmetrical form , and has a concise physical meaning .
It does not , however , become unity when * In fact ( xxxvii .
) gives us either , but not both , and vanish , nor does it , unless we multiply it by and take its sine , equal the coefficient of correlation when and Again , we might deduce a fairly simple approximation to the coefficient of correlation from the Equation ( xxiv .
) for , using only its fir.st few terms .
Thus we find Sin , where as an expression which vanishes with the transfer , and will be fairly close to the coefficient of correlation .
It is not , however , exactly unity when either or is zero .
But without entering into a discussion of such expressions , we can write several down which fully satisfy the three conditions:\mdash ; Vanishing with the transfer .
( ii .
) Being equal to unity if or ( iii .
) Being equal to the correlation for median divisions .
Such are , for example:\mdash ; , , ( lvi .
) , ( lvii .
) , where Only by actual examination of the numerical results has it seemed possible to pick out the most efficient of these coefficients .
was found of little service .
The following table gives the values of , and in the case of series selected to cover a fairly wide range of values:\mdash ; TO THE THEORY OF EyOLUTION .
Now an examination of this table shows that notwithstanding the extreme elegance and simplicity of Mr. YULE'S coefficient of association , the coefficients , and , which satisfy also his requirements , are much nearer to the values assumed by the correlation .
I take this to be such great gain that it more than counterbalances the somewhat greater labour of calculation .
If we except cases ( 6 ) and ( 10 ) , in which or take a large value exceeding unity , we find that , and in the fifteen cases hardly differ by as much as the probable error from the value of the correlation .
If we take the mean percentage error of the difference between the correlation and these coefficients , we find Mean difference of per cent. , , , , , , Thus although there is not much to choose between and , we can take as a good measure of the degree of independent variation .
The reader may ask : Why is it needful to seek for such a measure ?
Why cannot we always use the correlation as determined by the method of this paper ?
The answer is twofold .
We want first to save the labour of calculating for cases where the data are comparatively poor , and so reachin a fairly approximate result rapidly .
But labour-saving is never a wholly satisfactory excuse for adopting an inferior method .
The second and chief reason for seeking such a coefficient as lies in the fact that all our reasoning in this paper is based upon the normality of the equency .
We require to free ourselves from this assumption if possible , for the difficulty , as is exemplified in Illustration V. below , is to find material which actually obeys within the probable errors any such law .
Now , by considering the coefficient of regression , , as the slope of the line which best fits the series VOL CXCV.\mdash ; A. of points determined as the means of arrays of .
for given values of , we have once and for all freed ourselves from the difficulties attendant upon assuming normal frequency .
We become indifferent to the deviations from that law , merely observing how closely or not our means of arrays fall on a line .
When we are not given arrays but gross grouping under certain diyisions , we have seen that the " " transfer\ldquo ; is also a physical quantity of a significance independent of normality .
We want accordingly to take a function which vanishes with the transfer , and does not diverge widely from the correlation in cases that we can test .
Here the correlation is not taken as something peculiar to normal distributions , but something significant for all distributions whatever .
Such a function of a suitable kind appears to be given by S6 .
On the " " Excess\ldquo ; and its Relation to Correlation and Relative Variability .
There is another method of dealing with the correlation of characters for which we cannot directly discover a quantitative scale which deserves consideration .
It is capable of fairly wide application , but , unlike the methods previously discussed , it requires the data to be collected in a special manner .
It has the advantage of not applying only to the normal surface of frequency , but to any surface which can be converted into a surface of revolution by a slide and two stretches .
It is well known that not only the normal curve but the normal surface has a type form from which all others can be deduced by stretching or stretching and sliding .
Thus in 1895 the Cambridge Instrument Company made for the instrument room at University College , London , " " biprojector an instrument for giving arbitrary stretches in two directions at right angles to any curve .
In this manner by the use of type-templates we were able to draw a variety of curves with arbitrary parameters , e.g. , all ellipees from one circle , parabolas from one parabola , normal curves ffom one normal curve template .
Somewhat later Mr. G. U. YULE commenced a model of a normal frequency surface on the BRILL system of interlaced curves .
This , by the variable amount of slide given to its two rectangular systems of normal curves , illustrated the changes from zero to perfect correlation .
This model was exhibited at a College soire'e in June , 1897 .
Geometrically this property has been taken by Mr. W. F. SHEPPARD as the basis of his valuable paper on correlation the 'Phil .
Trans , vol. 192 , pp. 101-167 .
It is a slight addition to , and modification of , his results that I propose to consider in this section .
The equation to the normal frequency surface is , as we have seen in S1 , expt .
\mdash ; Now write .
This is merely giving the surface two uniform stretches ( or squeezes ) parallel to the coordinate axes .
We have for the frequency of pairs lying between , and , S TO THE THEORY OF EVOLUTION .
expt .
\mdash ; Now give the area a uniform slide parallel to the axis of defined by at unit distance from that axis .
This will not change the basal unit of area , and analytically we may write Whence we find zSxSy expt .
) .
This is the mechanical changing of the -BRILL model analytically represented .
The surface is now one of revolution , and the proof would have been precisely the same if we had written in the above results any function , instead of the exponential .
* It is easy to see that any volume cut off by two planes through the axis of the surface is to the whole volume as the angle between the two planes is to four right angles .
Further the corresponding volumes of this surface and the original surface are to each other as unity to the product of the two stretches .
Lastly , any plane through the -axis of the original solid remains a plane through the -axis after the two stretches and the slide .
These points have all been dealt with by Mr. SHEPPARD ( p. 101 et seq. , .
cit I will here adopt his notation , and term with him the dirergence .
Thus is ( in the language of the theory of strain ) the slide , and is the angle between the strained positions of the original and directions .
Now consider any plane which makes an angle with the plane of before strain .
Then , since the contour lines of the correlation surface are ellipses , the volumes of the surface upon the like shaded opposite angles of the plan diagram below will be equal ; and if they be and , then .
If and be the volumes after strain , then by what precedes we shall have and *The generalisation is not so great as might at first appeitr , for I have convinced myseIf that this property of conversion into a surface of revolution by stretches and slides does not hold for actual of markedly skew correlation .
Now and will be as the angles between the strained positions of the planes bounding and .
Ox does not change its direction .
Oy is turned through an angle clockwise , and becomes say .
Hence or Let us write and term it the excess for the -character for the line AB .
Then we easily find : ( lviii .
) .
It remains to determine and substitute .
The stretches alter into , such that Further , by the slide D. Hence we have by ( lviii .
) above , or , . . . . .
( lix .
) .
Now the excess is the difference of the frequencies in the sum of the strips of the volume made by planes parallel to the plane on the two sides of the plane ABz ( defined by ) , taken without regard to sign .
For on one side of the mean this is , and on the other .
Hence we have this definition of , the column excess for any line through the mean of a correlation table : Add up the frequencies above below the line in column and take their differences tuithout to sign , their sum is the column excess .
If we are dealing with an actual correlation table and not with a method of collecting statistics , then care must be taken to properly proportion the frequencies in the column in which the mean occurs , and also in the groups which are crossed by the line .
It is the difficulty of doing this satisfactorily , especially if the grouping , as in eye and coat colour , is large and somewhat rough , that hinders the effective use of the method , if the statistics have not been collected hoc .
Now let be the row excess for the line AB , defined in like manner , then we have in the same way TO THE THEORy OF EVOLUTION .
( lix .
Now eliminate between ( lix .
) and ( lix.bis ) ; then Whence we deduce and , therefore , .
Substituting for in ( lix .
) we find further ( lxi .
) .
Thus Equations and ( lxi .
) give the coefficient of lation and the relative variability of the two characters .
The latter is , I believe , quite new , the former novel in form .
If we call the frequency in the angle ( AOx of the figure above ) , then it is easy to see that , and similarly .
Thus .
But is the frequency in the first quadrant .
This Mr. SHEPPARD terms , while that the second he terms R. We have thus , or ( lxii .
) , i. e. , Mr. SHEPPARD'S fundamental Phil. Trans .
We can , of course , get Mr. SHEPPARD'S result directly if we put , wh.en we have at once , , and the result follows .
Equation ( lxi .
) may also be written in the form ( lxiii .
) .
If we put , then becomes zero , and the right-hand side of ( lxiii .
) is indeterminate .
If we proceed , however , to the limit by evaluating the frequency in an indefinitely thin wedge of angle , we reach merely the identity Hence there is no result corresponding to ( lxi .
) to be obtained by taking Mr. SHEPPARD'S case of The following are the values of the probable errors of the quantities involved : * In the actual classification of data and ( lxii .
suggest quite different processes .
We can where ( lxii .
is difficult or impossible , e.g. , correlation in shading of birds ' eggs from the same clutch .
Probable error of ( lxiv .
) . . . . .
( lxv .
) .
, Correlation between errors in and .
( lxvi .
) .
Probable error in ( lxvii .
) , where SHEPPARD , .
Probable error in ratio .
( lxviii .
) .
The application of the method here discussed to statistics without quantitative scale can now be indicated .
If the characters we are dealing with have the same scale , although it be unknown , then , if the quantitative order be maintained , i. e. , individuals arranged in order of lightness or darkness of coat or eye-colour , the diagonal line on the table at will remain unchanged , however we may suppose parts of the scale to be distorted , for the distortion will be the same at corresponding points of both axes .
Further , if we suppose the mean of the two characters to be the same , this line will pass through that mean , and will serve for the line AB of the above investigation .
In this case we must take , and consequently ( lxi .
) becomes , ( lxix .
) .
We can even , when the mean is a considerable way off the line , get , in some cases , good results .
Thus , the correlation in stature of husband and wife worked out by the ordinary product moment process is 2872 .
But in this case , and this gives the correlation2994 .
the other hand , the actual ratio of variabilities is , while ( lxix .
) makes it This arises from the fact that the errors in and , due to the mean being off the line , tend to cancel in , but tend in directly opposite directions in the ratio of the cosines .
Similarly the correlation between father and son works out , which may be compared with the values given in Illustration V. below , ranging from 5198 to 5939 .
Again , correlation in eye-colour between husband and wife came out by the excess process 0986 , and by the process given earlier in the present Memoir .
But all these are vourable examples , and many others gave much worse results .
We ought really only to apply it to find when the means are on the line , as in the correlation of the TO THE THEORY OF EVOLUTION .
same character in brethren , and even in this case the statistics ought to be collected hoc , i.e. , we ought to make a very full quantitative order , and then notice for each individual case the number above and below the type .
For example , suppose we had a diagram of some twenty-five to thirty eye tints in order ( e.g. , like BERTRAND'S ) , then we take any individual , note his tint , and observe how many relatives of a particular class\mdash ; brethren or cousins , have lighter and how many darker eyes ; the difference of the two would be the excess for this individual .
The same plan would be possible with horses ' coat-colour and other characters .
After trying the plan of the excesses on the data at my disposal for horses ' coat-colour and human eye- colour ( which were not collected ad hoc ) , I abandoned it for the earlier method of this Memoir ; for , the classification being in large groups , the proportioning of the excess ( as well as the differences in the means ) introduced too great errors for such investigations .
S 7 .
On Generalisation of Fundamental Theorem of the Present Memoir .
If we measure deviations in units of standard deviations , we may take for the equation to the correlation surface for variables ( lxx .
) , where and is the minor obtained by striking out the row and qth column .
is , of course , the correlation between the and variables , and equals denotes a summation for from 1 to , and a summation of every possible pair out of the quantities 1 to Now take the logarithmic differential of with regard to .
We find For and , generally , whether is or is not , or these are or are not and , we have . . . . .
( lxxi .
) .
This follows thus : ' or we have to show that where is the minor corresponding to the term in , and the minor corresponding to the term But this last result is obvious because , only contains in two places , i. e. , as and Putting , we have the other identity required above , i.e. , Returning now to the value for on the previous page , we see that the two sum terms may be expressed as a product , or we may put .
Now write Then , and Hence Now differentiate with regard to .
Then See also Scorr , ' Theory of Determinants , ' p. 59 .
TO THE THEORY OF UTION .
Thus finally .
( lxxii .
) .
In other words , the operator acting on can always be replaced by the operator Let denote the effect of applying the operator to , and putting zero after all differentiations have been performed , then the effect of this operator will be the same as if we used on , putting zero before differentiation .
Generally , let be any series of operations like , then we see that .
let ) function which gives the operation of expanding by theorem in powers of the correlation coefficients , i. e. , then This is the generalised of result reached ) Now let then is the ordinate of a frequency surface of the nth order , in which the distribution of the variables is absolutely independent .
We have accordingly the extremely interesting geometrical interpretation that the operator ' applied to a surface of frequency for independent variables converts it into a of frequency for dependent variables , the correlation between the and s'th variables being * I should like to suggest to the pure mathematician the interest which a stndy of such tors onld have , and in particular of the generalised form of projection hyperspace ted by then .
VOL. CXCV.\mdash ; A. Expanding , we have . . . .
Our next stage is to evaluate the operation Let us put and function of as defined by .
Let be a symbol such that represents .
Then we shall show that ( lxxiv .
) .
We shall prove this by induction .
By ( xii .
) or and by ( xvi .
) , or Now , let be any function of , if we suppose it can be expanded in powers of Then . . . . . . .
( lxxv .
) .
Similarly ( lxxvi .
) .
Now suppose that TO THE THEORY OF EVOLUTION .
then where stands for Hence , remembering that which had to be proved .
But it is easy to show by simple differentiation that . . . . .
( lxxvii .
) .
Hence the theorem is generally true .
Thus we conclude that . . . . . . .
( lxxviii .
) .
It is quite straightforward , if laborious , to write down the expansion for any number of variables .
Now let be the total frequency of complices of variables with lying between and between and , . .
between and , . . .
between and ; and let be the frequency of such complices if there were no correlations .
Then . . . . . . . . .
Now let where ' ( lxxix .
) .
We have But by ( xviii .
) where ( lxxx .
) , and as above , . . . . .
( lxxxi .
) .
Thus . . . . . . . . . . . . . . . . . . . . . . . . . .
or . .
( lxxxii .
) .
where denotes a product of for any number of 's with any and .
The rule , therefore , is very simple .
We must expand the value of in 's as given by ( lxxviii .
) above , then the multiple integral of this will be obtained by lowering every 's righthand subscript by unity ( remembering that ) , and further dividing by the the left-hand subscript .
The general expression up to terms of the fourth order has been written down ; it involves thirty-four sums , each represented by a type term All these would only occur in the case of the correlation of eight organs , or when we have to deal with twenty-eight coefficients of correlation .
Such a number seems beyond our present power of arithmetical manipulation , so that I have not printed the general expressions .
At the same time , the theory of multiple correlation is of such great importance for problems of evolution , in which over and over again we have three or four correlated characters to deal with , *that it seems desirable to place on record the expansion for these cases .
I give four variables up to the fourth and three variables up to the fifth order terms .
Afterwards I will consider special cases .
*In my memoir on Prehistoric Stature I have dealt with five correlated organs , i.e. , ten coefficients .
In some barometric investigations now in hand we propose to deal with at least fifteen coefficients , while Mr. in the correlation of parts of the skeleton , has , in a memoir not yet published , dealt with between forty and fifty cases of four variables or six coefficients .
TO THE THEORy OF EVOLUTION .
Value of the Quadruple in the Case of Four * To simplify the notation , have been used for ' / \fnof ; m TO THE THEORY OF ]VOLUTION .
.31 ' ' ' ' .
( lxxxiii .
) .
In the case of three variables , we must cancel in the above expression all terms involving .
Thus we shall have 3 instead of 6 first order terms , 6 instead of 21 second ordel terms , 10 instead of third order terms , and 15 instead of 126 fourth order terms\mdash ; a much more manageable series .
I give below the extra term necessary for ting the value of as far as the fifth order terms in the case of three variables .
Fifth Order Terms for Three riables .
\ldquo ; ' .
( lxxxiv .
) .
TO THE THEORy OF EVOLUTION .
A numerical illustration of these formulae will be given in the latter part of this Memoir .
It will , however , be clear that what we want are tables of log } , including or for a series of values of .
Such tables would render the computation of fairly direct and rapid ; they could be fairly easily from existing tables for the ordinate and area of the normal curve , and I hope later to find some one willing to undertake them .
Meanwhile let us look at special cases .
In the first place , suppose , in the case of three variables , that the division of the groups is taken at the mean , i.e. , .
Then we have Hence we have . . .
( lxxxv .
) .
Let , and let be the spherical excess of the spherical triangle whose angles are the divergences .
Then we have E. Or : ( lxxxvi .
) .
Now take the case of four variables .
Here we have / \fnof ; m and all the odd v 's zero .
Hence VOL CXCV.\mdash ; A. * . . . . . . . . . . . . . . . .
( lxxxvii .
) .
This is the correct value including terms of the fourth order , but to this order of approximation we can throw it into a much simpler form .
Let , then ( lxxxviii .
) .
We may write this ( lxxxix .
) where The expressions and of ( lxxxvi .
) and ( lxxxix .
) are of considerable interest , for they enable us to express the area of a spherical triangle in three-dimensioned space , TO THE THEORY EVOLUTION .
and ( up to the above degree of approximation ) the volume of a " " tetrahedron\ldquo ; on a " " sphere\ldquo ; in hyperspace of four dimensions .
In fact , the whole theory of hyperspace " " spherical trigonometry\ldquo ; needs investigation in relation to the properties of multiple correlation .
In our illustrations ( viii .
) and ( ix .
) will be found examples of the above formulae applied to important cases in triple and quadruple correlation in the theory of heredity .
I consider that the formulae above given will cover numerous novel applications , for many of which greater simplicity will be introduced to the choice of special values for the 's or for the correlation coefficients .
( 8 .
) Illustrations of the New I. Inheritance of Coat-colour in Horses.\mdash ; The following represents the ibution of sires and fillies in 1050 cases of thoroughbred racehorses , the grouping being made into all coat-colour classed as " " bay and darker " " chesnut and lighter\ldquo ; :\mdash ; Then we require the correlation between sire and filly in the matter of coat-colour , and also the probable error of its determination .
We have from and Hence from the probability integral tables We have then : HK by ( xvii .
) , Thence ( xxi .
) .
Calculating out the coefficients of the series in in ( xix .
) we find 619 , ectino .
powers of above the second , we find by solving the quadratic and taking the ) ositive root Solving by two pproximations the sextic we finally determine , I , to four places of figures .
now to the probable error as givcn by Equation ( 1 .
) , I find ( xlix .
) : Henco ( xlvii .
) and ( xlyiii .
) we find by means of the probability integral table By substituting in ( 1 .
) , we find error of From ( xxxiv .
) and ( xxxv .
) we find .
of Thus , ally , we may sum up our results p.e. of The of this , if we had been able to find it from the product lnomont , would have be , or only about one-half its present value .
TO THE THEORY OF EyOLUTION .
.\mdash ; Our analysis opens a large field suggested by the following problem :is the chance an man is of exceptional father ?
Of course much depends on how we define " " exceptional and any numerical measure of it must be quite arbitrary .
As an illustration , let us take a man who possesses a character only possessed by one man in twenty as exceptional .
For example , only one man in twenty is more than 6 feet inches in height , and such a stature may be considered " " exceptional In a class of twenty students we generally find one of\ldquo ; exceptional\ldquo ; ability , and so on .
Accordingly we have class fathers and sons who possess characters only possessed by one man in twenty as exceptional .
We first determine and , so that the tail of the frequency curve cut off is of its whole area .
This gives us Next we determine HK , and find HK Then we calculate the coefficients of the various powers of in ( xix .
) .
We find It remains to determine what value we shaU give to , the paternal correlation .
It ranges from3 to 5 for my own measurements as we turn from blended to exclusive inheritance .
Taking these two extreme values we find or 0096779 .
But , and the second term is the chance of exceptional fathers with exceptional sons , when variation is independent , i.e. , when there is no heredity , Thus or 012178 ; accordingly or 037822 .
Hence we conclude that of the 5 per cent. of exceptional men per cent. in the first case , and per cent. in the second case , are born of exceptional fathers , and per cent. in the first case and per cent. in the second case of llon-exceptional fathers .
In other words , out of 1000 men of mark we may expect 142 in the first case , 244 in the second , to be born of exceptional parents , in the first and 756 in the second are born of undistinguished fathers .
In the former case the odds are about 6 to 1 , in the latter 3 to 1 against a distinguished son having a distinguished father .
This result confirms what I have elsewhere stated , that we trust to the great mass of our population for the bulk of our distinguished men .
On the other hand it does not invalidate what I have written on the importance of creating good stock , for a good stock means a bias largely above that due to an exceptional father alone .
In addition to this the of the population forming the exceptional fathers produce142 or 244 exceptional sons to compare with the 858 or 756 exceptional sons produced by the of the population who are non-exceptional .
That is to say that the relative production is as 142 to , or 244 to , i. e. , in the one case as more than 3 to 1 , in the other case as more than 6 to 1 .
In other words , exceptional produce sons at to 6 times as as non-exceptional .
It is only because exceptional fathers are themselves so rare that we must trust for the bulk of our distinguished men to the non-exceptional class .
Illustration lII .
Heredit -colour of Hounds.\mdash ; To find the correlation in coat-colour between Basset hounds which are half-brethren , say , offspring of the same dam .
Here the classification is simply into lemon and white ( lw ) and lemon , black and white or tricolour ( t ) , The following is the table for 4172 cases : \mdash ; Proceeding precisely in the same way as in the first illustration we find : .
It will be sufficient now to go to .
We have TO THE THEORY OF EVOLUTION .
The quadratic gives .
Using the NEWTONIAN method of approximating to the root we find Summing up as before , after finding the probable errors , we have itance of Man.\mdash ; To find the correlation in eye-colour between a maternal grandmother and her granddaughter .
Here the classification is into eyes described as grey or lighter , and eyes described as dark grey or darker .
* As before , we find HK Series for up to The quadratic give , and the biquadratic the value of the term in being 000,00366 , so that higher terms may be neglected .
Determining the probable errors as in Illustration I. , we sum *According to GALTON'S classification , the first gronp contnins cyes described as light blue , blue , dark blue , blue-green , grey ; and the sccond eyes described as dark grcy , hazel , light brown , ) rown , dnrk brown , very dark brown , black .
Illustration of Stature.\mdash ; The following data have been found for the inheritance of stature between father and son from my Family Data cards , 1078 cases:\mdash ; Mean stature of father . . .
son . . . .
Standard deviation of father .
son . . .
Correlation Now for purposes of comparison of methods the correlation has been determined for this material from various groupings of fathers and sons : ( A. ) athers .
( B. ) Fathers .
TO THE THEORY OF EVOLUTION .
( C. ) Fathers .
( D. ) Fathers .
( E. ) Fathers .
( F. ) Fatloers .
VOL. CXOV .
of Results .
Classificatiom .
274,30 ) Now these Tesillts are of quite peculiar interest .
They show That the probable error of , as found by the present method , increases with and But the increase is not very rapid , so that the probable errors of the series range only between and 031 .
Hence while it is an advantage , it is not a very great advantage , to take the divisions of the groups near the medians .
It is an advantage which may be easily counterbalanced by some practical gain in the method of observation when the division is not close to the medians .
( ii .
) While the probable error , as found from the present method of calculation , is to 2 times the probable error as found from the product moment , it is by no means so large as to seriously weigh against the new process , if the old is unavailable .
It is quite true that the results given by the present process for six arbitrary divisions differ very considerably among themselves .
But a consideration of the probable errors shows that the differences are sensibly larger than the probable error of the differences , even in some case double ; hence it is not the method but the assumption of normal correlation for such distributions which is at fault .
As we shall hardly get a better variable than stature to hypothesise normality for , we see the weakness of the position which assumes without qualification the generality of the GAUSSIAN law of frequency .
( iii .
) We cannot assert that the smaller the probable error the more nearly will the correlation , as given by the present process , agree with its value as found by the product moment .
If we did we should discard 5213 , a very accordant result , in favour of 5529 , or even 5939 .
The fact is that the higher the correlation the lower , ceteris , the probable error , and this fact may obscure the really best result .
Judging by the smallness of and and of the probable error , we should be inclined to select or the value .
This only differs from by slightly more than the probable error of the difference ( as compared with029 ) ; but since both are found from the same statistics , and not from different samplings the same population , this forms sufficient evidence in itself of want of normality .
The approximate character of all results based on the theory of normal frequency must be carefully borne in mind ; and all we ought to conclude from the present TO THE THEORY OF EVOLUTION .
data for inheritance of stature from father to son would be that the correlation , while the product moment method would tell us more definitely that its value was .
There is no question that the latter method is the better , but this does not hinder the new method from being extremely serviceable ; for many cases it is the only one available .
Illustration .
Effectiveness of Vaccination.\mdash ; To find the correlation between strength to resist small-pox and the degree of effective vaccination .
We have in the earlier illustrations chosen cases in which in all probability a scale of character might possibly , if with difficulty , be determined .
In the present case , the relationship is a very important one , but a quantitative scale is hardly discoverable .
Nevertheless , it is of great interest to consider what results flow from the application of our method .
We may consider our two characters as strength to resist the ravages of small-pox and as degree of effective vaccination .
No quantitative scales are here available ; all the statistics provide are the number of recoveries and deaths from small-pox , and the absence or presence of a definite vaccination cicatrix .
Taking the Metropolitan Asylums Board statistics for the epidemic of 1893 , we have the table given below , where the cases of " " no evidence\ldquo ; have been omitted .
Proceeding in the usual manner we find Hence the equation for is whence Summing up we have , after calculating the probable errors , Strength to resist Small-pox when incurred .
We see accordingly that there is quite a large correlation between recovery and the presence of the cicatrix .
The two things are about as closely related as a child to its " " mid-parent.\ldquo ; While the correlation is very substantial and indicates the protectlve character of vaccination , even after small-pox is incurred , it is , perhaps , smaller than some over.ardent supporters of vaccination would have led us to believe .
lllnstrntion VII .
Effectiveness of Antitoxin Treatment.\mdash ; To measure quantitatively the effect of antitoxin in diphtheria cases .
In like manner we may find the correlation between recovery and the administration of antitoxin in diphtheria cases .
The statistics here are , however , somewhat difficult to obtain in a form suited to our purpose .
The treatment by antitoxin began in the Metropolitan Asylums Board hospitals in 1895 , but the serum was then administered only in those cases which gave rise to anxiety .
Hence we cannot correlate recovery and death with the cases treated or not treated in that year , for those who were likely to recover were not dosed .
In the year majority of the cases were , on the contrary , treated with antitoxin , and those not treated were the slight cases of very small risk ; hence , again , we are in great difficulties in drawing up a table .
* Further , if we compare an antitoxin year with a non-antitoxin year , we ought to compare the cases treated with antitoxin in the former year with those which would probably have been treated with it in the latter year .
Lastly , the dosage , nature of cases treated , and time of treatment have been modified by the experience gained , so that it seems impossible to club a number of years together , and so obtain a satisfactorily wide range of statistics .
In 1897 , practically all the laryngeal cases were treated with antitoxin .
Hence the best we can do is to compare the laryngeal cases in two years , one before and one after the introduction of antitoxin .
The numbers available are thus rather few , but will help us to form some idea of the correlation .
I take the following data from p. 8 of the Metropolitan Asylums Board ' Report upon the Use of Serum for 1896 ' :\mdash ; * When a ncw or process is introduced the medical profession are naturally anxious to give every patient the )eIlefit of it , and patients of course rush to those who first adopt it .
But if the real efficiency of the process or drng is to be measured this is very undesirable .
No definite datt ) by which to measure the effectiveness of the novelty are thus available .
TO THE THEORY OF EVOLUTION .
Here I find A further table is of interest:\mdash ; equiring In this case we have Lastly , I have drawn up a third table :\mdash ; Total Infantile Cases , Ages 0\mdash ; 5 years .
Here we The three coefficients are all sensible as compared with their probable errors , and that between the administration of antitoxin and recovery in laryngeal cases is substantial .
But the relationship is by no means so great as in the case of vaccination , and if its magnitude justifies the use of antitoxin , even when balanced against other ills which may follow in its train , it does not justify the sweeping statements of its effectiveness which I have heard made by medical friends .
It seems until wider statistics are forthcoming a case for cautiously feeling the way forward rather than for hasty generalisations .
Illustration VIII .
Effect on Produce of Superior Stock.\mdash ; To find the effect of superiority of stock on percentage goodness of produce .
To illustrate this and also the formula ( lxxxiii .
) for six correlation coefficients , we will investigate the effect of selecting sire , dam , and one grandsire on the produce when there *The values of for all the three cases of this Illustration were determined with great ease from Equation ( xxiv .
) .
is selective pairing of dam and sire .
We will suppose grandsire , dam , and sire to be above the average , and investigate what proportion of the produce will be above the average .
As numbers very like those actually occurring in the case of dogs , horses , and even men , we may take Correlation of grandsire and offspring sire or dam and offspring in both cases sire and grandsire . . .
Selective mating for sire and dam . . .
We will suppose zero correlation between paternal grandsire and dam , although with selective mating this may actually exist .
* We have then the following system : Hence , these values in ( lxxxvii .
) , we find\mdash ; after some arithmetic : But is the chance of produce above the average if there were no heredity between grandsire , sire , and dam , and no assortatiye mating .
Hence it equals .
N. Or , of the produce aboye the average , instead of are born of the superior stock owing to inheritance , \amp ; c. In other words , out of the above the average , are produced by the stock in sire , dam , and grandsire above the average , or by 1827 of the total stock .
The remaining8173 only produce3447 or the superior stock produces produce above the average at over twice the rate of the inferior stock .
Absolutely , the inferior stock being seven times as numerous produces about seven-tenths of the superior offspring .
IX .
Effect of xceptional Parentage .
-Chance of an exceptional man being born of exceptional parents .
Let us enlarge the example in Illustration II .
, and seek the proportion of exceptional men , defined as one in twenty , born of exceptional parents in a community with assortative mating .
*A correlation , if there be substantial selective mating , may exist between a man and his mother-inlaw .
Its rumonred absence , if established scientifically , would not , however , prove the non-existence of selective matiIlg , for A may be correlaced with and , but these not correlated with each other .
The propot'tion of pairs of parents associated with a grandsire above the average was found by putting , 2 , and for the threo correlation coefficients in ( lxxxv .
) .
In comparing with Illustration II .
, the reader must remember there doalt with an exceptional father , 1 in 20 , here only with relatives above the vely less gent selection .
TO THE THEORY OF EVOLUTION .
Here we take for father and son , for mother and son , and for assortative mating , We have then to apply the general formulae ( lxxxiii .
) and ( lxxxiv .
) for the case of three variables .
We have Whence , after some arithmetical reduction , we find But N. Hence N. We must now distinguish between the absolute and relative production of exceptional men by exceptional and non-exceptional parents .
The exceptional pairs of parents are obtained by ( xix , ) , whence we deduce , putting Whence the number of pairs of parents , both exceptional .
Thus , pairs of exceptional parents produce exceptional sons , and pairs of parents , non-exceptional in character , produce exceptional sons , i.e. , the remainder of the N. The rates of production are thus as 5014 to 0476 .
Or : of produce exceptional sons at rate more than ten times as great as pairs of non-exceptional .
At the same time , eighteen times as many exceptional sons are born to non-exceptional as to exceptional parents , for the latter form only about per cent. of the community .
The reader who will carefully investigate Illustrations II .
, VIII .
, and IX .
will grasp fully why so many famous men are born of undistinguished parents , but will , at the same time , realise the overwhelming advantage of coming of a good stock .

