]\gt ; [ 343 ] X. to the Theory of Evolution.\mdash ; II .
Skew in Material .
KARL PEARSON , University College , London .
Communicated by HENRICI , Received December 19 , 1894 , \mdash ; Read , 1895 .
PLATES 7-16 .
CONTENTS .
PART flCAL .
Page .
Section 1.\mdash ; Classification of iricaI frequency curves in general .
Types actually . . . . . . . . . .
344 Sections 2-5.\mdash ; Fitting of point-binomials by methods of quadratic and cubic . . . . .
345 Section 6.\mdash ; IIlustrations in cases of barometric and cral ) measurements . . . . .
351 Section 7.\mdash ; Fundamental geometrical relation between symmetrical binomial and normal frequency curve . . . . . . . . . . .
Section 8.\mdash ; Extension of this relation to the deduction of a skew frequency curve from the asymmetrical binomial . . . . . . . . . . .
356 Sections 9-10.\mdash ; General remarks on the defects of the normal fiequency curve as applied to actual statistics .
Skewness and limited range .
five types of theoretical curves 357 Sectiom ll.\mdash ; The hypergeometrical serios as replacing the point-binomial .
Cul'ves related to the hypergeometrical series in the same manner as the normal curve to the symmetrical point-binomial . . . . . . . . . . . .
360 Seotions 12-13.\mdash ; Equations to the possible frequency curves deduced from the hypergeometrical series . . . . . . . . . .
362 Section 14.\mdash ; Curve of Type I. Skewness with range limited in both direotions .
Criterion the existence of this type and method of fitting . . . . . . . .
367 Section 15.\mdash ; Special case of range being given . . . . . . . . .
370 Section 16.\mdash ; Special case of one end of range being given . . . . . . . . . .
371 Section \mdash ; Curve of Type .
Sytical with range limited in both directions .
This curve better than the normal curve , if observations the side of a point-binomial from normality . . . . . . . . . . . . . . . . . . .
372 Section 18-18 bis.\mdash ; Cnrve of Type III .
Skewness and range limited in one direction only .
Criterion and method of fitting . . . . . . . . . . . . . . . . . . .
373 Sections 19-20.\mdash ; Curve of Type unlimited in both directions , with vness .
and method of fitting .
Discussion of the -integral Sta fistic.al Examples .
Page .
tion 2 of romefer . . . . . . . . . . . . . . .
381 Section Crat)-measnrements ( Weldon No. 4 ) . . . . . .
, . . .
rebion 2 in height of ecruits . . . . .
ction 2tion in eight of ins n. . . . .
.386 Section ) \mdash ; Variation in length-breadth indcx of 900 sknlls . . . . .
Section enteric fever with age . . . . . .
390 tion 27.\mdash ; Distribution of guesses at mid-tints . . . . .
392 Section 28.\mdash ; Distincfion between " " skowness\ldquo ; and " " compoundness\ldquo ; in case of crabs ' " " foreheads\ldquo ; . . . . . . .
394 Section 29.\mdash ; Frequency of divorce with duration of marriagoe . . . . .
395 Section 30.\mdash ; Frequency of houses with given valuations . . . . .
396 in number of petals of buttercups . . . . .
399 Section 32.\mdash ; Variation in number of projecting blossoms in clover . . . . . .
402 Section iation in nIlmber of dorsal teeth on rostrum of prawn . . . . .
403 Section.34 .
iation in pauper.percentages in England and Wales . . . . .
404 Section 35.\mdash ; Resolution of mortality curve for English males into components . . . . . .
406 Section 36.\mdash ; Concluding remal.ks on skewness , variation , and tion ( correlation for whist ) . . . . . . . . . .
410 Note on 1ELE'S heatment of skew fi.equency . . . . .
412 Asymm itrical Frequency AN asymlY)etrical frequency curve may arise from two quite distinct dasses of causeS .
In the first place the material measured may be heterogeneous and may consist of a.mixture of two or more homogeneous materials .
Such frequency curves , for example , arise when we have a mixed population of two diffel.ent races , a homogeneous population with a sprinkling of diseased or deformed members , a curve for the frequency of matrimony covering more than one class of the population , or in economics a frequency of interest curve for securities of different types of stability\mdash ; railways and stocks mixed with and financial companies .
The treatment of this class of frequency curves requires us to break up the original curve into component parts , or simple frequency curves .
This branch of the subject ( for the special case of the compound being the sum of two normal curves ) been treated in a paper presented to the oyal S iety by the author , on October 18 , 1893 .
The second class of frequency curves arises in the case of homogeneous material when the teudeJlcy to deviation on side of the mean is unequal to the tendency to deviation on the other side .
Such curves arise in many physical , economic and biological investigations , for example , in frequency curves for the height of the , rometer , in those for pl.ices and rates of interest of securities of the same class , in mortality curves , especially the of to cases in all kinds of MR. K. PEARSON ON THE MATHEMATICAL THEORY OF EVOLUTION .
345 fevers , in income tax and house duty returns , and in various types of urements .
It is this class of curves , which are dealt with in the present paper .
The general type of this class of frequency curve will be found to vary ( see Plate 7 , fig. 1 ) through all phases the form close to the negative exponential curve : to a form close to the normal frequency curVe where and are constants .
Hence any theory which is to cover the whole series of these curves must give a curve capable of varying from one to another of these types , , from a type in which the maximum practically coincides with the extreme ordinate , to a type in which it coincides with the central ordinate as in the normal frequency curve .
It is well known that the points given by the point-binomial coincide very closely with the contour of a normal frequency curve when is ) moderately large .
For example , the 21 points of lie most closely on a normal frequency curve , and the author has devised a probability machine , which by continually bisecting streams of sand or rape seed for 20 successive falls gives good normal frequency curve by the heights of the resulting 21 columns .
Set to any other ratio of division other than bisection , the machine gives the binomial , or indeed any less power and thus a wide range of asymmetrical point-binomials .
Plate 7 , fig. 2 , represents , diagramatically , a 14-power binomial machine .
Just as the normal frequency curve may be obtained by running a continuous curve through the point-binomial when is fairly large , so a more general form of the curve may be obtained by running a continuous curve the general binomial .
As the great and only true test of the normal curve is : Does it really fit observations and measurements of a symmetrical kind ?
so the best argument for the generalised probability curve deduced in this paper is that it does fit , and fit surprisingly accurately observations of an character .
Indeed , there are very few results which have been represented by the curve which do not better fit the generalised probability curve , \mdash ; a slight of asymmetry being probably characteristic of nearly all groups of measurements .
Before the generalised probability curve , it may be well to show how ally asy1metrical curve may be fitted with its closest .
This will be the topic of the following five articles .
( 2 .
) Consider series of rectangles on equal base ) whose heights are resl ) tively the successive terms of the binomial , where .
Here clearly the area of the entire system .
Choose as origin a distant from the * I have found it convenient to use the term mode for the abscissa corrcsponding to the of mitximum frequency .
Tbus the " " mean the " " mode and the " " media1J " " ve all actcrs important to bhe statistician .
346 MR. K. uSON ON THE MATHEMATICAL THEORY OF EyOLUTION .
mdar .
of the rectangle , the line of common bases , and let be the height of the , or while Let us find the values of where is aIly integer , for values of from to 4 .
It is easy to see that where the operation is repeated times .
The operations indicated can easily be performed by when and the successive values can be found by theorem .
After differentiation we may put or .
There resnlts : Let NG be the vertical the ntroid of the system of rectangles , then dearly ON MR. K. PEARSON ON TIE MATICAL THEORY OF UTTON .
347 We shall now proceed to find the first four moments of the system of rectangles round .
If the inertia of idcred as its llltd , we old 1 for moment round , The resulting are whence , remembering that , we find and are roots of \ldquo ; Thus , when , and have been calculated for the frequency curve , the elements of the point-binomial are known .
These results were given by me letter to ' Nature , ' October 26 , 1893 .
They give quite a fair solution so as is large and small , i. e. , so as the symmetry and the " " excess Phil. Trans , sured respectively by and ( which vanish for the curve ) not ] In many cases , however , they siderable , and the following solution is general .
* If denote the largest term in and cyoI ) it , then nn of heorem\mdash ; if bc \mdash ; shows that Takc I og the right side in we Hencc , Now , mnking use of the lucs given in S ] wo ( 3 .
) To find the moment of a trapezium ABCD about a line parallel to its sides , the of the parallel sides , , their distances the Let be the moment .
Then . . .
( 4 .
) Now consider a curye of observations made up of a series of trapezia on equal bases , as in the accompanying figure : etc. hcre and This appears to be ) more cneral form of result given by Professor , 'Boy .
Soc. , p. 271 .
For the normal lence , if does not differ much from and ] ] bc hmall , and we may } their produck with .
Thus approximately .
rees wibh ecitd case if we expand the second exponential .
His ( negave frequency\ldquo ; is tccolulted for by fact that he has only { aken the first terms of long , i. e. , I have not considered this form of the sk-ew-cu at length , because it is only a first approximation to the mor general con sidcred this , and further , because it is only pplicable in practice in limits .
MR. K. PEARSON ON ] THEORY OF EVOLUTION .
,349 Here are the frequencies of deviations falling within the ranges . .
, and the tops of the ordinates are joined to form a frequency-curve in the usual manner .
Let be the nth moment of the system of trapezia about the Oy , then In particular , if we take Oy in the position at distance from , we , and , where .
In particular , , , , .
When we put , and , these reduce to , , , .
Now let be the value of the nth moment of the trapezia system about tho vertical through its centroid divided by its area .
We Thus we find : , , , .
Comparing these sults with those given in the 'Phil .
Trans vol. 185 , p. 79 , Eq. ( 4 ) , we see that treating the curve as built-up of trapezia instead of loaded ordinates introduces the parts into the values of the 's enclosed in ] brackets .
These additions small , but in many cases quite sensible .
Sirlce the series of trapezia giyes in general a closer approach than the series of loaded ordinates to the frequency curve , and , further , since the calculation of these additional terms is not very laborious , it be better for the future to calculate the moments of any equency curve from above modified formulae .
( 5 .
) now to the point-binomial , we have : ?
: , , .
If , instead of taking brapezia , had taken a series of rectangles , but not , as in S2 , concen t their areas along their axes , we should have found the following system : , , MR. K. PEARSON ON THE MATHEMATICAL LVOLUTION .
351 Hence if we write : , , we have : For trapezia : For rectangles : For loaded ordinates : the above general system may be applied to all cases .
Writing , and we have by elimination the cubic for The reJnaining constants of the binomial are : ' , and ( 6 .
) Let us illustrate these results by a numerical example .
gives Dr. VENN'S curve for baromef heights .
Alongr the onbal , 1 cm .
equals of height of barometer , and the scale of frequency is 1 sq .
The centroid vertical and the second , third , and bout it found for by the graphical process descl.ibed , 'Phil .
Trans vol. 185 , p. 79 .
We ve the following results:\mdash ; *This sult scems of able importance , and 1 do llot belve it has ct noticed .
It givcb the square for binomial distribufion , and we scc that for most practical poscs it identical with the value hitherto deduced as an result , by nssuming the to bc tcly a llormal curve .
If we the fundamental cubic reduccs to a form in which the coefficients aro easily calculated and the naturc of the roots iminated .
By .
G. U. YULE , who has given me very great in the laborious in the reduction of frequency curves .
We have used , with much economy of time , the " " \ldquo ; 10.14 , 15.95 , all in centimetre umits .
These give Hence for trapezia , and for rectangles , These give the following solutions:\mdash ; Here gives the distance of the start of the point-binomial from the centroid vertical .
The three point-binomials are therefore 77 .
86 .
respectively .
These omials represented in Plate 8 , fig. 3 .
It will be noticed that they very close to the barometric curve ; they would be still closer if that curve were a real curve and not a polygonal line .
The total areas between binomialpolygons and observation curves , treating all parts as positive , are for the three cases , sq .
centims. respectively , or taking the base range to be 23 centims. , we mean iations from the observation curve , 457 , 478 in the three cases respectively .
Thus the method of trapezia gives slightly the best result ; the method of along ordinates the worst result .
The total area of the curve being , we from another standpoint , mean percentage errors*in the ordinates of about , and espectively .
The generalised probability curve , if fitted to the observations , gives an areal deviation of 7 .
centims. , or a percentage error of about 4 .
Thus very nearly one-third as close again as the point-binomials .
* The " " percentag.e error\ldquo ; in ordinate is , of a rough test of the goodness of fit , but I have used it in default of a better .
MR. K. PEARSON ON THE THEORY EVOLUTION .
353 As typical samples of mean percentage errors considered by various statisticians to give good results , I may note the following , the frequellcy being about 1,000 or npwards : , 9 ; RRIMAN , ; GALTON ( Anthropometric ) , 7 to 15 ; WELDON ( Crabs ) , , ( Shrimps ) , ; STIEDA ( Skulls ) , ; PORTER ( School Girls ) , ( Recruits ) , ; BRADLEY'S observations , ; PEARSON ( Lottery ) , , ( Tossing ) , It is therefore clear that our point-binolnials and generalized curve may be considered to give good results .
* It will be noticed , however , a little difference in the method of calculating the point-binomials leads , without much alteration of the percentage error , to a considerable change in their centroid-positions and the gnitude of their constants .
Generally speaking we may conclude that in round numbers the barometric frequency corresponds to the binomial , or to the distribution of zeros when 20 ten-sided teetotums , marked , 1 9 , are spun together .
There is an apparent } ) limit to the height of the barometer , and its deviation below the mean much greater than its deviation above .
At the same time within the narrower range round the mean , the frequency of a high barometer is greater the frequency of a low barometer ; the odds against a " " contributory cause\ldquo ; tending to a low barometer being about 9 to 1 .
I propose to investigate a wider series of barometric observations , in order to test how far the conclusions which may be drawn from Dr. VENN'S statistics are general .
A rather interesting point ) be considered at this stage .
Is it always possible to fit a point-binomial to a series of observations with a chance frequency ?
Can we better the normal curve by a point-binomial The answer is Yes , if the fundamental cubic in ( second footnote , p. 351 ) , has a real positive root .
Now for the normal curve 2 , or is zero .
For the loaded ordinates will only be real if this expression be positive .
It may , however , take small negative values for the trapezia , in which case itself will be small and only within narrow limits give suitable values for Hence , for real values of and , it is impossible to fit a point-binomial to a series of observations for has a large negative value .
The normal curve , for which , is nearer to any such obseryations a ] For example , by aid of the modified explessions given in this paper , p. 350 , we have * As another manner of , compare the ten-points of the point-binomial for lines obscrvntions : \mdash ; Theory 19 7 2 03 Observation 12 2 00 A curve drawn ough the points of the three point-binomials woulCl be very close to fhc obscrvations .
As a matter of fact , the skew probability curve passes very near to a1130 points .
MisH A. Lec has since calculated the of three yoars of stbourne baromctric obscrvations for While / and differ widcly from ths values , she finds striking and suggestive agreement .
] for the data given for Professor WELDON'S Crab Measurements , No. 4 , ' Phil. Trans 18 p. 96 .
Hence , Thus is positive , and accordingly no rational point-binomial is likely to ft well as the normal curve .
As a matter of ct the fundamental cubic is now The two negative roots of this equation give imaginary value for and .
The small positive root gives greater than unity and negative , is also negative .
I can ) interpretation to these results , it seemed well to complete in the latter case the solution test how the point-binomial fitted the curves .
I found These give for the binomial or , Thus the sensible part of the binomial to the scale of our figure is a triangle .
I have drawn this binomial , see Plate 8 , fig. 4 .
The reader will mark a fit very close on the whole to the observations .
We have the folowing percentage mean errors of the ordinates:\mdash ; Normal curve . . . .
67 , Skew bility curve .
4.4 , Bin omial . . . .
10.5 .
We conclude , therefore , that even if our binomial constants have unintelligible values , yet onr method will give , in many , a closely-fitting polygonal figure .
This remark should be read in connection with Professor ] somewhat ( on fitting chance distributions with curves other than the normal error .
It is ) ossibl in ahnost every case to find simple combinations of ' Phil. vol. 334 , p. 24 , seq. , 1887 .
MR. K. PEARSON ON THE MATHEN[ATICAL THEORy OF EVOLUTION .
355 circles , or parabolas of various degrees which , give results extremely to any given set of observations .
For example , taking the range of frequency to be sensibly times the standard deviation , we have the following close expression for the error function by harmonic analysis Here is the maximum ordinate , any deyiation , and the ndard deyiation .
A couple of wave curves will thus very frequently give us a close approximation to a set of statistical measurements , quite as close as statistical practice shows the error curve to be .
The above expression further allows the normal curve to be constructed by aid of scale and cetrically , or its ordinates calculated from a table of cosines .
AnotlIer example of the fitting of a point-binomial will be found in Part 2 , S 34 , Pauper Percentages .
( 7 .
) Consider the point-binomial , where is any constant , and suppose a polygon formed by plotting up the terms of the binomial at distance from each other .
Then , esponding to , we have and if .
Now is the slope of the polygon corresponding to the mean ordinate , , writing * It is often sufficienh to takc The divergence of this value of from the ordinary value is to be noted .
sensibly if be great .
on a large scale , however , the point-binomial and the two normal with standard deviations of , I find that the Iatter has a error of only as compared with of the former .
Thus it .
that the corresponding to pq fits the point-binomial closer than one with ] usually dopted .
] Now compare this property of the polygon with that of the curve : We have by differentiation : Hence : this binomial polygon the curve of frequency very dose to each other , of geometrical nature , which quite independent of the ntagnitude .
In short their slopes are given by an identical relation .
By a proper choice of and , we can get the normal curve to fit closely the pointbinomial , owing to this slope property , ssumpt nto the indefinitely value of .
It is this geometrical property which is gely the ification for the manner in which statist , icians apply , and apply with success , the normal curve to cases in which is undoubtedly small .
No stress seems hitherto to have been laid upon the fact the normal curve of errors besides being the limit of a symmetrical point-binomial ha , also intimate geometrical relationship with it .
( 8 .
) Now let us deal with the skew point-binomial in precisely the same manner as we have dealt with the binomial .
Taking its form to be , we have , if and : Let us write X , .
Then , and : following table bhows the closeuess of frequency within a given as determined by the binomials : \mdash ; Here the distribution of 100 groups each of 100 evonts is seen to be actically the same whether we take or MR. K. PEARSON ON THE MATHEMATICAL THEORy OF EVOLUTION .
357 ' or , if if and The curve which has the same law of slope as this skew binolnial is : ( 9 .
) This curve accordingly stands in the same relationship to the skew the normal curve to the symmetrical binomial .
There are several points , however , to be considered with regard to it .
In the first place it is usually assumed that is indefinitely great and indefinitely small , and then it is supposed that we may neglect as compared with , and so we deduce tlJe normal error curve whether be equal to or not .
But I contend that this is unjustifiable except for very small values of .
When the deviation X ' is considerable and vanishingly small , X ' will be an indefinitely great multiple of must be in fact the unit in which X ' is measured and unless , the ordinary normal curve is only approximation , even if be large , near the maximum frequency .
In the next place , when we speak of being large , are we quite clear as to what we mean in the case of physical or biological frequency curves ?
We speak of a multiplicity of small " " causes\ldquo ; determining the actual dimensions of an organ , or the size of a physical error , or the height of the barometer .
But it is less clear why this multiplicity should be identified with the infinite greatness of .
If we take Dr. VENN'S frequency curve for barometric height , we see that the closest point-binomial is by no means consistent with either , or with being indefinitely great , many statistical results in games of chance are given with great exactness by the normal curve , although we are then able to show that is quite moderate .
Now , it is true that the biological and physical statistics to which we are ring , give essentially continuous curves , but it does not seem to follow of necessity that must be infinite ; while their frequent skewness sufficiently that the neglect * Note again the deviation of the constant its usually adopted val ae of , as compared with is unjustifiable .
Thus , the maximum of a fever mortality curve cannot be an infinite distance from birth , which limits the curve in one direction , nor an age-at-nlarriage curve have a maximum frequency infinitely distant from the age of puberty , nor a frequency of interest curve separate its maximum , between 3 or 4 per cent. , by an infinite distance from per cent. It is clear , therefore , that if such frequency curves as those referred to are to be treated as chance distributions at all , it would be idle to compare them to the limit of a symmetrical binomial .
really quite ignorant as to the nature of the contributory " " causes\ldquo ; in biological , physical , or economic frequency curves .
The continuity of such frequency curves may depend upon other features than the magnitude of If I toss twenty coins , a discrete series of , 1 , 2 , 3 , . .
20 , heads is the only possible range of results .
Each individual coin , here representing a " " contributory cause\ldquo ; can only give head or tail , and so many whole coins give head , so many tail .
If I want to make any ratio of head to tail , I have to take an indefinitely great number of coins , for each " " contributory cause\ldquo ; must give a unit to the total .
But it may possibly be that continuity in biological or physical frequency curves may arise from a limited number of " " contributory causes\ldquo ; with a power of the result .
We cannot conceive on the tossing of 20 coins that will give heads and will give tails , we are obliged to deal with 200 coins , 135 giving heads and 65 tails .
Yet the two things are not identical .
The former corresponds to a value intermediate between two ordinates of , and the latter to a definite ordinate of .
So long as we remain in ignorance of the nature and number of " " contributory causes ' in physics and biology , so long as we do find markedly skew distributions , it seems to me that we must seek more general resuJts than flow from the assumption that p and , The form of curve given in S8 above is suggested as a possible form for skew frequency curves .
Its justification lies essentially , like that of the normal curve , in its capacity to express statistical ( 10 .
) But it must be noted that the generalised probability curve in S although it contains the normal curve as a special case , is not sufficiently general .
It is hmited in direction , indefinitely extended in the other .
This limitation at one end only , corresponds theorctically to many cases in economics , physics , and biology .
But there are a great variety of cases in which there is theoretical limitation at both ends ; that is to say , there is a limited of possible deviations .
For example , let a f , rapezlum , ABCD , of white paper be pasted on a cylinder of black surface with MR. K. PEARSON ON MATHEMATICAL THEORY EyOLUTION .
359 , the axis of symmetry parallel to the axis of the cylinder .
Then , if the cylinder be rotated , we shall have a series of grey tints from a darkish to a lighter Now , if we ask several hundred persons to select a tint which would result from mixing the tints at and , we shall obtain a continuous frequency curve , falling , however , entirely within the range to .
Or , suppose a frequency curve obtained by plotting up the frequency of a given ratio of -length to total bodyor of carapace to body-length .
Here the lie between and 1 .
It is not that other values are excessively improbable , they are by the of the problem absolutely impossible .
Hence , it is clear that the curves obtained by Professor WELDON Mr. H. THOMPSON in the case of shrimps , crabs , and prawns , can only be approximately normal curves , even if it were possible for the ratios to run to 1 .
But as a matter of fact , the possible range is very much .
We may not be able to assert , priori , what it is , but for an adult prawn to have a carapace or of its body-length , or a man leg or of his -length , may be regarded as impossibilities ; they are abnor- malities , which could hardly to the adult condition .
Precise]y the same remarks apply to skull indices , and probably to the tive size of all of organs in the adult condition .
We may not know the range , priori , but we are quite certain that one exists , and it is a quantity to be determined\mdash ; just as the mean or the standard from our measurements themselves .
We may take it that in most biological measurements of adults there is a range of stability , so to speak , organs not falling within this range are inconsistent with the continued existence of the individual , with the that he has lived to be an adult .
Nor is this question of range confined to biological statistics .
A barometric frequency curve must show the same peculiarity ; there are excessively low and excessively high barometric heights which would be not only inconsistent with the survival of any meteorological observer , but also with the existing features of physical natuYe on this earth .
In vital statistics we find precise]y the same thing , a curve of percentages of mothers of different ages for the children born during any year in a country would be definitely limited by the ages of puberty and the cteric , which cannot be pushed indefinitely towards childhood and senility respectively .
Again in disease and mortality curves , while the lower limit of life is clear , it is highly probable that an upper limit exists , if we can only fix it by investigation of our statistics themselves .
A man of the present day , as now organised , may be able to live 120 years , perhaps , but we have exceeded his vital possibilities if we take , say , 200 years .
Thus the problem of range seems a very important one , it theoretically excludes the use of the normal curve in many classes of statistics ; it is quite true that , for many practical purposes , frequency curves of limited ramge may be sensibly identical either with unlimited curves , or even with normal curves , but , in other this * Absolute malformations , congcnital , or dnc to accidcnt are excluded .
or ampntations would be natnxally excluded from our measurements .
is not so , and under any circumstances the limited curve may actually give information as to the possible range\mdash ; the " " limits of stability\ldquo ; \mdash ; which is itself of great value .
We have , then , reached this point : that to effectively we require probability curves which include the factors of skewness range .
The generalised curve we have already reached , possesses skewness , but its range is limited in one ection only .
, we require the following types of frequency curyes:\mdash ; Type I.\mdash ; Limited ran in both directions , and skewness . .
\mdash ; Limited range and symmetry .
Type III .
\mdash ; Limited range in one direction only and skewness .
Unlimited range in both directions skewness .
V.\mdash ; Unlimited range in both directions and symmetry .
Type .
is the normal curve ; Type IV .
, with slight skewness , has been dealt with by POISSON in the form of an approximative series .
* Type III .
has been given above , it was first published by me wichout discussion in ' Roy .
Soc. Proc vol. 54 , We can now turn to the general problem .
( 11 .
) A very simple example will illustrate how a frequency curve , with limited range and skewness , may be considered to arise .
Take balls in a bag , of which are black , and are white , and let balls be drawn and the number of black be recorded .
If , the range of black balls will lie between and ; the resulting polygon will be skew and limited in range .
This polygon , which is given by a hypergeometrical series , leads us to generalised probability curves , in the manner as the symmetrical and skew binomials lead us to special cases of such curves .
If we consider our balls to become fine shot , or ultimately sand , and suppose each individual grain to have an equal chance of being drawn , we obtain a continuous curve .
It is not , however , impossible that , could we measure with sufficient accuracy , many physical as well as statistics might be found to proceed by units , much as in certain types of economic statistics we are not troubled with fractions of a penny .
For this reason we shall keep our results in the most , general form , and obtain a curve approximating to the hypergeometrical series referred to without any assumptions as to the relative magnitude of the quantities involved .
We easily obtain for the series giving the chances of , black balls being draw out of a bag containing , black , and , white , the expression Probabilite des Jugements chapter 3 .
pints of red sand and pints of white put into a vessel , and pints are withdrawn .
We if , a perfectly continuous frequency curve for red sand withdrawn ranging between and pints .
We here supposing no " " pel .
fact mixture\ldquo ; of the two kinds of sand , bnt theoretical equality of cllances for each MR. .
PEARSON ON THE MATHEMATICAL OF VOLUTION . .
If be the ordinate of this polygon , and we suppose these ordinates plotted up at distances apart , we have .
Thus Write and we find with our previous notation where Now , if we attempt to find the curve which has the same geometrical relation for the slope as the above hypergeometrical polygon , we see that it will change its type according to the sign of After some reductions we have MR. K. PHiARSON ON THE ICAL THEORY OF EyOLUTION .
will be real , according as lies outside or between the limits If lies outside these limits , then the integral of the right-hand side of equation is purely logarithmic ; if it lies between these limits , the integral is in part trigonometrical .
Since must be less , it follows that the integral must be trigonometrical if these limits are respectively and , i.e. , if or lie between For example , if 100 , then , if lies between and , the integral must be If lies outside these limits , say for example , then the integral will be logarithmic if does not lie between and 96 , i.e. , if we draw a small or large proportion of the total contents .
Let us treat the trigonometrical and logarithmic cases separately .
( 12 .
) Case I. The curve having the same geometrical slope relation is constant Write the origin ; further put for } , for , and for , then we have , being a constant of integration , This frequency curve is asymmetrical and has an unlimited range on either side of the .
It corresponds to the curve required as Type IV .
Here .
MR. K. PEARSON ON THE MATICAL T OF EVOLUTION , 363 cases .
Suppose , and very , then , say , Thus we have which reduces to the normal type by a change of origin .
It is to notice , however , that the standard deviation of this normal type and is very from the value , nearly , which is the usual form .
Only when we put and make small do they agree .
We thus conclude : That the form may fit , but it not follow that the standard deviation is of the binomial type generally cyssuraed .
( ii .
) Suppose , corresponding to the withdrawal of one-half of the contents of a vessel , then where This is an unlimited and frequency curve approaching more and more nearly to the normal form as we increase .
It has , however , a standard deviation , while the normal curve would give .
( iii .
) Suppose , we sgain reach the form where Make infinite and we have again the normal type , but a standard deviation of the form ouly approaching the usual value when is We postpone until we have discussed the remaining types the problem of fitting curve of Type .
to a series of observations .
( 13 .
) Case II .
Let and be the roots of .
Then the curve having the same geometrical relation for its slope is or , if , ' by constants .
Assuming that and can take any sign whatever , we see that there are three fundamental subtypes of this frequency curve , This is an asymmetrical curve with limited range and maximum towards mediocrity .
As a rule and are fractional and the curve becomes imaginary beyond the limits and ( ii .
) Here the ordinate between and varies from infinity to zero , and resembles the frequency curves given by wealth\ldquo ; distribution or infant mortality .
( iii .
) This is an curve with limited ) , mediocrity being in a minimum .
disappearance of mediocrity is not a very uncommon feature of statistics ; the MR. K. PEARSON ON THE MATHEMATICAL THEORY OF EVOLUTION .
365 " " prevalence of extremes\ldquo ; may appear not only in meteorological phenomena but in competitive examinations , where the mediocre have occasionally sufficient wisdom to refrain from entering .
The type is that of Mr. F. GALTON'S curve of\ldquo ; consumptivity ' The curve contains an interesting number of less fundamental subtypes .
Make in , This is the limit to the asymmetrical binomial , which has been already referred to in S8 .
Make This is the symmetrical frequency curve of limited range .
Make negative in , This is a symmetrical frequency curve , with limited range , and minimum of mediocrity .
( vii .
) Pub in and make , This is the normal curve .
* ' Naturnl Inheritance , ' 1889 , p. 174 .
( viii .
) Put in , ?
This is an asymmetrical fi:equency curve , with an ordinate varying from to along an infinite ange .
All eight of the above types are included in the single form or if we give positive , negative , or limiting values to the constants .
But to do this we require to give values to and in the expressions for , and , which are not easily intelligible , if we rigidly adhere to our example of drawing a definite quantity of sand from a limited mixture of two kinds of sand .
The last type of curve given is , however , the frequency curve for a priori probabilities , readily admits of a direct interpretation of the following kind .
Given a line of length , and suppose points placed on it at random ; what is the frequency with which the point from one end and from the other of the series of points falls on the element Sx of the line ?
The is clearly or , we have a frequency curve of the type We may express the problem a little differently .
Take cards and slip them at random between the pages of a book , the frequency of the page succ , eeding the card is given by the above curve .
CROFTON , " " Probability S 17 , ' Encycl .
Brit. ' The important point to be noticed hore i that we are dealing with a tion in which conh.ibutory causes are inte]'-dependent .
MR. K. PEARSON ON THE MATH EfICAL THEORY UTION .
367 Until we know very much more definitely than we do present , how the an organ in any individual , say , depends on the sizes of the same organ in its ancestors , or what are the nature of the causes which lead to the determination of ices , or of in come , or of mortality at a given age , I do see that we have any right to select as our sole frequency curve the normal type in preference to the far more general which not only includes the former , but supplies the element of skewness which is undoubtedly present in manly statistical frequency distributions .
As we may look upon the former as a limit to a coin-tossing series , so the latter represents a limit to teetotum-spinning and card-drawing experiments .
It is not easy to realise why nature or economics should , from the standpoint of chaI ) , be more akin to tossing than to teetotum-spinning or card-dealing .
At any rate , from purely utilitarian and prudent motives , we are tified so long as the analysis is manageable , in the more general form .
It will always give us a measure of the divergence of particular statistics from the normal type , and in maYly cases of skew , it can be used when it would be the of absurdity to apply the normal curve at all .
Since Types I. , II .
, III .
, and V. are all represented by the curve Type IV .
by the we have only to deal with these two cases in general .
We shall refer , in the course of our work , to special simplifications arising in particular sub-cases .
After a description of the manner which these generalised probability curves may be fitted to , we shall indicate , by examples , their practical applications .
the Generalised Probability Curve .
Typc Let the , whence and Further let thus Let be the of the curve between and moment l.ound parallel to bhe axis of through , and its moulent round the centroid vertical .
Then have , Thus , by the fundamental property of the function , we have , From these we easily deduce by the connecting and , if we write for brevity , , and ] ' : Now , , and are to be found by the methods indicated in Art .
4 from the polygon of observations , ) may be supposed known quantities , when we are dealing the fitting of frequency-curve to observations .
Then , if we have : : whence , eliminating , we : MR. K. PEARSON ON THE MATHEMATICAL HEORY OF EyOLUTION .
369 This gives , then : ' Since and ) are roots of Thus and are determined .
Further , , and are all determined .
Lastly : , give : \ldquo ; ' which completes the solution , if a Table of funcfions is to Remarks.\mdash ; It is clear that the solution is unique .
Tt is necessary in order that the solution may be real , that and old be real or .
Hence , if be negative , there is certainly a solution , because is always real .
The solution forms , however , one of the sub-types referred in our , ( ii ) and ( iii ) .
If be positive , we must have positive , or Now it is easy to prove that for any curve or is positive , for is always greater than Thus , we must have or *Very often with sufficient accuracy we may take : JDCCCXCV .
MR. K. iSON ON VOLUTION .
Now it is npossible to fit a normal to a frequency distribution for which .
It is , however , to fit this generalised curve of I. , bc , provided there is sulficient skewness to render .
the in the type of curve suitable for a given set of tjons is to ascertain the value of ression be sec a limitc ) ossibility .
ssing f to loness we remark that the distance between the centroid vertical the ximum ordinate Now it yonld good sure of vness , and it would if .
But , have not , it seems to me bettcl to sure of skewness the of the distance between the the centroid to the of the swing radius of the curye about the roid vertical , , the quantity In our we have skewness , in and skewness are determined in Type I. ( 15 .
) very considerable simplification of the bove analysis arises when the range is ( rlven by ) conditions of problem itself , between two given tints .
this we require oments one end of the rauge , amd the solulion ) as as in ) case of fitting a normal curve .
incc known , let and .
, of fioxion of ) 1 ) on cithcl sidct of MR. ARSON ON MATHEMATICAL THEORY OF EVOLUTION .
371 Then , and we have at once Then , and give and .
Finally given as before by ( 16 .
) A perhaps still more interesting and usual case arises when one end of the range is given , i.e. , when , but not , is known .
For example , a ) curve of bution of disease with age , the liability to the disease starting with birth .
Here we require to calculate from the observations and .
The solution is as follows : Let ; then and .
Solving Thus , determines the range .
Hence , since , and ' we have , with the aid of the previous expression for , the solution of the problem .
MR. K. PEARSON ON THE MATHEMATICAL THEORY OF EVOLUTION .
( 17 .
) bility curve of Type .
Limited Range and Symmetry .
The solution in this case follows very easily from ( 14 ) by putting , we have at once or Since , and dearly we have or ' the normal frequency curve , for a symmetrical point-polygon Hence , whenever a symmetrical frequency curve differs from the normal curve on the side of the -binomia we can better the normal solution by taking a symmetrical frequency curve oflimited range .
Since and if , we easily trace the transition from the limited symmetrical curve to the normal curve with infinite range .
Quite apart from extremely interesting problem of finding the range , it is clear that better fits will be obtained symmetrical distributions by the aid of this limited range curve for all cases in which MR. K. PEARSON ON THE MATHEMATICAL HEORY OF EVOLUTION .
373 ( 18 .
) bility Curve of the Type III .
Range limited in one direction only .
In this case we have no need to determine the value of , and the analysis is much simplified by the replacement of the function by a single function .
Take and , we have Further , .
Thus we find Hence , whence Or , transposing to the centroid-vertical , we have The first two results give us at once whence This compleies the solution of the problem , which is seen to require only the determination of and Remarks.\mdash ; The distance of the centroid-vertical from the axis of or maximum ordinate , is given by skewness MR. K. PEARSON ON THE MATHEMATTCAL OF EVOLUTION .
If we transfer th origin to the centroid-vertical we have where It is interesting to note how this skew curve passes into the normal curve when is made vanishingly small , or By WALLIS'S theorem the limit to It remains to find the limit of Now the limit of ) for is easily found to be , hence , the normal form .
Returning to the value we have found for .
and eliminat , ing and between , and we find is the expression ( see p. 398 ) which must be positive in the case of limited range .
It is zero also for the normal curve , because both and vanish .
Hence the more nearly the quantity approaches to zero , the more nearly are we able to fit our statistics with a skew frequency-curve having a range limited in one direction only .
.\mdash ; The skew frequency-curve of Type III .
deserves especial notice .
It is intermediate between those of Type I. and } ) , and they differ very little from it in appearance .
Hence , if the reader has once studied the various forms which Type III .
can take as we alter its constants , he will grasp at once the forms taken by I. and , by slmply considering the range doubly limited or doubly unlimited .
To the process of realising Type III .
, Plate 9 , fig. 5 , has been constructed ; it contains seven sub-types of this species , varying from , in which the curye is asymptotic to the ) aximum frequency-ordinate to fig. , which is practically identical with the normal curve .
Taking for the equation to the curve , we have tlJe following values for the constants and :\mdash ; MR. K. PEARSON ON THE TICAL THEORY EVOLUTION .
I. I I. 001 III .
265 IV .
V. 1 VI .
6.5625 3 2505 363 7676 5 In tlJe diagrams vertical and horizontal ( ?
and a ) have been chosen so as to illustrate best the changes of in the curve .
The general correspondence cf 1 ; with actual of frequency curVe , in Plate 7 , fig. 1 , will at ollce strike the reader .
The mean , median , and the mode or marked by bb , , respectively , as soon as the curves were , a remarkable elation manifested itself between che position of these three quantities : the median , so long as was positive , was seen to be about one-third from the towards maximum .
For negative and between and \mdash ; 1 , this relation was not true .
The distance between the maximum-ordinate and the mean is , if the equation to the curve be equal to .
Now the maximum cannot be accurately from observation , fair approximation can be made to the median .
Hence the constant could , if the relation were shown to be always true , be determined approximately as ]of thrice the Now distance of mean from and , , maximum , , Hence , supposing distance of median , we should ) to find about .
Equating the integral which gives the area up to the median to half the total , we have or , This is the equation for .
Unable to it generally I a series of erer values and found in all cases nearly 67 .
Its value , however , as creased .
, therefore , assumed to be really of the form , and determining and by the method of least squares , found Probably this is only the beginning of a rapidly converging series in inverse powers of , but it would appear to suffice for most practical purposes .
It is only true for and does not explain why , when is and fractional , is still apparently near ; thus its value for has only risen to 6931 .
We have then the following fairly means of determining roughl .
the constants of a skew curve of this type : ( 1 .
) Find the mean and th median ; these gives , approximately .
( 2 .
) Find for the mean ; this gives , since ( 3 .
) Knowing , correct the value of by using the above value for , and so obtain a ) corrected ( 4 .
) Determine from the area .
This method is not very laborious and may be of service in some cases .
It will , of course , fail for any curves in which is negative , and must only be applied when the curve is known to be of Type III .
If the beginning of the range is definitely known , we may save stage ( 2 ) above and find from the distance of the mean from the start of the range .
( 19 .
) bility Curve of Type unlimited , but form skew .
Put , hence , if provided * The points of inflexion may also occasionally be found from the observations ; they are at distances on either side of the maximum ordinate .
MR. K. PEARSON ON THE ATHEMATICAL THEORY OF EyOLUTION .
377 Thus , if we know and , we can find the successive .
Now and depends on the integral , which I propose to write .
The result above for shows us that the more general integral always be expressed in terms of -functions .
Further : Thus we find by the formula of reduction above : Referring to centroid vertical , we have : These may be rewritten , if As before , putting and , we have .
CCXCV .
378 MR. K. ARSON ON NATHEMATICAL OF EyOLUTION .
dividing out by , we have hence is known .
Further is known , whence is given .
and completely determine the problem .
be solution is clearly uniqu To determine the skewness we must find the position of the ordinate for which ; this is .
But Hence skewness ( cf. p. 370 ) .
( ii .
) We further notice thab Hence , since is always ( see p. 369 ) , it follows , since , that we must have or Whcthcl we ve the \mdash ; or sign will depend upon the sign of in the actual statistics .
MR. K. PEARSON ON THE MATHEMATICAL OF EVOLUTION .
379 Thus this expression is again critical for the class of curve with which we are dealing .
We may say that a skew frequency curve will have limited range , range limited in one direction only , or unlimited ange according as is greater than , equal to or less than zero .
Thus the calculation of this expression is the first step towards the classification of a frequency curve given by observation .
( iii .
) It is noteworthy that the values we have obtained for and will be real and possible if .
On the other hand we have quired in our work that should be .
I propose now to return to this point .
So long as the values of both and ] be finite , but the values of and and consequently of and will be infinite if be .
That is to say , the third and fourth moments of the curve about the centroid vertical become infinite .
This is quite conceivable from the geometrica ] standpoint , and various interesting questions , of purely theoretical value however , arise according as and , i.e. , and are both infinite , or and , i.e. , alone is infinite .
The solution we lave given fails in these cases .
We should obtain , however , finite relations between the four constants of the equation to the curve by taking the first and second moments and round the axis of ; we find in this case or , , .
results together with , are theoretically sufficient to determine the four constants and ; they would hardly be of service without very elaborate tables of the functions .
As a matter of fact , we are very unlikely in dealing with actual statistics to meet with cases in which and become infinite , because neither the range of observations , nor the size of the groups observed at great distances from the origin can be infinite .
With finite values of and , it is , however , easy to see that we always obtain from our solution on page 377 a value of , so that the solution is selfconsistent .
( iv .
) It remains to say a few words about the integral Provided , we find a formula of reduction .
Thus the value of the integral from to only will be required for diverse values of .
The integral does not yet appear to have been studied at length or tabulated .
Dr. A. R. FORSYTH* has kindly answered my for a fairly easy method of reducing for purposes of calculation , by sending me the formula ' where is GAUSS'S function such that II .
Taking as definition of that limit of when is infinite , we can reduce the aboye expression to the form Here , sirce 2 can always be supposed to lie between and 2 , when is small a few terms of the product will generally suffice for the calculation of to the degree of accuracy required in statistical practice .
On the other hand when is , i. e. , generally in cases of slight skewness , I find if very nearly .
Hence very nearly .
* ' valuation of wo Dcfinite Integrals ' Quarterly Journal of Mathematics , ' Jannary , 1895 .
MR. K. PEARSON ON THE MATHEMATICAL THEORY OF UTION .
381 ( 20 .
) We have now considered methods for investigating whether a given system of measurements has a limited range , and for ascertaining the degree of skewness of the system .
Analytically , our work may be expressed as follows:\mdash ; The slope of the normal curve is given by a relation of the form The slope of the curVe correlated to the skew binomial as the normal curve to the symmetrical binomial is given a relation of the form .
Finally , the slope of the curve correlated to the hypergeometrical series ( which expresses probability distribution in which the " " contributory causes\ldquo ; are not independent , and not equally likely to giye equal deyiations in excess and defect ) as the above curves to their respective binomials is given by a relation of the This latter curve comprises the other two as special cases , and so far as my investigations have yet gone practically covers all homogeneous statistics that have had to deal with .
Something still more general may be conceivable , but I have hitherto found no necessity for it .
To its fitness and the importance of these generalised frequency distributions for various problems in physics , economics , and , I have devoted the remainder of this paper to the consideration of special cases .
actual statistics .
PART II.\mdash ; STATISTICAL EXAMPLES .
( 21 .
) TELET , who often foreshadowed statistical adyances without the method by which they might be scier dealt with , ] treated of the subject of limits in Lettre XXII of his ' Lettres sir la heorie des Probabilites\ldquo ; ( 1846 ) .
He seems to have been conscious certain variations in excess or defect might biologica ] lie or physically be impossible , and he accordingly introduces the terms Limits ordinaires en plus et en raoins to mark the range of possible variation .
He makes no attempt to show how this range may be found from a given set of " " Lorsqu'on suppose le nombre des observations infini , ou peut porter les ecarts a des distances egalement infinies de la moyenne , et trouver toujours des probabilites qui correspondent .
Cette conception mathematique no peut evidemment s'accorder avec ce qui est dans nature . . . .
Les limits extraordinaires au dela desquelles se trouyent les nonstruosites , me semblent plus difficiles a fixer Indeed QUETELET'S attempt to { ix these limits in the case of the height of human at 2801 and 433 metres is purely empirica and scientifically worthless .
I propose in this the first section of the practical part of this paper to consider how far the theory we have developed in the part , enables us to find the range in various groups of physical and biological phenomena .
I. The Range of the orneter.\mdash ; The following results for curve of barometric heights are given on p. 352 .
10.14 We have accordingly : that is , this expression is positive , and we have a limited range .
We have further : Hence , determining the constants in the manner described , we have : Next to find , giving the distances of the centroid from the igin , or the distance barometer between mean and maximum , we have by p. 370 Thus of barometer above mean below , , Now , in the upon which our curve is drawn in , fig. 6 , each centimetre equals inch , and the mean barometer in Dr. VENN'S results equals about Thus the maximum possible and the ninimum possible ; the range of the barometer being about .
Now , the highest barometer in Dr. VENN'S record , and the lowest ; it is clear , therefore , that we reach much nearer in MR. K. PEARSON ON THE HEMATICAL THEORy OF EVOLUTION .
383 ctice to the upper than to the lower limit of the barometric The result here obtained for the range is of course only tentative and Faar larger statistics must be dealt with , and for a greater variety of places , we shall then be better to judge how far the , as ascertained from Dr. VLNN'S statistics , is local , or if general , what modification or correction may be required .
Calculating value , we iind for the curve of barometric heights : This curve is on Plate 10 , fig. 6 .
It will be seen to be extremely close to the tions .
Although expression 2 is not zero , it is interestin to see what closeness the skew curve which is the limit to a point can be fitted to the barometric observations .
This is the curve of Type III .
ting its constants by aid of , we find while , the distance between the maximum ordinate and the centroid-vertical , .
This gives maximum possible height of the bal.ometer of instead of , there being of course no lower limit .
The curve is shown in Plate 10 , fjob. .
6 , and will be seen to give a very close correspondence with the observations .
The " " skewness\ldquo ; of barometric results as given by the curve with limited range , and as given by the curve of Type III .
\mdash ; no very great difference .
The areal deviations of the two curves are almost exactly the same , being about sq .
centims. or tage error of .
The normal curve is also drawn on same plate .
It diverges widely from the observations , the areal deviation sq .
centims. or the percentage error , \mdash ; about times as great as in the case of either skew probability curve .
Till a ) wider range of barometric observations have been analysed , it may be wiser not to draw too definite conclusions from the above results , contenting ourselves with remark that the new skew curye gives far better results than the old normal curve of errors .
am if results are reduced to sea-level .
The vcst r ) height the British Isles reduced to sea-level is ertyre , Perthshire , January and the ighest ( at Roche 's Point , Cork , February 20 , 1882 ) is .
A statement that the barometer stood at Gordoll Castle , in January .
1820 , has hardly sufficient evidcnce .
Supposing Dr. statisbics to bo um.educed Cambridge tistics , the expression theoretically found for bnrolnctric soems to be on the whole satisfactory .
I at present in hand other serics of romctric hcights .
384 MR. K. PEARSON ON THE MATHEMATICAL THEORY OF EiVOLUTION WELDON'S Measurements No. 4 .
The details of these given in ' Phil. Trans vol. , p. 96 .
We have In this case and is accordingly negative .
In J. of the barometric heights we had Since , in the latter case , this value was sufficiently small to give a good curve of Type III .
we ) expect the like result in this case .
There is , indeed , a slight but sensible skewness even in this the most symmetrical of all Professor WELDON'S crab measurements , and the ske curve of Type III .
is really a better fit than the normal curve .
But dearly since the critical function is negative , we are dealing with a case of a curve of Type .
The ratio of the organs dealt with in No. 4 series of measurements does not give a " " limited range\ldquo ; of variation .
Proceeding by the method indicated in S19 , we find for the constants \mdash ; 21407 , Skewness 077267 , Thus the equation to the curve is : To trace the curve , .
take : If we take a skew curve of Type III .
, we find for its equation : , for the centroid and the skewness For the normal curve we have : was calculated by aid of the on p. 380 .
MR. ON THE MATHEMAT ] THEORY EyOLUTION 385 All three curves are drawn in fig. 4 of Plate 8 .
It will be seen that they are all very close to the observations .
So far as skewness is concerned , curves of Types .
and IV .
practically the same result ( and 077 ) ; in both cases the skewness is small .
The areal deviations are in the three cases respectively : sq .
centims. , sq .
centims. , and sq .
centims. , or we have mean percentage errors in frequency of , and nearly ; the percentage error for the dosest point binomial is We thus conclude that even in a case which has been selected as the most typically symmetrical series of measurements out of a very considerable set of careful statistics , the generalised probability curve is as good again as the normal curve , while the special case of that generalised probability curve\mdash ; which is not the most appropriate to our observations\mdash ; is itself distinctly better than the normal curve .
This result has been confirmed by a considerable application of these generalised curves ; in good cases of normal curve fitting , the generalised curves are always sensibly better ; in cases where the normal curve is almost useless , as in the case of barometric observations , the new curve , if of the appropriate type , represent with a 4 to 5 per cent. mean accuracy man observations not yet reduced to statistical theory .
It is , perhaps , unnecessary to repeat that this mean percentage is much less than the of what has been allowed to pass muster hitherto in both physical and biological measurements .
Professor EDGEWORTH'S vie thus seems untenable ; curve with a comparatively easy theory of its constants has been found which the accuracy of the hitherto adopted normal curve .
And this for the simple reason that it would pass into the normal curve , if that curve were itself ) best fit .
23 .
Example III.\mdash ; The following statistics of for recruits in the United States Army , are given by J. H. BAXTER , 'Medical Statistics of the Provost-Marshal-General 's ' vol. 1 , Plate 80 , 1875 .
78-77 2 77-76 6 76-75 9 75-74 42 74-73 118 73-72 343 680 71-70 1485 70-69 2075 69-68 3133 68-67 3631 67-66 4054 66-65 3475 65-64 3019 64-63 1947 63-62 1237 526 61-60 50 60-59 15 59-58 10 58-57 6 57-56 7 56-55 3 55-54 1 54-53 2 53-52 1 52-51 1 * ' Phil. Mag ol .
24 , p. 334 , 1887 .
MR. K. ARSON ON MATICAL T or EVOLUTION I : Mean height Standard deviation ximum ordinate , This gives a very close-fitting normal curve .
The for a generalised curve are Thus , positive , we see the curve belongs to Type IV .
There is , thus , exactly as in the previous examples of crab measurements , no range of alimited character for these statistics of true curve , ) to be and 3 respectively ; we therefore a Htill closer oach than in the case of the crabs to lity .
is about 400 , and on reasonable scale , there is no sensible between the normal and the generalised curves .
The skewness is very slight , about , or about half its value in the case of the crabs .
24 .
of 2192 St. Louis Girls , 8.\mdash ; The foll-owing statistics given by W. T. PORTER , " " The Growth of St. Louis Children ' Trans. of Acad. of Sci. of .
Louis , ' vol. , p. 279 , 1894 .
The following are the lculated values of the constants :\mdash ; take of ) find the ange limited on 'dwarf ' side 1about unit of theso constants centims. , except in the case of the mean height .
The lion ccntims .
, which rives / probable tion of centims. Tho mean MR. K. PEARSON ON THE MATHEMATICAL THEORY OF VOLUTION .
387 , Mean height centims. , , Standard deviation for normal curve Thus is positive , and the curve is again of Type IV .
We have 135606 , Skewness 04886 , 4 .
or , for the eqnation to ) curve:\mdash ; the axis of being positive towards dwarfs and the origin on the positive side of the centroid-vertical .
The maximum ordinate and occurs at The curve of Type IV .
, together with the normal curve , is drawn ( Plate 10 , fig. 7 ) .
If we attempt to fit a curve of Type III .
, we find about , and the range limited on the dwarf side at about centims. from the mean , or at a height of about centims. The largeness of causes this curve to coincide with the normal curve to the scale of our diagram .
The areal deyjations are for the curve of Type and for the normal curve 6.1 and 8.3 centims. , giving percentage mean errors of and in the ordinates respectively .
The advantage is again on the side of the generalised curve .
It wil ] be seen at once that the normal curve by no means well represents the number of girls of giant height .
The theoretical probability that these giants should occur small , and their actual redundancy over the numbers indicated by the normal curve suggests some peculiarity in this direction ; it is fully lnet by the curve of Type .
The asymmetry of the curves given by anthropometrical measurements on children has been noted both by BOWDITCH and PORTER , but in their published papers , to which I have had access , they do give their material , only the ogive Ctlrye arising from GALTON'S method of percentiles .
UnfortuIIately , theoretical ation of the skewness of anthropometric statistics can only be applied or verified when we have raw material , and not integral frequency and probable deviation , as given by Mr. , are and latter is ined from the mean deviation , but I do not know how the formel ' is to be accounted for .
* ' Growth of Children , studied by GALTON'S Method of Perccntilcs .
' Boston , 1891 , p. 496 .
'Growth of St. Louis Children .
' St. Louis , 1894 , p. 299 .
curves , the integral of the frequency in all suggested forms of the frequency curve being not expressible in terms of ermined constants .
] ] as is the method of percentiles for representing popularly the numerical facts of anthropometry , it is to be retted that percentile statistics are replacing the raw material in so many publications .
The raw material of Professor WELDON'S -measurements BOWDITCH and PORTER'S child-measurements ought to be preserved and in print , as a means of developing and testing statistical theory .
( 25 .
) .
Length-Breadth of 900 ) Skulls.\mdash ; The following statistics are taken from Tables I , VIII .
, inclusive , of J. RANI ' Beitrage sir physischen Anthropologie der Baiern , Miinchen , 1883 .
' They include all the material , which may } ) treated as typically " " Alt-Baierisch , \ldquo ; both male and female skulls .
We find , as before , ition of centroid-vertical , ( for normal curve ) , , 3.707179 , 3.649553 , 12.42734 , 111388 , Skewness 0321186 , Thus we see that the curve is again of Type .
This result seems of considerable , but it requires , of course , wider examination of cases than I have yet been able to make .
But , so as I have gone , in both anthropometric and biological statistics , whether relative or absolute measurements of organs , the frequency curves all deviate the normal curve\mdash ; howeyer slight the deyiation\mdash ; in the direction of Type IV .
That is to say , the distribution of chances upon which the frequency of variation of an organ depends , appears to resemble the drawing of a Indiccs sucl ] been divided veen 7 and 74 groups .
MR. K. PEARSON ON THE MATHEMATICAL THEORY VOLUTION .
389 limited amount from a limited mixture .
So far as this goes , it is evidence against the usual hypothesis that iu biological matters the chances of deyiations on either side of the are equal , and the " " contributory causes\ldquo ; independcnt and indefinitely in number .
Thus we appear in biological statistics to be dealing with a chance system corresponding , not to a binomial , but to a hypergeometrical series , such as that in S 11 .
If it be remarked that Type IV .
dismisses at once the problem of range from biological investigations , we must notice that , while this theoretically correct so long as we are dealing with the continuous by which we replace the ergeometrical series , it is not true the moment we fall back from the curve on the point series ( see p. 361 ) .
If the of that page ( or the ) be an integer , the is limited in range .
It seems very sible that discreteness , rather than continuity , is characteristic of the ultimate elements of variation ; in other words , if we replaced the curve by a discrete series of points , we should find a limited range .
It is the analytical transition from this series to a closely fitting curve which replaces the limited by an unlimited range .
Exactly the same transition occurs when we pass from the symmetrical point binomial to the normal curve .
Thus , while Type I. marks an absolutely limited range , the occurrence of Type .
does not necessarily mean that the range is actually unlimited .
For the equation to the curve we have the origin being at a distance on the positive side of the centroid vertical .
The normal curve as well as the curve of Type .
are shown ( Plate 11 , fig. 8 ) .
The result in both cases is quite good for this type of statistics\mdash ; i. , the skulls came eight rent districts and include 100 female skulls .
With the planimeter deviation in both cases square centims. , giving in either case an pererror of .
That the generalised curve does not in this case give decidedly better result than the normal curve I attribute to the heterogeneity of the material .
It clearly accounts better for the extreme dolichocephalic and brachycephalic skulls than the normal curve .
The same 900 skulJs have been fitted with a normal curve by rIEDA , but neither the constants of his normal distribution nor eserve for the present the fitting of hypergeometrical point sel.ies to statistical .
The discussion is related to curves of Type , as the fitting of point binomials to cnrves of Typc .
It will , I think , throw considerable on the nature of chance in the field of biological va ] tiou , especially with regard to limitation of the material to be drawn upon , to which I referred abo ve , and which , I believe , { inds confirmation in skull statistics .
" " die Anwendung der Wahrscheinlichkeitsrechnung in der anthropologischen Statistik ' Archiv fur Anthropologie , ' Bd .
14 .
Braunschweig , 1882 .
:390 MR. K. PEARSON ON MATHEMATICAL THEORY OF EVOLUTION .
his of 's observations with mine .
He has added together under 83 , for indices frotYl 83 to .
Thus , the indices 81 , 82 , he gives the frequencies 106 , 92 , 111 , 99 , while I find 82 , , 98 , 107 , a very sensible difference .
STIEDA'S method can introduce very sensible errors .
In this particular case it transfers the maximum frequency of observation from 82 to 84 .
The last four examples have dealt with cases where the has hitherto been content to assume symmetry .
They have been given to indicate an apparently uniform trend in ] ogical statistics of tion , and ( ii .
) the improved fitting of theory to practice which arises from the ised curve .
I now pass to of obvious skewness , where the ician has hitherto had no satisfactory theory .
( 26 .
) Vl .
Distribution of 8689 of nteric Fever ceived into the Asylums Fever Hospitals , 1871-93.I considered that the 13 cases " " above 60 " " might be distributed as follows : 60-65 , 8 ; 65-70 , 4 ; 70-75 , 1 .
Taking five years as unit I found The centroid-vertical is at years , Thus the curve is of Type I. Since , however , is small , a curve of Type III .
will also be a good fit .
We have for the other constants 259 .
2.79291 , 3.07801 , , Skewness 488922 , * I class as 83 all from , dividing veen 8 and 83 evenly , and 83.5 between 83 and 84 evenly .
Thus in the Table above certain frequencies will be found lvith such values as or skulls .
MR. K. PEARSON ON THE MATHEMATICAL THEORY EVOLUTION .
391 Thus we have for the curve of Type I. where the centroid is 98643 unit from axis of The curve of Type III .
is The centroid is in this case unit on the positive side of the origin and the skewness It will be noticed that the curve of Type I. extends unit or years , and the curve of III .
5676 unit or years before birth .
In both cases the chances of an " " antenatal\ldquo ; death from entel.ic fever are very , very small .
Curve of Type I. is in respect better than the curve of Type .
The latter curve gives maximum limit , the former a limit of about 77 units or 385 years .
In both cases , however , the nces of a case of entel.ic fever with the )ject over 100 years are nall .
These statistics of enteric fever set a maximum limit to the duration of life , but it is a limit so as to have little suggestiveness .
In order to see what is the nature of the difference made , when we suppose the liabiiity to enteric fever to commence with birth , I will treat these statistics as case falling under S 16 .
If then , and be the first three moments about the vertical through we have , whence we have for the curve Here the duration of life is 200 years about , and } maximum incidence of the disease is at years .
Lastly for the normal curve , we have the constants units years and All the above four curves are drawn .
Plate 12 , fig. 9 .
We see once that the normal curve is perfectly incapable of expressing statistical results like these .
It gives an average error in the ordinate of per cent. and no less than 260 antenatal deaths !
For the remaining three curves we have the following results:\mdash ; osest fbirth .
The percentage errors here are ] within those usually passed by statisticians .
If they are slightly larger than what we have found in previous cases the source of the error is not far to seek .
We have combined both ] and female cases , but the distributions of enteric fever for both sexes are not the same .
The fever curves for eithel sex .
in some cases markedly , less for eric fever than for dipbtheria , for example .
We have thus , in leality , a compound curve .
I have found for about 700 male cases only a percentage error of ) Another point needing notice is the question of antenatal cases , which may at first strike the reader as absurd .
The dosest fitting curve of Type I. runs , as we have seen , about before birth , and gives three antenatal cases .
Three cases indeed , case curve III .
) is a very small percentage of 8689 cases , and not of importance from the statistician 's standpoint .
But the fact that a curve starting before birth gives a better fit than one starting at birth , is significant , and there is every probability that a curve starting from about - year would give a still less centage error than one from year or from birth .
In dealing with mortality culves for infancy I have found it impossible to get good fitting theoretical curves , without carrying these curves backward to a limit of something less than a year .
The " " theoretical ' statistics thus obtained of antenatal deaths seem to be fairly well in accordance with the actual statistics of maternity chsrities .
In vital statistics therefore we must be prepared in most diseases for small percentages of tenaC al cases and antenatal deaths , and it is just possible that theory in this matter will be able to indicate lines of profitable inquiry to the medical statistician .
( 27 .
) Example VII.\mdash ; As an example of the method of Section 15 , I take the following statistics of guessing a tint .
Nine mixtures of bl-ack and white were taken , pose on another occasion to de with the age distribution of fever cases .
lly object at present is on l.y to give typical illustrations of the method of calculating skow curves .
In fact the case of a pregnant woman with enteric fever is to be considered as a also of fever .
MR. K. ON THE MATHEMATICAL THEORY OF EVOLUTION .
393 so as to get a series of tints in arithmetical progression 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , and 9 .
These tints were then placed in non-consecutive order , and 231 persons asked to guess a tint by affixed letters lying between 1 and 9 .
The results were as follows:\mdash ; Now , obviously , the number of tints and the number of persons guessing were far too limited to draw any definite conclusions as to the distribution of tint guesses .
* I propose here merely to use these statistics to illustrate the calculation of a skew frequency curve with a given limited range .
I do not wish to propound any theory of tint guessing , nor to assert that these guesses actually distribute themselves according to the curves dealt with in this paper .
Calculating the moments about the centroid in the usual manner , we have Centroid lies distance ofrom t units We easily find , or the observations fall in to curve of Type I. , that is to say , have a nited We obtain hence the range Further 4.858705 , 099765 , , Skewness Thus the range of the theoretical curve runs from a -point units before tint 1 , and concludes at a point.734674 unit before tint 9 .
The curve is , however .
* I hope latel ' to deal with the subject of tint guesses falling within a limited range , as my material increases in bulk .
I would only note here , that the geometrical mean frequenoy does not to give ding well with experiment .
394 K PEARSON ON THE MATHEMATICAL THEORY hVOLUTION practically inse1lsible before tint 1 .
Considering the roughness of the experimental method , the obtaining an actual range of about 11 instead of 8 , and its covering very nearly the range of 8 must be held to be fairly encouraging for the method .
I shall accordingly calculate the constants of the curve on the assumption that the range lies between Tints 1 and 9 , using the method We find Whence , and Thus we may take for the curve The curve is figured , Plate 11 , fig. 10 , with the first " " smooth\ldquo ; of the observations .
It will be seen to give the general character of the distribution , but much more elaborate experiments would be required before sny statement could be made as to whether frequency of tint guesses really does follow a curve with limited range of I. On the same plate the frequency of 128 guesses distributed over 18 tints is given , the approximation to a curve of Type I. is fairly close considering the of guesses .
( 28 .
) VIII.\mdash ; The question may be raised , how are we to discriminate between a true curve of skew type and a compound curve , supposing we have no reason to suspect our statistics priori of mixture .
I have at present been unable to find any general lJdition among the moments , which would be impossible for a skew curve and possible for a compound , and so indicate compoundedness .
I do not , however , despair of one being found .
It is a fact , possibly of some significance , that the best fitting skew curve to several compound curves that I have tested is a curve of Type I. , and that of Type IV .
which appears to be the more usual type in biological statistics .
ing , as an example , the statistics for the\ldquo ; foreheads\ldquo ; of Naples crabs due Co Professor WELDON , and resolved into their components in my memoir , ' Phil. Trans. ' , vol. 185 , p. 85 , et seq. , I find for the best fitting skew curve the equation the origin at units from in the MR. K. PEARSON ON THE MATHEMATICAL THEORY OF EVOLUTION .
395 positive sense of the horizontal scale .
If , now , we this skew curve and the compound curve of Plate 1 , ' Phil. Trans vol. 185 , of the observations Plate 13 , fig. 11 ) , we see at ce how much better is the fit of the compound curve .
skew curve gives a mean error in the ordinates of , the compound curve of only .
The determination of the best skew curve , when the compound curve is known , is easy , for all its details are already practically calculated .
A criterion of whether a compound skew curve is to be sought for ab initio , would be , however , of great value .
( 29 .
) Example IX.--A more markedly skew than any we have yet dealt with is that giving frequency of divorce with duration of marriage .
I take my statistics from a paper by Dr. W. F. WILLCOX , entitled " " The Divorce Problem , a Study in Statistics\ldquo ; ( ' Studies in History , Economics , and Public Law , ' Columbia College , vol. 1 , p. 25 ) .
They are as follows:\mdash ; Duration of marriage Duration of in years .
Divorces ( 1882-6 ) .
years .
Divorces ( 1882-6 ) .
ftnd over 9401 Total numI ) of divorces granted , 109,966 .
Now these statistics suffer from a defect common to many of the class\mdash ; th want of careful enumeration of the frequencies near the beginning and end of the It cannot be too often insisted upon that careful details of the frequencies in the start and finish of the distribution are site if we are to fit skew distributions with their appropriate skew curves .
How , in this case for example , are we to distribute the 9401 divorces which occur after 21 years of married life How , on the other hand , does the curve start ?
Tt is impossible to place 5314 divorces at the mean\mdash ; 6 months\mdash ; of the one year duration .
It is obvious that the applications for divorce will be far more numerous in the last half-year than the first half-year of matrimony .
The very time required to institute legal proceedings and get a diyorcc granted must ensure this if nothing else did .
Yet these two tails of 5314 and of which the accurate distributions are not given , are between and of the number of divorces , and until we know how they are exactly distributed , we cannot ope for the very exact fitting of a theoretical curve .
MR. K. PEARSON ON THF MATICAL THEORY OF EVOLUTION .
In order to make the best of the " " tails\ldquo ; under the circumstances , their moments were calculated on two hypotheses , that they wsre triangles , ( ii .
) that they were logarithmic curves , and the mean of these extreme results taken .
I found Distance of centroid from start of curve maximum , , , , maximum frequency Here the curve is assumed , owing to the obviously long tail to the and the abrupt start to the left , to be of Type III .
Its equation is , Skewness The curve is figured , Plate 11 , fig. 12 , and will be seen to rise abrupt ] at about 47 of a year 's duration .
It may be doubted whether legal proceedings even in America so rapid that a divorce suit can be complete within six months of marriage .
The iYfes fairly well the general form of the frequency statistics .
Could the moments have beeLl determined with greater accuracy , most probably a better fit vould have resulted .
As it is the mean percentage error is above 6 .
( 30 .
) X.\mdash ; A still more extreme case may be selected from the field of economics .
I take the following numbers from the 1887 Presidential Address of Mr. GOSCHEN to the Royal Statistical Society ( ' Journal , ' vol. 50 , Appendix II .
pp. 610-2 ) .
I have grouped together both houses and shops , because the details of the two are not in Mr. GOSCHEN'S returns separated for values under VALUATION of House Property , England and Wales , years 1885 to 1886 .
Here clearly the curve starts with the maximum frequency , and further to any MR. K. ARSON ON THE OF EVOLUTION .
397 scale to which the curve can be drawn , it tails away indefinitely to the right .
This justifies us in the assumption that the curye will be fairly approximated to by a form of type where would turn out to be a negative quantity lying between and 1 .
But the details given us of the start and finish of the curve are far too scanty to allow us to proceed by moments .
In the first place , to measure an element of area of the frequency curve by an element of value into its mid-ordinate is perfectly legitimate at such a point as ; it fails entirely , however , at such a point as , which Jdes the part of the curve which is asymptotic to the ordinate of maximum frequency .
The area at such a poin is much greater than the element into the mid-ordinate , and the calculation of moments on the assumption that 3,174,806 may be concentrated at , is purely idle .
The ordinate obtained from the area in this manner may often differ 30 per cent. from the true ordinate , and yet about threefifths of the total number of houses fall into this first group .
Further treating the area as ordinate into element of value is also true only if the element of value be small .
For " " elements\ldquo ; such as X150 , X200 , or even S500 , which are all that are given in the tail of these statistics , it is perfectly idle to concentrate the area at the mid-ordinate .
The centroid of a piece of tail such as the accompanying figure suggests lies far to the lsft of the mid-ordinate In other words , to attack the problem by the method of moments , we require to have the " " tail\ldquo ; as carefully recorded as the body of statistics .
Unfortunately the practical collectors of statistics often neglect this first need of theoretical investigation , and proceed by a method of " " lumping together\ldquo ; at the extremes of their statistical series .
Still three further points in regard to the present series of atlstics .
First , they are 398 MR. K. PEARSON ON THE MATHEMATICAL THEORy OF EVOLUTION very unlikely to be homogeneous .
Houses with an annual valuation of over.9300 hardly fffi under the same series of causes as the bulk of houses in the kingdom which fall under X100 .
Secondly , when we are told that 3 , 174,806 houses are valued under , it hardly mean that any houses are valued at , certainly not the maximum number .
Hence our frequency curve in theory must not be expected to rise from zero , but from some point between and , which corresponds to the customary minimum at which a cottage can be rented .
Lastly , there is one special cause at work tending to upset , about the value of the general distribution due to a great variety of small causes .
This is the value at which taxation commences , and we should expect a larger proportion of houses to be built just undel .
the taxable value than is given by a chance distribution .
Notwithstanding the many disadvantages of these , I determined to obtain if possible a skew curve approximating to the main portion of the distribution .
I took as my unit of value and 1000 houses as my unit of frequency .
I started with the ordinary method of moments , concentrating each area at its centroid as given by the total valuation of the group , also recorded by Mr. GOSCHEN , and found a curve of the type with This was so far satisfactory that it showed even by this rough method that was ative , and between and 1 .
Thus the theoretical curve gave an infinite ordinate , but at its start .
A laborious method of trial and error was then adopted , and by varying and slightly , as well as and the origin of the curve , I sought to improve the fit given by the rough method ( in this case ) of moments .
The fundamental consideration was to keep the total areas under SI00 value as neally as possible the same in the curve and the statistics .
portion of the curve I treated practically referring to eneouso material .
Ultimately I following curve : with the origin as 45 unit from zero .
Thus the minimum annual valuation was , or , to a weekly valuation , of .
This would connote probably a weekly rental of ls . .
to .
The total area of this theoretical curve was 5795 in thousands of houses ; of 5729 had a valuation under,9100 and 66 over , ; the corresponding numbers for the themselves are 5720 and 110 .
The additional 44 over XI00 I assume be due to the heterogeneity of the statistics\mdash ; high values corresI)onding to blocks of chambers , large hotels and other buildings hardly falling into the same cat , as the small house under X100 in value .
Unfortunately the " " tail\ldquo ; of the statistics is so defectively recorded that there is no hope of reaching a separate distribution for this high class property .
MR. K. PEARSON ON THE MATHEMATICAL THEORY OF EVOLUTION .
Returning now to the curve and statistics , we have the following comparative results :\mdash ; The general accordance here is very marked , the chief divergences being accounted for by the special causes to which we have referred above , i.e. ( i ) the crowding of houses just below the limit of taxation , and ( ii ) the divergent character of the causes at work determining the frequency of low and high class house property .
The results are depicted , Plate 14 , fig. 13 .
It will be observed that so far as the observations can be plotted to the theoretical curve , it leaves little to be desired .
The shows , however , the amount of deviation at the extremes of the curve .
( 31 .
) .\mdash ; Frequency curves of the type considered in Example X. are so common that it is needful to make a few further remarks with regard to them , and illustrate them by fu1ther examples .
Such curves occur in many economical instances ( income tax , house valuation , probate duty ) , in vital stabistics ( infantile mortality ) , and not uncommonly in botanical statistics of the frequency of variations in the petals or other characteristics of flowers .
As we have noted , the method of moments developed in this memoir cannot be directly applied , or only applied to obt , ain a first approximation to the constants required .
Tbis first approximation , however , will often assist us to obtain with quite sufficient accuracy the value of the moments of portions of the area , especially if the position of the imtial or asymptotic ordinate is known .
For example , consider the curve of limited range : where lies between and 1 .
Then if be its area , moment about the asymptotic ordinate of the area up to : troduced by the writer in his lectures on atistics as a term for a ommon form of graphical opresentation , i.e. , by columns marking as frequoncy to the of Hence , if the range be large and x be small , this series converges very rapidly , and we may often take with sufficient approximation even only its first term .
Thus early .
Now is given by the statistics , and we note that if has been determined to a first approximation by the method of moments , we can now improve the values of the moments of the near the asymptotic ordinate by the use of above For example , if as a first approximation , we have Concentration along the mid-ordinate in the usual manner would have given us as the up to a short distance from the asymptotic ordinate is generally a considerable proportion of the total area , the ) values very considerably modify calculated moments .
In the case of the curve we have the result Hence , as before , if and be small ] , MR. K. PEARSON ON THE MATHEMATICAL OF EVOhUTTON .
401 , approximately .
* Results such a , s the above enable us to approximate fairly rapidly to the constants of a frequency curve .
As a special example , I take the following .
In 1887 , .
H. VBIES transferred several plants of Ranunculus bulbosus to his flower garden , and counted the petals of 222 of their flowers in the following year .
He found Berichte deutschet ] botanischen Gesellschaft , ' Jahrg .
12 , pp. 203-4 , 1894 ) Petals 5 6 7 8 10 Frequency .
133 55 23 7 2 Now the series here proceeds by discrete units , and corresponds probably to a hypergeometrical series , but remembering how closely the results of tossing ten coins call be represented by a normal frequency curve , I was not without ) that the areas of a skew frequency would give results close to these numbers .
buttercups start with 5 petals and run to 10 , I therefore took my origin at and determined the constants to a second approximation in the manner above indicated .
There resulted , a curve of Type I. , with limited range , the asymptotic ordinate being petals , or practically a distribution ranging from 5 to 11 petals .
Calculating the areas , there Jllts , . . . . . .
5 6 7 8 9 10 11 Frequency ) The agreement here is very satisfactory considering the comparative pauci of the observations .
The results are exhibited by curve and histogram , Plate 15 , fig. 14 ; the two points on the " " observation curve\ldquo ; corresponding to five and six petals are deduced from the areas given by the stat , istics by the same percentage reduction as * Another very serviceable formula is due to NlLCTI .
It ives the area of the " " tnil " " of from to in a rapidly converging sericf ; , area } 2048 tosses of 10 f:hillings at a time gave a mean 3 per oent .
deviation between eory , 100 tosses gave oout 9 per cent. The above series corl.esponds bo about per ccll t thus is quite within the range of accuracy of coin-tossing experiments , converts the theoretical areas into the ordinates of the theoretical curve .
For other petals , ordinates and areas practically coincide in value .
( 32 .
) XII.\mdash ; Another example of a similar kind may be taken from Herr DE VRIES ' memoir .
He cultivated under the name of perumbellatum a race of Trifoliura repens , in which the axis is very frequently prolonged beyond the head of the flower , and bears one to ten blossoms .
In the summer of 1892 he had a bed of such clover , produce of a single plant , and in July counted the extent of this variation on 630 flowers .
In 325 cases the axis , according to VRIES , had not grown through the head of the flower , in 83 cases it had grown through and bore one blossom , in 66 cases two blossoms , and so on .
The complete statistics are follows :\mdash ; High blossomfi 1 2 3 4 5 6 7 8 9 10 Frequency .325 83 66 36 36 18 7 6 1 1 Taking moments in the manner of the earlier part of this memoir , I found as a first approximation to the frequency curve : with the origin at 47813 to left of maximum ordinate .
This first approximation seemed to justify three things : starting to the left of the maximum ordinate ; assuming a range , 11 , which just coyered the whole series of observations , i.e. , from to ; and ( iii .
) that the of the areas might be found from a value of not far A second approximation was then made , and taking moments round the asymptotic ordinate , I found : whence , in the manner of S1C ) , we have : and ultimately : and The equation to the frequency curve is therefore : The value found for , i.e. , 493 , justifies our calculation of the moments on the assumption that it was5 .
MR. K. PEARSON ON THE ICAL THEORY OF EVOLUTION .
403 Placing statistics and theory side by side , we have : 1 2 3 4 5 6 7 8 9 10 Statistics 325 83 66 51 36 36 18 761 Theory 52 The agreement between theory and observation is here all that could be desired , except in the case of and 1 high blossoms .
Here 22 blossoms have in actual counting been transferred from the theoretical group of 1 to the theoretical group of zero high blossom .
I consider it highly probable that the theory here better results than the actual statistics ; and this , for the simple reason that it must be very difficult to distinguish between any one of the low blossoms and a very slightly extended axis bearing only one blossom , that is to say , the extension of the axis passes insensibly into one of the low blossoms , or vice , and in a certain proportion of cases it must be diflicult to distinguish between the categories and 1 .
The comparison between theory and observation is represented by curve and histogram , Plate 15 , fig. 15 .
Examples X. to XII .
will suffice to illustrate the application of our theory to extreme cases of skew distribution .
( 33 .
) XlII.\mdash ; It must not be supposed that in every case of variation units ( as in the buttercup and clover examples ) , the curve will be found to be of Types I. or .
It is impossible to illustrate , in anything short of a treatise on statistics , the infinite variety of statistical distributions , but the occurrence of Type IV .
in zoological , as distinguished from botanical measurements , is so persistent that it seems well to illustrate this for the special case of discontinuous variation .
Professor WELDON has kindIy given me the following statistics of dorsal teeth on the rostrum of 915 8 and 9 specimens of Paloemonetes from Saltram Park , Plymouth .
The centroid-vertical here lies of a tooth beyond 4 , i.e. , at teeth .
The following are the moments about centroid-vertical:\mdash ; the umit the normal curve these give Standard deviation 9544 , Maximum ordinate the skew we have Hence \mdash ; .122702 , or , we have a curve of Type IV .
The values of and , however , show that will nol differ very widely from the normal type .
Proceeding to determine the other constants we find ( is negative smce is positive ) , Distance of origin from centroid-vertical Thus { an give the of the curve .
This curve , the normal curve , and the tions are drawn , Plate 13 , fig. 16 .
A comparison of the observations and the normal curve shows an amount of skewness in the tails of the former , which would be very improbable if the normal curve really the distribution .
The skew curve really accounts for this divergence and is a sensibly better fit .
The mean percentage errors in the ordin ates are for the two cases and curve is thus an excellent fit .
The discontinuity in these teeth probably corresponds to a hypergeometrical polygon , of which the skew curve is a limiting form .
( 34 ) .\mdash ; Another extremely interesting illustration of skew variation will be found in the statistics of pauperism for and Wales , to which my attention was by Mr. G. U. YULE , who Iotted the statistics from the raw material provided in Appcndix I. of Mr. CHARLES BOOTH'S ' Aged Poor ; Condition ' In Plate , we unions distributed over a range of pauperism varying from 100 to per 10,000 of the population for the yeal 189 ] .
The observ tions MR. K. PEARSON ON THE MATHEMATICAL OF EVOLUTION .
405 are at once seen to give a markedly skew distribution .
Taking 50 paupers as unit of variation , we find .
Hence or the curve is of Type I. The other constants were found to be 281650113 , 20 .
69714 , 99.9065 .
7.199312 Range Maximum to left of centroid vertical .
Skewness The equation to the curve is thus For the normal curve , Standard deyiation Maximum ordinate Both skew curve and normal curve are drawn on Plate 14 , fig. 13 .
The former is at once seen to be an excellent fit .
We might fairly have simplified our work by taking zero paupers as the commencement of our range , but preference was given to the more general results in order to demonstrate that they give no amount of " " negative pauperism The range determines a limit of about 15 per cent. as the greatest possible amount of pauperism .
The normal curve is seen to diverge very widely from the statistics besides giving an appreciable amount ( 3 to 4 unions ) with " " negative pauperism The point-binomial for these statistics is also figured on the plate .
Its constants are , the start of the binomial being to the left of the centroid-vertical : see S 5 .
The fit is a very close one , the mean error of ordinate , and the suggestiyeness of such results for social problems needs no emphasising .
The case is of peculiar interest , because the statistics of pauperism are known to give a definite trend to the distribution , i.e. , if the statistical curve of pauperism for 1881 be compared with that of 1891 , for example , the maximum frequency of the earlier will be found at a much higher percentage .
The whole frequency curve is sliding across from right to left .
Now it is of interest to notice that in this , as in other cases where the trend of the variation is known a priori , the skew curve is shifted away from the normal curve in the direction in which variation is taking place with lapse of time .
It is not safe at present to extend this to all biological instances , but the result suggests , for example , that there is a secular progression towards brachycephaly in Bavarian skulls ( fig. 8 ) , towards reduced antero-lateral margin in crabs ( fig. 4 ) , towards increased height in St. Louis school-girls ( fig. 7 ) , and towards longsightedness in Marlborough School boys .
* I believe most suggestive and important results might be obtained for the theory of evolution , if we only had the series of skew curves for a biological case of progressive variation in the same manner as we have for pauper percentages .
( 35 .
) Example XV .
The theoretical resolution of heterogeneous material into two components , each having skew variation , is not so hard a problem as might at first appear , and I propose to deal at length with the subject later .
If there be more than two components , the equations become unmanageable .
In this case however , if the components have rather divergent means , a tentative process ] often lead to practically useful results .
To illustrate this I propose to conclude this paper by an example of a mortality curve resolved into its chief components .
By a mortality curve I understand one in which frequency of death ( for 1,000 , 10,000 , or 100,000 born in the same year ) is plotted up to age .
I have worked out the resolution for English males , and for French of both sexes .
The generally close accordance the results for both cases has given me confidence in their approximate accuracy .
The method adopted was the following : An attempt was made to fit a generalised frequency curve to the old age portion of the whole mortality curve , the constants of this curve being determined from the data for four or five selected ages by the method of least squares ; the frequency curve so determined was subtracted from the total curve .
and a frequency curve fitted by the same method to the tail of the remainder .
This second component was again subtracted and the process repeated , until the emaindel .
left could itself be expressed by a single frequency curve .
The components thus obtained were added together , and a tentative process adopted of slightly modifying their constants and position , so that the total areas of components and of the whole mortality curve coincided .
It was soon obvious that no very great change either in the constants or position was permissible , if the sum of the components was to give the known resultant curve , hence I feel very confident that whatever be the combination of causes which result in the mortality curve , that curve is very approximately to be considered as the compound of five types of mortality centering about five different ages .
The allied character of the results obtained for both French and English statistics confirms this view .
* Dr. ROBERTS ' statistics , which I have reduced to skew curvos , but have not eproduced in this memoir .
MR. K. ARSON ON THE MATICAL T OF VOLUTION .
407 Professor LEXIS has already suggested that the old age distribution of mortality is given by a normal curve .
* Now , although the rougher French statistics give a fair approximation to a normal curve , this is not true for English males .
The curve for old age is of Type I. , but for all practical purposes it may be treated as one of Type III .
Whatever be the chief causes of old age mortality , they extend very sensibly through middle life , and less sensibly through youth , only becoming inappreciable in childhood .
Hence , if we speak of our first as the " " mortality of old age the name is to be understood as referring to a group of causes especially active in old age mortality , but excluded from other portions of life .
The second and third components I found to be skew curves , but so nearly normal that to my degree of approximation no stress could be laid on the skewness obtained .
The fourth component was a markedly skew curve , also closely given by a curve of Type III .
, and corresponding in general shape to the mortality curves of fevers peculiarly dangerous in childhood ( e.g. , diphtheria , scarlet fever , enteric fever , \amp ; c These three components I have termed respectively the mortality of middle Jife , of youth , and of childhood .
I found it impossible to fit the remainder of the original mortality curve with any type of generalised curve , so long as I supposed the mortality frequency to commence with birth .
I was therefore compelled to suppose the set of causes giving rise to " " infantile mortality\ldquo ; extended into the period of gestation , and I obtained a satisfactory fit for the infantile mortality frequency , when the range of the curve started about75 of a year before birth .
The form taken by the curve is the extreme type in which the curve is asymptotic to the ordinate of maximum frequency ( cf. Examples X.-XII .
) .
The five fundamental components of the mortality curve for English males are the following , the numbers referring to 1000 contemporaries , or persons born in same year : ( A. ) Old Age Mortality .
Total frequency Centroid-vertical at 67 years .
Maximum mortality at years .
The equation is the axis of being the maximum ordinate and the positive direction of towards age .
The skewness of the curve , and its range concludes at years .
The corresponding French component , but the maximum mortality occurs at years .
* Sir Theory der Massenerscheinungen in der Gesellschaft , ' S 46 .
1877 .
Unit of year , unit of death per year .
( B. ) Mortality of JIiddle Total frequency Centroid-vertical at years .
Maximum mortality The curve is very approximately normal , and has standard deviation of years .
The corresponding French component deaths , dard deviation years , with a maximum of 6 at 45 years .
( C. ) Illortality of Youth .
Total frequency Centroid-vertical at years .
Maximum mortality The curve is very approximately normal , with a standard deviation of The corresponding French component gives a total mortality of 78 , standard deviation of 6 years , and a maximum of at The greater and more concentrated French mortality of youth is noteworthy .
( D. ) Mortality of Childhood .
Total frequency Centroid-vertical at years .
Maximum mortality at 3 years .
The equation to the curve , the axis of being maximum ordinate , is Thus the skewness of the curve , and the range commences at 2 years .
The French component appears to be shifted further towards youth .
It gives a total of 47 deaths , centroid at years , and a maximum of at years , skewness .
Childish mortality is the , refore , if these results be correct , more vated , and at an earlier age in England than in France .
( E. ) Infantile Total frequency birth Maximum frequency after birth occurs in first year and equals The equation to the frequency curve is the being at birth , the skewness the centroid at 083 year , month neally , before birth .
Taking the corresponding French component , we have a total frequency after birth of 284 , with 186 deaths in the first year of life .
Infantile mortality is therefore considerably greater in France .
* The mortality of youth would be better expressed by a curve of type : see our S .
MR. K PEARSON ON THE MATHEMATICAL THEORY OF EVOLUTION .
409 If we investigate the areas of our infantile mortality curve , we have the following deaths : 1st year of life . . .
2nd year of life After this the mortality of childhood begins to sensibly increase the infanbile mortality .
Turning to the " " antenatal\ldquo ; portion of the curve , we have the following results , of course not verifiable from ordinary mortality statistics:\mdash ; The total " " antenatal\ldquo ; deaths for the 9 months preceding birth are 605 for every 1000 actually born and registered .
( ii .
) " " Antenatal\ldquo ; deaths for the 6 months immediately preceding birth are 214 for every 1000 born .
( iii .
) " " Antenatal\ldquo ; deaths the 3 months immediately preceding birth are 83 for evely 1000 born at the proper period .
The 391 " " deaths\ldquo ; of the first three months of pregnancy would not be recorded , and in many cases possibly pass without notice .
The 214 deaths of the remaining six months would be considered as miscarriages or still-births .
The proportion of 1 in 6 of such accidents to births of the normal kind does not appear excessive .
On the average Dr. GALAPIN says such an occurrence is " " the experience of every woman who has borne children and reached the limit of the child-bearing age So far then there appears nothing to contradict our theoretical results in what is known of the first six months of antenatal life .
For the lasf , three monbhs we have more deiinite data .
According to our curve we have 83 deaths ( per 1000 born ) in the last three months before birth , or 83 in 1083 pregnanciesabout per cent. Now this must consist of two factors\mdash ; still-born children and children born before their time , die shortly after birth , and who would not be recorded in any proper proportions in statistics based on census returns , nor as a rule in the returns of maternity charities .
For statistics of still-births , I find : per cent. Dublin Rotunda Hospital ( lS47-54 ) . . . . .
( l871-75 ) . . . . .
: , Dr. J. H. DAVIS for 14,000 births for a large maternity charity in St. Pancras Guy 's Hospital Lying-in arity , 25,777 births , 1,127 born dead or died within a few hours , 1000 corresponding to births in the last three months of pregnancy .
3.84 NEWSHOLME'S ' Vital Statistics\ldquo ; ( no authorit , y cited ) .
It would thus ) pear thab there are 4 to 5 per cent. of still-births , thus leaving to per cent. of deaths to be accounted for\mdash ; if there is any validity in our analysis\mdash ; by deaths of children born before their proper time and dying before their proper birthdays .
Such deaths would not appear in the category of still-born children in the returns of the maternity charities , nor in any true proportion in the census returns .
Thus , while it is impossible to assert any validity for the antenatal part of our curve of infantile mortality , while , indeed , the constants of that curve , and consequently the percentages of antenatal deaths , might be considerably modified had we surer data of the actual deaths in the first year of life ; still there appears to be nothing wildly impossible in the results obtained , and they may at any rate be suggestive , if only as to the nature of those statistics of " " antenatal\ldquo ; deaths , which it would be of the greatest interest to procure .
The absolute necessity of skew curves in all questions of vital statistics is sufficiently evidenced in this resolution of the general mortality curve .
A complete picture of the resolution into components of the mortality curve is given ( Plate 16 , fig. 18 ) , with a separate figure on an enlarged scale of infantile mortality .
( 36 .
) In conclusion , there are several points on which it seems worth while to insist .
The normal curve of errors connotes three equally important principles : An indefinitely great number of\ldquo ; contributory\ldquo ; causes .
( ii .
) Each contributory cause is in itself equally likely to give rise to a deviation of the same magnitude in excess and defect .
( iii .
) The contributory causes are independent .
The frequency of each possible number of heads in repeatedly throwing several hundred coins in a group together , practically fulfils all the above three conditions .
Condition is not , however , fulfilled if a number of be thrown or a number of teetotums of the same kind be spun together .
Condition ( iii .
) is still fulfilled .
Condition ( iii .
) is not fulfilled if drawn out of a pack of cards containing equal suits , supposing the cards to be drawn at one time .
Now , it appears to me that we cannot say priori whether the example of tossin , of teebotumspinning , or of card-drawing is more likely to fit the proceedings of nature .
There is , I think , now sufFicient evidence to show that the conditions to ( ui .
) are not fulfilled , or not exactly fulfilled , in many cases\mdash ; in economic , in physical , in zoometric , and botanical statistics .
We are , therefore , justified in seeing what results we shall obtain by supposing one or more of the above conditions which lead to the normal curve to be suspended .
The analogy of teetotums and cards leads us to a system of skew frequency curves which in this paper have been shown to give a very close approximation to observed frequency in a wide number of -an approximation quite as as the writer has himself obtained between theory and lment in very wide experiments in tossing , card-drawing , ball-drawing , and MR. K. PEARSON ON THE MATHEMATICAL THEORY OF EVOLUTION .
411 lotteries .
But the introduction of these skew curves leads us to two portant conclusions : \mdash ; If a material be heterogeneous we have no right to suppose it must be made up of groups of homogeneous material each obeying the law of distribution .
Each homogeneous group may follow its own skew distribution .
( ii .
) If material obeys a law of skew distribution , the theory of correlation as developed by GALTON and DICKSON requires very considerable modification .
We may note two points bearing on these two conclusions , which do not seem without interest for the general problem of evolution .
Feyer mortality curves are skew curves .
The general mortality curve\mdash ; frequency of death at different ages\mdash ; is a compound of many diseases , but with sufficient approximation , it can be , resolved into five components ; three of these components are markedly skew , the other two less so .
Selection , locording to age , is thus distributed with different degrees of skewness about five stages in life ; this at least suggests that selection to the size .
weight of an organ may be compound , if we take a considerable range of size , and that the components may have varying degrees of skewness .
The correlation of the ages of husband and wife at marriage is a subject with to which we have a very fair amount of .
For a given age of the husband , the frequency of marriage with the age of the wife fits very closely a curve of Type , and with sufficient exactness very often a curve of Type III.$ The sections of the surface of frequency are oval curves differing entirely from the ellipses of the theory , but resembling in general the " " oval\ldquo ; polygons obtained by taking horizontal sections of the frequency polyhedron for the correlation of cards of the same suit in two players ' hands at whist .
Plate 9 , fig. 19 , shows how widely these differ from ellipses .
There seems therefore to be considerable danger in assuming in vital statistics , whether in or the lower animals , that the " " contributory\ldquo ; causes are independent .
All the statistics for sizes of organs in ammals , which I have yet analysed , if they are not compound , seem to agree in following a curve of Type , and suggest this kind of inter-dependence of the " " contributory\ldquo ; causes .
Their correlation surfaces of frequency will thus have for lines of level skew what for want of a better name may be termed ' whist ovals\ldquo ; as distinguished from the ellipses which flow from the normal frequency surface .
The from quite a different standpoint of RANK on skull measuren } ents seem to lead to the same conclusion .
I propose on another occasion to illustrate the resolution of compound curves into skew components , and further to deal with the main features of correlation in cases of a skew frequency distribution .
* I have fitted some of PNROZZO'S marriage statislics wibh skew curves , but resve their ( liscussion .
the , as they belong properly to the theory of skcw correlation .
NOTE .
Added May 24 , 1895 .
[ Since Wl.iting the above memoir my attention has been drawn to a note in Dr. WESTEBGAARD'S " " Theorie der Statistik referring to Professor T. N. THIELE'S treatment of skew frequency curves .
I have procured and read his book , ' Forelaesninger over Almindelig Iagttagelseslaere , ' Kj .
1889 .
It seems to me a very valuable work , and is , I think , suggestive of several lines for new advance .
It does not cover any of the essential parts of the present memoir .
Dr. THIELE does indeed suggest the formation of certain " " half-invariants , \ldquo ; which are functions of the highermoments of the observation\mdash ; quantities corresponding to the \amp ; c. , of the above memoir .
He further states that a study of these halfinvariants for any series of observations would us with information as to the nature of the frequency distribution .
They are not used , however , to discriminate between various types of generalised curves , nor to calculate the constants of such types .
A method is given of expressing any frequency distribution by a series of differences of inverse with arbitrary constants .
Thus if and we can express any law of frequency by , where the constants can be determined numerically when the frequency of chosen derivation-elements is known .
1 see a possibility of more than one theoretical development of interest , especially in relation to compound material , from this development of Dr. THIELE'S , but I doubt whether it can be of practical statistical service even as an empirical expression for frequency .
Instead of having the 3 to 5 constants of our generalised curves , the full value of Dr. expression as many constants as there are recorded frequencies , and then expresses the result in functions like , by no means easily ised or likely to appeal to the practical statistician .
It is true the complete series gives absolutely accurately the frequency of all the points used in the calculation , but it does not , like the generalised curves , indicate the purely accidental variations of the frequency .
If , on the other hand , we take , as Dr. THIELE suggests , some halfdozen terms only of the which give the really essential character of the MR. K. PEARSON ON THE MATHEMATICAL THEORY OF EVOLUTION .
413 frequency\mdash ; we obtain results which , although more complex in form , are not as satisfactory as those given by the generalised curve .
For example , Dr. THIELE gives the following series His " " Faktiske Fejllove\ldquo ; gives .
He tells us that 6 terms practicaliy suffice , the additional terms merely accounting for the individual irregularities of this particular 500 observations .
Without specifyin what the observations are , he tells us that the possible values run from 4 to 28 , or that the range is really limited .
If we fit our generalised curve of Type I. , we find for its equation : the origin is at , or the range runs from 6.6715 to , i.e. , is a range of instead of 25 , but is shifted some 2 to 3 units .
Considering the small nulnber of observations , this is not a bad approximation to a marked feature of the distribution not indicated on the surface by the observations , nor discoyerable from the " " Faktiske Fejllove Comparing our curve with ( ii .
) the actual statistics\mdash ; all 13 terms of the " " Faktiske Fejllove\ldquo ; series , and with ( iii .
) the first 6 terms of the same series , we have the following results:\mdash ; The generalised curve here gives slightly the better results in addition to its more easily realised form , and its fewer constants ( iv .
) .
On the other band , there are , I think , some points of first-class theoretical importance in the mode adopted by Dr. THIELE for expressing frequency ; it gives us a means of expanding all varieties of frequency curves in a series of ffictorial functions which may Jead to important theorems in the analysis of heterogeneous material .
] PLATES .
The scale of the nying figures is not that of the original drawings , and the clearness and distinctness of the several curves of the same figure have been , in several instances , partially lost by the process of reproduction and reduction .
In every case the square element of the figure corresponds to the square centimetre of the original diagram , and is spoken of both in the text of the memoir and on the figures themselves as a square centimetre .
The scale of actual reduction is indicated by a fraction placed at the lower right-hand corner of the figure .

