]\gt ; 4-13 ] XI .
VIathematical ( to the of .\mdash ; X. Supplement to a Iemoir on Skew Variation .
* KARL PEARbON , F. R. S. , University Received May 22 , \mdash ; Read , June , 1901 .
( 1 .
) a memoir on Skew Variation published in the 'Phil .
, vol. 186 , , a series of curves are discussed which are integrals of the differential equation . . . . . .
( See p. 381 of the memoir .
) The discussion of four main types is given in detail , and a brief reference is made to various sub-types which may occur .
The types considered in that memoir covered at the time all the frequency ries , and they were fairly numerous , that I had had occasion to deal with .
In the course of the last few years , however , I have been somewhat puzzled by frequency distributions for which the criterion ( see p. 378 ) was positive , and therefore priori a curve of the type was to be expected , but which on calculation gave imaginary .
The frequency distributions in question arose occasionally in sociological statistics , but also in 'Phil .
Trans , vol. 186 , p. 343 .
Some other frequency distributions , which on first investigation fell under Types .
and .
of the present paper , were with improved values for the moments to fall uIldcr types already discussed .
Mr. W. F. SHEPPARD'S values for the moments ( Lond. Math. Soc vol. 29 , p. 369 , foImnla 30 ) should certninly be used in preference to those given by me Phil. Trans whenever we aro calculating the moments of a curve from areas and not from ordinates .
I hope shortly to publish a paper on this point , which is one really of quadrature formulae .
Meanwhile for every true frequency curve unfh high ronfacf bofh tcrminals we ought to use , instead of the values given on p. 350 , unchanged .
( 297 ) 39.11.1901 biological investigations .
It seemed , therefore , desirable to enter a little more fully into the analysis of the cases in which the criterion was positive but imaginary , and discover what types of frequency curves had escaped my attention .
* The key to the solution lies in the fact noted on p. 369 of the memoir , namely , that even if the criterion be positive , there will still be a solution akin to Type I. and not to Type Iy .
if be negative .
No frequency series satisfying these conditions had at that time come under my notice , and later , when collecting data of floral variability , my own remark as to had slipped from my memory .
It is the object of this supplement to obtain an improved criterion of type , to discuss the nature of the curves which fill the gap observed , and to illustrate by one or two examples the fitting of such curves to actual statistics .
( 2 .
) The Two riteria .
Throughout this supplement the notation of the previous memoir will be assumed to be familiar to the reader .
Turning to p. 378 of that memoir , we note that since and are necessarily positive , if positive must be Hence can only become imaginary if be negative , or Substitute in this the value of ) it becomes .
Hence the complete condition that a curve of Type IV .
shall give the distribution of frequency is not only but also Turning back to p. 369 , we see that being positive the complete conditions for curve of Type I. giving the distribution of frequency are * I was very loath to dopt Professor method of invenling new frequency curves by putting in a normal frequency distril)ution , Besides strong theoretical objections to this process , I had found Equation so sufficient for a great variety of cases that I felt confident it must cover the newly discovered outstanding oases , and this confidence seems justified by the resuIt .
CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
and The latter condition will always satisfied since and are positive any distribution whatever , and is negative by hypothesis .
Further , in the previous case is seen to be essentially positive .
Hence the criteria written down cover all possible cases but those for which Sub-cases which arise from transition curves just at the limits will , however , be likely to be of interest , What happens when and when ?
The only possibility for is , or .
But this curve has been fully treated under Type III .
in the memoir .
We shall see later that leads us up to a novel transition curve of considerable interest .
To ascertain something about the general case in which , let us return to the memoir again and examine the value of on p. 369 .
It can only be negative if be where is here Substituting , we find at once which in itself inyolves Hence the gap corresponds to those cases in which is negative .
It will be clear that , although in form giving a more complex criterion than is really more effective , as covering all the possible cases .
We have then the following scheme : \mdash ; Transit .
Type Vorresp , The object of this supplement is to discuss the calculation of curves of Type V. , to consider those of Type VI .
somewhat more at length , they being only briefly referred to on p. 369 of the memoir .
It will be seen that Type I. of the memoir has broken up into two divisions .
One portion is the old Type I. passing into the llormal curve on one side and Type III .
on the other .
This Type III .
separates the second portion , Type VI .
, of the old Type I. from the first portion .
Type VI .
passes from Type III .
to the new tiion curve Type , which , like Type III .
, will be found to have a range limited in one direction only .
Finally this new Type V. is the transition to the old Type IV .
bounded on the other side by the sub-curve , the old II .
, and beyond that the normal curve .
Thus we see that Types I. and IV .
do not pass directly into each other through Type III .
, as might be supposed by the criterion , but that there are a series of intervening curves , two of which , Types V. and VI .
, require further consideration , if we are to complete the whole round of frequency distributions embraced under the differential equation .
( 3 .
) On the Curve of Type Returning to the fundamental differential equation , let us consider what transformation takes place when the denominator on the right has equal roots .
* We may then write it in the form Hence \mdash ; const .
Thus where , is a constant , and .
Thus changing the origin we may write the curve : . . . . . . . . . .
( iii .
) , where gives the distance of the mode from the new origin .
To find the moments about this origin , we notice that , and being positive , when and when .
Thus as in the curve of Type III .
we have a range limited at one end only .
To find the moments we have , if be the area , .
I owe to Miss AGNES KELLY , Ph. D. , the suggestion that this type of frequency curve deserved fuller ireatment .
CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
Put , and we find . . . . . .
Thus : . . . . .
( vi ) . .
Transferring to the centroid we find ( viii .
) . .
Eliminating between and we find after some reductions : or , . . . . .
( xi .
) .
Clearly , since this is the condition for Type V. , that transition curve is none other than the curve obtained by making the denominator of the right-hand side of the differential equation have equal roots .
The curve is clearly of considerable interest , and its existence had not been noticed in the previous series of frequency curves .
The manner of fitting it is now easily described .
Equation ( ix .
) gives us a quadratic to find : \mdash ; ( xii .
) .
The positive root of this is the required solution .
is then found from the first of equations ( viii .
) , or if be the standard deviation , then . . . . . .
( xiii .
) .
: . . . . .
, which determines the remaining constant for the shape of the curve .
For the position of the curve , we have for the distance from origin to mean , from the first of equations : . . . . . .
If be the distance from mode to mean we have : \mdash ; ( xvi .
) .
Further , the skewness : Sk .
( xvii .
) .
Thus the solution is completed .
( 4 .
) On the Curve oj ) Type , as we have seen , corresponds to the case in which Type I. of the memoir has its negative .
Hence either or is negative and the curve transferring the origin takes the form . . . . .
xviii .
Now it is possible that this curve falls under the limited range type of a frequency from to , but as we see that the criterion places Type VL between two curves of range limited in one direction only , we expect Type VI .
also to be of that character , and a complete solution is obtained by taking the range from to ; this indeed fills up the gap for and , and ( xviii .
) with this range is seen to pass into one or other of the two transition curves or according as we allow the first or second factor to approach a limit .
* The sign of will determine the sig of , or , what may be taken as the same thing , the direction of the axis of Write : const .
) and make , and but finite .
Or , const .
) , and make together with finite .
CONTRIBUTIONS TO THE THEORy OF EVOLUTION .
Accordingly we shall write Type VI .
in the form . . . . .
( xix .
) , and take the range from to Differentiating to find the position of the mode we have ( xx .
) .
For the moments about the origin : Put , hence Hence we deduce ( xxi .
) . .
( xxii .
) .
Now if we compare these results with those on p. 368 of the earlier memoir we see that the one set can be at once deduced from the other by writing Thus with this interchange the whole of that solution holds , if we bear in mind that the range is now from to We easily find : V0L .
cxcvn .
and and are the roots of . . . . . .
( xxiii .
) , where and are to be determined as in that memoir , pp. 368-369 .
We have : ( xxiv .
) , where and are both negatiye .
This giyes Thus and are known , and Equation ( xxi .
) we find the remaining unknown constant the shape of the curve , .
As before , various approximations may be used to the values of the functions when either or or both are large .
We easily obtain for the distance between mode and mean ( xxvi .
) , and for the skewness : Sk .
( xxvii .
) .
( 5 .
) A special case of some interest arises when the start of the curve is a priori known .
Suppose its distance from the mean to be and let ( using moments about centroid ) . . . . .
( xxviii .
) .
Then we easily find : .
eing ntive , is negative , and accordingly by what goes before lies between 1 and The for curves of Type I. , if but large ( ' Phil. Trans , vol. 186 , p. 369 , foot-note ) , is and this can ) easily modified to suit ( xxv .
) above .
A very convenient and exact formula for , if be large , is that given by FORSYTH ( ' B.A. Report , ' 1883 , p. 47 ) : error being less than of the whole , TO TH1 THEORY OF EyOLUTION .
Whence we deduce to determine and : ( xxix .
) , and the solution proceeds as before .
( 6 .
) .\mdash ; I propose to llote a few distributions of frequency in which I have come aclOSS Types .
and VI .
( A. ) Statistics of Age of Bride at , the Bridegroom 's Age ) The observations given in the table , p. 454 , are taken from PEROZZO'S memoir : " " Nuove Applicazioni del Calcolo dell .
'Reale Accademia dei Lincei , ' Anno CCLXXIX .
, 1881-2 , Tayola I. The total number of recorded marriages is 28,454 .
The moments were calculated by using SHEPPARD'S COl.rections London Math. Proc , and are as follows:\mdash ; Mean age of bride Whence : 6.8873 1.9558 1.1094 Thus by p. we see that fype VI .
is the frequency to be selected , but as .
does not diffel widely unity , we shall obably get a good fit from Type V. as well .
Taking Type .
first , we find : The quadratic ( xxiii .
) is accordingly : *I selected this example at random , as one out of several leading to the curve types il my object to illustrate .
There is so much tampering with statistics , however , wheuevcr they refer to es of women , that it would probably been better to ] } used the men .
: Hence by ( xxiv .
) and by We have dingly for the equation to the curve : The distance from the ol.igin to the mean is given by the first equation of ( xxii .
) : or , the theoretical range starts with brides of years .
This is an excellent underlimit to the age of women marrying men of 24 to 25 in a country like Italy .
Our first group is at , and the above start is just two base units before this initial group .
The skewness , and the distance from mode to mean , or the mode is at Turning now to Type V. we have the following results:\mdash ; Hence Equation ( xii .
) is : Thus the positive value of is : Equation ( xiii .
) gives : Then gives : Thus the equation to the curve is : To find the position of its start we have by : or , since the mean age of brides is , the youngest possible theoretical bride is years .
This is probably a worse determination of the underlimit than in the case of Type VI .
At the same time I notice that out of about 1@0,000.women , 101 CONTRIBUTIONS TO THE THEORy OF EyOLUTION .
married between 14 and 15 years of age , and all the curves ) with sensibly finite ordinate at ; it is accordingly possible that a somewhat lowel age than actually occurs in Italy .
Equation ives us for the distance from mode to mean : or the modal age at marriage is years .
This is only about of a year or about 19 days different from the modal age as given by Type VI .
, a most satisfactory agreement .
For the skewness we have from Equation ( xvii .
) : Sk .
or , it differs by less than 3 per cent. from the skewness as given by Type ( VI .
) .
The diagram ( fig. ) shows the two and the table compares the results obtained from either with the Type Type . . . .
It is clear that for all practical purposes the curve of Type V. is as good as that of Type VI .
Indeed , there is practically no difference between them except for the * The observation data are really areas , while to save lengthy calcnlations we have compared both in diagram and table the ordinates of the theoretical curves .
This is general , if , its in this cuse , the 1lnmbe of is very ages 15 to 17 .
fit is , however , not a very good one , and although it is indefinitely better than a nol.mal curve , and we see why in the absence of these types the statistics could not be fitted with any of the first series of skew curves , yet we are compelled to consider that there are causes other than chance at work very definitely affecting the frequency of the recorded ages .
Thus the bridegrooms being 24 to 25 , the desire of the bride to be recorded as younger than her husband probably fully ccounts for the bulk of the preponderance of observation over theory TABLE of Obsel.ved and Calculated Frequencies .
in the frequency of the brides of 22 to 24 .
The defect of ides between 17 and 20 may be again due to the tendency to state the age as over 21 , and so free the woman from the need for parental sanction .
These causes , giving a false displacement of age frequency , are probably in themselves sufficient to account for the theoretical defect in brides of 15 to 17 .
( 7 .
) ( B. ) On the Variation in the Number of Lips of the P. Pentata .
My data are the following , taken from a paper by ALFRED GOmSBOROUGH MAYER : " " The Variations of a Newly Arisen Species of Medusa 'Science Bulletin of the Museum of the Brooklyn Institute , ' vol. , p. 1 , 1901 .
I have fouud in llngland the stateme1lt of the bride 's age in the marriage licence is for the same occasionally not in accord with the year of birth shown by the parish register .
CONTRIBUTIONS TO TH THEORy OF EVOLUTION Frequency .
No. of lips Total 996 Mr. MAyER ( p. 12 ) notes the failure of my curve of Type IV .
I find for he constants : Mean lips .
4.16834 Since is so nearly unity we may use Type V. Hence I find : ( must be negative since is negative ) Thus the curve starts at lips : or the one medusa with seven lips is theoretically excluded .
Here I have worked with the uncorrected moments because the lips are discontinuous variants .
Working with SHEPPARD'S corrective terms the limit is about six lips , and with the corrective terms suggested in my memoir on skew variation the limit is .
Further we have : distance from mean to mode Sk .
The mode is thus at , in good agreement with observation .
The equation to the curve is , taking positive from lips towards lesser values : \mdash ; K. PFARSON ON MATHEMATICAL This curve was drawn on a large scale and its areas read off with an integrator .
The following theoretical frequencies were obtained : No. of lips .
Observation .
Caloulation .
6 and over 50 47 There would not be any serious divergence here , were it not for the group with jour lips , which ) servation shows to be much under-represented .
But it must be remembered that we have only seven groups , and that such a number is very insufficient for a good determination of the moments of a curve .
Further , the variation is not really continuous , as indicated by the curve , but discrete .
We have at present no clear statement as to how the moments of a discrete system of variation should be modified or corrected so as to give the best results for the moments of the continuous curve which is to theoretically represent the series .
I am doubtful 6 of ' whether corrections\mdash ; the best for continuous variation\mdash ; are equally appropriate in this case .
Above I have used merely the rough moments , but I CONTRIBUTIONS TO THE OF EyOLUTION .
have found by considerable experience that in the case of the system as a polygon and correct , as in my memoir on Skew iation ( Trans vol. ) , appears to give the best results the lnpared with the discrete groups .
The point wants there ition ; when we have a large number of gl'onps it is of little importance , but it a able diflerence in these excessively skew distributions of discrete variables number of are small .
Above all , the diagram ( fig. ii .
) shows how all important it is to compare and not merely the ordinates of frequency curve with the blocks the discrete frequencies in such a case as this .
The wide-spread custoll foreign investigators of comparing merely the ordinates of the theoretical frequency curve with the observed frequencies leads in such cases to most acious results .
( 8 .
) ( C. ) On the Distribution of Incidence of Fever Age .
It seems sirable to give an illustration of the method of dealing with a distribution which falls under the class dealt with in Section ( 5 ) of this MACDONELL , in dealing with the intensity of incidence of different diseases at various ages , has come across in scarlet fever a good illustration of curves of the types now under consideration .
The whole of the arithmetical v-ork on the present example is due to him , and I have to thank him very heartily for allowing me to it here .
The statistics are taken from the ' Report of the MetropolitaIl Asylums } ' ( Statistical Part , 1899 ) .
They involve 39,253 ?
cases , distributed as follows:\mdash ; Yefi of life .
requency .
Under 1 443 1-2 1456 3-4 3599 4-5 3862 5-10 15-20 Year of life .
ency .
926 420 30-35 35-40 91 40-45 45-50 I7 55-60 data grouped partly one and partly in five-year the meant had to be calculated with caution , ) arating the into five years as the unit , Dr. MACDONELL found for the uncorrected nlents : , petals of buttercnps , teeth the carapncc of prawns , lips of , :as ith cillS on chestnut leaves , florets on ox-eyed , \amp ; c. VOL. CXCVII.\mdash ; A. 3 K. ON MATHEMATICAL Mean age of incidence , years .
The were not modified corrections , these suppose contact of a at terminals { he curve , and it was quite apparent that the curve a finite the side .
The wlng additional constants then ] ' : \mdash ; Thus is and and the distribution is of Type Vl .
Now let us suppose the incidence of scarlet fever to start with birth , although there might , as in the of enteric fever , be really some antenatal cases .
* Tul.ning to Section ( 5 ) we have : distance from birth to mean years units .
Hence deduce so ( xxix .
) or , ; Then from we find and , finally , after ] bermining y from ( xxi .
) , Thus the values of the frequency are given by The of the curve is thus , 104 before birth .
The mode is given by Thus : ' Phil. , vol. 186 , p. 390 .
The remarkal ) shnrp risc of the soarlet-fever ) cd with the cnteric is , however , mnch inst.this .
IBUTIONS TO THL OF ] This gives the value 3892 .
Distance between mode and mean Whence we find skewness the value Sk .
The tljat the fit may ) considered a good ( 9 .
) The conclusions of this paper are , I think , of some interest from the standpoin of scientific investigation .
A ) of frequency had been , not only by my co-WOl.kel .
S and myself here , but by biologists in America , not to fit into the general system of skew distributions dealt 11-ith by ] memoir .
first conclusion that ever w system appeal.ed to be , it ilure for a few skcw distributions .
But on more careful investigation of the differential it appeared that two ) of solution had been left out of consideration , and these were precisely those needed in the recorded cases of failure .
I owe some to authors like Professor DAVENPORT and Dr. DUNCKER , who have recently issued text-books on the application of statistical methods to biological variation , because although we known used these for some past , no account has hitherto been published of them , ctud , , biological using their of my methods have been , and I still may be , ally 1 zled . .
A. G. in tho on Iodnsae referred to )

