]\gt ; On the Conditions under which the " " Probable Errors\ldquo ; of Frequency Distributions Real Significance .
By L. ISSERLIS , B.A. ( Communicated by Prof. Karl Pearson , F.R.S. Received June 21 , 1915 .
) When we seek the value of a statistical constant , we may either consider the whole aggregate of individuals pessinl characteristics of which the constant in question is a function , or we may limit ourselves , from choice or necessity , to the consideration of a random sample of the whole population .
The mean height of of military , at a given instant , is a constant which could be determined by considering the whole population , but , in practice , would be determined from a random sample .
On the other hand , the mean weight of adult frequenting the North Sea is necessarily to be determined only by a consideration of a sample of the whole population .
Statistical constants calculated from a sample give us little information unless we know , at the same time , the manner in which the values may be expected to vary from random sample to random sample , i.e. the frequency distribution of the constant in many samples .
The universal custonl is to state the " " probable error \ldquo ; the constant , which is equivalent to giving the parameter of the normal curve of errors representing the frequencies of the deviations of the values of the constant , calculated from the various samples , from the value of the constant in the population as a whole .
The parameter\mdash ; the standard deviation of the frequency distribution\mdash ; therefore ceases to provide an adequate description of the facts if the frequency distribution differs sensibly from the normal .
Now any statistical constant may be determined by means of the moment coefficients of its frequency distribution , and it is well known that the mean and the standard deviation , i.e. the first and second momeut coefficients of a frequency distribution , themselves have frequency distributions which rapidly approach normality with the increase in the size of the sample , at any rate , when the sample is itself of small size compared with the size of the whole population .
When a frequency distribution ceases to be norma ] , it can , in the * The statement that a statistical constant has a value with a " " probable error \ldquo ; means that if the constant be determined many times from random samples of the relevant material , the value obtained will differ from by less than , in half the number of cases .
is always taken equal to times the standard deviation .
this is based on the assumption of normal distribution , and it is this value of that ceases to have real significance if the distribution be not normal .
24 Mr. L. Isserlis .
Conditions which " " Probable Errors\ldquo ; overwhelming majority of cases , be represented by one of Pearson 's well known frequency curves , the constants of such a curve requiring in addition to the mean and standard deviation , a of the -coefficients and , where , are the third and fourth moment coefficients .
For a Gaussian distribution is zero and is three .
I propose , in the present paper , to discuss the values of these -coefficients for a moment coefficient of order in the case of a sample of size ?
extracted out of a finite population of size N. It will appear thaC in the case of moment coefficients of moderate order , and approximate to their normal values as the size of sample increases , especially if the sample is itself small compared to the general population .
For coefficients of however , the sample has to an inconveniently fraction of the popnlation itself if and are to approach even approximately to their Gaussian values .
The usual " " probable error\ldquo ; ceases for such moment coefficients to have Gaussian significance and statistical constants , founded on such moments , cease to be of valne , unless they are accompanied by the constants of their skew frequency oution .
It t.o be hoped that the values of , which are here given without approximation , for moments of any order , ma prove useful in applications in which is not small and moments of order necessarily occur .
The results of the present paper suggest , howevel , that there is little value in graduation which , thanks to the employment of high moments , represent a iven sample with raphic accuracy .
Now the moment coefficient of order obtained from a ) of drawn a population of is given by the formula where is the frequency of the roup of character in the sample , being measured from some fixed origin .
The mean value of or will clearly be the value of for the population , or where is the frequency of the group of character in the whole populatiou where deviation in the sample .
We have now to find the monlentsM , of the distribution of to ascertain the values of and and ascertain how these approach the values zero and three , wluch would accord with normality as the value of is increased .
It is clear that we shall * For moments of odd tends to zero , howe high the order may be .
of Frequency Distrributions require for this purpose a knowledge of the mean values of , as well as the mean values of various products of the deviations of type , and in general of products of type , of yree not higher than the fourth .
In the next place when the moment coefficients are taken about the mean of the sample , so that and we shall require for the determination of the moments of the deviation , a of the mean values of products ( of not higher than the fourth ) of type We proceed in the next two sections to a systematic ation of the mean values of products of these two types .
2 ( i ) .
The Moment of the Frequencics.\mdash ; If a sample of be extracted from a population of containing Np with and Nq without a given characteristic , then the moment coefficients are the moments of the terms of the hypergeometrical series , ( 1 ) where The moments of this series are known to be as follows:* In the case of a character occurring in a group of and , whence by substitution in the above values we obtain the moment coefficients of occurring in the sample / Let us write , ( 3 ) 'Phil .
Mag 1899 , pp. 236-246 , Prof. Karl Pearson , " " On Certnin Properties of the Hypergeometrical Series Mr. L. Isserlis .
Conditions under which " " Probable Errors\ldquo ; and When the sample is small compared with ] the total population , so that is ible , we have and Then the mean value of is .
Or Mean to the best of our knowledge , as is the besb value we can put for Similarly Mean value of , ( 7 ) Mean value of ( ii ) The of the Frequencies.\mdash ; Let a populatiou co1ltain dividuals of type 1 , N2individuals of type 2 and others .
A sample of size being taken , the probability that it contains individuals of type 1 , individuals of type 2 , others , is clearly : where , , while Thus the product moment coefficients are those of the tel.ms of the double hypergeometrical series The sno1nents of this series have been calculated by the author* 'Phil .
Mag 1914 , pp. 386-387 , L. Isserlis , ' Application of Solid Hypergeometrical Serics to Frequency Distributions in Space of Frequency butions hSigni.ficance .
and the values required in the present paper can be obtained by replacing the quantities , and of the ' Phil. Mag. ' article by and respectively .
We thus obtain the additional mean values\mdash ; Mean value of Mean value of .
( 10 ) It was unnecessary for the purpose of fitting the double hypergeometrical series to calculate the product moments and , but as the mean values of and are required in the present investigation , we will indicate briefly the methods by which these product moments are obtained .
The notation is that employed in the ' Phil. Mag. ' article previonsly quoted .
Operate on Equations ( 11 ) and ( 12 ) with and respectively and we obtain , ( A ) and Divide by and put , then equation ( B ) becomes or replacing the raw moments by their equivalents in terms of moments about the mean If we now use the moments already found , we obtain on reduction .
If we treat ( A ) in the same way , it becomes 'Phil .
Mag 1914 , p. 380 .
28 Mr. L. Isserlis .
Conditions under tnhach " " Probable Errors\ldquo ; A similar equation will hold in which the suftixes and accents are interchanged , is replaced by by and by .
Let second equation be subtracted from the one last written ; the factor will cancel , and , after considerable reduction , we find ( D ) From ( C ) and ( D ) , we iind Mean value of ( 11 ) Mean value of .
( 12 ) ( iii ) We shall require also the values of , and Now , for given values of and , the mean value of be iven by the regression plane Mean Now from ( 6 ) and ( 9 ) and .
Hence mean for constant and mean mealt mean d .
( 13 ) The mean value of similarly seen to be mean mean d , and by the use of ( 11 ) and ( 12 ) we , after some reduction , Mean value of .
( 14 ) inally , to find the meall value of , we use the regression equation Mean for constant ' Phil. Trans , vol. 18 p. 302 .
of Frequency Distributions have Real Significance .
where denotes the minor with its proper sign ) of in the determinant Using the values of the 's and 's as above , we find mean value of ( mean mean dmean d ) , and with the aid of ( 14 ) this reduces to mean value of When the total population is infinite , the frequency distributions can be represented by binomial series in place of the hypergeometrical series employed above .
If , in addition , we suppose the sample to be so large that may be neglected , we get the case considered by Mr. H. E. Soper* in his discussion of the probable error of the correlation coefficient , and we may vote , as a check on the results ( 6 ) to , that when we put , snd neglect terms in , they agree with the values in equations ( xviii ) to ( xxvii ) of his paper .
For the special case of a normal distribution .
Ghe mean value of can be from a esult published by Prof. Karl Pearson leading to , reeing with our equation ( 12 ) .
We may note also that the methods of the latter part of this section may also be employed to obtain the mean values of and directly from the single series , but fail to give mean value of .
This necessitates the employment of the double series .
3 .
We are now in a position to calculate the exact values of the moments and the coefficients of the distribution of , the characteristics , , being measured from some fixed origin 'Biometrika , ' vol. 9 , p. 95 ( 1913 ) , H. E. Soper , M.A. , " " On the Probable Error of the Correlation Coefficient to a Second Approximation .
' Drapers ' Company Research Memoirs , ' XIV , " " the General Theory of Skew Correlation p. 20 , footnote .
30 Mr. L. Isserlis .
under which " " Probable Errors\ldquo ; and meall value for all samples of or mean value of using ( 6 ) and ( 9 ) , or .
( 16 ) For the third moment coefficient about a fixed origin , we have mean value of mean value of using ( 7 ) , ( 10 ) , and ( 13 ) , this becomes reducing to Thus .
( 18 ) If the population out of which the sample is compared to the sample , we may put , and ( 18 ) may be put in the form ( 19 ) The fourth momeltt about a fixed origin is mean)alue of mean value of value of insert this expression the mean values of the fourth order moments and product 1noments of the frequency deviations given by ( 8 ) , ( 11 ) , ( 12 ) , ( 14 ) , and ( 15 ) , and find of Distributions Real Significance .
' or This be reWl.itten in the form But we have already shown that , ( 16 ) therefore the " " second \ldquo ; of the distribution of the mcment of order about a fixed Origin is given by the elegant formula Now for an infinite population and .
For such a population , therefore , Let us put in equations ( 19 ) and ( 23 ) and we obtain vell known result : If , B2 be th constants of the of the of of drawn at ra , ndom , corrcsponding to in the original f ) ibntion ( uith infinite population ) , In other words the and tend to the normal values and 3 respectively as the size of the sample increases .
* Henderson , R. , JourlL Inst. of Actuaries , ' , pp. 429-442 ; cf. also " " Student ' Biometrika , ' vol. 7 , p. 210 . .
Isserlis .
Conditions under which " " Probable Errors\ldquo ; 4 .
With a view to their use in values of the moments , and t , he -coefficients B2 of the moment coefficients , no about a fixed ; but about the mean of the sample , we proceed in this section to the calculation of the products of various powers of the deviations in the frequencies of one or more of the characters and various verse o the deviation in the rneaIl .
, meall , by ( 6 ) and ( 9 ) , or mean .
( 24 ) When this reduces to the well known formula for the probable error of a mean : Of course the result in ( 24 ) can be obtained directly by putting in equation ( 16 ) .
In the same way we obtain directly from ( 17 ) and ( 21 ) the esults : Mean value of .
( 25 ) Mean value of .
( 26 ) Mean of mean value of by ( 6 ) and ( 9 ) We shall find it convenient , and an aid to brevity to denote , etc. , by , , hence mean value of .
( 27 ) Mean value of mean value of by and ( 10 ) ( 28 ) of Frequency Distri utions have Mean value of mean value of mean of Using ( 7 ) and ( 10 ) and ( 13 ) , this reduces to .
( 29 ) The mean value of mean of and using ( 10 ) and ( 1:3 ) this reduces to .
( 30 ) For the mean value of we use ( 8 ) and ( 11 ) and find mean of or Mean .
( 31 ) The mean value of mean of using equations ( 11 ) , ( 12 ) , and ( 14 ) , we VOL. XCIL\mdash ; A. 34 Mr. L. Isserlis .
Conditions under which " " Probable Errors\ldquo ; and obtain after a further reduction Mean .
( 32 ) The mean value of mean of by ( 14 ) and ( 15 ) , reducing to Mean .
( 33 ) The mean value of mean of With the aid of equations ( 8 ) , ( 11 ) , ( 12 ) , and ( 14 ) this can be reduced to - - so that Mean value of ( 34 ) The mean value of mean of of Frequency ibutions have To reduce this we use ( 11 ) , ( 12 ) , ( 14 ) , and ( 15 ) , obtaining from which we deduce Mean value of Finally we shall require the mean value of , i.e. of This can be reduced by means of equations ( 8 ) , ( 11 ) , , and ( 15 ) .
We give the-final result Mean value of .
( 36 ) 5 .
We proceed to apply the results of the preceding section to the calculation of the constants of the distribution of a moment coefficient when it is taken not about a fixed origin , but about the mean of the sample .
We have , or .
( 37 ) mean value of mean of 36 Mr. L. Isserlis .
Conditions under ?
which Errors\ldquo ; Substituting the meall values found above , of and we get .
( 38 ) Secondly value of , and by ( 37 ) . .
Thus .
( 40 ) For we have again , from ( 37 ) , of Frequency Distnbutions have ignifican c and mean value of for all samples .
Now the mean value of the first term is clearly obtained by dropping the accents in ( 20 ) , and so equals The mean value of the second term is The third term has for mean value The fourth term has a mean value whilst the mean value of the last term is given by ( 26 ) and equals On adding the various terms together , we find that can be written in the form , ( 41 ) where , ( 42 ) ( 43 ) 38 Mr. L. Isserlis .
Conditions under which " " Probable Errors\ldquo ; From equation ( 38 ) we see that P. ( 44 ) Hence .
( 45 ) Also ( 40 ) may be written , ( 46 ) where .
( 47 ) 6 .
When is small compared with so that ( 46 ) and ( 45 ) become , and rapidly approach normal values zero and three as increases , provided and do not become large .
Now we cannot calculate either of these expressions unless we know the actual distribution , but , for large samples , they occur as coefficients of the small quantity and we may get a good idea of their order of magnitude by considering their values for the case of a normal distribution .
Now for such distributions , ( 48 ) and , ( 49 ) Hence for odd values of , ( 50 ) .
( 51 ) For even values of , ( 52 ) , ( 53 ) .
( 54 ) of Frequency Distributions .
39 To calculate these values in numerical cases , we use , when is odd , the formula : and when is even both being simple consequences of equation ( 49 ) .
In practice moments of higher order than the fourth are rarely employed .
The following Tables give the values of , for the moments up to the sixth order and the resulting values of Tabl I. 'Probable \ldquo ; .
Frequency Distributions .
It must be remembered that when the ratio of to , i.e. size of sample to total population , is small , the quantities and all approach unity .
The deviation from normality of the fifth and sixth moment coefficients is considerable , even for samples of 10,000 , and the of the fourth moment differs very significantly from three for samples of 1000 .
The size of the salnple is more influential than the ratio of the sample to the popnlation .
Consider a sample of 1000 , extracted from a population of 1,000,000 , to six places of decimals : , so that to three and For such a saml ) the -coefficients of the first four moments are given in the following Table:\mdash ; Table III.\mdash ; Approximate Valnes of for Samples of 1000 out of a Population of 1,000,000 .
To three places of decimals , values in this Table agree with the values we should obtain by ptltting , or , for a sample of 1000 , population of 1,000,000 may be considered infinite as far as its influence on the first four moments is concerned .
The first four moments are those most frequently employed and a sample of 1000 is not unusually small\mdash ; it is , therefore , of great importance to notice that for such a sample the frequency distribution of the fourth moment is decidedly skew .
7 .
Finally , it is interesting to examine the values we obtained in ( 19 ) and ( 22 ) for the and and of moment coefficient about a fixed , for the case of crtn infinite population .
They may obviously be written:\mdash ; , ( 58 ) where , and are the -coefficients of a frequency curve obtained by Some on the Properties of electrics .
distorting the frequency curve of the original distribution , the distortion consisting in raising each ordinate to the power It seems probable that the values and B2\mdash ; which are given by ( 45 ) and ( 46 ) for an infinite population admit of a similar interpretation .
In fact ( 52 ) , ( 53 ) , and ( 54 ) show that this is the case for even , when the original distribution is normal , but the general case appears at present to be intractable .
Some Experiments on the Properties of Dielectr.ics .
By SPENCER W. RICHARDSON , ( Camb .
) , D.Sc .
( Lond. ) , formerly Principal of , and Professor of Physics at , the University , Southampton .
( Communicated by Sir Joseph J. Thomson , O.M. , F.R.S. Received July 19 , 1915 .
Introdnction.\mdash ; The phenomenon of the residual discharge has attracted the attention of many eminent physicists , and a large number of papers on this subject have been published .
In 1854 Kohlrausch* showed that the instantaneous discharge apparently independent of the residual discharge and that , for a given jar left to self for a given time after charging , the residual discharge was proportional to the initial potential ; and in 1873 Maxwell put forward the view that the existence of the residual discharge could be explained by the assumption that it was due to the heterogeneity of the medium exhibiting it .
In 1881 and asserted that Iceland Spar did not exhibit the phenonlenon of the residual discharge , and claimed that its absence was a confirmation of Maxwell 's view .
I , however , having performed a considerable number of experiments on Iceland Spar , have found evidence of the phenomenon in this case .
The specimen I have studied was a carefully polished slice\mdash ; cut parallel to a cleavage \mdash ; in which no flaw of any kind could be detected by eye .
This specimen was silvered on both surfaces and round the edge\mdash ; a gap being left between the silvered part of each surface and the edge for insulation see .
The charge , as it appeared on one of the surfaces , was conducted away to a subsidiary condenser of much greater capacity than the specimen , and the 'Poggendorff 's Annalen , ' 1854 .
'Treatise , ' vol. 1 , 1873 .
'Phil .
Mag 1881 .

