]\gt ; Captain J. W. Bispham .
Distribution of the the velocity of sound .
It must not be supposed that this represents the al aocuracy of the results , especially at velocity ratios above or , where the effect of the zero point may be serious .
It does seem , however , possible to claim that the tables and curves of this paper give values of these coefficients of the true order of magnitude , and show correctly the type of change that takes place as the velocity of the shell passes through the velocity of sound .
An Determination of the Distribution of the Partial Correlation Coefficient in of Thirty .
By Captain J. W. BISPHAM , E. Communicated Prof. AIartin , Received November PART I.\mdash ; Samples an Uncorrelated Universe .
1 .
IntroductorlJ In .
" " Stndent\ldquo ; dealt experimentally with the distribution of the total correlation coefficient of small samples .
In particular , he dealt with values of as low as 4 for the case of zero correlation in the sampled population .
In 1913 H. E. SoperT theoretically determined the mean correlation and the standard deviation of the distribution of correlations to second approximations .
In 1915 .
A. an equation for the frequency distribution of , and in 1917 , as a result of a co-operative study by H. E. Sopel , A. W. Young , B. M. Cave , A. Lee and K. Pearson , S this was reduced to stitable form for numerical manipulation , and the frequency distributions and frequency constants for samples of size ranging from to were given for values of the correlation in the sampled population from to .
The present expel.imental investigation was commenced in 1914 , but had to be put aside during the war .
It was intended to determine whether the distribution of partial correlation coefficients for samples as small as 30 showed greater dispersion than is observed for total correlation coefficients .
has shown that for normal distributions and samples the standard deviations of the distributions should be of the 'Biometrika , ' vol. 6 , p. 302 , et seq. ' Biometrika , ' vol. 9 , p. 91 , et seq. Biometrika , ' vol. 10 , p. 507 , et S 'Biometrika , ' vol. 11 , p. 328 , et seq. G. U. Yule , ' Roy .
Soc. Proc , vol. ( 1907 ) .
Partial Correlation Coefficient in Samples of Thirty .
219 same magnitude .
The experiment can now be related to the complete evaluation of the distributions of total correlations referred to above .
* 2 .
of Sampted The population to be sampled was obtained artiiicially as follows : Thirty counters\ldquo ; bearing the numbers to and \mdash ; 1 to were drawn in random order from a container and the numbers were written down in the order drawn .
This set of thirty numbers represents the varying values of the first attribute .
The values of the second attribute were obtained in exactly the same way , viz. , the counters were returned to the container , shaken up , and drawn again in random order , the numbers drawn being written beside the first set of thirty .
Similarly , the varying values of the third attribute were obtained by a third draw .
This artificial population has the advantage that the mean value of each attribute in each sample is zero , and further , the standard deviations are invariable from sample to sample .
This saved a good deal of arithmetical work .
The frequency distribution of the attributes in the artificial population is very far from normal , being of rectangular form with a central strip missing , since any of the numbers except zero occurs as often as any other , while the number zero does not appear at all .
A further advantage of so irregular a distribution lay in the fact that the investigation was intended , in part , to test the reliability of partial correlation coefficients obtained in espect of groups of registration districts used in a previous investigation , and in such cases the frequency distribution is , of course , irregular .
3 .
Observed Distributions of Totad and Partial Correlation cients .
The values of the total correlation coefficient between the various pairs of attributes will be referred to generally as , and .
The partial correlation coefficient determined was that between attributes 1 and 2 for 3 constanlt .
notation will be used .
In all , 1000 cases were worked and the observed distributions are shown in detail in Table I. The iven in column 6 is that deduced from the investigation referred to above .
R. A. equation for the distribution is .
shown to lead to 'Biometrika , ' vol. 11 , p. 328 , et seq. ' Co-operative Study , ' .
cit. ptain J. W. Bispham .
Distribution of the where and This is a very convenient form for the evaluation of the distributions for successive values of Table I.\mdash ; Total and Partial Correlation Coefficients : * These are frequencies for the whole range to the limits of the distribution .
'Co-operative Study , ' .
cit. the paper in question the ordinates are given for and ordinates for were calculated by the method indicated above , checked to six significant figures by the alternative method given on .
The intermediate ordinates forming the innings and ends of the Partial elation Coefficient in Samples of Thirty .
221 frequency groups were calculated by the latter method , and Simpson 's formula was used for finding the areas , viz. , An inspection of Table I leads one to expect a reasonably good fit in all cases .
Before referring to the determination of oodness of fit it may well to give particulars of the frequency constants of the observed distributions and of those calculated from the theoretical These are given below:\mdash ; Table II.\mdash ; Values of Frequency Constants of Distributions in Table I. On p. 371 of the paper in question , it is stated that the condition for Gaussian ibution is ) attained for samples of 25 or 50 .
The distributions considered , , are concerned with values of , as high as ; and , for ] of , the tions of are markedly skew .
In the case now under consideration , however , the distribution is symmetrical , and an examination of Table II shows that the departure from the Gaussian type is not .
In fact , if the theoretical distribution be compared with the normal curve that fits it best\mdash ; that is , the normal curve with the same standard is found that , grouping the frequencies as below , the " " goodness of fit as measured by , is .
This expresses in another that , for the special case and , the distribution of 's approaches closely to the normal , and , consequently , the standard deviation may be regarded as good measure of the dispersion .
The tabulated values of are subject to a probable error , and it may be noted that they do not depart significantly from the valu , obtained from the usual expression , which is itself identical to five places the theoretical value .
'Co-operative Study , ' loc. Captain J. W. Bispham .
Distribution of the .
C'urnc : Total In the following Table III will be found the frequency distributions of able I on ranges centled about the tabulated value , together with the corresponding best fitting normal curve frequencies and the theoretical frequencies .
Table III refers to the three sets of 1000 total correlations .
The frequency distributions of the best fitting normal curves to the observed distributions are referred to as follows:\mdash ; \mdash ; best fitting normal curve for \mdash ; best fitting normal curve for \mdash ; best fitting normal curve for ) \mdash ; normal curve based on the theoretical value of as determined from the equations of the co-operative study .
\mdash ; theoretical distribution .
Table \mdash ; Distribution of Total Correlation Coefficients .
Values of are given at the foot of the columns as .
Comparing the values of the actual distributions columns ( 1 ) , ( 3 ) and ( 5 ) with the theoretical distribution given in column ( 8 ) we get for the various values of P. with with with The value of as between columns ( 7 ) and ( 8 ) , i.e. , between the normal * The probable error of is rather large , as has been shown by Prof. K. Pearson , ' Phil. Mag vol. 31 , April , 1916 .
Partial Correlation Coefficient in of Thirty .
223 curve based on the theoretical value of and the actual theoretical distribution is as stated above Cur : Partial tations .
The ensuing Table gives the corresponding frequencies of the correlations as observed , together with the corresponding values of , the best fitting normal curve to the distribution , the best fitting Pearson curve type 2 and the values of and as used above .
In each case the values of are given at the foot of the columns .
Table \mdash ; Frequency Distribution of Partial Correlation Coefficientb .
5 .
Sumrvary .
It will be seen that the distribution of partial correlation coefficients is , the whole , as well fitted by the theoretical distribution as are the distributions of total correlations .
Comparing the four observed distributions with the distribution referred Go as , we have\mdash ; For distribution of For distribution of For distribution of For distribution of 224 Partial Coefficient in Samples of Thirty .
The 's all lie within the ranges of each other's-probable errors so that it is not .
possible to say that one distribution is significantly better fitted than another .
The experiment indicates , inter alia , that values of the partial correlation coeflicient ( as determined by samples of 30 ) as great as may be expected when the true value is zero , about 4 times per 1000 , and further that the dispersion of partial correlation coefficients for and is not significantly different from that evaluated according to the co-operative study ( loc. cit. ) referred to above .
The results obta , ined above are of considerable practical importance .
The partial correlation coefficient has been recently used in the analysis of many types of data .
In the current psychological controversy*as to the existence of a central factor common to all forms of intellectual activity the method of partial correlation has been used in respect of groups of individuals of the size now under consideration .
Considerable use has also been made of the partial correlation coefficient in analysing vital statistics and to refer to only one other sphere of application of the method a recent published by the Office indicates the extent to which total and partial correlations have used in analysing weather data .
A further experimental inyesGigation is respect of samples from a highly correlated population has been in progress concurrently with the above and it is hoped to publish the results in due course as Part II of this investigation .
The numerical work in both cases has been extremely laborious .
In conclusion I beg to state that the investigation has been conducted under the boeneral supervision formerly of Dr. E. C. Snow and latterly of Captain NL Greenwood and owes much to their help and advice .
Prof. Spearman , 'American Journal of Psychology , ' vol. 15 , p. 284 ; Cyril Burt , ' British Journal of Psychology , ' vol. 3 , pp. 94-177 ; W. Brown , " " The Essentials of Mental Measurement 'Camb .
Univ. Press , ' .
Dr. E. Webb , " " Character and Intelhgence 'Brit .
Jour .
Psychology Monograph , ' 1915 ; G. H. Thomson and J. C. M. Garnett , 'Brit .
Jour .
of Psychology , ' vol. 9 , p. 321 , et seq. ; J. C. Maxwell Garnett , 'Roy .
Soc. Proc Series , vol. 96 ( 1919 ) ; Godfrey H. Thomson , 'Roy .
Soc. Proc Series , vol. 95 , 1918 .
M.O. 223 , The Computer 's Handbook , ' sec. , 3 .
A Collection of Correlation Coefficients from Meteorological Papers , and a Note on the Partial Correlation Coefficient .
By Captain E. H. Chapman , R.E. I am especially indebted to Miss A. Fowler and to Mr. M. E. Wilson for help in this respect both for original working and for checking .
Most of the original drawings were made by school children .

