337 On the Influence of Bias and of Personal Equation in Statistics of Ill-defined Qualities : An Experimental Study .
By G.Udny Yule , Newmarch Lecturer in Statistics , University College , London .
( Communicated by Professor 0 .
Henrici , F.R.S. Received November 4 , \#151 ; Read December 7 , 1905 .
) ( Abstract .
) I. This experiment was undertaken to elucidate the real character of such statistics as those of eye-colour , hair-colour , temper , health , etc. , which have been given , e.g. , by Mr. Galton and by Professor Pearson .
The statistics are , it should be noted , not merely statistics of qualities , but of ill-defined qualities , the only guidance to the use of the terms of classification being\#151 ; with some exceptions\#151 ; common usage .
Strictly speaking we must remember that data so collected are statistics , not of qualities themselves , but of names assigned thereto .
It was desired to determine how far the distinction is of importance ( 1 ) as regards the naming of single samples : ( 2 ) as regards the naming of pairs , two samples of a quality being named more or less together , by themselves , for forming a contingency-table .
A matt-surfaced photographic paper was printed by successive exposure to 16 depths of tint , from a slightly impure white to a deep blackish-brown .
Small scraps of about S-inch square were cut from each tint , and mounted on cards , two scraps being placed on each card , combined in such a way that every possible combination occurred , making 10 x 16 = 256 cards .
Observers were then asked to name the tints on each card under one or more of the following schemes of classification , each observer naming the whole pack:\#151 ; Series A.\#151 ; 1 .
Light .
2 .
Dark .
Series B.\#151 ; 1 .
Light .
2 .
Medium .
3 .
Dark .
Series C.\#151 ; 1 .
Yery light to light .
2 .
Rather light .
3 .
Medium .
4 .
Rather dark .
5 .
Dark to very dark .
The cards in the pack were arranged , by shuffling , in a more or less random order .
Returns wTere obtained from 34 volunteer observers , who sent in 17 schedules under Series A , 20 under Series B , and 30 under Series C. II .
As regards the way in which single tints alone are named : ( 1 ) No observer , as might be expected , is quite self-consistent in his naming ; ( 2 ) the inconsistencies are greater for Series B than for A , and greater for C than for B ; ( 3 ) the observers attach very sensibly different meanings to the terms used for classification ; ( 4 ) As a combined result of ( 1 ) and ( 3 ) the terms used for classification do not determine discrete classes , but 338 Mr. G. Udny Yule .
Influence of and of [ Nov. 4 , very widely overlapping frequency distributions .
In Series B , for example , the following is the distribution of the actual tints under the names :\#151 ; Table showing Distribution of Tints under Names for the whole of the Twenty Observers who made Beturns under Series B. 10,240 Observations .
Name Number of tint ( 1 = white , 16 = very dark brown ) .
to tint .
1 .
2 .
3 .
4 .
5 .
6 .
7 .
8 .
9 .
10 .
11 .
12 .
13 .
14 .
15 .
16 .
Total .
Light ... 640 640 636 618 521 368 269 186 93 23 15 4009 Medium \#151 ; \#151 ; 4 22 119 272 366 451 524 564 516 227 193 21 6 1 3286 Dark ... 5 3 23 53 109 413 447 619 634 639 2945 The result suggests that methods which treat as discrete classes determined only by names in common use are not strictly applicable , and that quantitative results obtained by such methods can be regarded only as useful illustrative analogies .
III .
Contingency-tables were formed from each observer 's schedules for the names assigned to the upper and lower tints on each card .
If an observer returned quite without bias , the frequencies in the compartments of his table should be given by the rule of independence ( total of row x total of column -r- by whole number of observations ) .
There proved , however , to be a distinct tendency to .
return an excess of pairs of the same name ; this tendency , though vanishingly small for Series A , became marked for Series B , and more marked still for Series C. This feature was emphasised when different observers ' results were pooled , as the pooling of results of different observers who are quite unbiassed , tends in itself to give an excess of homonymous pairs .
In Series C there was also an excess of contrasted pairs .
The following table gives the actual aggregate of returns for Series B. The first number is the theoretically correct frequency , the number after the sign the excess or deficiency of the actual returns .
Name assigned to lower tint .
Name assigned to upper tint on card .
Light .
Medium .
Dark .
Light 785 + 65 653-35 570-30 633-62 527 + 66 460- 4 583- 3 486-31 423 + 34 Medium Dark 1905 .
] Personal Equation in , etc. 339 The above table includes returns from the 20 observers ; separate tables for the first and second ten in alphabetical order gave reasonably consistent results , the frequencies of homonymous pairs being in excess in every case .
An experiment was also tried , eliminating the returns for certain cards in the pack with contrasted tints so as to make the distribution a correlated instead of an independent one .
In this case the contrasted pairs returned for Series B seem to be in excess instead of in defect as above , the correlation coefficient calculated for a division between light and medium for the one tint and medium and dark for the other , being slightly lower than the true value for a similar division of the actual data .
For symmetrical division the coefficient was slightly higher than the true .
IY .
Certain of the contingency-tables given by Professor Pearson were examined to see how far the observed peculiarities might be due to subjective influences .
The excess of homonymous pairs , and , indeed , the correlation , in the eye-colour table , for homogamy between husband and wife would seem to be largely due to such influences .
In eye-colour tables for brother and brother , and for father and son , the divergencies of the distribution from normality are of the same type as the divergencies , in the above experiment , of observation from truth , but much larger .
For the brother-brother tables for temper and for curliness of hair , speaking generally , the same thing holds .
There is very possibly , accordingly , some real objective effect , the nature of which requires elucidating , but considerable reserve is necessary in view of the qualitative similarity to effects of subjective origin .
A collection of data is required in which these are eliminated by the use of good representative scales , or possibly by the naming of the two members of a pair quite independently by different observers .
VOL. LXXVII..A .

