38 
Notes on the History of Correlation 
Another feature of Galton's work at this time must be noted. He worked with 
the median instead of the mean, and he i;sed probable errors or quartile values 
instead of standard deviations. Fixrther, to obtain r, he somewhat laboriously 
expressed both variates in terms of their quartile deviations : thus r became the 
slope of his regression line. It was then determined by graphically fitting a good 
line, or from certain chosen arrays. Thus he worked with somewhat primitive 
statistical tools, and the wonder is that he achieved as much by their aid as he did. 
Given A and B with regression r„,,, B and G with regression riy^, Gal ton assumed 
rf,c=')'aij X i\. to obtain his kinship relations. A nephew is the son of a brother. Hence 
r for uncle and nephew = r for brothers x r for father and son. 
This of course is incorrect ; it implies the vanishing of the corresponding partial 
correlation coefficient. Again, I think, his mid-parental correlation is not theoretically 
consonant with his parental correlations. 
Another noteworthy point of the 1877 R. I. and the 1885 A. I. papers is the 
ample provision of mechanical models to illustrate by dropping shot or seeds the 
propei"ties of bi-variate frequency. One wonders whether these elaborate quin- 
cunxes have been preserved and if so where they are at the present time. I repro- 
duce one of them by permission from the Journal of the Royal Institution. 
DC. VIA TlON 
POSITIVE 
DEVIATION 
COcrr^OF 
REVERSION 
OKl_ OBl 
OA Oa 
In 1886 Galton published a paper in the Royal Society Proceedings* on 
" Family Likeness in Stature." This contains Hamilton Dickson's note and further 
data from Galton's Family Records. 
* Vol. XL. pp. 42—66. 
