|
BGonline.org Forums
Conclusion
Posted By: Maik Stiebler In Response To: Conclusion (Casper Van der Tak)
Date: Sunday, 14 August 2016, at 3:46 p.m.
Let X the raw outcome of a trial and Y the measured luck of a trial. Let var the variance, cov the covariance, sig the standard deviation and cor the correlation. Then literally, the variance reduction in absolute terms is
var_X - var_{X+Y}.
Using the definitions and elementary algebra, this simplifies to2*cov_{X,Y}-var_Y.
In terms of standard deviation and correlation,absolute var. red. = 2*cor_{X,Y}*sig_X*sig_Y - sig_Y*sig_Y =sig_Y*(2*cor_{X,Y}*sig_X - sig_Y).
So variance reduction 'works' when2*cor_{X,Y} > sig_Y/sig_X.
It is interesting to look at the relative variance reduction, which is obtained by dividing the absolute one by (sig_X)^2:rel. var. red. = sig_Y/sig_X*(2*cor_{X,Y} - sig_Y/sig_X)
For a given correlation, this is optimized whensig_Y/sig_X=cor_{X,Y}.
As simple linear scaling of Y changes sig_Y/sig_X, but not the correlation, I wonder if that can actually be used to fine-tune variance reduction.
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.