[EMAIL PROTECTED] (Robert J. MacG. Dawson) wrote in message news:<[EMAIL PROTECTED]>... > Jay Warner wrote: > > > > Simplest approach would be to transform the x-axis data so that the > > desired fixed point was at x = 0, and transform the y-axis so the desired > > fixed point was at 0, then tell Excel to force the line to go through > > (x,y) = (0,0). > > > > You can do this by subracting (or adding) a fixed value to the x's and > > another fixed value to the y's. > > > Another way of getting this result would be to use weighted regression > and put a very high weight on the point you wanted to be sure to hit, or > (if Excel doesn't support weighted regression) to cut and paste 1000 > copies of the point you want to hit into the data set. > > But this misses the question that I was trying to raise: why is it > desirable to do this? Most natural situations in which one would take > it for granted that f(a) = b would also be such that you would take it > for granted that Var(f(x)) is less for x near a than for x far from a. > In such a case you might do better to subtract (a,b) from each vector > and do a log-log transform.
Certainly there are many situations where this is the case. Mostly this happens at (0,0) and people still do lines through the origin rather than transform and estimate the average difference in the logs. Which you should do (or whether something else is better) does depend on what you're measuring (e.g. if the responses are counts, you may not want to do either!) Glen . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
