John Randall wrote: > Having learned back in those days, the fewer operations message may have > stuck. I still think there is a problem if a lot of values are close to > the mean, since calculating the difference will cause loss of precision, > but the justification I had was false. Standard deviations are used in a > way that does not require much accuracy, so it may not matter anyway.
The problem arises when the measurements involve a large number of significant figures and the variation is only in the last few digits. It is well known that for such data using data relative to some value close to the mean greatly improves the accuracy in calculating the sd.. Replacing the measurements by the deviations from the first value will often do. Then adjusting for the mean and using the older common 'desk calculator version' sd2 works fine. There are algorithms which use less operations than sd but with modern machines that only matters in specialised cases. Fraser ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
