Subject:
reg stability
From:
Peter Bancel <pabancel@club-internet.fr>
Date:
Sat, 15 Nov 2003 14:02:55 +0100
To:
Roger Nelson <rdnelson@Princeton.EDU>

Hi Roger,
 
Here are some plots of normalized reg data. I don't think I sent you these. Format is gif.
 
I normalize the data separately for each reg and plot the cumdev of the squared z-scores. Data is from my CD's which run through Jan 2003. The normalization of course removes all data specified in the rotten egg file and also removes raw trials that deviate from 100 by more than 50. Horizontal tick marks on the plots are 3-month intervals.
 
It is interesting to stare at these plots. For instance, reg 28 has a section of bad data in 2001 that is not excluded by the rotten egg file. You can also see where the reg has been replaced by a mindsong later in the second half of 2002. Both of these problems are immediately evident. (the normalization assumes each reg number refers to a unique device. This isn't the case for 28, so we see it in the cumdev. This is basically revealing to the eye the difference in variance bias of the two devices.)
 
Some regs like 1021,116,118 and 119 seem to have slow non-linear drifts. You'll find others that are questionable in this way, too.
 
I think there is some bad data at the begining of reg 230 and reg 226 doesn't look good.
 
Regs 2165 through 2236 should also be looked at carefully. I'll be interested to see more data from these devices.
 
Anyway, it's fun to look at these, so enjoy. I hope all's well. I will be in New York 15-29 dec. Maybe we can meet and also get together with Greg, if that's workable.
 
cheers,
 
Peter