Re: Data reduction methodology for V-I colors

[Andrew Bennett <andrew.bennett@ns.sympatico.ca>]

On Tue, 31 Aug 2004 12:20:51 -0500, Tom wrote:

>Mike,
>
>Yep, this and a lot more.  I cannot identify a "bad" night.  I did a lot of
>work on the assumption that there was such a thing as a bad night.  For
>example, go through all the data and look for stars with big deviations
>from the mean.  (This should work because most stars are not variable.) Now
>sort all frames by fraction of stars with big deviations from the mean. 
>Now eliminate these from the data set and look at the result. I did this
>eliminating the noisiest 10%, 20%, 50% ...  No improvement at any cut
>level.  Conclusion:  Eliminating frames with lots of deviant measurements
>does not improve the quality of the data as a whole.

Not for the first time, may I point out that Tom's
repeated conclusion disagrees with my analysis as
reported in TN94. I don't know why.

I was using an ensemble calibration.
I was "correcting" - I use the term loosely -
for gradients within each image using linear
(3 coefficients: zero point and two gradients)
and, on occasion, quadratic correction (6
coefficients, adding curvature in each coordinate
and cross gradient.)

Eliminating images with high scatter gave a
significant improvement.
Eliminating stars with high scatter was also
included - later versions that I have not
written up do not eliminate high scatter stars
and appear none the worse ... not sure why!

The improvement from eliminating high scatter
images was nowhere near as great as I had hoped.
The residual errors are very far from Gaussian
and include far too many large errors. The
detection of variable stars remains horribly
unreliable as compared to naive calculations 
using the observed rms errors assuming a Gaussian
distribution.

Andrew Bennett, Avondale Vineyard