st: estimating measurement error with interval censoring

["Rosy Reynolds" <rr@dandr.demon.co.uk>]

Hi,
I am working with a rather wobbly measurement method where the readings are 
also interval-censored. (It is a doubling-dilutions titration, for anyone 
interested. After transforming to a log2 scale, the measurements appear as 
integers where a reading of e.g. x=4 actually means 3<x<=4.)

I am interested in characterising the repeatability of the measurements, and 
have some replicates to work with. I have about 300 samples, each of which 
has been measured on up to 8 occasions, typically 6-7 occasions, but 
sometimes only once or twice. This has some similarities with panel data 
with N around 300, T around 6. The actual measurement will be different in 
each of the 300 samples, but I'm not interested in that. I am only 
interested in the variation between replicates, and I am prepared to assume 
that the variation is normally distributed and the same for all samples.

My idea so far is to use -intreg- and include dummies for sample. I was 
concerned about using dummies, and such a large number of them, having been 
warned frequently about the dangers of group dummies with logistic 
regression, but I thought it should be OK as this is linear regression (see 
e.g. the footnote in this post from Bill Gould 
http://www.hsph.harvard.edu/cgi-bin/lwgate/STATALIST/archives/statalist.0710/date/article-934.html). 
However, a very quick simulation suggests that the estimate of the variation 
(obtained from e(sigma) following -intreg-) is biased downwards when the 
number of replicates per sample is small-ish.

I'd be very grateful if someone on the list with experience in this sort of 
thing could comment about whether the -intreg- plus dummies approach is 
reasonable, or suggest something more appropriate.

Thank you.
Rosy Reynolds 

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/