In today's digest three items caught my attention.
1. Someone posted a one-line message saying "thank you, X" followed by
about 50 lines of prior dialogue. This is a waste of our time and download.
Please say thank you directly. Also, in posting replies, it isn't usually
needed to keep the full text of 5 or 10 previous messages. Some prior text
is usually needed to establish context (I can't figure out an answer that
say "yes, that's OK" or "no, you should try X, Y, and Z any better than you
can), but all previous messages aren't usually needed.
2. In selecting variables for a regression, there is a lot of
discussion that shows stepwise and other 'best' regressions are hopelessly
biased and should be treated very cautiously at best. Frank Harrell posted
something a few years ago, but his book of about 2 years ago in the Springer
series (Regression modeling strategies) is excellent and worth reading if
you aren't convinced by my little tantrum.
3. Looking at correlation coefficients is also fraught with peril
(how's that for a clich�?). Any tests and confidence intervals are rely
very heavily on the normality assumption (and no outliers, etc.). A paper
by E. Pearson in Biometrika (late 1960s or early 70s) did a large simulation
study that suggested this was a problem. I would suggest either using a
non-parametric correlation, or bootstrapping to get the confidence intervals
or tests. Interestingly, in my experience, when I compare Spearman and
Pearson correlations, they usually are consistent with each other (in the
non-statistical sense)
Peter A. Lachenbruch
Director, Division of Biostatistics, OBE/CBER
Tel: (301) 827-3320
FAX: (301) 827-5218
*
* 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/
