Thanks to Kit Baum, a new module -corrtable- is available
from SSC. Stata 8 is required. Use -ssc- to install if
interested, and even if not interested.
-corrtable- is for presenting correlation matrices
as graphical tables: not as "heat maps" or the like, but
as tables produced by -graph-, with the possibilities that allows.
-corrtable- arose because a colleague was somewhat frustrated
in getting a correlation matrix in the form he wanted out
of Stata and into his favourite word processor (not a word shall
escape me on which that is, but it isn't my own favourite way
of producing documents).
As often happens, a different approach to get round his problems
turned out to be something that might be more generally useful,
to me and perhaps others.
The basic idea is very simple. Each cell in a correlation matrix
is a correlation like say (to 3 d.p.)
0.420
which we could put on a single graph like this:
twoway scatteri 1 1 "0.420", ms(none) mlabsize(*7) xscale(r(0 2) off)
yscale(r(0 2) off) mlabpos(0)
It is unlikely that you would want to do this for an individual
correlation,
but looping over pairs of variables would give you a matrix of graphs
that
you can then -combine-. And other modifications then are relatively
easy,
including formatting with different numbers of decimal places, adding
information on sample size (if that differs between correlations) and
P-value, changing the font according to the correlation and changing
the background colour to indicate different classes.
-corrtable- is the result. The help file is quite detailed and lays
down all kinds of reservations, not least that this can be
slooooooooooooooooooooooowwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww.
To repeat another: -corrtable- does not
support starring systems such as * ** *** for different levels of
significance and I have no intention of implementing any. I don't have,
or even want, the power to stop others cloning the program and
implementing that, but I decline to do it myself. Most P-values
for correlations are not worth much any way insofar, as the assumptions
of mutual independence and bivariate normality are often a matter of
faith.
Also, correlation matrices are almost always much less informative
than scatter plot matrices, but you know that anyway.
But I shouldn't disparage my own program. I guess some people
will enjoy playing with it a bit. If you ever need to give
a talk including little correlation matrices, this is a way
of going beyond a very mundane table -- and, positively,
highlights and patterns in your table can be made clearer
without graphamatazz.
I showed the results to someone
who uses quite different software and was told, "Oh, that
would be _very_ fiddly programming in ??????". I am not
clear how to quantify that against the few hours I spent
on this piece of Stata, but there you are.
Nick
[email protected]
*
* 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/
