[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
RE: st: question on chow test
At 11:19 AM 2/20/2005 -0800, Daniel Schneider wrote:
No, the tests are different. -testparm g2x1 g3x1 g4x1- tests whether there
are differences in the effects of x1 between any of the 4
groups. -testparm g2x1 g3x1 g4x1, equal- allows the three groups to differ
from the reference category; but it tests whether the effect of x1 is the
same in groups 2, 3 and 4. For example, suppose your groups are white,
black, other, with white as the reference category. The effect of x1 for
blacks and others might significantly differ from the effect for white; but
the effect of x1 may not be significantly differ between blacks and others.
You might do things like this if you are trying to identify what the
important contrasts are. Is it sufficient to simply distinguish between
whites and non-whites, or are there differences among the subgroups of the
If you say "testparm g2x1 g3x1 g4x1, equal" tests whether the effects of
x1 are the same in all groups, what is the difference to "testparm g2x1
g3x1 g4x1" which should test the same thing (at least that is what I
figured out and what you confirmed some lines above)?
Well (a) X1 and X2 have to be measured in the same metric of course, and
(b) you'd still want to do significance tests as to whether, say, g4x2
really was larger than g4x1. Doesn't much appeal to me, but maybe if I
knew what the variables were and the underlying theory I'd think differently.
Lets assume I have following coefficients:...
I also have a second group of coefficients:...
Again, differences among groups are statistically significant.
Can I make a meaningful discussion of the values of the differences
between the coefficents, for example claiming that the differences for
x2 are stronger because the differences are larger? I am personally feel
uncomfortable with just talking about statistical significance without
talking about effect strength and in this case the effect strength
should be buried in the difference, but I am not sure if it is correct
to really discuss the differences in this way.
I'd probably want to run the single regression. I'd also probably want to
focus on the most critical contrasts, e.g. do the effects of x1-x4 differ
across groups or don't they? If you do want to try out all the different
possibilities, you should go for more stringent significant levels. But,
the more complicated you make this, the more difficult it will be to
interpret, and the more likely you are to come up with results that may be
ideosyncratic to your data set. Just my opinion, others may disagree.
> There are potentially zillions of tests you can be doing
> here, so you want
> to be careful you aren't just capitalizing on chance.
Sure, that's correct. My only problem is that I have four variables that
all could be different across groups, but don't have to (actually I have
more, but I am willing to hold the others constant based on theoretical
assumptions and my research interests).
Or is it better to run several separate regressions and leting only one
of the variables vary across groups while holding the others constant?
In that case I could use the simple testparm command as outlined in the
FAQ to test the differences of a specific variable.
* For searches and help try: