First, a couple of questions: did you enter the dummy, the trade variable, and the interaction of the dummy with the trtade variable in the whole-sample analysis?
If you did, then the implied intercepts and slopes in the 2 subsamples will be identical whether U est in 1 smpl or split smpls.
I say implied slopes & intercepts meaning the ones implied for the subsamples & not just the coefficients if U entered the variables in the way I just said. And that you are then comparing the comparable combination of terms in the pooled estimation to the proper std err for that combination.
If you did enter things that way, & you understand properly what combinations to compare, if all that, and implied slopes or intercepts are actually not same in the two estimations, then:
Second, did Stata (I am assuming) automatically "help" you by dropping some variables or observations in one or the other? If no to 2 but yes to 1, & slopes & int's not the same, then throw away your software and never use it again.
If not 1) or 2), & the software is right (slopes & intercepts same), and not some other simple mistake of the sort I usually make, then I can guess one more way its actually "correctly" possible, but probably it's something like 1 or 2. The slopes are the same but s.e. for one is much grtr, yes? It's not the different N in the subsamples, b/c this should be properly accounted either way. But sigma-hat, the estimated error variance in the two subsamples could be radically different. The only diff b/w split smpl & one smpl, each done properly, is that latter will give just one estimate for one sigmahat. That one will basically wtd-average the two you get by split-smpl. If the one in the early developers smpl is very high & the other much lower, what U describe could happen & would be right. The right way to do the joint sample in that case would be to do some sort of wtd least squares, wtng by development.
But, like I said, in my personal experience, some much more mundane sort of error on my part is usually the culprit...
Hope that all made sense from my blackberry thumbing, & then is helpful.
Rob
Sent via BlackBerry by AT&T
-----Original Message-----
From: "Carlos Rodriguez" <[log in to unmask]>
Date: Mon, 28 Jul 2008 18:03:40
To: <[log in to unmask]>
Subject: [POLMETH] interactions and analyses for subsets of the sample
Dear POLMETH list,
I am working with TSCS data and would appreciate some advice on this issue.
I have run some regressions with an interaction between trade and a
dummy for late industrializers. The interaction was not significant.
However, I reran the analysis dividing my dataset into two groups:
early industrializers and late industrialzers. It turns out that
trade iIS significant in the set of late industrializers. Why would
the interaction between trade and late industrializers in the whole
sample fail to achieve significance whereas trade comes up as highly
significant when the regression is run separtely on the group of
countries that are late indutrialers? Can the difference be put down
to the dissimialr Ns? Moreover, which result shall I "believe"?
thanks for your advice.
Carlos Rodriguez
**********************************************************
Political Methodology E-Mail List
Editors: Melanie Goodrich, <[log in to unmask]>
Delia Bailey, <[log in to unmask]>
**********************************************************
Send messages to [log in to unmask]
To join the list, cancel your subscription, or modify
your subscription settings visit:
http://polmeth.wustl.edu/polmeth.php <http://polmeth.wustl.edu/polmeth.php>
**********************************************************
|
