You should never pay much attention to the alpha in the extended model,
unless you believe the model holds exactly. The best use of the extended
model, as the basic model, is to go all the way to the precinct-level
estimates, since these (and not alpha) rely on the bounds which make
inferences less sensitive to incorrect modeling assumptions. Then from
these, compute any quantity you may be intersted in.
I don't see any point in a test of "logical consistency" except for some
theoretical purpose unrelated to inference. The only relevant question is
how far the estimate is from the truth, which will depend on the width of
the bounds and, if the bounds are wide, on the modeling assumptions.
(Was it Huey Long who accused a campaign opponent of being a "known
matriculator"?)
Gary
On Thu, 25 Sep 2003, Kenneth R. Benoit wrote:
Hi Gary and list --
I have a few questions about EI-R and how to implement practically the
AKHS prescriptions on using precinct-level EI estimates in second-stage
regressions. for those of you who might know of it, I also refer below
to a forthcoming article by HS in AJPS that revisits poor Burden and
Kimball's 1998 APSR article. It develops a logical consistency test for
potential EI-R applications, also makes strongly the case for the
dominance of the extended EI model over EI-R.
I would like to use extended EI directly, but am having some
difficulties in accessing and interpreting the alpha parameters, also in
fitting this into the 2x3 ei/ei2 framework.
My problem attempts to exstimate and explain differences between list PR
ballot voting and candidate-based single-member district ballot voting
in the 1996 italian election. To summarize, here's my setup. I have a
2x3 table where I am interested in the total table values of the top row
only. I run ei and ei2 to estimate beta^b_i and lambda^b_i, then
combine them to get precinct-level parameters of interest. e.g. row 1
col 1 = beta^b_i*lambda^b_i. (I rely on the vector of _Esim simulated
precinct-level parameters to do the multiplication, then take the mean
for the combined parameter of interest.)
for covariates I use zb = a scalar measure of policy position (the
left-right midpoint between the two candidates in the single-member
district). No covariate for zw since this is not a quantity of direct
interest. The zb covariate is included in both ei and ei2.
those combined params of interest are my precinct-level estimates of
split voting. I then want to estimate the effect of policy positions of
the precinct-candidates -- which varies by precinct -- on split voting.
this involves using the estimates of the precinct-level params of
interest as a dependent variable in EI-R. My reading of the lit I refer
to above, and my e-mail exchanges with M Herron, all indicate that i
should avoid EI-R and simply rely on the estimates of alpha from the
extended EI model. I am willing to follow this advice, but have a few
issues to think about and resolve.
1) According to AKHS, the bias from the HS/HS2-type problem can be
detected based on the width of the bounds. My mean bound on lambda^b_i
for example is wide at .82, putting it squarely into the "bias zone".
This indicates I should avoid EI-R.
2) According to the test for logical consistency in HS-AJPS, I can
detect a problem by regressing Zb_i on X_i and conducting a Wald test.
My application actually passes this test, since the wald test fails to
be significant for either X on zb from ei, or on (estimated) X from zb
from ei2. This indicates possibly that EI-R would be okay. btw gary
the test also promotes hair growth.
3) It seems from the interpretation of alpha^b in HS-AJPS that it can be
directly interepreted as the effect of zb on shifts in the mean
distribution from which beta^b_i is drawn. (For the superpopulation
case which is what I am interested in.) That's okay with me. But where
do I obtain this estimate? Is this the output column from
eiread(db,"sum") labeled:
Maximum likelihood results in scale of estimation (and se's)
Zb0 Zb1
this would not seem to be alpha since the Zb0 seems to be a constant,
and alpha is after entering zb as a deviation from its mean vector
(thereby removing the constant). Do I have to estimate the alphas by
hand based on p170 of your book? If so this is hardly something most
people would be able to readily do.
4) AK states that "2nd stage analyses...have the advantage of being
easier to understand, use, and evaluate" which is certainly true, all
the more so for the 2x3 case as in my application. For me the best
feature is that I have combined the beta^b_i and lambda^b_i estimates to
get my overall split voting quantities of interest, and then can regress
these directly. If I scrap the EI-R and use the extended model's alpha
estiamtes directly, how can I combine the alphas from ei and ei2 to get
measures of influence on my overall 2x3 table quantities of interest?
(and still preserve standard errors?)
Advice is much appreciated!!
Ken
................................................................
Kenneth Benoit
http://benoit.tcd.ie
Department of Political Science mailto:kbenoit@tcd.ie
Trinity College Tel: 353-1-608-2491
Dublin 2, Ireland...........................Fax: 353-1-677-0546
-
EI mailing list served by Harvard-MIT Data Center
Subscribe/Unsubscribe:
http://lists.hmdc.harvard.edu/?info=ei