If you stick to the solely aggregate analysis, you can avoid ecological
inference altogether. This won't work (for good reasons) with your
reviewers, but here's how it would work: you'd say that %indigenous
causes %voting. you'd then be indifferent between (a) the indigenous
people voting for the indigenous candidate and (b) and the nonindigenous
people who happen to live in areas with lots of indigenous people are the
ones voting for the indigenous candidate (controlling for the other
variables in your regression). This would be fine, but the different
stories here are pretty massive in most applications. So what you might
do is to first do your analysis and explicitly say that, since you're not
doing ecological inference, you can't distinguish between (a) and (b).
And so you would then move to a method of ecological inference to
distinguish between the two.
Best of luck with your research,
Gary
: Gary King, King(a)Harvard.Edu
http://GKing.Harvard.Edu :
: Center for Basic Research Direct (617) 495-2027 :
: in the Social Sciences Assistant (617) 495-9271 :
: 34 Kirkland Street, Rm. 2 HU-MIT DC (617) 495-4734 :
: Harvard U, Cambridge, MA 02138 eFax (928) 832-7022 :
I think the way I'd present it if I were you
On Wed, 20 Nov 2002, Beck, Scott H. wrote:
Professor King:
My colleague, Ken Mijeski (a political scientist) and I (sociology)
have been analyzing election data from Ecuador the past few years. This
is aggregate data in that the voting results and other variables are
only available at the level of the county (in Spanish, canton). One
primary issue we have dealt with in our analyses concerns whether
indigenous voters (Indians) in Ecuador more often vote for candidates
sponsored by the indigenous political party. Using the aggregate data,
we ran correlations and regressions and not surprisingly found that in
almost all cases there is a substantial positive correlation between the
%Indigenous in a county and the %Voting for Indigenous Party Candidate.
So, the possibility of the ecological fallacy raises its very ugly head
and journal reviewers are quick to jump on the conclusion that
indigenous voters much more often vote for those candidates (since we
know quite a lot about Ecuador and its politics we don't feel that the
alternative hypothesis, i.e., non-indigenous voters in counties with
higher proportions of indigenous people much more often vote for
indigenous party candidates). So, we have been told to learn and then
use your innovative method of estimating and then inferring the
proportions of indigenous (black) and non-indigenous (white) who vote
for candidate x. Though our aging brains struggled through your book
(despite your clear exegesis) and we have downloaded EZi and used it, we
think, successfully (at least we can understand most of the output).
That output backs up our original conclusion - a higher proportion of
indigenous voters voted for indigenous party candidates than did
non-indigenous voters.
Our question is this: can we properly use the results of your
method as a verification technique of the hypothesis, and once verified,
go back to our previous line of analysis where we are more interested in
the size of the correlation (standardized beta) between %Indigenous and
%Voting for Indigenous Party candidate, including OLS multiple
regression where we introduce "control variables" to determine whether
the net effect of %Indigenous remains intact? That is, our interest
ultimately is more of a causal analysis at the aggregate level where
competing factors, such as poverty rate and illiteracy rate at the
county level, can be taken into account.
Your answer to this issue would be greatly appreciated and we thank
you in advance.
Scott Beck
-
EI mailing list served by Harvard-MIT Data Center
Subscribe/Unsubscribe:
http://lists.hmdc.harvard.edu/?info=ei