Leslie,
I agree with Brandon -- that sounds a complicated problem and there might
not be a way to do it. To take a stab at it, maybe you could take the
predicted values from the first stage and put them on the x-axis. For the
y-axis, you could bin the observations to get probabilities, so that the
x-axis is the predicted values from the first stage of your independent
variable and the y-axis is the binned probabilities given the value of the
predicted independent variable. I'm not sure it would work, especially
because you might want to reflect the uncertainty from your predicted
values, but give it a try.
Best,
Molly
On Tue, Apr 26, 2011 at 8:05 PM, Brandon Stewart
<brandonmstewart(a)gmail.com>wrote;wrote:
Leslie,
I don't know; my take on this would be that you just want to graph
this the same way you graph normal causal inference estimates. IV is
just the identification strategy. I'm not sure there is anything to
graph after you've made the estimate. Its sort of like graphing the
coefficient in a logit model. Its not clear that this is the quantity
of interest. You need to convince people that the assumptions hold
relative to the instrument but this is a qualitative not a
quantitative argument. Maybe I am missing something about what you
are intending though?
Gelman and Hill (2007) have a fantastic appendix on statistical
graphics which I would encourage everyone to read. Its only a few
pages but it does a great job of breaking down the basic wisdom of
these things.
Brandon
On Tue, Apr 26, 2011 at 6:01 PM, Leslie Finger <lfinger(a)fas.harvard.edu>
wrote:
Hey class,
Does anybody know of any awesome ways to graph IV results? We're dealing
with a
binary outcome (school enrollment), and we'd
love to include some
powerful
visual of our results, but I'm having
trouble conceptualizing what an IV
graph
might look like, especially with a binary outcome
(i.e. the standard
bivariate
plot doesn't do it). I have been playing
with graphing the data in ways
that
speak to what we do with the IV, but don't
actually use the IV results.
The
internet hasn't been too helpful.
Looking forward to any thoughts you have.
Best,
Leslie
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu