Hello,
I am hoping that someone could shed some light on a puzzle that I have
discovered when using Clarify to interpret results from a logit model.
It is well known that the effect of a changing the value an independent
variable x1 in a logit model depends on this variable's starting value,
as well as the values of the other independent variables in the model.
Clearly, the value of the lower and upper bounds of the p% confidence
interval surround x1's estimated effect must vary as well. Intuitively,
one would expect that x1's significance level might change too.
My question regards this last point. Specifically, I have found whenever
I use Clarify to assess the first difference (in the predicted
probability) of a given change in the value of x1, the significance
level of this effect never varies, regardless of what values the other
independent variables in my model take. However, when I use conventional
analytical techniques to calculate the same first difference and assess
whether this estimated effect differs significantly from zero, I find
that the significance level change of the estimated effect does change.
I understand that the simulation-based approach is more precise because
the delta method produces an approximation, and also because the
simulation-based method implicitly corrects for bias in the logit
formula for calculating predicted probabilities. However, because the
results produced using the simulation-based method conflict with my
intuition, I want to make sure that I am not missing something.
One thought that occurs to me is that, depending on the mechanics of the
-setx- command, the uncertainty surrounding the values of the other
independent variables is somehow not being taken into account when
assessing a change in the value of x1. This would not be an issue with
the delta method because, even though this technique produces an
approximation, it is one that reflects the multivariate nature of the
equation.
I have constructed a simple example using a publicly available Stata
dataset to illustrate.
sysuse auto, clear
set seed 12345
gen x1 = round(uniform())
logit for x1 mpg length price
mfx, at(mpg=17)
mfx, at(mpg=30)
The output shows that the p-value associated with a unit change in the
dummy variable x1 is .576 when mpg = 17, and .604 when mpg = 30.
estsimp logit for x1 mpg, sims(1000)
forvalues clev = 61/63 {
di `clev'
setx mpg 17
simqi, fd(prval(1)) changex(x1 0 1) l(`clev')
setx mpg 30
simqi, fd(prval(1)) changex(x1 0 1) l(`clev')
}
The outputs shows that, regardless of the value of mpg, the effect of a
unit change in x1 differs significantly from zero for p = .61, but not
when p >= .62.
I realize that this example is trivial, but I have found more
substantial differences using my own data. I'd very much appreciate it
if someone could tell me whether the problem is with Clarify or
with my intuition.
Regards,
Bennet Zelner
--
Bennet A. Zelner
Duke University
Fuqua School of Business
Box 90120
Durham, NC 27708-0120
bzelner(a)duke.edu
Tel +1 919 660-1093
Fax +1 919 681-6244
--
Clarify mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive: http://lists.gking.harvard.edu/?info=clarify
I was hoping somebody could detect the source of an annoying Stata error
when I try to use estsimp with mlogit and multi-imputed datasets. Here's
the command (edited to reduce the number of independent variables):
estsimp mlogit gltialar age education female nonwhite, noconstant
basecategory(2), mi(outdat1 outdat2 outdat3 outdat4 outdat5)
This gives me the generic Stata error, "no variables defined" and the 11
return code.
I have Stata set on the proper directory and it contains the files
outdat1.dta through outdat5.dta. Before running estsimp, I can run the
miest from the same directory and it does just fine.
It seems to me that it's not reading the datasets but I can't figure out
why from the Clarify documentation.
Thanks.
Ken Wald
--
Kenneth D. Wald
Distinguished Professor of Political Science
University of Florida
POB 117325
Gainesville, FL 32611-7325
Voice Mail: 352-273-2391
Home Page: web.clas.ufl.edu/users/kenwald
--
Clarify mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive: http://lists.gking.harvard.edu/?info=clarify