Re: st: MI margins

[Austin Nichols <austinnichols@gmail.com>]

Dana Goin <DGoin@urban.org>:
It is a little hard to follow what you are doing, or even what you are
asking, with all the -mi- and -svy- clutter. But it seems you are
comparing mean number of doctor visits for the uninsured to those with
Medicaid, then comparing mean number of doctor visits for the
uninsured to those with private coverage.  The mean number of doctor
visits for the uninsured is apparently .8088408 but the standard error
on that estimate depends on what other observations are included in
your regression.  The set of observations (not "population" as in
"initial regression includes different populations") giving that
estimate is the same in both regressions (i.e. cases with
unin_flag==1), so you should hope to get the same estimates in each
case.

However, a linear regression of a count variable on predictors may not
give you the high-quality inference you seem to be aiming for; -glm-
with a log link is much more appropriate for a count outcome like
number of visits.

On Mon, Oct 29, 2012 at 9:46 AM, Goin, Dana <DGoin@urban.org> wrote:
> Im getting unexpected results when using the margins command with multiply imputed data. Im using Stata 12.1 for Windows.
>
> I’m estimating two linear probability models. One estimates the likelihood of a doctor’s visit for those without health insurance and for those who have Medicaid. The other model does the same (i.e. estimates the likelihood of a doctor’s
>  visit) for those without health insurance and for those with private health insurance. I’m using -margins- to predict the likelihood of a doctor’s visit if the previously uninsured got Medicaid or private coverage.
>
>
> After running the regression, I’ve used an ado file to make -margins- compatible with the mi commands. Here’s the code I used for the ado file:
>
>
> program uninmargins, eclass properties(mi)
>   version 12
>   args var cov_flag
>   svy, subpop(if (unin_flag==1 | `cov_flag'==1)): reg `var' `cov_flag' $regvars
>   margins if unin_flag==1, subpop(if (unin_flag==1 | `cov_flag'==1)) at(`cov_flag'=(0 1)) post
>
> end
>
> where `var’ is the various doctor visit indicators and `cov_flag’ is first for Medicaid and then for private coverage.
>
>
> And here’s the code in my do file:
>
> mi estimate, post esampvaryok cmdok: uninmargins `var’ `cov_flag’
>
>
> Here is my problem: the margins assigning everyone to be uninsured (i.e. `cov_flag’==0) are
> exactly the same for both models (but the standard errors are not). Does this make sense? It seems to me that the initial regression includes different populations and has different coefficients, so even though I’m restricting the margins to the uninsured
>  in both scenarios they should not be the same. Is that wrong?
>
> Here’s the first margin, for the uninsured and Medicaid:
> ------------------------------------------------------------------------------
>              |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
> -------------+----------------------------------------------------------------
>          _at |
>           1  |   .8088408   .0150372    53.79   0.000     .7793684    .8383132
>           2  |   .9046395   .0107063    84.50   0.000     .8836547    .9256242
>
> And here’s the second, for the uninsured and private health insurance:
>
> ------------------------------------------------------------------------------
>              |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
> -------------+----------------------------------------------------------------
>          _at |
>           1  |   .8088408    .014756    54.81   0.000     .7799196     .837762
>           2  |   .9119163   .0149138    61.15   0.000     .8826857    .9411469
>
> Any help would be greatly appreciated.
>
> Thank you,
> Dana Goin

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/