Re: st: adjust vs. margins revisited

[Tim Wade <wadetj@gmail.com>]

Everyone, thanks for your thoughts on this issue, I also wrote to tech
support with this query and received a thoughtful response, but  the
main points were also covered in this thread.  Tech support emphasized
the flexibility of -margins-, and discussed the ability of -margins
-to calculate the average of the predicted values (the "AME" approach
referred to by Michael Mitchell and discussed by others in this
thread), which I agree is a very useful addition.   Here is a snippet
of the response:

"
As he mentioned, the -adjust- command with the -pr- option calculates
confidence intervals by transforming the endpoints from confidence intervals
that are created based on the linear prediction, xb.  On the other hand, the
-margins- command creates a symmetric confidence interval around the
calculated margin.

Both formulations of the confidence interval are valid and are based on an
assumption of being quadratic in some space.  The difference is that -adjust-
command is much more limited than -margins- in the types of predictions that
it can work with.

...

In addition, when you specify the -atmeans- option with -margins- (or use
-adjust-), transforming the end points of a confidence interval created for xb
makes sense.  However, -margins- is also more flexible in that it allows you
to, and by default does, calculate the average of the predicted values instead
of the predictions where every covariate is set to a particular value.  It is
not clear that transforming the end points of a confidence interval is valid
in when we take the mean of the predictions.

I mentioned above that both formulations of the confidence interval are valid.
However, when the predictions are close to 0 or 1, it is certainly possible to
obtain confidence intervals that lie outside the range of 0 to 1 with the
methodology used by -margins-.  When the confidence intervals are created
first in the -xb- space, the boundary conditions are acknowledged.

Lastly, I realize that the help file for -adjust- says that -margins- can do
everything it did.  These transformed confidence intervals are one case where
this is not true.  I have discussed this with our developers, and we will be
removing this statement from the -adjust- help file in order to help eliminate
confusion on this issue."


Thanks again, Tim



On Fri, Jul 30, 2010 at 6:53 PM, Alan Neustadtl
<alan.neustadtl@gmail.com> wrote:
> I had asked the list (with no response) and then Stata technical
> support about using -margins- to replicate the standard error of the
> forecast given by -adjust-.  Currently, -margins- does not do this
> directly.
>
> The response from Stata Corp can be found at
> http://www.stata.com/statalist/archive/2010-03/msg00942.html.
>
> Best,
> Alan
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>

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