Re: st: nlcom question

[Austin Nichols <austinnichols@gmail.com>]

john metcalfe <johnzmetcalfe@gmail.com>:
The calculations will all be easier if you run a -poisson- regression
of y on X instead of OLS regression of ln(y) on X, since the
predictions of mean y given X no longer include an estimate of the
variance of error terms. To replicate -margins- output, you can simply
-predict- after you -replace- X variables with appropriate patterns as
in e.g.
http://www.stata.com/statalist/archive/2009-02/msg00081.html
and if you wrap all the predictions and -summarize- commands in a
-program-, you can bootstrap the whole thing, or you can use the delta
method to calculate SEs by hand.  See also
http://gelbach.eller.arizona.edu/~gelbach/ado/margfx.pdf
or -margeff- (use -findit margeff- to find the latest version).

On Thu, Mar 4, 2010 at 4:23 PM, john metcalfe <johnzmetcalfe@gmail.com> wrote:
> Thanks, Stas. How could I replicate Stata 11's -margins- command using Stata
> 10? (I went ahead and ordered the upgrade but it won't arrive until next
> week.)
> Thanks,
> John
>
> On Wed, Mar 3, 2010 at 7:46 PM, Stas Kolenikov <skolenik@gmail.com> wrote:
>
>> If the distributions of the dependent variable are the same for two levels
>> of the categorical factor, then they will be the same no matter whether you
>> transformed them or you did not. Hence it suffices to use -test- rather
>> than
>> -nlcom-. The latter will be answering a more subtle question of whether the
>> means are the same (you implicitly put zeroes for all other variables,
>> which
>> may or may not be appropriate); you still may have differences in variance,
>> skewness, kurtosis, etc. between groups even if you find the means to be
>> the
>> same.
>>
>> Stata 11 has new -margins- command; have you looked at it?
>>
>> On Wed, Mar 3, 2010 at 9:31 PM, john metcalfe <johnzmetcalfe@gmail.com
>> >wrote:
>>
>> > Dear Statalist,
>> > I have a simple question I hope someone can help me with.
>> > I am using OLS with robust variance estimators to model a continuous,
>> > log-transformed DV ranging from 0 to 10 in increments of 0.01 (this is
>> > an immunologic test in common use in the U.S.). My goal is to
>> > determine whether or not there are differences in this test
>> > performance according to a categorical independent variable (rax, 4
>> > levels) with an interaction term (nt, 3 levels) and other categorical
>> > covariates, as below:
>> >
<snip>
*
*   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/