On Wed, Dec 30, 2009 at 11:16 PM, Austin Nichols
<[email protected]> wrote:
> Skipper Seabold <[email protected]> :
> The -margins- manual entry is exhaustive and a good place to start,
> but in short:
> Don't divide by y but by yhat. Particularly for a zero/one depvar,
> where dividing by y just sets the negative outcomes to missing
> (dividing by zero) and leaves the positives unchanged (dividing by
I looked through it if you mean the main reference pdf, I think it was
just a notational thing for me and not actually thinking about what I
was doing.
> one). Also no need to use Mata here... and easier to see a mistake
> in Stata probably.
>
Thanks this is much cleaner.
> clear all
> use http://www.ats.ucla.edu/stat/stata/examples/greene/tbl19-1
> qui logit grade gpa tuce psi
> set type double
> predict xb, xb
> g f=invlogit(xb)/(1+exp(xb))
> g fy=f/grade
> predict p
> g fp=f/p
> su
> mat b=r(mean)*e(b)
> margins, eydx(*)
> mat li b
>
>
> On Wed, Dec 30, 2009 at 10:50 PM, Skipper Seabold <[email protected]> wrote:
>> If eydx is defined as dy/dx * (1/y), then I don't understand how the
>> following values are found by margins (using Stata 11):
>>
>> use http://www.ats.ucla.edu/stat/stata/examples/greene/tbl19-1, clear
>> logit grade gpa tuce psi
>>
>> <snip>
>>
>> margins, eydx
>>
>> Average marginal effects Number of obs = 32
>> Model VCE : OIM
>>
>> Expression : Pr(grade), predict()
>> ey/dx w.r.t. : gpa tuce 1.psi
>>
>> ------------------------------------------------------------------------------
>> | Delta-method
>> | ey/dx Std. Err. z P>|z| [95% Conf. Interval]
>> -------------+----------------------------------------------------------------
>> gpa | 1.854637 .8479031 2.19 0.029 .1927771 3.516496
>> tuce | .0624472 .0930901 0.67 0.502 -.1200061 .2449006
>> 1.psi | 1.554903 .7283702 2.13 0.033 .127324 2.982483
>> ------------------------------------------------------------------------------
>>
>>
>> Presumably it would give something like the following, if I had
>> treated psi as continuous (Please excuse the hackish mata. I'm still
>> learning my way around...). Are the missing values (NaNs or Infs,
>> depending on your take) handled in some odd way or is something
>> entirely different going on? I didn't see anything in the manual on
>> this or with a (quick) google search.
>>
>> mata
>> exog = st_data(., ("gpa", "tuce", "psi"))
>> stata(". gen cons = 1")
>> cons = st_data(., "cons")
>> exog = (exog,cons)
>> params = st_matrix("e(b)")'
>> xb = exog * params
>> pdf = exp(-xb):/(1:+exp(-xb)):^2
>> dydx = pdf * params'
>> endog = st_data(., ("grade"))
>> mfx = mean(effects :/ endog)
>> end
>>
>> Any clarifications or further references would be appreciated.
>>
>> Cheers,
>> Skipper
>>
>
> *
> * 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/
