 Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

Re: st: Margins after mlogit with survey data

 From kmacdonald@stata.com (Kristin MacDonald, StataCorp) To statalist@hsphsun2.harvard.edu Subject Re: st: Margins after mlogit with survey data Date Wed, 08 Sep 2010 14:31:31 -0500

Joao Lima <jricardofl@gmail.com> used -margins- after -svy:mlogit- and the
marginal effects were reported as "not estimable".

This may be caused by numerical instabilities in determining estimable
functions with some -svy:mlogit- model fits.  We believe we have a fix for
this; the fix should be available in the next executable update.

In the meantime, Joao may be able to obtain the marginal effects by specifying
a different base outcome in the -svy:mlogit- command.  Although the estimated
coefficients change based on the choice of base outcome, the marginal effects
will not change.  Using an example similar to Joao's, we can see that the
marginal effects are not estimated when specifying the -baseoutcome(1)- option
in -svy:mlogit-.  However, if we specify -baseoutcome(4)-, -margins- will
calculate the marginal effects.

***** BEGIN:
. webuse nhanes2f

. svyset psuid [pweight=finalwgt], strata(stratid)

pweight: finalwgt
VCE: linearized
Single unit: missing
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>

. svy: mlogit  region i.sex i.race c.age##c.age sizplace, baseoutcome(1)
(output omitted)

. margins, vce(unconditional) dydx(*) predict(outcome(1)) atmeans

Conditional marginal effects                    Number of obs      =     10337

Expression   : Pr(region==NE), predict(outcome(1))
dy/dx w.r.t. : 2.sex 2.race 3.race age sizplace
at           : 1.sex           =    .4795785 (mean)
2.sex           =    .5204215 (mean)
1.race          =    .8790162 (mean)
2.race          =    .0956152 (mean)
3.race          =    .0253686 (mean)
age             =    42.23732 (mean)
sizplace        =    4.808137 (mean)

------------------------------------------------------------------------------
|             Linearized
|      dy/dx   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
2.sex |  (not estimable)
|
race |
2  |  (not estimable)
3  |  (not estimable)
|
age |  (not estimable)
sizplace |  (not estimable)
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.

. svy: mlogit  region i.sex i.race c.age##c.age sizplace, baseoutcome(4)
(output omitted)

. margins, vce(unconditional) dydx(*) predict(outcome(1)) atmeans

Conditional marginal effects                    Number of obs      =     10337

Expression   : Pr(region==NE), predict(outcome(1))
dy/dx w.r.t. : 2.sex 2.race 3.race age sizplace
at           : 1.sex           =    .4795785 (mean)
2.sex           =    .5204215 (mean)
1.race          =    .8790162 (mean)
2.race          =    .0956152 (mean)
3.race          =    .0253686 (mean)
age             =    42.23732 (mean)
sizplace        =    4.808137 (mean)

------------------------------------------------------------------------------
|             Linearized
|      dy/dx   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
2.sex |  -.0110787   .0073234    -1.51   0.140    -.0260149    .0038575
|
race |
2  |  -.1496754    .029835    -5.02   0.000    -.2105243   -.0888265
3  |  -.1779278   .0299411    -5.94   0.000    -.2389932   -.1168625
|
age |   .0006879   .0004459     1.54   0.133    -.0002216    .0015974
sizplace |  -.0198073   .0087242    -2.27   0.030    -.0376004   -.0020143
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.
***** END:

If changing the base outcome does not produce estimable marginal effects, Joao
can email tech-support@stata.com for more support on this.

-- Kristin					-- Jeff