Re: st: mlogit coefs

[Chiara Mussida <cmussida@gmail.com>]

Unfortunately the estimates are not robust.
My trouble is to understand why the results of the same mlogit carried
out on the same covariates and by assuming the same baseline category
do differ if based on the overall sample with respect to the ones
based on a subsample. Might this be due to the issue of sample
selection?

On 20/04/2012, Chiara Mussida <cmussida@gmail.com> wrote:
> Thanks for all suggestions. I checked my dataset containing all the
> employed covariates:
>
> mlogit transition male_unmarried female_married female_unmarried age
> agesq ncomp child northw northe centre Ubenef edu1 edu2 health
> qu1nolav qu3nolav qu2nolav nopersincnolav noothersineq qu1ot qu2ot
> qu3ot if age>=15 & age<=64, b(3)
>
>
> the dummy indicators for quantile of non labour income "qu1nolav
> qu3nolav qu2nolav" takes the value one only for few individuals n each
> subsample of analysis. e.g. 9 indivv in qu1nolav for the first outcome
> (transition==1), etc. I'm going to try to substitute the indicators
> for the quantiles with a dummy variable for the presence absence of
> non labour income.
> My question is more general: Is it possible that specific dummy
> variables that take the value 1 for few few indviduals do gen not
> robust results? I mean The above mlogit results differ to the ones
> obtained on only a subsample of the initial sample (e.g. 3 transitions
> instead of 9).
> I tried to re-estimate the model without the specific dummies quoted,
> and the results semm to be robust to the two alternative model
> specifications mentioned.
>
> Thanks
> Chiara
>
>
>
> On 17/04/2012, David Hoaglin <dchoaglin@gmail.com> wrote:
>> Dear Chiara,
>>
>> I have a comment, much more minor than Maarten's, but still useful.
>>
>> If the contribution of age is nonlinear, it may not be satisfactory to
>> assume that the nonlinearity is quadratic (in practice it often is
>> not).  You did not mention the number of observations; but since you
>> have the entire labor force, you may have enough data to approach the
>> functional form of age empirically.  One strategy would separate the
>> values of age into disjoint intervals (as narrow as the data will
>> support), include in the model a dummy variable for each interval
>> except one, and plot the fitted coefficients of those dummy variables
>> against the midpoints of the intervals.  If that plot looks quadratic,
>> fine.  But it may suggest that a linear spline would be a better
>> summary of the contribution of age (taking into account the other
>> variables in the model).
>>
>> David Hoaglin
>>
>> On Tue, Apr 17, 2012 at 10:00 AM, Chiara Mussida <cmussida@gmail.com>
>> wrote:
>>> Dear All,
>>> I run a mlogit model for 9 labour market outcomes (transitions between
>>> the three states of employment unemployment and inactivity, therefore
>>> 6 transitions and 3 permanences), like:
>>>
>>> mlogit transition male_unmarried female_married female_unmarried age
>>> agesq ncomp child northw northe centre Ubenef edu1 edu2 health
>>> qu1nolav qu3nolav qu2nolav nopersincnolav noothersineq qu1ot qu2ot
>>> qu3ot if age>=15 & age<=64, b(3)
>> *
>> *   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/
>>
>
>
> --
> Chiara Mussida
> PhD candidate
> Doctoral school of Economic Policy
> Catholic University, Piacenza (Italy)
>


-- 
Chiara Mussida
PhD candidate
Doctoral school of Economic Policy
Catholic University, Piacenza (Italy)
*
*   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/