st: Re: st: BOOTSTRAP: the standard errors of marginal effects of MIXLOGIT‏

[Arne Risa Hole <arnehole@gmail.com>]

Dear Kayo

Your procedure is incorrect as the model needs to be re-estimated for
each bootstrap sample. In other words your “marginal_inc” program
should include the call to -mixlogit-. Whether this is practical or
not depends on how long it takes to estimate your model – you may be
in for a long wait!

See this thread
<http://www.stata.com/statalist/archive/2010-11/msg01025.html> for an
example of how -bootstrap- can be used with -mixlogit-.

Arne

PS Note that you can bootstrap several statistics in one go – you
don’t need to run -bootstrap- for each marginal effect.

On 3 June 2012 09:01, nagi kayo <kayonagi@hotmail.co.jp> wrote:
> Dear Professor Arne Risa Hole and all
>
> I estimated mixed logit model using the command -mixlogit- and
> calculated the marginal effects using the command -mixlpred-.
> (Once again, I greatly appreciate that professor Arne Risa Holl
>
> gave me advice on how to calculate the marginal effects using
>
> "mixlpred" last february)
>
> However, now I am having trouble obtaining the standard errors and
> p-value of marginal effects using the bootstrap.
>
> First, please let me explain the data.
> My data set is like below.
> (Actually, my data set inlude 10,000 id, but here I show you the data only on two id
> for simplicity.  In addition, the data I used in the estimation include
>
> the data on many household characteristics such as age, wealth, education level.)
>
> id  alt  d   d1   d2   d3  d1inc  d2inc  d3inc  d4inc    p
>  1   1   1    1    0    0    665     0      0      0   0.214
>  1   2   0    0    1    0      0   665      0      0   0.186
>  1   3   0    0    0    1      0     0    665      0   0.381
>  1   4   0    0    0    0      0     0      0    665   0.219
>  2   1   0    1    0    0    779     0      0      0   0.553
>  2   2   1    0    1    0      0   779      0      0   0.301
>  2   3   0    0    0    1      0     0    779      0   0.107
>  2   4   0    0    0    0      0     0      0    779   0.039
>
>
>
> id: households
> alt: 1=households are worried about their retirement life
>       because the pension benefit is NOT enough.
>     2=households are worried about their retirement life
>       for some reasons other than pension.
>     3=households are NOT worried about their retirement life
>       because the pension benefit is enough.
>     4=households are NOT worried about their retirement life
>       for some reasons other than pension.
> d: dummy which equals one if households choose "alt" in the same row.
>    (so id 1 chose alt 1, and id 2 chose alt 2.)
> d1-d3: intercepts
> d1inc-d4inc: real households income
> p: an alternative specific variable
>
>
>
> using the above data, i did the estimation as follows:
>
>
>
> **********************************************************************
> mixlogit d d1 d2 d3 d1inc d2inc d3inc, group(id) rand(p)
> mixlpred prep_base
> preserve
> quietly replace d1inc=d1inc+1 if alt==1
> quietly replace d2inc=d2inc+1 if alt==2
> quietly replace d3inc=d3inc+1 if alt==3
> quietly replace d4inc=d4inc+1 if alt==4
> mixlpred prep_inc
> cap program drop marginal_inc
> program marginal_inc, rclass
>  quietly sum prep_base if alt==1
>  local av_prep_base1 = r(mean)
>  quietly sum prep_inc if alt==1
>  local av_prep_inc1 = r(mean)
>  return scalar marge_inc1 = `av_prep_inc1' - `av_prep_base1'
>  quietly sum prep_base if alt==2
>  local av_prep_base2 = r(mean)
>  quietly sum prep_inc if alt==2
>  local av_prep_inc2 = r(mean)
>  return scalar marge_inc2 = `av_prep_inc2' - `av_prep_base2'
>  quietly sum prep_base if alt==3
>  local av_prep_base3 = r(mean)
>  quietly sum prep_inc if alt==3
>  local av_prep_inc3 = r(mean)
>  return scalar marge_inc3 = `av_prep_inc3' - `av_prep_base3'
>  quietly sum prep_base if alt==4
>  local av_prep_base4 = r(mean)
>  quietly sum prep_inc if alt==4
>  local av_prep_inc4 = r(mean)
>  return scalar marge_inc4 = `av_prep_inc4' - `av_prep_base4'
> end
> bootstrap r(marge_inc1), reps(1000): marginal_inc
> bootstrap r(marge_inc2), reps(1000): marginal_inc
> bootstrap r(marge_inc3), reps(1000): marginal_inc
> bootstrap r(marge_inc4), reps(1000): marginal_inc
> restore
> ************************************************************************
>
>
> The results of bootstrap is as follows:
>
>                            Standard errors       P-value
> marge_inc1      -0.0011        4.44e-06           0.000
> marge_inc2      -0.0008        2.70e-06           0.000
> marge_inc3       0.0013        5.41e-06           0.000
> marge_inc4       0.0007        3.36e-06           0.000
> ************************************************************************
>
>
> Apparently, there is no problem in the above results.
> However, all the standard errors of marginal effects of ANY explanatory variable
> (e.g. age, education level, risk aversion) are very small,
> and all p-values of marginal effects of ANY explanatory variables are 0.000.
> I am wondering if my program has something wrong.
>
> Could anyone give me any comments?
> I greatly appreciate it.
>
> with best wishes,
> KAYO
>
> *
> *   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/