Re: st: problem with marginal effect after running a logit regression

[Rieza Soelaeman <rsoelaeman@gmail.com>]

Hi Jeremy,
Your advisor is correct that the coefficients of a logistic regression
cannot be interpreted in the same way as OLS.  Using the margins
command allows for an estimation of the marginal effect (e.g. the
increase in probability of your outcome = 1, here I assumed outcome is
binary). One question for you: when your advisor meant by "at median,"
did he mean at median values for all the characteristics in your
model, or just the median level of education?

If the specific effect of interest is going from mstudymid to
mstudyhigh, I would suggest making mstudymid the reference category in
your set of dummy variables for education.  Here I assume you have
mstudylow as the reference (excluded) category.  If you make mstudymid
your reference, then the marginal effect of mstudyhigh would be the
marginal effect of going from mstudymid to mstudyhigh.  Similarly, the
marginal effect of mstudylow would be the marginal effect of going
from mstudylow to mstudymid.

Typically, if your predictors are continuous, it makes sense to have
Stata calculate marginal effects at the means of each value of your
predictors. This can be achieved by executing the following command
after running your regression:

margins, atmeans

However, because your predictors are categorical (or if you are using
a version of Stata before Stata 12), you may be able to get away with
specifying criteria for the "typical" individual in your dataset for
which you are calculating the marginal effect.  Then justify the
choices you made in describing the "typical" individual.

For example, in your dataset, the "typical" individual may be a 35
year old, male, who is a chief wage earner, with high education,
mintpol = "mid", mpol = "right", and mincome = "high," then the
command you would run would be something like:

mfx, at (mstudymid=0 mstudyhigh=1 mhomme=1 mchiefwageearner=1 mage28_37=1
mage38_47=0 mage48_57=0 .............. mincomehigh=1)

*Note the ........... means you should assign a 0 or 1 value for your
categorical predictors as appropriate to describe your person.

I see there are several variables in your dataset that could benefit
from being continuous, though.  If age were continuous, you can simply
plug in the average age (from any of the univariate commands you can
use to describe the mean of a vbl).  Same thing with income.  I think
it would make your regression more robust to use the continuous.

Of course using this method (with -mfx-) is complicated by the
clustering in your data and the interactions between the cluster
variables S003 and S002 (it appears to me these are polychotomous
categorical variables, as you have used the i. in adding them to your
regression).  Because I don't know what they represent and how many
levels of each they are, I am not sure how they would be specified in
the -mfx- command.  Do you absolutely need to know the marginal effect
of each of those clusters, or were they included just so you can
control for them?  If you included them just to control for them,
consider using -xtmelogit- (mixed effects logit) instead, and specify
S003 and S002 for random intercept calculation.

HTH,
Rieza

*I invite other statalisters to correct me if I have said something in error
above.

On Thu, Jul 26, 2012 at 2:17 PM, Jeremy Franklin <jfrankli@ulb.ac.be> wrote:
> Dear all,
>
> Here is my little trouble:
>
> For my master degree thesis I decided to test for the role of education level in assession the importance of fighting inflation.
>
> Here is my final regression formula:
>
> xi: logit mfirstchoice  mstudymid mstudyhigh mhomme mchiefwageearner mage28_37 mage38_47 mage48_57 mage58 mintpollow mintpolmid mintpolhigher mpolleft mpolright  mincomemid mincomehigh i.s003 i.s002 i.s003*i.s002, vce(cluster s003)
>
> I hate the results but my thesis coordinator told me that the results of logit regression cannot be interpreted like coefficients of a linear regression. Therefore, he suggested me to check for the marginal effects at the median in order to see the marginal effects of one individual coming from mstudymid to mstudyhigh
>
> I googled everything, i tried hundreds of formulas, both with mfx and margins but i still cannot find the correct one in order to interpret my results.
>
> Can ANYONE help me please.
>
> ps: a robustness test included in my thesis include the following formula (this time with ologit)-
>
> xi: ologit minflation  mstudymid mstudyhigh mhomme mchiefwageearner mage28_37 mage38_47 mage48_57 mage58 mintpollow mintpolmid mintpolhigher mpolleft mpolright x047 i.s003 i.s002 i.s003*i.s002, vce(cluster s003)
>
> *
> *   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/