[Urgent]How to estimate marginal effects at means in Probit?

[bengal007]

Hello,

I would appreciate some hands-on advice on the following.

I am estimating impact of a number of factors on likelihood of school attendance (such as parent's education, household wealth etc). School attendance is a binary variable which =1 if child is currently attending for school, hence a Probit model. In order to interpret values of the estimated coefficients, I am looking to estimate an average partial effect for Probit. I had a look at the threads attached below but I am still having difficulty estimating the marginal effects in EViews.
http://forums.eviews.com/viewtopic.php? ... fect#p7846
http://forums.eviews.com/viewtopic.php?f=4&t=659

I am using eViews 6.

A generous pint for the best answer.

Best wishes,
Dominik

[EViews Glenn]

The easiest way is to use the forecast to create your average value and then go from there.

From your estimated equation:

1. Proc/Forecast, choose Index. This will create the XB series with values for each person in your forecast sample. If you do it by command, you can use:

Code: Select all

eq1.forecast(i) xbf

2. Compute the mean of this series. You can put it in a scalar if you want

Code: Select all

scalar meanxb = @mean(xbf)

Note that because of the linearity of the index, the mean of the index is the same as the index evaluated at the means.

3. Compute the marginal effects weighting term. Again, putting into a scalar,

Code: Select all

scalar meffectw = @dnorm(-meanxb)

Note that you could combine 2 and 3 into one step

Code: Select all

scalar meffectw = @dnorm(-@mean(xbf))

Note also that the sign doesn't matter for probits and logits. It does for the extreme value form of the binary model.

4. Compute the marginal effects for each coefficient

Code: Select all

vector meffects = meffectw*eq1.@coefs

Your desired result is in the vector meffects.

[bengal007]

Glen,

Thank you very much for your prompt reply! I have now obtained the results and can confirm that they are very much identical with the manual check against the normal distribution function after rescaling.
While this is already done by EViews, what approach do you reckon is the best to manually calculate the scaling factor to enable direct comparison of Probit and LPM probability estimates?

Best wishes,
Dominik

[EViews Glenn]

It's a difficult question. It sounds like what you want are values computed at a "representative individual". One way to do the representative individual is to evaluate at means, as you have done. One might also want to evaluate at various quantiles of the distribution. An alternative is to evaluate at each of your data points, then to compute moments of the marginal effects. But there's really no right answer. The nature of the problem might give you some guidance, but it's really your call.

[kamarinette]

hi

i have the same problem. i'm using eviews 7 and its doesn't recognise scalar and vector function!!!

how can i do?

[zain k]

Thank you for your help in determining marginal effects, it worked with me :D . I just have two questions: does it still work if I have dummy variables in my set of regressors and not just continuous variables? Also when I come to interpret the value for continuous variables, the marginal effect computed shows the effect at the mean value of the variable, while for the dummy variable it shows the effect of change in the dummy variable from 0 to 1, right?

Thanks alot :)

[EViews Glenn]

For dummy variables, you'd have to fix all of the other variables at their means, and then compute the finite difference between the value with the dummy at 0 and and 1.

[zain k]

Ok, so to analyze dummy variable of interest I need to multiply each marginal effect with its mean for continuous variables and for all other dummies which I am not interested in analyzing I will make them zero, right?. And when I come to interpret the marginal effects for continuous variables, do I also need to multiply them by their mean value in order to analyze them, or did it already do that in the calculation?

Thanks a lot for your help :D

[EViews Glenn]

For the marginal effect of a dummy. You should use means for the X's for all of the non-relevant dummy variables and evaluate the XB corresponding to those values.

Set the dummy variables so that they correspond to your first configuration (say "male=1", or "red=0 and blue=1"). Add in the dummy values times the corresponding coefficients to the XB computed above to obtain the full index, and evaluate the probability. Set the dummy variables so that they correspond to a different configuration (say "male=0", or "red=0 and blue=0" or "red=1 and blue=0"), multiply by the dummy coefficients and add to the original XB computed in the first step to compute the full index, and evaluate the probability. The difference is the change in probabilities as you move configurations.

Note that the above might not be quite what you want if you have more than one set of dummy variables as evaluating at means for one set while of dummies while you compute differences for another is something you might not want to do. You can always, for moderate numbers of dummies, treat the entire set of dummies as a single configuration.

[zain k]

Ok so to check if I understood right, I will multiply the dummy value times its probit coefficient times its marginal effect in the computed xb vector for all the dummies (given I fix the nonrelevant dummies to a certain value, 1 or 0) and then I change the dummy of interest from 0 to 1 to check for the effect, right?. Last question: I also have continuous variables, so do I also need to multiply the mean value of each one times their probit coefficient times their value in the xb vector computed to get the marginal effect of each at the average level?

Thank you so much for your help and sorry for this confusion, :)

[EViews Glenn]

Sounds about right. Just make sure you think about what evaluating things at means or 0 or 1 or whatever means conceptually.

[kemp40]

Hi! I have been trying to emulate the marginsplot command from stata in eviews 8. Failed.
I am trying to plot the marginal effect around confidence intervals but I just cannot included these in a single graph.
Is it possible to do it?
I'd appreciate your help or opinion.
Many thanks!