Question on interpreting probit effect

[Lindsey Buck]

Hi all,

I have a quick question on probit effects.
When I run my probit y xvar othervars
I get my xvar has a coefficient of -.36 with a p value of 0.004. I wanted to transform this into a probability so I used the command:

margins, dydx(*) atmeans post

and it showed that my xvar had a dy/dx of -.021 with a p-value of 0.946.
Why would it change to something that is not significant if it was significant in the original probit? Can you help explain? My understanding was that this means that a one unit increase in my xvar decreases my yvar by 2.1% but I may be wrong here. Confused on the signficance thing as well.
Thanks!
L

[Clyde Schechter]

So, a few different things are being confused here.

The command -margins, dydx(*) at means post- doesn't show you probabilities. It shows marginal effects, which are differences in probabilities. Now, -probit- being a non-linear model, there is no unique marginal effect on probabilities for any variable. Rather there are infinitely many different marginal effects on probabilities that depend on the particular values of all of the variables in question. You have stipulated in your command that you want the marginal effect conditional on all of the predictor variables being at their means. As it turns out, in that configuration of predictors, the marginal effect of xvar is small and turns out not to be statistically significant. The coefficient of xvar in a probit model is really rather difficult to interpret in concrete terms. It doesn't represent anything that we have another name, or simple description for. In any case, it is a unique value for each predictor variable, and it reflects the slope of the xb (to which the probit transform is then applied to get a probability) on xvar--the relationship of this to the marginal effect at any particular set of values of xvar and the other vars is quite indirect and there is no reason to expect them to be similar.

[Lindsey Buck]

Thank you so much! Is there any way that I can get my x coefficient (-.36) in terms of a percentage? I am not sure!

[Clyde Schechter]

I'm not sure what you have in mind, but I think the answer is no.