xtprobit, marginal effects the same as coefficients?

[Alex Jeffries]

Hi guys,

Running an xtprobit, and then computing the marginal effects, and marginal effects at means. Both return just the coefficients from the probit however, can someone shed any light on where I could be going wrong here?

Regards

[Jeff Wooldridge]

Yes, this is unfortunate, and something Stata should fix in the future. The user-written command mfx does it "almost" correctly, in that it inserts the mean value of heterogeneity (zero) before average across the covariates. Or, you can use pooled probit -- which is more robust because it assumes nothing about serial correlation -- and use the margins command after "probit."

If you want to use xtprobit and get the proper average partial effects, you can write the formula yourself. Attached are some slights that accompany my MIT Press book that show how. But to get a standard error you probably have to bootstrap, which is time consuming. Search for xtprobit in the PDF. I have a static and dynamic example. The former shows how to do the calculation for the continuous case, the latter for the discrete case.

JW

[Richard Williams]

Also see -help xtprobit_postestimation- , especially the sections on predict and margins (hopefully you have Stata 14 as the documentation is clearer than it used to be). The default predict option is -xb- and you may want to use something else, e.g. pu0.

[Alex Jeffries]

Yes, this is unfortunate, and something Stata should fix in the future. The user-written command mfx does it "almost" correctly, in that it inserts the mean value of heterogeneity (zero) before average across the covariates. Or, you can use pooled probit -- which is more robust because it assumes nothing about serial correlation -- and use the margins command after "probit."

If you want to use xtprobit and get the proper average partial effects, you can write the formula yourself. Attached are some slights that accompany my MIT Press book that show how. But to get a standard error you probably have to bootstrap, which is time consuming. Search for xtprobit in the PDF. I have a static and dynamic example. The former shows how to do the calculation for the continuous case, the latter for the discrete case.

JW

Thanks very much!

[Jeff Wooldridge]

I know this is an old thread, but I should've thanked Rich, and at least partly retracted my previous statement. The command

margins, dydx(x1) predict(pu0)

"almost" gets the average marginal effects. The difference is that it plugs in zero for the heterogeneity (ui in Stata notation) and then averages across the observed covariates. What I define as the Average Partial (or Marginal) Effect averages across the joint distribution of (ui,xi). I believe this is preferred (perhaps only slightly in practice), but the "pu0" option has the benefit of providing valid standard errors without bootstrapping (which can be very time consuming). Jeff