I've run a probit regression in R with a random effect and can find no way to get the marginal effects with s.e. and p values. I have therefore tried to calculate the marginal effects 'by hand' by using the probit scalars and regression coefficients. However, I do not know how to get p values or standard errors and as far as I have found there is no easy way to do this for a mixed effects probit regression.

My model m1 is

  • $\begingroup$ What do you mean by marginal effects? The estimated probits for each Name.Origin value? Or something else? (In fact, is Name.Origin a factor or a covariate?) $\endgroup$ – Russ Lenth Aug 19 '15 at 0:03
  • $\begingroup$ @rvl Name.Origin is a factor so is basically a dummy variable. No, I calculated marginal effects by multiplying ProbitScalar by the regression coefficients as I do not know (and cannot find) any other way to calculate marginal effects for a glmer which also show standard errors and p values. I'm not sure it is even possible unless you write your own code/function from scratch which I have no idea how to do....do you know of any way? $\endgroup$ – Lola2000 Aug 19 '15 at 0:26

Well, if I understand it correctly, you just want the prediction from the model at each level of Name.Origin. The predict function for merMod objects should in principle work, but I don't see an argument in it to obtain SEs.

However, the lsmeans package works nicely:

lsmeans(m1, "Name.Origin")           # predictions, SEs, CIs
pairs(lsmeans(m1, "Name.Origin"))    # pairwise comparisons of above
lsmeans(m1, "Name.Origin", type = "response")  # transform to probabilities
| cite | improve this answer | |
  • $\begingroup$ thank you, the answers do report S.E ! Would this be interpreted as the chance each level has of getting a job (which is what I am looking into)? (irrespective of the 'reference' group). Do you know of anyway of exploring the probabilities between groups? e.g. how much more likely one level of my factor would be in getting a job than the other? $\endgroup$ – Lola2000 Aug 19 '15 at 10:00
  • $\begingroup$ The one with type="response" gives the estimated probabilities. The pairs one does tests comparing the different factor levels. It appears you are flying somewhat blind here, and I suggest you find a stat consultant (say at a stat dept at a nearby university) and get them to sit down with you and explain this stuff more carefully than can be done here. $\endgroup$ – Russ Lenth Aug 19 '15 at 21:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.