What scale does STATA use to compute the average marginal effects (AME) for metobit? I have a model with two binary predictors and an interaction term:
metobit y x1##x2 || r1: || r2:, ll(0) nolog
I then get the average marginal effect for x2:
margins, dydx(x2)
This gives me .1448938, slightly less than the coefficient for x2, which is 0.17. I thought that the AME was the difference in the average predicted probability between each level of x2. To check, I tried:
predict xb
gen prob = normprob(xb)
mean prob xb, over(x2)
Neither difference in xb or prob from x2 = 0 to x2 = 1 results in .144. I also tried incorporating the random effects.
predict eta
gen prob_eta = normprob(eta)
mean prob_eta eta, over(x2)
I’ve also tried converting xb to predicted probabilities by hand using exp(xb)/1+exp(xb), but the probabilities were basically the same at x2 = 0 and x2 =1. The AME seems to approximate the coefficient, but is slightly less. How would I get 0.144 by hand so I understand what is going on?
metobit y x1##x2 || r1: || r2:, ll(0) nolog
I then get the average marginal effect for x2:
margins, dydx(x2)
This gives me .1448938, slightly less than the coefficient for x2, which is 0.17. I thought that the AME was the difference in the average predicted probability between each level of x2. To check, I tried:
predict xb
gen prob = normprob(xb)
mean prob xb, over(x2)
Neither difference in xb or prob from x2 = 0 to x2 = 1 results in .144. I also tried incorporating the random effects.
predict eta
gen prob_eta = normprob(eta)
mean prob_eta eta, over(x2)
I’ve also tried converting xb to predicted probabilities by hand using exp(xb)/1+exp(xb), but the probabilities were basically the same at x2 = 0 and x2 =1. The AME seems to approximate the coefficient, but is slightly less. How would I get 0.144 by hand so I understand what is going on?
