I have asked this question on Cross Validated, but think I might not get help as this is more of a programming question rather then theory/interpretation of the statistics.
I am trying to use the mlogit package in R and have been following the vignette trying to figure out how to get the marginal effects for my data. The example provided uses continuous variables, but I am wondering how to do this with categorical explanatory variables.
I have a value of risk which is continuous as a covariate, but I also have age, class, and gender as covariates. I want to see the marginal effects of "females" only or of "Young - females" in regard to risk. How would I do this?
The help documents say:
z <- with(Fish, data.frame(price = tapply(price, index(m)$alt, mean),
catch = tapply(catch, index(m)$alt, mean),
income = mean(income)))
# compute the marginal effects (the second one is an elasticity
effects(m, covariate = "income", data = z)
effects(m, covariate = "price", type = "rr", data = z)
effects(m, covariate = "catch", type = "ar", data = z)
I'm not sure how to manipulate the z data frame to get the mean risk for females or young females to then be able to calculate the marginal effects. Would I do them all separately? Do I somehow divide the data frame by age class (say I have just 2 age classes: young and old) so that I have 1 data frame for the young, and a separate new data frame for old, then calculate mean risk?
What I am hoping to get from my own data is to be able to interpret the magnitude of the likelihood in producing my categories of offspring. As an example, what I want to say is that if there is a 1 unit increase in risk, it is 10% more likely for older females to produce 2 offspring. As there is a 1 unit increase in risk, younger females are 15% more likely to produce 2 offspring.
I am not sure how to calculate the marginal effects by hand, and therefore am confused as to how to get a package to do it for me. Ive also been trying in the nnet library or the VGAM, but neither of these seem to give a great deal of help either.
I sort of got an answer - not sure if its the best, but it worked. My covariate that I was interested in just so happened to be 2 classes - which means I could turn the covariate into a binary 0,1 numeric response. Therefore, when I rerun the code, I could then calculate the mean for this "categorical" variable.
However, I think that I am missing the point with this marginal effect and am likely using it or trying to interpret it incorrectly.