I'm looking for the most user-friendly way of calculating predicted probabilities with confidence intervals from a logistic regression model using the marginaleffects package in R.
Let's consider a toy model:
mod <- glm(am ~ factor(cyl), data = mtcars, family = binomial)
I could do
invlogit <- \(x) exp(x)/(1+exp(x))
hypotheses(mod, hypothesis = c("invlogit(`(Intercept)`) = 0",
"invlogit(`(Intercept)` + `factor(cyl)6`) = 0",
"invlogit(`(Intercept)` + `factor(cyl)8`) = 0"))
to get the predicted probabilities in the three groups given by the cyl-variable with corresponding confidence intervals (yeah, I know they're bad here because of the low sample, but let's forget about that).
Is there some other way of doing this where I don't have to make the invlogit-function and perhaps where the specification of the three groups is done a bit easier. With Stata's margins function you can simply write margins cyl or margins var1#var2 to get the predicted probabilities for each combination of two categorical variables.
Aren't you just looking for marginaleffects::marginal_means?
mod <- glm(am ~ factor(cyl), data = mtcars, family = binomial)
marginaleffects::marginal_means(mod)
#> Term Value Mean Pr(>|z|) 2.5 % 97.5 %
#> cyl 4 0.727 0.147 0.414 0.910
#> cyl 6 0.429 0.706 0.144 0.770
#> cyl 8 0.143 0.019 0.036 0.427
#>
#> Results averaged over levels of: cyl
#> Columns: rowid, term, value, cyl, estimate, p.value, conf.low, conf.high, am, wts
Note that in any case you don't need to define invlogit; it does the same thing as base R's plogis