Marginal effects for sample selection model

[reddylee]

Hello,
I'm encountering challenges in calculating marginal effects for a sample selection model using the sampleSelection package. Currently, I can only obtain marginal effects for the selection equation. I'm wondering if it's possible to calculate marginal effects for the outcome equation as well. Many thanks.

[vincentarelbundock]

Have you tried the type argument? If so, please supply a minimal replicable example and explain exactly why the type argument fails.

Here are some tips on asking good questions on software help forums: https://marginaleffects.com/vignettes/help.html#getting-help

[reddylee]

Thanks for the quick reply. I tried type argument:

type = "link" get the coefficients for selection equation
type = "unconditional" get the coefficients for outcome equation
While type = "response" (which is the default) only get one outcome of marginal effects, which I have no idea it is for selection or for outcome.

library(sampleSelection)
library(marginaleffects)
data(Mroz87)

m <- selection(lfp ~ educ + age + kids5 + kids618 + nwifeinc,
               wage >= 5 ~ educ + exper, data = Mroz87)

avg_slopes(m) |> 
  view()

[vincentarelbundock]

Thanks for the example.

I looked into this and the issue is that you need to specify both the type argument and the part argument. This is documented in ?predict.sampleSelection

avg_slopes(m, type = "unconditional", part = "outcome")
avg_slopes(m, type = "response", part = "selection")
avg_slopes(m, type = "link", part = "selection")

[reddylee]

Thanks, Vincent. Your code can not get the marginal effects of the outcome. Please see the comments on the code below. ?predict.sampleSelection does not work, it reads "No documentation for ‘predict.sampleSelection’ in specified packages and libraries:". I did not find predict.sampleSelection in the codebook (https://cran.r-project.org/web/packages/marginaleffects/marginaleffects.pdf, June 14 2024). Do I miss something?

avg_slopes(m, type = "unconditional", part = "outcome")#coefficients of outcome

avg_slopes(m, type = "response", part = "selection")# marginal effects of selection

avg_slopes(m, type = "link", part = "selection")#coefficients of selection

[vincentarelbundock]

oops, it's ?sampleSelection::predict.selection and in the sampleSelection package codebook.

[vincentarelbundock]

What do you mean by "coefficient of outcome"? The outcome component is linear, so the slopes correspond to the coefficients.

[reddylee]

Thanks, Vincent. Yes, for normal selection model, the outcome equation is linear. However, I am using two binary model, which both selection and outcome equation are probit model. Therefore, the outcome model is not linear. I guess that's the reason we get the same. Wonder if it can deal with this double probit model.

[vincentarelbundock]

This is not a marginaleffects question. If the functionality is supported through the type argument in the sampleSelection package, then it should be supported in marginaleffects as well.