Marginal Effects after Multinomail Logit model estimated using Control Function Approach

[Anna Maccagnan]

Dear Statalist,


I am estimating a multinomial logit model using Control Function Approach, as one of my predictors is endogenous. I follow the procedure described by Woolridge 2015 ("Control Function Methods" published in Applied Econometrics).

reg xend z X  // first stage - where xend is my endogenous predictor, z my IV and X a vector of exogenous predictors
predict xend_res, residuals
mlogit y xend xend_res  X  // 2nd stage (CFA)

I am however unsure about how to compute the marginal effects of my predictors.

I have been using the command margins to compute marginal effects for all exogenous predictors:

margins,  dydx(*) post

While for the endogenous predictor I follow the procedure described in the STATA manual on ivprobit postestimation:

margins, at(xend = generate(xend)) at(xend = generate(x_end*1.10)) contrast(at(r) nowald) predict(pr outcome(1)) // and so on for the other possible outcomes of my dependent variable

Does this procedure to estimate marginal effects (both for the exogenous and endogenous variables) seem correct to you? Very grateful for any advice.

Thanks

Anna