Dear all,
I have a regression with binary outcomes and two endogenous continuous regressors.
When I use the maximum likelihood (ML) method to run the regression, it always has a problem to converge. I have tried all the 4 algorithms, only one of them worked when I don't include more than 10 variables into the model (but my model regression need at least 30 variables). I searched for this problem online and find many other people having the same experience. I suspect that it may be a shortcoming of the algorithm of ML of Stata.
So my first question is that is there anything else I can do to get my estimation converged?
Since I failed to get my estimation converged with ML, I turned to Newey's two-step approach. It quickly generates an outcome. I compared the coefficients of the outcome with the ones of the ML approach (although the estimation did not converge, it stopped somewhere and reported the coefficient at that point). The two outcomes are very similar to each other.
My second question is that, in this case, is the outcome generated by Newey's two-step approach trustworthy?
Another question is that if my endogenous variable is not exactly continuous but discrete, what parametric model should I use? I know it is not proper to use IVprobit in this case since it assumes the endogenous regressors to be continuous and only run OLS in the first stage. But it seems it is also not proper to use ordered probit in the first stage. I remember that Angrist proposed to use 2SLS when having a binary dependent variable and binary endogenous variable in his textbook. I am not sure if I should use 2SLS when having a binary dependent variable and discrete endogenous variable.
Thank you for all your response.
I have a regression with binary outcomes and two endogenous continuous regressors.
When I use the maximum likelihood (ML) method to run the regression, it always has a problem to converge. I have tried all the 4 algorithms, only one of them worked when I don't include more than 10 variables into the model (but my model regression need at least 30 variables). I searched for this problem online and find many other people having the same experience. I suspect that it may be a shortcoming of the algorithm of ML of Stata.
So my first question is that is there anything else I can do to get my estimation converged?
Since I failed to get my estimation converged with ML, I turned to Newey's two-step approach. It quickly generates an outcome. I compared the coefficients of the outcome with the ones of the ML approach (although the estimation did not converge, it stopped somewhere and reported the coefficient at that point). The two outcomes are very similar to each other.
My second question is that, in this case, is the outcome generated by Newey's two-step approach trustworthy?
Another question is that if my endogenous variable is not exactly continuous but discrete, what parametric model should I use? I know it is not proper to use IVprobit in this case since it assumes the endogenous regressors to be continuous and only run OLS in the first stage. But it seems it is also not proper to use ordered probit in the first stage. I remember that Angrist proposed to use 2SLS when having a binary dependent variable and binary endogenous variable in his textbook. I am not sure if I should use 2SLS when having a binary dependent variable and discrete endogenous variable.
Thank you for all your response.
Comment