Help with cdeco and Heckman

[Alina Faruk]

Dear All,

I am new to the forum, so please pardon any mistake on my part as this is my first post.

I am trying to do quantile decompostion with the cdeco command, and would like to take selection effect into account.

However, unlike oaxaca, where I could just use the option ",adjust(mills)", I cannot do so here.

What can be the solution here? Any help is very much appreciated.

[Phil Bromiley]

I can't find the cdeco command - following the FAQ on asking questions, it would help if you let us know where you found this software.
This appears to be a user-written command. I assume you've looked hard to make sure a built-in Stata command won't do what you want. You may need to contact the authors.

[Alina Faruk]

I can't find the cdeco command - following the FAQ on asking questions, it would help if you let us know where you found this software.
This appears to be a user-written command. I assume you've looked hard to make sure a built-in Stata command won't do what you want. You may need to contact the authors.

Thanks for replying Mr Bromiley.

This is the official page for the command, and yes, it is a user-written one:

https://sites.google.com/site/blaise...-distributions

[Felipe Salce]

Dear Alina,

I currently have the same question, did you find a solution?

[Sven-Kristjan Bormann]

I can only offer how I have attempted in the past to deal with sample selection and quantile regression. First run some kind semi-parametric binary-choice model (for example with the user provided -sml- package) similar to the probit estimation for the Heckman sample selection model. Then calculate the inverse mills ratio and add it to the quantile regression as an additional variable.
The same approach might work as well for the -cdeco- command, but I have not tried it yet.
I got my inspiration for this estimation approach from this article and some older posts to the predecessor of this forum. But I don't remember anymore where exactly I got the idea from.

But I have not seen yet any other implementations of sample selection correction for quantile regression.

[Alina Faruk]

Dear Alina,

I currently have the same question, did you find a solution?

Adding to what Mr Bormann said, you might find the following thread helpful:

https://www.stata.com/statalist/arch.../msg01309.html

[farah aslam]

I can only offer how I have attempted in the past to deal with sample selection and quantile regression. First run some kind semi-parametric binary-choice model (for example with the user provided -sml- package) similar to the probit estimation for the Heckman sample selection model. Then calculate the inverse mills ratio and add it to the quantile regression as an additional variable.
The same approach might work as well for the -cdeco- command, but I have not tried it yet.
I got my inspiration for this estimation approach from this article and some older posts to the predecessor of this forum. But I don't remember anymore where exactly I got the idea from.

But I have not seen yet any other implementations of sample selection correction for quantile regression.

Can you please guide me about the syntax and commands that are used to get the sml estimator. The way I am trying is taking too long ,still processing

[Sven-Kristjan Bormann]

You could show me the way that you used until now. Then I can try to spot the difficulties. Below is some example code for quantile regression with sample selection which might work for you:

ssc install st0226 // get the package containing the snp-command if not previously installed
snp <dummy variable> <selection variables>, robust
predict index_variable if e(sample), xb
gen inverse_mills_ratio = normalden(index_variable)/normal(index_variable)
qreg <dependent_variable> <independent_variables> <inverse_mills_ratio>

[Alina Faruk]

You could show me the way that you used until now. Then I can try to spot the difficulties. Below is some example code for quantile regression with sample selection which might work for you:

ssc install st0226 // get the package containing the snp-command if not previously installed
snp <dummy variable> <selection variables>, robust
predict index_variable if e(sample), xb
gen inverse_mills_ratio = normalden(index_variable)/normal(index_variable)
qreg <dependent_variable> <independent_variables> <inverse_mills_ratio>

Hi,

Can you please tell me why one should choose -snp- over -sml-? All the papers using Buchinsky decomposition I am coming across seems to have used Klein and Spady, which is implemented by -sml-. I read the Stata Journal article regarding the -snp- command, but it does not answer this specific context. Moreover, my sml estimation is taking forever whereas snp gave output immediately.

Appreciate your help. Thank you.

[Sven-Kristjan Bormann]

If I remember correctly, then I have chosen -snp- over -sml- for the same reason as you did. -sml- did not converge or ran forever whereas -snp- converged after some time. When I estimated these models I thought that my dataset and the chosen set of variables were not suitable for -sml-. That's why I tried -snp- instead. I would need to re-run my old attempts of these models to see what exactly the problem was.

[Kadir Ruslan]

I am doing the decomposition of wages difference using cdeco. How do I get the results of decomposition for all independent variables not only the total difference, explained and explained? I want also to identify the contributing factor along the distribution.