Using fixed effects with the CMP command and question about output

[Pamela Kuperstein]

Hi everyone,

I'm attempted to use Roodman's CMP command to run SUR with a probit and OLS regression. I've run into two problems and haven't found an answer I quite understand. First, I added in year fixed effects to both regressions by putting in i.year. Whenever I do this, the second stage fails to converge. Does this mean that the CMP command cannot have fixed effects?

Since fixed effects didn't work, I took them out and was able to get results. However, in terms of the output, in the second stage mixed process regression I only get coefficient estimates, and do not see any standard errors or probability statistics. How do I get the remaining information?

This is my first time posting, so please let me know if I need to provide additional information.

[Alfonso Sánchez-Peñalver]

Hi Pamela, without further information about your estimations, i.e. code and results, it's very hard to tell you exactly what's wrong. I'll address a couple of things you mention here.

Yes, cmp can take fixed year effects, no problem.

That you don't see standard errors and p-values is a sign that something may be wrong with your specification.

This is where we can't venture a guess as to what exactly is wrong because we lack further information.

[Pamela Kuperstein]

Hopefully this will be enough additional information. Please let me know if there is additional information you would need to help me. Your help already is much appreciated!

The code I'm trying to run is as follows:

cmp (rpt= post totalcomp iroa itrs sales bm rd missing insiderown totaldir dual founder tenure numberdir percentinsidedir percentlinkeddir percentbusydir percentolddir percentappoint b l i.gvkey i.year) (totalcomp = rpt ceopost post iroa sroa itrs strs sales bm insiderown totaldir dual founder tenure numberdir percentinsidedir percentlinkeddir percentbusydir percentolddir percentappoint i.gvkey i.year), ind($cmp_probit $cmp_cont)

where rpt is a binary variable and totalcomp is a continuous variable.

When I run this code, the program will run for a long time with many iterations. At some point, the log likelihood stops changing, but it still does not converge.

When I take out the fixed effects (i..e i.gvkey and i.year) I will eventually get results, but they show up like this (I'm only showing the second stage results for the first regression, the second stage results for the second regression looks the same such in that it has coefficient estimates but no other information):

I also get the following error after each of the first stages (the variable "rpt" in the warning copied below is "totalcomp" for the first stage of the second regression):

Warning: regressor matrix for rpt equation appears ill-conditioned. (Condition number = 1754.2837.)
This might prevent convergence. If it does, and if you have not done so already, you may need to remove nearly
collinear regressors to achieve convergence. Or you may need to add a nrtolerance(#) or nonrtolerance option to the command line.

I read through the help information, but I don't really understand what the options ntolerance or nonrtolerance would do or how to determine what number to use.

Thank you!

[Alfonso Sánchez-Peñalver]

Hi Pamela,

I see that you're including rpt as an explanatory variable for totalcomp and totalcomp as an explanatory variable for rpt. This seems more a system of equations than a seemingly unrelated regression. My first thought is that cmp is not appropriate for your estimation. You may want to contact its author, David Roodman, and ask him if it is.

[David Roodman]

Alfonso is right that cmp was not designed for simultaneous models. This is the cause of all the problems. However, features added some years ago might suffice here. Try making one change. Add a "#" suffix to the reference to totalcomp:

cmp (rpt= post totalcomp# iroa itrs sales bm rd missing insiderown totaldir dual founder tenure numberdir percentinsidedir percentlinkeddir percentbusydir percentolddir percentappoint b l i.gvkey i.year) (totalcomp = rpt ceopost post iroa sroa itrs strs sales bm insiderown totaldir dual founder tenure numberdir percentinsidedir percentlinkeddir percentbusydir percentolddir percentappoint i.gvkey i.year), ind($cmp_probit $cmp_cont)

The help file discusses the use of this suffix. Leaving aside #-references, the model must be triangular/recursive. But there can be simultaneity involving the # references.

[lao wu]

David Roodman , sir how to run fixed effect (fe) in cmp regression command for instrumental variable order probit model (IVOPROBIT).?

[David Roodman]

Use the "i." notation to generate fixed effect dummies and include them in your model.

[lao wu]

David Roodman , sir can you please recommend me any book to understand cmp modules for better understanding?

[David Roodman]

I wrote a paper about it. http://www.stata-journal.com/article...article=st0224. Also, type "help cmp".

[lao wu]

David Roodman , Sir i use instrumental variable order probit model (ivoprobit) for my regression analyses with the help of cmp command. Now i want to test my instrumental variable (valid/strong/weak). How can i do it after cmp command?

Thanks

[lao wu]

David Roodman , sir how to run different diagnostic test (i-e heteroskedasticity, Sargen-hensen over identified instrument variables, endogenity test, VIF, Normality test etc.) after using ivoprobit (cmp) command??

[David Roodman]

Those are not really applicable because they are designed for longrest models. If you are concerned that an instrument is invalid and your regression is overidentified, you can add it to the second stage and see if it is significant.

[lao wu]

David Roodman , sir in ivoprobit model, is it not necessary to run heteroskedasticity, Sargen-hensen over identified instrument variables, endogenity test, VIF, Normality test etc.???

[lao wu]

David Roodman , Sir can i use cmp command on pool cross sectional data after controlling year effect (i.year)??

[Jeff Wooldridge]

The responses here so far focus on the computational issues. But there are even more important statistical issues. While I can't tell for sure, you seem to have lots of firms that you are including fixed effects for. Except in very special cases -- the linear model is the leading case -- including fixed effects causes an incidental parameters problem. All estimators are inconsistent and likely pretty biased. You might try a correlated random effects approach instead.