No convergence with Xtheckman (Can't see where I have gone wrong)

[Ruari McColl]

Hello.

I am currently writing my dissertation on the political influences on the distribution of FEMA natural disaster expenditure within states in the US. I have data from 2001 to 2020 on all counties on the following states: New York, Florida, Louisiana and Texas.

The process of counties getting FEMA expenditure is as follows:
1. The county receives a disaster declaration from the President
2. If the county receives a disaster declaration, it can then apply for Public Assistance (i.e. relief) expenditure
3. If the county receives Public Assistance expenditure it can apply for Hazard Mitigation Grant Programme expenditure which is capped at 15% of Public Assistance (and Individual Assistance) expenditure

I have ran a Heckman two stage regression for stages 1 and 2, and stages 1 and 3 but I have not managed to get convergence. Even when I only include a small number of variables and have a theoretically valid variable which affects declaration but not expenditure (property damage in this case) I don't get convergence. What am I doing wrong?

[FernandoRios]

There is probably nothing on your side.
xtheckman simply requires very specific type of data to converge. Perhaps you can give a try to xtheckmanfe (ssc). See the references in the help file to understand what the command does.
F

[Ruari McColl]

FernandoRios Hello thank you for the help. However, I have carried out a Hausman test on the xtreg of the expenditure and have found that I should be doing random effects and not fixed effects. Therefore, I cannot use the xtheckmanfe?

[FernandoRios]

Fixed effects are not invalid, simply less efficient.
and xtheckmanfe is not really a fixed effect model, but a correlated random effects model.
See references in helpfile for details

[Ruari McColl]

FernandoRios Thanks so much for the replies, it has helped A LOT! I have three final questions:
1. When I run the xtheckman results I don't get convergence but I do get a table with some values. However, these values are very different to the xtheckmanfe results. Why is this?
2. What do the _mn_... coefficients mean?
3. I have looked across the internet but am a first time user of Statalist so am unsure what you mean by the help file and the references in this. Could you direct me to this?

[Ruari McColl]

You can see the xtheckmanfe results at the top and the xtheckman results at the bottom:

[FernandoRios]

Type help xtheckmanfe
scroll all the way down to references. There are two papers by Prof Wooldridge cited that the command was based on. Read the. And understand them, then search for references to those papers
that will help you getting applied work where the terms are Interpreted

[Luis Ortiz]

Dear Fernando,

I am becoming desperate using 'xtheckman' with household panel data (Understanding Society). Convergence takes ages, and sometimes it does not converge at all, without giving an error message. Often, the iteration process gets stalled when there's a change in the optimisation technique. I am not being to find what is wrong (it may be something wrong with my data). For instance, I have tried to get rid of all the variables and make the model quite simple... but to no avail.

Could convergence be faster achieved with 'xtheckmanfe' than with 'xtheckman'?

Thanks

Luis Ortiz

[FernandoRios]

Hi Luis
so, Xtheckman is a MLE model, so it could be very very sensitive to initial conditions and overall stability of the model.
xtheckmanfe is kind of a CRE model, and for most practical purposes its like a Fixed Effect Heckman selection model.
It will give you results much faster, unless you use bootstrap, or the MLE option I added. But that MLE is different from xtheckman.
See Wooldridge paper for references.
F