Pooled Mean Group (PMG) "Initial values not feasible"

[Jose Tuano]

Hi! I am currently attempting to run a PMG estimator (Pesaran, Shin, & Smith, 1999) with a balanced panel data set of 83 countries over 52 years. Although it is balanced in terms of having the same number of years for each panel, there is a lot of missing data for different variables since this dataset was a combination of different sources (I attached the excel file if you need to take a quick look at the data structure). I am attempting to estimate the short/long-run relationship between a country's economic complexity (eci) and trade flows (im / ex), whilst controlling for other variables such as GDP per capita, total population, etc.

The main issue is that after trying to include more than 3 variables in the command, I obtain the error message "Initial values not feasible r(1400)". For example:

Entering the following seems to work fine!

(1) xtpmg d.eci d.ln_ex d.ln_im, lr(l.eci ln_ex ln_im) ec(ec) pmg

or (2) xtpmg d.eci d.ln_im d.ln_gdppc, lr(l.eci ln_im ln_gdppc) ec(ec) pmg

or (3) xtpmg d.eci d.ln_tpop d.ln_gdppc, lr(l.eci ln_tpop ln_gdppc) ec(ec) pmg

So I do not believe that it has to do with the lack of observations for some variables.

However, attempting to run a command along the lines of:

(4) xtpmg d.eci d.ln_ex d.ln_im d.ln_gdppc, lr(l.eci ln_ex ln_im ln_gdppc) ec(ec) pmg

Always results in the error message. Alternatively, I attempted to calculate Mean Group (MG) estimators or Dynamic Fixed Effects (DFE) estimators using the options available with the xtpmg command, and they seem to work, albeit the MG estimator is limited by matsize when I try to add too many regressors (8 variables).

I suppose my question is whether anybody has an idea about what may be causing the problem and if you have any ideas on how I can work around it. I am unsure if the issue is with a large amount of missing observations in my dataset, or whether there is some limitation faced by the xtpmg programme that I am unaware of.

Best,
Jose Tuano

[Jesse Wursten]

User-written commands are always hard to debug. One start could be to do the following

set trace on
set tracedepth 2
xtpmg d.eci d.ln_ex d.ln_im d.ln_gdppc, lr(l.eci ln_ex ln_im ln_gdppc) ec(ec) pmg

You might need to increase the tracedepth. It will give a bit more feedback on what's going on behind the scenes which might help in pinpointing where the problem arises.

[Jose Tuano]

Hi Jesse,

Thanks for the help. Here is what is returned after using the trace command. Since I'm quite new to Stata, I'm unsure on how to interpret most of the text. I included only the text that was returned after "initial values not feasible" since the entire text is quite lengthy and I am unsure if all of it is relevant.

initial values not feasible
    matrix `theta'=e(b)
    matrix `thV'=e(V)
    quie predict double `xb' if `touse'
    quie gen double `ec'=$LRy-`xb' if `touse'
    matrix `thV'=J(1,colsof(e(b)),.) \ `thV'
    matrix `thV'=J(colsof(e(b))+1,1,.) , `thV'
    }

Perhaps you can provide me some more insight on what this means? And of course, let me know if I should include more or less of the information provided by the trace command.

[Jesse Wursten]

I'll be honest, I don't really know how to fix this. Have you tried using fe/mg instead of pmg? Alternatively, you might want to contact the authors.

[Jose Tuano]

The dfe and mg options did appear to work, however, there were additional problems with the use of mg when adding more regressors (even after increasing matsize to 11,000). I suspect that the issue is similar to the problem I faced with the pmg option, and I believe it may be a compatibility issue between my data and the command. But I can't be sure...

I've since decided to move back to the Arellano-Bond estimator (which estimates just fine), so thanks for trying anyway!