Problem with sureg: last estimates not found

[Giovanni Angioni]

Dear All,

I have a panel with the following structure: N=6 T=53 (monthly data). I am trying to estimate the impact of opinion polls over the frequency of tweets about inequality by 6 parties. Given that T>N, I discarded FE model in favour of Zellner's seemingly unrelated regressions (sureg). Hence, I reshaped my panel wide and now it looks like this:

* Example generated by -dataex-. To install: ssc install dataex
clear
input float w_freq_ineq1 double share1 float w_freq_ineq2 double share2 float w_freq_ineq3 double share3 float w_freq_ineq4 double share4 float w_freq_ineq5 double share5 float w_freq_ineq6 double share6
        0                  .  6.289308                  .         0                  .         0                  .         0                  .         0                  .
        0               21.6       3.2                1.9         0                4.1         0               25.2  3.243243               25.4         0                 .9
        0               22.6 2.3255813                1.5         0                3.9 13.043478               26.7  10.81081                 26         0                 .7
        0               26.2  2.631579                1.7         0                4.1       2.5               24.8  47.61905               25.2         0                 .8
 6.896552               27.7 2.2727273                  2         0                4.1  3.076923               22.7 1.4492754               24.2         0                 .9
        0               27.5  8.163265                2.2       2.5                3.7 1.6949153               18.3        25               27.9         .                  .
        0               26.9 3.5714285                2.7  .9174312                  4 10.227273               18.3  2.857143               26.9         .                  .
        0                  .         0                  .         0                  .     9.375                  . 33.333332                  .         .                  .
 2.857143               26.2         0                2.4         0                4.1  7.042253               20.1 14.285714               28.1         0                 .9
 3.658537                 24 3.7037036                2.6 1.0204082                4.3  52.63158               21.6 3.3613446               29.2         0                 .9
 5.128205                 22         0                2.8        12                  4  22.61905                 22      6.25               28.7        20                 .9
     12.5               21.3  3.846154                2.5  1.234568                3.9  16.86747               21.8  6.779661               30.1         0                 .8
 5.084746                 22  1.724138                2.4  8.333333                3.9  4.878049               21.4 4.5454545               31.5         0                 .7
     12.5               22.4         0                2.5         0                3.9  8.450705               22.7         0               30.4         0                 .9
    31.25               22.3         0                2.8 4.5454545                4.1  10.67961               21.8  2.631579               31.1         .                  .
end

However, after setting a macro and trying sureg:

 
 tsset time  forvalues i = 1/6 {      local vars  "`vars'  (w_freq_ineq`i' share`i')"       }     
 sureg `vars'

I get the following error: last estimates not found r(301). Many thanks for your help! Giovanni

[Carlo Lazzaro]

Giovanni:
what follows is probably not the reply you were asking for, but I do hope it might be helpful: if you have a large T, small N panel dataset, why not considering something like -xtregar,fe- instead of -sureg-?

[Giovanni Angioni]

Dear Carlo,

Thank you very much for your reply. This is very much useful indeed as I am struggling to choose the right estimator. I guess -xtregar,fe- could be a valid solution, I am thinking about setting-up a model with lags of independent variables only with it, and then try xtpcse introducing a lag of the dependent as well.

A slightly unrelated question: the main dep variable (opinion polls' shares) is non stationary as it is, do you think it could be a major problem in this panel setting? First-differencing makes it stationary, but the interpretation of the coefficient becomes less straightforward.

Thank you very much again!
Giovanni

[Carlo Lazzaro]

Giovanni:
1) unit root and non stationarity: how to deal with this issue mostrly depends on what others did in your research field;
2) are you sure that by lagging the dependent variable you do not jump into the tricky realm of dynamic panel data models?

[Giovanni Angioni]

Dear Carlo:

Regarding (2), yes, I am aware that by including a lag of the dependent I am moving to dynamic panel data models, that's why I was considering using -xtpcse- which should be capable to do so and should be a good solution with T>N (as opposed to, for instance, xtabond etc). Don't you think it's a good solution?
Many thanks!
Giovanni

[Carlo Lazzaro]

Giovanni:
most part of the literature I've read on dynamic panel regression models deals with the small N, large T case.
I do not know whether -xtpcse- can replace -xtabond- when you have large T, small N datasets, as I'm not familiar with the latter when they go dynamic