estimation using 3 sls and sampling

[smathey]

Hi All,

Can anyone help in three-stage least square specification problems, Eviews 6

I used (I have more variables at the RHS but following the same logic)

growth-period1 c0+c1*gdp1+c2*democ1 (1)
growth-period2 co+c1*gdp2+c2*democ2 (2)
growth-period3 co+c1*gdp3+c2*democ3 (3)

I have 213 countries observations but I have missing observations (eg 98 gdp1 observations while 62 observations for gdp2).
At the end I found eviews regressing on the minimum observations (at times 25 observations, as eviews pick up only countries with observations in all series).

How can I "force" eviews to use the max number of observations- using in the regression countries that have missing observations (I was reading on individual sampling but it worked only for data statistics).

I am running on the side endogeneity with ols, eg gdp1_hat co+C1*FDI

Is there a better way to specify the equations (1) to (3)?

[startz]

I don't believe you can do 3SLS with missing observations because you need the contemporaneous covariance matrix. You might try 2sls in the system object.

[smathey]

Thanks a lot.
In fact I am mimicking a Barro growth's 3 sls.

I have this quoted remarks from professor. Please, excuse my stupidity but can anyone indicate how that can be performed in a PRACTICAL way?
PROF COMMENTS
First, for some specifications you endogenize X1 and X2. How did you do this? Did your software do this automatically, or did you perform indirect least squares, estimating the first stage regression and then using the fitted values in subsequent estimates. If your software estimated the first stage as part of the 3SLS estimates, there is no problem. If you estimated the first stage “by hand” and then used the fitted values as a new variable, did you correct your variance-covariance matrix?
If you perform IV estimates “by hand,” the resulting variance-covariance estimate is incorrect. To estimate the coefficient of an endogenous variable, the fitted values from the first stage estimates are used. For the variance-covariance matrix, the actual, not fitted, values of the endogenous variable are used. Software programs do this automatically, but if you conduct indirect least squares estimates, performing IV “by hand,” you must estimate the variance-covariance matrix yourself.