There is a bug in the option "cluster" for clustered standard errors in OLS estimates: they are not computed in panel data. This is a simple code to reproduce the issue: open greene14_1.gdt logs C Q PF LF series d = (unit <= 3) ols l_C 0 l_Q l_PF l_LF setobs unit year --panel-vars ols l_C 0 l_Q l_PF l_LF --cluster=d setobs 1 1 ols l_C 0 l_Q l_PF l_LF --cluster=d The second and third estimates should give the same result. However, the s.e. in the second estimate are equal to the first one...

Moreover, the same option cluster does not work and returns an error if there is at least one missing observation.

On Fri, 2013-08-02 at 17:30 +0100, Allin Cottrell wrote: > On Thu, 1 Aug 2013, Giuseppe Vittucci wrote: >> Moreover, the same option cluster does not work and returns an error if >> there is at least one missing observation. > > I cannot replicate that -- could you give an example? It was really hard to replicate the issue.... Here is a simple example: nulldata 10000 series y = normal(0,1) series x = 0 series d = 0 loop for i=1..100 --quiet smpl index > 100*(i-1) && index <= 100*i --restrict --replace d = i y = y + i x = randgen1(n,0,1) endloop smpl full ols y 0 x ols y 0 x --cluster=d This works. There are 100 clusters each one made up of 100 obs. Now, if you remove from y the first observation of each cluster, i.e. loop for i=1..100 --quiet j = 100*(i-1)+1 y[j] = NA endloop or the i-th obs in the i-th cluster: loop for i=1..100 --quiet j = 100*(i-1)+i y[j] = NA endloop and run: ols y 0 x --cluster=d gretl returns an error: Missing values encountered Error executing script: halting This instead works: smpl ok(y) --restrict ols y 0 x --cluster=d