"variance matrix is nonsymmetric or highly singular" in Stata; no SEs in table I get the above warning in Stata when I run a (clustered-panel pseudo-Poisson ML gravity-) regression. 
Observations are pairs of countries and the regression includes dummies for each combination of country and year (2005-2012) as well as a lagged variable of one continuous independent variable.
If I drop either the set of dummies or the lagged variable, it works fine. But of course I'd like the dummies AND the lag to be in that regression...
Can anyone tell me what causes that warning and how I can work around it?
Results can be found here: http://pastebin.com/BieZCZJL
(note: I am aware that the dependent should not be logged)
 A: Judging by your output, you have 4 country pairs with only one observation. In this case including dummy variable is wrong, since it explains all the variation in the cluster, making the variance of disturbance zero. Removal of such cluster sometimes help.
A: I was also getting this error for some regressions but occasionally with equivalent but different specifications Stata would provide Standard Errors suggesting some issue with Stata not my specification. My advisor suggested using a 2-step residualization method (see the answer by Tyrel Stokes here) when Stata gets overwhelmed.
Where the form of the regression is as follows: Y = X + Other_Controls
Step 1: Regress Dependent Variable (aka the Left Hand Side (LHS)) on All Controls Except Key Dependent Variable (X) for which you want standard errors (reg y Other_Controls). Then obtain the residuals (predict LHS_residuals, residuals)
Step 2: Regress Key Dependent Variable (X) (aka the Right Hand Side (RHS)) on the Other_Controls (reg X Other_Controls) then obtain the residuals from this regression (predict RHS_residuals, residuals)
Step 3: Regress LHS_Residuals on RHS_Residuals and you should get your standard errors for the key dependent variable of interest.
