Hi,
I am comparing a treatment effect of a labour market policy in one country while keeping another country's observations as a control group. I decided to cluster the standard errors by country id, to allow for correlations in local labour markets. If I don't cluster, I get estimates which are not significant statistically. When I cluster, my standard errors are reduced greatly, so that they are reduced by about 10^-12, which makes my coefficients extraordinarily significant. I deduce that such reductions are due to negative intracluster correlations: I was wondering whether I should be satisfied with such reduction, or am I doing something wrong? There are some other group variables like employer id within these countries which I can cluster by, but they don't differ too much from the unclustered SEs.
Thanks,
Alex
I am comparing a treatment effect of a labour market policy in one country while keeping another country's observations as a control group. I decided to cluster the standard errors by country id, to allow for correlations in local labour markets. If I don't cluster, I get estimates which are not significant statistically. When I cluster, my standard errors are reduced greatly, so that they are reduced by about 10^-12, which makes my coefficients extraordinarily significant. I deduce that such reductions are due to negative intracluster correlations: I was wondering whether I should be satisfied with such reduction, or am I doing something wrong? There are some other group variables like employer id within these countries which I can cluster by, but they don't differ too much from the unclustered SEs.
Thanks,
Alex