I've run into the situation of not getting an overall F stat when
running a
regression with clustered standard errors. Previous postings and
Stata help
say that:
1) the number of estimated coefficients must be lower than the number of
clusters
2) each variable cannot have a non-zero value for just one observation
My final sample fulfils the above two.
I think the problem lies in having some of the dummy variables
corresponding to only one cluster. My data is clustered by 'case
investigation' (with several observations for each case) and each
case falls
into a particular industry. I have industry dummy variables.
Some of these industry dummies have only one corresponding 'case
investigation'. When I get rid of these 'one case' industry dummies,
I get
an F stat.
In a cluster covariance matrix estimator, you are essentially running
a regression using one observation per cluster, which is why
condition (1) above is important. In a standard regression, a dummy
with a single 1 will essentially remove that data point from the
analysis--that is, if you run the same regression without the dummy
and without the observation to which it pertains, you will get the
same results for the other parameters (and N-k, Root MSE, etc. will
be unchanged). But the ANOVA F is messed up because it miscounts the
slopes, considering that dummy to be a meaningful regressor rather
than a nuisance. Try
g dum=(_n==10)
reg hours kidslt6 kidsge6 dum
reg hours kidslt6 kidsge6 if _n!=10
The clickable help for that missing F-stat when you cluster with a
dummy that is only nonzero for one cluster says
Is there a regressor that is nonzero for only one observation?
The VCE you have just estimated is not of sufficient rank to
perform the model test. This
can happen if there is a variable in your model that is nonzero
for only a single observation
in the estimation sample. In that case the derivative of the
sum-of-squares or likelihood
function with respect to that variable's parameter is zero for
all observations. That
implies that the outer-product-of-gradients (OPG) variance
matrix is singular. Since the OPG
variance matrix is used in computing the robust variance matrix,
the latter is therefore
singular as well.
I think that StataCorp might want to expand this to include "is there
a regressor which is nonzero for only one cluster when you are using
the cluster option?"
