[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: Why no sum of squared residuals in reg output after robust or cluster?
At 08:41 PM 3/3/2004 -0600, firstname.lastname@example.org wrote:
This is probably easy, but I have been wondering...
This is a good chance to try out my scanner and optical character
recognition software. Here is what the manual says:
When you add the comand robust or cluster after the regression you don't
the regression output the information about the explained sum of squares, the
residual sum of squares and the total sum of squares, information that you
usually get if you perform regress without the robust or the cluster command.
Why is it so?
Is it that the residual sum of square is not reliable when you correct for
heteroscedasticity with the Huber/White/Sandwich estimator, or if you cluster
the standard errors?
But then, how is the R-square (whch is reported, but not the adjusted one)
From p. 337, Stata 8 Reference Manual N-R (from the documentation for the
When you specify robust, regress purposefully suppresses displaying the
ANOVA table, as it is no longer appropriate in a statistical sense even
though, mechanically, the numbers would be unchanged. That is, sums of
squares remain unchanged, but the meaning you might be tempted to give
those sums is no longer relevant. The F statistic, for instance, is no
longer based on sums of squares; it becomes a Wald test based on the
robustly estimated variance matrix. Nevertheless, regress continues to
report the R^2 and the root MSE even though both numbers are based on sums
of squares and are, strictly speaking, irrelevant. In this, the root MSE is
more in violation of the spirit of the robust estimator than is R^2. As a
goodness-of-fit statistic, R^2 is still fine; just do not use it in
formulas to obtain F statistics because those formulas no longer apply. The
Root MSE is valid as long as you take its name literally - it is the square
root of the mean square error. Root MSE is no longer a prediction of sigma
because there is no single sigma; the variance of the residual varies
observation by observation.
* For searches and help try: