On Feb 5, 2005, at 12:40 PM, Apostolos Ballas wrote:
> Is there a simple way to test equality of coefficients from
regressions
> estimated over a number of sub-samples using the by command?
>
> To be more specific, lets say I estimate
>
> By group:regress y x1 x2
>
> How can I test (automatically) whether
>
> b11=b12=b13=...
>
> Where bij refers to the coefficients of the ith independent estimated
> in the
> jth sub-sample.
and Michael replied
Would it be econometrically equivalent to create dummy
variables for
(n-1) of your sub-samples (equivalently, I think, your groups),
interact them with x1 & x2, and test the joint significance of the set
of interactions with each x? This approach could be made more compact
with the -xi- command. (You'll likely want dummies on the intercept
for the same (n-1) sub-samples as well.) HTH.
It would not be econometrically equivalent since the single regression,
fully interacted including the constant, is being estimated with a
constraint of a single sigma^2 across the groups. Unless that is a
tenable assumption, it will lead to biased standard errors. Since the
resulting heteroskedasticity is of known form, handling it with
-robust- is not the best idea.
I believe this is a proper application of _suest_ (not be be confused
with _sureg_). whelp suest. Rather than using _by_, you need to use
_forvalues_ or _foreach_ to iterate over the groups and run the
regressions and store estimates for each: e.g.
webuse grunfeld,clear
drop if company>4
estimates clear
qui forv i=1/4 {
regress invest mvalue kstock if company==`i'
estimates store co`i'
}
suest *
test [co1_mean]mvalue=[co2_mean]mvalue
test [co1_mean]mvalue=[co3_mean]mvalue,accum
test [co1_mean]mvalue=[co4_mean]mvalue,accum
