[R] Robust vce for heckman estimators
Achim.Zeileis at uibk.ac.at
Mon Jul 11 23:40:29 CEST 2011
On Mon, 11 Jul 2011, Mateus Rabello wrote:
> When using function heckit() from package ???sampleSelection???, is
> there anyway to make t-tests for the coefficients using robust
> covariance matrix estimator? By ???robust??? I mean something like if a
> had an object ???lm??? called ???reg??? and then used:
>> coeftest(reg, vcov = vcovHC(reg)).
You can do essentially the same for selection models with sandwich
standard errors. For example:
data("PSID1976", package = "AER")
PSID1976$nwincome <- with(PSID1976, (fincome - hours * wage)/1000)
reg <- selection(participation ~ nwincome + education + experience +
I(experience^2) + age + youngkids + oldkids,
log(wage) ~ education + experience + I(experience^2),
data = PSID1976)
coeftest(reg, vcov = sandwich)
Simple "sandwich" standard errors are available while other "vcovHC"
standard errors (such as HC2, HC3, etc.) are not available for many models
beyond linear regression.
Note also that I used the selection() function above which is typically
preferable to heckit(), because the former produces the maximum likelihood
foot. The latter by default produces the 2-step estimator which is
nowadays usually only of interest in replication studies.
> I???m asking this because in Stata we could use function heckman and
> then use vce option ???robust???. We could do the same for cluster.
> In a more general way, is there anyway to use another covariance matrix
> to make t-test (e.g. linear hypothesis) for heckit (selection) models?
> Mateus Rabello
> [[alternative HTML version deleted]]
More information about the R-help