sandwich package: HAC estimators
On Wed, 1 Jun 2016, T.Riedle wrote:
Thank you very much. I have applied the example to my case and get
following results:
crisis_bubble4<-glm(stock.market.crash~crash.MA+bubble.MA+MP.MA+UTS.MA+UPR.MA+PPI.MA+RV.MA,family=binomial("logit"),data=Data_logitregression_movingaverage)
summary(crisis_bubble4)
Call:
glm(formula = stock.market.crash ~ crash.MA + bubble.MA + MP.MA +
UTS.MA + UPR.MA + PPI.MA + RV.MA, family = binomial("logit"),
data = Data_logitregression_movingaverage)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.7828 -0.6686 -0.3186 0.6497 2.4298
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -5.2609 0.8927 -5.893 3.79e-09 ***
crash.MA 0.4922 0.4966 0.991 0.32165
bubble.MA 12.1287 1.3736 8.830 < 2e-16 ***
MP.MA -20.0724 96.9576 -0.207 0.83599
UTS.MA -58.1814 19.3533 -3.006 0.00264 **
UPR.MA -337.5798 64.3078 -5.249 1.53e-07 ***
PPI.MA 729.3769 73.0529 9.984 < 2e-16 ***
RV.MA 116.0011 16.5456 7.011 2.37e-12 ***
---
Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 869.54 on 705 degrees of freedom
Residual deviance: 606.91 on 698 degrees of freedom
AIC: 622.91
Number of Fisher Scoring iterations: 5
coeftest(crisis_bubble4)
z test of coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -5.26088 0.89269 -5.8933 3.786e-09 ***
crash.MA 0.49219 0.49662 0.9911 0.321652
bubble.MA 12.12868 1.37357 8.8300 < 2.2e-16 ***
MP.MA -20.07238 96.95755 -0.2070 0.835992
UTS.MA -58.18142 19.35330 -3.0063 0.002645 **
UPR.MA -337.57985 64.30779 -5.2494 1.526e-07 ***
PPI.MA 729.37693 73.05288 9.9842 < 2.2e-16 ***
RV.MA 116.00106 16.54560 7.0110 2.366e-12 ***
---
Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1
coeftest(crisis_bubble4,vcov=NeweyWest)
z test of coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -5.26088 5.01706 -1.0486 0.29436
crash.MA 0.49219 2.41688 0.2036 0.83863
bubble.MA 12.12868 5.85228 2.0725 0.03822 *
MP.MA -20.07238 499.37589 -0.0402 0.96794
UTS.MA -58.18142 77.08409 -0.7548 0.45038
UPR.MA -337.57985 395.35639 -0.8539 0.39318
PPI.MA 729.37693 358.60868 2.0339 0.04196 *
RV.MA 116.00106 79.52421 1.4587 0.14465
---
Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1
waldtest(crisis_bubble4, vcov = NeweyWest,test="F")
Wald test Model 1: stock.market.crash ~ crash.MA + bubble.MA + MP.MA + UTS.MA + UPR.MA + PPI.MA + RV.MA Model 2: stock.market.crash ~ 1 Res.Df Df F Pr(>F) 1 698 2 705 -7 2.3302 0.02351 * --- Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1
waldtest(crisis_bubble4, vcov = NeweyWest,test="Chisq")
Wald test Model 1: stock.market.crash ~ crash.MA + bubble.MA + MP.MA + UTS.MA + UPR.MA + PPI.MA + RV.MA Model 2: stock.market.crash ~ 1 Res.Df Df Chisq Pr(>Chisq) 1 698 2 705 -7 16.311 0.02242 * --- Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1 Do you agree with the methodology?
Well, this is how you _can_ do what you _wanted_ to do. I already expressed my doubts about several aspects. First, some coefficients and their standard errors are very large which may (or may not) hint at problems that are close to separation. Second, given the increase in the standard errors, the autocorrelation appears to be substantial and it might be good to try to capture these autocorrelations explicitly rather than just correcting the standard errors.
I read in a book that it is also possible to use vcov=vcovHAC in the coeftest() function.
Yes. (I also mentioned that in my e-mail yesterday, see below.)
Nevertheless, I am not sure what kind of HAC I generate with this command. Which weights does this command apply, which bandwith and which kernel?
Please consult vignette("sandwich", package = "sandwich") for the details.
In short: Both, vcovHAC and kernHAC use the quadratic spectral kernel with
Andrews' parametric bandwidth selection. The latter function uses
prewhitening by default while the latter does not. In contrast, NeweyWest
uses a Bartlett kernel with Newey & Wests nonparametric lag/bandwidth
selection and prewhitening by default.
Kind regards
________________________________________
From: Achim Zeileis <Achim.Zeileis at uibk.ac.at>
Sent: 31 May 2016 17:19
To: T.Riedle
Cc: r-help at r-project.org
Subject: Re: [R] sandwich package: HAC estimators
On Tue, 31 May 2016, T.Riedle wrote:
Many thanks for your feedback.
If I get the code for the waldtest right I can calculate the Chi2 and
the F statistic using waldtest().
Yes. In a logit model you would usually use the chi-squared statistic.
Can I use the waldtest() without using bread()/ estfun()? That is, I
estimate the logit regression using glm() e.g. logit<-glm(...) and
insert logit into the waldtest() function.
Does that work to get chi2 under HAC standard errors?
I'm not sure what you mean here but I include a worked example. Caveat:
The data I use are cross-section data with an overly simplified set of
regressors. So none of this makes sense for the application - but it shows
how to use the commands.
## load AER package which provides the example data
## and automatically loads "lmtest" and "sandwich"
library("AER")
data("PSID1976", package = "AER")
## fit a simple logit model and obtain marginal Wald tests
## for the coefficients and an overall chi-squared statistic
m <- glm(participation ~ education, data = PSID1976, family = binomial)
summary(m)
anova(m, test = "Chisq")
## replicate the same statistics with coeftest() and lrtest()
coeftest(m)
lrtest(m)
## the likelihood ratio test is asymptotically equivalent
## to the Wald test leading to a similar chi-squared test here
waldtest(m)
## obtain HAC-corrected (Newey-West) versions of the Wald tests
coeftest(m, vcov = NeweyWest)
waldtest(m, vcov = NeweyWest)
Instead of NeweyWest other covariance estimators (e.g., vcovHAC, kernHAC,
etc.) can also be plugged in.
hth,
Z
________________________________________
From: Achim Zeileis <Achim.Zeileis at uibk.ac.at>
Sent: 31 May 2016 13:18
To: T.Riedle
Cc: r-help at r-project.org
Subject: Re: [R] sandwich package: HAC estimators
On Tue, 31 May 2016, T.Riedle wrote:
I understood. But how do I get the R2 an Chi2 of my logistic regression
under HAC standard errors? I would like to create a table with HAC SE
via e.g. stargazer().
Do I get these information by using the functions
bread.lrm <- function(x, ...) vcov(x) * nobs(x)
estfun.lrm <- function(x, ...) residuals(x, "score")?
Do I need to use the coeftest() in this case?
The bread()/estfun() methods enable application of vcovHAC(), kernHAC(),
NeweyWest(). This in turn enables the application of coeftest(),
waldtest(), or linearHypothesis() with a suitable vcov argument.
All of these give you different kinds of Wald tests with HAC covariances
including marginal tests of individual coefficients (coeftest) or global
tests of nested models (waldtest/linearHypothesis). The latter can serve
as replacement for the "chi-squared test". For pseudo-R-squared values I'm
not familiar with HAC-adjusted variants.
And I'm not sure whether there is a LaTeX export solution that encompasses
all of these aspects simultaneously.
________________________________________
From: R-help <r-help-bounces at r-project.org> on behalf of Achim Zeileis <Achim.Zeileis at uibk.ac.at>
Sent: 31 May 2016 08:36
To: Leonardo Ferreira Fontenelle
Cc: r-help at r-project.org
Subject: Re: [R] sandwich package: HAC estimators
On Mon, 30 May 2016, Leonardo Ferreira Fontenelle wrote:
Em S?b 28 mai. 2016, ?s 15:50, Achim Zeileis escreveu:
On Sat, 28 May 2016, T.Riedle wrote:
I thought it would be useful to incorporate the HAC consistent
covariance matrix into the logistic regression directly and generate an
output of coefficients and the corresponding standard errors. Is there
such a function in R?
Not with HAC standard errors, I think.
Don't glmrob() and summary.glmrob(), from robustbase, do that?
No, they implement a different concept of robustness. See also
https://CRAN.R-project.org/view=Robust
glmrob() implements GLMs that are "robust" or rather "resistant" to
outliers and other observations that do not come from the main model
equation. Instead of maximum likelihood (ML) estimation other estimation
techniques (along with corresponding covariances/standard errors) are
used.
In contrast, the OP asked for HAC standard errors. The motivation for
these is that the main model equation does hold for all observations but
that the observations might be heteroskedastic and/or autocorrelated. In
this situation, ML estimation is still consistent (albeit not efficient)
but the covariance matrix estimate needs to be adjusted.
Leonardo Ferreira Fontenelle, MD, MPH
______________________________________________
R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
______________________________________________
R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
______________________________________________
R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
______________________________________________
R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
______________________________________________
R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.