2:05 a.m.
Dear everyone,
I conducted a random effects panel model (90 observations, 15 cross-sectional units, 6
time-series length). The model has only one dichotomous predictor (coded 0 and 1). May i
know why Hausman Test is not generated by Gretl?
The Breusch-Pagan test is significant (p<.001). I would appreciate any suggestion on
the appropriate model for my data. That is, is it safe to use random effects or fixed
effects? Again, why Hausman Test is not computed by gretl?
Thank you.Lema
4:41 a.m.
Lema,
The short answer is not bother too much with Hausman tests.
The long answer is:
(1) Re-parameterize your model as a multilevel model;
(2) Decompose your variables into 'fixed' (within) and 'random' (between)
components;
(3) Run it in R (through gretl, if you wish), and then;
(4) Do a joint F test of the equality of coefficients.
In R, it would look something like this:
<R>
library(car,lme4) # assuming you have these packages installed in your R
mydata<-read.csv("/.../mydata.csv",header=T,sep=",",fill=T)
attach(mydata) # R purists like to lecture you about not using -attach()-;
for now, ignore them
m_x1=rep(tapply(x1,groupvar,mean),tapply(x1,groupvar,length))
d_x1=x1-m_x1
m_x2=rep(tapply(x2,groupvar,mean),tapply(x2,groupvar,length))
d_x2=x2-m_x2
fit1<-lmer(y~(1|groupvar)+(1|timevar),REML=T)
summary(fit1)
fit2<-lmer(y~d_x1+m_x1+d_x2+m_x2+(1|groupvar)+(1|timevar),REML=T)
summary(fit2)
anova(fit1,fit2)
linearHypothesis(fit2,c("d_x1=m_x1","d_x2=m_x2"))
</R>
(NB: -timevar- should be ordered by way of a time counter starting at 1
*within* each of your -groupvars-.)
If you have run the above routine successfully, it means that you have now
tested for the equality (or not) of your fixed and random effects, thus
obviating the need for any Hausman tests.
I hope that helps and good luck.
C
On 26 April 2018 at 03:05, 3jlema <3jlema(a)gmail.com> wrote:
Dear everyone,
I conducted a random effects panel model (90 observations, 15
cross-sectional units, 6 time-series length). The model has only one
dichotomous predictor (coded 0 and 1). May i know why Hausman Test is not
generated by Gretl?
The Breusch-Pagan test is significant (p<.001). I would appreciate any
suggestion on the appropriate model for my data. That is, is it safe to use
random effects or fixed effects? Again, why Hausman Test is not computed by
gretl?
Thank you.
Lema
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
--
Clive Nicholas
"My colleagues in the social sciences talk a great deal about methodology.
I prefer to call it style." -- Freeman J. Dyson
10:11 a.m.
Am 26.04.2018 um 06:41 schrieb Clive Nicholas:
The short answer is not bother too much with Hausman tests.
Such an answer is not really helpful, and I don't think this personal
opinion of yours can be regarded as a professional consensus.
(1) Re-parameterize your model as a multilevel model;
(2) Decompose your variables into 'fixed' (within) and 'random'
(between) components;
(3) Run it in R (through gretl, if you wish), and then;
(4) Do a joint F test of the equality of coefficients.
now tested for the equality (or not) of your fixed and random
effects,
Typically the fixed or random effects are regarded as nuisance
parameters in these contexts. It presume it is not really interesting
for the application whether they are equal. That's one advantage of the
Hausman test which directly targets the coefficients of interest.
That said, your idea certainly has merit. There's nothing wrong with
using one of R's packages (especially with lme4 I've done it myself),
but providing more hints as to how to achieve it in gretl would even be
more welcome on a gretl list.
cheers,
sven
7:41 p.m.
On 26 April 2018 at 11:11, Sven Schreiber <svetosch(a)gmx.net> wrote:
Am 26.04.2018 um 06:41 schrieb Clive Nicholas:
The short answer is not bother too much with Hausman tests.
>
Such an answer is not really helpful, and I don't think this personal
opinion of yours can be regarded as a professional consensus.
*Frankly, I could care less whether such things are "professional
consensus" or not. It is my personal view and I'm happy to give it. If the
OP finds it useful, so much the better. If they prefer to chuck my post
straight in the bin, that's for them. Moreover, my view is supported by a
highly readable paper on the subject by Bell and Jones (2012) "Explaining
Fixed Effects: Random Effects Modelling of Time-Series Cross-Sectional and
Panel Data" (available via Google Scholar), which promotes this approach
over Hausman tests and explains why.*
(1) Re-parameterize your model as a multilevel model;
> (2) Decompose your variables into 'fixed' (within) and 'random'
(between)
> components;
> (3) Run it in R (through gretl, if you wish), and then;
> (4) Do a joint F test of the equality of coefficients.
>
now tested for the equality (or not) of your fixed and random effects,
>
Typically the fixed or random effects are regarded as nuisance parameters
in these contexts. It presume it is not really interesting for the
application whether they are equal. That's one advantage of the Hausman
test which directly targets the coefficients of interest.
That said, your idea certainly has merit. There's nothing wrong with using
one of R's packages (especially with lme4 I've done it myself), but
providing more hints as to how to achieve it in gretl would even be more
welcome on a gretl list.
*I explained to the OP that s/he could run the R code (if s/he so wished)
through -gretl-. You may have missed that bit. I would welcome any effort
to see HLMs introduced in -gretl-, but it's not really my place to ask for
it. If people want it badly enough, they can learn to code it themselves. I
can't, so I won't.*
cheers,
sven
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
--
Clive Nicholas
"My colleagues in the social sciences talk a great deal about methodology.
I prefer to call it style." -- Freeman J. Dyson
7:11 a.m.
Am 26.04.2018 um 21:41 schrieb Clive Nicholas:
On 26 April 2018 at 11:11, Sven Schreiber <svetosch(a)gmx.net
<mailto:svetosch@gmx.net>> wrote:
*Frankly, I could care less whether such things are
"professional
consensus" or not. It is my personal view and I'm happy to give it. If
the OP finds it useful, so much the better. If they prefer to chuck my
post straight in the bin, that's for them.
Well, at the risk of starting religious wars, I'm always happy to hear
arguments and reasons for personal views. But a concrete answer to a
concrete question ("why isn't the Hausman test shown") would also be
nice. For example, what if you asked a question about a multilevel model
and somebody just answers "Bayesian is better anyway" ? My guess is that
that wouldn't help you too much.
*I explained to the OP that s/he could run the R code (if s/he so
wished) through -gretl-. You may have missed that bit.
In fact I haven't. But this line:
mydata<-read.csv("/.../mydata.csv",header=T,sep=",",fill=T)
wouldn't work from gretl, or at least you would have to prepare stuff first.
Gretl has the "foreign" block apparatus which could be used here, no doubt.
I would welcome
any effort to see HLMs introduced in -gretl-, but it's not really
my
place to ask for it. If people want it badly enough, they can learn to
code it themselves. I can't, so I won't.*
Fair enough. I'm not an expert in that specific area, either. However,
it seems that at least Stata is using max likelihood there (judging from
their ME manual), so my guess is that for someone who knows these models
better it's not prohibitive, using the ML apparatus (mle block, see ch.
23 of the guide).
But that's not a recommendation for applied users, of course.
Thanks,
sven
8:39 a.m.
Am 26.04.18 um 04:05 schrieb 3jlema
Dear everyone,
I conducted a random effects panel model (90 observations, 15 cross-sectional units, 6
time-series length). The model has only one dichotomous predictor (coded 0 and 1). May i
know why Hausman Test is not generated by Gretl?
The Breusch-Pagan test is significant (p<.001). I would appreciate any suggestion on
the appropriate model for my data. That is, is it safe to use random effects or fixed
effects? Again, why Hausman Test is not computed by gretl?
If I understand correctly you only have one regressor. Could it be that it’s constant over
time such that fixed effects are impossible?
Thanks
Sven
2423
days inactive
2424
days old
5 comments
3 participants
participants (3)
-
3jlema -
Clive Nicholas -
Sven Schreiber