From: gaetan on
Dear all,

I have data with four class dimensions: countries, year, sector, size
of company (qualitative variable).

I construct dummies for the industry and use PROC REG to estimate an
equation in the form Y = X + industry dummies. I would like to make a
Hausman test to check whether my industry fixed effects are not in fact
random effects. Unfortunately, because I have more than two dimensions,
neither PROC TSCSREG nor PROC PANEL accept to do the regressions (there
is more than one observation per country-year). It seems that my only
option is PROC MIXED.

PROC MIXED does not allow to compute Hausman tests but I try the
significance of a random effect in two ways (as suggested by a SAS
employee):

(a) I run the model as a pure fixed effect model. The type III F-test
shall give me the significance of the the fixed effect. If it is
significantly different from zero, then I prefer the fixed effect
model.

(b) I run the model as fixed effect and then as random effect. I look
at the difference in the -2 Log Likelihood values. The difference is
distributed as a chi-square with one degree of freedom. I look for the
1-prob of the distribution of the chi-square. The test rejects or does
not reject random effects.

I would like to have your opinion. I am doing things in a correct way?

Thank you.

Gaetan

From: gaetan on
Dear all,

Is there a way to test in proc mixed whether an effect is fixed or
random (besides the BIC and AIC criteria)?

Thank you.
Best,
Gaetan

gaetan a écrit :

> Dear all,
>
> I have data with four class dimensions: countries, year, sector, size
> of company (qualitative variable).
>
> I construct dummies for the industry and use PROC REG to estimate an
> equation in the form Y = X + industry dummies. I would like to make a
> Hausman test to check whether my industry fixed effects are not in fact
> random effects. Unfortunately, because I have more than two dimensions,
> neither PROC TSCSREG nor PROC PANEL accept to do the regressions (there
> is more than one observation per country-year). It seems that my only
> option is PROC MIXED.
>
> PROC MIXED does not allow to compute Hausman tests but I try the
> significance of a random effect in two ways (as suggested by a SAS
> employee):
>
> (a) I run the model as a pure fixed effect model. The type III F-test
> shall give me the significance of the the fixed effect. If it is
> significantly different from zero, then I prefer the fixed effect
> model.
>
> (b) I run the model as fixed effect and then as random effect. I look
> at the difference in the -2 Log Likelihood values. The difference is
> distributed as a chi-square with one degree of freedom. I look for the
> 1-prob of the distribution of the chi-square. The test rejects or does
> not reject random effects.
>
> I would like to have your opinion. I am doing things in a correct way?
>
> Thank you.
>
> Gaetan

From: Peter Flom on
>>> gaetan <ganicode(a)ULB.AC.BE> 11/28/2006 11:41 am >>> wrote
<<<
Is there a way to test in proc mixed whether an effect is fixed or
random (besides the BIC and AIC criteria)?
>>>

At least one of us is confused :-)

If I am interpreting your question, you want to tell if an effect is
fixed or random. But this is determined by the user, based on the
design.

Am I interpreting your question correctly?

Peter
From: David L Cassell on
ganicode(a)ULB.AC.BE wrote (again):
>
>Dear all,
>
>Is there a way to test in proc mixed whether an effect is fixed or
>random (besides the BIC and AIC criteria)?
>
>Thank you.
>Best,
>Gaetan
>
>gaetan a �crit :
>
> > Dear all,
> >
> > I have data with four class dimensions: countries, year, sector, size
> > of company (qualitative variable).
> >
> > I construct dummies for the industry and use PROC REG to estimate an
> > equation in the form Y = X + industry dummies. I would like to make a
> > Hausman test to check whether my industry fixed effects are not in fact
> > random effects. Unfortunately, because I have more than two dimensions,
> > neither PROC TSCSREG nor PROC PANEL accept to do the regressions (there
> > is more than one observation per country-year). It seems that my only
> > option is PROC MIXED.
> >
> > PROC MIXED does not allow to compute Hausman tests but I try the
> > significance of a random effect in two ways (as suggested by a SAS
> > employee):
> >
> > (a) I run the model as a pure fixed effect model. The type III F-test
> > shall give me the significance of the the fixed effect. If it is
> > significantly different from zero, then I prefer the fixed effect
> > model.
> >
> > (b) I run the model as fixed effect and then as random effect. I look
> > at the difference in the -2 Log Likelihood values. The difference is
> > distributed as a chi-square with one degree of freedom. I look for the
> > 1-prob of the distribution of the chi-square. The test rejects or does
> > not reject random effects.
> >
> > I would like to have your opinion. I am doing things in a correct way?
> >
> > Thank you.
> >
> > Gaetan

I agree with Peter.

If you are making inferences only about the given set of industries
and no others, then you have a fixed effect. If you are making
statistical inferences across a much wider population of industries,
then you have a random effect.

Also, I would suggest that you do NOT build a pile of dummy variables
for the industries. SAS procs tend to do a much better job of this
than people do. And if you want ot use one of the stat procs which
doesn't do this, then you can use a tool like PROC GLMMOD or PROC
TRANSREG to do the hard work for you.

I don't think that you really have 4 orthogonal class dimensions.
You have year and industry. Granted, industries are clustered
in some way under your variable SECTOR, and they have an
additional variable of company size tacked on. But you should
be able to start with a year-by-industry analysis and see how
that handles the fundamental estimation process first.

HTH,
David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330

_________________________________________________________________
Share your latest news with your friends with the Windows Live Spaces
friends module.
http://clk.atdmt.com/MSN/go/msnnkwsp0070000001msn/direct/01/?href=http://spaces.live.com/spacesapi.aspx?wx_action=create&wx_url=/friends.aspx&mk