[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
RE: st: RE: Missing Wald test with -cluster- or -robust- option [further explanations?]
> -----Original Message-----
> From: firstname.lastname@example.org
> [mailto:email@example.com] On Behalf Of Mirko
> Sent: 26 September 2006 20:51
> To: firstname.lastname@example.org
> Subject: Re: st: RE: Missing Wald test with -cluster- or
> -robust- option [further explanations?]
> Thank you Mark,
> Just adding this. You are right when you say that the Wald
> test is missing only after a adding some parameters. The
> variable was not a dummy though, but a squared variable.
> Your suggestion is to drop that variable even if that
> variable is very important for the analysis, right?
No. Or, more precisely, not necessarily.
The fact that the vcv isn't full rank is a warning that something might
be wrong, but it isn't guaranteed that you have a problem. To use the
example you originally cited, #parameters>#clusters only means that you
can't test the hypothesis that all the parameters are jointly zero (your
Wald test). But you can still test individual parameters.
An example that cuts the other way: if the problem were a singleton
dummy, you could test all the other parameters in the model, but you
couldn't get a sensible test out of your coefficient on this dummy,
because, after all, you basically have only one observation on it.
You need to look at the variable that is causing the vcv to be less than
full rank, work out why, and then decide what to do.
Prof. Mark Schaffer
Department of Economics
School of Management & Languages
Heriot-Watt University, Edinburgh EH14 4AS
tel +44-131-451-3494 / fax +44-131-451-3296
> It might be useful to know that I think there is no need to
> implement the loop you suggest. Stata suggests which
> parameter is causing troubles after testing by hand all the
> parameters in the model using -test-. I am not completely
> sure about this. Here is an example of Stata output.
> . oprobit y x1 x2 ... x34, cluster(c)
> (Output omitted)
> . test x1 x2...x34
> (1) x1=0
> (2) x2=0
> (34) x34=0
> Contraint 22 dropped
> chi2( 33) = 396.35
> Prob > chi2 = 0.0000
> So, Stata automatically add "Constraint 22 dropped" before the Chi2.
> Dropping manually the variable number 22 (i.e. x22) from my
> model I obtained an oprobit model where the Wald test is not
> missing anymore.
> On 9/26/06, Schaffer, Mark E <M.E.Schaffer@hw.ac.uk> wrote:
> > Mirko,
> > > -----Original Message-----
> > > From: email@example.com
> > > [mailto:firstname.lastname@example.org] On Behalf Of Mirko
> > > Sent: Tuesday, September 26, 2006 5:38 PM
> > > To: email@example.com
> > > Subject: st: Missing Wald test with -cluster- or -robust- option
> > > [further explanations?]
> > >
> > > Hi,
> > >
> > > I am estimating an Ordered probit model with standard
> error adjusted
> > > for clustering using -cluster- option. Unfortunately the
> model does
> > > not provide the (adjusted) Wald test for the model and
> the relative
> > > Prob>Chi2.
> > >
> > > According to Stata help this could be related to the number of
> > > parameters and the number of clusters (i.e., the number of
> > > parameters must be less than the clusters).
> > >
> > > Unfortunately, this is not my case. My sample size is 1,400
> > > observations, I have 94 clusters and my model contains
> around 40-50
> > > variables (most of them dummy 1/0).
> > >
> > > The same problem occurs if I use -robust- option istead of
> > > -cluster- or if I run -regress- (missing F and Prob>F
> > > missing) instead of -oprobit-.
> > My guess is that you have a singleton dummy (or something similar).
> > This causes the robust covariance matrix to be less than full rank.
> > You can confirm this by trying to do the Wald test by hand using
> > -test- or -testparm-. I suspect that you will get
> plausible looking
> > test statistics if you test most but not all of the
> parameters, and if
> > you keep adding parameters to the set you are testing, at
> some point
> > the test statistic will be missing. The last parameter you
> add will
> > correspond to the variable that causes the vcv to be less than full
> > rank.
> > Come to think of it, you could probably do this using a
> -foreach- loop
> > that loops over the variables in the estimation, and a
> -test- command
> > with the -accum- option so that you keep adding variables
> to the set
> > being tested.
> > Cheers,
> > Mark
> > Prof. Mark E. Schaffer
> > Director
> > Centre for Economic Reform and Transformation Department of
> > School of Management & Languages Heriot-Watt University
> Edinburgh EH14
> > 4AS UK
> > 44-131-451-3494 direct
> > 44-131-451-3296 fax
> > http://www.sml.hw.ac.uk/cert
> > > Does anybody know why this is happening? Or at least can
> give me an
> > > alternative explanation?
> > >
> > > Thanks for your help.
> > >
> > > Mirko Moro
> > > University College Dublin
> > > *
* For searches and help try: