Identification Restrictions in Ordered Probit Model

[BigUnit2]

I want to estimate an ordered probit model where the dependent variable take the values -1, 0, 1.
I want to regress the dependent variable only on a constant term and have created a dummy=1 for that purpose.
When I run this, EV returns the error message:

"WARNING: Singular covariance - coefficients are not unique Covariance matrix computed using second derivatives"

Is it possible to make some identification restrictions that would allow me to estimate this? For instance, can I set the variance in the Likelihood Function equal to 1, so as to identify my thresholds properly?
thanks,
BU


EViews 6; July 12 2007 build.

[EViews Glenn]

Just use "C" as your regressor. We automatically add it to the specification if it is not there, and are not detecting that your dummy variable is really the same thing.

As an aside, I would suggest that you recode your indicator variable as 0, 1, 2 (you can do this on-the-fly by adding 1 to your dependent variable, as in the specification "(y+1) c". Most things should work fine even if you don't do the recoding, but there are a couple of naming things that don't like having the negative indicator.

[BigUnit2]

Glenn, thanks for the reply.
I tried using just "C" instead of my dummy variable but I got an error message saying:

"Regressors specified by list contains coefficients. Specify by explicit equation, i.e. Y=C(1) + C(2)*X or correct error in"

When I specify explicitly as Y=C(1) I get an error message as follows:

"Illegal specification for ordered equation. The derivative of C(1) is a constant."

When I specify as Y=C(1)*Dummy, I get the same problem as before that the coefficients aren't properly identified.

Also, note that I recoded Y into 0,1,2 as suggested. However, that didn't change the result.

So for now, I'm still stuck.

[EViews Glenn]

What was the first specification you tried?

[BigUnit2]

First specification was:
_ Y Dummy
_ where Dummy = 1.

thanks,
BU

[EViews Glenn]

Just do

Y C

[BigUnit2]

I tried just Y C.
See my post above for the error message I got.
(Postby BigUnit2 on Wed Jul 21, 2010 3:04 pm)

[startz]

I hate to be a party-pooper, but the reason that BigUnit2 is having a problem is most likely that EViews is busted.
Here's a screenshot.

Capture.JPG (48.74 KiB) Viewed 171 times

I hit OK and the next thing I knew.

Capture.JPG (22.98 KiB) Viewed 171 times

Here's the code

Code: Select all

create u 100
series x=nrnd
series y = 0
smpl if x<-0.5
y= -1
smpl if x>0.5
y=1
smpl @all
ls y c

[BigUnit2]

Thank you Startz... I don't know if EV is busted, but you seem to have identified the same problem that I am having.

Furthermore, in the example you gave, specifying Y=C(1) returns the following error:

error.JPG (10.67 KiB) Viewed 180 times

[startz]

On further reflection, maybe it's just an uninformative error message.

I think that in an ordered probit the intercept isn't identified, because it's effectively collinear with the cut points.

Glenn? You're the expert.

[EViews Glenn]

Sorry guys. Does look like a bug to me, but I'll investigate.

The perfect collinearity of C with the cutpoints is automatically handled by EViews. That's what I thought the original problem was with the user specifying their own constant term in the form of dumm=1. I'll take a look.

[EViews Glenn]

Apologies. A bug, with a cryptic error message. I'll get a fix posted soon.

[EViews Glenn]

The fix has been posted.

[BigUnit2]

Glenn,
The patch has only been posted for EV 7.
My organization is still on EV 6 and that's what I'm using.
Please provide a patch for EV 6 as well.
Thanks,
BU

[EViews Glenn]

I'll see what I can do, but it may take some time before I can get a 6 patch put together.