Dear group,

I use multinomial logit model to predict directional change in interest
rate. The output from LIMDEP shows that the standard errors of all the
coefficients for one of the categories are extremely large, resulting
in virtually zero t-ratios. Does anyone know why this might be the
case? Any thought is much appreciated. Thank you.

Thanaset

Dear group,

I use multinomial logit model to predict directional change in interest
rate. The output from LIMDEP shows that the standard errors of all the
coefficients for one of the categories are extremely large, resulting
in virtually zero t-ratios. Does anyone know why this might be the
case? Any thought is much appreciated. Thank you.

Look at the data for those categories: you might simply have few cases.

Thanaset wrote:
Dear group,

I use multinomial logit model to predict directional change in interest
rate. The output from LIMDEP shows that the standard errors of all the
coefficients for one of the categories are extremely large, resulting
in virtually zero t-ratios. Does anyone know why this might be the
case? Any thought is much appreciated. Thank you.

Look at the data for those categories: you might simply have few cases.
Another alternative is that the cases are all either successes or
failures, so the point estimates of the coefficients are either plus or
minus infinity. Software packages will get as close as they can, and
then give up. You should have got a warning about this:I don't know
LIMDEP, so I don't know if it does this. The estimates of the standard
errors are then awful, because they depend on the likelihood to be well
enough behaved.

Bob

In article <engloc$266$1@oravannahka.helsinki.fi>,
Bob O'Hara <bob.ohara@helsinki.fi> wrote:
Thanaset wrote:
Dear group,

I use multinomial logit model to predict directional change in interest
rate. The output from LIMDEP shows that the standard errors of all the
coefficients for one of the categories are extremely large, resulting
in virtually zero t-ratios. Does anyone know why this might be the
case? Any thought is much appreciated. Thank you.

Look at the data for those categories: you might simply have few cases.
Another alternative is that the cases are all either successes or
failures, so the point estimates of the coefficients are either plus or
minus infinity. Software packages will get as close as they can, and
then give up.

In that case, the coefficients should be 0, except for the
constant term; but the estimated variance should also be zero.
Something close to that might occur.

You should have got a warning about this; I don't know
LIMDEP, so I don't know if it does this. The estimates of the standard
errors are then awful, because they depend on the likelihood to be well
enough behaved.

This is something not realized; the estimates of standard
error, and the t-tests, etc, are based on asymptotic theory;
how good this is for samples of reasonable size is not clear.
One can look at the behavior of the likelihood function,
which gives a better idea than merely its second derivative
matrix at the maximum.

Bob


--
This address is for information only. I do not claim that these views
are those of the Statistics Department or of Purdue University.
Herman Rubin, Department of Statistics, Purdue University
hrubin@stat.purdue.edu Phone: (765)494-6054 FAX: (765)494-0558

In article <engloc$266$1@oravannahka.helsinki.fi>,
Bob O'Hara <bob.ohara@helsinki.fi> wrote:
Thanaset wrote:
Dear group,

I use multinomial logit model to predict directional change in interest
rate. The output from LIMDEP shows that the standard errors of all the
coefficients for one of the categories are extremely large, resulting
in virtually zero t-ratios. Does anyone know why this might be the
case? Any thought is much appreciated. Thank you.

Look at the data for those categories: you might simply have few cases.
Another alternative is that the cases are all either successes or
failures, so the point estimates of the coefficients are either plus or
minus infinity. Software packages will get as close as they can, and
then give up.

In that case, the coefficients should be 0, except for the
constant term; but the estimated variance should also be zero.
Something close to that might occur.

I don't follow this: in effect you're suggesting that the fitted

You should have got a warning about this; I don't know
LIMDEP, so I don't know if it does this. The estimates of the standard
errors are then awful, because they depend on the likelihood to be well
enough behaved.

This is something not realized; the estimates of standard
error, and the t-tests, etc, are based on asymptotic theory;
how good this is for samples of reasonable size is not clear.
One can look at the behavior of the likelihood function,
which gives a better idea than merely its second derivative
matrix at the maximum.

True: I think the likelihood surface flattens out when the fitted

Herman Rubin wrote:

You should have got a warning about this; I don't know
LIMDEP, so I don't know if it does this. The estimates of the
standard errors are then awful, because they depend on the likelihood
to be well enough behaved.

This is something not realized; the estimates of standard error, and
the t-tests, etc, are based on asymptotic theory;
how good this is for samples of reasonable size is not clear.
One can look at the behavior of the likelihood function,
which gives a better idea than merely its second derivative
matrix at the maximum.

True: I think the likelihood surface flattens out when the fitted
probabilities approach zero, and the recommendation is to use a profile
likelihood approach to get confidence intervals. It may be that the OP
doesn't need to go this far, though.

Just to correct myself: the quadratic approximation to the likelihood is

Bob O'Hara wrote...

"The estimates of the standard
errors are then awful, because they depend on the likelihood to be well

enough behaved."

That's a strange statement. OMU

I was trying to avoid explaining the technicalities: see my reply to

Dear group,

I use multinomial logit model to predict directional change in interest
rate. The output from LIMDEP shows that the standard errors of all the
coefficients for one of the categories are extremely large, resulting
in virtually zero t-ratios. Does anyone know why this might be the
case? Any thought is much appreciated. Thank you.

Herman Rubin wrote:
In article <engloc$266$1@oravannahka.helsinki.fi>,
Bob O'Hara <bob.ohara@helsinki.fi> wrote:
Thanaset wrote:
Dear group,

I use multinomial logit model to predict directional change in interest
rate. The output from LIMDEP shows that the standard errors of all the
coefficients for one of the categories are extremely large, resulting
in virtually zero t-ratios. Does anyone know why this might be the
case? Any thought is much appreciated. Thank you.

Look at the data for those categories: you might simply have few cases.
Another alternative is that the cases are all either successes or
failures, so the point estimates of the coefficients are either plus or
minus infinity. Software packages will get as close as they can, and
then give up.

In that case, the coefficients should be 0, except for the
constant term; but the estimated variance should also be zero.
Something close to that might occur.

I don't follow this: in effect you're suggesting that the fitted
probability of successes would be the same as at the intercept. But if
the intercept gives a fitted probability that is non-zero, then for the
classes where there are no successes, then the fitted probability should
be zero, i.e. the effect size (on the logit scale) would be minus a
large number, not the intercept probability.

You should have got a warning about this; I don't know
LIMDEP, so I don't know if it does this. The estimates of the standard
errors are then awful, because they depend on the likelihood to be well
enough behaved.

This is something not realized; the estimates of standard
error, and the t-tests, etc, are based on asymptotic theory;
how good this is for samples of reasonable size is not clear.
One can look at the behavior of the likelihood function,
which gives a better idea than merely its second derivative
matrix at the maximum.

True: I think the likelihood surface flattens out when the fitted
probabilities approach zero, and the recommendation is to use a profile
likelihood approach to get confidence intervals. It may be that the OP
doesn't need to go this far, though.

Bob

Some who posted to this OP are rambling on about the symptoms, not the
cure. The fundamental fact underlying this seems to be "there's very
little informatiokn about that model coefficient(s) in the data."

summary(glm(cbind(10000,0)~1, family=binomial()))