LOGL Error Message

[henry]

Could someone please tell me what this error message means?

"Missing values in @LOGL series at current coefficients at observation 1"

I'm trying to estimate a LOGL function. What might I have done wrong?

In case it helps, the logl function is as follows:

@logl l2
l2 = (1-index)*log(1-@cnorm(xalpha)*@cnorm(xbeta/stdev)) + index*(-0.5*log(1+theta(1)^2+alcohol^2) + log(@cnorm(xalpha)) - log(stdev) + log(@dnorm((alcihs - xbeta)/stdev)))

Henry

[EViews Gareth]

The error message usually indicates that at the starting values you provided, there are missing values in the logl function. This could be for a number of reasons, although common ones are dividing by zero (say you're dividing by a coefficient, and the starting values for coefficient is zero), log of a negative number of something...


Looking at your logl definition, it looks as though you're only maximising over one coefficient (theta(1)), and that there are a lot of additive terms that won't affect the maximisation problem. You could probably simplify the problem a great deal by taking out those additive terms.

[henry]

Thanks. I think there's a log of 0 causing it.

And I'm maximising over 82 coefficients - the others are contained in xalpha, xbeta and stdev.

Henry

[EViews Gareth]

Fair enough.


Good luck with 82 coefficients. It might be tricky to get it to converge.

[henry]

OK well it only seems to be maximising over theta. I was under the impression that since xalpha, xbeta and stdev each include coefficients, that it would maximise over these as well?

xalpha = alpha(1).a1 + alpha(2).a2 + ... + alpha(35).a35
xbeta = beta(1).a1 + beta(2).a2 + ... + beta(35).a35
stdev = gamma(1).a1 + gamma(2).a2 + ... + gamma(11).a11

In the user's guide it does the following:

@logl logl1
res = y - c(1) - c(2)*x - c(3)*z
var = c(4) * z^c(5)
logl1 = log(@dnorm(res/@sqrt(var))) - log(var)/2

Do I need to specify xalpha, xbeta and stdev in the LOGL object or something?

Henry

[EViews Gareth]

Yes, you do.

[henry]

Done, thanks.

But I've got a new error message now =):

"Warning: Singular covariance - coefficients are not unique"

What does this one mean?

Henry

[EViews Gareth]

It means that the covariance matrix is singular. Therefore the estimated coefficients may not be unique (there may be other values that will satisfy the same conditions as those that were reported).

[henry]

Is there any need to worry about it?

[henry]

Or can I do anything to make sure the coefficients are unique?