a question on significant of hyperparamter in DLM

[hardmann]

Dear Tom:

In page 158, rats programing manual 2.0, when checking the significant of coeff in Linreg, it use codes as follows

compute signif=%ttest(%tstats(%nreg),%ndf)

we want to know how to compute the significant of hyperparameters in DLM instruct, for example, sigxi, sigzeta.
In other word, how to get the significant of hyperparameters from DLM defined varibles.

In reference manual, dlm does not define %ndf variable.

Best regard
Hardmann

[TomDoan]

In general, those should be done using %ztest rather than %ttest since most estimates have distributions which are asymptotically Normal. The %ttest has somewhat fatter tails, and is used, when appropriate, for a more conservative distribution, but isn't strictly appropriate except in a relatively small number of cases. At any rate, the component variances don't satisfy all the conditions even for asymptotically Normality as they have to be non-negative.

[hardmann]

Dear Tom:

When we estimated complicated DLM model by using Extended kalman smoother with iteration, a couple of hyperparameters maybe wrongly initialize, the model will not converge and some hyperparameter are no significent. We plan to use grid search to find out global maxamium. So, we should check the significant of hyperparameters.
Please Tom help us.

Best Regard
Hardmann

[TomDoan]

Estimate with and without a peg to zero and see whether it makes any real (economic) difference. You have been trying to estimate models which have way too many components to be estimated from data.