The RATS Software Forum

dear
I run the textexample koop_bp200.rpf
I find the resluts esitmated by the maximum likelihood and Gibbs sampler are so different in the aspect of of space state curves (a1 in the program)?
I think two methods should not get such a big different results
May you tell me the reasons:?:
thank you very much

iloverats wrote:dear
I run the textexample koop_bp200.rpf
I find the resluts esitmated by the maximum likelihood and Gibbs sampler are so different in the aspect of of space state curves (a1 in the program)?
I think two methods should not get such a big different results
May you tell me the reasons:?:
thank you very much

There's quite a lengthy discussion in the comments near the end of that about how the prior isn't really appropriate. Even though the priors are quite loose (1 degree of freedom), their scales are so far out of line with the appropriate values that they force way too much time variation.

thank you

Are nulambda and s2lambda the shape parameter and the scale parameter respectively

or What?

Shape and scale on the variance of the coefficient drift.

in this model
i think we hard ,even might not to get hyperparameter lambadas' nulambda and s2lambda prior information
can s2lambda = 1.0/the number of observations be considered as general rule for such case?

how can we choose the proper prior information for such case?

iloverats wrote:in this model
i think we hard ,even might not to get hyperparameter lambadas' nulambda and s2lambda prior information
can s2lambda = 1.0/the number of observations be considered as general rule for such case?

how can we choose the proper prior information for such case?

There's a fairly lengthy discussion of TVP models at:

http://www.estima.com/forum/viewtopic.php?f=5&t=381

The value for s2lambda is appropriate for an autoregressive parameter, which has a natural scale on the order of 1 because it's scale-independent. It's not as clear with other regression coefficients.

dear

how can i set the unconditional prior information to do bayesian gibbs estimation in such case?
thank you

The value for s2lambda is appropriate for an autoregressive parameter, which has a natural scale on the order of 1 because it's scale-independent. It's not as clear with other regression coefficients.[/quote]
Dear
which textbook or papers i should refer to ?
i need more details
thank you very much

iloverats wrote:how can i set the unconditional prior information to do bayesian gibbs estimation in such case?
thank you

That's a question for which I have no really good answer. There are some coefficients (autoregressive coefficients, stock beta's) which have a natural scale. Many others don't. Using the scaling information from a full-sample linear regression will generally be the best course when you have nothing else. This is similar to what's done with time-varying parameter VAR's, where the covariance matrix of the increments is a fraction (like .01) times a prior based upon empirical estimates.

dear
what is the scaling information?

iloverats wrote:dear
what is the scaling information?

The prior for a variance is an inverse gamma with degrees of freedom and a scale. It's the scale that often isn't obvious from the context, and if it's way off (particularly on the high side) you can end up with too much time-variation.