PRESAMPLE .. Bayesian. ODDS ratio

Hi all,

I am estimating a dsge by setting different values in the PRESAMPLE option of ESTIMATION command .

My question is:
Can I compare the estimated models with different PRESAMPLE=values (everything else the same) ?
What is the role of this PRESAMPLE. does it imply that the higher PRESAMPLE=value the shorter the data sample ? (then LAPLACE Approx values can not be compared for different PRESAMPLE values ) ??

it is not clear in the documentation the implications of different values of PRESAMPLE-option.

The manual for the unstable version states for presample:

[quote]The number of observations after [first_obs], page 60 to be skipped before evaluating
the likelihood. These presample observations do not enter the likelihood, but are
used as a training sample for starting the Kalman filter iterations.[/quote]

That means these observations are effectively dropped from the sample. Hence, different presample values when when using the same dataset with the same nobs and first_obs implies that the sample relevant for the likelihood, the posterior density, and marginal data density will be different and computing odds-ratios is not feasible. However, I am not sure when such a case would ever arise.

Thanks Professor,

It’s cleatr that the ODDS ratio can not be used as the data for Likelihood is different (for different PRESAMPLE values)
Thyen it’s just unclear what is the right number to set in the PRESAMPLE ?

do we just try different values and pick the one we like the results ? (ODDs ratio is of no help obviously)

thanks

You simply use the longest available sample while leaving the effective sample used for likelihood evaluation the same. Say you want to compare the model fit for a country between 1960 and 2015. If you have good data before 1960, say from 1950 onwards that you can use for initializing the Kalman filter, you want to use it.

Now it might be that you are not using that initial 10 years of data for estimation, because you think there was a structural break in 1960. In that case, you might want to work without a presample, because there is effectively no data available before 1960. That is the reason most people work without a presample, because they argue that they are already using the longest available sample for estimation so that there is no presample they could use.

Thanks Proffessor,

In the and it seems one can make his own judgement, while the least problematic is to work wihtout presample.

Ps. and sorry for the late reply