Sspace: does estimation use one-step, filtered or smoothed states from the Kalman filter?

[LaPadre]

Hi,

When estimating a sspace model, does the MLE routine use the one-step, filtered or smoothed states from the Kalman filter to evaluate the likelihood?

I am looking at this in the context of adding individual quarterly dummies into a signal equation.
When I do so, the one-step signal prediction residuals are non-zero, but the smoothed "disturbance" estimates appear to be zero.
This suggests to me that the MLE routine might be using the smoothed states.
I suspect that this is also why the filtered estimates of my states change over the dummied period (since the one-step prediction error is non-zero, the filter will update the state).
I am open to any other thoughts on knocking out information from a signal equation for a particular period.

Cheers,
-Alex

[EViews Glenn]

Likelihoods in state space models are all based on the prediction error decomposition, which uses the one-step ahead values as given in

https://www.eviews.com/help/helpintro.h ... round.html

If you want to knock out signal information for a period, just recode the dependent variable to have a missing in that period, and use the recoded value.

[LaPadre]

Many thanks.
I had read that documentation but failed to note that the tilde definitions used in the likelihood were defined above as the one-step predictions.

Also thanks for the suggestion about replacing with missing values - that would not have occurred to me but is very simple!