Missing Values in the Signal Eq of the State Space Model

[dhanssen]

I just wanted to double check -- within the state space model (sspace), Eviews allows for missing values in the state equation, but not in the signal equation, is this correct?

[dhanssen]

It actually seems that it doesn't allow for missing values in either the state or signal equation -- which means that I must interpolate all values since I'm running it with panel data so if I delete one date observation I must do it for all cross sections. Is this the only method of making this work? Any advice is much appreciated!

[EViews Glenn]

Not true, EViews should handle missings in the signal equation.

[dhanssen]

Thank you very much for your help! I was hoping this is the case, but am wondering why I get the following error when I run the state space without filling in missing values:

" Procedure does not allow breaks in sample (after excluding observations with state equation NAs"

I first tried to fill in the missing values for just the variables in the in the signal equation, and still got this error, and now have filled in missing values for variables in both the state and signal and then it will begin solving.

Thank you once more for your assistance!

[EViews Glenn]

It's the missings in the explanatory variables in the state equation that are causing problems.

[dhanssen]

So are you saying that Eviews will not allow for missing values for variables in my state equation then? For example, if I have:

y = c(1) + sv1 + c(5)*x + c(6)* z

@state sv1 = c(2)*sv1(-1) + c(3)*x+ c(4)*z+ [var=1]

Then if I have missing values in my x or z's, I will get an error, but it is fine to have missing values for my y?

[EViews Glenn]

Yes.
The state equations are the unobserved components in your model. We have to compute estimates of them somehow, and we can't if there are missings in their observed components. If it's a purely unobserved component, we can do the updating.

The Kalman Filter is designed for unobserveds in the signals, but we can't do anything if we can't update the state estimates.

[dhanssen]

Great -- thanks so much for your help.

[dhanssen]

One final follow-up question here. Is there a way to possibly tell the state to stay "as is" for observations where it encounters missing values ( i.e. some setting or coding that will accomplish this)? I am trying to avoid having to fill in these missing with an estimated value.

[startz]

One final follow-up question here. Is there a way to possibly tell the state to stay "as is" for observations where it encounters missing values ( i.e. some setting or coding that will accomplish this)? I am trying to avoid having to fill in these missing with an estimated value.

Possibly...

Code: Select all

x = @nan(x,x(-1))

(so long as the first observation isn't missing.

[dhanssen]

OKay, so I will just fill in the missing values with the most recent existing value. This is better than estimating it, so I'll use it. Thanks!

[dhanssen]

I lied about the final follow - up, sorry. One more question -- I've filled in all the values now, per the suggestion above, and am getting the same error message. I know longer have any missing values, so is there any reason this could happen? I have a lot of zero values and times when the observations don't change too much, that's all I can think of. It's driving me crazy!

[startz]

You might want to post your workfile. (And, of course, be sure your copy of EViews is uptodate.)

[dhanssen]

Thank you! The workfile just contains an example of what is happening, the specification doesn't make sense right now. It is a panel data set. Please let me know if you need more information.

[startz]

Thank you! The workfile just contains an example of what is happening, the specification doesn't make sense right now. It is a panel data set. Please let me know if you need more information.

The exogenous variables in the state equation seem to still have missing values.