Using the State Space Framework on Panel Data

[dhanssen]

Hello,

I using the sspace object on my workfile which has a balance panel structure. I want to use a fixed effects model for the cross sections. As far as I understand, this must be done "manually" using the sspace object, so I have done this both by creating a set of dummies for the cross sections and also (for comparison purposes) by standardizing the variables. However, when I look at the estimates for for the state space variable i create (called sv1) it doesn't appear that the state space/Kalman filter is recognizing that this is a panel data set. If this was the case, the "starting" value that I have for my sv1 would be the same for each cross section, correct? When I look at the state variable forecasts, it appears the estimates continue from one cross section to the other.

As such, my question is the following. Does the sspace object actually work with panel data? If so, how do I get it to recognize that I am working with it?

Thank you very much for your help!

Heidi

[EViews Gareth]

Top of the head recollection is that it re-starts the filter at the start of each cross-section.

Could you post your example?

[dhanssen]

Thank you for your help! I have posted an example of what I am working with. Unless I am doing something wrong (very possible) it appears through my state variable forecast that the estimates are continuing across cross sections (meaning it is not recognizing the panel structure), since if they weren't the state variable should start at zero in each case. The example I have posted controls for fixed effects by standardizing the variables (by cross section), as this allows for a much smaller model that will run quicker. A model I created that controls for FE by using dummy variables also has the same issues.

Thanks again!

[EViews Glenn]

You are correct that the sspace object isn't using the panel structure. Panel structured workfiles were introduced after the sspace and it appears that we didn't add the requisite checks for the structure to the existing routines.

We probably shouldn't be allowing it to run at all. I'm not certain that there's an easy way to do what you want, but I'll give it some thought.

[dhanssen]

Thank you very much for clearing this up. I will limit my work to simpler data sets without the panel structure. In the meantime, could you please tell me or refer me to a reference which will clear up the difference between the options under the Proc-->Make State Series menu? There are 1) Predicted States 2)Filtered State Estimates 3)Smoothed State Estimates, the latter two seem very similar, but the first is very different, and I am unsure how to reconcile what they are doing. Thanks!

[EViews Glenn]

Predicted and filtered are described on p. 385 of the manual. Smoothing on p. 385-386.

The Cliff Notes version is that the predicted are the one-step ahead estimates, obtained using information up through t-1 to predict the value in each period t. The filtered use information through t to estimate the value in each period t. The smoothing use the entire sequence of data through the end of the smoothing sample, say T, to estimate the value in each period t. The idea behind the latter is that subsequent information can help you refine your estimates for period t.

I hope that this answers your question.

[dhanssen]

Yes it does, thank you very much. One final question -- would one expect the smoothed/filtered estimates to begin at the initial value specified with @mprior statement (or at 0 if not specified)? It seems that this happens for the step ahead predicted states, but not for the other two, and I am trying to understand why it would be different. Thanks again.

[EViews Glenn]

No, since they estimates are updated using the contemporaneous data for the former, and all of the data for the latter.