Kalman Filter and Missing Values

[scunthorpe]

Hi, I was wondering if there is any reference material on how Eviews handles missing values in the signal equation.

I am estimating common factors on the basis of an unbalanced panel (as in the following paper: http://rbnz.govt.nz/research/discusspapers/dp07_13.pdf). In the paper the missing values are handles by giving the idiosyncratic variance of a variable with missing values an infinte value. So the filter puts no weight on the missing value variables when computing the common factor.

I've played around with estimating the common factors with missing values in the signal equations and it calculates, just wondering how it does it.

Thanks

[EViews Glenn]

Missing values in the signal equation are treated as, well just that, missing. The Kalman filter algorithm is specifically designed to handle missing values in the signal equation. in essence, you run the filter an additional period, but no additional signal information is used in obtaining the posteriors. Durbin and Koopman's describe the updates in this case in some detail. I am certain that Harvey also has a discussion.

I haven't thought about these models in some time, but my instinct is that the infinite variance approach is a computational trick to get the desired behavior in Kalman filter implementations that don't directly handling missings. The only concern I would have with the trick is that unless you are careful it might leak into the estimate of the idiosyncratic variance. My recollection is that algorithmically we set the signal residual and Kalman gain to zero; this probably works out to be the same as the trick...but don't hold me to either of these statements :)

But to answer your question. EViews uses the standard filtering mechanism for handling missings. Any of the cited references in our chapter should indicate how this is handled.

[scunthorpe]

Thanks for your help Glenn