EViews.com

Hello,

I am attempting to apply a kalman filter (via state space object) to detrend the solow residual. To run the filter, I rely on HP-filtered series as proxies for the initialisation vector, to set the variances of the state variables, and to set the variances of some of the error terms. Because setting the variances of all error terms produces implausible results, I would like to leave the variance of E2 diffuse.

My problem is that I get different outputs depending on the sequence in which I run the filter. When I try to run the filter, setting the variances of E1 and E3, and leaving E2 diffuse, I get a message saying 'Failure to improve the log likelihood ...] However, when I first run the filter with all error variances specified (which produces convergence), and then go back and change the specification to set E2 to diffuse, convergence is achieved. Can anyone please suggest why this is the case?

Trend/cycle models using the State Space Model can sometimes (in EViews, Gauss, Matlab etc) be hard to estimate due to the fact that to can be hard to find the desired structure in the data. With specific formaulations of the state space model it can be harder och easier.

One thing that I have found useful is to include the square instead of the exponential function in the estimation of the variances. This, of course, opens up for a multimodal likelihood, but these multimodalities should be predictable and observationally equivalent.

Hence, you could try out

var=c(10)^2 instead of var=exp(c(10))

to see what happens.

/K

Thanks for the suggestion