State space estimation - one state series captures another one

[Max]

Hello,

I am familiar with state space models, but I recently came across a problem. My model converges and gives sensible parameters, but it seems that one state estimate is only driven by another one. To explain, I aim at estimating a common trend from a few countries time-series ( lets say inflation). If you look at the smoothed state estimates below, you see that the global factor is (nearly) only explained by one country : Italy.

forum.jpg (60.08 KiB) Viewed 3929 times

Italian idiosyncratic term has nearly no variance anymore, it feels like all variance in the Italy timeseries has been passed on to the common trend, which is odd since I used the same specification for the 3 countries, as you can see in the following code:

Code: Select all

@signal ca= c(3) +c(4)*sv1 + err_ca @signal fr= c(7) + c(8)*sv1 +err_fr @signal it= c(11) + c(12)*sv1 +err_it @state sv1= log(c(19))*sv1(-1) + [var=1] @state err_ca= log(c(22))*err_ca(-1) + [var = exp(c(23))] @state err_fr= log(c(26))*err_fr(-1) + [var = exp(c(27))] @state err_it= log(c(30))*err_it(-1) + [var = exp(c(31))]

Starting values have been estimated by OLS, I use a tolerance of 1e-09, with Marquardt ML method.

Does anyone had to deal with that kind of issue?

Thanks a lot in adavnce

[Max]

Any clues?