Time series of observations $y_t$. The proposed model is that there's unobservable scalar series $x_t$: $$x_t=\phi x_{t-1}+B_tu_t+w_t$$ where $u_t$ - vector predictirs and $w_t$ - noise.

Then there's observable scalar: $$y_t=Z_t\times(x_t\beta_1+(1-x_t)\beta_2)+v_t$$ where $Z_t$ - another vector of regressors, potentially, could be the same as $u_t$ if it helps; $v_t$ - noise, and $\beta_1,\beta_2$ - parameter vectors I'm trying to estimate.

Is this model identifiable? Can estimate betas and the unobservable $x_t$?


It could be depending on what your state regressors are ($u_t$). Consider the flipped around latent process $\tilde{x}_t = 1-x_t$ whose dynamics can be found using your first equation: $$ \tilde{x}_t= (1-\phi) + \phi \tilde{x}_{t-1} + \tilde{B}u_t + \tilde{w}_t. $$ where $\tilde{B}_t = -B_t$ and $\tilde{w}_t = -w_t$.

The associated observation equation is $$ y_t=Z_t\times((1-\tilde{x}_t)\beta_1+ \tilde{x}_t\beta_2 )+v_t. $$

If you forced the state process to be zero mean, and not have an intercept, then this flipped process wouldn't yield a state space model of the same form, so you would be okay.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.