Bayesian Estimation and Blanchard-Kahn conditions

Dear all

I am estimating a standard RE model


x(1) = a11x + a12y(-1) + a13eps;
y = a21
x + a22y(-1) + a23eta;


I specify priors in the for of (for example)


where a110, as I understand, is used to compute initial likelihood. Suppose the Blanchard-Kahn conditions are satisfied for a110 (and other initial values). Does this imply that posterior distributions are such that the Blanchard-Kahn conditions are satisfied? When the dynare takes a random draw of aij coefficients and finds that BK conditions are NOT satisfied, what does it do? Since it never crashes I suspect it assumes the likelihood is minus (or plus) large number and ignores this combination. Am I right?


Theoretically it should, however more than once I have left a series of optimizers running overnight, just to find out in the morning that the last mode candidate saved by the routine did not satisfy the Blanchard-Kahn conditions.

I am wondering if this is a rounding problem of coefficients that live close to an indeterminacy region and are stored with a lower degree of precision after the last iteration of the optimizer. Is there any way to overcome it?

I would be grateful if anyone had any thoughts about this matter. Thanks,