Differences between 4.1.2 and 4.3.1

Dear all,
I would like to know why when I run my model with 4.1.2 dynare’s version, all is fine, and when I run it with 4.3.1, it says a Blanchard and Kahn conditions are not verified ?

I have missing observations in my samples but it seems not to be the problem with 4.3.1 (I have tested without missing observations and it does the same error).

Thank you in advance for your answers and tell me which version is the best.

Jonathan

Hello to all,
I have just finished two estimations of the same mod file (Iacoviello Neri 2011, 3 chains of 15000 iterations each) and the same data set, one using 4.1.2 and the other using 4.3.1.

The results are completely differents… so my question is simple: which version is the best ? and why of course…

Thank you a lot for your answers.

Jonathan

Most likely, both are wrong. You are talking about a Monte Carlo Method, which is always stochastic. We only know that in infinite time, we will converge to the true distribution. In practical applications like yours, diverging results usually means at least one of the runs has not found the correct mode around which to center the MCMC. What do convergence diagnostics says? Did you try an extensive run of mode_compute=6?

Thank you for your answer Johannes.

I use mode compute = 4, and the convergence diagnostics are bad for both estimations (red and blue line converge but not by an horizontal way). Moreover, when I am running twice the same estimation of the same mode file with the same data, I obtain the same results (maybe because there is sufficiently iterations, 15000). I will redo this exercise with mode compute = 6.

So my first question becomes important now:

when I run my model with 4.1.2 dynare’s version, all is fine, and when I run it with 4.3.1, it says Blanchard and Kahn conditions are not verified. Why ? Which version can be trusted ? (this error appears before Monte Carlo simulations, the rank condition is verified for the two versions of dynare).

Thank you in davance for all of your answers

UPDATE

I have found an unstable process in my model (x=0.999x(-1)+shock) so when I have modified it, my model runs also on Dynare 4.3.1

Why Dynare 4.1.2 is able to run my mod file with this (unstable) process whereas Dynare 4.3.1 not ?

???

Could you send me the mod-file and the data-file?

I am sorry, I’m sure you understand that I cannot do that.

But I only need to know if the computation of the solver or other techniques used to solve and estimate a model has changed between these two versions. If it is the case (i feel it is), which version is the best and why.

If you want to know the differences between the versions, you can look at the following pages:

dynare.org/DynareWiki/KnownBugs
dynare.org/DynareWiki/NewFeatures

Or you can even directly look at the source code.

Since we (hopefully) fix more bugs than we introduce over time, I strongly encourage you to use the latest version (4.3.1).

Thank you Sebastien