When I use Bayesian approach to estimate the parameters in the model, I choose, for example, 6 observation series as there are 6 shocks and get results which seem ideal. However, when I change one of the series to a new one (e.g. I substitute output series for interest rate series), the results still seem acceptable while the values of estimated parameters are a little different and the conditional variance decomposition of variables is quite different.
Is this normal when doing bayesian estimation? Does this mean the results are not robust ? Or I just need to compare the model implied moments with data moments and choose the more fitted results ?
Thank you for your advice. I read this paper and find that the author seems not reconmending the omission of variable like interest rate. I don’t know if this conclusion is pervasive to all kinds of dsge models. I tried my model and find that the model implied moments fit much better to the empirical data moments when interest rate is not included in the observation series.
In this case, is it OK for me to accept the bayesian estimation results without interest rate in observation series ? Or I must use interest rate as observation variable and retry the estimation to get fitted moments?
Note that the goal of full information estimation is not matching particular moments that you are interested in, but rather matching all moments.
But if your estimation without the interest rate looks good, that’s the way to go.