Hi,

This is a problem I am currently facing in my research...
I have a dataset collected over time for time=0 to T
I need to estimate some parameters (regarding this data) using Bayesian Inference. I use data for each time instance for Bayesian inference with prior as the posterior obtained in the previous time instance. The posterior obtained in this time instant is used as prior for the next time instance and so on...
However I do not have any priors available to start with. I wanted to know if the following methodology is correct...

(1) Divide the dataset into two parts say from time=0 to t and from time =t to T
(2) Use the dataset from t=0 to t to determine prior probabilities for my parameters.
(3) Use dataset from time= t to T for Bayesian inference with priors obtained in previous step.

Let me preface my statements by saying that I'm not a Bayesian statistician, but Bayes' methods are useful in some situations....

=========

That's the thing about priors - they're pretty subjective.

Your method sounds fine to me - I would try a variety of time periods (0 to t --> change what "t" is) to see what impact each of them have on the Bayesian inference.

This is often done with regular multiple regression to check the robustness of the model. The sample is randomly split in half - one half is used to generate the model or equation parameters, then it is checked for "goodness-of-fit" with the other half.

Hi,

Thanks a lot for your reply....

I'll try out what you told me. Let you know if I have any probems.

regards,
abcd