Bayesian linear regression for non-stationary time series?
Greetings everyone! Im total newbie and here is my questions:
1. Is it appropriate to use JASP software to conduct Bayesian Linear Regression for non-stationary(can't be transformed) small sized (N=21) time series data? The purpose of such estimation is to see possible impact of certain exogenous variable on the dependent one. Furthermore, I have included three additional variables (also potentially related to the dependent variable) for robustness.
2. If Yes, how should i choose priors and model priors in the setup section? ( For now, im leaning towards AIC with Beta binominal a 1 b 1, as i've found info that AIC helps to select a model that has enough complexity to capture the data patterns in non-stationary time series and Beta binominal considering the size of the model, assuming the whole combination of variables more robust can explain the dependent one )
Thanks in advance!
Comments
I'm not an expert on the topic, but it seems to me that if you have non-stationary time series data, you need to use a time series model as your base --perhaps a state-space model?-- and then model the exogeneous factors on top of that...JASP does have state-space models but I don't think we have the option to include covariates yet.
EJ
Thank You very much for the answer! What if I use both models? Is it possible to check them in JASP to see if there are any potential violations or issues? I mean, besides plots and results analyse (like GOF for BSSM )
What do you mean with "both models"? Usually we offer the opportunity to check for violations in all analyses we have -- if we don't you can suggest that we add this by posting on our GitHub page (for details see https://jasp-stats.org/2018/03/29/request-feature-report-bug-jasp/)
Cheers,
E.J.
Dear Professor Wagenmakers, many thanks for another response!
When I meant by two models, I intended to use BSSM first, and then BLR. Maybe add a third one, like Bayesian Lasso Regression from other software. This combo is supposed to validate each other's results, because relying only on BSSM, I can't get appropriate results (due to data issues, there will be wide CI including zero, low mean output, etc.). So, I assume that if all 2 (or 3) models will show similar results, even with issues, it can be an extra proof for my research question or at least an argument not to reject the hypothesis. Yet, I'm still questioning this approach, especially using BLR with time series. (I've found many opinions for it and against it)
Warm regards,