blog/2018-04-05-gamm-random-effects/

[giscus[bot]]

blog/2018-04-05-gamm-random-effects/

https://drmowinckels.io/blog/2018-04-05-gamm-random-effects/

[afustamo]

Hi! very interesting thank you! Currently, I am working on GAMM compared to Random Forest and HDGM in a spatiotemporal framework. I am wondering to know how to predict value from the GAMM model which I fit to compare RMSE to the other two models. In particular, I did not add any random effects explicitly but a corAR1. When I check the autocorrelation of residuals from the model$gam I see that it is not disapperead and not even decreased. I think I am predicting without considering the autocorrelation and indeed the performance are very bad compared to the other two models. So my question is: how can I extract predicted values not just from the GAM part but considering also the correlation structure?

Thank you so much in advance!!
I am quite stack :(

Ale

[drmowinckels]

Hi Ale!

I'm not familiar with working with corAR1, and I am not sure its doing what you want it to if you are not seeing the autocorrelation at least decreased. It might be that gamm is not doing what you are intending, and completely ignoring your matrix. Have you tried using gam in stead, which I know supports this. basically, I'm unsure whether gamm does support this way of working.

[afustamo]

Thank you for you reply! By the way I have found the solution (S Wood helped me): the GAMM object has two objects inside: GAM and LME. Basically, predict with GAM part and calculate residuals (observed - predicted). Then, multiply the residuals by the "phi" coefficienti (mine was 0.66) estimated in the LME part. Finally, add to the predicted value at time t the "modified" residual of t-1.

Now they don't show autocorrelation anymore and performances have increased.

Guessing will help someone!

Thank you again!
Ale