Process error too large

[amael-ls]

Despite an informative narrow prior, the process error sigma_proc is too large. Its posterior average is 6 times the prior average (that is not necessarily a problem), and is around 8mm (that is the problem)! Plan to solve this (got to step i + 1 if step i fails):

1. Try observation error as function of size (maybe it will be larger), see Rüger 2011 --> currently an experimental branch growth_structure_logN_sizeDependenceError
2. Fix process error to a sensible value and check which parameter (especially error and proba) try to compensate for the 'low' (i.e., sensible) value of sigma proc
3. Less informative prior on routine observation error. Watch identifiability out!
4. Take holidays and then come back to the problem

[amael-ls]

With a less informative prior on sigmaObs, I get a sigmaProc that 'makes more sense'. However, I get an unrealistic routine observation error sigmaObs. So I am going to do two test here:

1. Same as bullet 3 in first message, but this time with a quadratic term in front of dbh
2. Prior less informative on etaObs (extreme error), and more informative on sigmaObs (though less than what it used to be!). Test for Fagus sylvatica (slow growing) and Abies grandis (fast growing)

[amael-ls]

The bug is apparently mostly because my model is not adapted to the structure of the French data. Working on the German data only greatly improved the results. That being said, the mixing of sigmaProc for the tested species (Betula pendula, Fagus sylvatica, Picea abies, Pinus sylvestris, and Quercus robur) is still not very good. I think a reparametrisation is necessary (testing on the branch reparametrisation)

[amael-ls]

For the mixing part, I had to fix the observation error sigmaObs and also do many other stuff. Today, everything seems as ok as it can be!