DO.mod in predict_DO for Bayesian model

[Tzu-Yao]

Hi Alison and Bob,

First of all thank you so much for answering my questions in #390. Your insights are really helpful!
Here I'd like to raise another question about DO.mod in predict_DO. As I was trying to calculate the metric for goodness of fit using DO.mod and DO.obs, as described in #367, I'm not sure if this DO.mod from predict_DO has been corrected by process error already (I want to know what kinds of error the residuals contain).
Later I followed #222 and produced two plots, as below. The blue dashed line is DO.mod from mcmc results and the brown line is DO.mod calculated by calc_dDOdt. For the observation model ("b_Kb_oi_tr_plrckm.stan"), the two lines match, but for the state-space model ("b_Kb_oipi_tr_plrckm.stan"), they don't. So I was wondering if this difference is caused by the correction of process error in mcmc calculation. In other words, DO.mod in predict_DO (brown line) does not account for process error?

The observation error Bayesian model

The state-space Bayesian model

Thank you,
Tzu-Yao

[aappling-usgs]

Your interpretation of the software behavior is correct - predict_DO returns predictions for DO that are not corrected for process error, which is why the observation-error-only models show better matches between the observations and the output of that function.

I've had open issues for a couple of years to add in process error corrections within predict_DO (#294, #330). Given that I am no longer actively developing the package, they're not likely to be resolved soon. Fortunately, you can already get the process-error-corrected predictions by requesting that they be tracked in Stan (they're called DO_mod):

revise(specs, params_out=c(params_out, 'DO_mod'))

Then you just need to pull these DO_mod values out of the instantaneous data.frame within the model fit.

[Tzu-Yao]

Thank you so much Alison!