How can i see the model parameters

[MerveKaptan]

Dear @AuguB Stijn,

I hope you are doing well!
Sometimes, I get unreasonable z-scores (more than 10.000) within my model (even after widening my bspline limits).
I checked the raw data and it seems fine.
In order to understand what is going on I wanted to save the whole output of model (sigma and residuals).
I was hoping that I will have these values when I save the results to a .csv as I did below. However, I am not seeing sigma there.

What would you recommend?

I see all of these --> but not sigma..

X,X,X,X,Y,Y_harmonized,Z,batch_effects,batch_effects,subjects,centiles,centiles,centiles,centiles,centiles
age,height,weight,z_slice,MEAN(area),MEAN(area),MEAN(area),sex,vendor,subjects,"('MEAN(area)', 0.05)","('MEAN(area)', 0.25)","('MEAN(area)', 0.5)","('MEAN(area)', 0.75)","('MEAN(area)', 0.95)"

this is my code:

transfer_dir = os.path.join(out_dir, "transfer")
# Run transfer
model.savemodel = True
model.saveresults = True
model.saveplots = True
model.evaluate_model = True
model.save_dir = transfer_dir
transfer_pred = model.predict(transfer_data)

transfer_data.to_dataframe().to_csv(os.path.join(save_dir, f"transfer_{metric_prefix}.csv"))

# inspect/save from the returned NormData
df = transfer_pred.to_dataframe()
print(df.columns)  # this replaces: print(model.results.columns)


out_path = Path(transfer_dir, f"{metric_prefix}_transfer_full_results.csv")
df.to_csv(out_path, index=False)
print("✓ saved transfer outputs to", out_path)

Thanks so much!!!
Merve

[AuguB]

Hi Merve,

Sorry to hear that you are still having problems, I hope we can figure out what's causing this, and fix it.

So, the parameters of the HBR models are stored as MCMC chains in the idata.nc files. These can be loaded like:

import arviz as az
idata = az.from_netcdf(path)

The paths can be found in the model save dirs:

model
├── normative_model.json
├── Right-Lateral-Ventricle
│   ├── idata.nc
│   └── regression_model.json
└── WM-hypointensities
    ├── idata.nc
    └── regression_model.json

The chains can be thoroughly inspected using the tools in the Arviz library, it is really very useful.

Good luck with the forensics. Let me know if you need any further help, or if you would like to discuss during a call. Also keep me updated about any findings, it would be very nice if we can improve the toolkit based on your experience.

Stijn

EDIT:

To create a visualization of the graphical model, from which the relations between the parameters in the MCMC chains should become clear, you can do:

NormativeModel.load(path)
model.predict(test_data) # <- this ensures that the regression models have a compiled`pymc.Model` 
model.model_specific_evaluation() # <- this creates plots of the graphical models in the save dir (you may need to install some additional libraries, but that should become clear from the error message that you get if this is the case)

[MerveKaptan]

Hi Stijn,

Thanks a lot, as usual, this is super helpful! I will try and see if I can understand the model parameters!
Yes, I am slowly getting there :) Once I try a few more things systematically, I definitely would love to have a call! I will send you and email.

Thanks so much again!
Merve