Interpret.plot_predictions with transformed variables

[SanBertero]

Hey there!

I had a question regarding the use of bmb.intrepret when the model includes transformations (scale(X), log(X) or both).

By default Bambi seems to plot the covariable back into its original scale/values, while the prediction in its transformed scale. Is there a way to control this either way? Say, keep Y(X) in its original scale, or plot it as a function of the transformed X? The latter is manageable by adding a new column in my dataframe, but the former isn't so direct.

More importantly, I had a case come up where I wanted to normalize log(x). scale(log(x)) seems to work for the inference model, but then bmb.interpret.plot_predictions fails/returns an error. Is it something related to my input or a limitation of the function?

Attached is a short Jupyter notebook with a quick example if it helps.
ExampleInterpret.zip

Thank you!

[GStechschulte]

Hey @SanBertero thanks for the question and notebook.

Is there a way to control this either way? Say, keep Y(X) in its original scale, or plot it as a function of the transformed X?

To keep Y on its original scale, you could pass an inverse transform function to the transforms parameter, e.g.

# model_5.formula = Formula('scale(Y) ~ scale(X)')

y_mean = np.mean(Y)
y_std = np.std(Y)

inv_transform = lambda x: x*y_std + y_mean

fig, ax = bmb.interpret.plot_predictions(
    model_5,results_5,
    'X',
    pps=True,
    prob=0.9, 
    transforms={"scale(Y)": inv_transform}
)

I had a case come up where I wanted to normalize log(x). scale(log(x)) seems to work for the inference model, but then bmb.interpret.plot_predictions fails/returns an error

What version of Bambi are you using? We pushed changes to address this bug, but I am not sure it is in the latest release.

[SanBertero]

Thank you, I now got farther out into the example.

I notice that now the model builds a deterministic 'mu' compared to the latest release, which I'll figure out how it affects things/examples later...

However, and sorry to keep bothering, now bmb.interpret.plot_predictions fails. It seems that the variable pps=True has been changed to something else. I can make it work if I remove or set it to False to see the mean deviation. The code provides a warning saying it's now been moved to "response", but it doesn't seem to be a valid keyword argument.

[GStechschulte]

I notice that now the model builds a deterministic 'mu' compared to the latest release

That's correct. This was introduced in #804. It shouldn't affect your results in a negative way. We wrap all the parameters of the likelihood into a deterministic node to make model graphs easier to read.

It seems that the variable pps=True has been changed to something else. I can make it work if I remove or set it to False to see the mean deviation. The code provides a warning saying it's now been moved to "response", but it doesn't seem to be a valid keyword argument.

The changing of the parameter names is expected due to the commits in the merged PR linked above. This is weird because I am running your example on the import-error branch with pps=True and I don't get an error. Mmmm.

Perhaps it is an xarray version problem. In my environment I have

(bambi_main) ➜  ~ pip list | grep xarray
xarray              2024.6.0
xarray-einstats     0.7.0

[GStechschulte]

By the way, thanks for reaching out with questions and problems. The communication and feedback are much appreciated.

[SanBertero]

Hmm... I don't know. Here's the error I get when running predict on my own (I switch kind="pps" to "response" to avoid the warning, but everything else is the same), if it helps.

Ahh, let me check that about xarray... I'm on 2023.7.0...

[SanBertero]

That's right! I installed a newer version of xarray from the github (2024.6.1.dev32+ga86c3ff4) and now everything worked out! Thank you so much for the help. Seems to be a dependency thing