Plot comparisons notebook for docs

[GStechschulte]

This PR adds a demo notebook for the Bambi docs explaining the following:

• what the methodology (average predictive comparisons) behind comparisons and plot_comparisons is
• how to use comparisons and plot_comparisons and the various quantities of interest: unit level, grid, average

Ultimately, this docs example builds off of this blog post, but will go into more detail along with more examples.

This is a draft PR and can only be merged once #684 is merged.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[codecov-commenter]

Codecov Report

Merging #695 (6f269e0) into main (0fa0b6f) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##             main     #695   +/-   ##
=======================================
  Coverage   87.91%   87.91%           
=======================================
  Files          43       43           
  Lines        3318     3318           
=======================================
  Hits         2917     2917           
  Misses        401      401           

Impacted Files	Coverage Δ
bambi/plots/plotting.py	88.77% <ø> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[tomicapretto]

Reads soooo well! Amazing work!!!

[GStechschulte]

Reads soooo well! Amazing work!!!

Thank you! I'm excited to get user feedback and tests.

[tomicapretto]

Just left a very small comment. It looks so good. If you want, feel free to add a note at the end saying this was authored by you as part of your GSoC project.

I only have two questions (for now)

• Should the results of comparisons contain the computed operation? diff, ratio, etc.
• In the "Default contrast and conditional values" it says "Bambi computes the mean and then subtracts and adds 2.5 units..." Does it work the same way in marginaleffects? I'm thinking about how well it can work with different scales of data.

edit I think this is ready to be merged

[aloctavodia]

Line #6.    from bambi.plots import comparisons, plot_comparison

We should be able to import bambi once and then call the plots submodel directly, like

import bambi as bmb

bmb.comparisons or bmb.plot_comparison

Reply via ReviewNB

[aloctavodia]

It is possible to specify 1 to 2 and 1 to 4 comparisons by writing something like contrast={"persons": [1, (2, 4)]} or with some other syntax?

Reply via ReviewNB

[GStechschulte]

Yes, this can be achieved by passing

contrast={"persons": [1, 2, 4]}

and since we use list(itertools.combinations(contrast.values, 2)) where contrast.values = [1, 2, 4], the contrasts end up being a pairwise combination:

(1, 2), (1, 4), (2, 4)

[aloctavodia]

That will generate all combinations, I am asking if it is possible (or even desireable) to write something like this (not necessarily this) contrast={"persons": [1, (2, 4)]} and get [(1,2), (1,4)] .

[GStechschulte]

Just left a very small comment. It looks so good. If you want, feel free to add a note at the end saying this was authored by you as part of your GSoC project.

I only have two questions (for now)

• Should the results of comparisons contain the computed operation? diff, ratio, etc.
• In the "Default contrast and conditional values" it says "Bambi computes the mean and then subtracts and adds 2.5 units..." Does it work the same way in marginaleffects? I'm thinking about how well it can work with different scales of data.

edit I think this is ready to be merged

Thank you! :)

Should the results of comparisons contain the computed operation? diff, ratio, etc.

Currently it does because: (1) if the user wants the estimates, etc. without plotting then this function offers that possibility, and (2) if the user wants to perform their own plotting (for whatever reason) then they have access to the same data we use for plotting. What are your thoughts?

In the "Default contrast and conditional values" it says "Bambi computes the mean and then subtracts and adds 2.5 units..." Does it work the same way in marginaleffects? I'm thinking about how well it can work with different scales of data.

Very good point. By default, I use a one unit centered difference, i.e., $\pm$ $0.5$. For example, if the user does not pass any value for contrast="persons", then $\pm$ $0.5$ is applied to the mean of persons. I added additional wording to clarify the distinction between default centered differences and centered differences when a user passes their own value.

[aloctavodia]

LGTM

[GStechschulte]

LGTM

Thanks for the review! Much appreciated!

[tomicapretto]

@GStechschulte now we have some conflicts. I can merge once they are fixed :)

[GStechschulte]

@GStechschulte now we have some conflicts. I can merge once they are fixed :)

Good to go now :) Thanks!