Paging @jnothman and @RianneSchouten.

It might be good to amend that first commit with --author 'Rianne Schouten' etc. Thanks for this! Will look soon!

OK, I think that worked.

I suspect that without examples/impute/README.txt the example won't render.

I've created examples/impute/README.txt in the iterativeimputer branch.

Oh, no. Did I break the doctest again?

@jnothman I wondered back to this PR, and now it passes tests!

Any interest in picking up work on this? I feel like it was in a pretty good place already and we were just unsure about the extremely long runtimes.

@jnothman @glemaitre Any thoughts on my last comment? Repeated here: "Any interest in picking up work on this? I feel like it was in a pretty good place already and we were just unsure about the extremely long runtimes."

I found the current docs ambiguous and believe the community would value this work.

I'm not familiar with these build trigger checks. Can anyone please suggest how to fix it?

Can anyone please suggest how to fix it?

Syncing with the main branch should fix the build trigger error.

Thank you @thomasjpfan! Any idea if we can get this merged once all the tests pass?

Any idea if we can get this merged once all the tests pass?

This PR still need two approvals to get merged and I do not have a time estimate for that to happen.

At a glance, I see two big tasks for this PR:

1. It needs to use another dataset instead of load_boston because load_boston has been deprecated and will be removed in 1.2
2. In the recent years, we have been moving to more narrative driven examples, where text to place together with code. For example: Common pitfalls in the interpretation of coefficients of linear models. This means moving text from the big paragraph in the beginning and placing them into the text to create a narrative.

Thanks! I can make those changes. What regression dataset do you recommend to replace Boston? Smaller is best because this example is hefty, but I can always subsample any dataset.

Any idea if we can get this merged once all the tests pass?

I would like to see how this example integrates within the proposal in #21967.

@glemaitre MICE is in the family of multiple imputation - perform imputation multiple times, then apply your subsequent pipeline multiple times also, and then have multiple solutions. For sklearn users the subsequent pipeline will often be "train/val/test a ML alg". I read through #21967 and multiple imputation isn't mentioned at all, but it is common in the stats world as far as I understand. This example would be useful because people coming from stats might want to do what the mice R package does, but don't know how with IterativeImputer as it does not do it out of the box.

To summarize:

• Documenting missing-values practices #21967 is about how to make a guide to imputation, (mostly) for the purpose of ML.
• This example is about multiple imputation.
• The two don't have an clear bridge at the moment.

I agree that having an example defining what is "multiple imputations" is important to remove the confusion with the iterative procedure of the IterativeImputer and that there is no confusion. In this regard, I find the example too complex.

In this regard, I would prefer to have a single pipeline to make single imputation and then create a specific estimator to show how to make multiple imputations. We would not even need to use an IterativeImputer in this case. This would speed up the example that is currently taking up to 3 minutes while we usually try to have examples running under 30 seconds.

I think that this is super important to point out in the discussion that the example stands at providing a definition of "multiple imputations with code" rather than showing that multiple imputations work better. I am not sure that currently in ML setting there is any evidence that multiple imputations are working better than using a strong learner (@GaelVaroquaux and @marineLM have better insights than me on this).

Cc @A-pl (we need to put your paper on HAL)

I'm a bit confused. We do have a single pipeline in example 2: https://github.com/sergeyf/scikit-learn/blob/iterativeimputer_mice_example/examples/impute/plot_multiple_imputation.py#L303

And it's used multiple times to do MICE: https://github.com/sergeyf/scikit-learn/blob/iterativeimputer_mice_example/examples/impute/plot_multiple_imputation.py#L315

Can you please clarify what you'd like changed?