Add chapter 11 of Rethinking 2

[AlexAndorra]

Ready for review.
There are two main points of attention:

1. The scientific model of the number of tools for the Pacific Islands (page 356 in book, cell Code 11.49 in notebook) runs smoothly but makes terrible predictions, especially for high-contact islands. I suspect it's because the b coefficient for high contact societies is unidentifiable at high population ranges (because there are no samples), but I tried a bunch of different Exponential and HalfNormal priors for this parameter and nothing seems to work. The weird thing is that I think I used the same parametrization as the book.

2. The first multinomial model behaves weirdly and doesn't replicate the book's results -- possible bug? I added more comments in the notebook.

Thanks a lot for the reviews ✌️ I'll re-run the whole NB once we've fixed all the issues.

[review-notebook-app]

Check out this pull request on 

You'll be able to see Jupyter notebook diff and discuss changes. Powered by ReviewNB.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-03-22T18:58:37Z
----------------------------------------------------------------

Nice trick

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-03-22T18:58:38Z
----------------------------------------------------------------

Did the chapters switch here?

AlexAndorra commented on 2020-03-22T19:08:50Z
----------------------------------------------------------------

If you're referring to all the cells Code 10.XX, they are just code from the first edition that I have here for reference -- as I'm still working on the end of this chapter they could prove useful. I'll delete them when I'm finished.

[AlexAndorra]

@aloctavodia and @canyon289 I found a way to improve the scientific model of Oceanic tools 🎉
I used a different trick to scale the population data, and it seems to work -- although I'm not sure why the difference is so big compared to the previous method 🤷
I would very much appreciate @junpenglao's opinion on this, to make sure we didn't miss something -- see the PR's description for more details or tell me if you have questions. And a big thank you in advance!

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-04-14T03:30:50Z
----------------------------------------------------------------

Is the +1 missing here because we changed actor to 0 index?

AlexAndorra commented on 2020-04-14T09:19:14Z
----------------------------------------------------------------

Yes, we don't need it here because of zero-indexing

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-04-14T04:20:42Z
----------------------------------------------------------------

I only did cursory googling but for 11.3 xtabs would pandas cross tabs reproduce the same result? https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.crosstab.html

AlexAndorra commented on 2020-04-14T10:17:27Z
----------------------------------------------------------------

Don't know why I forgot this cell -- it's done now ;)

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-04-14T04:20:43Z
----------------------------------------------------------------

What is m11_1bis? It looks like the dnorm(0, 1.5) model but since it doesnt appaer in the code block its a bit confusing. Maybe add comment saying not part of code block but used to recreate figure 11.3?

Also out of curiosity was does bis mean?

AlexAndorra commented on 2020-04-14T10:22:39Z
----------------------------------------------------------------

You're right -- just added this precision.

"bis" is a latin prefix we use in French, Spanish, Italian, etc. to indicate something is repeated, or is the second instance of something.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-04-14T04:20:43Z
----------------------------------------------------------------

Similar note here. The model is defined in 11.7. Would add comment or label to explain some of the code has switched blocks from the book

AlexAndorra commented on 2020-04-14T10:25:43Z
----------------------------------------------------------------

Done

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-04-14T04:20:44Z
----------------------------------------------------------------

This is my own curiosity , what does the intX and pm.Data do in these lines? actor_id = pm.intX(pm.Data("actor_id", actor_idx))

AlexAndorra commented on 2020-04-14T09:28:24Z
----------------------------------------------------------------

Currently, you can't use shared variables for indexing in a model, because index variables have to be ints but pm.Data casts them to floats under the hood.

There is a PR currently working on that, but in the meantime the workaround is what I'm doing above.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-04-14T04:20:45Z
----------------------------------------------------------------

Nothing you need to change but ths margin in this forest plot on top is excessive. Something we should address in arviz code

AlexAndorra commented on 2020-04-14T09:29:14Z
----------------------------------------------------------------

Yeah I agree -- Osvaldo made the same comment.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-04-14T04:20:45Z
----------------------------------------------------------------

Truly a masterpiece

AlexAndorra commented on 2020-04-14T09:30:15Z
----------------------------------------------------------------

Ha ha ha, thanks! This was a monster!

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

canyon289 commented on 2020-04-14T04:20:46Z
----------------------------------------------------------------

Being picky here this would be easier to read if the data load occurred again like it is in the book even if the outcome is the same

AlexAndorra commented on 2020-04-14T10:31:48Z
----------------------------------------------------------------

Done

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

aloctavodia commented on 2020-04-20T12:57:00Z
----------------------------------------------------------------

seaborn: unused import

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

aloctavodia commented on 2020-04-20T12:57:01Z
----------------------------------------------------------------

Here you can also add one dimension to avoid the warnings as I mentioned in my other review

[AlexAndorra]

• Blackified the code and adressed @aloctavodia's and @canyon289's comments -- about the ArviZ warning, I chose not to add one dimension, as it clutters the code and it probably comes from a bug. Once this bug is resolved I can rerun the NB.

• I solved the problem with the first Multinomial model by taking the first category as pivot instead of the last one (as the last one is very probable compared to the others, it switches the sign of the slope when taken as the pivot). However, this is still a mystery to me why this works in the book -- is it a typo in the book or a bug on PyMC3? 🤷‍♂️ I think we can merge, but maybe we'll have to revisit that.

If all this is good for you, I think this is ready to merge 🎉

[aloctavodia]

It would be a good idea to open issues so we do not forget about this stuff.

[AlexAndorra]

True, but I don't know if these are still open issues: maybe what I did solved the problems 🤷‍♂️ I'm counting on the community's peer-reviewing to find out 😅