Add Besag-York-Mollie notebook #566
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
|
View / edit / reply to this conversation on ReviewNB fonnesbeck commented on 2023-08-21T19:00:01Z Line #10. from icardist import ICAR Can we change this to a PyMC import now that it's merged? daniel-saunders-phil commented on 2023-08-22T00:38:37Z I was going to just wait until the next release. Do you know if it's possible to install versions directly from github without them being released? I searched a bit for instructions but came up empty handed. bwengals commented on 2023-08-24T18:39:51Z assuming you have a conda env that you're working out of where you've already installed pymc, you can pip install from a specific git repo branch. Add |
|
View / edit / reply to this conversation on ReviewNB fonnesbeck commented on 2023-08-21T19:00:02Z "so it would receive a large penalty from each OF them"
also, missing a closing |
|
View / edit / reply to this conversation on ReviewNB fonnesbeck commented on 2023-08-21T19:00:03Z In the equation maybe just use a greek letter instead of
The last paragraph is a duplicate of the third-to-last paragraph.
Again missing a closing |
|
View / edit / reply to this conversation on ReviewNB fonnesbeck commented on 2023-08-21T19:00:04Z maybe add a |
|
View / edit / reply to this conversation on ReviewNB fonnesbeck commented on 2023-08-21T19:00:05Z What if there are covariates? Do we smooth then? e.g. if a covariate might explain why a tract differs from its neighbors daniel-saunders-phil commented on 2023-08-21T21:50:04Z Yeah that's a good question. My hunch is that everything proceeds the same. The ICAR distribution would capture all the unexplained spatial covariance. If the covariates explain some of the spatial variance, that's great. But I won't think that we require in a change in how we approach the BYM component of the model. The Stan case study (published paper version) includes four covariates and they just stick them into the linear model like:
y = b0 + b1 * x1 + b2 * x2 + ... + bym_components
I omitted adding predictors in my case study to keep things short but maybe its worth including one predictor variable to give us an opportunity to address it. bwengals commented on 2023-08-24T18:46:38Z I think it'd be helpful to include other covariates. If you don't have other covariates and your goal is just smoothing, there are tons of other ways to do that. Once you're trying to factor out the effects of other covariates then things like ICAR and GPs become a lot more useful. |
|
Yeah that's a good question. My hunch is that everything proceeds the same. The ICAR distribution would capture all the unexplained spatial covariance. If the covariates explain some of the spatial variance, that's great. But I won't think that we require in a change in how we approach the BYM component of the model. The Stan case study (published paper version) includes four covariates and they just stick them into the linear model like:
y = b0 + b1 * x1 + b2 * x2 + ... + bym_components
I omitted adding predictors in my case study to keep things short but maybe its worth including one predictor variable to give us an opportunity to address it. View entire conversation on ReviewNB |
…termark, Chris' comments
…ying something with internal section references
|
I was going to just wait until the next release. Do you know if it's possible to install versions directly from github without them being released? I searched a bit for instructions but came up empty handed. View entire conversation on ReviewNB |
|
@daniel-saunders-phil to answer your install question, since the PR was merged into main, then installing from pip via: should do the trick. |
|
assuming you have a conda env that you're working out of where you've already installed pymc, you can pip install from a specific git repo branch. Add View entire conversation on ReviewNB |
|
I think it'd be helpful to include other covariates. If you don't have other covariates and your goal is just smoothing, there are tons of other ways to do that. Once you're trying to factor out the effects of other covariates then things like ICAR and GPs become a lot more useful. View entire conversation on ReviewNB |
|
View / edit / reply to this conversation on ReviewNB bwengals commented on 2023-08-24T18:47:00Z while naive BYM models are often not.
Might be better to say "original" instead of naive BYM model. Alternatively you could make a distinction between BYM1 and BYM2 |
Okie I'll add a covariates. If both you and Chris are interested in seeing that example, I can imagine other folks will too. It will take a couple days I reckon. The raw stan data files are a bit of mess 🫠 |
|
@fonnesbeck @bwengals I finished adding the covariate business and made a number of other small tweaks. It's ready for another look.
Fixed. I think we'll keep it short by not detailing the BYM1 vs 2 distinction - it's well covered in the linked papers. |
|
Thanks @daniel-saunders-phil! |
|
Really nice explanation of the scaling factor, TY @daniel-saunders-phil! |
I've been working on a notebook that demonstrates how to build and interpret the Besag-York-Mollie model. The BYM model is a popular choice for spatial data, especially in epidemiology. PyMC just gained an ICAR distribution, which is a key ingredient in the BYM model. There is a well-developed literature on Bayesian approaches to BYM, written largely by folks in the Stan and r-INLA community:
By contrast, there is very little tutorial material written for Python users who want to do Bayesian spatial modeling. So this notebook would help rectify that and maybe make PyMC appealing to a new segment of the data science community.
To do:
📚 Documentation preview 📚: https://pymc-examples--566.org.readthedocs.build/en/566/