phenolology 🐢 @scisus@scholar.social

Spent two days testing model configurations. Got everything working. Saved + version control. Now I’m running the same code and it’s not working and I can’t figure out what happened 😞

WHYYYY are there divergences

concurrency errors are not the best kind of errors. I think I'm going to have to do this problem the slow way :( :( :(

I love JEH Smith so much. A snippet to entice you:

"This new monotremous existence of ours leads to many moments of pure surrealism, where, for example, a massive government grant agency requires as a condition of funding that you include a Twitter presence as part of the public-outreach dimension of your research, and before you know it you are fielding feedback on your work in deep-ocean stratigraphy or the geopolitics of natural-gas pipelines from @maoistassplay69."

https://justinehsmith.substack.com/p/hobby-horses

@SummerSchool Hi! I have an accessibility question - is it possible to submit a pre-recorded presentation and do asynchronous text questions on here instead of "at" the conference? I have a chronic illness and don't ever know if I'll be well enough to give a presentation on any given day.

Good morning all! The storm last night kept me awake with it's beating at the walls (and dramatic blowing around of construction supplies) so I expect I'll require a nap later. But for now, a bit of writing and then and then onwards to a refactor of my data processing code so that I can better incorporating censoring into my models.

I added censorship to my model because a bunch of my data is censored, but there are an awful lot of divergences and I'm not sure what else I can do to fix them. I've already carefully non-centered a bunch of stuff to try to make the geometry nicer.

🤦‍♂️ If the parameters in the log-likelihood function in the test dataset aren't calculated from the training dataset then your log likelihood calculation won't work.

I think this means that for a hierarchical model with grouped folds, this means, for example, that if you calculate SiteA parameter with the training data and then test with SiteB data, you gotta use the superpopulation Site param, not the specific Site param.

I'm feeling grumpy today, but I'm going to sit down and see if I can finish my k-fold cross-validation before lunch. And then I'm going to spend a couple hours in my garden.