/ lab
Statistical experiments
Stats ideas I wanted to see drawn out. Each one goes after a single thing about a model and stops there, so they're rougher than the rest of the site (on purpose, mostly), and there are two so far.
GAMLSS
Same scatter, three fits
Four scatter plots with three fits stacked on each, and you get to watch each fit run out of road. A line only gets a slope. A GAM can bend to follow the average. A GAMLSS can also let the spread grow and shrink as you move along, which real data does basically all the time.
explore →
Bayesian
Three problems, three posteriors
Lost keys, Bay Area earthquakes, and a yes-or-no call. Each one ends with a distribution over the thing you didn't know, and once you have that, every question you actually care about is an integral over it.
explore →
source fugue