Post

Did you do some kinda z-order curve here or what?

Tried a Hilbert ordering once, but it ends up being pretty dumb. Then I got into TSP solvers, and those outlined some general strategies -- but I found that really you have to bounce between assigned hierarchies (for states, US census regions) and TSP solutions. observablehq.com/@bmschmidt/u...

There are tradeoffs though -- above map cuts off the Baltics from Scandanavia, so sometimes I like this path instead.

umap to 1d works really well for that! Kinda by design, umap tries to preserve both local and global distances, which gives a pretty reasonable ordering, all things considered

I'd love to see the rows individually reshuffle with a transition effect.

I think the problem here is not only geographical, but also statistical/demographic. Isn't the noisyness also an effect of the unit sizes (Bayesian Surprise)? 14M vs 15 states with <1M in 1950. Like group by regions first, then drill down. en.wikipedia.org/wiki/1950_Un...

I learned that lesson in spades during COVID. Plotting stats by county was very noisy, but when I went to Texas Public Health Regions, the plots cleaned up significantly.

Maybe even revert everything back from relatives to absolutes and do a sort of per-region-swimlane beeswarm-plot with dots every 1000x cases. If the point is to communicate this cut-off at introduction state borders matter little, if there was no difference in policies.