Estimating the heights of New Yorkers from their scuff marks

33 points by veggie 20 hours ago on lobsters | 3 comments

n1000 | 4 hours ago

This is tricky as the unobserved effect (leg angle) is correlated with the independent variable (scuff height), but perhaps we can assume a structural model

h_scuff = x * height - cos(knee bend) * y * height

But x and y are related, Your knee is some fraction of your height, and also determines the amount the scuff height varies with angle

h_scuff = x * height - cos(knee bend) * x * height

cosine is probably unrealistically simple but probably fair for now. The key point is the produced scuff mark will be drawn from a distribution centered at knee height, and bounded below by the floor (presumably where your straight leg reaches) and above by 2*knee height. The logit-normal is a typical distribution used in these situations.

So maybe it can be structured

h_scuff = K * height + F * x * height = (1 + F) * x * height

F is the foot factor, a logit-normal random variable representing how high up it is comfortable to rest and K is the fraction of your height up to your knee. At this point I think one can start trying to get rid of the too many free parameters but I realize we still have a key degeneracy, each observation has multiple unknowns (height, knee angle). You were right to call for a Bayesian approach. Here's a stab at it:

height ~ N(μ_h, σ_h²)                   #height prior
A      ~ N(0, σ²)                         #comfortable angle (who knows)
F      = expit(A)                         # foot factor
K      ~ N(0.27, something)      # maybe truncated or logit-normal: small variance knee height as proportion of total
h_scuff = 2K · height · F

cyplo | 11 hours ago

delightful !

brokebit | 8 hours ago

This made my day. Reminds me of looking at which buttons are worn on a keypad to isolate potential digits in the passcode.