Possible correlates of percentile about scadsanalysis HOT 5 CLOSED

I STRONGLY SUSPECT S/N, spread, and mean will be correlated with each other, which raises the question which one(s) to use as predictors for percentile.

from scadsanalysis.

Statistical approaches

• lm
  • pro: super straightforward, r2 computes automatically
  • con: not actually appropriate for response variable bounded 0-1 (or 0-100); estimates go outside range of possible values AND fails to accommodate expected bunching up at the extremes
• binomial glm
  • pro: easy and automatically bendy
  • con: r2?
• beta regression
  • pro: from googling, developed specifically for modeling probability as a response. Use coefficients and pseudo-R2? See https://cran.r-project.org/web/packages/betareg/vignettes/betareg.pdf
  • con: looks bayesian (not freq), not sure how to implement or interpret (but learnable)

from scadsanalysis.

Beta regression can't handle 0 or 1, and I'm getting failures even when I fudged it (changed 0's to v small, etc)

Try zero-one inflated beta regression in brms? https://cran.r-project.org/web/packages/brms/vignettes/brms_families.html

from scadsanalysis.

These all covary so strongly with S and N I no longer have much confidence in teasing apart relationships.

from scadsanalysis.

I think it's worth doing a zero one inflated brms just to see if %ile value changes with S and N, or maybe N/S. We know there are complicated relationships between S and N and FS characteristics, but it is hard to tell if there's really any cascading effecton %iles.

from scadsanalysis.