Overall, `sqrt(x)` seems better than `log(x+1)`.
- count variables: log(x+1) better than sqrt(x), but not very different
- summed variables: sqrt(x) is better
- fraction variables: sqrt(x) is better
Four outcome variables shown below (pc1, fc, mbfc, afm). All are imputed ratings. Left peak is due to imputing 0, right peak is due to winsorizing.
![[1666304072.png]]
![[1666303910.png]]
![[1666304000.png]]
![[1666304042.png]]
---
# exposure distributions
- sqrt(x) is better for both count and summed variables
![[1666306211.png]]
![[1666306275.png]]
![[1666306312.png]]
![[1666306312.png]]