A low-ceremony alternative to pmx_covariates() for a long list of
covariates whose fidelity does not matter. Instead of a public range or level
set per column, the columns are named and their values are drawn directly
from the source data: a uniform draw over the (clipped) observed range for
continuous columns, and a proportional resample for categorical ones. Column
type is detected from the data at fit time.
Usage
pmx_covariates_auto(names, clip = c(0.01, 0.99))Arguments
- names
Character vector of covariate column names.
- clip
Two probabilities giving the quantiles a continuous column is clipped to before its range is taken, so the exact minimum and maximum are not exposed. Defaults to the 1st and 99th percentiles. Pass
NULLto use the raw observed minimum and maximum, matchingsynadamexactly.
Details
This is the approach used by Novartis's synadam, and it is not
differentially private: it exposes the data-derived support of each column. A
model that uses it is marked as having non-private covariates, and its
privacy report says so. Use it only where the source data's own access
controls and confidentiality obligations still apply, and never when the
covariate columns may reach anyone the source data could not.