# Covariate sets, keyed by grade band.
# numeric    -> continuous controls (demographic shares, 0..1; enrollment as count)
# categorical -> dummy-encoded controls (first level dropped as reference)
#
# Note on collinearity: race/ethnicity shares are compositional and sum to ~1,
# so we deliberately OMIT pct_white (it becomes the reference) to avoid a
# perfectly collinear design.
sets:
  es_base:
    numeric: [pct_econ, pct_ell, pct_swd, pct_black, pct_hispanic, pct_asian, enrollment]
    categorical: [admission_bucket]

  ms_base:
    numeric: [pct_econ, pct_ell, pct_swd, pct_black, pct_hispanic, pct_asian, enrollment]
    categorical: [admission_bucket]

  hs_base:
    numeric: [pct_econ, pct_ell, pct_swd, pct_black, pct_hispanic, pct_asian, enrollment]
    categorical: [admission_bucket]   # method1 bucket; dominant predictor at HS