Rank ANCOVA: cohort effect after adjusting for covariates (EXPLORATORY)

WHAT: fits lm(rank(y) ~ covariates + group) per population and reports the partial F test for group. Ranking the response first is the Conover-Iman rank-transform approach: it keeps the robustness of a rank test while allowing the design matrix a parametric test cannot otherwise get from these n.

Usage

stats_rank_ancova(
  freq,
  patients,
  smap,
  group_of,
  covariates = c("age", "sex"),
  value_col = NULL,
  min_resid_df = 5L,
  min_n = 3L
)

Arguments

freq

Population frequency table, one row per sample x population.

patients

Patient metadata table, keyed by patient_id.

smap

Sample map joining sample_id to patient_id.

group_of

Named character vector mapping sample_id to group label.

covariates

Column names in the patient table to screen as confounders. Default c("age", "sex").

value_col

Column to test or plot instead of the default measure.

min_resid_df

Minimum residual degrees of freedom before a model is fitted. Default 5L.

min_n

Minimum samples per group before a test is attempted. Default 3L.

Details

WHY IT IS LABELLED EXPLORATORY IN THE OUTPUT ITSELF, not merely in the docs: the rank transform is not exact in the presence of covariates, its Type I error drifts at small n, and – the binding constraint here – if a covariate is near-collinear with cohort the adjusted estimate is extrapolation. The min_resid_df guard refuses to fit rather than return a number that looks like the others in the results folder but is not comparable to them.

READ IT ALONGSIDE stats_confounding(): if that table shows no covariate is both unbalanced and associated, this adjustment should barely move the unadjusted p-value, and agreement between the two is the reassurance. A large disagreement is a finding about the DESIGN, not about the biology.