icc_from_replicates

icc_from_replicates(data, subject_col, *, features=None)

Per-probe ICC(1,1) computed from the user’s own technical replicates.

For labs that ran duplicates. A one-way random-effects, single-measurement ICC is the right model when the “raters” are interchangeable array positions rather than named assessors:

ICC(1,1) = (MSB - MSW) / (MSB + (k-1)·MSW)

with k the average replicates per subject. Negative values are kept, not clipped to zero: a negative ICC means the within-subject spread exceeded the between-subject spread, which is a real and reportable state of affairs for a probe that measures nothing.

Preferred over the bundled table whenever it is available, because it is this laboratory’s noise on this platform rather than a published cohort’s.