Leave-One-Out Diagnostic for interaction effects — loo_interaction

Computes leave-one-out diagnostics for each feature by dropping each sample one at a time from each ancestry dataset. Shows how much each sample influences the interaction effect estimate.

Usage

loo_interaction_effect(X, Y, MX, MY, g_col, a_col, feature)

Arguments

X: A numeric matrix or data.frame of expression values for ancestry A.
Y: A numeric matrix or data.frame of expression values for ancestry B.
MX: A data.frame of metadata for X. Must include a column with condition/group labels.
MY: A data.frame of metadata for Y. Must include a column with condition/group labels.
g_col: String. Name of the column in `MX` and `MY` that defines the condition/group (must be a factor with exactly 2 levels).
a_col: String. Name of the column in `MX` and `MY` that defines ancestry.
feature: Character or numeric vector. Names or indices of features (columns) to test.

Value

A tidy data.frame with columns:

feature: Feature name or index.
sample: Sample name that was left out.
T_loo: Interaction effect with that sample left out.
T_obs: Original full-sample interaction effect.
ancestry: Ancestry of the sample ("eur" or "afr").