Module influence 

Influence diagnostics: which individual observations disproportionately move the fit.

This is a genuinely different question from “is the overall fit good” (crate::fit_statistics) or “are the residual assumptions met” (crate::residuals). An observation can have an unremarkable residual yet bend the whole regression line toward itself.

• leverage — unusualness of an observation’s predictor values alone.
• cooks_distance — leverage and residual size combined into one overall influence measure.
• dffits — how much the fitted value for an observation moves when that observation is dropped.

Threshold guidance (convention, not mathematical fact)

• Leverage hᵢ > 2p/n (some use 3p/n) is often called high.
• Cook’s distance Dᵢ > 4/n is a commonly cited flag.
• |DFFITSᵢ| > 2·√(p/n) is a commonly cited flag.

These vary by source; treat them as convention.

Functions

cooks_distance

Cook’s distance Dᵢ for each observation.

dffits

DFFITS ᵢ for each observation.

leverage

Leverage hᵢ (the diagonal of the hat matrix) for each observation.