xend/yend is a positional aesthetic on purpose: that is what gets the
prediction transformed by its scale and trained into the panel's range.
Exactly one of the two arrives, on the axis the plot put the model's
outcome on.
Format
A ggplot2::Stat object.
Details
compute_layer is overridden to a pass-through, the same shape
ggplot2::StatIdentity uses, so an observation with an NA on it (a
predictor the model dropped) survives the stat instead of being removed
before the position runs. The residual's jitter is a function of the seed
and the rows it is handed, exactly as the point layer's is, so losing a row
here would hand it a different sequence of draws and the segment would land
away from its point.
See also
gf_resid() and gf_square_resid(), which pair this stat and a
geom for you.