Check that a segmented model's cached knot still matches a fresh MARS estimate
Source:R/modelling_tools.R
check_knot_still_matches.RdThe segmented brms model's breakpoint prior is seeded from a fast,
frequentist MARS knot search (earth::earth()), then the model is fit once
and cached to disk with file_refit = "never" (see
fit_brms_model() and the "Continuous alternatives" section of
vignette("model-comparison", package = "aphantasiaEmotions")). Because the
cached model is never automatically refitted, it can silently go stale if
the underlying data changes - the site would keep reporting a knot fit
around old data without any visible sign that anything is wrong.
Arguments
- data
The data the live model-comparison page is using right now (e.g.
all_data).- seed_knot
The knot value that was used to seed the cached model's prior when it was originally fit - a fixed constant recorded at fit time, not read from the model object itself.
- formula
Formula passed to
earth::earth(). Default istas ~ vviq, matching the segmented model's own formula.
Details
This function is a safeguard against exactly that. It re-runs the cheap
earth::earth() search on the data supplied now, and compares the knot
it finds against seed_knot - the knot value that was actually used to
seed the prior for the cached model at the time it was fit (not the
brms model's own posterior estimate of the knot, which is expected to
differ from the earth seed even on unchanged data, since brms re-estimates
it from a prior rather than reproducing earth's point estimate exactly).
If the two don't match exactly, the underlying data has almost certainly
changed since the model was cached, and the model should be refit.