You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Two things #193 scoped for example7 that did not land in PR #243, split out so
that PR can close its issue.
A dispersion sub-model as a vignette section.Rewrite example7 as a growth-data case study using the alga dataset #193 called this "the
payoff": example7 fits growth-rate data where the variance is not constant
in the mean, and disp() is the package feature that addresses it, but the
vignette does not demonstrate it.
Re-run the precision sweep under disp("loglinear"). The sweep in the
vignette was run with a pooled sigma.
Why it matters
The bounded check recorded in notes/disp_sensitivity_example7.md shows that arm B2's headline numbers are largely a pooled-sigma artefact. That is one
scenario at n = 50, so it is not enough to re-score the study, but it is enough
to say that the vignette's claims about B2 are conditioned on a variance
assumption the vignette does not state and does not test. Either the sweep is
re-run under a dispersion sub-model, or the claim is qualified.
Evidence
notes/disp_sensitivity_example7.md, on the branch of PR #243. One scenario,
n = 50, bounded response.
Implementation detail
Ordering: this is downstream of PR #243 merging, and downstream of the ECx
redefinition in #281, which changes every ECx example7 reports and therefore
requires a re-render regardless. Do not start it before both have landed.
The re-run is the expensive half: example7 alone takes about 137 minutes
locally and about 2 h 56 m in CI, for 24 alga fits, and the sweep multiplies
that. Read PR #243's handoff note before starting a precompile -- it records two
traps that cost a previous session time, the stale installed library and the
figure paths.
The third item PR #243 flagged, that example7 is about 14,700 words and the
longest vignette in the package, is a judgement for RF and is not part of this
issue. Adding a dispersion section makes it longer.
What
Two things #193 scoped for
example7that did not land in PR #243, split out sothat PR can close its issue.
A dispersion sub-model as a vignette section. Rewrite
example7as a growth-data case study using thealgadataset #193 called this "thepayoff":
example7fits growth-rate data where the variance is not constantin the mean, and
disp()is the package feature that addresses it, but thevignette does not demonstrate it.
Re-run the precision sweep under
disp("loglinear"). The sweep in thevignette was run with a pooled
sigma.Why it matters
The bounded check recorded in
notes/disp_sensitivity_example7.mdshows thatarm B2's headline numbers are largely a pooled-
sigmaartefact. That is onescenario at n = 50, so it is not enough to re-score the study, but it is enough
to say that the vignette's claims about B2 are conditioned on a variance
assumption the vignette does not state and does not test. Either the sweep is
re-run under a dispersion sub-model, or the claim is qualified.
Evidence
notes/disp_sensitivity_example7.md, on the branch of PR #243. One scenario,n = 50, bounded response.
Implementation detail
Ordering: this is downstream of PR #243 merging, and downstream of the ECx
redefinition in #281, which changes every ECx
example7reports and thereforerequires a re-render regardless. Do not start it before both have landed.
The re-run is the expensive half:
example7alone takes about 137 minuteslocally and about 2 h 56 m in CI, for 24
algafits, and the sweep multipliesthat. Read PR #243's handoff note before starting a precompile -- it records two
traps that cost a previous session time, the stale installed library and the
figure paths.
The third item PR #243 flagged, that
example7is about 14,700 words and thelongest vignette in the package, is a judgement for RF and is not part of this
issue. Adding a dispersion section makes it longer.