Reports whether the grid actually brackets the minimum. A profile whose lowest point is its first or last grid value has not found the optimum, it has run out of grid, and the curve drawn from it understates how far the parameter can move. That is the failure the numbers alone hide, since a partial profile plots as a perfectly ordinary line.
Usage
# S3 method for class 'Rceattle_profile'
print(x, cutoff = 1.92, ...)Arguments
- x
A
"Rceattle_profile"object fromprofile.Rceattle().- cutoff
Objective units above the minimum that bound the reported interval. Default
1.92, the 95% profile-likelihood cutoff for one parameter.- ...
Currently unused.
Details
For a one-dimensional profile the interval reported is the usual
profile-likelihood one: the grid values within cutoff of the minimum, where
the default 1.92 is \(\chi^2_1(0.95)/2\). It is read off the grid, so it is
no finer than the spacing of values, and it is reported as open on either
side the grid does not close. It is also referenced to the best GRID point
rather than to the unconstrained MLE, which the object does not hold: the
grid minimum sits at or above the MLE, so the interval errs wide. And it is
reported as a range, so a profile with a second basin, or a failed point
inside the range, is called out as not contiguous rather than left to read
as one interval. No interval is given for a cross-profile over
two or more cells, the cutoff would be \(\chi^2_k(0.95)/2\) and the region
is not an interval.
Under random_rec = TRUE the objective is the Laplace-approximated marginal
likelihood, so this is a profile of that, with the usual caveat that the
approximation is what is being profiled.