Measure, and optionally correct, the Laplace approximation error in a fit
Source:R/ctBackendLaplaceCheck.R
ctLaplaceCheck.RdRecomputes a backend='julia', intoverpop='laplace' fit's log
marginal likelihood by adaptive Gauss-Hermite quadrature over the same
random-effect integral the Laplace approximation approximates, and reports
how far the estimate would move if that integral were taken exactly.
Usage
ctLaplaceCheck(
fit,
nodes = 5L,
correction = TRUE,
step = 0.001,
refine = FALSE,
maxiter = 50L,
cores = getOption("mc.cores", 2L),
verbose = 0L
)Arguments
- fit
A
ctJuliaFitfitted withintoverpop='laplace'.- nodes
Quadrature nodes per random effect. Cost is
nodes^kprocess log likelihoods per block, wherekis that level's number of random effects; 5 is enough to locate the maximum in the cases tested, 9 to settle the value.- correction
Compute the first-order correction to the estimate. Costs
2 * nparquadrature evaluations, and needs the fit's Hessian, so it is skipped when uncertainty was not computed.- step
Finite-difference step for the gap gradient, on the raw scale.
- refine
Continue the fit on the quadrature objective rather than correcting linearly: the refined fit
intoverpop = 'quadrature'runs at fit time (seectFit), from this fit's estimate, with the quadrature rules placed at every point. Its gradient is exact, so it costs a quadrature gradient per step rather than the2 * nparquadrature evaluations per gradient the finite-difference refinement this replaced cost, and it reports the point it reaches inrefined.- maxiter
Iteration cap for the climb of
refine.- cores
Engine threads for the quadrature.
- verbose
Integer; 1 or more prints progress.
Value
A list with gap (quadrature minus Laplace log marginal at the
estimate), gap_per_subject, nodes, and, when
correction is TRUE, a parameters data frame with one
row per raw parameter giving estimate, delta,
corrected, se and delta_se. With refine it also
carries refined. at says whether the estimate checked was the
Laplace optimum ('laplace') or already quadrature-corrected
('corrected').
dropped_directions counts the directions of the information matrix
the correction does not step along, whose correction is reported as zero
rather than as the arbitrarily large step an unguarded solve would
produce: exactly the directions fit$identifiability names, those
whose curvature is negligible or not positive and those the likelihood was
measured flat along. On a model where the Laplace approximation is
already exact the gap gradient is floating-point noise, and dividing that
by a near-singular curvature is how a meaningless correction gets a
plausible-looking number.
Details
The quadrature is adaptive in the usual sense: it is centred at each subject's inner mode and scaled by the inverse of the inner curvature there, both of which the fit has already computed. The curvature's eigenvalues are clipped from below at one, the prior's own curvature, before scaling: where the likelihood has gone convex in the random effects an unclipped rule puts its nodes far outside the prior and misses the mass it is meant to measure (1.8 nats too high on average on near-singular units in simulation, 0.06 clipped). Where every eigenvalue is at least one nothing changes.
The gap is quadrature minus the fit's own Laplace term, the quantity the fit
maximised. With nodes = 1 that gap is zero wherever the curvature is
at least the prior's, which is what makes the comparison meaningful – the
two are integrating the same thing by different rules. On a unit with an
eigenvalue below one it is not zero even at one node: the one-point clipped
rule is the eigenwise-floored term, below the fit's.
The correction is a single Newton step, delta = solve(-hessian,
gradient of (quadrature - laplace)), evaluated at the estimate. Read it
against the standard errors: delta / se well below one says the
approximation is not what limits the answer, and near or above one says the
point estimate is approximation-limited and the interval will not cover
whatever width it has.
A fit can be refined, with its units' random effects integrated by
quadrature, when it is fitted
(ctFit(intoverpop = 'quadrature')) or after
(ctFitUncertainty(fit, intoverpop = 'quadrature')); by default it
only runs the random-effects check, which warns where integrating would
move an estimate. On a corrected fit this check is made at the corrected estimate,
which is not a Laplace optimum: the correction reported is then the
further Newton step on the quadrature objective from there, with the
Laplace gradient included, and at on the result is
'corrected'. A small further step says the correction got there; a
large one says it did not. On any other fit it is the first step.
Nested groupings are handled. A group's integral does not factor over its members, but it does factor conditionally – given the group effect the members are independent – so the rule recurses over the same block tree the fit already builds, at a cost linear in the number of groups at each level rather than exponential in the number of members.
See also
ctFitUncertainty with method = 'sample'
and intoverpop = 'none' removes the approximation
instead of measuring it, by sampling the joint posterior; much slower,
and the right answer when delta_se says a linear correction is not
credible.
Examples
if (FALSE) { # \dontrun{
data <- ctstantestdat
model <- ctModel(type = 'ct', manifestNames = 'Y1', latentNames = 'eta1',
LAMBDA = matrix(1))
model$pars$indvarying <- model$pars$matrix %in% 'MANIFESTMEANS'
fit <- ctFit(data, model, backend = 'julia', intoverpop = 'laplace')
check <- ctLaplaceCheck(fit)
check # gap, and the largest corrections in standard errors
check$parameters # the full per-parameter table
} # }