Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

add CAIC and EDF calcs #387

Merged
merged 3 commits into from
Dec 12, 2024
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions NAMESPACE
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
# Generated by roxygen2: do not edit by hand

S3method(cAIC,sdmTMB)
S3method(coef,sdmTMB)
S3method(confint,sdmTMB)
S3method(df.residual,sdmTMB)
Expand All @@ -25,6 +26,7 @@ S3method(vcov,sdmTMB)
export(Beta)
export(add_barrier_mesh)
export(add_utm_columns)
export(cAIC)
export(censored_poisson)
export(delta_beta)
export(delta_gamma)
Expand Down
129 changes: 129 additions & 0 deletions R/caic.R
Original file line number Diff line number Diff line change
@@ -0,0 +1,129 @@
#' Calculate conditional AIC
#'
#' Calculates the conditional Akaike Information criterion (cAIC).
#'
#' @param object Output from [sdmTMB()].
#' @param what Whether to return the cAIC or the effective degrees of freedom
#' (EDF) for each group of random effects.
#' @param ... Other arguments for specific methods. Not used.
#'
#' @details cAIC is designed to optimize the expected out-of-sample predictive
#' performance for new data that share the same random effects as the in-sample
#' (fitted) data, e.g., spatial interpolation. In this sense, it should be a
#' fast approximation to optimizing the model structure based on k-fold
#' cross-validation.
#'
#' By contrast, [AIC()] calculates the marginal Akaike Information Criterion,
#' which is designed to optimize expected predictive performance for new data
#' that have new random effects, e.g., extrapolation, or inference about
#' generative parameters.
#'
#' cAIC also calculates the effective degrees of freedom (EDF) as a byproduct.
#' This is the number of fixed effects that would have an equivalent impact on
#' model flexibility as a given random effect.
#'
#' Both cAIC and EDF are calculated using Eq. 6 of Zheng, Cadigan, and Thorson
#' (2024).
#'
#' For models that include profiled fixed effects, these profiles are turned
#' off.
#'
#' @return
#' Either the cAIC or the effective degrees of freedom (EDF) by group
#' of random effects depending on the argument `what`.
#'
#' @references
#' **Deriving the general approximation to cAIC used here:**
#'
#' Zheng, N., Cadigan, N., & Thorson, J. T. (2024).
#' A note on numerical evaluation of conditional Akaike information for
#' nonlinear mixed-effects models (arXiv:2411.14185). arXiv.
#' \doi{10.48550/arXiv.2411.14185}
#'
#' **The utility of EDF to diagnose hierarchical model behaviour:**
#'
#' Thorson, J. T. (2024). Measuring complexity for hierarchical
#' models using effective degrees of freedom. Ecology,
#' 105(7), e4327 \doi{10.1002/ecy.4327}
#'
#' @examples
#' mesh <- make_mesh(dogfish, c("X", "Y"), cutoff = 15)
#' fit <- sdmTMB(catch_weight ~ s(log(depth)),
#' time_varying = ~1,
#' time_varying_type = "ar1",
#' time = "year",
#' spatiotemporal = "off",
#' mesh = mesh,
#' family = tweedie(),
#' data = dogfish,
#' offset = log(dogfish$area_swept)
#' )
#' cAIC(fit)
#' cAIC(fit, what = "EDF")
#' AIC(fit)
#' @export
cAIC <- function(object, what = c("cAIC", "EDF"), ...) {
UseMethod("cAIC", object)
}

#' @exportS3Method
cAIC.sdmTMB <- function(object, what = c("cAIC", "EDF")) {
what <- match.arg(what)
what <- tolower(what)
tmb_data <- object$tmb_data

## Ensure profile = NULL
if (is.null(object$control$profile)) {
obj <- object$tmb_obj
} else {
obj <- TMB::MakeADFun(
data = tmb_data,
parameters = object$parlist,
map = object$tmb_map,
random = object$tmb_random,
DLL = "sdmTMB",
profile = NULL #<
)
}

## Make obj_new
tmb_data$weights_i[] <- 0
obj_new <- TMB::MakeADFun(
data = tmb_data,
parameters = object$parlist,
map = object$tmb_map,
random = object$tmb_random,
DLL = "sdmTMB",
profile = NULL
)

par <- obj$env$parList()
parDataMode <- obj$env$last.par
indx <- obj$env$lrandom()
q <- length(indx)
p <- length(object$model$par)

## use '-' for Hess because model returns negative loglikelihood
Hess_new <- -Matrix::Matrix(obj_new$env$f(parDataMode, order = 1, type = "ADGrad"), sparse = TRUE)
Hess_new <- Hess_new[indx, indx] ## marginal precision matrix of REs

## Joint hessian etc
Hess <- -Matrix::Matrix(obj$env$f(parDataMode, order = 1, type = "ADGrad"), sparse = TRUE)
Hess <- Hess[indx, indx]
negEDF <- Matrix::diag(Matrix::solve(Hess, Hess_new, sparse = FALSE))

if (what == "caic") {
jnll <- obj$env$f(parDataMode)
cnll <- jnll - obj_new$env$f(parDataMode)
cAIC <- 2 * cnll + 2 * (p + q) - 2 * sum(negEDF)
return(cAIC)
} else if (what == "edf") {
## Figure out group for each random-effect coefficient
group <- factor(names(object$last.par.best[obj$env$random]))
## Calculate total EDF by group
EDF <- tapply(negEDF, INDEX = group, FUN = length) - tapply(negEDF, INDEX = group, FUN = sum)
return(EDF)
} else {
cli_abort("Option not implemented")
}
}
75 changes: 75 additions & 0 deletions man/cAIC.Rd

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

16 changes: 16 additions & 0 deletions scratch/caic-demo.R
Original file line number Diff line number Diff line change
@@ -0,0 +1,16 @@

library(sdmTMB)

# Build a mesh to implement the SPDE approach:
mesh <- make_mesh(pcod_2011, c("X", "Y"), cutoff = 20)

# Fit a Tweedie spatial random field GLMM with a smoother for depth:
fit <- sdmTMB(
density ~ s(depth),
data = pcod_2011, mesh = mesh,
family = tweedie(link = "log"),
control = sdmTMBcontrol(profile="b_j")
)

CAIC.sdmTMB(fit, what="CAIC")
CAIC.sdmTMB(fit, what="EDF")
21 changes: 21 additions & 0 deletions tests/testthat/test-cAIC.R
Original file line number Diff line number Diff line change
@@ -0,0 +1,21 @@
test_that("cAIC and EDF work", {
skip_on_cran()
skip_on_ci()
mesh <- make_mesh(dogfish, c("X", "Y"), cutoff = 15)
suppressMessages(
fit <- sdmTMB(catch_weight ~ s(log(depth)),
time_varying = ~1,
time_varying_type = "ar1",
time = "year",
spatiotemporal = "off",
mesh = mesh,
family = tweedie(),
data = dogfish,
offset = log(dogfish$area_swept)
)
)
expect_equal(AIC(fit), 12192.9613, tolerance = 1e-1)
expect_equal(cAIC(fit), 12089.4289, tolerance = 1e-1)
edf <- cAIC(fit, what = "EDF")
expect_equal(sum(edf), 54.3870, tolerance = 1e-2)
})
Loading