Log-likelihood, and hence AIC and BIC

Returns the log-likelihood of the data at the fitted coefficients, with df set to the total effective degrees of freedom. That is deliberately not the fitting criterion: under REML the criterion is a restricted one, is not a likelihood, and is not comparable across different mean structures, so using it for AIC would be wrong. It remains available as fit$objective.

Usage

# S3 method for class 'gamRTMB'
logLik(object, ...)

Arguments

object

A gamRTMB fit.

...

Ignored.

Value

An object of class logLik.

Details

With df = total EDF this is the convention used by mgcv::gam() and by GAMLSS's GAIC, so AIC() and BIC() are comparable across models fitted to the same response. Under method = "ML", where effective degrees of freedom are unavailable, df falls back to counting the fixed coefficients and free smoothing parameters.

Examples

set.seed(1)
d <- data.frame(x = runif(200)); d$y <- rnorm(200, sin(2 * pi * d$x), 0.3)
fit <- gamRTMB(y ~ list(mean = ~ s(x, k = 8)), data = d)
logLik(fit)
#> 'log Lik.' -34.98567 (df=8.392431)
AIC(fit)
#> [1] 86.7562