Models fitted by gamm4::gamm4() (version 0.2-7 or later) have class
"gamm4". Pass the fitted object directly to gratia methods; historical
unclassed lists are not supported.
Details
Smooths, parametric effects, predictions, derivatives, simulations,
and posterior draws describe the population GAM component. Terms supplied
through random are excluded, including their realized effects, predictive
variation, and uncertainty. conditional_values() and
conditional_differences() condition on covariate values, not on realized
group effects. Group variables occurring only in random are not required
in prediction data, so new groups receive the same population predictions.
Use the mer component with lme4 for predictions conditional on group
effects and simulations including their variation. Diagnostics and response
simulations use the full fitted mixed model by default. With
conditioning = "conditional", all fitted smooth and additional random
effects are held fixed for response references. "marginal" draws new effects
from their fitted distribution. "gam_component" explicitly selects the
extracted GAM calculations used by earlier versions.
Marginal GLMM checks average over random effects using lme4's joint response
simulations. Residuals then use ranks in a shared reference sample, as
described in diagnostic_support. Gaussian identity-link models instead
use analytic CDFs of the marginal response distribution.
Posterior response draws also default to the full conditional model. They
draw fixed and fitted random effects together, keeping covariance parameters
at their fitted values. These Gaussian approximations currently use fitting
observations only. To supply user coefficient draws or prediction data to
posterior_samples(), select conditioning = "gam_component" explicitly.
Coefficient draws from post_draws(), fitted_samples(), predicted_samples()
and smooth_samples() still use the fitted GAM covariance matrix, including
its uncertainty due to fitting the mixed model. User draws must have one
column per GAM coefficient. Metropolis-Hastings sampling is unsupported:
the GAM likelihood does not include the additional random effects.
Smoothness uncertainty-corrected covariance matrices and alternative EDF
are unavailable from gamm4; requested corrections warn and fall back to
the available covariance matrix or EDF. Variance components are extracted
from the mixed-model component, with smooth labels and random-effect term
names (separate penalties on a tensor smooth retain their grouping suffix);
interval bounds are NA because they are not supplied by the fit.