Fixed terms and an aggregate random-effect contribution use the same joint native conditional covariance. Variance/dispersion parameters are held fixed; GLMM covariance is a local Gaussian approximation, not posterior draws. Singular fits remain flagged in metadata, rather than acquiring invented variance. Dense joint covariance calculations are limited to 256 MiB. Non-log natural differences are evaluated at the full fitted reference predictor, including random effects. Analytic difference intervals condition on that reference; joint simulations also propagate its coefficient dependence. Influence reference grids require focus values and existing grouping levels; response-index reference grids instead omit year and set group effects to zero. For a single random intercept, aggregate random-effect CDI categories use the fitted grouping labels. More complex random structures retain the aggregate contribution categories rather than inventing separate coefficients.
Usage
# S3 method for class 'merMod'
influ(
model,
focus,
data = NULL,
weights = NULL,
reference_data = NULL,
reference_weights = NULL,
uncertainty = "auto",
retain = "summary",
probs = c(0.025, 0.975),
ndraws = 1000L,
seed = NULL,
draws_path = NULL,
keep_model = FALSE,
response_transform = "identity",
...
)
# S3 method for class 'lmerMod'
influ(
model,
focus,
data = NULL,
weights = NULL,
reference_data = NULL,
reference_weights = NULL,
uncertainty = "auto",
retain = "summary",
probs = c(0.025, 0.975),
ndraws = 1000L,
seed = NULL,
draws_path = NULL,
keep_model = FALSE,
response_transform = "identity",
...
)
# S3 method for class 'glmerMod'
influ(
model,
focus,
data = NULL,
weights = NULL,
reference_data = NULL,
reference_weights = NULL,
uncertainty = "auto",
retain = "summary",
probs = c(0.025, 0.975),
ndraws = 1000L,
seed = NULL,
draws_path = NULL,
keep_model = FALSE,
response_transform = "identity",
...
)Arguments
- model
A fit from [lme4::lmer()], [lme4::glmer()], or [lme4::glmer.nb()]. Nonlinear `nlmer()` and arbitrary custom families are not supported.
- focus
Name of the focus variable, usually year.
- data
Optional model data. The model frame is used by default.
- weights
Optional numeric observation weights or the name of a weight column in `data`.
- reference_data
Optional prediction grid defining the common standardisation distribution. By default the observed data are used.
- reference_weights
Optional numeric weights, or a column name, for `reference_data`.
- uncertainty
One of `"auto"`, `"none"`, `"analytic"`, or `"simulation"`.
- retain
One of `"summary"`, `"derived_draws"`, or `"disk"`.
- probs
Lower and upper interval probabilities.
- ndraws
Number of joint coefficient draws for simulation uncertainty.
- seed
Optional simulation seed.
- draws_path
Output path used when `retain = "disk"`.
- keep_model
Retain the fitted model inside the diagnostic object.
- response_transform
For lme4 only, `"identity"` (default) or `"log"`. Declare `"log"` for a Gaussian lmer model of `log(response)`, or a named precomputed log-response column. Influence ratios are on the natural scale. Response indices use `exp(eta + sigma^2/2)`; sigma and random-effect variance parameters are treated as known. Other transformations are not supported.
- ...
Reserved for future backend options.