mmes.specification: Results for bounded formula specification models

mmes.specificationR Documentation

Results for bounded formula specification models

Description

Point outputs, simulation and explicitly requested local inference for formula expressions fitted by the bounded dense reference engine.

Usage

## S3 method for class 'mmes.specification'
coef(object, ...,
  type=c("mean", "all", "working", "group"))
## S3 method for class 'mmes.specification'
fitted(object, type=c("response", "link"), ...)
## S3 method for class 'mmes.specification'
residuals(object,
  type=c("response", "pearson"), ...)
## S3 method for class 'mmes.specification'
predict(object, newdata=NULL,
  type=c("response", "link"), target=c("conditional", "zero", "population"),
  se.fit=FALSE, cov.fit=FALSE,
  interval=c("none", "confidence", "prediction"), level=.95, trials=NULL, ...)
## S3 method for class 'mmes.specification'
simulate(object, nsim=1, seed=NULL,
  target=c("conditional", "zero", "population"), ...)
## S3 method for class 'mmes.specification'
summary(object, ..., inference=FALSE)
## S3 method for class 'mmes.specification'
vcov(object, ...,
  type=c("mean", "all", "working", "random", "conditional", "group"))
## S3 method for class 'mmes.specification'
confint(object, parm, level=.95, ...,
  type=c("mean", "all", "working"))

Arguments

object

A fitted mmes.specification model.

type

Output scale or parameter/covariance layout.

newdata

Optional nonempty prediction data frame. Random levels must be known for conditional and population targets.

target

Conditional fitted modes, zero deviations, or the integrated prior population. Population link prediction integrates the nonlinear link predictor, not its zero-deviation value.

se.fit, cov.fit

Request standard errors and, optionally, full prediction covariance ordered as unlist(fit, use.names=FALSE).

interval

No interval, local normal confidence interval, or local normal future-response interval. The latter requires scalar response output.

level

Confidence level strictly between zero and one.

trials

Positive integer trials for binomial prediction intervals; one or one per prediction row. Required for new-data grouped-binomial intervals.

nsim

Positive integer simulation count.

seed

Optional seed. Seeded calls restore the caller's RNG state.

inference

Explicitly request coefficient uncertainty in the summary.

parm

Unique parameter names or indices. Omitted selects all.

...

No additional arguments are accepted.

Details

Default coefficients are fixed regression coefficients on their declared working/link scales. type="working" includes every free optimizer coordinate; "all" reports mean coefficients and natural covariance and family nuisance coordinates, including fixed entries with zero variance. coef(type="group") returns natural parameter values per retained predictor row, including fitted parameter deviations.

Scalar response predictions are vectors. A single ordinal or nominal family returns a probability matrix with declared category ordering; mixed categorical families return trait-named lists. Link output preserves nominal logit matrices. Categorical response residuals subtract fitted category indicators (nominal) or expected category codes (ordinal).

Population integration uses the joint covariance of all active predictor coordinates, including correlated nominal logits within a record. Separate parameter records remain independent. Quadrature doubles order until the requested tolerance or fails explicitly at its order/memory limit. Simulation draws the fitted conditional modes, zero deviations or independent declared prior covariance blocks before applying parameter links. Binomial predictions and simulated binomial responses are proportions, using the declared integer trials.

Conditional curvature composes symbolic expression derivatives with the full native family curvature. Outer scores use central numerical differences. Legacy noncanonical scalar links use a numerical second derivative of the inverse link.

post_nonlinear() and covariance/interval requests verify reconstruction, convergence, stationarity, numerical derivative agreement, boundaries and positive identified curvature. Conditional modes are re-certified at tighter tolerances for post-fit derivatives, without changing the saved fitting controls or statistical model. Singular information is not repaired using a pseudoinverse. vcov(type="conditional") holds all marginal parameters fixed; "random" adds mode sensitivity and is a local empirical Bayes PEV approximation, not an exact frequentist PEV. "random", "conditional", and "group" are unavailable for ordinary fixed-only fits. "group" propagates parameter links and mode motion to the natural group parameters. No uncertainty is computed during fitting or ordinary summary. All interval results are local normal approximations, not exact discrete or simultaneous intervals. Probability confidence limits may extend outside zero and one. Population prediction intervals include both integrated family variance and prior variation.

Ordinary fixed-only models retain their lm/glm covariance semantics, with fitted ML dispersion when appropriate. For nonlinear models, all estimated covariance and nuisance coordinates enter the checked observed information. vcDelta() names the non-mean natural reporting coordinates V1, V2, and so on, and returns their mapping as an attribute. Comparisons admit only regular nested ordinary fixed terms with unchanged nonlinear, family, random and observation semantics; Wald tests require named mean coordinates and explicit finite nulls. Boundary-variance tests are not calibrated.

Value

Coefficient vectors or matrices, predictions in the fitted family layout, simulation data frames (trait lists for mixed categorical responses), covariance matrices, confidence intervals, or a summary. Prediction uncertainty returns fit, se.fit, an approximation label, optional cov.fit, and optional lower/upper with matching shapes.

See Also

parameter, nonlinear_curve, post_nonlinear, profile_nonlinear


sommer documentation built on Nov. 14, 2026, 5:07 p.m.