hessian2: Hessian With Small Sample Correction.

In lavaSearch2: Tools for Model Specification in the Latent Variable Framework

Hessian With Small Sample Correction.

Description

Extract the hessian from a latent variable model, with small sample correction

Usage

hessian2(object, indiv, cluster, as.lava, ...)

## S3 method for class 'lvmfit'
hessian2(
  object,
  indiv = FALSE,
  cluster = NULL,
  as.lava = TRUE,
  ssc = lava.options()$ssc,
  ...
)

## S3 method for class 'lvmfit2'
hessian2(object, indiv = FALSE, cluster = NULL, as.lava = TRUE, ...)

Arguments

object	a lvmfit or lvmfit2 object (i.e. output of lava::estimate or lavaSearch2::estimate2).
indiv	[logical] If TRUE, the hessian relative to each observation is returned. Otherwise the total hessian is returned.
cluster	[integer vector] the grouping variable relative to which the observations are iid.
as.lava	[logical] if TRUE, uses the same names as when using stats::coef.
...	additional argument passed to estimate2 when using a lvmfit object.
ssc	[character] method used to correct the small sample bias of the variance coefficients: no correction (code"none"/FALSE/NA), correct the first order bias in the residual variance ("residual"), or correct the first order bias in the estimated coefficients "cox"). Only relevant when using a lvmfit object.

Details

When argument object is a lvmfit object, the method first calls estimate2 and then extract the hessian.

Value

An array containing the second derivative of the likelihood relative to each sample (dim 3) and each pair of model coefficients (dim 1,2).

See Also

estimate2 to obtain lvmfit2 objects.

Examples

#### simulate data ####
n <- 5e1
p <- 3
X.name <- paste0("X",1:p)
link.lvm <- paste0("Y~",X.name)
formula.lvm <- as.formula(paste0("Y~",paste0(X.name,collapse="+")))

m <- lvm(formula.lvm)
distribution(m,~Id) <- Sequence.lvm(0)
set.seed(10)
d <- lava::sim(m,n)

#### latent variable models ####
e.lvm <- estimate(lvm(formula.lvm),data=d)
hessian2(e.lvm)

lavaSearch2 documentation built on April 12, 2023, 12:33 p.m.