fitADMG: Fit data to an ADMG model

In rje42/ADMGs2: Functions for Describing and Fitting Discrete Mixed Graphs

Fit data to an ADMG model

Description

Fit discrete data to the Markov structure implied by an acyclic directed mixed graph or summary graph.

Usage

fitADMG(
  dat,
  graph,
  r = TRUE,
  tol = sqrt(.Machine$double.eps),
  SEs = TRUE,
  sparse = FALSE,
  quietly = TRUE,
  use_optim = TRUE
)

Arguments

dat	The data, as an array of counts or a data frame whose final column contains the counts.
graph	An ADMG (or summary graph), as an object of class graph.
r	Logical indicating whether or not recursive factorizations should be used.
tol	Numeric; if log-likelihood increases by less that tol in one step, procedure stops.
SEs	logical: should standard errors be calculated?
sparse	Should sparse matrices be used? Requires package Matrix.
quietly	Logical indicating whether output should be suppressed.
use_optim	should optim be used for fitting?

Details

Fits data using coordinate-wise block descent algorithm, with an Armijo line search. Details in Evans and Richardson (2010).

Recursive factorizations allow for modelling of Verma-constraints. See Richardson et al (2012).

Value

An object of class mixed_fit. This is a list containing (amongst other things):

params	An object of class Mparams, containing the Moebius parameters which maximise the likelihood.
ll	Value of the log-likelihood at the maximum.

Warning

For the algorithm to be guaranteed to work correctly, all counts for marginal tables consisting of districts and their parents should be positive. A warning will be produced if this is not so.

Author(s)

Robin Evans, Ilya Shpitser

References

Evans, R.J. and Richardson, T.S. (2010) - Fitting acyclic directed mixed graphs to binary data. UAI-10.

Richardson, T.S., Robins, J.M., Shpitser, I., Evans, R.J. (2012) -

See Also

summary.mixed_fit autoFit

Examples

data(gss_small)
data(gr1, package="MixedGraphs")

out = fitADMG(gss_small, gr1)
summary(out)
# not a good fit!

rje42/ADMGs2 documentation built on Sept. 3, 2024, 7:39 p.m.