Compute BIC for one or several fitted model objects for which a log-likelihood value can be obtained.
Arguments
- object
A fitted model object from
splm()
,spautor()
,spglm()
, orspgautor()
whereestmethod
is"ml"
or"reml"
.- ...
Optionally more fitted model objects.
Value
If just one object is provided, a numeric value with the corresponding BIC.
If multiple objects are provided, a data.frame
with rows corresponding
to the objects and columns representing the number of parameters estimated
(df
) and the BIC.
Details
When comparing models fit by maximum or restricted maximum
likelihood, the smaller the BIC, the better the fit. The theory of
BIC requires that the log-likelihood has been maximized, and hence,
no BIC methods exist for models where estmethod
is not
"ml"
or "reml"
. Additionally, BIC comparisons between "ml"
and "reml"
models are meaningless – comparisons should only be made
within a set of models estimated using "ml"
or a set of models estimated
using "reml"
. BIC comparisons for "reml"
must
use the same fixed effects. To vary the covariance parameters and
fixed effects simultaneously, use "ml"
.
BIC is defined as \(-2loglik + log(n)(estparams)\), where \(n\) is the sample size
and \(estparams\) is the number of estimated parameters. For "ml"
, \(estparams\) is
the number of estimated covariance parameters plus the number of estimated
fixed effects. For "reml"
, \(estparams\) is the number of estimated covariance
parameters.