Skip to content

Problem comparing Akaike's AIC - nlme package

2 messages · Santiago Beguería, Brian Ripley

Santiago Beguería
# 
Hello,

I am comparing models made with nlme functions and non-nlme functions, based
on Akaike's AIC. The AIC values I get for exactly the same model formulation
--for example a linear model with no random effects fit with gls and lm,
respectively-- do not fit, although the values of the four model parameters
are exactly the same. For example:

m1 <- gls(height ~ age, data = Loblolly)
m2 <- lm(height ~ age, data = Loblolly)

m1$coefficients
(Intercept)         age
  -1.312396    2.590523
m2$coefficients
(Intercept)         age
  -1.312396    2.590523

But then:

AIC(m1)
[1] 428.9243
AIC(m2)
[1] 423.9153

I am trying to compare between more complex models, i.e. different ways of
incorporating spatial self-correlation, and this issue with the AIC is
really making me silly!

Thanks,

    S. Begueria
Brian Ripley
# 
The problem is that AIC is only defined for ML fitting, and gls defaults 
to REML.

I have always maintained that it is a bug that nlme's logLik function 
returns a log restricted likelihood for the default fits, and that this 
is converted to a mis-named AIC.  If you use

m1 <- gls(height ~ age, data = Loblolly, method="ML")

you wil get agreement.
On Tue, 11 Mar 2008, sbegueria wrote: