On Fri, 24 Oct 2003, Volker Bahn wrote:

>

> | On Wed, 22 Oct 2003, Volker Bahn wrote:

> |

> | > Dear Roger,

> | >

> | > thanks for your tips. I will check Haining. Concerning your latter

> comment:

> | > |

> | > | Perhaps just fit lm(), AIC should be OK, as should the likelihood

> ratio

> | > | test. I don't have access to the software, so I'm generalising from

> how

> | > | similar classes work in R.

> | >

> | > I initially fit my "null" model in lm() but the resulting log-lik was so

> | > different from the corresponding slm() log-lik that a comparsion was not

> | > possible. I figured that this must be because of the different

> optimisation

> | > processes in lm() and slm().

> |

> | I wouldn't think so - the basic R logic in logLik.lm() is:

> |

> | val <- 0.5 * (sum(log(w)) - N * (log(2 * pi) + 1 - log(N) +

> | log(sum(w * res^2))))

> |

> | where w is a vector of 1's if there are no weights. val is adjusted for

> | REML estimators, but otherwise is as you would expect. Maybe simply the

> | slm() is a much better fit?

> |

>

> This is starting to be out of my league - I don't really know or understand

> the formula or its consequences. What I do know, though, is that my loglik

> of "null" models calculated with lm() was actually around 0 while the loglik

> of the equivalent slm() model was around -2000 suggesting that the slm() was

> way worse. I then used the following formula to calculate the loglik of

> slm() myself from the RSE, but, as I learned later, this formula really only

> applies to OLS regressions:

>

> Loglik = (-n / 2) * LN(RSE^2 * df / n) - (n / 2) * LN(2*PI()) - (n / 2)

>

> The results were pretty credible, though: The slm() were just about as much

> better than the corresponding lm() as I had expected. My next line of attack

> was to set the spatial coefficient rho or just "parameter" as it is called

> in Splus to 0 to get a comparable loglik for the non-spatial model. I

> accomplished

> this indirectly through a likelihood ratio test (lrt.slm), which allows me

> to control the parameter settings for the reduced model. However, I would

> prefer to be able to do this directly so that I can see whether the

> resulting model is really similar in coefficients and all to an lm model.

> Looking through the help files didn't show me an obvious way to accomplish

> this and looking through the functions themselves is just a little bit over

> my head. If you would happen to know how to control the spatial parameter in

> slm() or would have other suggestions I would appreciate to know.

Since I don't have access to S-PLUS myself, I don't think I can see how to

help. Does anyone have access to SpatialStats and feel able to help (CC-ed

to r-sig-geo)? I'm pretty uncomfortable with your logLik by hand, because

the Jacobian seems to be omitted, is that correct?

Roger

>

> Cheers

>

> Volker

>

>

>

--

Roger Bivand

Economic Geography Section, Department of Economics, Norwegian School of

Economics and Business Administration, Breiviksveien 40, N-5045 Bergen,

Norway. voice: +47 55 95 93 55; fax +47 55 95 93 93

e-mail: Roger.Bivand at nhh.no

Roger Bivand

Department of Economics

Norwegian School of Economics

Helleveien 30

N-5045 Bergen, Norway