search

QCBSRworkshops / workshop08

Workshop 8 - Generalized additive models (GAMs)
Other
11 stars 7 forks source link

Fitting methods (REML should be preferred) #10

Closed dschoenig closed 3 years ago

dschoenig commented 3 years ago

Simon Wood recommends using ML or REML for fitting GAMs, as it is less prone to overfitting and because the Gaussian approximation to the model posterior will be more accurate (e.g. the confidence intervals are more robust).

REML will become the default method very soon in mgcv, it only hasn't been until now for backwards compatibility.

Because of that I think we should use method = "REML" for all gam() calls.

This change, however, would also affect the explanation of the fitting process, which currently explains GCV.

dschoenig commented 3 years ago

0ea20b83fb91a78531df5682347d0d1ed8e6ebff now uses method = "REML" for all calls to gam() in section 3 and later. The argument is not introduced earlier because theory is explained at the end of section 2. Slides explaining GCV have been hidden.