Timeline for GAM (mgcv): AIC vs Deviance Explained
Current License: CC BY-SA 3.0
5 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Feb 1, 2018 at 4:30 | comment | added | jon_simon | Another thought: to determine whether the observed effect is due to unexpected behavior on the AIC side side or the Deviance side, replace AIC and Deviance, each in turn, by logLik and see what happens to the plots | |
| Feb 1, 2018 at 0:27 | comment | added | jon_simon | @L.Steele My initial answer was mistaken, and I've updated to clarify. I would recommend reading this post for information on how deviance is calculated for GAMs: stats.stackexchange.com/questions/190172/… | |
| Feb 1, 2018 at 0:24 | history | edited | jon_simon | CC BY-SA 3.0 |
added 755 characters in body
|
| Jan 31, 2018 at 6:58 | comment | added | L.Steele | Thank you, Jon. This suggests both are penalized by a constant (e.g., 2*df model) correct? I'm a bit confused as I've also read that AIC is "based on the Deviance, but penalizes you for making the model more complicated." This inquiry began with a reviewer insisting that AIC was penalized and deviance is not, and this seems to agree with gam() help which indicates model aic is and deviance is not penalized. Howerver, the results above support your comment, and/or that the aic penalty is fairly negligible here. Still reading & learning; thanks again for any thoughts. | |
| Jan 30, 2018 at 11:27 | history | answered | jon_simon | CC BY-SA 3.0 |