Skip to main content
Cross Validated

Return to Answer

BIC seems to be a big question mark for GAMMs, I therefore removed it as a possible answer
Source Link
Nate
Nate
  • 2.1k
  • 8
  • 25

What I found:

  1. Select should be FALSE when comparing with AIC
  2. Select=TRUE ONLY when we have a single model in mind (instead of comparing many models with AIC)

    Select should be FALSE when comparing with AIC

  3. All models being compared should be fit with method=ML when unpenalized fixed effects (parametric terms) are present (like here)

    Select=TRUE ONLY when we have a single model in mind (instead of comparing many models with AIC)

  4. Method=REML is fine though for comparing "nested" GAM(M) models in AIC when all effects are fully penalized

    All models being compared should be fit with method=ML when unpenalized fixed effects (parametric terms) are present (like here)

  5. Same for BIC

    Method=REML is fine though for comparing "nested" GAM(M) models in AIC when all effects are fully penalized

Found a great tutorial on GAMs from Gavin here which covers most of this.

What I found:

  1. Select should be FALSE when comparing with AIC
  2. Select=TRUE ONLY when we have a single model in mind (instead of comparing many models with AIC)
  3. All models being compared should be fit with method=ML when unpenalized fixed effects (parametric terms) are present (like here)
  4. Method=REML is fine though for comparing "nested" GAM(M) models in AIC when all effects are fully penalized
  5. Same for BIC

Found a great tutorial on GAMs from Gavin here which covers most of this.

What I found:

  1. Select should be FALSE when comparing with AIC

  2. Select=TRUE ONLY when we have a single model in mind (instead of comparing many models with AIC)

  3. All models being compared should be fit with method=ML when unpenalized fixed effects (parametric terms) are present (like here)

  4. Method=REML is fine though for comparing "nested" GAM(M) models in AIC when all effects are fully penalized

Found a great tutorial on GAMs from Gavin here which covers most of this.

Corrected some statements
Source Link
Nate
Nate
  • 2.1k
  • 8
  • 25

What I found:

  1. Select should be FALSE when comparing with AIC, but can be  
  2. TRUESelect=TRUE in "best"ONLY when we have a single model for variable selectionin mind (as seen in every example I've found on this siteinstead of comparing many models with AIC).
  3. ModelsAll models being compared should be fit with MLmethod=ML when unpenalized fixed effects (parametric terms terms) are missing in one model vs. anotherpresent (like here)
  4. REMLMethod=REML is fine though for comparing "nested" GAM(M) models in AIC withwhen all effects are fully penalized fixed effects
  5. Same for BIC

Found a great tutorial on GAMs from Gavin here which covers most of this.

What I found:

  1. Select should be FALSE when comparing with AIC, but can be TRUE in "best" model for variable selection (as seen in every example I've found on this site).
  2. Models should be fit with ML when unpenalized fixed effects (parametric terms) are missing in one model vs. another (like here)
  3. REML is fine for comparing "nested" GAM(M) models in AIC with fully penalized fixed effects
  4. Same for BIC

What I found:

  1. Select should be FALSE when comparing with AIC 
  2. Select=TRUE ONLY when we have a single model in mind (instead of comparing many models with AIC)
  3. All models being compared should be fit with method=ML when unpenalized fixed effects (parametric terms) are present (like here)
  4. Method=REML is fine though for comparing "nested" GAM(M) models in AIC when all effects are fully penalized
  5. Same for BIC

Found a great tutorial on GAMs from Gavin here which covers most of this.

EDIT: REML requires fully penalized fixed effects in AIC
Source Link
Nate
Nate
  • 2.1k
  • 8
  • 25

What I found:

  1. Select should be FALSE when comparing with AIC, but can be TRUE in "best" model for variable selection (as seen in every example I've found on this site).
  2. Models should be fit with ML when theunpenalized fixed effects (parametric terms) are differentmissing in one model vs. another (like herehere)
  3. REML is fine for comparing "nested" GAM(M) models in AIC with the samefully penalized fixed effects
  4. Same for BIC

OP w/ answer from Gavin

What I found:

  1. Select should be FALSE when comparing with AIC, but can be TRUE in "best" model for variable selection (as seen in every example I've found on this site).
  2. Models should be fit with ML when the fixed effects are different (like here)
  3. REML is fine for comparing "nested" GAM(M) models in AIC with the same fixed effects
  4. Same for BIC

OP w/ answer from Gavin

What I found:

  1. Select should be FALSE when comparing with AIC, but can be TRUE in "best" model for variable selection (as seen in every example I've found on this site).
  2. Models should be fit with ML when unpenalized fixed effects (parametric terms) are missing in one model vs. another (like here)
  3. REML is fine for comparing "nested" GAM(M) models in AIC with fully penalized fixed effects
  4. Same for BIC
Source Link
Nate
Nate
  • 2.1k
  • 8
  • 25