AIC and BIC

Question 1

AIC and BIC (model comparison)

Answer 1

used for non-nested and nested models
combines information about sample size, number of model parameters & residual sums of squares
lower values == better and include a penalty for the number of predictors in the model (BIC is harsher)
AIC has no cut-offs, BIC uses a difference of 10 to suggest one model is better than other

Question 1

AIC & BIC (deviance)

Answer 1

AIC = deviance + 2p
BIC = deviance + p log(n)
n = sample size, p = number of regression coefs in model
smaller values preferred
BIC favours models with fewer regression coefs

Question 2

parsimony corrections

Answer 2

the penalising of models for being too complex
help us avoid overfitting

Question 3

overfitting (model comparison)

Answer 3

adding arbitrary predictors to try and improve model fit