Week 6 Flashcards
if model SS> residual/error SS
good model if p also <.001
hierarchical compression
special care of model comparison when the two models being compared are nested
nested models
one includes all the variables of another model
model specification
picking what goes into the model including form (linear vs non linear) and what parameters to icnlude
log likelihood
most models work by maximising the LL
trying to find a peak that uses parameters that will maximise the likelihood of getting observed samples
Higher LL= better goodness of fit of that model
LL in linear models
negative LL is often SS deviations
LL provides a direct measures
how likely is it that we could observe these data given some parameter estimate?
AIC and BIC
measures of a good model that includes a penalty for model complexity
help selevt best model
lower scores are better, pick the model (M1, M2) with lowest AIC/BIC
best model=
highest likelihood and fewest number of parameters/more parsimonious
AIC
akaike information criterion
2k-2LL
BIC
Bayesian information criterion
In(N)k-2LL