AIC Flashcards
(15 cards)
What is AIC?
AIC is a model selection tool that balances fit vs. complexity. The lower the AIC, the better.
It measures the information lost when we fit each candidate model to the data.
What is the equation for AIC?
-2log L + 2K
-2log L = measuring the fit of the model to the maximum likelihood parameter values
2K = penalising the model for the number of parameters that are included
What is step wise model selection?
Simplifying a ‘global model’ by removing parameters until we are left with a model that has the smallest possible AIC (best compromise between fit and complexity)
What is multi model inference?
Developing a set of candidate models either defined by an a prior (representing different hypotheses), or models that represent all possible combinations/subsets of predictors, ranking models using AIC
What is delta AIC?
The difference in AIC between the top model and all subsequent models
What is the akaike weight?
The relative probability that a model is the best explanation of the data among the given set of models.
What is the variance importance?
The sum of akaike weights for each variable in the model set
How do you extract parameter estimates from a model selection set?
Extract estimates from the best model
Model-averaged estimates (and standard errors): average estimates across models, weighted by their akaike weight.
What do you do to continuous variables if you are including all interaction terms?
Centre them
What do you do when there is no clear top model?
Model averaging
- Use AIC weights to average the coefficients across models
- Calculate confidence intervals for the coefficients
How do you compare coefficients for AIC?
Coefficients are interpreted in the same way as our glm/ms — but, if we standardise our predictors so they are on the same scale, we can interpret the relative size of the effects
What else do you need to check for with glm/ms?
Collinearity: VIF - less than 3-10 is good
Overdispersion: <1.5 (quasi poisson)
How can you deal with collinearity?
Remove the less biologically relevant variable
What do you do if you have a strongly weighted model?
You can use the estimates from the model (AICc, delta AICc, Akaike weights)
When do you add interactions and what should you do with them?
Center continuous variables if interaction terms are included. Only include interaction terms if they are needed.
