Model selection

(Jan 28, 2020)

When fitting a mortality model, analysts are faced with the decision of which risk factors to include or exclude.  One way of doing this is to look for the improvement in an information criterion that balances the fit against the number of parameters.  The bigger the improvement in the information criterion, the more strongly the model with the smaller value is preferred.

One natural question is "how big an improvement is significant?".  In mortality models using individual lives we often use Akaike's Information Criterion (AIC), and in Macdonald et al (2018, page 98) we wrote that a difference of 4 AIC units could be regarded as a threshold.  But where did the number 4 come from?

The answer comes from the conceptů

Read more

Tags: information criterion, AIC, relative likelihood

Find by key-word


Find by date


Find by tag (show all )