next up previous contents index
Next: ANALYSIS PHASE Up: NOTES Previous: WORK CODE   Contents   Index


See Kao, Zeng and Basten (1999) for more detailed information on the information criteria. We use

        IC(k) = -2(log(L) - k c(n) / 2)

where L is the likelihood for a k-parameter model and log is the natural log function. The penalty function c(n) takes one of six forms:

1. c(n) = log(n)
2. c(n) = 2
3. c(n) = 2 log(log(n))
4. c(n) = 2 log(n)
5. c(n) = 3 log(n)
6. c(n) = 0

Use the numbers above with the -S option to indicate which information criterion you want to use. If you use penalty functions 1 through 5 above, then you should also specify a threshold of 0.0 with the -L function. Penalty function 6 is equivalent to no penalty function and requires an experimentwise threshold value that might be obtained via a permutation test.

Christopher Basten 2002-03-27