Hello list,
I'm trying to carry out a global sensitivity analysis using the
"sensitivity" package. I have a model with 26 paramters, mostly
normally distributed (with a few truncated to not go below 0). I've been
trying both the fast99 and sobol2007 functions, but having problems with both:
* with fast99, points seem to be generated at the minimum and maximum values a
distribution can take, which is fine for qunif, but causes problems with qnorm,
as it generates +-Inf values, which don't work with the model.
* with sobol2007, I've done runs using 2x500 point samples, resulting in 14k
runs. This gives quite wacky output, with both negative and very high values
appearing in the first order and total indices (and associatedly large
confidence intervals).
As far as I can tell, my options are:
* find a way to truncate the distributions for use with fast99
* increase the number of runs with sobol2007
Does anyone have any advice on which of these is more likely to work, or if
there is an alternative route which would be better?
Thanks!
Dave Murray-Rust
--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
