RE: $OMEGA blocks and log-likelihood profiling

From: mark.e.sale@gsk.com Subject: RE: [NMusers] $OMEGA blocks and log-likelihood profiling Date: Tue, June 1, 2004 6:55 am Nick, I like your suggestion that bootstrap is a better way to determine CI than LLP. The other advantage is speed, one set of bootstrap samples will give you CI for all parameters, rather than need to do a set for each parameter. But a question (or two). I know you've been an advocate that the covariance step success isn't always necessary to believe the results (and I'm beginning to agree with you). So, I assume you'd be willing to include NONMEM runs that failed covariance step in your bootstrap interval. But, what about other failures? Do you include termination due to rounding error? How do you protect against local minima? And then there are the ones that simply crash. We do this a lot as well (so far we have discarded the runs that fail covariance step), but are concerned that the runs that fail are the very ones that we're most interested in (the ones that define the tails where one or more parameters is poorly estimated), so eve if they are only 2%, they may be critical in defining the CI. Mark