I'm maximising a reasonably complex function using nlme (version 3.1-65, have also tried 3.1-66) and am having trouble with fixed parameter estimates slightly away from the maximum of the log likelihood. I have profiled the log likelihood and it is a parabola but with sum dips. Interestingly changing the parameterisation moves the dips around slightly. Unfortunately the PNLS step is finding a maximum at the dips rather than the mle. I have tried using starting values for the fixed parameters without change. Any ideas ? Ken
Ken Beath <kbeath <at> efs.mq.edu.au> writes:> > I'm maximising a reasonably complex function using nlme (version > 3.1-65, have also tried 3.1-66) and am having trouble with fixed > parameter estimates slightly away from the maximum of the log > likelihood. I have profiled the log likelihood and it is a parabola > but with sum dips. Interestingly changing the parameterisation moves > the dips around slightly. Unfortunately the PNLS step is finding a > maximum at the dips rather than the mle. I have tried using starting > values for the fixed parameters without change. Any ideas ?Ken, you should not use nlme for "maximising a complex function", because it's a rather specialized tool for mixed-model statistical analysis. Try to use optim directly, which has quite a few methods to choose from, and one of them might work for your problem. Dieter
I meant fitting not maximising, it is a nonlinear mixed effects model, with both fixed and random effects. My assumption is that for the function I am using the approximation approach used in nlme is not quite close enough, and nothing much that I can do, except for looking at starting values. I was hoping that someone would have other suggestions, so I will keep attempting to understand the control parameters. I can add an extra parameter to the model and obtain a worse fit. Ken Dieter Menne writes:> >> >> I'm maximising a reasonably complex function using nlme (version >> 3.1-65, have also tried 3.1-66) and am having trouble with fixed >> parameter estimates slightly away from the maximum of the log >> likelihood. I have profiled the log likelihood and it is a parabola >> but with sum dips. Interestingly changing the parameterisation moves >> the dips around slightly. Unfortunately the PNLS step is finding a >> maximum at the dips rather than the mle. I have tried using starting >> values for the fixed parameters without change. Any ideas ? > > Ken, > > you should not use nlme for "maximising a complex function", > because it's a > rather specialized tool for mixed-model statistical analysis. Try > to use optim > directly, which has quite a few methods to choose from, and one of > them might > work for your problem. > > Dieter >