Problems with convergence

[rsb21]

I attended the mark workshop this june so i am terribly sorry to be bothering people so soon...

I am currently trying to run a multistrata live encounters only model with 14 states (yes i know this is a lot!).

The model i wish to run has 171 parameters with 2 of them fixed. I realise that this model will take a long time to run given its complexity, however at the moment having waited an awfully long time it fails to converge. I know that the model is a reasonable model (it has successfully been fitted with other software) and it is biologically reasonable and I have triple checked that the model structure has been correctly enetered.

I presume the problem is that the likelihood must be flat in a number of the parameters. Is there anyway of investigating why this is happening and is there any way I can overcome this problem?

Many thanks, Rachel

[cooch]

I presume the problem is that the likelihood must be flat in a number of the parameters. Is there anyway of investigating why this is happening and is there any way I can overcome this problem?

There are a couple of things you could/should try:

1. using starting values - run a reduced parameter model first, and then use parameter estimates from that model to 'seed' your more paramterized model.

2. use the simulated annealing approach for maximizing the likelihood (rather than the default numerical method).

Both are described in the 'sidebar' beginning on p. 30 of Chapter 9 (the most recent version of Chapter 9, which is online at

http://www.phidot.org/software/mark/docs/book/