Chris Jackson

That works for me, using `Q

@Heart-wang A reproducible example would help others to give advice. Though please first try some of the tricks above in the thread, and in the "Convergence failure" section of the...

Thanks for the reproducible example. The problem there is that the model is not identifiable. The consequence is that the optimiser is exploring a log-likelihood surface that is a flat...

I expect Bayesian approaches would help if the data are weak, but only if you have substantive information to go into the prior. I'm not aware of any package that...

@HaniyehDanesh This model is almost certainly not identifiable from the data. You are allowing transitions between every state, assuming the state is intermittently observed, and assuming the states are misclassified....

@HaniyehDanesh As I said, please work through the [course notes](https://chjackson.github.io/msm/msmcourse/) to understand all this better. Consider what transitions in continuous time can happen, and consider what your observations tell you...

Hi Manuel - Thanks for this work. I'm unsure about putting it in though, because it results in a lot of duplicated code, and it makes the resulting function very...

It is just a different parameterisation of the same model. `flexsurv` parameterises all distributions in a way that is consistent with the distribution functions (such as `dexp`, `dweibull`) in base...

@mattwarkentin would you be able to look at this, as the original author? I have no idea about the pros and cons of this.

Thanks both, and sorry for leaving this issue so long! I would be happy to go with this suggested change, so feel free to send a PR @hfrick .