1.Rémi wrote:- First this feature works only conditional on the estimates of our model.
Have you done a model before calling this tool ?
Yes, I have. I have tried also by first selecting the just run model in the OUTPUT space and after press the Count transition numbers and it still gives me the same error.
Anyway, if it is conditional on the estimates of one specific model perhaps I haven't understood what it is counting. My idea was that it counts the number of known transitions between states in the dataset, i.e. how many times in the whole dataset there are two contiguous cells in the same row like, say, 12.
This would give a snapshot of how frequent the transition is in the dataset and, consequentially, how good the software should be to estimate that transition. I looked for a similar tool to decide if a priori avoid modeling time variation in the transition rates but this lead to the second point of the above post.
Meanwhile, if someone might be interested in it, there is a way to count those transitions (in the sense I thought) in an excel sheet by doing something like this:
Where "SUMAPRODUCTO" is "SUMPRODUCT" in the english version.
In case of a multievent data set where you have an event meaning state unknown (my case) and one want to see how many times one individual is seen in state 1 and after (no matter how much after) is seen in state 2 you could eliminate all the zero and unknown event, shift the cells leftward, fill the empty cells with zero and repeat the same procedure as above.
2.Rémi wrote:- Secondly, the best to test for a transition to be zero is to make the two models
with and without this transtion to zero and to do an LRT test knowing that in that
case the difference of the two deviances follow a mixture of khi-square.
And, after this, in case I get a positive result (
psi i=0), should I fix to zero that transition rate in order to improve the optimization process?
3. This is a very important point to my analyses. I have been checking how many known transitions I had in my data set and I have seen that they are very few, just 3-6 if I consider contiguous cells and 3-13 if I consider the soon or later seen in the other state - I have two real states. For the rest I have more than 1000 individuals captured at least once, 196 captures in one state, 232 in the other, 2293 in unknown state.
Even if a priori this sounds dramatic to me, I have run my analyses (by avoiding time variation on transition rates) and (i) I get no troubles with convergence, (ii) I get no troubles with parameters estimability, and (iii) I get more or less acceptable estimates (in terms of CIs) of almost all the biological parameters in my more parsimonious models. Particularly I have been impressed from the fact that I get acceptable estimates (again in terms of CIs) of the transition rates.
Does it make sense? May I feel comfortable with these analyses?
4. I have two steps for both transition and event, and also I have uneven time intervals, should I set
Setting/Unequal time interval: # of steps/ 1 2 (more than just upload the file with the time intervals)?
Thanks,
Simone