| Very good advice by <@U01UMBR0K38> Take some time to write your obs/states and transition/observation matrices on paper before working in R/Nimble. |
|
| NA |
OK, the war of memes is open :hugging_face: |
| NA |
:rolling_on_the_floor_laughing::rolling_on_the_floor_laughing::rolling_on_the_floor_laughing: |
| What about situations when the number of sites is large and not all of them were sampled to mark the individuals? |
Two questions in one I guess. If you have many sites, then it gets difficult to estimate all transition parameters, unless some transitions can be assumed never to happen and the transition matrix simplified. If some sites are not sampled in some years, then the corresponding detection probability can be fixed to 0, it will affect the estimation of survival/transition though. Does it help? |
| NA |
Yep! Thx. |
| NA |
This paper by a former student of mine suggests a solution when you have many sites to deal with
|
| NA |
If you're interested in, check out follow-up work by Hugo Cayuela. |
| NA |
There is also the possibility to include ‘unobersvable states’ we use it when we know birds/animals can be elsewhere but we can’t sample there, for eg wintering areas for birds |
| Are there obvious (or non obvious) advantages / disadvantages of the Dirichelet vs multinomial implementation that <@U01UMBR0K38> just showed us? |
When it gets to include covariates, the multinomial logit is easier to use. |
| Are dirichlet priors the priors of choice for multistate param? There is no such things as multi-logit? |
you can use the multinomial logit (it can be understood as logit for each category of the multinomial). It might be personal choice, but some people find it easier to use multinomial than Dirichlet when we want to use covariates. Does it answer your question ? |
| NA |
yep, thanks |
| Could it cause problems that the zinits==0 are assigned to either of the 2 sites outside of the initial.values function? It seems that this random assignation could potentially assign observations to implausible states, and this sampling process won't be regenerated (ie like the other initial values) when the model is rerun. |
When creating the zinits we want to make sure to avoid impossible transitions - which can be painful if you have several sites, and some impossible transitions among them... in this case, you have to generate the zinits making sure to avoid impossible transitions. But apparently Nimble has a way aournd this issue (see later Perry’s class about marginalization) |
| NA |
great, thanks |
| NA |
The idea is that with marginalisation, you get rid of the latent states in the likelihood of the HMM, so no need to initialize them anymore. NimbleEcology implements the marginalized likelihood for HMM, more in lecture 8 this afternoon with . More about nimbleEcology
|
| Can I ask why we need to specify initial state probabilities (i.e. delta)? I’ve been using CJS/multistate models in JAGS using base scripts from the Kery & Schaub 2012 book, which don’t require initial state probabilities to be specified, so I’m not clear on what we gain by doing so? Sorry if this is something I missed yesterday! |
Sorry for the confusion. We use delta as a placeholder for more complex models to come this afternoon in lecture 7. For future workshops, we might decide to forget about that as it generates confusion. |
| NA |
Ah okay! No problem :slightly_smiling_face: So we can in theory remove it in the models covered so far with no effect? |
| NA |
Sure. |
| shouldn't the (2,1) entry of the transition matrix not be phi? |
it corresponds to the probability of remaining alive and coming back from state 2 to state 1, so survive and back to state 1, phi * psi(comes in again) |
| Titi song for Chloe <https://www.youtube.com/watch?v=gF48b-FxZ8Y> |
hearing Manu’s bird songs is the best moment of every workshop :heart: |
| NA |
Random fact: In Filipino, "titi" means penis! :joy: |
| NA |
we are on a slippery slope here! |
| NA |
Omg that is fantastic. Birds are such weirdos!:exploding_head: |
| NA |
I purposely avoided saying titis because a little voice in my head goes "Haha, titties, titties, titties..." :sweat_smile: |
| Here is <https://oliviergimenez.github.io/bayesian-cr-workshop/slides/applistHMM.html|a searchable list> of HMM analyses of capture-recapture data. I'll refer to that list again in the last lecture <https://oliviergimenez.github.io/bayesian-cr-workshop/slides/applistHMM.html> |
Email me if you'd like me to add references to the list, it's far from exhaustive :wink: |
| NA |
yes please |
| Sorry if I missed this, is there a repository with all the datasets? |
Yes, from the website in the live demos pull-down menu, "Download all datasets at once". Direct link
|
| Will you keep traces of all questions/answers of each session ? |
Yes. We'll organise them in a FAQ section in the website. Also the Slack will remain open for some time after the workshop. |
| NA |
Thanks |
| In the last example, how would you test for a difference between breeders and non-breeder to 'confirm' a trade-off?! |
you can look at the estimates for both states and see they overlap, or you can monitor directly the difference between the two parameters and look at its posterior (do this by monitoring the difference as an additional parameter) |
| NA |
another option would be to compare a model constraining the survival to be equal among the two states, with a model in which survival is different among these two states, and see which one is best supported by the data |
| Random fact, we thought our focal species', eastern barred bandicoots, maximum age was about 3 years, then we translocated them to new sites, we captured one at 5 years old last year, approaching ages seen in captivity. |
Oh his name is Francois? :rolling_on_the_floor_laughing: Love it! |
| NA |
But about the topic: when I set the "last" occasion, I always am quite generous. Usually i make it something like: last observation + 2*max observed age |
| NA |
Yes, this is from my masters in an earlier translocation to an island called French Island in Victoria, I don't think the name has anything to do with the country, the bandicoots mostly had French names :flag-fr: |
| NA |
Very cute. I actually think we may have met at a conference not too long ago (BES?). I remember talking to someone about CMR of banded bandicoots... |
| NA |
:laughing: there's only me |
| One comment and a couple of questions. Much of what we have learned so far can be easily implemented in MARK/RMark or ESURGE. I am looking forward to learning about Bayesian magic that can cannot be done in a frequentist framework. Second, is there is framework for handling missing encounter history data? E.g., when a survey was missed. |
On that note I did show in class 5 how to implement time random effect which is hard in max. likelihood because of the integrals. More about the hard stuff coming in the afternoon :smiley: |
| NA |
See here about Bayesian magic:
|
| NA |
And here about missed surveys:
|
| NA |
Also, for more magic, wait for ’s lecture this afternoon! |
| Working on a multi event model with three disease states S (susceptible), A (acute recently infected), P (prior infection) (there is no recovery in this population). Using multi event since disease state can also be unknown. The transition matrix is very depleted as can only move forward and can’t jump ahead skipping a step. So only SS, SA, AA, AP, PP transitions exist. And psiPP = 1. So psiSS and psiSA= 1- psiSS. And psiAA and psiAP = 1-psiAA. So in this case I am concluding that I am only monitoring psiSS and psiAA and therefore am already restricting the transition probability from each state to 1 and do not need to use the dirichlet prior or multinomial logit link. Is my logic correct or am I missing something? |
Nop you’re correct ;-) |