| Anyone working on CMR for otters? Otters are great! :heart: |
hopefully beginning next year ! |
| NA |
Oooh exciting :star-struck: |
| NA |
Indeeeeeeeeeeeeeed neighbour :slightly_smiling_face: |
| NA |
Giant otters in Dam lake in French guiana.. |
| NA |
Olivier, not neighbour for long time, coming back to French guiana in september |
| NA |
Bummer, I didn't know... |
| NA |
will come back in Montpellier often |
| I guess 'observation' here is synonym for 'event'? |
YES |
| NA |
you define events based on your field observations |
| NA |
Uh oh, going all caps now :stuck_out_tongue: |
| NA |
hahaha on :fire: |
| NA |
call your firemen guysss |
| NA |
hahaha |
| NA |
:rolling_on_the_floor_laughing: |
| What if the individual is a breeder and I mistakenly mark it as a non-breeder ? Why this probability is 0? |
I'm not quite sure, but I think is that model design there was not account for the error of mistakenly marking a breeders as non-breeder. I think it should be needed another parameter/matrix. But I don't know a lot about this, I am just starting ^^U |
| NA |
*in that model |
| NA |
yeah... model knows betaB and betaNB parameters... but it seems that we might assign the status correctly or we might not be sure... but we can't make a mistake. I just wonder if this is a simplification assumption? |
| NA |
it's possible to set an uncertain state at the first capture. It will be necessary to write the initial probs with a state assignment probability |
| NA |
Andjela, do you mean if an individual is considered non-breeder (because it was not breeding during capture occasion) but is in fact breeder during the same season but few days later? If it is your question sorry... i don't have the answer but i also ask to myself how to consider that. |
| NA |
In my case, for example, you may assign the status 'non breeder' for birds that are not nesting in a nest box. However, in our study area there are few natural cavities in which a little proportion of our 'non-breeders' may be actually breeding, so the assigment is not perfect. |
| NA |
Yeah.. or now... we study healthy/ill birds... we can clearly catch a bird that is ill and we don't notice... and we mark it as healthy. There is a "measurement mistake" let's say.
It seems in this model it is not considered. Or I missed something. |
| NA |
any thoughts on this? :sweat_smile: |
| NA |
You can include ‘error of assignment’ for this you would need to decompose the observation matrix in several steps 1° prob of detection 2° prob of correct status assignement, check the list of HMMs applications, there are some examples very similar to your question |
| NA |
:smiling_face_with_3_hearts: thank you! |
| NA |
Here is the link to the HMMs examples :
|
| NA |
Ok thanks. |
| random question: where does the name multi-event come from to describe uncertainty on state assignment? why is that a multi-event? |
I approve of this question ! No idea, and not necessarily intuitive… where does it come from ? |
| NA |
Roger has the answer: |
| NA |
|
| NA |
Cool - so that is the rationale, stolen from the pdf: "The main aim of this article is to show how uncertainty in the assessment of state can be incorporated into the analysis of “multistate” CR data. The solution I propose puts an emphasis on the real nature of the data. Conceptually, it is not states that are observed but rather something, say an “event,” which reflects to some extent the underlying state, which is the ultimate object of the study. The new models are called “multievent” to reflect the nature of the data." |
| NA |
I think that this is what Roger Pradel had in mind: 1 true state->2 possible obs (events, like in ESURGE) if the individuals is resighted/recaptured. |
| Can these multievent models also deal with different types of observation? Could you combine visual e.g. resight and recapture data or would this break the sampling assumption of the model? |
Yes, you can include several events, eg resighted, recaptured, then even several more steps because you cannot always make all the measurements at each capture |
| NA |
so you can even decompose the observation matrix in several steps |
| NA |
For combining different types of observations, i.e. resightings and captures, you do not need a multi-event model. This can be included into a multi-state model, or even a CJS model. It will be an extension of your observation matrix. |
| NA |
But, if there is uncertainty in those observations, e.g. you cannot assign breeding state with resightings, only with captures, then you need a multi-event. |
| NA |
multievent is useful if one type of obs leads to some uncertainty in status, eg if you can’t know sex from resight but only from recaptures that would lead to unknown status :smiley: |
| NA |
That's great thanks :slightly_smiling_face: |
| again curiosity: I wonder why would the survival be higher for infected individuals (from an ecological point of view I mean?) |
I don’t know this particular study but maybe an element of answer here ?
|
| NA |
Thanks . (when you happen to take a break from talking, I wonder if) do you happen to know in this specific case ! |
| NA |
I think so - but also I don't think it necessarily kills them directly (in the short term at least which I think this dataset is) but it does prevent them from doing much else. |
| NA |
I've read some of the papers as I'm working on a vaguely similar system but can't remember the specifics off the top of head! |
| NA |
Faustino et al. (2004) and Conn & Cooch (2009) concluded that disease is associated with lower survival but they reached this conclusion based on model-averaged estimates. Their papers deal with weekly survival probability, and results differed according to year. I don't know which year is included in this workshop example, but in one year the Cornell team also found higher survival in infected birds... |
| NA |
It's the year with fewer data, and the results are not very convincing I'd say. Conn and Cooch got more reliable results with the other year, which we don't have. I might ask Paul or Evan :slightly_smiling_face: |
| About the finches : I thought the probability of remain sick was 1 ? So why recovery rate is 46% ? Did I miss something ? |
I think the example in the coding tutorial assumes that ill individuals can become healthy again. |
| NA |
If you check in the transition matrix (again!), you can see that it is not 0 in that example :slightly_smiling_face:
That being said: Olivier also talked about a case with an in-curable disease. In that case, you are absolutely correct that recovery rate should be set to 0. |
| NA |
Sorry for the confusion. Recovery is possible in the house finch example. I showed the transition matrix for an incurable disease for the sake of illustration only. |
| NA |
ok. Thank you |
| In the housefinch illness example, didn't we assume that the prob. of transition between ill and health is zero? Why do we have estimates of psiIH in the last row of the results? I think I am missing something. |
Hey Fabi :wave: you’re right there might be a problem here..maybe the results were from the model from the live demo assuming individuals can recover an idea ? |
| NA |
See answers above:
|
| NA |
Sorry for the confusion. Recovery is possible in the house finch example. I showed the transition matrix for an incurable disease for the sake of illustration only. |
| Is there a preferred/standardised way to present outputs from these models for paper writing? I have been looking at how recent papers have presented them but as we're discussing things thought I'd ask! |
Good question, not that I am aware of... nice to have the posterior distrib of important parameters somewhere, or at least a summary of it. |
| NA |
There are no “official” guidelines. How you present results should depend on the question you want to answer / result you want to show.
Two things I think are important:
1. ALWAYS plot uncertainty
2. Have plots of posterior distributions, and a table with summary stats, of all important parameters in the Supplementary |
| NA |
Thank you both ! |
| NA |
Also, in-text we typically present results as either:
mean [95% CI]
or
median [95% CI] |
| NA |
Depending on your results, you may also want to use a different CI (credible interval). |
| NA |
Hi, a short follow-up question on the one from Katherine: If one of the parameters of interest is for instance phi[k,t,i] where k=9 states, t=10 time occasions, i=1000 individuals, would you need to provide traceplots & Rhats for all combinations (9x10x1000)? Or would an alternative be to select randomly 20 combinations of phi and show those only? |
| Thank you :slightly_smiling_face: |
here we considered two alive states - not sure what example it was but if it’s the wolves it might refer to dominant and subordinates. So here pi is the probability for an individual to belong to state dominant, 1-pi to belong to subordinate |
| NA |
Thank you! I see it now |
| NA |
Just bear in mind that the interpretation of the classes of heterogeneity is made a posteriori. The model doesn't know about dominant/subordinate, just alive in class 1 or class 2. |
| NA |
If p1 is telling us that the detection probability for individuals in class 1 is 0.38, pi is telling us the same, i.e., that a 62% of individuals belonging to class 1 is being non-detected? I am not sure if I got it right |
| NA |
pi is telling us that 62% of individuals belong to class 1 so they have 0.38 detection rate |
| NA |
which also means here that 0.38% of individuals belong to class 2 with 0.5 detection rate |
| NA |
Ok Sarah, thank you. I am just a bit confuse by the fact that the detection probability for ind at class 1 is 1 - the probability for an ind to belong to class 1 |