Closed marksorel8 closed 4 years ago
I am receiving data on efficiency for individual days in the Wenatchee
Do you really have daily capture efficiency data? I thought there were only a few times during the outmigration season when WDFW conducted those experiments.
I have also used lifestage (subyearling vs yearling) as a covariate to capture some of the seasonal length and behavioral effects
To clarify, in this case you essentially model capture efficiency differently for different seasons corresponding to the outmigration windows for the different life stages?
What would be the best way to assess different capture efficiency models?
I would caution against some search for the "best" capture efficiency model, and instead I suggest writing out the full form for the model that represents the data in hand. For example, if you indeed have capture efficiency data on daily time steps, but with lots of missing data, then adding random effects for day and year will likely result in near perfect fits, as was the case previously.
@mdscheuerell ,
Do you really have daily capture efficiency data?
No, but the few mark-recapture efficiency trials that I do have are referenced to individual days. Of course those days only represent the first day when fish were recaptured, which I've been told is the day when the majority of fish are recaptured, but some fish are also recaptured on subsequent days. Unfortunately I don't have the data on how many fish were recaptured on each day after release, so I was planning on assuming that they were all recaptured in a single time period (day or week).
I suggest writing out the full form for the model that represents the data in hand
The main relevant covariates for capture efficiency are 1) lifestage (or equivalent), 2) discharge, 3) trap position (upper or lower) although I am missing trap position for some catch observations, and 4) year effects. There could certainly be interactions, but I suspect I would need to limit the complexity somewhat. Perhaps a penalized-complexity prior could be used to improve predictive power.
adding random effects for day and year will likely result in near perfect fits, as was the case previously
I recently realized that I was previously modeling the random effects on capture efficiency very incorrectly/incompletely, because I wasn't including the random effects for the capture efficiencies that were being used to predict catch in the model. In fact, I wasn't even using the same capture efficiencies when fitting the capture-efficiency model as when fitting the daily-migrant model. Oy! Assuming I was modeling it correctly, I could see the model-predicted capture efficiency differing from the data if, for example, there was a very high catch that could be explained by a combination of 1) process error in the daily migrant model and 2) higher than expected capture efficiency.
Now that I have realized that I was modeling capture efficiency incorrectly, maybe the new process would go something like this: 1) Fit the model using the full capture-efficiency model that represents the data in hand. 2) Compare the predicted capture efficiency to the data and assess whether there is overdispersion that needs to be dealt with. 3) Add random effects if necessary to model overdispersion.
Does that sound reasonable?
Hi @mdscheuerell and @ebuhle , I'm thinking about how to structure the model of trap capture efficiency for the Wenatchee. In @ebuhle 's Tucannon model, capture efficiency is modeled as constant within weeks, which is likely due to the data being aggregated within weeks, but I am receiving data on efficiency for individual days in the Wenatchee. Therefore, I could model efficiency for each day.
In @ebuhle 's Tucannon model, efficiency is modeled as a function of covariates with random effects for year and week of year (i.e. the random effect of week is the same across years as I understand it). I am guessing that the random effect for week of year is meant to capture effects of length and seasonal fish behavior.
I could adopt a similar approach for the Wenatchee, potentially using day of year as an alternative to week of year. But I have also used lifestage (subyearling vs yearling) as a covariate to capture some of the seasonal length and behavioral effects. An alternative is to use a random effect for day or week that is not constant among year, to model any overdispersion in the binomial model capture efficiency.
Obviously, the choice of capture efficiency model is not insignificant and will affect the rest of the model. What would be the best way to assess different capture efficiency models? Some sort of posterior predictive check?
Thanks, Mark