www.phidot.org

Forum,

I am currently debating using a model in order to find possible predictors or effects on the survival of GPS radio-collared ungulates that were translocated for restoration efforts. I have 2 years of data: 2 separate groups of animals (of the same species) were translocated from north Texas to west Texas during late-January and were equipped with GPS collars set to drop-off during late-November (300 days before drop-off post-release) of each year for 2013 and 2014, respectively. Each release site was spaced appr. 60 miles between each-other, and habitat was very similar. 59 animals were collared in 2013 and 62 were collared in 2014. We used aerial telemetry at least once per week for the 300 day period to monitor mortality (therefore all collared animals were encountered at least 1 time per week). GPS collars allowed us to determine appr. date of death. Therefore, we only monitored by flight to determine specific cause of mortality and locate animals.

After reading through the first 7 chapters of the Program MARK book, I learned a lot more than I expected to know about models, model selection, AIC, etc. Thus, I am still perplexed with the notion of even using MARK for estimating survival. I mainly want to find out if age (sub-adult, less than 4; adult, greater than or equal to 4), sex (Male or Female), study area, season (dry or monsoonal), or possibly daily/weekly rainfall, are predictors of survival and if using a model which of these effects play the largest role in survival of translocated animals. I know I will have to use a staggered-entry design, hence animals were collared and released within a 3 day period for each release year.

I have 3 main questions: (1) If I were to carry on with the previously mentioned question of survival, would I use the Known-Fate or the Nest Survival method in Program Mark; and (2) would going back to the ancient Kaplan-Meier staggered entry design (Pollock et al. 1999) or using some sort of ordinal logistic regression make more since with my question? To make a note on question (2), age and sex are the more important variables and to use precipitation or climate in my model I would just be pulling from NCDC data (in the given areas) for rainfall throughout the 300 day period (I didn't take rainfall or climate data during the study). And (3), Do I have enough data or variables to input in the model to do Known Fate or Nest Survival analysis?

Your comments and answers are greatly appreciated,

Taylor

tgar3881 wrote:Forum,

I am currently debating using a model in order to find possible predictors or effects on the survival of GPS radio-collared ungulates that were translocated for restoration efforts. I have 2 years of data: 2 separate groups of animals (of the same species) were translocated from north Texas to west Texas during late-January and were equipped with GPS collars set to drop-off during late-November (300 days before drop-off post-release) of each year for 2013 and 2014, respectively. Each release site was spaced appr. 60 miles between each-other, and habitat was very similar. 59 animals were collared in 2013 and 62 were collared in 2014. We used aerial telemetry at least once per week for the 300 day period to monitor mortality (therefore all collared animals were encountered at least 1 time per week). GPS collars allowed us to determine appr. date of death. Therefore, we only monitored by flight to determine specific cause of mortality and locate animals.

After reading through the first 7 chapters of the Program MARK book, I learned a lot more than I expected to know about models, model selection, AIC, etc. Thus, I am still perplexed with the notion of even using MARK for estimating survival. I mainly want to find out if age (sub-adult, less than 4; adult, greater than or equal to 4), sex (Male or Female), study area, season (dry or monsoonal), or possibly daily/weekly rainfall, are predictors of survival and if using a model which of these effects play the largest role in survival of translocated animals. I know I will have to use a staggered-entry design, hence animals were collared and released within a 3 day period for each release year.

Hey Taylor,
Ok, well, you are where most studies start and the fact that you have read the first several chapters is a good start. First thing you have to figure out is what your objective is. It seems that you want to estimate survival based off a candidate model set that is a function of your 'predictors' and see which candidate model best fits the data you collected, and then make inference from that/that combination (including model selection uncertainty) of 'highest ranked models conditional on the model set' to make some statement about survival. I am going to lay out a few things here and below that you will need think about/discuss with your major prof.

1) You collected 'weekly' data on fate via your sampling design, thus the minimum period for which you could estimate S is weekly. However what is a relevant period to estimate survival for your species (which I am inferring is pronghorn as that is the only thing moved in Texas in that large of a quantity lately)? Typically, the sampling period for your data needs to have at least 1 'event' in it (e.g., death of a individual in your case), else the parameter of interest (S) is 1.00 for that period (and is technically not estimable). So, if say very few of your individuals die, then you are going to have a whole lot of weekly survival estimates of 1.0, which does not tell you much. Use of a wider interval (say month for instance) is fine, and may make more sense. But, I don't know if moving your critters from N. Tx to W. Tx just causes them to fall over or what. So, first thing, decide on appropriate interval for estimating survival.

2) It seems as if most of your predictors are binary (age, sex, season, site) with one that may be a continuous covariate (precip). I assume you have read the individual covariates chapter, as 99% of what you are asking about is detailed in there. What your specific analytical targets are should be decided by your major prof and you.

3) You mention staggered design, but I don't think you are understanding what it is. A staggered entry design is where you are capturing and bringing individuals into your sample over time, working under the assumption that the survivorship function for any animal entering the sample at time t is equal to the survivorship function of individuals entering the sample at t. All of your individuals enter in a period <3 days (less than the length of time of you sampling intervals), so this is not anything you need to worry about.

I have 3 main questions: (1) If I were to carry on with the previously mentioned question of survival, would I use the Known-Fate or the Nest Survival method in Program Mark; and (2) would going back to the ancient Kaplan-Meier staggered entry design (Pollock et al. 1999) or using some sort of ordinal logistic regression make more since with my question? To make a note on question (2), age and sex are the more important variables and to use precipitation or climate in my model I would just be pulling from NCDC data (in the given areas) for rainfall throughout the 300 day period (I didn't take rainfall or climate data during the study). And (3), Do I have enough data or variables to input in the model to do Known Fate or Nest Survival analysis?

Your comments and answers are greatly appreciated,

Taylor

1) Based on what you have told us, known fate models seem the obvious choice.

2) Calling 'Pollock et al. 1999' ancient is a little rough on those authors as it provides the foundation for the known fate approach as you see it. But, to answer your question, you can get the same answers using the approaches in MARK, using the 'ancient' Pollock approach, or using ordinal logistic regression, if you know how they all work.

3) We have no idea if you have enough data, I mean, if all of them died, then survival for the period is 0, if none died its 1, otherwise we this goes back to my 1 above and it depends on your design and question. Do you have enough to estimate S, probably. As for variable(s), they really don't play a role in this. As for models, known fate is what you want based on what you said.

The analysis you are detailing is pretty straightforward with tons of examples in the book and elsewhere. Good luck.

\bret

Hi Bret,

Thanks for your response. Yes my study species is pronghorn. I will attempt to answer your questions and discuss the project in more detail.

1.) Here is a rough idea of the data. For 2013, we collared 59 pronghorn whereas 13 individuals died following release during the study. About 6 died during capture or at the release, so these individuals will not be used (right?). For 2014, we collared 62 pronghorn whereas 21 individuals died following release. 5 died at capture or at release. The only issue with using groups is sex; only 6 adult males were collared in 2013 and 4 adult males were collared in 2014. In 2014, an additional 6 fawns (which became yearlings half-way through the study) were collared. Most of the mortalities (>80%) occurred during the first 4-6 months after release for both study years. Therefore in the last 4 months of the study (until collar drop-off) only 1-3 more individuals died in both study years. Would using monthly intervals suffice for this?

2.) Our predictors will be a combination of age, sex?, season (dry versus wet), precipitation, average monthly or weekly temperature, and maybe study area. If I were to use monthly intervals, I am assuming I would using monthly precipitation and temperature estimates. Also, I can only use to climatic seasons (dry and wet) based on our study period: April to June (dry) and July to October (wet). How would I analyze this into the model if these seasons only occurred during a portion of the study?

3.) Yes you are correct, staggered entry would not be applicable.

4.) As for the ancient Pollock approach comment, I was only using sarcasm. I applaud and have used the Pollock approach.

5.) After speaking with my advisor, I will carry on with building models to predict survival for my study.

Thanks for your response, and any more help will be appreciated.

Taylor

tgar3881 wrote:Hi Bret,

Thanks for your response. Yes my study species is pronghorn. I will attempt to answer your questions and discuss the project in more detail.

Ok then, tell Louis I said hey.

1.) Here is a rough idea of the data. For 2013, we collared 59 pronghorn whereas 13 individuals died following release during the study. About 6 died during capture or at the release, so these individuals will not be used (right?). For 2014, we collared 62 pronghorn whereas 21 individuals died following release. 5 died at capture or at release. The only issue with using groups is sex; only 6 adult males were collared in 2013 and 4 adult males were collared in 2014. In 2014, an additional 6 fawns (which became yearlings half-way through the study) were collared. Most of the mortalities (>80%) occurred during the first 4-6 months after release for both study years. Therefore in the last 4 months of the study (until collar drop-off) only 1-3 more individuals died in both study years. Would using monthly intervals suffice for this?

Assuming you are just removing the obvious deaths due to capture and release, then I would think monthly intervals would work fine for the first 4-6 months, but if you only lost a couple of individuals in the later months then monthly intervals might still have estimates of 1. But, and I cannot answer this for your, does a 'monthly' interval tell you what you need to know from a management/scientific standpoint?

2.) Our predictors will be a combination of age, sex?, season (dry versus wet), precipitation, average monthly or weekly temperature, and maybe study area. If I were to use monthly intervals, I am assuming I would using monthly precipitation and temperature estimates. Also, I can only use to climatic seasons (dry and wet) based on our study period: April to June (dry) and July to October (wet). How would I analyze this into the model if these seasons only occurred during a portion of the study?

For weather stuff (precip), a monthly average or some other metric representing the month is fine. I would expect the response to be lagged for your species, meaning that rain in May impacts survival in June. For climatic season, you treat it like any other variable, so if for instance you think that survival is constant within a climatic period but different between climatic periods, then you can use the PIMS to set one group to 1 and the other to 2, lots of examples on this in MARKBOOK, you might look at the flood/no flood example in there.

3.) Yes you are correct, staggered entry would not be applicable.

Awesome.

4.) As for the ancient Pollock approach comment, I was only using sarcasm. I applaud and have used the Pollock approach.

my sarcasm filter was apparently turned the wrong way, applaud away.

5.) After speaking with my advisor, I will carry on with building models to predict survival for my study.

Thanks for your response, and any more help will be appreciated.
Taylor

Go forth and make science! If you hit a rough patch holler.

\bret

Bret,

Doc said hello back!

Sorry for the late response. I think using the same seasons (climatic seasons derived from a climate model) I used for (GPS collar) movement analysis would suffice for survival. However the last season has only 1 death. Acclimation period (1/28/13-3/31/2013): 7 deaths; Dry season (4/1/2013-6/30/2013): 4 deaths; and Wet season (7/1/2013-10/31/2013): 1 death.

Based on this, can I treat these 3 seasons as time intervals? Also, I would only have 3 intervals, but would that be sufficient for capturing and figure out predictions of survival?

Thanks!

Taylor

Oops, here are the data for 2014: Acclimation: 11 deaths; Dry: 8 deaths; Wet: 2. 2014 might seem to work better?

Thanks,

Taylor

tgar3881 wrote:Bret,

Doc said hello back!

Sorry for the late response. I think using the same seasons (climatic seasons derived from a climate model) I used for (GPS collar) movement analysis would suffice for survival. However the last season has only 1 death. Acclimation period (1/28/13-3/31/2013): 7 deaths; Dry season (4/1/2013-6/30/2013): 4 deaths; and Wet season (7/1/2013-10/31/2013): 1 death.

Based on this, can I treat these 3 seasons as time intervals? Also, I would only have 3 intervals, but would that be sufficient for capturing and figure out predictions of survival?

Thanks!

Taylor

Taylor,
Sorry slow reply, a sandy beach in FL with no internet called and I could not resist.

So, short answer is yes, season would work for estimating survival as you have at least 1 event (death) in each interval.

Now, would that be sufficient for estimating survival, dunno, that is between you and your committee. With such a large sampling period (couple months for each) you are only going to get a survival estimate for each period, so if that is enough, good.

If you have 2 years of data and the periods are the same, and if annual effects are not important you might be able to group the data over the 2 years and increase the value of the estimates. But, given you have some pretty significant differences between years in your study that might not be a good idea.

I am back if you have additional questions so my response will be quicker.

\bret

Hey Bret,

Sorry for the late response.

We have decided to go with the 3 seasons (Acclimation, Dry, and Wet) as our intervals. These seem to work nicely for estimating survival. My groups are: Sub-adult males, Adult males, Sub-adult females, and Adult females. Here is my data input:

known fate group=1;
1 0;
1 0;
1 0;
known fate group=2;
5 0;
5 0;
5 1;
known fate group=3;
12 1;
11 0;
11 0;
known fate group=4;
41 6;
35 4;
31 0;

I am now running into issues with figuring out the goodness of fit (GOF) for my global model.

I have developed my candidate model list (where you see season, that is the time interval of 3 encounter occasions): all equal (sex, age, season), sex only, age only, season only, age x sex, sex x season, age x season, age x sex x season.

In this case, the global model would be age x sex x season (12, all parameters included) correct? In my model list deviance is showing up as 0.000 for the global model and 0.767 x E006 in the output editor of results. In order to calculate GOF (or c-hat) I need to find the chi-square of the model divided by the degrees of freedom of the model, whereas this value should be less than equal to 3 right? From the output editor there is no given chi-square. Is this because the deviance is 0.000?

Sincerely,

Taylor

tgar3881 wrote:Hey Bret,

I am now running into issues with figuring out the goodness of fit (GOF) for my global model.

Sincerely,

Taylor

Taylor:

And that is because there are not GOF tests for known fate data (TM, Section 16.).

\bret