Problem Covariable Rmark

[bfol]

Hi everyone,

I'm pretty novice in Rmark and I have to use it to estimate nesting survival. My dataset is 6 years (2008-2013). We found 539 nests and we have 20 variables (habitats, meteorological ect ...). However, I have a problem when I try to include meteorological variables (= covariables). These variables are not taken into account. I don't know if a covariable need to have a specific parametrization ? I have for example one measure of temperature, of rainfall or water level per year. But when I try to include one or more covariables, I have this warning :

Note: only 9 parameters counted of 10 specified parameters
AICc and parameter count have been adjusted upward

Moreover, I tried (just to see the consequences) to change these covariables in individual covariables (multiplicate by a number distributed by a normal distribution (mean = 1, var = 0.001) ). Even if I choose a low variance, covariables are not taken into account. So I am a little bit lost ...

Does anybody had this problem ?
I hope anyone may help me, thanks a lot

[jlaake]

I'm not sure what you mean when you say "These variables are not taken into account". You didn't provide any details from the results so it is hard to tell. The message you got simply means that MARK found that one of the parameters is either at a boundary or is confounded with another parameter. Please provide the code for the call to the mark function and the results of summary(model) where model is the name of your model object. Please read the newest RMark Workshop notes at http://www.phidot.org/software/mark/rmark/RMarkDocumentation.zip.

--jeff

[bfol]

Thank you very much for your answer.
I begun to read the workshop notes.

My dataset is named "Milouin". The Year variable is declared as a factor.

This is my R code :

Code: Select all

load("donnees_Rmark5.0.RData")
milouin$Year=factor(milouin$Year)

run.milouin=function()
    {
Model1=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~Year + Predation + Water20 + Willow + Meadow20)),
                groups=c("Year"))

Model2=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~Year + Predation + Water20 + Willow + Meadow20 + Precipitation)),  groups=c("Year"))

Model3=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~Year + Predation + Water20 + Willow + Meadow20 + Water_level)),  groups=c("Year"))

Model4=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~Year + Predation + Water20 + Willow + Meadow20 + Temperature)),  groups=c("Year"))

 return(collect.models() )
}

milouin.results=run.milouin()
milouin.results



The Model 1 is my best model. When I try to add my weather variables (one value for each year), Model 2, 3 and 4 have the same deviance than the Model 1. But I have the warning "Note: only 9 parameters counted of 10 specified parameters _ AICc and parameter count have been adjusted upward
So that is why I thought my weather covariables were not taken into account.

However, when I only try to test the weather variables separately, parameters are well estimated

Code: Select all

Model5=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~ Temperature)),
                groups=c("Year"))

 Model6=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~ Precipitation)),
                groups=c("Year"))

 Model7=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~ Water_level)),
                groups=c("Year"))

Model8=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~Year + Temperature)),
                groups=c("Year"))

 Model9=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~Year + Precipitation)),
                groups=c("Year"))

 Model10=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~Year + Water_level)),
                groups=c("Year"))

 Model11=mark(milouin,nocc=386,model="Nest",
                model.parameters=list(S=list(formula=~Year)),
                groups=c("Year"))


Model5, Model 6 and Model 7 have a different deviance (respectively 1178.810, 1165.858 and 1185.273).
If I try to add the Year effect: Model8, 9 and 10 have the same deviance than Model 11.

model npar AICc DeltaAICc weight Deviance
7 S(~Year) 6 1102.308 0.000000 4.758550e-01 1090.296
1 S(~Year + Temperature) 7 1104.312 2.003914 1.747150e-01 1090.296
2 S(~Year + Precipitation) 7 1104.312 2.003914 1.747150e-01 1090.296
3 S(~Year + Water_level) 7 1104.312 2.003914 1.747150e-01 1090.296
5 S(~Precipitation) 2 1169.859 67.551139 1.020778e-15 1165.858
4 S(~Temperature) 2 1182.811 80.502939 0.000000e+00 1178.810
6 S(~Water_level) 2 1189.275 86.966839 0.000000e+00 1185.273


I don't know why weather variables cannot be estimated. Do you have an hypothesis ? Or is it my mistake ?
Thank you very much for your time

Ben

[Dan_Gibson]

I believe your problem is related to the fact you are trying explain annual variation in your data with a covariate in a model that has already assigned all the annual variation to the year factor. I think this pitfall is highlighted somewhere in the Program MARK manual, but I forget where.

If you are interested in the relative explanatory power of your weather variables, I believe the current recommended approach is to do an Analysis of Deviance, which essentially provides an estimate of the variance explained by your covariate relative to the total amount of variance explained by time. This feature is built into MARK, but I am not sure if it is currently in RMARK.

-Dan

[jlaake]

Dan is correct. Because there is only a single value of weather for each year and you have included a Year factor variable, there is no contrast for the weather variable to explain. If you had different areas within each year that experienced different weather, then you could have included weather and Year. You got the message because the weather covariate is completely confounded with Year and it cannot be estimated. You can use a formula with weather if you drop Year. These become competing models.

--jeff

[bfol]

Indeed ! I also thought of this mistake this morning. I started again my analysis without "year" effect when I wanted to test my weather variables and it worked well !
Good idea for the ANODEV !

Thanks a lot Jeff and Dan for your answers !

Ben