www.phidot.org

I'm trying to integrate concurrent telemetry data into a capture recapture dataset. When fitting without telemetry data, estimates from the most supported model seem reasonable but when using the dataset where telemetry is added, I get nonsensical estimates. I am curious if I am doing something wrong or if it's an artefact of the data (relatively small dataset). Estimates from RPSV are the same for both capture histories (with and without telem). This doesn't seem like it should be the case- should it? I have adjusted telemetry scale multiple times and altered estimation method to no avail. The summary appears to be correct. Any suggestions?


Object class capthist
Detector type multi (6), telemetry
Telemetry type concurrent
Detector number 110
Average spacing 136.2351 m
x-range 305198 307255 m
y-range 3419021 3421681 m

Usage range by occasion
1 2 3 4 5 6 7
min 0 0 0 0 0 0 0
max 1 1 1 1 1 1 1

Counts by occasion
1 2 3 4 5 6 7 Total
n 4 3 6 20 4 8 17 62
u 4 3 6 19 0 7 11 50
f 39 10 1 0 0 0 0 50
M(t+1) 4 7 13 32 32 39 50 50
losses 0 0 0 0 0 0 0 0
detections 4 3 6 20 4 8 71 116
detectors visited 2 2 1 4 1 3 0 13
detectors used 110 110 110 110 110 110 0 660

Empty histories : 11
17 telemetered animals, 6 detected
1-5 locations per animal, mean = 4.18, sd = 1.42

Individual covariates
Age Sex
A:28 F:24
J:22 M:26

Code: Select all

rm(list=ls())

#load in packages
library(secr)
library(rgdal)
library(sp)
library(raster)

options(scipen = 999)


#Read trap files
fall2023_detectors<- read.traps(file = "RCR_fall_2023trapsites.txt", detector = "multi", binary.usage = TRUE)
# Get a summary from trap files
summary(fall2023_detectors)

#read in telemetry data
CHt <- read.telemetry("RCR_fall23_telem.txt", verify = T, covnames = c("Age", "Sex"))
head(CHt)
summary(CHt)

#Read in captures file
fall2023_captures<-read.delim(file = "RCR_fall23_captures.txt")
fall2023_captures
bobCH2023<-make.capthist(captures = fall2023_captures, traps = fall2023_detectors)
summary(bobCH2023)

#merge capture data and telemetry data
mytelem <- addTelemetry(bobCH2023, CHt, type ="concurrent", verify = T, collapsetelemetry = T)
summary(mytelem)


###############
#Sigma estimate -- these are the SAME?
initialsigma_telem <- RPSV(mytelem, CC = TRUE)
cat("Quick and biased estimate of sigma =", initialsigma_telem, "m\n")

initialsigma <- RPSV(bobCH2023, CC = TRUE)
cat("Quick and biased estimate of sigma =", initialsigma, "m\n")


#Mask
centroids <- data.frame(t(sapply(telemetryxy(CHt), apply, 2, mean)))
tmpxy <- rbind(centroids, data.frame(traps(bobCH2023)))
mask2 <- make.mask(tmpxy, buffer = 4 * initialsigma, type = 'trapbuffer', nx=32)

# fit models
## THESE DO NOT FIT
model.HHN.2023_telem<-secr.fit(mytelem, trace=T,groups= c('Age', 'Sex'),
                               detectfn= "HHN", mask=mask2)
model.HEX.2023_telem<-secr.fit(mytelem, trace=T, groups= c('Age', 'Sex'),
                               detectfn= "HEX", mask=mask2)

# # # baseline detection function results
det_results_2023_telem<-AIC(model.HHN.2023_telem, model.HEX.2023_telem)
det_results_2023_telem
model.HEX.2023_telem


#Baseline Detection Function Models- No telem-- THESE FIT FINE
model.NEGEX.2023<- secr.fit(bobCH2023, groups= c('Age', 'Sex'), detectfn = "EX", trace=T, mask = mask2)
model.HR.2023 <- secr.fit(bobCH2023, trace=T, groups= c('Age', 'Sex'), detectfn= "HR", mask = mask2) #can't be used with telem data
model.HN.2023<-secr.fit(bobCH2023, trace=T, groups= c('Age', 'Sex'), detectfn= "HN", mask = mask2)
# baseline detection function results
det_results_2023<-AIC(model.NEGEX.2023, model.HR.2023, model.HN.2023)
det_results_2023



I'm sorry nothing came to mind immediately.

As you discovered, RPSV does not attempt to combine regular detections and telemetry data (it reports just the detection component of the combined capthist), although it will summarise a telemetry capthist alone. There may have been a reason - I don't remember - but this should at least be documented.

I've always found telemetry data difficult to fit in an SECR context. A possible reason that occurred to me was that we attempt to fit a single sigma common to all individuals, when the telemetry data (usually) provide overwhelming evidence of differences that really should be allowed via a random effect. Given some time and motivation I could include this in the likelihood, but until then, if it really matters, you may have to go Bayesian.

Murray

Thanks for the reply, Murray. I've fitted closed, non-spatial models using both MLE and bayesian approaches and get similar estimates as those estimated in secr. I might try the Bayesian spatial approach next. Not sure the telem will add much information and make a substantial difference but it would be a good exercise nonetheless.

Great. I would be interested to hear if including a random effect for sigma (in the Bayesian approach) delivers sensible results for the combined data that are not available from the vanilla ML approach. That would be good motivation.